From 10bb545c5b54175ed9874ad8d187effa2bcb4b5f Mon Sep 17 00:00:00 2001 From: 0cc4m Date: Thu, 19 Jun 2025 09:15:42 +0200 Subject: [PATCH 01/43] Vulkan: Set device max size for host memory to avoid OOM warning and fallback to CPU buffer (#14249) --- ggml/src/ggml-vulkan/ggml-vulkan.cpp | 8 +++++++- 1 file changed, 7 insertions(+), 1 deletion(-) diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp index 8d62303aa..1375bfeb9 100644 --- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp +++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp @@ -9495,6 +9495,12 @@ static size_t ggml_backend_vk_host_buffer_type_get_alignment(ggml_backend_buffer UNUSED(buft); } +static size_t ggml_backend_vk_host_buffer_type_get_max_size(ggml_backend_buffer_type_t buft) { + return vk_instance.devices[0]->suballocation_block_size; + + UNUSED(buft); +} + // Should be changed to return device-specific host buffer type // but that probably requires changes in llama.cpp ggml_backend_buffer_type_t ggml_backend_vk_host_buffer_type() { @@ -9503,7 +9509,7 @@ ggml_backend_buffer_type_t ggml_backend_vk_host_buffer_type() { /* .get_name = */ ggml_backend_vk_host_buffer_type_name, /* .alloc_buffer = */ ggml_backend_vk_host_buffer_type_alloc_buffer, /* .get_alignment = */ ggml_backend_vk_host_buffer_type_get_alignment, - /* .get_max_size = */ NULL, // defaults to SIZE_MAX + /* .get_max_size = */ ggml_backend_vk_host_buffer_type_get_max_size, /* .get_alloc_size = */ ggml_backend_cpu_buffer_type()->iface.get_alloc_size, /* .is_host = */ ggml_backend_cpu_buffer_type()->iface.is_host, }, From faed5a5f5dde4d816adecdccb4f4682a04126f92 Mon Sep 17 00:00:00 2001 From: Aaron Teo Date: Thu, 19 Jun 2025 17:48:54 +0800 Subject: [PATCH 02/43] llamafile : support s390x SIMD instruction set (#14273) --- ggml/src/ggml-cpu/llamafile/sgemm.cpp | 55 ++++++++++++++++++++++++++- ggml/src/ggml-cpu/llamafile/sgemm.h | 5 +++ 2 files changed, 59 insertions(+), 1 deletion(-) diff --git a/ggml/src/ggml-cpu/llamafile/sgemm.cpp b/ggml/src/ggml-cpu/llamafile/sgemm.cpp index 1c545f803..7ed3874af 100644 --- a/ggml/src/ggml-cpu/llamafile/sgemm.cpp +++ b/ggml/src/ggml-cpu/llamafile/sgemm.cpp @@ -62,7 +62,7 @@ #define NOINLINE __attribute__((__noinline__)) #endif -#if defined(__ARM_NEON) || defined(__AVX512F__) +#if defined(__ARM_NEON) || defined(__AVX512F__) || defined(__VXE__) || defined(__VXE2__) #define VECTOR_REGISTERS 32 #else #define VECTOR_REGISTERS 16 @@ -109,6 +109,12 @@ inline float16x8_t sub(float16x8_t x, float16x8_t y) { return vsubq_f16(x, y); } inline float16x8_t mul(float16x8_t x, float16x8_t y) { return vmulq_f16(x, y); } #endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC +#if defined(__VXE__) || defined(__VXE2__) +inline float32x4_t add(float32x4_t x, float32x4_t y) { return vec_add(x, y); } +inline float32x4_t sub(float32x4_t x, float32x4_t y) { return vec_sub(x, y); } +inline float32x4_t mul(float32x4_t x, float32x4_t y) { return vec_mul(x, y); } +#endif + #if defined(__MMA__) typedef vector unsigned char vec_t; typedef __vector_quad acc_t; @@ -162,6 +168,13 @@ inline float16x8_t madd(float16x8_t a, float16x8_t b, float16x8_t c) { #endif #endif +#if defined(__VXE__) || defined(__VXE2__) +template <> +inline float32x4_t madd(float32x4_t a, float32x4_t b, float32x4_t c) { + return vec_madd(a, b, c); +} +#endif + //////////////////////////////////////////////////////////////////////////////////////////////////// // VECTORIZED HORIZONTAL SUM @@ -178,6 +191,13 @@ inline float hsum(float16x8_t x) { } #endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC +#if defined(__VXE__) || defined(__VXE2__) +inline float hsum(float32x4_t x) { + float32x4_t tmp = x + vec_reve(x); + return tmp[0] + tmp[1]; +} +#endif + #if defined(__SSE__) || defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) inline float hsum(__m128 x) { #if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) @@ -227,6 +247,21 @@ template <> inline float32x4_t load(const ggml_fp16_t *p) { #endif // _MSC_VER #endif // __ARM_NEON +#if defined(__VXE__) || defined(__VXE2__) +template <> inline float32x4_t load(const ggml_fp16_t * p) { + float tmp[4]; + + for (int i = 0; i < 4; i++) { + tmp[i] = GGML_FP16_TO_FP32(p[i]); + } + + return vec_xl(0, (const float *)(tmp)); +} +template <> inline float32x4_t load(const float * p) { + return vec_xl(0, p); +} +#endif + #if defined(__SSE__) || defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) template <> inline __m128 load(const float *p) { return _mm_loadu_ps(p); @@ -3319,6 +3354,14 @@ bool llamafile_sgemm(const struct ggml_compute_params * params, int64_t m, int64 (const float *)B, ldb, (float *)C, ldc}; return tb.matmul(m, n); +#elif defined(__VXE__) || defined(__VXE2__) + if (n < 4) + return false; + tinyBLAS<4, float32x4_t, float32x4_t, float, float, float> tb{ params, + k, (const float *)A, lda, + (const float *)B, ldb, + (float *)C, ldc}; + return tb.matmul(m, n); #elif defined(__MMA__) if (k % 8) return false; @@ -3410,6 +3453,16 @@ bool llamafile_sgemm(const struct ggml_compute_params * params, int64_t m, int64 (float *)C, ldc}; return tb.matmul(m, n); } +#elif defined(__VXE__) || defined(__VXE2__) + if (n < 4) + return false; + if (Btype == GGML_TYPE_F16) { + tinyBLAS<4, float32x4_t, float32x4_t, ggml_fp16_t, ggml_fp16_t, float> tb{ params, + k, (const ggml_fp16_t *)A, lda, + (const ggml_fp16_t *)B, ldb, + (float *)C, ldc}; + return tb.matmul(m, n); + } #endif return false; } diff --git a/ggml/src/ggml-cpu/llamafile/sgemm.h b/ggml/src/ggml-cpu/llamafile/sgemm.h index 3d2909515..729e8853d 100644 --- a/ggml/src/ggml-cpu/llamafile/sgemm.h +++ b/ggml/src/ggml-cpu/llamafile/sgemm.h @@ -1,6 +1,11 @@ #pragma once #include #include + +#if defined(__VXE__) || defined(__VXE2__) +#include +#endif + #ifdef __cplusplus extern "C" { #endif From 5fc7856815920f828f9e90cb759ac82c7f0c1ea5 Mon Sep 17 00:00:00 2001 From: pqnet <119850+pqnet@users.noreply.github.com> Date: Thu, 19 Jun 2025 12:21:40 +0200 Subject: [PATCH 03/43] convert : fix remote option in Windows (#14100) --- convert_hf_to_gguf.py | 20 +++++++++++--------- 1 file changed, 11 insertions(+), 9 deletions(-) diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py index b754dd815..2e08db345 100755 --- a/convert_hf_to_gguf.py +++ b/convert_hf_to_gguf.py @@ -6389,8 +6389,8 @@ def parse_args() -> argparse.Namespace: help="model is executed on big endian machine", ) parser.add_argument( - "model", type=Path, - help="directory containing model file", + "model", type=str, + help="directory containing model file or huggingface repository ID (if --remote)", nargs="?", ) parser.add_argument( @@ -6493,18 +6493,20 @@ def main() -> None: else: logging.basicConfig(level=logging.INFO) - dir_model = args.model - if args.remote: + hf_repo_id = args.model from huggingface_hub import snapshot_download local_dir = snapshot_download( - repo_id=str(dir_model), + repo_id=hf_repo_id, allow_patterns=["LICENSE", "*.json", "*.md", "*.txt", "tokenizer.model"]) dir_model = Path(local_dir) logger.info(f"Downloaded config and tokenizer to {local_dir}") + else: + hf_repo_id = None + dir_model = Path(args.model) if not dir_model.is_dir(): - logger.error(f'Error: {args.model} is not a directory') + logger.error(f'Error: {dir_model} is not a directory') sys.exit(1) ftype_map: dict[str, gguf.LlamaFileType] = { @@ -6524,9 +6526,9 @@ def main() -> None: if args.outfile is not None: fname_out = args.outfile - elif args.remote: + elif hf_repo_id: # if remote, use the model ID as the output file name - fname_out = Path("./" + str(args.model).replace("/", "-") + "-{ftype}.gguf") + fname_out = Path("./" + hf_repo_id.replace("/", "-") + "-{ftype}.gguf") else: fname_out = dir_model @@ -6555,7 +6557,7 @@ def main() -> None: split_max_tensors=args.split_max_tensors, split_max_size=split_str_to_n_bytes(args.split_max_size), dry_run=args.dry_run, small_first_shard=args.no_tensor_first_split, - remote_hf_model_id=str(args.model) if args.remote else None) + remote_hf_model_id=hf_repo_id) if args.vocab_only: logger.info("Exporting model vocab...") From fffcce535ebbdc25f81966a15b758658788e7466 Mon Sep 17 00:00:00 2001 From: bashayer hijji Date: Thu, 19 Jun 2025 13:24:12 +0300 Subject: [PATCH 04/43] llama-bench : add --no-warmup flag (#14224) (#14270) Add no_warmup parameter to cmd_params struct and command-line parsing to allow users to skip warmup runs before benchmarking. - Add no_warmup boolean field to cmd_params struct - Add --no-warmup command-line argument parsing - Add help text documentation for the new flag - Wrap existing warmup logic in conditional check - Maintain full backward compatibility (warmup enabled by default) Addresses #14224 --- tools/llama-bench/llama-bench.cpp | 44 ++++++++++++++++++------------- 1 file changed, 26 insertions(+), 18 deletions(-) diff --git a/tools/llama-bench/llama-bench.cpp b/tools/llama-bench/llama-bench.cpp index e59d61f19..b80e984d0 100644 --- a/tools/llama-bench/llama-bench.cpp +++ b/tools/llama-bench/llama-bench.cpp @@ -267,6 +267,7 @@ struct cmd_params { int delay; bool verbose; bool progress; + bool no_warmup; output_formats output_format; output_formats output_format_stderr; }; @@ -303,6 +304,7 @@ static const cmd_params cmd_params_defaults = { /* delay */ 0, /* verbose */ false, /* progress */ false, + /* no_warmup */ false, /* output_format */ MARKDOWN, /* output_format_stderr */ NONE, }; @@ -325,6 +327,7 @@ static void print_usage(int /* argc */, char ** argv) { output_format_str(cmd_params_defaults.output_format_stderr)); printf(" -v, --verbose verbose output\n"); printf(" --progress print test progress indicators\n"); + printf(" --no-warmup skip warmup runs before benchmarking\n"); printf("\n"); printf("test parameters:\n"); printf(" -m, --model (default: %s)\n", join(cmd_params_defaults.model, ",").c_str()); @@ -425,6 +428,7 @@ static cmd_params parse_cmd_params(int argc, char ** argv) { params.prio = cmd_params_defaults.prio; params.delay = cmd_params_defaults.delay; params.progress = cmd_params_defaults.progress; + params.no_warmup = cmd_params_defaults.no_warmup; for (int i = 1; i < argc; i++) { arg = argv[i]; @@ -798,6 +802,8 @@ static cmd_params parse_cmd_params(int argc, char ** argv) { params.verbose = true; } else if (arg == "--progress") { params.progress = true; + } else if (arg == "--no-warmup") { + params.no_warmup = true; } else { invalid_param = true; break; @@ -1925,25 +1931,27 @@ int main(int argc, char ** argv) { llama_attach_threadpool(ctx, threadpool, NULL); // warmup run - if (t.n_prompt > 0) { - if (params.progress) { - fprintf(stderr, "llama-bench: benchmark %d/%zu: warmup prompt run\n", params_idx, params_count); + if (!params.no_warmup) { + if (t.n_prompt > 0) { + if (params.progress) { + fprintf(stderr, "llama-bench: benchmark %d/%zu: warmup prompt run\n", params_idx, params_count); + } + //test_prompt(ctx, std::min(t.n_batch, std::min(t.n_prompt, 32)), 0, t.n_batch, t.n_threads); + bool res = test_prompt(ctx, t.n_prompt, t.n_batch, t.n_threads); + if (!res) { + fprintf(stderr, "%s: error: failed to run prompt warmup\n", __func__); + exit(1); + } } - //test_prompt(ctx, std::min(t.n_batch, std::min(t.n_prompt, 32)), 0, t.n_batch, t.n_threads); - bool res = test_prompt(ctx, t.n_prompt, t.n_batch, t.n_threads); - if (!res) { - fprintf(stderr, "%s: error: failed to run prompt warmup\n", __func__); - exit(1); - } - } - if (t.n_gen > 0) { - if (params.progress) { - fprintf(stderr, "llama-bench: benchmark %d/%zu: warmup generation run\n", params_idx, params_count); - } - bool res = test_gen(ctx, 1, t.n_threads); - if (!res) { - fprintf(stderr, "%s: error: failed to run gen warmup\n", __func__); - exit(1); + if (t.n_gen > 0) { + if (params.progress) { + fprintf(stderr, "llama-bench: benchmark %d/%zu: warmup generation run\n", params_idx, params_count); + } + bool res = test_gen(ctx, 1, t.n_threads); + if (!res) { + fprintf(stderr, "%s: error: failed to run gen warmup\n", __func__); + exit(1); + } } } From 600e3e9b50c1f0c9fc4a70356241fd87f00e8e14 Mon Sep 17 00:00:00 2001 From: Anton Mitkov Date: Thu, 19 Jun 2025 11:40:21 +0100 Subject: [PATCH 05/43] sycl: Cleanup codepaths in Get Rows in sycl backend (#14215) Addresses unused reorder path --- ggml/src/ggml-sycl/getrows.cpp | 98 +--------------------------------- 1 file changed, 2 insertions(+), 96 deletions(-) diff --git a/ggml/src/ggml-sycl/getrows.cpp b/ggml/src/ggml-sycl/getrows.cpp index 4a7712781..03f8dd907 100644 --- a/ggml/src/ggml-sycl/getrows.cpp +++ b/ggml/src/ggml-sycl/getrows.cpp @@ -60,54 +60,6 @@ static void k_get_rows( dst_row[iybs + iqs + y_offset] = v.y(); } -template -static void k_get_rows_reorder( - const void * src0, const void *src0_dq, const int32_t * src1, dst_t * dst, - int64_t ne00, /*int64_t ne01, int64_t ne02, int64_t ne03,*/ - /*int64_t ne10, int64_t ne11,*/ int64_t ne12, /*int64_t ne13,*/ - /*size_t s0,*/ size_t s1, size_t s2, size_t s3, - /*size_t nb00,*/ size_t nb01, size_t nb02, size_t nb03, - size_t s10, size_t s11, size_t s12, - const sycl::nd_item<3> &item_ct1/*, size_t s13*/) { - - const int i00 = (item_ct1.get_group(2) * item_ct1.get_local_range(2) + - item_ct1.get_local_id(2)) * - 2; - const int i10 = item_ct1.get_local_range(1) * item_ct1.get_group(1) + - item_ct1.get_local_id(1); - const int i11 = (item_ct1.get_group(0) * item_ct1.get_local_range(0) + - item_ct1.get_local_id(0)) / - ne12; - const int i12 = (item_ct1.get_group(0) * item_ct1.get_local_range(0) + - item_ct1.get_local_id(0)) % - ne12; - - if (i00 >= ne00) { - return; - } - auto ncols = ne00; - const int i01 = src1[i10*s10 + i11*s11 + i12*s12]; - - dst_t * dst_row = dst + i10*s1 + i11*s2 + i12*s3; - - const int src0_off = i01 * ncols + i00; - const int ib = src0_off / QK4_0; // block index - const int iqs = (i00%qk)/qr; // x quant index - const int iybs = i00 - i00%qk; // dst block start index - const int y_offset = qr == 1 ? 1 : qk/2; - - // dequantize - dfloat2 v; - dequantize_kernel_recorder((const void *)src0_dq, ib, (const void *)src0, src0_off/2, v); - - dst_row[iybs + iqs + 0] = v.x(); - dst_row[iybs + iqs + y_offset] = v.y(); - - GGML_UNUSED(nb01); - GGML_UNUSED(nb02); - GGML_UNUSED(nb03); -} - template static void k_get_rows_float( const src0_t * src0, const int32_t * src1, dst_t * dst, @@ -177,47 +129,6 @@ static void get_rows_sycl(ggml_backend_sycl_context & ctx, const ggml_tensor *sr GGML_UNUSED(ctx); } -template -static void get_rows_sycl_reorder(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1, - ggml_tensor *dst, const void *src0_dd, - const int32_t *src1_dd, float *dst_dd, - queue_ptr stream) { - - GGML_TENSOR_BINARY_OP_LOCALS - - const sycl::range<3> block_dims(1, 1, SYCL_GET_ROWS_BLOCK_SIZE); - const int block_num_x = (ne00 + 2*SYCL_GET_ROWS_BLOCK_SIZE - 1) / (2*SYCL_GET_ROWS_BLOCK_SIZE); - const sycl::range<3> block_nums(ne11 * ne12, ne10, block_num_x); - - // strides in elements - //const size_t s0 = nb0 / ggml_element_size(dst); - const size_t s1 = nb1 / ggml_element_size(dst); - const size_t s2 = nb2 / ggml_element_size(dst); - const size_t s3 = nb3 / ggml_element_size(dst); - - const size_t s10 = nb10 / ggml_element_size(src1); - const size_t s11 = nb11 / ggml_element_size(src1); - const size_t s12 = nb12 / ggml_element_size(src1); - //const size_t s13 = nb13 / ggml_element_size(src1); - - GGML_ASSERT(ne00 % 2 == 0); - - const uint8_t* src0_q = (const uint8_t*)src0_dd; - const size_t ncols = ne00; - const size_t nrows = ne01; - const sycl::half* src0_dq = (const sycl::half*)(src0_q + nrows * ncols / 2); - stream->parallel_for(sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]]{ - k_get_rows_reorder( - src0_dd, src0_dq, src1_dd, dst_dd, ne00, ne12, s1, s2, - s3, nb01, nb02, nb03, s10, s11, s12, item_ct1); - }); - - GGML_UNUSED(dst); - GGML_UNUSED(ctx); -} - - template static void get_rows_sycl_float(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1, ggml_tensor *dst, @@ -277,13 +188,8 @@ void ggml_sycl_op_get_rows(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { src1_i32, (float *)dst->data, ctx.stream()); break; case GGML_TYPE_Q4_0: - if (ctx.opt_feature.reorder && dst->op == GGML_OP_MUL_MAT) { - get_rows_sycl_reorder(ctx, dst->src[0], dst->src[1], dst, (const float *)dst->src[0]->data, - src1_i32, (float *)dst->data, ctx.stream()); - } else { - get_rows_sycl(ctx, dst->src[0], dst->src[1], dst, (const float *)dst->src[0]->data, - src1_i32, (float *)dst->data, ctx.stream()); - } + get_rows_sycl(ctx, dst->src[0], dst->src[1], dst, (const float *)dst->src[0]->data, + src1_i32, (float *)dst->data, ctx.stream()); break; case GGML_TYPE_Q4_1: get_rows_sycl(ctx, dst->src[0], dst->src[1], dst, (const float *)dst->src[0]->data, From 456af35eb70177b8dd5779b6d4c21bb020f9cebd Mon Sep 17 00:00:00 2001 From: fanyang Date: Thu, 19 Jun 2025 20:49:48 +0800 Subject: [PATCH 06/43] build : suppress gcc15 compile warnings (#14261) * Change _contains_any() substrs to std::string_view and fix the find comparison logic. --- common/common.cpp | 6 ++++++ ggml/src/ggml-backend-reg.cpp | 5 +++++ src/llama-vocab.cpp | 4 ++-- src/unicode.cpp | 5 +++++ 4 files changed, 18 insertions(+), 2 deletions(-) diff --git a/common/common.cpp b/common/common.cpp index eb80cee08..c2c94e7ae 100644 --- a/common/common.cpp +++ b/common/common.cpp @@ -706,11 +706,17 @@ bool fs_validate_filename(const std::string & filename) { // disable C++17 deprecation warning for std::codecvt_utf8 # pragma clang diagnostic push # pragma clang diagnostic ignored "-Wdeprecated-declarations" +#elif defined(__GNUC__) +# pragma GCC diagnostic push +# pragma GCC diagnostic ignored "-Wdeprecated-declarations" #endif + std::wstring_convert, char32_t> converter; #if defined(__clang__) # pragma clang diagnostic pop +#elif defined(__GNUC__) +# pragma GCC diagnostic pop #endif filename_utf32 = converter.from_bytes(filename); diff --git a/ggml/src/ggml-backend-reg.cpp b/ggml/src/ggml-backend-reg.cpp index 405d8e315..2d93771fd 100644 --- a/ggml/src/ggml-backend-reg.cpp +++ b/ggml/src/ggml-backend-reg.cpp @@ -69,6 +69,9 @@ #if defined(__clang__) # pragma clang diagnostic push # pragma clang diagnostic ignored "-Wdeprecated-declarations" +#elif defined(__GNUC__) +# pragma GCC diagnostic push +# pragma GCC diagnostic ignored "-Wdeprecated-declarations" #endif namespace fs = std::filesystem; @@ -91,6 +94,8 @@ static std::string path_str(const fs::path & path) { #if defined(__clang__) # pragma clang diagnostic pop +#elif defined(__GNUC__) +# pragma GCC diagnostic pop #endif #ifdef _WIN32 diff --git a/src/llama-vocab.cpp b/src/llama-vocab.cpp index dd2251ef3..d90f1d6b1 100644 --- a/src/llama-vocab.cpp +++ b/src/llama-vocab.cpp @@ -2060,9 +2060,9 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) { //NOTE: Per token attributes are missing from the GGUF file. //TODO: Extract attributes from GGUF file. { - auto _contains_any = [] (const std::string & str, const std::vector & substrs) -> bool { + auto _contains_any = [] (const std::string & str, const std::vector & substrs) -> bool { for (const auto & substr : substrs) { - if (str.find(substr) < std::string::npos) { + if (str.find(substr) != std::string::npos) { return true; } } diff --git a/src/unicode.cpp b/src/unicode.cpp index e63bb4ab0..43a4581b9 100644 --- a/src/unicode.cpp +++ b/src/unicode.cpp @@ -204,12 +204,17 @@ static inline std::wstring unicode_wstring_from_utf8(const std::string & s) { // disable C++17 deprecation warning for std::codecvt_utf8 # pragma clang diagnostic push # pragma clang diagnostic ignored "-Wdeprecated-declarations" +#elif defined(__GNUC__) +# pragma GCC diagnostic push +# pragma GCC diagnostic ignored "-Wdeprecated-declarations" #endif std::wstring_convert> conv; #if defined(__clang__) # pragma clang diagnostic pop +#elif defined(__GNUC__) +# pragma GCC diagnostic pop #endif return conv.from_bytes(s); From d67341dc18fc5cc63362880ab2f8f9ecfc7932e7 Mon Sep 17 00:00:00 2001 From: aa956 Date: Thu, 19 Jun 2025 16:01:03 +0300 Subject: [PATCH 07/43] server : add server parameters for draft model cache type (#13782) Co-authored-by: aa956 <27946957+aa956@users.noreply.github.com> --- common/arg.cpp | 26 ++++++++++++++++++++++++++ common/common.h | 3 +++ tools/server/README.md | 2 ++ tools/server/server.cpp | 6 ++---- 4 files changed, 33 insertions(+), 4 deletions(-) diff --git a/common/arg.cpp b/common/arg.cpp index 231de227a..3dfaa71ef 100644 --- a/common/arg.cpp +++ b/common/arg.cpp @@ -3210,6 +3210,32 @@ common_params_context common_params_parser_init(common_params & params, llama_ex params.speculative.model.path = value; } ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_MODEL_DRAFT")); + add_opt(common_arg( + {"-ctkd", "--cache-type-k-draft"}, "TYPE", + string_format( + "KV cache data type for K for the draft model\n" + "allowed values: %s\n" + "(default: %s)", + get_all_kv_cache_types().c_str(), + ggml_type_name(params.speculative.cache_type_k) + ), + [](common_params & params, const std::string & value) { + params.speculative.cache_type_k = kv_cache_type_from_str(value); + } + ).set_env("LLAMA_ARG_CACHE_TYPE_K_DRAFT")); + add_opt(common_arg( + {"-ctvd", "--cache-type-v-draft"}, "TYPE", + string_format( + "KV cache data type for V for the draft model\n" + "allowed values: %s\n" + "(default: %s)", + get_all_kv_cache_types().c_str(), + ggml_type_name(params.speculative.cache_type_v) + ), + [](common_params & params, const std::string & value) { + params.speculative.cache_type_v = kv_cache_type_from_str(value); + } + ).set_env("LLAMA_ARG_CACHE_TYPE_V_DRAFT")); add_opt(common_arg( {"-mv", "--model-vocoder"}, "FNAME", diff --git a/common/common.h b/common/common.h index 00b6ca03a..5710c4e97 100644 --- a/common/common.h +++ b/common/common.h @@ -199,6 +199,9 @@ struct common_params_speculative { float p_split = 0.1f; // speculative decoding split probability float p_min = 0.75f; // minimum speculative decoding probability (greedy) + ggml_type cache_type_k = GGML_TYPE_F16; // KV cache data type for the K + ggml_type cache_type_v = GGML_TYPE_F16; // KV cache data type for the V + struct cpu_params cpuparams; struct cpu_params cpuparams_batch; diff --git a/tools/server/README.md b/tools/server/README.md index 06533c172..43aa65d50 100644 --- a/tools/server/README.md +++ b/tools/server/README.md @@ -187,6 +187,8 @@ The project is under active development, and we are [looking for feedback and co | `-devd, --device-draft ` | comma-separated list of devices to use for offloading the draft model (none = don't offload)
use --list-devices to see a list of available devices | | `-ngld, --gpu-layers-draft, --n-gpu-layers-draft N` | number of layers to store in VRAM for the draft model
(env: LLAMA_ARG_N_GPU_LAYERS_DRAFT) | | `-md, --model-draft FNAME` | draft model for speculative decoding (default: unused)
(env: LLAMA_ARG_MODEL_DRAFT) | +| `-ctkd, --cache-type-k-draft TYPE` | KV cache data type for K for speculative decoding model
allowed values: f32, f16, bf16, q8_0, q4_0, q4_1, iq4_nl, q5_0, q5_1
(default: f16)
(env: LLAMA_ARG_CACHE_TYPE_K_DRAFT) | +| `-ctvd, --cache-type-v-draft TYPE` | KV cache data type for V for speculative decoding model
allowed values: f32, f16, bf16, q8_0, q4_0, q4_1, iq4_nl, q5_0, q5_1
(default: f16)
(env: LLAMA_ARG_CACHE_TYPE_V_DRAFT) | | `-mv, --model-vocoder FNAME` | vocoder model for audio generation (default: unused) | | `--tts-use-guide-tokens` | Use guide tokens to improve TTS word recall | | `--embd-bge-small-en-default` | use default bge-small-en-v1.5 model (note: can download weights from the internet) | diff --git a/tools/server/server.cpp b/tools/server/server.cpp index 721d09182..9d55b3338 100644 --- a/tools/server/server.cpp +++ b/tools/server/server.cpp @@ -1969,10 +1969,8 @@ struct server_context { params_dft.n_ctx = params_base.speculative.n_ctx == 0 ? params_base.n_ctx / params_base.n_parallel : params_base.speculative.n_ctx; params_dft.n_gpu_layers = params_base.speculative.n_gpu_layers; params_dft.n_parallel = 1; - - // force F16 KV cache for the draft model for extra performance - params_dft.cache_type_k = GGML_TYPE_F16; - params_dft.cache_type_v = GGML_TYPE_F16; + params_dft.cache_type_k = params_base.speculative.cache_type_k; + params_dft.cache_type_v = params_base.speculative.cache_type_v; llama_init_dft = common_init_from_params(params_dft); From 381174bbdaf10d6a80dc2099f284b20544d86962 Mon Sep 17 00:00:00 2001 From: Alex Trotta <44127594+Ahajha@users.noreply.github.com> Date: Thu, 19 Jun 2025 09:56:12 -0400 Subject: [PATCH 08/43] gguf-py : make sentencepiece optional (#14200) * Make sentencepiece optional * Bump to 0.18.0 * Bump patch instead of minor Co-authored-by: compilade --------- Co-authored-by: compilade --- gguf-py/gguf/vocab.py | 8 +++++++- gguf-py/pyproject.toml | 4 ++-- 2 files changed, 9 insertions(+), 3 deletions(-) diff --git a/gguf-py/gguf/vocab.py b/gguf-py/gguf/vocab.py index cca097986..44d066ee7 100644 --- a/gguf-py/gguf/vocab.py +++ b/gguf-py/gguf/vocab.py @@ -7,7 +7,10 @@ import os from pathlib import Path from typing import Any, Callable, Sequence, Mapping, Iterable, Protocol, ClassVar, runtime_checkable -from sentencepiece import SentencePieceProcessor +try: + from sentencepiece import SentencePieceProcessor +except ImportError: + SentencePieceProcessor = None import gguf @@ -302,6 +305,9 @@ class SentencePieceVocab(Vocab): name = "spm" def __init__(self, base_path: Path): + if SentencePieceProcessor is None: + raise RuntimeError("sentencepiece is not installed") + added_tokens: dict[str, int] = {} if (fname_tokenizer := base_path / 'tokenizer.model').exists(): # normal location diff --git a/gguf-py/pyproject.toml b/gguf-py/pyproject.toml index f11351cba..0f3a1eeee 100644 --- a/gguf-py/pyproject.toml +++ b/gguf-py/pyproject.toml @@ -1,6 +1,6 @@ [tool.poetry] name = "gguf" -version = "0.17.0" +version = "0.17.1" description = "Read and write ML models in GGUF for GGML" authors = ["GGML "] packages = [ @@ -22,7 +22,7 @@ python = ">=3.8" numpy = ">=1.17" tqdm = ">=4.27" pyyaml = ">=5.1" -sentencepiece = ">=0.1.98,<=0.2.0" +sentencepiece = { version = ">=0.1.98,<=0.2.0", optional = true } PySide6 = { version = "^6.9", python = ">=3.9,<3.14", optional = true } [tool.poetry.dev-dependencies] From 8f71d0f3e86ccbba059350058af8758cafed73e6 Mon Sep 17 00:00:00 2001 From: Diego Devesa Date: Thu, 19 Jun 2025 12:24:14 -0700 Subject: [PATCH 09/43] ggml-cpu : remove unnecesary arm feature detection (#14281) Support for Arm runtime feature detection has now been added to GGML_CPU_ALL_VARIANTS. This removes the old and not very functional code. --- ggml/src/ggml-cpu/arch/arm/repack.cpp | 2002 ++++++++++++------------- ggml/src/ggml-cpu/ggml-cpu.c | 95 +- 2 files changed, 1004 insertions(+), 1093 deletions(-) diff --git a/ggml/src/ggml-cpu/arch/arm/repack.cpp b/ggml/src/ggml-cpu/arch/arm/repack.cpp index 9337e01b6..39a0dd301 100644 --- a/ggml/src/ggml-cpu/arch/arm/repack.cpp +++ b/ggml/src/ggml-cpu/arch/arm/repack.cpp @@ -256,45 +256,43 @@ void ggml_gemv_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo UNUSED(blocklen); #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD) - if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) { - const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx; + const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx; - for (int c = 0; c < nc; c += ncols_interleaved) { - const block_q8_0 * a_ptr = (const block_q8_0 *) vy; - float32x4_t acc = vdupq_n_f32(0); - for (int b = 0; b < nb; b++) { - int8x16_t b0 = vld1q_s8((const int8_t *) b_ptr->qs); - int8x16_t b1 = vld1q_s8((const int8_t *) b_ptr->qs + 16); - int8x16_t b2 = vld1q_s8((const int8_t *) b_ptr->qs + 32); - int8x16_t b3 = vld1q_s8((const int8_t *) b_ptr->qs + 48); - float16x4_t bd = vld1_f16((const __fp16 *) b_ptr->d); + for (int c = 0; c < nc; c += ncols_interleaved) { + const block_q8_0 * a_ptr = (const block_q8_0 *) vy; + float32x4_t acc = vdupq_n_f32(0); + for (int b = 0; b < nb; b++) { + int8x16_t b0 = vld1q_s8((const int8_t *) b_ptr->qs); + int8x16_t b1 = vld1q_s8((const int8_t *) b_ptr->qs + 16); + int8x16_t b2 = vld1q_s8((const int8_t *) b_ptr->qs + 32); + int8x16_t b3 = vld1q_s8((const int8_t *) b_ptr->qs + 48); + float16x4_t bd = vld1_f16((const __fp16 *) b_ptr->d); - int8x16_t a0 = vld1q_s8(a_ptr->qs); - int8x16_t a1 = vld1q_s8(a_ptr->qs + qk/2); - float16x4_t ad = vld1_dup_f16((const __fp16 *) &a_ptr->d); + int8x16_t a0 = vld1q_s8(a_ptr->qs); + int8x16_t a1 = vld1q_s8(a_ptr->qs + qk/2); + float16x4_t ad = vld1_dup_f16((const __fp16 *) &a_ptr->d); - int32x4_t ret = vdupq_n_s32(0); + int32x4_t ret = vdupq_n_s32(0); - ret = vdotq_laneq_s32(ret, b0 << 4, a0, 0); - ret = vdotq_laneq_s32(ret, b1 << 4, a0, 1); - ret = vdotq_laneq_s32(ret, b2 << 4, a0, 2); - ret = vdotq_laneq_s32(ret, b3 << 4, a0, 3); + ret = vdotq_laneq_s32(ret, b0 << 4, a0, 0); + ret = vdotq_laneq_s32(ret, b1 << 4, a0, 1); + ret = vdotq_laneq_s32(ret, b2 << 4, a0, 2); + ret = vdotq_laneq_s32(ret, b3 << 4, a0, 3); - ret = vdotq_laneq_s32(ret, b0 & 0xf0U, a1, 0); - ret = vdotq_laneq_s32(ret, b1 & 0xf0U, a1, 1); - ret = vdotq_laneq_s32(ret, b2 & 0xf0U, a1, 2); - ret = vdotq_laneq_s32(ret, b3 & 0xf0U, a1, 3); + ret = vdotq_laneq_s32(ret, b0 & 0xf0U, a1, 0); + ret = vdotq_laneq_s32(ret, b1 & 0xf0U, a1, 1); + ret = vdotq_laneq_s32(ret, b2 & 0xf0U, a1, 2); + ret = vdotq_laneq_s32(ret, b3 & 0xf0U, a1, 3); - acc = vfmaq_f32(acc, vcvtq_n_f32_s32(ret, 4), - vmulq_f32(vcvt_f32_f16(ad), vcvt_f32_f16(bd))); - a_ptr++; - b_ptr++; - } - vst1q_f32(s, acc); - s += ncols_interleaved; + acc = vfmaq_f32(acc, vcvtq_n_f32_s32(ret, 4), + vmulq_f32(vcvt_f32_f16(ad), vcvt_f32_f16(bd))); + a_ptr++; + b_ptr++; } - return; + vst1q_f32(s, acc); + s += ncols_interleaved; } + return; #endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD) float sumf[4]; int sumi; @@ -341,50 +339,48 @@ void ggml_gemv_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo UNUSED(blocklen); #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD) - if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) { - const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx; + const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx; - for (int c = 0; c < nc; c += ncols_interleaved) { - const block_q8_0 * a_ptr = (const block_q8_0 *) vy; - float32x4_t acc = vdupq_n_f32(0); - for (int b = 0; b < nb; b++) { - int8x16_t b0 = vld1q_s8((const int8_t *) b_ptr->qs); - int8x16_t b1 = vld1q_s8((const int8_t *) b_ptr->qs + 16); - int8x16_t b2 = vld1q_s8((const int8_t *) b_ptr->qs + 32); - int8x16_t b3 = vld1q_s8((const int8_t *) b_ptr->qs + 48); - float16x4_t bd = vld1_f16((const __fp16 *) b_ptr->d); + for (int c = 0; c < nc; c += ncols_interleaved) { + const block_q8_0 * a_ptr = (const block_q8_0 *) vy; + float32x4_t acc = vdupq_n_f32(0); + for (int b = 0; b < nb; b++) { + int8x16_t b0 = vld1q_s8((const int8_t *) b_ptr->qs); + int8x16_t b1 = vld1q_s8((const int8_t *) b_ptr->qs + 16); + int8x16_t b2 = vld1q_s8((const int8_t *) b_ptr->qs + 32); + int8x16_t b3 = vld1q_s8((const int8_t *) b_ptr->qs + 48); + float16x4_t bd = vld1_f16((const __fp16 *) b_ptr->d); - int8x16_t a0 = (int8x16_t) vld1q_dup_s64((const int64_t *) a_ptr->qs); - int8x16_t a1 = (int8x16_t) vld1q_dup_s64((const int64_t *) a_ptr->qs + 1); - int8x16_t a2 = (int8x16_t) vld1q_dup_s64((const int64_t *) a_ptr->qs + 2); - int8x16_t a3 = (int8x16_t) vld1q_dup_s64((const int64_t *) a_ptr->qs + 3); - float16x4_t ad = vld1_dup_f16((const __fp16 *) &a_ptr->d); + int8x16_t a0 = (int8x16_t) vld1q_dup_s64((const int64_t *) a_ptr->qs); + int8x16_t a1 = (int8x16_t) vld1q_dup_s64((const int64_t *) a_ptr->qs + 1); + int8x16_t a2 = (int8x16_t) vld1q_dup_s64((const int64_t *) a_ptr->qs + 2); + int8x16_t a3 = (int8x16_t) vld1q_dup_s64((const int64_t *) a_ptr->qs + 3); + float16x4_t ad = vld1_dup_f16((const __fp16 *) &a_ptr->d); - int32x4_t ret0 = vdupq_n_s32(0); - int32x4_t ret1 = vdupq_n_s32(0); + int32x4_t ret0 = vdupq_n_s32(0); + int32x4_t ret1 = vdupq_n_s32(0); - ret0 = vdotq_s32(ret0, b0 << 4, a0); - ret1 = vdotq_s32(ret1, b1 << 4, a0); - ret0 = vdotq_s32(ret0, b2 << 4, a1); - ret1 = vdotq_s32(ret1, b3 << 4, a1); + ret0 = vdotq_s32(ret0, b0 << 4, a0); + ret1 = vdotq_s32(ret1, b1 << 4, a0); + ret0 = vdotq_s32(ret0, b2 << 4, a1); + ret1 = vdotq_s32(ret1, b3 << 4, a1); - ret0 = vdotq_s32(ret0, b0 & 0xf0U, a2); - ret1 = vdotq_s32(ret1, b1 & 0xf0U, a2); - ret0 = vdotq_s32(ret0, b2 & 0xf0U, a3); - ret1 = vdotq_s32(ret1, b3 & 0xf0U, a3); + ret0 = vdotq_s32(ret0, b0 & 0xf0U, a2); + ret1 = vdotq_s32(ret1, b1 & 0xf0U, a2); + ret0 = vdotq_s32(ret0, b2 & 0xf0U, a3); + ret1 = vdotq_s32(ret1, b3 & 0xf0U, a3); - int32x4_t ret = vpaddq_s32(ret0, ret1); + int32x4_t ret = vpaddq_s32(ret0, ret1); - acc = vfmaq_f32(acc, vcvtq_n_f32_s32(ret, 4), - vmulq_f32(vcvt_f32_f16(ad), vcvt_f32_f16(bd))); - a_ptr++; - b_ptr++; - } - vst1q_f32(s, acc); - s += ncols_interleaved; + acc = vfmaq_f32(acc, vcvtq_n_f32_s32(ret, 4), + vmulq_f32(vcvt_f32_f16(ad), vcvt_f32_f16(bd))); + a_ptr++; + b_ptr++; } - return; + vst1q_f32(s, acc); + s += ncols_interleaved; } + return; #endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD) float sumf[4]; int sumi; @@ -432,7 +428,7 @@ void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) #if defined(__ARM_FEATURE_SVE) - if (ggml_cpu_has_sve() && ggml_cpu_get_sve_cnt() == QK8_0) { + if (ggml_cpu_get_sve_cnt() == QK8_0) { const void * b_ptr = vx; const void * a_ptr = vy; float * res_ptr = s; @@ -547,54 +543,52 @@ void ggml_gemv_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const UNUSED(blocklen); #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD) - if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) { - const int8x16_t kvalues = vld1q_s8(kvalues_iq4nl); - const block_q8_0 * a_ptr = (const block_q8_0 *) vy; - float * res_ptr = s; + const int8x16_t kvalues = vld1q_s8(kvalues_iq4nl); + const block_q8_0 * a_ptr = (const block_q8_0 *) vy; + float * res_ptr = s; - for (int x = 0; x < nc / ncols_interleaved; x++) { - const block_iq4_nlx4 * b_ptr = (const block_iq4_nlx4 *) vx + (x * nb); + for (int x = 0; x < nc / ncols_interleaved; x++) { + const block_iq4_nlx4 * b_ptr = (const block_iq4_nlx4 *) vx + (x * nb); - float32x4_t sumf = vdupq_n_f32(0); - for (int l = 0; l < nb; l++) { - uint8x16_t b_0 = vld1q_u8(b_ptr[l].qs + 0); - uint8x16_t b_1 = vld1q_u8(b_ptr[l].qs + 16); - uint8x16_t b_2 = vld1q_u8(b_ptr[l].qs + 32); - uint8x16_t b_3 = vld1q_u8(b_ptr[l].qs + 48); + float32x4_t sumf = vdupq_n_f32(0); + for (int l = 0; l < nb; l++) { + uint8x16_t b_0 = vld1q_u8(b_ptr[l].qs + 0); + uint8x16_t b_1 = vld1q_u8(b_ptr[l].qs + 16); + uint8x16_t b_2 = vld1q_u8(b_ptr[l].qs + 32); + uint8x16_t b_3 = vld1q_u8(b_ptr[l].qs + 48); - int8x16_t b_0_hi = vqtbl1q_s8(kvalues, b_0 >> 4); - int8x16_t b_0_lo = vqtbl1q_s8(kvalues, b_0 & 0x0F); - int8x16_t b_1_hi = vqtbl1q_s8(kvalues, b_1 >> 4); - int8x16_t b_1_lo = vqtbl1q_s8(kvalues, b_1 & 0x0F); - int8x16_t b_2_hi = vqtbl1q_s8(kvalues, b_2 >> 4); - int8x16_t b_2_lo = vqtbl1q_s8(kvalues, b_2 & 0x0F); - int8x16_t b_3_hi = vqtbl1q_s8(kvalues, b_3 >> 4); - int8x16_t b_3_lo = vqtbl1q_s8(kvalues, b_3 & 0x0F); + int8x16_t b_0_hi = vqtbl1q_s8(kvalues, b_0 >> 4); + int8x16_t b_0_lo = vqtbl1q_s8(kvalues, b_0 & 0x0F); + int8x16_t b_1_hi = vqtbl1q_s8(kvalues, b_1 >> 4); + int8x16_t b_1_lo = vqtbl1q_s8(kvalues, b_1 & 0x0F); + int8x16_t b_2_hi = vqtbl1q_s8(kvalues, b_2 >> 4); + int8x16_t b_2_lo = vqtbl1q_s8(kvalues, b_2 & 0x0F); + int8x16_t b_3_hi = vqtbl1q_s8(kvalues, b_3 >> 4); + int8x16_t b_3_lo = vqtbl1q_s8(kvalues, b_3 & 0x0F); - int8x16_t a_0 = vld1q_s8(a_ptr[l].qs + 0); - int8x16_t a_1 = vld1q_s8(a_ptr[l].qs + 16); + int8x16_t a_0 = vld1q_s8(a_ptr[l].qs + 0); + int8x16_t a_1 = vld1q_s8(a_ptr[l].qs + 16); - int32x4_t sumi = vdupq_n_s32(0); - sumi = vdotq_laneq_s32(sumi, b_0_lo, a_0, 0); - sumi = vdotq_laneq_s32(sumi, b_0_hi, a_1, 0); - sumi = vdotq_laneq_s32(sumi, b_1_lo, a_0, 1); - sumi = vdotq_laneq_s32(sumi, b_1_hi, a_1, 1); - sumi = vdotq_laneq_s32(sumi, b_2_lo, a_0, 2); - sumi = vdotq_laneq_s32(sumi, b_2_hi, a_1, 2); - sumi = vdotq_laneq_s32(sumi, b_3_lo, a_0, 3); - sumi = vdotq_laneq_s32(sumi, b_3_hi, a_1, 3); + int32x4_t sumi = vdupq_n_s32(0); + sumi = vdotq_laneq_s32(sumi, b_0_lo, a_0, 0); + sumi = vdotq_laneq_s32(sumi, b_0_hi, a_1, 0); + sumi = vdotq_laneq_s32(sumi, b_1_lo, a_0, 1); + sumi = vdotq_laneq_s32(sumi, b_1_hi, a_1, 1); + sumi = vdotq_laneq_s32(sumi, b_2_lo, a_0, 2); + sumi = vdotq_laneq_s32(sumi, b_2_hi, a_1, 2); + sumi = vdotq_laneq_s32(sumi, b_3_lo, a_0, 3); + sumi = vdotq_laneq_s32(sumi, b_3_hi, a_1, 3); - float32x4_t a_d = vcvt_f32_f16(vld1_dup_f16((const float16_t *)&a_ptr[l].d)); - float32x4_t b_d = vcvt_f32_f16(vld1_f16((const float16_t *)b_ptr[l].d)); - float32x4_t d = a_d * b_d; + float32x4_t a_d = vcvt_f32_f16(vld1_dup_f16((const float16_t *)&a_ptr[l].d)); + float32x4_t b_d = vcvt_f32_f16(vld1_f16((const float16_t *)b_ptr[l].d)); + float32x4_t d = a_d * b_d; - sumf = vmlaq_f32(sumf, d, vcvtq_f32_s32(sumi)); - } - - vst1q_f32(res_ptr + x * 4, sumf); + sumf = vmlaq_f32(sumf, d, vcvtq_f32_s32(sumi)); } - return; + + vst1q_f32(res_ptr + x * 4, sumf); } + return; #endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) { float sumf[4]; @@ -643,465 +637,463 @@ void ggml_gemm_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo UNUSED(ncols_interleaved); UNUSED(blocklen); -#if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) - if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) { - const void * b_ptr = vx; - const void * a_ptr = vy; - float * res_ptr = s; - size_t res_stride = bs * sizeof(float); +#if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD) + const void * b_ptr = vx; + const void * a_ptr = vy; + float * res_ptr = s; + size_t res_stride = bs * sizeof(float); - __asm__ __volatile__( - "mov x10, %x[nr]\n" - "mov x9, #0x88\n" - "cmp x10, #0x10\n" - "mul x9, %x[nb], x9\n" - "blt 4f\n" - "1:" // Row loop - "add x28, %x[b_ptr], #0x8\n" - "mov x27, %x[nc]\n" - "add x26, %x[res_ptr], %x[res_stride], LSL #4\n" - "2:" // Column loop - "add x25, %x[a_ptr], #0x8\n" - "movi v15.16b, #0x0\n" - "movi v19.16b, #0x0\n" - "mov x24, %x[nb]\n" - "add x23, x25, x9\n" - "movi v18.16b, #0x0\n" - "movi v14.16b, #0x0\n" - "add x22, x23, x9\n" - "movi v11.16b, #0x0\n" - "movi v13.16b, #0x0\n" - "add x21, x22, x9\n" - "movi v23.16b, #0x0\n" - "movi v16.16b, #0x0\n" - "movi v25.16b, #0x0\n" - "movi v7.16b, #0x0\n" - "movi v0.16b, #0x0\n" - "movi v4.16b, #0x0\n" - "movi v5.16b, #0x0\n" - "movi v21.16b, #0x0\n" - "movi v8.16b, #0x0\n" - "movi v1.16b, #0x0\n" - "3:" // Block loop - "ldr q3, [x28, #0x0]\n" - "ldr q31, [x25, #0x0]\n" - "movi v28.16b, #0x4\n" - "movi v10.4s, #0x0\n" - "ldr q22, [x28, #0x10]\n" - "ldr q6, [x25, #0x10]\n" - "movi v29.4s, #0x0\n" - "movi v9.4s, #0x0\n" - "ldr q27, [x28, #0x20]\n" - "ldr q30, [x28, #0x30]\n" - "movi v20.4s, #0x0\n" - "movi v24.16b, #0xf0\n" - "ldr d2, [x25, #-0x8]\n" - "ldr d26, [x23, #-0x8]\n" - "sshl v12.16b, v3.16b, v28.16b\n" - "sub x20, x28, #0x8\n" - "ldr d17, [x20, #0x0]\n" - "and v3.16b, v3.16b, v24.16b\n" - "subs x24, x24, #0x1\n" - "add x28, x28, #0x48\n" - ".inst 0x4f9fe18a // sdot v10.4s, v12.16b, v31.4b[0]\n" - ".inst 0x4fbfe19d // sdot v29.4s, v12.16b, v31.4b[1]\n" - ".inst 0x4f9fe989 // sdot v9.4s, v12.16b, v31.4b[2]\n" - ".inst 0x4fbfe994 // sdot v20.4s, v12.16b, v31.4b[3]\n" - "sshl v31.16b, v22.16b, v28.16b\n" - "and v22.16b, v22.16b, v24.16b\n" - "fcvtl v17.4s, v17.4h\n" - "fcvtl v2.4s, v2.4h\n" - "fcvtl v26.4s, v26.4h\n" - ".inst 0x4f86e3ea // sdot v10.4s, v31.16b, v6.4b[0]\n" - ".inst 0x4fa6e3fd // sdot v29.4s, v31.16b, v6.4b[1]\n" - ".inst 0x4f86ebe9 // sdot v9.4s, v31.16b, v6.4b[2]\n" - ".inst 0x4fa6ebf4 // sdot v20.4s, v31.16b, v6.4b[3]\n" - "sshl v6.16b, v27.16b, v28.16b\n" - "sshl v28.16b, v30.16b, v28.16b\n" - "and v27.16b, v27.16b, v24.16b\n" - "and v30.16b, v30.16b, v24.16b\n" - "ldr q24, [x25, #0x20]\n" - ".inst 0x4f98e0ca // sdot v10.4s, v6.16b, v24.4b[0]\n" - ".inst 0x4fb8e0dd // sdot v29.4s, v6.16b, v24.4b[1]\n" - ".inst 0x4f98e8c9 // sdot v9.4s, v6.16b, v24.4b[2]\n" - ".inst 0x4fb8e8d4 // sdot v20.4s, v6.16b, v24.4b[3]\n" - "ldr q24, [x25, #0x30]\n" - ".inst 0x4f98e38a // sdot v10.4s, v28.16b, v24.4b[0]\n" - ".inst 0x4fb8e39d // sdot v29.4s, v28.16b, v24.4b[1]\n" - ".inst 0x4f98eb89 // sdot v9.4s, v28.16b, v24.4b[2]\n" - ".inst 0x4fb8eb94 // sdot v20.4s, v28.16b, v24.4b[3]\n" - "ldr q24, [x25, #0x40]\n" - ".inst 0x4f98e06a // sdot v10.4s, v3.16b, v24.4b[0]\n" - ".inst 0x4fb8e07d // sdot v29.4s, v3.16b, v24.4b[1]\n" - ".inst 0x4f98e869 // sdot v9.4s, v3.16b, v24.4b[2]\n" - ".inst 0x4fb8e874 // sdot v20.4s, v3.16b, v24.4b[3]\n" - "ldr q24, [x25, #0x50]\n" - ".inst 0x4f98e2ca // sdot v10.4s, v22.16b, v24.4b[0]\n" - ".inst 0x4fb8e2dd // sdot v29.4s, v22.16b, v24.4b[1]\n" - ".inst 0x4f98eac9 // sdot v9.4s, v22.16b, v24.4b[2]\n" - ".inst 0x4fb8ead4 // sdot v20.4s, v22.16b, v24.4b[3]\n" - "ldr q24, [x25, #0x60]\n" - ".inst 0x4f98e36a // sdot v10.4s, v27.16b, v24.4b[0]\n" - ".inst 0x4fb8e37d // sdot v29.4s, v27.16b, v24.4b[1]\n" - ".inst 0x4f98eb69 // sdot v9.4s, v27.16b, v24.4b[2]\n" - ".inst 0x4fb8eb74 // sdot v20.4s, v27.16b, v24.4b[3]\n" - "ldr q24, [x25, #0x70]\n" - "add x25, x25, #0x88\n" - ".inst 0x4f98e3ca // sdot v10.4s, v30.16b, v24.4b[0]\n" - ".inst 0x4fb8e3dd // sdot v29.4s, v30.16b, v24.4b[1]\n" - ".inst 0x4f98ebc9 // sdot v9.4s, v30.16b, v24.4b[2]\n" - ".inst 0x4fb8ebd4 // sdot v20.4s, v30.16b, v24.4b[3]\n" - "fmul v24.4s, v17.4s, v2.s[0]\n" - "scvtf v10.4s, v10.4s, #0x4\n" - "scvtf v29.4s, v29.4s, #0x4\n" - "scvtf v9.4s, v9.4s, #0x4\n" - "scvtf v20.4s, v20.4s, #0x4\n" - "fmla v15.4s, v10.4s, v24.4s\n" - "ldr q24, [x23, #0x0]\n" - "fmul v10.4s, v17.4s, v2.s[1]\n" - "fmla v19.4s, v29.4s, v10.4s\n" - "ldr q10, [x23, #0x10]\n" - "fmul v29.4s, v17.4s, v2.s[2]\n" - "fmul v2.4s, v17.4s, v2.s[3]\n" - "fmla v18.4s, v9.4s, v29.4s\n" - "movi v9.4s, #0x0\n" - "movi v29.4s, #0x0\n" - ".inst 0x4f98e189 // sdot v9.4s, v12.16b, v24.4b[0]\n" - ".inst 0x4fb8e19d // sdot v29.4s, v12.16b, v24.4b[1]\n" - "fmla v14.4s, v20.4s, v2.4s\n" - "movi v20.4s, #0x0\n" - "movi v2.4s, #0x0\n" - ".inst 0x4f98e994 // sdot v20.4s, v12.16b, v24.4b[2]\n" - ".inst 0x4fb8e982 // sdot v2.4s, v12.16b, v24.4b[3]\n" - "ldr q24, [x23, #0x20]\n" - ".inst 0x4f8ae3e9 // sdot v9.4s, v31.16b, v10.4b[0]\n" - ".inst 0x4faae3fd // sdot v29.4s, v31.16b, v10.4b[1]\n" - ".inst 0x4f8aebf4 // sdot v20.4s, v31.16b, v10.4b[2]\n" - ".inst 0x4faaebe2 // sdot v2.4s, v31.16b, v10.4b[3]\n" - "ldr q10, [x23, #0x30]\n" - ".inst 0x4f98e0c9 // sdot v9.4s, v6.16b, v24.4b[0]\n" - ".inst 0x4fb8e0dd // sdot v29.4s, v6.16b, v24.4b[1]\n" - ".inst 0x4f98e8d4 // sdot v20.4s, v6.16b, v24.4b[2]\n" - ".inst 0x4fb8e8c2 // sdot v2.4s, v6.16b, v24.4b[3]\n" - "ldr q24, [x23, #0x40]\n" - ".inst 0x4f8ae389 // sdot v9.4s, v28.16b, v10.4b[0]\n" - ".inst 0x4faae39d // sdot v29.4s, v28.16b, v10.4b[1]\n" - ".inst 0x4f8aeb94 // sdot v20.4s, v28.16b, v10.4b[2]\n" - ".inst 0x4faaeb82 // sdot v2.4s, v28.16b, v10.4b[3]\n" - "ldr q10, [x23, #0x50]\n" - ".inst 0x4f98e069 // sdot v9.4s, v3.16b, v24.4b[0]\n" - ".inst 0x4fb8e07d // sdot v29.4s, v3.16b, v24.4b[1]\n" - ".inst 0x4f98e874 // sdot v20.4s, v3.16b, v24.4b[2]\n" - ".inst 0x4fb8e862 // sdot v2.4s, v3.16b, v24.4b[3]\n" - "ldr q24, [x23, #0x60]\n" - ".inst 0x4f8ae2c9 // sdot v9.4s, v22.16b, v10.4b[0]\n" - ".inst 0x4faae2dd // sdot v29.4s, v22.16b, v10.4b[1]\n" - ".inst 0x4f8aead4 // sdot v20.4s, v22.16b, v10.4b[2]\n" - ".inst 0x4faaeac2 // sdot v2.4s, v22.16b, v10.4b[3]\n" - "ldr q10, [x23, #0x70]\n" - "add x23, x23, #0x88\n" - ".inst 0x4f98e369 // sdot v9.4s, v27.16b, v24.4b[0]\n" - ".inst 0x4fb8e37d // sdot v29.4s, v27.16b, v24.4b[1]\n" - ".inst 0x4f98eb74 // sdot v20.4s, v27.16b, v24.4b[2]\n" - ".inst 0x4fb8eb62 // sdot v2.4s, v27.16b, v24.4b[3]\n" - "ldr q24, [x22, #0x0]\n" - ".inst 0x4f8ae3c9 // sdot v9.4s, v30.16b, v10.4b[0]\n" - ".inst 0x4faae3dd // sdot v29.4s, v30.16b, v10.4b[1]\n" - ".inst 0x4f8aebd4 // sdot v20.4s, v30.16b, v10.4b[2]\n" - ".inst 0x4faaebc2 // sdot v2.4s, v30.16b, v10.4b[3]\n" - "fmul v10.4s, v17.4s, v26.s[0]\n" - "scvtf v9.4s, v9.4s, #0x4\n" - "scvtf v29.4s, v29.4s, #0x4\n" - "scvtf v20.4s, v20.4s, #0x4\n" - "scvtf v2.4s, v2.4s, #0x4\n" - "fmla v11.4s, v9.4s, v10.4s\n" - "ldr q9, [x22, #0x10]\n" - "fmul v10.4s, v17.4s, v26.s[1]\n" - "fmla v13.4s, v29.4s, v10.4s\n" - "ldr d29, [x22, #-0x8]\n" - "fmul v10.4s, v17.4s, v26.s[2]\n" - "fmul v26.4s, v17.4s, v26.s[3]\n" - "fcvtl v29.4s, v29.4h\n" - "fmla v23.4s, v20.4s, v10.4s\n" - "movi v20.4s, #0x0\n" - "movi v10.4s, #0x0\n" - "fmla v16.4s, v2.4s, v26.4s\n" - "movi v26.4s, #0x0\n" - "movi v2.4s, #0x0\n" - ".inst 0x4f98e194 // sdot v20.4s, v12.16b, v24.4b[0]\n" - ".inst 0x4fb8e18a // sdot v10.4s, v12.16b, v24.4b[1]\n" - ".inst 0x4f98e99a // sdot v26.4s, v12.16b, v24.4b[2]\n" - ".inst 0x4fb8e982 // sdot v2.4s, v12.16b, v24.4b[3]\n" - "ldr q24, [x22, #0x20]\n" - ".inst 0x4f89e3f4 // sdot v20.4s, v31.16b, v9.4b[0]\n" - ".inst 0x4fa9e3ea // sdot v10.4s, v31.16b, v9.4b[1]\n" - ".inst 0x4f89ebfa // sdot v26.4s, v31.16b, v9.4b[2]\n" - ".inst 0x4fa9ebe2 // sdot v2.4s, v31.16b, v9.4b[3]\n" - "ldr q9, [x22, #0x30]\n" - ".inst 0x4f98e0d4 // sdot v20.4s, v6.16b, v24.4b[0]\n" - ".inst 0x4fb8e0ca // sdot v10.4s, v6.16b, v24.4b[1]\n" - ".inst 0x4f98e8da // sdot v26.4s, v6.16b, v24.4b[2]\n" - ".inst 0x4fb8e8c2 // sdot v2.4s, v6.16b, v24.4b[3]\n" - "ldr q24, [x22, #0x40]\n" - ".inst 0x4f89e394 // sdot v20.4s, v28.16b, v9.4b[0]\n" - ".inst 0x4fa9e38a // sdot v10.4s, v28.16b, v9.4b[1]\n" - ".inst 0x4f89eb9a // sdot v26.4s, v28.16b, v9.4b[2]\n" - ".inst 0x4fa9eb82 // sdot v2.4s, v28.16b, v9.4b[3]\n" - "ldr q9, [x22, #0x50]\n" - ".inst 0x4f98e074 // sdot v20.4s, v3.16b, v24.4b[0]\n" - ".inst 0x4fb8e06a // sdot v10.4s, v3.16b, v24.4b[1]\n" - ".inst 0x4f98e87a // sdot v26.4s, v3.16b, v24.4b[2]\n" - ".inst 0x4fb8e862 // sdot v2.4s, v3.16b, v24.4b[3]\n" - "ldr q24, [x22, #0x60]\n" - ".inst 0x4f89e2d4 // sdot v20.4s, v22.16b, v9.4b[0]\n" - ".inst 0x4fa9e2ca // sdot v10.4s, v22.16b, v9.4b[1]\n" - ".inst 0x4f89eada // sdot v26.4s, v22.16b, v9.4b[2]\n" - ".inst 0x4fa9eac2 // sdot v2.4s, v22.16b, v9.4b[3]\n" - "ldr q9, [x22, #0x70]\n" - "add x22, x22, #0x88\n" - ".inst 0x4f98e374 // sdot v20.4s, v27.16b, v24.4b[0]\n" - ".inst 0x4fb8e36a // sdot v10.4s, v27.16b, v24.4b[1]\n" - ".inst 0x4f98eb7a // sdot v26.4s, v27.16b, v24.4b[2]\n" - ".inst 0x4fb8eb62 // sdot v2.4s, v27.16b, v24.4b[3]\n" - "ldr q24, [x21, #0x0]\n" - ".inst 0x4f89e3d4 // sdot v20.4s, v30.16b, v9.4b[0]\n" - ".inst 0x4fa9e3ca // sdot v10.4s, v30.16b, v9.4b[1]\n" - ".inst 0x4f89ebda // sdot v26.4s, v30.16b, v9.4b[2]\n" - ".inst 0x4fa9ebc2 // sdot v2.4s, v30.16b, v9.4b[3]\n" - "fmul v9.4s, v17.4s, v29.s[0]\n" - "scvtf v20.4s, v20.4s, #0x4\n" - "scvtf v10.4s, v10.4s, #0x4\n" - "scvtf v26.4s, v26.4s, #0x4\n" - "scvtf v2.4s, v2.4s, #0x4\n" - "fmla v25.4s, v20.4s, v9.4s\n" - "ldr q9, [x21, #0x10]\n" - "fmul v20.4s, v17.4s, v29.s[1]\n" - "fmla v7.4s, v10.4s, v20.4s\n" - "ldr d20, [x21, #-0x8]\n" - "fmul v10.4s, v17.4s, v29.s[2]\n" - "fmul v29.4s, v17.4s, v29.s[3]\n" - "fcvtl v20.4s, v20.4h\n" - "fmla v0.4s, v26.4s, v10.4s\n" - "movi v26.4s, #0x0\n" - "movi v10.4s, #0x0\n" - "fmla v4.4s, v2.4s, v29.4s\n" - "movi v2.4s, #0x0\n" - "movi v29.4s, #0x0\n" - ".inst 0x4f98e19a // sdot v26.4s, v12.16b, v24.4b[0]\n" - ".inst 0x4fb8e18a // sdot v10.4s, v12.16b, v24.4b[1]\n" - ".inst 0x4f98e982 // sdot v2.4s, v12.16b, v24.4b[2]\n" - ".inst 0x4fb8e99d // sdot v29.4s, v12.16b, v24.4b[3]\n" - "ldr q12, [x21, #0x20]\n" - "fmul v24.4s, v17.4s, v20.s[0]\n" - ".inst 0x4f89e3fa // sdot v26.4s, v31.16b, v9.4b[0]\n" - ".inst 0x4fa9e3ea // sdot v10.4s, v31.16b, v9.4b[1]\n" - ".inst 0x4f89ebe2 // sdot v2.4s, v31.16b, v9.4b[2]\n" - ".inst 0x4fa9ebfd // sdot v29.4s, v31.16b, v9.4b[3]\n" - "ldr q9, [x21, #0x30]\n" - "fmul v31.4s, v17.4s, v20.s[1]\n" - ".inst 0x4f8ce0da // sdot v26.4s, v6.16b, v12.4b[0]\n" - ".inst 0x4face0ca // sdot v10.4s, v6.16b, v12.4b[1]\n" - ".inst 0x4f8ce8c2 // sdot v2.4s, v6.16b, v12.4b[2]\n" - ".inst 0x4face8dd // sdot v29.4s, v6.16b, v12.4b[3]\n" - "ldr q12, [x21, #0x40]\n" - "fmul v6.4s, v17.4s, v20.s[2]\n" - "fmul v20.4s, v17.4s, v20.s[3]\n" - ".inst 0x4f89e39a // sdot v26.4s, v28.16b, v9.4b[0]\n" - ".inst 0x4fa9e38a // sdot v10.4s, v28.16b, v9.4b[1]\n" - ".inst 0x4f89eb82 // sdot v2.4s, v28.16b, v9.4b[2]\n" - ".inst 0x4fa9eb9d // sdot v29.4s, v28.16b, v9.4b[3]\n" - "ldr q9, [x21, #0x50]\n" - ".inst 0x4f8ce07a // sdot v26.4s, v3.16b, v12.4b[0]\n" - ".inst 0x4face06a // sdot v10.4s, v3.16b, v12.4b[1]\n" - ".inst 0x4f8ce862 // sdot v2.4s, v3.16b, v12.4b[2]\n" - ".inst 0x4face87d // sdot v29.4s, v3.16b, v12.4b[3]\n" - "ldr q12, [x21, #0x60]\n" - ".inst 0x4f89e2da // sdot v26.4s, v22.16b, v9.4b[0]\n" - ".inst 0x4fa9e2ca // sdot v10.4s, v22.16b, v9.4b[1]\n" - ".inst 0x4f89eac2 // sdot v2.4s, v22.16b, v9.4b[2]\n" - ".inst 0x4fa9eadd // sdot v29.4s, v22.16b, v9.4b[3]\n" - "ldr q17, [x21, #0x70]\n" - "add x21, x21, #0x88\n" - ".inst 0x4f8ce37a // sdot v26.4s, v27.16b, v12.4b[0]\n" - ".inst 0x4face36a // sdot v10.4s, v27.16b, v12.4b[1]\n" - ".inst 0x4f8ceb62 // sdot v2.4s, v27.16b, v12.4b[2]\n" - ".inst 0x4faceb7d // sdot v29.4s, v27.16b, v12.4b[3]\n" - ".inst 0x4f91e3da // sdot v26.4s, v30.16b, v17.4b[0]\n" - ".inst 0x4fb1e3ca // sdot v10.4s, v30.16b, v17.4b[1]\n" - ".inst 0x4f91ebc2 // sdot v2.4s, v30.16b, v17.4b[2]\n" - ".inst 0x4fb1ebdd // sdot v29.4s, v30.16b, v17.4b[3]\n" - "scvtf v26.4s, v26.4s, #0x4\n" - "scvtf v10.4s, v10.4s, #0x4\n" - "fmla v5.4s, v26.4s, v24.4s\n" - "scvtf v2.4s, v2.4s, #0x4\n" - "scvtf v29.4s, v29.4s, #0x4\n" - "fmla v21.4s, v10.4s, v31.4s\n" - "fmla v8.4s, v2.4s, v6.4s\n" - "fmla v1.4s, v29.4s, v20.4s\n" - "bgt 3b\n" - "mov x20, %x[res_ptr]\n" - "subs x27, x27, #0x4\n" - "add %x[res_ptr], %x[res_ptr], #0x10\n" - "str q15, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q19, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q18, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q14, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q11, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q13, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q23, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q16, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q25, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q7, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q0, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q4, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q5, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q21, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q8, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q1, [x20, #0x0]\n" - "bne 2b\n" - "mov x20, #0x4\n" - "sub x10, x10, #0x10\n" - "cmp x10, #0x10\n" - "mov %x[res_ptr], x26\n" - "madd %x[a_ptr], x20, x9, %x[a_ptr]\n" - "bge 1b\n" - "4:" // Row loop skip - "cbz x10, 9f\n" - "5:" // Row tail: Row loop - "add x24, %x[b_ptr], #0x8\n" - "mov x23, %x[nc]\n" - "add x22, %x[res_ptr], %x[res_stride], LSL #2\n" - "6:" // Row tail: Column loop - "movi v15.16b, #0x0\n" - "movi v19.16b, #0x0\n" - "add x25, %x[a_ptr], #0x8\n" - "mov x21, %x[nb]\n" - "movi v18.16b, #0x0\n" - "movi v14.16b, #0x0\n" - "7:" // Row tail: Block loop - "ldr q7, [x24, #0x0]\n" - "ldr q5, [x25, #0x0]\n" - "movi v9.16b, #0x4\n" - "movi v4.4s, #0x0\n" - "ldr q3, [x24, #0x10]\n" - "ldr q2, [x25, #0x10]\n" - "movi v1.4s, #0x0\n" - "movi v0.4s, #0x0\n" - "ldr q13, [x24, #0x20]\n" - "ldr q31, [x25, #0x20]\n" - "movi v30.4s, #0x0\n" - "movi v29.16b, #0xf0\n" - "ldr q28, [x24, #0x30]\n" - "ldr q27, [x25, #0x30]\n" - "sshl v20.16b, v7.16b, v9.16b\n" - "sub x20, x24, #0x8\n" - "ldr q26, [x25, #0x40]\n" - "ldr q25, [x25, #0x50]\n" - "sshl v17.16b, v3.16b, v9.16b\n" - "and v7.16b, v7.16b, v29.16b\n" - "ldr q24, [x25, #0x60]\n" - "ldr q16, [x25, #0x70]\n" - "sshl v22.16b, v13.16b, v9.16b\n" - "and v3.16b, v3.16b, v29.16b\n" - "ldr d21, [x20, #0x0]\n" - "ldr d12, [x25, #-0x8]\n" - ".inst 0x4f85e284 // sdot v4.4s, v20.16b, v5.4b[0]\n" - ".inst 0x4fa5e281 // sdot v1.4s, v20.16b, v5.4b[1]\n" - ".inst 0x4f85ea80 // sdot v0.4s, v20.16b, v5.4b[2]\n" - ".inst 0x4fa5ea9e // sdot v30.4s, v20.16b, v5.4b[3]\n" - "sshl v9.16b, v28.16b, v9.16b\n" - "subs x21, x21, #0x1\n" - "and v13.16b, v13.16b, v29.16b\n" - "and v28.16b, v28.16b, v29.16b\n" - "add x25, x25, #0x88\n" - "add x24, x24, #0x48\n" - "fcvtl v21.4s, v21.4h\n" - "fcvtl v12.4s, v12.4h\n" - ".inst 0x4f82e224 // sdot v4.4s, v17.16b, v2.4b[0]\n" - ".inst 0x4fa2e221 // sdot v1.4s, v17.16b, v2.4b[1]\n" - ".inst 0x4f82ea20 // sdot v0.4s, v17.16b, v2.4b[2]\n" - ".inst 0x4fa2ea3e // sdot v30.4s, v17.16b, v2.4b[3]\n" - "fmul v11.4s, v21.4s, v12.s[0]\n" - "fmul v23.4s, v21.4s, v12.s[1]\n" - "fmul v17.4s, v21.4s, v12.s[2]\n" - ".inst 0x4f9fe2c4 // sdot v4.4s, v22.16b, v31.4b[0]\n" - "fmul v6.4s, v21.4s, v12.s[3]\n" - ".inst 0x4fbfe2c1 // sdot v1.4s, v22.16b, v31.4b[1]\n" - ".inst 0x4f9feac0 // sdot v0.4s, v22.16b, v31.4b[2]\n" - ".inst 0x4fbfeade // sdot v30.4s, v22.16b, v31.4b[3]\n" - ".inst 0x4f9be124 // sdot v4.4s, v9.16b, v27.4b[0]\n" - ".inst 0x4fbbe121 // sdot v1.4s, v9.16b, v27.4b[1]\n" - ".inst 0x4f9be920 // sdot v0.4s, v9.16b, v27.4b[2]\n" - ".inst 0x4fbbe93e // sdot v30.4s, v9.16b, v27.4b[3]\n" - ".inst 0x4f9ae0e4 // sdot v4.4s, v7.16b, v26.4b[0]\n" - ".inst 0x4fbae0e1 // sdot v1.4s, v7.16b, v26.4b[1]\n" - ".inst 0x4f9ae8e0 // sdot v0.4s, v7.16b, v26.4b[2]\n" - ".inst 0x4fbae8fe // sdot v30.4s, v7.16b, v26.4b[3]\n" - ".inst 0x4f99e064 // sdot v4.4s, v3.16b, v25.4b[0]\n" - ".inst 0x4fb9e061 // sdot v1.4s, v3.16b, v25.4b[1]\n" - ".inst 0x4f99e860 // sdot v0.4s, v3.16b, v25.4b[2]\n" - ".inst 0x4fb9e87e // sdot v30.4s, v3.16b, v25.4b[3]\n" - ".inst 0x4f98e1a4 // sdot v4.4s, v13.16b, v24.4b[0]\n" - ".inst 0x4fb8e1a1 // sdot v1.4s, v13.16b, v24.4b[1]\n" - ".inst 0x4f98e9a0 // sdot v0.4s, v13.16b, v24.4b[2]\n" - ".inst 0x4fb8e9be // sdot v30.4s, v13.16b, v24.4b[3]\n" - ".inst 0x4f90e384 // sdot v4.4s, v28.16b, v16.4b[0]\n" - ".inst 0x4fb0e381 // sdot v1.4s, v28.16b, v16.4b[1]\n" - ".inst 0x4f90eb80 // sdot v0.4s, v28.16b, v16.4b[2]\n" - ".inst 0x4fb0eb9e // sdot v30.4s, v28.16b, v16.4b[3]\n" - "scvtf v4.4s, v4.4s, #0x4\n" - "scvtf v1.4s, v1.4s, #0x4\n" - "scvtf v0.4s, v0.4s, #0x4\n" - "fmla v15.4s, v4.4s, v11.4s\n" - "scvtf v30.4s, v30.4s, #0x4\n" - "fmla v19.4s, v1.4s, v23.4s\n" - "fmla v18.4s, v0.4s, v17.4s\n" - "fmla v14.4s, v30.4s, v6.4s\n" - "bgt 7b\n" - "mov x20, %x[res_ptr]\n" - "cmp x10, #0x1\n" - "str q15, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "ble 8f\n" - "cmp x10, #0x2\n" - "str q19, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "ble 8f\n" - "cmp x10, #0x3\n" - "str q18, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "ble 8f\n" - "str q14, [x20, #0x0]\n" - "8:" // Row tail: Accumulator store skip - "subs x23, x23, #0x4\n" - "add %x[res_ptr], %x[res_ptr], #0x10\n" - "bne 6b\n" - "subs x10, x10, #0x4\n" - "add %x[a_ptr], %x[a_ptr], x9\n" - "mov %x[res_ptr], x22\n" - "bgt 5b\n" - "9:" // Row tail: Row loop skip - : [a_ptr] "+&r" (a_ptr), [res_ptr] "+&r" (res_ptr) - : [b_ptr] "r" (b_ptr), [nr] "r" (nr), [nb] "r" (nb), [res_stride] "r" (res_stride), [nc] "r" (nc) - : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28" - ); - return; - } + __asm__ __volatile__( + "mov x10, %x[nr]\n" + "mov x9, #0x88\n" + "cmp x10, #0x10\n" + "mul x9, %x[nb], x9\n" + "blt 4f\n" + "1:" // Row loop + "add x28, %x[b_ptr], #0x8\n" + "mov x27, %x[nc]\n" + "add x26, %x[res_ptr], %x[res_stride], LSL #4\n" + "2:" // Column loop + "add x25, %x[a_ptr], #0x8\n" + "movi v15.16b, #0x0\n" + "movi v19.16b, #0x0\n" + "mov x24, %x[nb]\n" + "add x23, x25, x9\n" + "movi v18.16b, #0x0\n" + "movi v14.16b, #0x0\n" + "add x22, x23, x9\n" + "movi v11.16b, #0x0\n" + "movi v13.16b, #0x0\n" + "add x21, x22, x9\n" + "movi v23.16b, #0x0\n" + "movi v16.16b, #0x0\n" + "movi v25.16b, #0x0\n" + "movi v7.16b, #0x0\n" + "movi v0.16b, #0x0\n" + "movi v4.16b, #0x0\n" + "movi v5.16b, #0x0\n" + "movi v21.16b, #0x0\n" + "movi v8.16b, #0x0\n" + "movi v1.16b, #0x0\n" + "3:" // Block loop + "ldr q3, [x28, #0x0]\n" + "ldr q31, [x25, #0x0]\n" + "movi v28.16b, #0x4\n" + "movi v10.4s, #0x0\n" + "ldr q22, [x28, #0x10]\n" + "ldr q6, [x25, #0x10]\n" + "movi v29.4s, #0x0\n" + "movi v9.4s, #0x0\n" + "ldr q27, [x28, #0x20]\n" + "ldr q30, [x28, #0x30]\n" + "movi v20.4s, #0x0\n" + "movi v24.16b, #0xf0\n" + "ldr d2, [x25, #-0x8]\n" + "ldr d26, [x23, #-0x8]\n" + "sshl v12.16b, v3.16b, v28.16b\n" + "sub x20, x28, #0x8\n" + "ldr d17, [x20, #0x0]\n" + "and v3.16b, v3.16b, v24.16b\n" + "subs x24, x24, #0x1\n" + "add x28, x28, #0x48\n" + ".inst 0x4f9fe18a // sdot v10.4s, v12.16b, v31.4b[0]\n" + ".inst 0x4fbfe19d // sdot v29.4s, v12.16b, v31.4b[1]\n" + ".inst 0x4f9fe989 // sdot v9.4s, v12.16b, v31.4b[2]\n" + ".inst 0x4fbfe994 // sdot v20.4s, v12.16b, v31.4b[3]\n" + "sshl v31.16b, v22.16b, v28.16b\n" + "and v22.16b, v22.16b, v24.16b\n" + "fcvtl v17.4s, v17.4h\n" + "fcvtl v2.4s, v2.4h\n" + "fcvtl v26.4s, v26.4h\n" + ".inst 0x4f86e3ea // sdot v10.4s, v31.16b, v6.4b[0]\n" + ".inst 0x4fa6e3fd // sdot v29.4s, v31.16b, v6.4b[1]\n" + ".inst 0x4f86ebe9 // sdot v9.4s, v31.16b, v6.4b[2]\n" + ".inst 0x4fa6ebf4 // sdot v20.4s, v31.16b, v6.4b[3]\n" + "sshl v6.16b, v27.16b, v28.16b\n" + "sshl v28.16b, v30.16b, v28.16b\n" + "and v27.16b, v27.16b, v24.16b\n" + "and v30.16b, v30.16b, v24.16b\n" + "ldr q24, [x25, #0x20]\n" + ".inst 0x4f98e0ca // sdot v10.4s, v6.16b, v24.4b[0]\n" + ".inst 0x4fb8e0dd // sdot v29.4s, v6.16b, v24.4b[1]\n" + ".inst 0x4f98e8c9 // sdot v9.4s, v6.16b, v24.4b[2]\n" + ".inst 0x4fb8e8d4 // sdot v20.4s, v6.16b, v24.4b[3]\n" + "ldr q24, [x25, #0x30]\n" + ".inst 0x4f98e38a // sdot v10.4s, v28.16b, v24.4b[0]\n" + ".inst 0x4fb8e39d // sdot v29.4s, v28.16b, v24.4b[1]\n" + ".inst 0x4f98eb89 // sdot v9.4s, v28.16b, v24.4b[2]\n" + ".inst 0x4fb8eb94 // sdot v20.4s, v28.16b, v24.4b[3]\n" + "ldr q24, [x25, #0x40]\n" + ".inst 0x4f98e06a // sdot v10.4s, v3.16b, v24.4b[0]\n" + ".inst 0x4fb8e07d // sdot v29.4s, v3.16b, v24.4b[1]\n" + ".inst 0x4f98e869 // sdot v9.4s, v3.16b, v24.4b[2]\n" + ".inst 0x4fb8e874 // sdot v20.4s, v3.16b, v24.4b[3]\n" + "ldr q24, [x25, #0x50]\n" + ".inst 0x4f98e2ca // sdot v10.4s, v22.16b, v24.4b[0]\n" + ".inst 0x4fb8e2dd // sdot v29.4s, v22.16b, v24.4b[1]\n" + ".inst 0x4f98eac9 // sdot v9.4s, v22.16b, v24.4b[2]\n" + ".inst 0x4fb8ead4 // sdot v20.4s, v22.16b, v24.4b[3]\n" + "ldr q24, [x25, #0x60]\n" + ".inst 0x4f98e36a // sdot v10.4s, v27.16b, v24.4b[0]\n" + ".inst 0x4fb8e37d // sdot v29.4s, v27.16b, v24.4b[1]\n" + ".inst 0x4f98eb69 // sdot v9.4s, v27.16b, v24.4b[2]\n" + ".inst 0x4fb8eb74 // sdot v20.4s, v27.16b, v24.4b[3]\n" + "ldr q24, [x25, #0x70]\n" + "add x25, x25, #0x88\n" + ".inst 0x4f98e3ca // sdot v10.4s, v30.16b, v24.4b[0]\n" + ".inst 0x4fb8e3dd // sdot v29.4s, v30.16b, v24.4b[1]\n" + ".inst 0x4f98ebc9 // sdot v9.4s, v30.16b, v24.4b[2]\n" + ".inst 0x4fb8ebd4 // sdot v20.4s, v30.16b, v24.4b[3]\n" + "fmul v24.4s, v17.4s, v2.s[0]\n" + "scvtf v10.4s, v10.4s, #0x4\n" + "scvtf v29.4s, v29.4s, #0x4\n" + "scvtf v9.4s, v9.4s, #0x4\n" + "scvtf v20.4s, v20.4s, #0x4\n" + "fmla v15.4s, v10.4s, v24.4s\n" + "ldr q24, [x23, #0x0]\n" + "fmul v10.4s, v17.4s, v2.s[1]\n" + "fmla v19.4s, v29.4s, v10.4s\n" + "ldr q10, [x23, #0x10]\n" + "fmul v29.4s, v17.4s, v2.s[2]\n" + "fmul v2.4s, v17.4s, v2.s[3]\n" + "fmla v18.4s, v9.4s, v29.4s\n" + "movi v9.4s, #0x0\n" + "movi v29.4s, #0x0\n" + ".inst 0x4f98e189 // sdot v9.4s, v12.16b, v24.4b[0]\n" + ".inst 0x4fb8e19d // sdot v29.4s, v12.16b, v24.4b[1]\n" + "fmla v14.4s, v20.4s, v2.4s\n" + "movi v20.4s, #0x0\n" + "movi v2.4s, #0x0\n" + ".inst 0x4f98e994 // sdot v20.4s, v12.16b, v24.4b[2]\n" + ".inst 0x4fb8e982 // sdot v2.4s, v12.16b, v24.4b[3]\n" + "ldr q24, [x23, #0x20]\n" + ".inst 0x4f8ae3e9 // sdot v9.4s, v31.16b, v10.4b[0]\n" + ".inst 0x4faae3fd // sdot v29.4s, v31.16b, v10.4b[1]\n" + ".inst 0x4f8aebf4 // sdot v20.4s, v31.16b, v10.4b[2]\n" + ".inst 0x4faaebe2 // sdot v2.4s, v31.16b, v10.4b[3]\n" + "ldr q10, [x23, #0x30]\n" + ".inst 0x4f98e0c9 // sdot v9.4s, v6.16b, v24.4b[0]\n" + ".inst 0x4fb8e0dd // sdot v29.4s, v6.16b, v24.4b[1]\n" + ".inst 0x4f98e8d4 // sdot v20.4s, v6.16b, v24.4b[2]\n" + ".inst 0x4fb8e8c2 // sdot v2.4s, v6.16b, v24.4b[3]\n" + "ldr q24, [x23, #0x40]\n" + ".inst 0x4f8ae389 // sdot v9.4s, v28.16b, v10.4b[0]\n" + ".inst 0x4faae39d // sdot v29.4s, v28.16b, v10.4b[1]\n" + ".inst 0x4f8aeb94 // sdot v20.4s, v28.16b, v10.4b[2]\n" + ".inst 0x4faaeb82 // sdot v2.4s, v28.16b, v10.4b[3]\n" + "ldr q10, [x23, #0x50]\n" + ".inst 0x4f98e069 // sdot v9.4s, v3.16b, v24.4b[0]\n" + ".inst 0x4fb8e07d // sdot v29.4s, v3.16b, v24.4b[1]\n" + ".inst 0x4f98e874 // sdot v20.4s, v3.16b, v24.4b[2]\n" + ".inst 0x4fb8e862 // sdot v2.4s, v3.16b, v24.4b[3]\n" + "ldr q24, [x23, #0x60]\n" + ".inst 0x4f8ae2c9 // sdot v9.4s, v22.16b, v10.4b[0]\n" + ".inst 0x4faae2dd // sdot v29.4s, v22.16b, v10.4b[1]\n" + ".inst 0x4f8aead4 // sdot v20.4s, v22.16b, v10.4b[2]\n" + ".inst 0x4faaeac2 // sdot v2.4s, v22.16b, v10.4b[3]\n" + "ldr q10, [x23, #0x70]\n" + "add x23, x23, #0x88\n" + ".inst 0x4f98e369 // sdot v9.4s, v27.16b, v24.4b[0]\n" + ".inst 0x4fb8e37d // sdot v29.4s, v27.16b, v24.4b[1]\n" + ".inst 0x4f98eb74 // sdot v20.4s, v27.16b, v24.4b[2]\n" + ".inst 0x4fb8eb62 // sdot v2.4s, v27.16b, v24.4b[3]\n" + "ldr q24, [x22, #0x0]\n" + ".inst 0x4f8ae3c9 // sdot v9.4s, v30.16b, v10.4b[0]\n" + ".inst 0x4faae3dd // sdot v29.4s, v30.16b, v10.4b[1]\n" + ".inst 0x4f8aebd4 // sdot v20.4s, v30.16b, v10.4b[2]\n" + ".inst 0x4faaebc2 // sdot v2.4s, v30.16b, v10.4b[3]\n" + "fmul v10.4s, v17.4s, v26.s[0]\n" + "scvtf v9.4s, v9.4s, #0x4\n" + "scvtf v29.4s, v29.4s, #0x4\n" + "scvtf v20.4s, v20.4s, #0x4\n" + "scvtf v2.4s, v2.4s, #0x4\n" + "fmla v11.4s, v9.4s, v10.4s\n" + "ldr q9, [x22, #0x10]\n" + "fmul v10.4s, v17.4s, v26.s[1]\n" + "fmla v13.4s, v29.4s, v10.4s\n" + "ldr d29, [x22, #-0x8]\n" + "fmul v10.4s, v17.4s, v26.s[2]\n" + "fmul v26.4s, v17.4s, v26.s[3]\n" + "fcvtl v29.4s, v29.4h\n" + "fmla v23.4s, v20.4s, v10.4s\n" + "movi v20.4s, #0x0\n" + "movi v10.4s, #0x0\n" + "fmla v16.4s, v2.4s, v26.4s\n" + "movi v26.4s, #0x0\n" + "movi v2.4s, #0x0\n" + ".inst 0x4f98e194 // sdot v20.4s, v12.16b, v24.4b[0]\n" + ".inst 0x4fb8e18a // sdot v10.4s, v12.16b, v24.4b[1]\n" + ".inst 0x4f98e99a // sdot v26.4s, v12.16b, v24.4b[2]\n" + ".inst 0x4fb8e982 // sdot v2.4s, v12.16b, v24.4b[3]\n" + "ldr q24, [x22, #0x20]\n" + ".inst 0x4f89e3f4 // sdot v20.4s, v31.16b, v9.4b[0]\n" + ".inst 0x4fa9e3ea // sdot v10.4s, v31.16b, v9.4b[1]\n" + ".inst 0x4f89ebfa // sdot v26.4s, v31.16b, v9.4b[2]\n" + ".inst 0x4fa9ebe2 // sdot v2.4s, v31.16b, v9.4b[3]\n" + "ldr q9, [x22, #0x30]\n" + ".inst 0x4f98e0d4 // sdot v20.4s, v6.16b, v24.4b[0]\n" + ".inst 0x4fb8e0ca // sdot v10.4s, v6.16b, v24.4b[1]\n" + ".inst 0x4f98e8da // sdot v26.4s, v6.16b, v24.4b[2]\n" + ".inst 0x4fb8e8c2 // sdot v2.4s, v6.16b, v24.4b[3]\n" + "ldr q24, [x22, #0x40]\n" + ".inst 0x4f89e394 // sdot v20.4s, v28.16b, v9.4b[0]\n" + ".inst 0x4fa9e38a // sdot v10.4s, v28.16b, v9.4b[1]\n" + ".inst 0x4f89eb9a // sdot v26.4s, v28.16b, v9.4b[2]\n" + ".inst 0x4fa9eb82 // sdot v2.4s, v28.16b, v9.4b[3]\n" + "ldr q9, [x22, #0x50]\n" + ".inst 0x4f98e074 // sdot v20.4s, v3.16b, v24.4b[0]\n" + ".inst 0x4fb8e06a // sdot v10.4s, v3.16b, v24.4b[1]\n" + ".inst 0x4f98e87a // sdot v26.4s, v3.16b, v24.4b[2]\n" + ".inst 0x4fb8e862 // sdot v2.4s, v3.16b, v24.4b[3]\n" + "ldr q24, [x22, #0x60]\n" + ".inst 0x4f89e2d4 // sdot v20.4s, v22.16b, v9.4b[0]\n" + ".inst 0x4fa9e2ca // sdot v10.4s, v22.16b, v9.4b[1]\n" + ".inst 0x4f89eada // sdot v26.4s, v22.16b, v9.4b[2]\n" + ".inst 0x4fa9eac2 // sdot v2.4s, v22.16b, v9.4b[3]\n" + "ldr q9, [x22, #0x70]\n" + "add x22, x22, #0x88\n" + ".inst 0x4f98e374 // sdot v20.4s, v27.16b, v24.4b[0]\n" + ".inst 0x4fb8e36a // sdot v10.4s, v27.16b, v24.4b[1]\n" + ".inst 0x4f98eb7a // sdot v26.4s, v27.16b, v24.4b[2]\n" + ".inst 0x4fb8eb62 // sdot v2.4s, v27.16b, v24.4b[3]\n" + "ldr q24, [x21, #0x0]\n" + ".inst 0x4f89e3d4 // sdot v20.4s, v30.16b, v9.4b[0]\n" + ".inst 0x4fa9e3ca // sdot v10.4s, v30.16b, v9.4b[1]\n" + ".inst 0x4f89ebda // sdot v26.4s, v30.16b, v9.4b[2]\n" + ".inst 0x4fa9ebc2 // sdot v2.4s, v30.16b, v9.4b[3]\n" + "fmul v9.4s, v17.4s, v29.s[0]\n" + "scvtf v20.4s, v20.4s, #0x4\n" + "scvtf v10.4s, v10.4s, #0x4\n" + "scvtf v26.4s, v26.4s, #0x4\n" + "scvtf v2.4s, v2.4s, #0x4\n" + "fmla v25.4s, v20.4s, v9.4s\n" + "ldr q9, [x21, #0x10]\n" + "fmul v20.4s, v17.4s, v29.s[1]\n" + "fmla v7.4s, v10.4s, v20.4s\n" + "ldr d20, [x21, #-0x8]\n" + "fmul v10.4s, v17.4s, v29.s[2]\n" + "fmul v29.4s, v17.4s, v29.s[3]\n" + "fcvtl v20.4s, v20.4h\n" + "fmla v0.4s, v26.4s, v10.4s\n" + "movi v26.4s, #0x0\n" + "movi v10.4s, #0x0\n" + "fmla v4.4s, v2.4s, v29.4s\n" + "movi v2.4s, #0x0\n" + "movi v29.4s, #0x0\n" + ".inst 0x4f98e19a // sdot v26.4s, v12.16b, v24.4b[0]\n" + ".inst 0x4fb8e18a // sdot v10.4s, v12.16b, v24.4b[1]\n" + ".inst 0x4f98e982 // sdot v2.4s, v12.16b, v24.4b[2]\n" + ".inst 0x4fb8e99d // sdot v29.4s, v12.16b, v24.4b[3]\n" + "ldr q12, [x21, #0x20]\n" + "fmul v24.4s, v17.4s, v20.s[0]\n" + ".inst 0x4f89e3fa // sdot v26.4s, v31.16b, v9.4b[0]\n" + ".inst 0x4fa9e3ea // sdot v10.4s, v31.16b, v9.4b[1]\n" + ".inst 0x4f89ebe2 // sdot v2.4s, v31.16b, v9.4b[2]\n" + ".inst 0x4fa9ebfd // sdot v29.4s, v31.16b, v9.4b[3]\n" + "ldr q9, [x21, #0x30]\n" + "fmul v31.4s, v17.4s, v20.s[1]\n" + ".inst 0x4f8ce0da // sdot v26.4s, v6.16b, v12.4b[0]\n" + ".inst 0x4face0ca // sdot v10.4s, v6.16b, v12.4b[1]\n" + ".inst 0x4f8ce8c2 // sdot v2.4s, v6.16b, v12.4b[2]\n" + ".inst 0x4face8dd // sdot v29.4s, v6.16b, v12.4b[3]\n" + "ldr q12, [x21, #0x40]\n" + "fmul v6.4s, v17.4s, v20.s[2]\n" + "fmul v20.4s, v17.4s, v20.s[3]\n" + ".inst 0x4f89e39a // sdot v26.4s, v28.16b, v9.4b[0]\n" + ".inst 0x4fa9e38a // sdot v10.4s, v28.16b, v9.4b[1]\n" + ".inst 0x4f89eb82 // sdot v2.4s, v28.16b, v9.4b[2]\n" + ".inst 0x4fa9eb9d // sdot v29.4s, v28.16b, v9.4b[3]\n" + "ldr q9, [x21, #0x50]\n" + ".inst 0x4f8ce07a // sdot v26.4s, v3.16b, v12.4b[0]\n" + ".inst 0x4face06a // sdot v10.4s, v3.16b, v12.4b[1]\n" + ".inst 0x4f8ce862 // sdot v2.4s, v3.16b, v12.4b[2]\n" + ".inst 0x4face87d // sdot v29.4s, v3.16b, v12.4b[3]\n" + "ldr q12, [x21, #0x60]\n" + ".inst 0x4f89e2da // sdot v26.4s, v22.16b, v9.4b[0]\n" + ".inst 0x4fa9e2ca // sdot v10.4s, v22.16b, v9.4b[1]\n" + ".inst 0x4f89eac2 // sdot v2.4s, v22.16b, v9.4b[2]\n" + ".inst 0x4fa9eadd // sdot v29.4s, v22.16b, v9.4b[3]\n" + "ldr q17, [x21, #0x70]\n" + "add x21, x21, #0x88\n" + ".inst 0x4f8ce37a // sdot v26.4s, v27.16b, v12.4b[0]\n" + ".inst 0x4face36a // sdot v10.4s, v27.16b, v12.4b[1]\n" + ".inst 0x4f8ceb62 // sdot v2.4s, v27.16b, v12.4b[2]\n" + ".inst 0x4faceb7d // sdot v29.4s, v27.16b, v12.4b[3]\n" + ".inst 0x4f91e3da // sdot v26.4s, v30.16b, v17.4b[0]\n" + ".inst 0x4fb1e3ca // sdot v10.4s, v30.16b, v17.4b[1]\n" + ".inst 0x4f91ebc2 // sdot v2.4s, v30.16b, v17.4b[2]\n" + ".inst 0x4fb1ebdd // sdot v29.4s, v30.16b, v17.4b[3]\n" + "scvtf v26.4s, v26.4s, #0x4\n" + "scvtf v10.4s, v10.4s, #0x4\n" + "fmla v5.4s, v26.4s, v24.4s\n" + "scvtf v2.4s, v2.4s, #0x4\n" + "scvtf v29.4s, v29.4s, #0x4\n" + "fmla v21.4s, v10.4s, v31.4s\n" + "fmla v8.4s, v2.4s, v6.4s\n" + "fmla v1.4s, v29.4s, v20.4s\n" + "bgt 3b\n" + "mov x20, %x[res_ptr]\n" + "subs x27, x27, #0x4\n" + "add %x[res_ptr], %x[res_ptr], #0x10\n" + "str q15, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q19, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q18, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q14, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q11, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q13, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q23, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q16, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q25, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q7, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q0, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q4, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q5, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q21, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q8, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q1, [x20, #0x0]\n" + "bne 2b\n" + "mov x20, #0x4\n" + "sub x10, x10, #0x10\n" + "cmp x10, #0x10\n" + "mov %x[res_ptr], x26\n" + "madd %x[a_ptr], x20, x9, %x[a_ptr]\n" + "bge 1b\n" + "4:" // Row loop skip + "cbz x10, 9f\n" + "5:" // Row tail: Row loop + "add x24, %x[b_ptr], #0x8\n" + "mov x23, %x[nc]\n" + "add x22, %x[res_ptr], %x[res_stride], LSL #2\n" + "6:" // Row tail: Column loop + "movi v15.16b, #0x0\n" + "movi v19.16b, #0x0\n" + "add x25, %x[a_ptr], #0x8\n" + "mov x21, %x[nb]\n" + "movi v18.16b, #0x0\n" + "movi v14.16b, #0x0\n" + "7:" // Row tail: Block loop + "ldr q7, [x24, #0x0]\n" + "ldr q5, [x25, #0x0]\n" + "movi v9.16b, #0x4\n" + "movi v4.4s, #0x0\n" + "ldr q3, [x24, #0x10]\n" + "ldr q2, [x25, #0x10]\n" + "movi v1.4s, #0x0\n" + "movi v0.4s, #0x0\n" + "ldr q13, [x24, #0x20]\n" + "ldr q31, [x25, #0x20]\n" + "movi v30.4s, #0x0\n" + "movi v29.16b, #0xf0\n" + "ldr q28, [x24, #0x30]\n" + "ldr q27, [x25, #0x30]\n" + "sshl v20.16b, v7.16b, v9.16b\n" + "sub x20, x24, #0x8\n" + "ldr q26, [x25, #0x40]\n" + "ldr q25, [x25, #0x50]\n" + "sshl v17.16b, v3.16b, v9.16b\n" + "and v7.16b, v7.16b, v29.16b\n" + "ldr q24, [x25, #0x60]\n" + "ldr q16, [x25, #0x70]\n" + "sshl v22.16b, v13.16b, v9.16b\n" + "and v3.16b, v3.16b, v29.16b\n" + "ldr d21, [x20, #0x0]\n" + "ldr d12, [x25, #-0x8]\n" + ".inst 0x4f85e284 // sdot v4.4s, v20.16b, v5.4b[0]\n" + ".inst 0x4fa5e281 // sdot v1.4s, v20.16b, v5.4b[1]\n" + ".inst 0x4f85ea80 // sdot v0.4s, v20.16b, v5.4b[2]\n" + ".inst 0x4fa5ea9e // sdot v30.4s, v20.16b, v5.4b[3]\n" + "sshl v9.16b, v28.16b, v9.16b\n" + "subs x21, x21, #0x1\n" + "and v13.16b, v13.16b, v29.16b\n" + "and v28.16b, v28.16b, v29.16b\n" + "add x25, x25, #0x88\n" + "add x24, x24, #0x48\n" + "fcvtl v21.4s, v21.4h\n" + "fcvtl v12.4s, v12.4h\n" + ".inst 0x4f82e224 // sdot v4.4s, v17.16b, v2.4b[0]\n" + ".inst 0x4fa2e221 // sdot v1.4s, v17.16b, v2.4b[1]\n" + ".inst 0x4f82ea20 // sdot v0.4s, v17.16b, v2.4b[2]\n" + ".inst 0x4fa2ea3e // sdot v30.4s, v17.16b, v2.4b[3]\n" + "fmul v11.4s, v21.4s, v12.s[0]\n" + "fmul v23.4s, v21.4s, v12.s[1]\n" + "fmul v17.4s, v21.4s, v12.s[2]\n" + ".inst 0x4f9fe2c4 // sdot v4.4s, v22.16b, v31.4b[0]\n" + "fmul v6.4s, v21.4s, v12.s[3]\n" + ".inst 0x4fbfe2c1 // sdot v1.4s, v22.16b, v31.4b[1]\n" + ".inst 0x4f9feac0 // sdot v0.4s, v22.16b, v31.4b[2]\n" + ".inst 0x4fbfeade // sdot v30.4s, v22.16b, v31.4b[3]\n" + ".inst 0x4f9be124 // sdot v4.4s, v9.16b, v27.4b[0]\n" + ".inst 0x4fbbe121 // sdot v1.4s, v9.16b, v27.4b[1]\n" + ".inst 0x4f9be920 // sdot v0.4s, v9.16b, v27.4b[2]\n" + ".inst 0x4fbbe93e // sdot v30.4s, v9.16b, v27.4b[3]\n" + ".inst 0x4f9ae0e4 // sdot v4.4s, v7.16b, v26.4b[0]\n" + ".inst 0x4fbae0e1 // sdot v1.4s, v7.16b, v26.4b[1]\n" + ".inst 0x4f9ae8e0 // sdot v0.4s, v7.16b, v26.4b[2]\n" + ".inst 0x4fbae8fe // sdot v30.4s, v7.16b, v26.4b[3]\n" + ".inst 0x4f99e064 // sdot v4.4s, v3.16b, v25.4b[0]\n" + ".inst 0x4fb9e061 // sdot v1.4s, v3.16b, v25.4b[1]\n" + ".inst 0x4f99e860 // sdot v0.4s, v3.16b, v25.4b[2]\n" + ".inst 0x4fb9e87e // sdot v30.4s, v3.16b, v25.4b[3]\n" + ".inst 0x4f98e1a4 // sdot v4.4s, v13.16b, v24.4b[0]\n" + ".inst 0x4fb8e1a1 // sdot v1.4s, v13.16b, v24.4b[1]\n" + ".inst 0x4f98e9a0 // sdot v0.4s, v13.16b, v24.4b[2]\n" + ".inst 0x4fb8e9be // sdot v30.4s, v13.16b, v24.4b[3]\n" + ".inst 0x4f90e384 // sdot v4.4s, v28.16b, v16.4b[0]\n" + ".inst 0x4fb0e381 // sdot v1.4s, v28.16b, v16.4b[1]\n" + ".inst 0x4f90eb80 // sdot v0.4s, v28.16b, v16.4b[2]\n" + ".inst 0x4fb0eb9e // sdot v30.4s, v28.16b, v16.4b[3]\n" + "scvtf v4.4s, v4.4s, #0x4\n" + "scvtf v1.4s, v1.4s, #0x4\n" + "scvtf v0.4s, v0.4s, #0x4\n" + "fmla v15.4s, v4.4s, v11.4s\n" + "scvtf v30.4s, v30.4s, #0x4\n" + "fmla v19.4s, v1.4s, v23.4s\n" + "fmla v18.4s, v0.4s, v17.4s\n" + "fmla v14.4s, v30.4s, v6.4s\n" + "bgt 7b\n" + "mov x20, %x[res_ptr]\n" + "cmp x10, #0x1\n" + "str q15, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "ble 8f\n" + "cmp x10, #0x2\n" + "str q19, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "ble 8f\n" + "cmp x10, #0x3\n" + "str q18, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "ble 8f\n" + "str q14, [x20, #0x0]\n" + "8:" // Row tail: Accumulator store skip + "subs x23, x23, #0x4\n" + "add %x[res_ptr], %x[res_ptr], #0x10\n" + "bne 6b\n" + "subs x10, x10, #0x4\n" + "add %x[a_ptr], %x[a_ptr], x9\n" + "mov %x[res_ptr], x22\n" + "bgt 5b\n" + "9:" // Row tail: Row loop skip + : [a_ptr] "+&r" (a_ptr), [res_ptr] "+&r" (res_ptr) + : [b_ptr] "r" (b_ptr), [nr] "r" (nr), [nb] "r" (nb), [res_stride] "r" (res_stride), [nc] "r" (nc) + : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28" + ); + return; #endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) { float sumf[4][4]; @@ -1160,404 +1152,402 @@ void ggml_gemm_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo UNUSED(blocklen); #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8) - if (ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) { - const void * b_ptr = vx; - const void * a_ptr = vy; - float * res_ptr = s; - size_t res_stride = bs * sizeof(float); + const void * b_ptr = vx; + const void * a_ptr = vy; + float * res_ptr = s; + size_t res_stride = bs * sizeof(float); - __asm__ __volatile__( - "mov x10, %x[nr]\n" - "mov x9, #0x88\n" - "cmp x10, #0x10\n" - "mul x9, %x[nb], x9\n" - "blt 4f\n" - "1:" // Row loop - "add x28, %x[b_ptr], #0x8\n" - "mov x27, %x[nc]\n" - "add x26, %x[res_ptr], %x[res_stride], LSL #4\n" - "2:" // Column loop - "add x25, %x[a_ptr], #0x8\n" - "movi v2.16b, #0x0\n" - "movi v10.16b, #0x0\n" - "mov x24, %x[nb]\n" - "add x23, x25, x9\n" - "movi v12.16b, #0x0\n" - "movi v28.16b, #0x0\n" - "add x22, x23, x9\n" - "movi v11.16b, #0x0\n" - "movi v13.16b, #0x0\n" - "add x21, x22, x9\n" - "movi v22.16b, #0x0\n" - "movi v23.16b, #0x0\n" - "movi v25.16b, #0x0\n" - "movi v5.16b, #0x0\n" - "movi v7.16b, #0x0\n" - "movi v4.16b, #0x0\n" - "movi v6.16b, #0x0\n" - "movi v30.16b, #0x0\n" - "movi v24.16b, #0x0\n" - "movi v14.16b, #0x0\n" - "3:" // Block loop - "ldr q21, [x28, #0x0]\n" - "ldr q16, [x28, #0x10]\n" - "movi v1.16b, #0x4\n" - "movi v19.4s, #0x0\n" - "ldr q27, [x25, #0x0]\n" - "ldr q15, [x25, #0x10]\n" - "movi v26.4s, #0x0\n" - "movi v18.4s, #0x0\n" - "ldr q29, [x28, #0x20]\n" - "ldr q3, [x28, #0x30]\n" - "movi v17.4s, #0x0\n" - "movi v0.16b, #0xf0\n" - "ldr d20, [x25, #-0x8]\n" - "ldr d9, [x23, #-0x8]\n" - "sshl v8.16b, v21.16b, v1.16b\n" - "sshl v31.16b, v16.16b, v1.16b\n" - "and v21.16b, v21.16b, v0.16b\n" - "and v16.16b, v16.16b, v0.16b\n" - "sub x20, x28, #0x8\n" - "subs x24, x24, #0x1\n" - "add x28, x28, #0x48\n" - ".inst 0x4e88a773 // smmla v19.4s, v27.16b, v8.16b\n" - ".inst 0x4e9fa77a // smmla v26.4s, v27.16b, v31.16b\n" - "ldr q27, [x25, #0x20]\n" - ".inst 0x4e88a5f2 // smmla v18.4s, v15.16b, v8.16b\n" - ".inst 0x4e9fa5f1 // smmla v17.4s, v15.16b, v31.16b\n" - "sshl v15.16b, v29.16b, v1.16b\n" - "sshl v1.16b, v3.16b, v1.16b\n" - "and v29.16b, v29.16b, v0.16b\n" - "and v3.16b, v3.16b, v0.16b\n" - "ldr q0, [x25, #0x30]\n" - "fcvtl v20.4s, v20.4h\n" - ".inst 0x4e8fa773 // smmla v19.4s, v27.16b, v15.16b\n" - "fcvtl v9.4s, v9.4h\n" - ".inst 0x4e81a77a // smmla v26.4s, v27.16b, v1.16b\n" - "ldr q27, [x25, #0x40]\n" - ".inst 0x4e8fa412 // smmla v18.4s, v0.16b, v15.16b\n" - ".inst 0x4e81a411 // smmla v17.4s, v0.16b, v1.16b\n" - "ldr q0, [x25, #0x50]\n" - ".inst 0x4e95a773 // smmla v19.4s, v27.16b, v21.16b\n" - ".inst 0x4e90a77a // smmla v26.4s, v27.16b, v16.16b\n" - "ldr q27, [x25, #0x60]\n" - ".inst 0x4e95a412 // smmla v18.4s, v0.16b, v21.16b\n" - ".inst 0x4e90a411 // smmla v17.4s, v0.16b, v16.16b\n" - "ldr q0, [x25, #0x70]\n" - "add x25, x25, #0x88\n" - ".inst 0x4e9da773 // smmla v19.4s, v27.16b, v29.16b\n" - ".inst 0x4e83a77a // smmla v26.4s, v27.16b, v3.16b\n" - "ldr d27, [x20, #0x0]\n" - ".inst 0x4e9da412 // smmla v18.4s, v0.16b, v29.16b\n" - ".inst 0x4e83a411 // smmla v17.4s, v0.16b, v3.16b\n" - "fcvtl v27.4s, v27.4h\n" - "uzp1 v0.2d, v19.2d, v26.2d\n" - "uzp2 v26.2d, v19.2d, v26.2d\n" - "fmul v19.4s, v27.4s, v20.s[0]\n" - "scvtf v0.4s, v0.4s, #0x4\n" - "scvtf v26.4s, v26.4s, #0x4\n" - "fmla v2.4s, v0.4s, v19.4s\n" - "ldr q19, [x23, #0x0]\n" - "uzp1 v0.2d, v18.2d, v17.2d\n" - "uzp2 v18.2d, v18.2d, v17.2d\n" - "fmul v17.4s, v27.4s, v20.s[1]\n" - "scvtf v0.4s, v0.4s, #0x4\n" - "scvtf v18.4s, v18.4s, #0x4\n" - "fmla v10.4s, v26.4s, v17.4s\n" - "ldr q17, [x23, #0x10]\n" - "fmul v26.4s, v27.4s, v20.s[2]\n" - "fmul v20.4s, v27.4s, v20.s[3]\n" - "fmla v12.4s, v0.4s, v26.4s\n" - "ldr d0, [x22, #-0x8]\n" - "ldr d26, [x21, #-0x8]\n" - "fcvtl v0.4s, v0.4h\n" - "fmla v28.4s, v18.4s, v20.4s\n" - "movi v20.4s, #0x0\n" - "movi v18.4s, #0x0\n" - ".inst 0x4e88a674 // smmla v20.4s, v19.16b, v8.16b\n" - ".inst 0x4e9fa672 // smmla v18.4s, v19.16b, v31.16b\n" - "ldr q19, [x23, #0x20]\n" - "fcvtl v26.4s, v26.4h\n" - ".inst 0x4e8fa674 // smmla v20.4s, v19.16b, v15.16b\n" - ".inst 0x4e81a672 // smmla v18.4s, v19.16b, v1.16b\n" - "ldr q19, [x23, #0x40]\n" - ".inst 0x4e95a674 // smmla v20.4s, v19.16b, v21.16b\n" - ".inst 0x4e90a672 // smmla v18.4s, v19.16b, v16.16b\n" - "ldr q19, [x23, #0x60]\n" - ".inst 0x4e9da674 // smmla v20.4s, v19.16b, v29.16b\n" - ".inst 0x4e83a672 // smmla v18.4s, v19.16b, v3.16b\n" - "uzp1 v19.2d, v20.2d, v18.2d\n" - "scvtf v19.4s, v19.4s, #0x4\n" - "uzp2 v20.2d, v20.2d, v18.2d\n" - "fmul v18.4s, v27.4s, v9.s[0]\n" - "scvtf v20.4s, v20.4s, #0x4\n" - "fmla v11.4s, v19.4s, v18.4s\n" - "ldr q18, [x22, #0x0]\n" - "fmul v19.4s, v27.4s, v9.s[1]\n" - "fmla v13.4s, v20.4s, v19.4s\n" - "movi v19.4s, #0x0\n" - "movi v20.4s, #0x0\n" - ".inst 0x4e88a633 // smmla v19.4s, v17.16b, v8.16b\n" - ".inst 0x4e9fa634 // smmla v20.4s, v17.16b, v31.16b\n" - "ldr q17, [x23, #0x30]\n" - ".inst 0x4e8fa633 // smmla v19.4s, v17.16b, v15.16b\n" - ".inst 0x4e81a634 // smmla v20.4s, v17.16b, v1.16b\n" - "ldr q17, [x23, #0x50]\n" - ".inst 0x4e95a633 // smmla v19.4s, v17.16b, v21.16b\n" - ".inst 0x4e90a634 // smmla v20.4s, v17.16b, v16.16b\n" - "ldr q17, [x23, #0x70]\n" - "add x23, x23, #0x88\n" - ".inst 0x4e9da633 // smmla v19.4s, v17.16b, v29.16b\n" - ".inst 0x4e83a634 // smmla v20.4s, v17.16b, v3.16b\n" - "uzp1 v17.2d, v19.2d, v20.2d\n" - "scvtf v17.4s, v17.4s, #0x4\n" - "uzp2 v20.2d, v19.2d, v20.2d\n" - "fmul v19.4s, v27.4s, v9.s[2]\n" - "fmul v9.4s, v27.4s, v9.s[3]\n" - "scvtf v20.4s, v20.4s, #0x4\n" - "fmla v22.4s, v17.4s, v19.4s\n" - "ldr q17, [x22, #0x10]\n" - "movi v19.4s, #0x0\n" - ".inst 0x4e88a653 // smmla v19.4s, v18.16b, v8.16b\n" - "fmla v23.4s, v20.4s, v9.4s\n" - "movi v20.4s, #0x0\n" - "movi v9.4s, #0x0\n" - ".inst 0x4e9fa654 // smmla v20.4s, v18.16b, v31.16b\n" - "ldr q18, [x22, #0x20]\n" - ".inst 0x4e88a629 // smmla v9.4s, v17.16b, v8.16b\n" - ".inst 0x4e8fa653 // smmla v19.4s, v18.16b, v15.16b\n" - ".inst 0x4e81a654 // smmla v20.4s, v18.16b, v1.16b\n" - "ldr q18, [x22, #0x40]\n" - ".inst 0x4e95a653 // smmla v19.4s, v18.16b, v21.16b\n" - ".inst 0x4e90a654 // smmla v20.4s, v18.16b, v16.16b\n" - "ldr q18, [x22, #0x60]\n" - ".inst 0x4e9da653 // smmla v19.4s, v18.16b, v29.16b\n" - ".inst 0x4e83a654 // smmla v20.4s, v18.16b, v3.16b\n" - "movi v18.4s, #0x0\n" - ".inst 0x4e9fa632 // smmla v18.4s, v17.16b, v31.16b\n" - "ldr q17, [x22, #0x30]\n" - ".inst 0x4e8fa629 // smmla v9.4s, v17.16b, v15.16b\n" - ".inst 0x4e81a632 // smmla v18.4s, v17.16b, v1.16b\n" - "ldr q17, [x22, #0x50]\n" - ".inst 0x4e95a629 // smmla v9.4s, v17.16b, v21.16b\n" - ".inst 0x4e90a632 // smmla v18.4s, v17.16b, v16.16b\n" - "ldr q17, [x22, #0x70]\n" - "add x22, x22, #0x88\n" - ".inst 0x4e9da629 // smmla v9.4s, v17.16b, v29.16b\n" - ".inst 0x4e83a632 // smmla v18.4s, v17.16b, v3.16b\n" - "uzp1 v17.2d, v19.2d, v20.2d\n" - "uzp2 v20.2d, v19.2d, v20.2d\n" - "fmul v19.4s, v27.4s, v0.s[0]\n" - "scvtf v17.4s, v17.4s, #0x4\n" - "scvtf v20.4s, v20.4s, #0x4\n" - "fmla v25.4s, v17.4s, v19.4s\n" - "ldr q19, [x21, #0x0]\n" - "fmul v17.4s, v27.4s, v0.s[1]\n" - "fmla v5.4s, v20.4s, v17.4s\n" - "ldr q17, [x21, #0x10]\n" - "uzp1 v20.2d, v9.2d, v18.2d\n" - "uzp2 v9.2d, v9.2d, v18.2d\n" - "fmul v18.4s, v27.4s, v0.s[2]\n" - "fmul v0.4s, v27.4s, v0.s[3]\n" - "scvtf v20.4s, v20.4s, #0x4\n" - "scvtf v9.4s, v9.4s, #0x4\n" - "fmla v7.4s, v20.4s, v18.4s\n" - "movi v20.4s, #0x0\n" - "movi v18.4s, #0x0\n" - ".inst 0x4e88a674 // smmla v20.4s, v19.16b, v8.16b\n" - ".inst 0x4e9fa672 // smmla v18.4s, v19.16b, v31.16b\n" - "ldr q19, [x21, #0x20]\n" - "fmla v4.4s, v9.4s, v0.4s\n" - "movi v9.4s, #0x0\n" - "movi v0.4s, #0x0\n" - ".inst 0x4e88a629 // smmla v9.4s, v17.16b, v8.16b\n" - "fmul v8.4s, v27.4s, v26.s[0]\n" - ".inst 0x4e9fa620 // smmla v0.4s, v17.16b, v31.16b\n" - "ldr q17, [x21, #0x30]\n" - ".inst 0x4e8fa674 // smmla v20.4s, v19.16b, v15.16b\n" - "fmul v31.4s, v27.4s, v26.s[1]\n" - ".inst 0x4e81a672 // smmla v18.4s, v19.16b, v1.16b\n" - "ldr q19, [x21, #0x40]\n" - ".inst 0x4e8fa629 // smmla v9.4s, v17.16b, v15.16b\n" - "fmul v15.4s, v27.4s, v26.s[2]\n" - "fmul v27.4s, v27.4s, v26.s[3]\n" - ".inst 0x4e81a620 // smmla v0.4s, v17.16b, v1.16b\n" - "ldr q1, [x21, #0x50]\n" - ".inst 0x4e95a674 // smmla v20.4s, v19.16b, v21.16b\n" - ".inst 0x4e90a672 // smmla v18.4s, v19.16b, v16.16b\n" - "ldr q26, [x21, #0x60]\n" - ".inst 0x4e95a429 // smmla v9.4s, v1.16b, v21.16b\n" - ".inst 0x4e90a420 // smmla v0.4s, v1.16b, v16.16b\n" - "ldr q21, [x21, #0x70]\n" - "add x21, x21, #0x88\n" - ".inst 0x4e9da754 // smmla v20.4s, v26.16b, v29.16b\n" - ".inst 0x4e83a752 // smmla v18.4s, v26.16b, v3.16b\n" - ".inst 0x4e9da6a9 // smmla v9.4s, v21.16b, v29.16b\n" - ".inst 0x4e83a6a0 // smmla v0.4s, v21.16b, v3.16b\n" - "uzp1 v29.2d, v20.2d, v18.2d\n" - "uzp2 v21.2d, v20.2d, v18.2d\n" - "scvtf v29.4s, v29.4s, #0x4\n" - "uzp1 v18.2d, v9.2d, v0.2d\n" - "uzp2 v16.2d, v9.2d, v0.2d\n" - "scvtf v21.4s, v21.4s, #0x4\n" - "fmla v6.4s, v29.4s, v8.4s\n" - "scvtf v18.4s, v18.4s, #0x4\n" - "scvtf v16.4s, v16.4s, #0x4\n" - "fmla v30.4s, v21.4s, v31.4s\n" - "fmla v24.4s, v18.4s, v15.4s\n" - "fmla v14.4s, v16.4s, v27.4s\n" - "bgt 3b\n" - "mov x20, %x[res_ptr]\n" - "subs x27, x27, #0x4\n" - "add %x[res_ptr], %x[res_ptr], #0x10\n" - "str q2, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q10, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q12, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q28, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q11, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q13, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q22, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q23, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q25, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q5, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q7, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q4, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q6, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q30, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q24, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q14, [x20, #0x0]\n" - "bne 2b\n" - "mov x20, #0x4\n" - "sub x10, x10, #0x10\n" - "cmp x10, #0x10\n" - "mov %x[res_ptr], x26\n" - "madd %x[a_ptr], x20, x9, %x[a_ptr]\n" - "bge 1b\n" - "4:" // Row loop skip - "cbz x10, 9f\n" - "5:" // Row tail: Row loop - "add x24, %x[b_ptr], #0x8\n" - "mov x23, %x[nc]\n" - "add x22, %x[res_ptr], %x[res_stride], LSL #2\n" - "6:" // Row tail: Column loop - "movi v2.16b, #0x0\n" - "movi v10.16b, #0x0\n" - "add x25, %x[a_ptr], #0x8\n" - "mov x21, %x[nb]\n" - "movi v12.16b, #0x0\n" - "movi v28.16b, #0x0\n" - "7:" // Row tail: Block loop - "ldr q6, [x24, #0x0]\n" - "ldr q5, [x24, #0x10]\n" - "movi v17.16b, #0x4\n" - "movi v8.4s, #0x0\n" - "ldr q4, [x25, #0x0]\n" - "ldr q13, [x25, #0x10]\n" - "movi v27.4s, #0x0\n" - "movi v0.4s, #0x0\n" - "ldr q31, [x24, #0x20]\n" - "ldr q14, [x24, #0x30]\n" - "movi v29.4s, #0x0\n" - "movi v22.16b, #0xf0\n" - "ldr q11, [x25, #0x20]\n" - "ldr q23, [x25, #0x30]\n" - "sshl v21.16b, v6.16b, v17.16b\n" - "sshl v16.16b, v5.16b, v17.16b\n" - "ldr q20, [x25, #0x40]\n" - "ldr q26, [x25, #0x50]\n" - "and v6.16b, v6.16b, v22.16b\n" - "and v5.16b, v5.16b, v22.16b\n" - "ldr q25, [x25, #0x60]\n" - "ldr q3, [x25, #0x70]\n" - "sshl v19.16b, v31.16b, v17.16b\n" - "sshl v18.16b, v14.16b, v17.16b\n" - "ldr d17, [x25, #-0x8]\n" - ".inst 0x4e95a488 // smmla v8.4s, v4.16b, v21.16b\n" - ".inst 0x4e90a49b // smmla v27.4s, v4.16b, v16.16b\n" - "and v31.16b, v31.16b, v22.16b\n" - ".inst 0x4e95a5a0 // smmla v0.4s, v13.16b, v21.16b\n" - ".inst 0x4e90a5bd // smmla v29.4s, v13.16b, v16.16b\n" - "and v14.16b, v14.16b, v22.16b\n" - "sub x20, x24, #0x8\n" - "ldr d16, [x20, #0x0]\n" - "subs x21, x21, #0x1\n" - "add x25, x25, #0x88\n" - "fcvtl v17.4s, v17.4h\n" - "add x24, x24, #0x48\n" - ".inst 0x4e93a568 // smmla v8.4s, v11.16b, v19.16b\n" - ".inst 0x4e92a57b // smmla v27.4s, v11.16b, v18.16b\n" - ".inst 0x4e93a6e0 // smmla v0.4s, v23.16b, v19.16b\n" - ".inst 0x4e92a6fd // smmla v29.4s, v23.16b, v18.16b\n" - "fcvtl v16.4s, v16.4h\n" - ".inst 0x4e86a688 // smmla v8.4s, v20.16b, v6.16b\n" - ".inst 0x4e85a69b // smmla v27.4s, v20.16b, v5.16b\n" - "fmul v23.4s, v16.4s, v17.s[0]\n" - "fmul v21.4s, v16.4s, v17.s[1]\n" - "fmul v1.4s, v16.4s, v17.s[2]\n" - "fmul v20.4s, v16.4s, v17.s[3]\n" - ".inst 0x4e86a740 // smmla v0.4s, v26.16b, v6.16b\n" - ".inst 0x4e85a75d // smmla v29.4s, v26.16b, v5.16b\n" - ".inst 0x4e9fa728 // smmla v8.4s, v25.16b, v31.16b\n" - ".inst 0x4e8ea73b // smmla v27.4s, v25.16b, v14.16b\n" - ".inst 0x4e9fa460 // smmla v0.4s, v3.16b, v31.16b\n" - ".inst 0x4e8ea47d // smmla v29.4s, v3.16b, v14.16b\n" - "uzp1 v19.2d, v8.2d, v27.2d\n" - "uzp2 v18.2d, v8.2d, v27.2d\n" - "scvtf v19.4s, v19.4s, #0x4\n" - "uzp1 v17.2d, v0.2d, v29.2d\n" - "uzp2 v16.2d, v0.2d, v29.2d\n" - "scvtf v18.4s, v18.4s, #0x4\n" - "fmla v2.4s, v19.4s, v23.4s\n" - "scvtf v17.4s, v17.4s, #0x4\n" - "scvtf v16.4s, v16.4s, #0x4\n" - "fmla v10.4s, v18.4s, v21.4s\n" - "fmla v12.4s, v17.4s, v1.4s\n" - "fmla v28.4s, v16.4s, v20.4s\n" - "bgt 7b\n" - "mov x20, %x[res_ptr]\n" - "cmp x10, #0x1\n" - "str q2, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "ble 8f\n" - "cmp x10, #0x2\n" - "str q10, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "ble 8f\n" - "cmp x10, #0x3\n" - "str q12, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "ble 8f\n" - "str q28, [x20, #0x0]\n" - "8:" // Row tail: Accumulator store skip - "subs x23, x23, #0x4\n" - "add %x[res_ptr], %x[res_ptr], #0x10\n" - "bne 6b\n" - "subs x10, x10, #0x4\n" - "add %x[a_ptr], %x[a_ptr], x9\n" - "mov %x[res_ptr], x22\n" - "bgt 5b\n" - "9:" // Row tail: Row loop skip - : [a_ptr] "+&r" (a_ptr), [res_ptr] "+&r" (res_ptr) - : [b_ptr] "r" (b_ptr), [nr] "r" (nr), [nb] "r" (nb), [res_stride] "r" (res_stride), [nc] "r" (nc) - : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28" - ); - return; - } + __asm__ __volatile__( + "mov x10, %x[nr]\n" + "mov x9, #0x88\n" + "cmp x10, #0x10\n" + "mul x9, %x[nb], x9\n" + "blt 4f\n" + "1:" // Row loop + "add x28, %x[b_ptr], #0x8\n" + "mov x27, %x[nc]\n" + "add x26, %x[res_ptr], %x[res_stride], LSL #4\n" + "2:" // Column loop + "add x25, %x[a_ptr], #0x8\n" + "movi v2.16b, #0x0\n" + "movi v10.16b, #0x0\n" + "mov x24, %x[nb]\n" + "add x23, x25, x9\n" + "movi v12.16b, #0x0\n" + "movi v28.16b, #0x0\n" + "add x22, x23, x9\n" + "movi v11.16b, #0x0\n" + "movi v13.16b, #0x0\n" + "add x21, x22, x9\n" + "movi v22.16b, #0x0\n" + "movi v23.16b, #0x0\n" + "movi v25.16b, #0x0\n" + "movi v5.16b, #0x0\n" + "movi v7.16b, #0x0\n" + "movi v4.16b, #0x0\n" + "movi v6.16b, #0x0\n" + "movi v30.16b, #0x0\n" + "movi v24.16b, #0x0\n" + "movi v14.16b, #0x0\n" + "3:" // Block loop + "ldr q21, [x28, #0x0]\n" + "ldr q16, [x28, #0x10]\n" + "movi v1.16b, #0x4\n" + "movi v19.4s, #0x0\n" + "ldr q27, [x25, #0x0]\n" + "ldr q15, [x25, #0x10]\n" + "movi v26.4s, #0x0\n" + "movi v18.4s, #0x0\n" + "ldr q29, [x28, #0x20]\n" + "ldr q3, [x28, #0x30]\n" + "movi v17.4s, #0x0\n" + "movi v0.16b, #0xf0\n" + "ldr d20, [x25, #-0x8]\n" + "ldr d9, [x23, #-0x8]\n" + "sshl v8.16b, v21.16b, v1.16b\n" + "sshl v31.16b, v16.16b, v1.16b\n" + "and v21.16b, v21.16b, v0.16b\n" + "and v16.16b, v16.16b, v0.16b\n" + "sub x20, x28, #0x8\n" + "subs x24, x24, #0x1\n" + "add x28, x28, #0x48\n" + ".inst 0x4e88a773 // smmla v19.4s, v27.16b, v8.16b\n" + ".inst 0x4e9fa77a // smmla v26.4s, v27.16b, v31.16b\n" + "ldr q27, [x25, #0x20]\n" + ".inst 0x4e88a5f2 // smmla v18.4s, v15.16b, v8.16b\n" + ".inst 0x4e9fa5f1 // smmla v17.4s, v15.16b, v31.16b\n" + "sshl v15.16b, v29.16b, v1.16b\n" + "sshl v1.16b, v3.16b, v1.16b\n" + "and v29.16b, v29.16b, v0.16b\n" + "and v3.16b, v3.16b, v0.16b\n" + "ldr q0, [x25, #0x30]\n" + "fcvtl v20.4s, v20.4h\n" + ".inst 0x4e8fa773 // smmla v19.4s, v27.16b, v15.16b\n" + "fcvtl v9.4s, v9.4h\n" + ".inst 0x4e81a77a // smmla v26.4s, v27.16b, v1.16b\n" + "ldr q27, [x25, #0x40]\n" + ".inst 0x4e8fa412 // smmla v18.4s, v0.16b, v15.16b\n" + ".inst 0x4e81a411 // smmla v17.4s, v0.16b, v1.16b\n" + "ldr q0, [x25, #0x50]\n" + ".inst 0x4e95a773 // smmla v19.4s, v27.16b, v21.16b\n" + ".inst 0x4e90a77a // smmla v26.4s, v27.16b, v16.16b\n" + "ldr q27, [x25, #0x60]\n" + ".inst 0x4e95a412 // smmla v18.4s, v0.16b, v21.16b\n" + ".inst 0x4e90a411 // smmla v17.4s, v0.16b, v16.16b\n" + "ldr q0, [x25, #0x70]\n" + "add x25, x25, #0x88\n" + ".inst 0x4e9da773 // smmla v19.4s, v27.16b, v29.16b\n" + ".inst 0x4e83a77a // smmla v26.4s, v27.16b, v3.16b\n" + "ldr d27, [x20, #0x0]\n" + ".inst 0x4e9da412 // smmla v18.4s, v0.16b, v29.16b\n" + ".inst 0x4e83a411 // smmla v17.4s, v0.16b, v3.16b\n" + "fcvtl v27.4s, v27.4h\n" + "uzp1 v0.2d, v19.2d, v26.2d\n" + "uzp2 v26.2d, v19.2d, v26.2d\n" + "fmul v19.4s, v27.4s, v20.s[0]\n" + "scvtf v0.4s, v0.4s, #0x4\n" + "scvtf v26.4s, v26.4s, #0x4\n" + "fmla v2.4s, v0.4s, v19.4s\n" + "ldr q19, [x23, #0x0]\n" + "uzp1 v0.2d, v18.2d, v17.2d\n" + "uzp2 v18.2d, v18.2d, v17.2d\n" + "fmul v17.4s, v27.4s, v20.s[1]\n" + "scvtf v0.4s, v0.4s, #0x4\n" + "scvtf v18.4s, v18.4s, #0x4\n" + "fmla v10.4s, v26.4s, v17.4s\n" + "ldr q17, [x23, #0x10]\n" + "fmul v26.4s, v27.4s, v20.s[2]\n" + "fmul v20.4s, v27.4s, v20.s[3]\n" + "fmla v12.4s, v0.4s, v26.4s\n" + "ldr d0, [x22, #-0x8]\n" + "ldr d26, [x21, #-0x8]\n" + "fcvtl v0.4s, v0.4h\n" + "fmla v28.4s, v18.4s, v20.4s\n" + "movi v20.4s, #0x0\n" + "movi v18.4s, #0x0\n" + ".inst 0x4e88a674 // smmla v20.4s, v19.16b, v8.16b\n" + ".inst 0x4e9fa672 // smmla v18.4s, v19.16b, v31.16b\n" + "ldr q19, [x23, #0x20]\n" + "fcvtl v26.4s, v26.4h\n" + ".inst 0x4e8fa674 // smmla v20.4s, v19.16b, v15.16b\n" + ".inst 0x4e81a672 // smmla v18.4s, v19.16b, v1.16b\n" + "ldr q19, [x23, #0x40]\n" + ".inst 0x4e95a674 // smmla v20.4s, v19.16b, v21.16b\n" + ".inst 0x4e90a672 // smmla v18.4s, v19.16b, v16.16b\n" + "ldr q19, [x23, #0x60]\n" + ".inst 0x4e9da674 // smmla v20.4s, v19.16b, v29.16b\n" + ".inst 0x4e83a672 // smmla v18.4s, v19.16b, v3.16b\n" + "uzp1 v19.2d, v20.2d, v18.2d\n" + "scvtf v19.4s, v19.4s, #0x4\n" + "uzp2 v20.2d, v20.2d, v18.2d\n" + "fmul v18.4s, v27.4s, v9.s[0]\n" + "scvtf v20.4s, v20.4s, #0x4\n" + "fmla v11.4s, v19.4s, v18.4s\n" + "ldr q18, [x22, #0x0]\n" + "fmul v19.4s, v27.4s, v9.s[1]\n" + "fmla v13.4s, v20.4s, v19.4s\n" + "movi v19.4s, #0x0\n" + "movi v20.4s, #0x0\n" + ".inst 0x4e88a633 // smmla v19.4s, v17.16b, v8.16b\n" + ".inst 0x4e9fa634 // smmla v20.4s, v17.16b, v31.16b\n" + "ldr q17, [x23, #0x30]\n" + ".inst 0x4e8fa633 // smmla v19.4s, v17.16b, v15.16b\n" + ".inst 0x4e81a634 // smmla v20.4s, v17.16b, v1.16b\n" + "ldr q17, [x23, #0x50]\n" + ".inst 0x4e95a633 // smmla v19.4s, v17.16b, v21.16b\n" + ".inst 0x4e90a634 // smmla v20.4s, v17.16b, v16.16b\n" + "ldr q17, [x23, #0x70]\n" + "add x23, x23, #0x88\n" + ".inst 0x4e9da633 // smmla v19.4s, v17.16b, v29.16b\n" + ".inst 0x4e83a634 // smmla v20.4s, v17.16b, v3.16b\n" + "uzp1 v17.2d, v19.2d, v20.2d\n" + "scvtf v17.4s, v17.4s, #0x4\n" + "uzp2 v20.2d, v19.2d, v20.2d\n" + "fmul v19.4s, v27.4s, v9.s[2]\n" + "fmul v9.4s, v27.4s, v9.s[3]\n" + "scvtf v20.4s, v20.4s, #0x4\n" + "fmla v22.4s, v17.4s, v19.4s\n" + "ldr q17, [x22, #0x10]\n" + "movi v19.4s, #0x0\n" + ".inst 0x4e88a653 // smmla v19.4s, v18.16b, v8.16b\n" + "fmla v23.4s, v20.4s, v9.4s\n" + "movi v20.4s, #0x0\n" + "movi v9.4s, #0x0\n" + ".inst 0x4e9fa654 // smmla v20.4s, v18.16b, v31.16b\n" + "ldr q18, [x22, #0x20]\n" + ".inst 0x4e88a629 // smmla v9.4s, v17.16b, v8.16b\n" + ".inst 0x4e8fa653 // smmla v19.4s, v18.16b, v15.16b\n" + ".inst 0x4e81a654 // smmla v20.4s, v18.16b, v1.16b\n" + "ldr q18, [x22, #0x40]\n" + ".inst 0x4e95a653 // smmla v19.4s, v18.16b, v21.16b\n" + ".inst 0x4e90a654 // smmla v20.4s, v18.16b, v16.16b\n" + "ldr q18, [x22, #0x60]\n" + ".inst 0x4e9da653 // smmla v19.4s, v18.16b, v29.16b\n" + ".inst 0x4e83a654 // smmla v20.4s, v18.16b, v3.16b\n" + "movi v18.4s, #0x0\n" + ".inst 0x4e9fa632 // smmla v18.4s, v17.16b, v31.16b\n" + "ldr q17, [x22, #0x30]\n" + ".inst 0x4e8fa629 // smmla v9.4s, v17.16b, v15.16b\n" + ".inst 0x4e81a632 // smmla v18.4s, v17.16b, v1.16b\n" + "ldr q17, [x22, #0x50]\n" + ".inst 0x4e95a629 // smmla v9.4s, v17.16b, v21.16b\n" + ".inst 0x4e90a632 // smmla v18.4s, v17.16b, v16.16b\n" + "ldr q17, [x22, #0x70]\n" + "add x22, x22, #0x88\n" + ".inst 0x4e9da629 // smmla v9.4s, v17.16b, v29.16b\n" + ".inst 0x4e83a632 // smmla v18.4s, v17.16b, v3.16b\n" + "uzp1 v17.2d, v19.2d, v20.2d\n" + "uzp2 v20.2d, v19.2d, v20.2d\n" + "fmul v19.4s, v27.4s, v0.s[0]\n" + "scvtf v17.4s, v17.4s, #0x4\n" + "scvtf v20.4s, v20.4s, #0x4\n" + "fmla v25.4s, v17.4s, v19.4s\n" + "ldr q19, [x21, #0x0]\n" + "fmul v17.4s, v27.4s, v0.s[1]\n" + "fmla v5.4s, v20.4s, v17.4s\n" + "ldr q17, [x21, #0x10]\n" + "uzp1 v20.2d, v9.2d, v18.2d\n" + "uzp2 v9.2d, v9.2d, v18.2d\n" + "fmul v18.4s, v27.4s, v0.s[2]\n" + "fmul v0.4s, v27.4s, v0.s[3]\n" + "scvtf v20.4s, v20.4s, #0x4\n" + "scvtf v9.4s, v9.4s, #0x4\n" + "fmla v7.4s, v20.4s, v18.4s\n" + "movi v20.4s, #0x0\n" + "movi v18.4s, #0x0\n" + ".inst 0x4e88a674 // smmla v20.4s, v19.16b, v8.16b\n" + ".inst 0x4e9fa672 // smmla v18.4s, v19.16b, v31.16b\n" + "ldr q19, [x21, #0x20]\n" + "fmla v4.4s, v9.4s, v0.4s\n" + "movi v9.4s, #0x0\n" + "movi v0.4s, #0x0\n" + ".inst 0x4e88a629 // smmla v9.4s, v17.16b, v8.16b\n" + "fmul v8.4s, v27.4s, v26.s[0]\n" + ".inst 0x4e9fa620 // smmla v0.4s, v17.16b, v31.16b\n" + "ldr q17, [x21, #0x30]\n" + ".inst 0x4e8fa674 // smmla v20.4s, v19.16b, v15.16b\n" + "fmul v31.4s, v27.4s, v26.s[1]\n" + ".inst 0x4e81a672 // smmla v18.4s, v19.16b, v1.16b\n" + "ldr q19, [x21, #0x40]\n" + ".inst 0x4e8fa629 // smmla v9.4s, v17.16b, v15.16b\n" + "fmul v15.4s, v27.4s, v26.s[2]\n" + "fmul v27.4s, v27.4s, v26.s[3]\n" + ".inst 0x4e81a620 // smmla v0.4s, v17.16b, v1.16b\n" + "ldr q1, [x21, #0x50]\n" + ".inst 0x4e95a674 // smmla v20.4s, v19.16b, v21.16b\n" + ".inst 0x4e90a672 // smmla v18.4s, v19.16b, v16.16b\n" + "ldr q26, [x21, #0x60]\n" + ".inst 0x4e95a429 // smmla v9.4s, v1.16b, v21.16b\n" + ".inst 0x4e90a420 // smmla v0.4s, v1.16b, v16.16b\n" + "ldr q21, [x21, #0x70]\n" + "add x21, x21, #0x88\n" + ".inst 0x4e9da754 // smmla v20.4s, v26.16b, v29.16b\n" + ".inst 0x4e83a752 // smmla v18.4s, v26.16b, v3.16b\n" + ".inst 0x4e9da6a9 // smmla v9.4s, v21.16b, v29.16b\n" + ".inst 0x4e83a6a0 // smmla v0.4s, v21.16b, v3.16b\n" + "uzp1 v29.2d, v20.2d, v18.2d\n" + "uzp2 v21.2d, v20.2d, v18.2d\n" + "scvtf v29.4s, v29.4s, #0x4\n" + "uzp1 v18.2d, v9.2d, v0.2d\n" + "uzp2 v16.2d, v9.2d, v0.2d\n" + "scvtf v21.4s, v21.4s, #0x4\n" + "fmla v6.4s, v29.4s, v8.4s\n" + "scvtf v18.4s, v18.4s, #0x4\n" + "scvtf v16.4s, v16.4s, #0x4\n" + "fmla v30.4s, v21.4s, v31.4s\n" + "fmla v24.4s, v18.4s, v15.4s\n" + "fmla v14.4s, v16.4s, v27.4s\n" + "bgt 3b\n" + "mov x20, %x[res_ptr]\n" + "subs x27, x27, #0x4\n" + "add %x[res_ptr], %x[res_ptr], #0x10\n" + "str q2, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q10, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q12, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q28, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q11, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q13, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q22, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q23, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q25, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q5, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q7, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q4, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q6, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q30, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q24, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q14, [x20, #0x0]\n" + "bne 2b\n" + "mov x20, #0x4\n" + "sub x10, x10, #0x10\n" + "cmp x10, #0x10\n" + "mov %x[res_ptr], x26\n" + "madd %x[a_ptr], x20, x9, %x[a_ptr]\n" + "bge 1b\n" + "4:" // Row loop skip + "cbz x10, 9f\n" + "5:" // Row tail: Row loop + "add x24, %x[b_ptr], #0x8\n" + "mov x23, %x[nc]\n" + "add x22, %x[res_ptr], %x[res_stride], LSL #2\n" + "6:" // Row tail: Column loop + "movi v2.16b, #0x0\n" + "movi v10.16b, #0x0\n" + "add x25, %x[a_ptr], #0x8\n" + "mov x21, %x[nb]\n" + "movi v12.16b, #0x0\n" + "movi v28.16b, #0x0\n" + "7:" // Row tail: Block loop + "ldr q6, [x24, #0x0]\n" + "ldr q5, [x24, #0x10]\n" + "movi v17.16b, #0x4\n" + "movi v8.4s, #0x0\n" + "ldr q4, [x25, #0x0]\n" + "ldr q13, [x25, #0x10]\n" + "movi v27.4s, #0x0\n" + "movi v0.4s, #0x0\n" + "ldr q31, [x24, #0x20]\n" + "ldr q14, [x24, #0x30]\n" + "movi v29.4s, #0x0\n" + "movi v22.16b, #0xf0\n" + "ldr q11, [x25, #0x20]\n" + "ldr q23, [x25, #0x30]\n" + "sshl v21.16b, v6.16b, v17.16b\n" + "sshl v16.16b, v5.16b, v17.16b\n" + "ldr q20, [x25, #0x40]\n" + "ldr q26, [x25, #0x50]\n" + "and v6.16b, v6.16b, v22.16b\n" + "and v5.16b, v5.16b, v22.16b\n" + "ldr q25, [x25, #0x60]\n" + "ldr q3, [x25, #0x70]\n" + "sshl v19.16b, v31.16b, v17.16b\n" + "sshl v18.16b, v14.16b, v17.16b\n" + "ldr d17, [x25, #-0x8]\n" + ".inst 0x4e95a488 // smmla v8.4s, v4.16b, v21.16b\n" + ".inst 0x4e90a49b // smmla v27.4s, v4.16b, v16.16b\n" + "and v31.16b, v31.16b, v22.16b\n" + ".inst 0x4e95a5a0 // smmla v0.4s, v13.16b, v21.16b\n" + ".inst 0x4e90a5bd // smmla v29.4s, v13.16b, v16.16b\n" + "and v14.16b, v14.16b, v22.16b\n" + "sub x20, x24, #0x8\n" + "ldr d16, [x20, #0x0]\n" + "subs x21, x21, #0x1\n" + "add x25, x25, #0x88\n" + "fcvtl v17.4s, v17.4h\n" + "add x24, x24, #0x48\n" + ".inst 0x4e93a568 // smmla v8.4s, v11.16b, v19.16b\n" + ".inst 0x4e92a57b // smmla v27.4s, v11.16b, v18.16b\n" + ".inst 0x4e93a6e0 // smmla v0.4s, v23.16b, v19.16b\n" + ".inst 0x4e92a6fd // smmla v29.4s, v23.16b, v18.16b\n" + "fcvtl v16.4s, v16.4h\n" + ".inst 0x4e86a688 // smmla v8.4s, v20.16b, v6.16b\n" + ".inst 0x4e85a69b // smmla v27.4s, v20.16b, v5.16b\n" + "fmul v23.4s, v16.4s, v17.s[0]\n" + "fmul v21.4s, v16.4s, v17.s[1]\n" + "fmul v1.4s, v16.4s, v17.s[2]\n" + "fmul v20.4s, v16.4s, v17.s[3]\n" + ".inst 0x4e86a740 // smmla v0.4s, v26.16b, v6.16b\n" + ".inst 0x4e85a75d // smmla v29.4s, v26.16b, v5.16b\n" + ".inst 0x4e9fa728 // smmla v8.4s, v25.16b, v31.16b\n" + ".inst 0x4e8ea73b // smmla v27.4s, v25.16b, v14.16b\n" + ".inst 0x4e9fa460 // smmla v0.4s, v3.16b, v31.16b\n" + ".inst 0x4e8ea47d // smmla v29.4s, v3.16b, v14.16b\n" + "uzp1 v19.2d, v8.2d, v27.2d\n" + "uzp2 v18.2d, v8.2d, v27.2d\n" + "scvtf v19.4s, v19.4s, #0x4\n" + "uzp1 v17.2d, v0.2d, v29.2d\n" + "uzp2 v16.2d, v0.2d, v29.2d\n" + "scvtf v18.4s, v18.4s, #0x4\n" + "fmla v2.4s, v19.4s, v23.4s\n" + "scvtf v17.4s, v17.4s, #0x4\n" + "scvtf v16.4s, v16.4s, #0x4\n" + "fmla v10.4s, v18.4s, v21.4s\n" + "fmla v12.4s, v17.4s, v1.4s\n" + "fmla v28.4s, v16.4s, v20.4s\n" + "bgt 7b\n" + "mov x20, %x[res_ptr]\n" + "cmp x10, #0x1\n" + "str q2, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "ble 8f\n" + "cmp x10, #0x2\n" + "str q10, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "ble 8f\n" + "cmp x10, #0x3\n" + "str q12, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "ble 8f\n" + "str q28, [x20, #0x0]\n" + "8:" // Row tail: Accumulator store skip + "subs x23, x23, #0x4\n" + "add %x[res_ptr], %x[res_ptr], #0x10\n" + "bne 6b\n" + "subs x10, x10, #0x4\n" + "add %x[a_ptr], %x[a_ptr], x9\n" + "mov %x[res_ptr], x22\n" + "bgt 5b\n" + "9:" // Row tail: Row loop skip + : [a_ptr] "+&r" (a_ptr), [res_ptr] "+&r" (res_ptr) + : [b_ptr] "r" (b_ptr), [nr] "r" (nr), [nb] "r" (nb), [res_stride] "r" (res_stride), [nc] "r" (nc) + : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28" + ); + return; #endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8) float sumf[4][4]; int sumi; @@ -1615,7 +1605,7 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) #if defined(__ARM_FEATURE_SVE) && defined(__ARM_FEATURE_MATMUL_INT8) - if (ggml_cpu_has_sve() && ggml_cpu_has_matmul_int8() && ggml_cpu_get_sve_cnt() == QK8_0) { + if (ggml_cpu_get_sve_cnt() == QK8_0) { const void * b_ptr = vx; const void * a_ptr = vy; float * res_ptr = s; @@ -2083,59 +2073,57 @@ void ggml_gemm_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const UNUSED(blocklen); #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD) - if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) { - const int8x16_t kvalues = vld1q_s8(kvalues_iq4nl); + const int8x16_t kvalues = vld1q_s8(kvalues_iq4nl); - for (int y = 0; y < nr / 4; y++) { - const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb); - for (int x = 0; x < nc / ncols_interleaved; x++) { - const block_iq4_nlx4 * b_ptr = (const block_iq4_nlx4 *) vx + (x * nb); + for (int y = 0; y < nr / 4; y++) { + const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb); + for (int x = 0; x < nc / ncols_interleaved; x++) { + const block_iq4_nlx4 * b_ptr = (const block_iq4_nlx4 *) vx + (x * nb); - float32x4_t sumf[4]; - for (int m = 0; m < 4; m++) { - sumf[m] = vdupq_n_f32(0); + float32x4_t sumf[4]; + for (int m = 0; m < 4; m++) { + sumf[m] = vdupq_n_f32(0); + } + + for (int l = 0; l < nb; l++) { + float32x4_t a_d = vcvt_f32_f16(vld1_f16((const float16_t *)a_ptr[l].d)); + float32x4_t b_d = vcvt_f32_f16(vld1_f16((const float16_t *)b_ptr[l].d)); + + int32x4_t sumi_0 = vdupq_n_s32(0); + int32x4_t sumi_1 = vdupq_n_s32(0); + int32x4_t sumi_2 = vdupq_n_s32(0); + int32x4_t sumi_3 = vdupq_n_s32(0); + + for (int k = 0; k < 4; k++) { + int8x16_t a_0 = vld1q_s8(a_ptr[l].qs + 16 * k + 0); + int8x16_t a_1 = vld1q_s8(a_ptr[l].qs + 16 * k + 64); + + uint8x16_t b = vld1q_u8(b_ptr[l].qs + 16 * k); + int8x16_t b_hi = vqtbl1q_s8(kvalues, b >> 4); + int8x16_t b_lo = vqtbl1q_s8(kvalues, b & 0xF); + + sumi_0 = vdotq_laneq_s32(sumi_0, b_lo, a_0, 0); + sumi_1 = vdotq_laneq_s32(sumi_1, b_lo, a_0, 1); + sumi_2 = vdotq_laneq_s32(sumi_2, b_lo, a_0, 2); + sumi_3 = vdotq_laneq_s32(sumi_3, b_lo, a_0, 3); + sumi_0 = vdotq_laneq_s32(sumi_0, b_hi, a_1, 0); + sumi_1 = vdotq_laneq_s32(sumi_1, b_hi, a_1, 1); + sumi_2 = vdotq_laneq_s32(sumi_2, b_hi, a_1, 2); + sumi_3 = vdotq_laneq_s32(sumi_3, b_hi, a_1, 3); } - for (int l = 0; l < nb; l++) { - float32x4_t a_d = vcvt_f32_f16(vld1_f16((const float16_t *)a_ptr[l].d)); - float32x4_t b_d = vcvt_f32_f16(vld1_f16((const float16_t *)b_ptr[l].d)); + sumf[0] = vmlaq_f32(sumf[0], vmulq_laneq_f32(b_d, a_d, 0), vcvtq_f32_s32(sumi_0)); + sumf[1] = vmlaq_f32(sumf[1], vmulq_laneq_f32(b_d, a_d, 1), vcvtq_f32_s32(sumi_1)); + sumf[2] = vmlaq_f32(sumf[2], vmulq_laneq_f32(b_d, a_d, 2), vcvtq_f32_s32(sumi_2)); + sumf[3] = vmlaq_f32(sumf[3], vmulq_laneq_f32(b_d, a_d, 3), vcvtq_f32_s32(sumi_3)); + } - int32x4_t sumi_0 = vdupq_n_s32(0); - int32x4_t sumi_1 = vdupq_n_s32(0); - int32x4_t sumi_2 = vdupq_n_s32(0); - int32x4_t sumi_3 = vdupq_n_s32(0); - - for (int k = 0; k < 4; k++) { - int8x16_t a_0 = vld1q_s8(a_ptr[l].qs + 16 * k + 0); - int8x16_t a_1 = vld1q_s8(a_ptr[l].qs + 16 * k + 64); - - uint8x16_t b = vld1q_u8(b_ptr[l].qs + 16 * k); - int8x16_t b_hi = vqtbl1q_s8(kvalues, b >> 4); - int8x16_t b_lo = vqtbl1q_s8(kvalues, b & 0xF); - - sumi_0 = vdotq_laneq_s32(sumi_0, b_lo, a_0, 0); - sumi_1 = vdotq_laneq_s32(sumi_1, b_lo, a_0, 1); - sumi_2 = vdotq_laneq_s32(sumi_2, b_lo, a_0, 2); - sumi_3 = vdotq_laneq_s32(sumi_3, b_lo, a_0, 3); - sumi_0 = vdotq_laneq_s32(sumi_0, b_hi, a_1, 0); - sumi_1 = vdotq_laneq_s32(sumi_1, b_hi, a_1, 1); - sumi_2 = vdotq_laneq_s32(sumi_2, b_hi, a_1, 2); - sumi_3 = vdotq_laneq_s32(sumi_3, b_hi, a_1, 3); - } - - sumf[0] = vmlaq_f32(sumf[0], vmulq_laneq_f32(b_d, a_d, 0), vcvtq_f32_s32(sumi_0)); - sumf[1] = vmlaq_f32(sumf[1], vmulq_laneq_f32(b_d, a_d, 1), vcvtq_f32_s32(sumi_1)); - sumf[2] = vmlaq_f32(sumf[2], vmulq_laneq_f32(b_d, a_d, 2), vcvtq_f32_s32(sumi_2)); - sumf[3] = vmlaq_f32(sumf[3], vmulq_laneq_f32(b_d, a_d, 3), vcvtq_f32_s32(sumi_3)); - } - - for (int m = 0; m < 4; m++) { - vst1q_f32(s + (y * 4 + m) * bs + x * 4, sumf[m]); - } + for (int m = 0; m < 4; m++) { + vst1q_f32(s + (y * 4 + m) * bs + x * 4, sumf[m]); } } - return; } + return; #endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) { float sumf[4][4]; diff --git a/ggml/src/ggml-cpu/ggml-cpu.c b/ggml/src/ggml-cpu/ggml-cpu.c index 2c12e493b..1bb9c4e36 100644 --- a/ggml/src/ggml-cpu/ggml-cpu.c +++ b/ggml/src/ggml-cpu/ggml-cpu.c @@ -74,13 +74,8 @@ #if defined(__ARM_ARCH) struct ggml_arm_arch_features_type { - int has_neon; - int has_dotprod; - int has_i8mm; - int has_sve; int sve_cnt; - int has_sme; -} ggml_arm_arch_features = {-1, -1, -1, -1, 0, -1}; +} ggml_arm_arch_features = { 0 }; #endif @@ -678,87 +673,15 @@ bool ggml_is_numa(void) { #if defined(__linux__) && defined(__aarch64__) #include -#elif defined(__APPLE__) -#include -#endif - -#if !defined(HWCAP2_I8MM) -#define HWCAP2_I8MM (1 << 13) -#endif - -#if !defined(HWCAP2_SME) -#define HWCAP2_SME (1 << 23) #endif static void ggml_init_arm_arch_features(void) { -#if defined(__linux__) && defined(__aarch64__) - uint32_t hwcap = getauxval(AT_HWCAP); - uint32_t hwcap2 = getauxval(AT_HWCAP2); - - ggml_arm_arch_features.has_neon = !!(hwcap & HWCAP_ASIMD); - ggml_arm_arch_features.has_dotprod = !!(hwcap & HWCAP_ASIMDDP); - ggml_arm_arch_features.has_i8mm = !!(hwcap2 & HWCAP2_I8MM); - ggml_arm_arch_features.has_sve = !!(hwcap & HWCAP_SVE); - ggml_arm_arch_features.has_sme = !!(hwcap2 & HWCAP2_SME); - -#if defined(__ARM_FEATURE_SVE) +#if defined(__linux__) && defined(__aarch64__) && defined(__ARM_FEATURE_SVE) ggml_arm_arch_features.sve_cnt = PR_SVE_VL_LEN_MASK & prctl(PR_SVE_GET_VL); #endif -#elif defined(__APPLE__) - int oldp = 0; - size_t size = sizeof(oldp); - if (sysctlbyname("hw.optional.AdvSIMD", &oldp, &size, NULL, 0) != 0) { - oldp = 0; - } - ggml_arm_arch_features.has_neon = oldp; - - if (sysctlbyname("hw.optional.arm.FEAT_DotProd", &oldp, &size, NULL, 0) != 0) { - oldp = 0; - } - ggml_arm_arch_features.has_dotprod = oldp; - - if (sysctlbyname("hw.optional.arm.FEAT_I8MM", &oldp, &size, NULL, 0) != 0) { - oldp = 0; - } - ggml_arm_arch_features.has_i8mm = oldp; - - if (sysctlbyname("hw.optional.arm.FEAT_SME", &oldp, &size, NULL, 0) != 0) { - oldp = 0; - } - ggml_arm_arch_features.has_sme = oldp; - - ggml_arm_arch_features.has_sve = 0; - ggml_arm_arch_features.sve_cnt = 0; -#else -// Run-time CPU feature detection not implemented for this platform, fallback to compile time -#if defined(__ARM_NEON) - ggml_arm_arch_features.has_neon = 1; -#else - ggml_arm_arch_features.has_neon = 0; -#endif - -#if defined(__ARM_FEATURE_MATMUL_INT8) - ggml_arm_arch_features.has_i8mm = 1; -#else - ggml_arm_arch_features.has_i8mm = 0; -#endif - -#if defined(__ARM_FEATURE_SVE) - ggml_arm_arch_features.has_sve = 1; - ggml_arm_arch_features.sve_cnt = 16; -#else - ggml_arm_arch_features.has_sve = 0; - ggml_arm_arch_features.sve_cnt = 0; -#endif - -#if defined(__ARM_FEATURE_SME) || defined(__ARM_FEATURE_SME2) - ggml_arm_arch_features.has_sme = 1; -#else - ggml_arm_arch_features.has_sme = 0; -#endif -#endif } -#endif + +#endif // __ARM_ARCH struct ggml_tensor * ggml_new_i32(struct ggml_context * ctx, int32_t value) { GGML_ASSERT(!ggml_get_no_alloc(ctx)); @@ -3443,7 +3366,7 @@ int ggml_cpu_has_vxe(void) { int ggml_cpu_has_neon(void) { #if defined(__ARM_ARCH) && defined(__ARM_NEON) - return ggml_arm_arch_features.has_neon; + return 1; #else return 0; #endif @@ -3451,7 +3374,7 @@ int ggml_cpu_has_neon(void) { int ggml_cpu_has_dotprod(void) { #if defined(__ARM_ARCH) && defined(__ARM_FEATURE_DOTPROD) - return ggml_arm_arch_features.has_dotprod; + return 1; #else return 0; #endif @@ -3459,7 +3382,7 @@ int ggml_cpu_has_dotprod(void) { int ggml_cpu_has_sve(void) { #if defined(__ARM_ARCH) && defined(__ARM_FEATURE_SVE) - return ggml_arm_arch_features.has_sve; + return 1; #else return 0; #endif @@ -3467,7 +3390,7 @@ int ggml_cpu_has_sve(void) { int ggml_cpu_has_matmul_int8(void) { #if defined(__ARM_ARCH) && defined(__ARM_FEATURE_MATMUL_INT8) - return ggml_arm_arch_features.has_i8mm; + return 1; #else return 0; #endif @@ -3483,7 +3406,7 @@ int ggml_cpu_get_sve_cnt(void) { int ggml_cpu_has_sme(void) { #if defined(__ARM_ARCH) && defined(__ARM_FEATURE_SME) - return ggml_arm_arch_features.has_sme; + return 1; #else return 0; #endif From 9eaa51e7f08593f123f00136591179a8f5956ecd Mon Sep 17 00:00:00 2001 From: Aman Gupta Date: Fri, 20 Jun 2025 09:50:24 +0800 Subject: [PATCH 10/43] CUDA: add conv_2d_dw (#14265) * CUDA: add conv_2d_dw * better naming * simplify using template * Review: fix operation ordering in ggml-cuda, use __forceinline__, use more const --- ggml/src/ggml-cuda/conv2d-dw.cu | 161 +++++++++++++++++++++++++++++++ ggml/src/ggml-cuda/conv2d-dw.cuh | 5 + ggml/src/ggml-cuda/ggml-cuda.cu | 5 + 3 files changed, 171 insertions(+) create mode 100644 ggml/src/ggml-cuda/conv2d-dw.cu create mode 100644 ggml/src/ggml-cuda/conv2d-dw.cuh diff --git a/ggml/src/ggml-cuda/conv2d-dw.cu b/ggml/src/ggml-cuda/conv2d-dw.cu new file mode 100644 index 000000000..7583233b1 --- /dev/null +++ b/ggml/src/ggml-cuda/conv2d-dw.cu @@ -0,0 +1,161 @@ +#include "conv2d-dw.cuh" + +struct conv_params { + int in_w, in_h; + int out_w, out_h; + int kernel_w, kernel_h; + int stride_x, stride_y; + int padding_x, padding_y; + int dilation_x, dilation_y; + int channels, batches; +}; + +struct kernel_bounds { + int y_min, y_max; + int x_min, x_max; +}; + +__device__ __forceinline__ kernel_bounds calculate_kernel_bounds(int out_x, int out_y, const conv_params & params) { + kernel_bounds bounds; + bounds.y_min = max(0, (params.padding_y - out_y * params.stride_y + params.dilation_y - 1) / params.dilation_y); + bounds.y_max = + min(params.kernel_h, + (params.in_h + params.padding_y - out_y * params.stride_y + params.dilation_y - 1) / params.dilation_y); + bounds.x_min = max(0, (params.padding_x - out_x * params.stride_x + params.dilation_x - 1) / params.dilation_x); + bounds.x_max = + min(params.kernel_w, + (params.in_w + params.padding_x - out_x * params.stride_x + params.dilation_x - 1) / params.dilation_x); + return bounds; +} + +__device__ __forceinline__ int calculate_input_coord(int out_coord, int kern_coord, int stride, int dilation, int padding) { + return out_coord * stride + kern_coord * dilation - padding; +} + +struct whcn_layout { + __device__ static int input_index(int n, int c, int y, int x, const conv_params & params) { + return n * (params.channels * params.in_w * params.in_h) + c * params.in_w * params.in_h + y * params.in_w + x; + } + + __device__ static int kernel_index(int c, int ky, int kx, const conv_params & params) { + return c * params.kernel_h * params.kernel_w + ky * params.kernel_w + kx; + } + + __device__ static int output_index(int n, int c, int y, int x, const conv_params & params) { + return n * (params.channels * params.out_w * params.out_h) + c * params.out_w * params.out_h + + y * params.out_w + x; + } + + __device__ static void unpack_indices(int global_idx, const conv_params & params, int & n, int & c, int & out_y, + int & out_x) { + out_x = global_idx % params.out_w; + out_y = (global_idx / params.out_w) % params.out_h; + c = (global_idx / (params.out_w * params.out_h)) % params.channels; + n = global_idx / (params.out_w * params.out_h * params.channels); + } +}; + +struct cwhn_layout { + __device__ static int input_index(int n, int c, int y, int x, const conv_params & params) { + return n * (params.channels * params.in_w * params.in_h) + (y * params.in_w + x) * params.channels + c; + } + + __device__ static int kernel_index(int c, int ky, int kx, const conv_params & params) { + return (ky * params.kernel_w + kx) * params.channels + c; + } + + __device__ static int output_index(int n, int c, int y, int x, const conv_params & params) { + return n * (params.channels * params.out_w * params.out_h) + y * (params.out_w * params.channels) + + x * params.channels + c; + } + + __device__ static void unpack_indices(int global_idx, const conv_params & params, int & n, int & c, int & out_y, + int & out_x) { + c = global_idx % params.channels; + out_x = (global_idx / params.channels) % params.out_w; + out_y = (global_idx / (params.channels * params.out_w)) % params.out_h; + n = global_idx / (params.channels * params.out_w * params.out_h); + } +}; + +template +__global__ void conv2d_dw_kernel(const T * __restrict__ input, const T * __restrict__ kernel, T * __restrict__ output, + const int in_w, const int in_h, const int out_w, const int out_h, + const int kernel_w, const int kernel_h, const int stride_x, const int stride_y, + const int padding_x, const int padding_y, const int dilation_x, const int dilation_y, + const int channels, const int batches) { + const int global_idx = blockIdx.x * blockDim.x + threadIdx.x; + const int total_elements = batches * channels * out_h * out_w; + + if (global_idx >= total_elements) { + return; + } + + conv_params params = { in_w, in_h, out_w, out_h, kernel_w, kernel_h, stride_x, + stride_y, padding_x, padding_y, dilation_x, dilation_y, channels, batches }; + + int batch_idx, channel_idx, out_y_idx, out_x_idx; + Layout::unpack_indices(global_idx, params, batch_idx, channel_idx, out_y_idx, out_x_idx); + + T accumulator = 0; + kernel_bounds bounds = calculate_kernel_bounds(out_x_idx, out_y_idx, params); + + for (int kern_y = bounds.y_min; kern_y < bounds.y_max; ++kern_y) { + int in_y_idx = calculate_input_coord(out_y_idx, kern_y, params.stride_y, params.dilation_y, params.padding_y); + + for (int kern_x = bounds.x_min; kern_x < bounds.x_max; ++kern_x) { + int in_x_idx = calculate_input_coord(out_x_idx, kern_x, params.stride_x, params.dilation_x, params.padding_x); + + const T input_val = input[Layout::input_index(batch_idx, channel_idx, in_y_idx, in_x_idx, params)]; + const T kernel_val = kernel[Layout::kernel_index(channel_idx, kern_y, kern_x, params)]; + + accumulator += input_val * kernel_val; + } + } + + output[Layout::output_index(batch_idx, channel_idx, out_y_idx, out_x_idx, params)] = accumulator; +} + +void ggml_cuda_op_conv2d_dw(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { + const ggml_tensor * kernel = dst->src[0]; + const ggml_tensor * input = dst->src[1]; + + GGML_ASSERT(kernel->type == GGML_TYPE_F32 && input->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32); + const float * w_d = (const float *) kernel->data; + const float * x_d = (const float *) input->data; + float * y_d = (float *) dst->data; + + const int32_t * p = (const int32_t *) dst->op_params; + const int stride_x = p[0]; + const int stride_y = p[1]; + const int padding_x = p[2]; + const int padding_y = p[3]; + const int dilation_x = p[4]; + const int dilation_y = p[5]; + + const int in_w = input->ne[0]; + const int in_h = input->ne[1]; + const int kernel_w = kernel->ne[0]; + const int kernel_h = kernel->ne[1]; + const int out_w = dst->ne[0]; + const int out_h = dst->ne[1]; + const int channels = dst->ne[2]; + const int batches = dst->ne[3]; + + cudaStream_t st = ctx.stream(); + + const int total = batches * channels * out_h * out_w; + const int blocks = (total + CUDA_CONV2D_DW_BLOCK_SIZE - 1) / CUDA_CONV2D_DW_BLOCK_SIZE; + + if (ggml_is_contiguous(input)) { + conv2d_dw_kernel<<>>( + x_d, w_d, y_d, in_w, in_h, out_w, out_h, kernel_w, kernel_h, stride_x, stride_y, padding_x, padding_y, + dilation_x, dilation_y, channels, batches); + } else if (ggml_is_contiguous_channels(input)) { + conv2d_dw_kernel<<>>( + x_d, w_d, y_d, in_w, in_h, out_w, out_h, kernel_w, kernel_h, stride_x, stride_y, padding_x, padding_y, + dilation_x, dilation_y, channels, batches); + } else { + GGML_ABORT("Unsupported memory layout for conv_2d_dw"); + } +} diff --git a/ggml/src/ggml-cuda/conv2d-dw.cuh b/ggml/src/ggml-cuda/conv2d-dw.cuh new file mode 100644 index 000000000..b5d5a69d3 --- /dev/null +++ b/ggml/src/ggml-cuda/conv2d-dw.cuh @@ -0,0 +1,5 @@ +#pragma once +#include "common.cuh" + +#define CUDA_CONV2D_DW_BLOCK_SIZE 256 +void ggml_cuda_op_conv2d_dw(ggml_backend_cuda_context & ctx, ggml_tensor * dst); diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu index 898b24341..80fe05073 100644 --- a/ggml/src/ggml-cuda/ggml-cuda.cu +++ b/ggml/src/ggml-cuda/ggml-cuda.cu @@ -11,6 +11,7 @@ #include "ggml-cuda/clamp.cuh" #include "ggml-cuda/concat.cuh" #include "ggml-cuda/conv-transpose-1d.cuh" +#include "ggml-cuda/conv2d-dw.cuh" #include "ggml-cuda/convert.cuh" #include "ggml-cuda/count-equal.cuh" #include "ggml-cuda/cpy.cuh" @@ -2310,6 +2311,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg case GGML_OP_IM2COL: ggml_cuda_op_im2col(ctx, dst); break; + case GGML_OP_CONV_2D_DW: + ggml_cuda_op_conv2d_dw(ctx, dst); + break; case GGML_OP_CONV_TRANSPOSE_1D: ggml_cuda_op_conv_transpose_1d(ctx,dst); break; @@ -3209,6 +3213,7 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g return op->src[0]->nb[0] == ggml_type_size(op->src[0]->type) && ggml_is_contiguous_2(op->src[0]); } case GGML_OP_IM2COL: + case GGML_OP_CONV_2D_DW: case GGML_OP_POOL_2D: case GGML_OP_SUM: case GGML_OP_SUM_ROWS: From 4c9fdfbe1580a66fd7d77c77418ce2c606a29fdd Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Fri, 20 Jun 2025 10:14:14 +0300 Subject: [PATCH 11/43] ubatch : new splitting logic (#14217) ggml-ci --- src/llama-batch.cpp | 924 +++++++++++++++++----------- src/llama-batch.h | 166 ++--- src/llama-context.cpp | 133 ++-- src/llama-context.h | 2 +- src/llama-graph.cpp | 264 +++----- src/llama-graph.h | 6 +- src/llama-hparams.cpp | 4 + src/llama-hparams.h | 2 + src/llama-kv-cache-unified-iswa.cpp | 40 +- src/llama-kv-cache-unified-iswa.h | 7 +- src/llama-kv-cache-unified.cpp | 136 ++-- src/llama-kv-cache-unified.h | 7 +- src/llama-kv-cells.h | 4 +- src/llama-memory-hybrid.cpp | 81 ++- src/llama-memory-hybrid.h | 9 +- src/llama-memory-recurrent.cpp | 76 ++- src/llama-memory-recurrent.h | 7 +- src/llama-memory.h | 7 +- tools/server/server.cpp | 32 - 19 files changed, 992 insertions(+), 915 deletions(-) diff --git a/src/llama-batch.cpp b/src/llama-batch.cpp index 8b6d14fe8..b3c996e18 100644 --- a/src/llama-batch.cpp +++ b/src/llama-batch.cpp @@ -1,7 +1,6 @@ #include "llama-batch.h" #include "llama-impl.h" -#include "llama-cparams.h" #include "llama-vocab.h" #include "llama-memory.h" @@ -10,282 +9,7 @@ #include #include -llama_ubatch llama_sbatch::reserve_ubatch(size_t n_ubatch, bool has_embd) { - // clear empty sequences - // the previous ubatch is assumed to be gone, - // so nothing should refer to values in these sequences anymore. - for (size_t i = seq.size(); i-- > 0;) { - if (seq[i].length == 0) { - seq.pop_back(); - } else { - break; - } - } - - udatas.push_back({}); - - auto & udata = udatas.back(); - - udata.token.resize(!has_embd ? n_ubatch : 0); - udata.embd.resize(has_embd ? n_embd * n_ubatch : 0); - udata.pos.resize(n_ubatch); - udata.n_seq_id.resize(n_ubatch); - udata.seq_id.resize(n_ubatch); - udata.output.resize(n_ubatch); - - llama_ubatch ubatch = { - /*equal_seqs =*/ true, - /*n_tokens =*/ 0, - /*n_seq_tokens =*/ 0, - /*n_seqs =*/ 0, - /*token =*/ !has_embd ? udata.token.data() : nullptr, - /*embd =*/ has_embd ? udata.embd.data() : nullptr, - /*pos =*/ udata.pos.data(), - /*n_seq_id =*/ udata.n_seq_id.data(), - /*seq_id =*/ udata.seq_id.data(), - /*output =*/ udata.output.data(), - }; - - return ubatch; -} - -void llama_sbatch::add_seq_to_ubatch(llama_ubatch & ubatch, llama_sbatch_seq & seq, size_t length) { - GGML_ASSERT(batch != nullptr); - GGML_ASSERT(length <= seq.length); - // Can only add sequences of equal lengths to a batch, - // otherwise it isn't clear to which sequence a token belongs - GGML_ASSERT(seq.n_seq_id == 0 || ubatch.n_seqs == 0 || length == (size_t) ubatch.n_tokens / ubatch.n_seqs); - GGML_ASSERT((seq.n_seq_id != 0) == ubatch.equal_seqs); - // NOTE: loops are separated for cache-friendliness - if (batch->token) { - if (ubatch.equal_seqs) { - for (size_t i = 0; i < length; ++i) { - ubatch.token[ubatch.n_tokens + i] = batch->token[ids[seq.offset + i]]; - } - } else { - // simple split - ubatch.token = batch->token + seq.offset; - } - } else { - ubatch.token = nullptr; - } - if (batch->embd) { - if (ubatch.equal_seqs) { - for (size_t i = 0; i < length; ++i) { - memcpy( - ubatch.embd + (n_embd * (ubatch.n_tokens + i)), - batch->embd + (n_embd * ids[seq.offset + i]), - n_embd * sizeof(float) - ); - } - } else { - // simple split - ubatch.embd = batch->embd + (n_embd * seq.offset); - } - } else { - ubatch.embd = nullptr; - } - if (ubatch.equal_seqs) { - for (size_t i = 0; i < length; ++i) { - ubatch.pos[ubatch.n_tokens + i] = batch->pos[ids[seq.offset + i]]; - } - } else { - // simple split - ubatch.pos = batch->pos + seq.offset; - } - if (ubatch.equal_seqs) { - ubatch.n_seq_id[ubatch.n_seqs] = seq.n_seq_id; - if (seq.seq_id) { - ubatch.seq_id[ubatch.n_seqs] = seq.seq_id; - } - } else { - // simple split - if (batch->n_seq_id) { - ubatch.n_seq_id = batch->n_seq_id + seq.offset; - } else { - for (size_t i = 0; i < length; ++i) { - ubatch.n_seq_id[ubatch.n_seqs + i] = 1; - } - } - if (batch->seq_id) { - ubatch.seq_id = batch->seq_id + seq.offset; - } - } - if (batch->logits) { - if (ubatch.equal_seqs) { - for (size_t i = 0; i < length; ++i) { - size_t id = ids[seq.offset + i]; - int8_t is_output = batch->logits[id]; - ubatch.output[ubatch.n_tokens + i] = is_output; - if (is_output) { out_ids.push_back(id); } - } - } else { - // simple split - ubatch.output = batch->logits + seq.offset; - for (size_t i = 0; i < length; ++i) { - if (ubatch.output[i] != 0) { out_ids.push_back(seq.offset + i); } - } - } - } else { - // only get last output - for (size_t i = 0; i < length; ++i) { - size_t id = ids[seq.offset + i]; - int8_t is_last = id == ids.size() - 1; - ubatch.output[ubatch.n_tokens + i] = is_last; - if (is_last) { out_ids.push_back(id); } - } - } - if (ubatch.n_tokens == 0 && ubatch.n_seqs == 0) { - ubatch.n_seq_tokens = ubatch.equal_seqs ? length : 1; - } - ubatch.n_tokens += length; - ubatch.n_seqs += ubatch.equal_seqs ? 1 : length; // virtual sequences for simple splits - seq.offset += length; - seq.length -= length; - n_tokens -= length; - GGML_ASSERT(ubatch.n_tokens == ubatch.n_seq_tokens * ubatch.n_seqs); -} - -llama_ubatch llama_sbatch::split_simple(size_t n_ubatch) { - n_ubatch = n_tokens < n_ubatch ? n_tokens : n_ubatch; - llama_ubatch ubatch = reserve_ubatch(n_ubatch, /* has_embd */ batch->embd != nullptr); - ubatch.equal_seqs = false; - if (!seq.empty()) { - llama_sbatch_seq & s = seq[0]; - size_t length = s.length < n_ubatch ? s.length : n_ubatch; - GGML_ASSERT(seq.size() == 1 && s.n_seq_id == 0); // don't mix with other splits - add_seq_to_ubatch(ubatch, s, length); - } - return ubatch; -} - -llama_ubatch llama_sbatch::split_equal(size_t n_ubatch) { - n_ubatch = n_tokens < n_ubatch ? n_tokens : n_ubatch; - llama_ubatch ubatch = reserve_ubatch(n_ubatch, /* has_embd */ batch->embd != nullptr); - if (!seq.empty()) { - size_t length = 0; - size_t n_tokens_in_ubatch = 0; - GGML_ASSERT(seq[0].n_seq_id > 0); // should not be mixed with simple splits - // smallest first, because it's easier to split this way; - // starting from the end to pop in constant time. - for (size_t i = seq.size(); i-- > 0;) { - llama_sbatch_seq & s = seq[i]; - GGML_ASSERT(s.length > 0); - if (length == 0) { - length = s.length < n_ubatch ? s.length : n_ubatch; - } - add_seq_to_ubatch(ubatch, s, length); - n_tokens_in_ubatch += length; - // shared prompts can't be mixed with any of their sequences, - // so it's safer to compute them in their own ubatch - if (s.n_seq_id > 1) { break; } - // stop when there isn't enough space for another sequence - if (length + n_tokens_in_ubatch > n_ubatch) { break; } - } - } - return ubatch; -} - -llama_ubatch llama_sbatch::split_seq(size_t n_ubatch) { - n_ubatch = n_tokens < n_ubatch ? n_tokens : n_ubatch; - llama_ubatch ubatch = reserve_ubatch(n_ubatch, /* has_embd */ batch->embd != nullptr); - if (!seq.empty()) { - llama_sbatch_seq & s = seq[seq.size() - 1]; - size_t length = s.length < n_ubatch ? s.length : n_ubatch; - GGML_ASSERT(s.n_seq_id > 0); // should not be mixed with simple splits - add_seq_to_ubatch(ubatch, s, length); - } - return ubatch; -} - -llama_sbatch::llama_sbatch(const llama_batch & batch, size_t n_embd, bool simple_split) { - GGML_ASSERT(batch.n_tokens >= 0); - this->batch = &batch; - this->n_embd = n_embd; - - n_tokens = batch.n_tokens; - ids.resize(n_tokens); - out_ids.clear(); - // TODO: reserve out_ids and seq - - for (size_t i = 0; i < n_tokens; ++i) { - ids[i] = i; - } - - if (simple_split) { - seq.resize(1); - llama_sbatch_seq & s = seq[0]; - s.n_seq_id = 0; - s.seq_id = nullptr; - s.offset = 0; - s.length = n_tokens; - return; - } - - std::sort(ids.begin(), ids.end(), - [&batch](size_t a, size_t b) { - int32_t n_seq_a = batch.n_seq_id ? batch.n_seq_id[a] : 1; - int32_t n_seq_b = batch.n_seq_id ? batch.n_seq_id[b] : 1; - // sort by seq_id, then by pos - if (n_seq_a == n_seq_b) { - if (batch.seq_id) { - for (int32_t i = 0; i < n_seq_a; ++i) { - llama_seq_id seq_id_a = batch.seq_id[a][i]; - llama_seq_id seq_id_b = batch.seq_id[b][i]; - // smaller seq_ids go first - if (seq_id_a != seq_id_b) { - return seq_id_a < seq_id_b; - } - } - } - // when all else is equal, sort by pos - if (batch.pos) { - return batch.pos[a] < batch.pos[b]; - } - // no pos, sort by id - return a < b; - } - // shared prompts go first - return n_seq_a > n_seq_b; - } - ); - - // init seq - llama_sbatch_seq * last_seq = nullptr; - - for (size_t i = 0; i < n_tokens; ++i) { - const size_t bi = ids[i]; - const int32_t n_seqs = batch.n_seq_id[bi]; - llama_seq_id * seq_ids = batch.seq_id[bi]; - if (last_seq != nullptr) { - bool same = n_seqs == last_seq->n_seq_id; - for (int32_t j = 0; same && j < n_seqs; ++j) { - if (seq_ids[j] != last_seq->seq_id[j]) { - same = false; - } - } - if (same) { - last_seq->length += 1; - continue; - } - } - llama_sbatch_seq new_seq = {n_seqs, seq_ids, i, 1}; - seq.push_back(new_seq); - last_seq = &seq.back(); - } - - // keep shared prompts first at the end, then sort by length descending. - std::sort(seq.begin(), seq.end(), - [](llama_sbatch_seq & a, llama_sbatch_seq & b) { - if (a.n_seq_id == b.n_seq_id) { - return a.length > b.length; - } - return a.n_seq_id < b.n_seq_id; - } - ); -} - -llama_batch_allocr::llama_batch_allocr() { +llama_batch_allocr::llama_batch_allocr(uint32_t n_pos_per_embd) : n_pos_per_embd(n_pos_per_embd) { const char * LLAMA_BATCH_DEBUG = getenv("LLAMA_BATCH_DEBUG"); debug = LLAMA_BATCH_DEBUG ? atoi(LLAMA_BATCH_DEBUG) : 0; @@ -294,17 +18,22 @@ llama_batch_allocr::llama_batch_allocr() { for (auto & cur : seq_cpl) { cur.resize(LLAMA_MAX_SEQ); } + + seq_idx.resize(LLAMA_MAX_SEQ, -1); } bool llama_batch_allocr::init( const llama_batch & batch_inp, const llama_vocab & vocab, const llama_memory_i * memory, - bool embd_all) { + uint32_t n_embd, + bool output_all) { clear(); batch = batch_inp; + this->vocab = &vocab; + GGML_ASSERT(batch.n_tokens > 0); // @@ -359,6 +88,7 @@ bool llama_batch_allocr::init( llama_pos p0[LLAMA_MAX_SEQ]; for (int32_t s = 0; s < LLAMA_MAX_SEQ; ++s) { if (!memory) { + // if no memory -> start from 0 p0[s] = 0; } else { p0[s] = memory->seq_pos_max(s) + 1; @@ -370,8 +100,11 @@ bool llama_batch_allocr::init( pos[i] = p0[seq_id]; + // update the starting position for all sequences that are assigned to the this token for (int32_t s = 0; s < batch.n_seq_id[i]; ++s) { - p0[batch.seq_id[i][s]] = pos[i] + 1; + const llama_seq_id seq_id = batch.seq_id[i][s]; + + p0[seq_id] = pos[i] + 1; } } @@ -379,7 +112,7 @@ bool llama_batch_allocr::init( } if (!batch.logits) { - if (embd_all) { + if (output_all) { // return the output for all tokens output.resize(batch.n_tokens, true); } else { @@ -389,7 +122,7 @@ bool llama_batch_allocr::init( } batch.logits = output.data(); - } else if (embd_all) { + } else if (output_all) { bool warn = false; for (int32_t i = 0; i < batch.n_tokens; ++i) { @@ -410,6 +143,9 @@ bool llama_batch_allocr::init( // compute stats // + this->n_embd = n_embd; + + // count the outputs in this batch for (int32_t i = 0; i < batch.n_tokens; ++i) { n_outputs += batch.logits[i] != 0; } @@ -417,85 +153,86 @@ bool llama_batch_allocr::init( // determine coupled sequences // these are pairs of sequences that have at least one token in the input batch that is assigned to both of them for (int32_t i = 0; i < batch.n_tokens; ++i) { + const llama_seq_id s0 = batch.seq_id[i][0]; + for (int32_t s = 0; s < batch.n_seq_id[i]; ++s) { - seq_pos[batch.seq_id[i][s]].insert(batch.pos[i]); + const llama_seq_id s1 = batch.seq_id[i][s]; + + seq_pos[s1].insert(batch.pos[i]); if (s > 0) { - const llama_seq_id s0 = batch.seq_id[i][0]; - const llama_seq_id s1 = batch.seq_id[i][s]; - // mark that sequence s1 is coupled to s0 seq_cpl[s1][s0] = true; - // note: the other way around is not necessary for now + // note: tracking the other way around is not necessary for now //seq_cpl[s0][s1] = true; } } } + // precompute the sequence sets for each token and determine the unique sequence ids that participate in the batch + { + seq_set_t seq_set_unq; + + for (int32_t i = 0; i < batch.n_tokens; ++i) { + seq_set_t cur; + for (int32_t s = 0; s < batch.n_seq_id[i]; ++s) { + const llama_seq_id seq_id = batch.seq_id[i][s]; + + cur .set(seq_id); + seq_set_unq.set(seq_id); + } + + seq_set.push_back(cur); + seq_set_map[cur].push_back(i); + } + + for (int32_t s = 0; s < LLAMA_MAX_SEQ; ++s) { + if (seq_set_unq.test(s)) { + seq_idx[s] = seq_id_unq.size(); + seq_id_unq.push_back(s); + } + } + } + if (debug > 0) { LLAMA_LOG_DEBUG("%s: input batch info:\n", __func__); - LLAMA_LOG_DEBUG("%s: n_tokens = %d\n", __func__, batch.n_tokens); - LLAMA_LOG_DEBUG("%s: token = %p\n", __func__, (void *) batch.token); - LLAMA_LOG_DEBUG("%s: embd = %p\n", __func__, (void *) batch.embd); - LLAMA_LOG_DEBUG("%s: pos = %p\n", __func__, (void *) batch.pos); - LLAMA_LOG_DEBUG("%s: n_seq_id = %p\n", __func__, (void *) batch.n_seq_id); - LLAMA_LOG_DEBUG("%s: seq_id = %p\n", __func__, (void *) batch.seq_id); - LLAMA_LOG_DEBUG("%s: logits = %p\n", __func__, (void *) batch.logits); - LLAMA_LOG_DEBUG("%s: n_outputs = %d\n", __func__, n_outputs); - if (debug > 1) { - int seq_id_max = 0; - for (int32_t i = 0; i < batch.n_tokens; ++i) { - for (int s = 0; s < batch.n_seq_id[i]; ++s) { - for (int s = 0; s < batch.n_seq_id[i]; ++s) { - seq_id_max = std::max(seq_id_max, batch.seq_id[i][s]); - } + llama_ubatch ubatch { + /*.equal_seqs =*/ false, + /*.n_tokens =*/ (uint32_t) batch.n_tokens, + /*.n_seq_tokens =*/ (uint32_t) 1, + /*.n_seqs =*/ (uint32_t) batch.n_tokens, + /*.n_seqs_unq =*/ (uint32_t) this->seq_id_unq.size(), + /*.token =*/ batch.token, + /*.embd =*/ batch.embd, + /*.pos =*/ batch.pos, + /*.n_seq_id =*/ batch.n_seq_id, + /*.seq_id =*/ batch.seq_id, + /*.seq_id_unq =*/ this->seq_id_unq.data(), + /*.seq_idx =*/ this->seq_idx.data(), + /*.output =*/ batch.logits, + }; + + ubatch_print(ubatch, debug); + + LLAMA_LOG_DEBUG("%s: seq = [\n", __func__); + for (int s0 = 0; s0 < (int) seq_pos.size(); ++s0) { + if (seq_pos[s0].empty()) { + continue; + } + + std::stringstream ss; + for (int s1 = 0; s1 < (int) seq_cpl[s0].size(); ++s1) { + if (seq_cpl[s0][s1]) { + ss << s1 << " "; } } - ++seq_id_max; - LLAMA_LOG_DEBUG("%s: token = [\n", __func__); - for (int32_t i = 0; i < batch.n_tokens; ++i) { - std::vector seq_id(seq_id_max); - - for (int s = 0; s < batch.n_seq_id[i]; ++s) { - seq_id[batch.seq_id[i][s]] = 1; - } - - std::stringstream ss; - for (int s = 0; s < seq_id_max; ++s) { - if (seq_id[s]) { - ss << s%10; - } else { - ss << "."; - } - } - - LLAMA_LOG_DEBUG("%s: %4d: id = %6d (%16s), pos = %4d, n_seq_id = %2d, seq_id = [%s], output = %d\n", - __func__, i, batch.token[i], vocab.token_to_piece(batch.token[i]).c_str(), - batch.pos[i], batch.n_seq_id[i], ss.str().c_str(), batch.logits[i]); - } - LLAMA_LOG_DEBUG("%s: ]\n", __func__); - - LLAMA_LOG_DEBUG("%s: seq = [\n", __func__); - for (int s0 = 0; s0 < (int) seq_pos.size(); ++s0) { - if (seq_pos[s0].empty()) { - continue; - } - - std::stringstream ss; - for (int s1 = 0; s1 < (int) seq_cpl[s0].size(); ++s1) { - if (seq_cpl[s0][s1]) { - ss << s1 << " "; - } - } - - LLAMA_LOG_DEBUG("%s: %4d: pos = [%4d, %4d], cpl = %s\n", - __func__, s0, seq_pos_min(s0), seq_pos_max(s0), ss.str().empty() ? "-" : ss.str().c_str()); - } - LLAMA_LOG_DEBUG("%s: ]\n", __func__); + LLAMA_LOG_DEBUG("%s: %4d: pos = [%4d, %4d], cpl = %s\n", + __func__, s0, seq_pos_min(s0), seq_pos_max(s0), ss.str().empty() ? "-" : ss.str().c_str()); } + LLAMA_LOG_DEBUG("%s: ]\n", __func__); } // @@ -507,9 +244,22 @@ bool llama_batch_allocr::init( continue; } - if (memory && seq_pos_min(s) != memory->seq_pos_max(s) + 1) { - LLAMA_LOG_ERROR("%s: sequence %d does not start from the last position stored in the memory\n", __func__, s); - return false; + if (memory) { + if (batch.token) { + if (seq_pos_min(s) != memory->seq_pos_max(s) + 1) { + LLAMA_LOG_ERROR("%s: sequence %d does not start from the last position stored in the memory\n", __func__, s); + return false; + } + } else { + assert(batch.embd); + + // for embeddings (typically used as vision input), we allow them to have repeating positions + // ref: https://github.com/ggml-org/llama.cpp/issues/13694#issuecomment-2983871762 + if (seq_pos_min(s) != memory->seq_pos_max(s) && seq_pos_min(s) != memory->seq_pos_max(s) + 1) { + LLAMA_LOG_ERROR("%s: sequence %d does not start from the last position stored in the memory\n", __func__, s); + return false; + } + } } if (seq_pos_max(s) - seq_pos_min(s) + 1 > (int) seq_pos[s].size()) { @@ -532,17 +282,120 @@ bool llama_batch_allocr::init( } } + // disallow partial sequence sub-sets: + // + // invalid: x + // i: 0 1 2 ... + // --------------------------------------- + // seq_id[i][0]: 0 0 1 + // seq_id[i][1]: 1 1 2 + // seq_id[i][2]: 2 + // + // disallow decreasing sequence positions: + // + // invalid: x + // i: 0 1 2 3 4 5 6 ... + // --------------------------------------- + // pos[i]: 4 5 0 1 6 2 3 + // seq_id[i][0]: 0 0 1 1 0 1 0 + // + { + seq_set_t cur_seq_set[LLAMA_MAX_SEQ]; + for (int32_t s = 0; s < LLAMA_MAX_SEQ; ++s) { + cur_seq_set[s].set(); + } + + llama_pos cur_seq_pos[LLAMA_MAX_SEQ]; + for (int32_t s = 0; s < LLAMA_MAX_SEQ; ++s) { + cur_seq_pos[s] = -1; + } + + for (int32_t i = 0; i < batch.n_tokens; ++i) { + const llama_pos pos = batch.pos[i]; + + for (int32_t s = 0; s < batch.n_seq_id[i]; ++s) { + const llama_seq_id seq_id = batch.seq_id[i][s]; + + cur_seq_set[seq_id] &= seq_set[i]; + + if (cur_seq_set[seq_id].none()) { + LLAMA_LOG_ERROR("%s: sequence %d belongs to incompatible sequence sets (not allowed)\n", __func__, seq_id); + return false; + } + + if (pos < cur_seq_pos[seq_id]) { + LLAMA_LOG_ERROR("%s: sequence %d positions are decreasing (not allowed)\n", __func__, seq_id); + return false; + } + } + } + } + + split_reset(); + return true; } +llama_ubatch llama_batch_allocr::ubatch_reserve(uint32_t n_seq_tokens, uint32_t n_seqs) { + const uint32_t n_tokens = n_seq_tokens*n_seqs; + + clear(); + split_reset(); + + ubatches.emplace_back(); + + auto & ubatch = ubatches.back(); + + ubatch.token .resize(n_tokens); + ubatch.embd .clear(); + ubatch.pos .resize(n_tokens); + ubatch.n_seq_id .resize(n_tokens); + ubatch.seq_id .resize(n_tokens); + ubatch.seq_id_unq.resize(0); + ubatch.seq_idx .resize(LLAMA_MAX_SEQ, -1); + ubatch.output .resize(n_tokens); + + for (uint32_t s = 0; s < n_seqs; ++s) { + ubatch.seq_idx[s] = s; + ubatch.seq_id_unq.push_back(s); + } + + llama_ubatch res { + /*.equal_seqs =*/ true, + /*.n_tokens =*/ n_tokens, + /*.n_seq_tokens =*/ n_seq_tokens, + /*.n_seqs =*/ n_seqs, + /*.n_seqs_unq =*/ n_seqs, + + /*.token =*/ ubatch.token.data(), + /*.embd =*/ nullptr, + /*.pos =*/ ubatch.pos.data(), + /*.n_seq_id =*/ ubatch.n_seq_id.data(), + /*.seq_id =*/ ubatch.seq_id.data(), + /*.seq_id_unq =*/ ubatch.seq_id_unq.data(), + /*.seq_idx =*/ ubatch.seq_idx.data(), + /*.output =*/ ubatch.output.data(), + }; + + return res; +} + const llama_batch & llama_batch_allocr::get_batch() const { return batch; } +uint32_t llama_batch_allocr::get_n_tokens() const { + return batch.n_tokens; +} + uint32_t llama_batch_allocr::get_n_outputs() const { return n_outputs; } +std::vector & llama_batch_allocr::get_out_ids() { + return out_ids; +} + llama_pos llama_batch_allocr::seq_pos_min(llama_seq_id seq_id) const { return seq_pos[seq_id].empty() ? -1 : *seq_pos[seq_id].begin(); } @@ -551,14 +404,188 @@ llama_pos llama_batch_allocr::seq_pos_max(llama_seq_id seq_id) const { return seq_pos[seq_id].empty() ? -1 : *seq_pos[seq_id].rbegin(); } +void llama_batch_allocr::split_reset() { + out_ids.clear(); + + used.clear(); + used.resize(get_n_tokens(), false); + + ubatches.clear(); +} + +llama_ubatch llama_batch_allocr::split_simple(uint32_t n_ubatch) { + // find the first unused token + uint32_t cur_idx = 0; + while (cur_idx < used.size() && used[cur_idx]) { + ++cur_idx; + } + + // we are done + if (cur_idx >= used.size()) { + return {}; + } + + std::vector idxs; + + while (true) { + idxs.push_back(cur_idx); + + used[cur_idx] = true; + + ++cur_idx; + + if (cur_idx >= used.size()) { + break; + } + + if (idxs.size() >= n_ubatch) { + break; + } + } + + return ubatch_add(idxs, idxs.size(), false); +} + +llama_ubatch llama_batch_allocr::split_equal(uint32_t n_ubatch) { + std::vector cur_seq_set; + + // determine the non-overlapping sequence sets participating in this ubatch + for (int32_t i = 0; i < batch.n_tokens; ++i) { + if (used[i]) { + continue; + } + + bool add = true; + + for (uint32_t s = 0; s < cur_seq_set.size(); ++s) { + // no overlap with existing sequence sets: + if (!(cur_seq_set[s] & seq_set[i]).none()) { + add = false; + break; + } + } + + if (add) { + cur_seq_set.push_back(seq_set[i]); + + if (cur_seq_set.size() > n_ubatch) { + break; + } + } + } + + const uint32_t n_seqs = cur_seq_set.size(); + + // we are done + if (n_seqs == 0) { + return {}; + } + + // the current batch index of each sequence set + std::vector cur_idx(n_seqs, 0); + + for (uint32_t s = 0; s < n_seqs; ++s) { + while (used[seq_set_map[cur_seq_set[s]][cur_idx[s]]]) { + ++cur_idx[s]; + } + } + + // the list of batch indices for each sequence set + // at the end we will concat these to get the final ubatch + std::vector idxs_per_seq(n_seqs); + + while (true) { + // we can only add new n_seq_tokens tokens if all the sequence sets have at least one more unused token and + // if we haven't reached n_ubatch + bool can_expand = true; + + for (uint32_t s = 0; s < n_seqs; ++s) { + if (cur_idx[s] >= (int32_t) seq_set_map[cur_seq_set[s]].size()) { + can_expand = false; + break; + } + } + + if (!can_expand) { + break; + } + + for (uint32_t s = 0; s < n_seqs; ++s) { + const int32_t idx = seq_set_map[cur_seq_set[s]][cur_idx[s]]; + + idxs_per_seq[s].push_back(idx); + + used[idx] = true; + + ++cur_idx[s]; + } + + if ((idxs_per_seq[0].size() + 1)*n_seqs > n_ubatch) { + break; + } + } + + // concat the per-sequence-set lists + std::vector idxs; + + for (uint32_t s = 0; s < n_seqs; ++s) { + idxs.insert(idxs.end(), idxs_per_seq[s].begin(), idxs_per_seq[s].end()); + } + + return ubatch_add(idxs, n_seqs, true); +} + +llama_ubatch llama_batch_allocr::split_seq(uint32_t n_ubatch) { + // find the first unused token + uint32_t cur_idx = 0; + while (cur_idx < used.size() && used[cur_idx]) { + ++cur_idx; + } + + // we are done + if (cur_idx >= used.size()) { + return {}; + } + + // this is the starting sequence set + // we allow adding tokens only if their sequence set is a subset of the current sequence set + auto cur_seq_set = seq_set[cur_idx]; + + std::vector idxs; + + while (true) { + idxs.push_back(cur_idx); + + used[cur_idx] = true; + + if (idxs.size() >= n_ubatch) { + break; + } + + do { + ++cur_idx; + } while (cur_idx < get_n_tokens() && (used[cur_idx] || ((cur_seq_set & seq_set[cur_idx]) != seq_set[cur_idx]))); + + if (cur_idx == get_n_tokens()) { + break; + } + + cur_seq_set = seq_set[cur_idx]; + } + + return ubatch_add(idxs, 1, true); +} + void llama_batch_allocr::clear() { n_outputs = 0; batch = {}; - pos.clear(); - n_seq_id.clear(); - seq_id.clear(); - output.clear(); + + pos .clear(); + n_seq_id .clear(); + seq_id .clear(); + seq_id_unq.clear(); + output .clear(); for (auto & cur : seq_pos) { cur.clear(); @@ -567,6 +594,177 @@ void llama_batch_allocr::clear() { for (auto & cur : seq_cpl) { std::fill(cur.begin(), cur.end(), false); } + + seq_set.clear(); + + seq_set_map.clear(); + + std::fill(seq_idx.begin(), seq_idx.end(), -1); +} + +llama_ubatch llama_batch_allocr::ubatch_add(const std::vector & idxs, uint32_t n_seqs, bool equal_seqs) { + const uint32_t n_tokens = idxs.size(); + + assert(n_tokens%n_seqs == 0); + + ubatches.emplace_back(); + + auto & ubatch = ubatches.back(); + + const int32_t n_pos_cur = batch.embd ? n_pos_per_embd : 1; + + const int64_t n_embd_all = batch.embd ? (int64_t) n_tokens*n_embd : 0; + const int64_t n_pos_all = (int64_t) n_tokens*n_pos_cur; + + ubatch.token .resize(n_tokens); + ubatch.embd .resize(n_embd_all); + ubatch.pos .resize(n_pos_all); + ubatch.n_seq_id .resize(n_tokens); + ubatch.seq_id .resize(n_tokens); + ubatch.seq_id_unq.resize(0); + ubatch.seq_idx .resize(LLAMA_MAX_SEQ, -1); + ubatch.output .resize(n_tokens); + + seq_set_t seq_set_unq; + + for (size_t i = 0; i < idxs.size(); ++i) { + if (batch.token) { + ubatch.token[i] = batch.token[idxs[i]]; + } + + if (batch.embd) { + memcpy(ubatch.embd.data() + i*n_embd, batch.embd + (int64_t) idxs[i]*n_embd, n_embd*sizeof(float)); + } + + for (int j = 0; j < n_pos_cur; ++j) { + ubatch.pos[j*n_tokens + i] = batch.pos[j*batch.n_tokens + idxs[i]]; + } + + ubatch.n_seq_id[i] = batch.n_seq_id[idxs[i]]; + ubatch.seq_id[i] = batch.seq_id[idxs[i]]; + ubatch.output[i] = batch.logits[idxs[i]]; + + for (int s = 0; s < ubatch.n_seq_id[i]; ++s) { + seq_set_unq.set(ubatch.seq_id[i][s]); + } + + if (ubatch.output[i]) { + out_ids.push_back(idxs[i]); + } + } + + for (int32_t s = 0; s < LLAMA_MAX_SEQ; ++s) { + if (seq_set_unq.test(s)) { + ubatch.seq_idx[s] = ubatch.seq_id_unq.size(); + ubatch.seq_id_unq.push_back(s); + } + } + + llama_ubatch res { + /*.equal_seqs =*/ equal_seqs, + /*.n_tokens =*/ n_tokens, + /*.n_seq_tokens =*/ n_tokens/n_seqs, + /*.n_seqs =*/ n_seqs, + /*.n_seqs_unq =*/ (uint32_t) ubatch.seq_id_unq.size(), + + /*.token =*/ batch.token ? ubatch.token.data() : nullptr, + /*.embd =*/ batch.embd ? ubatch.embd.data() : nullptr, + /*.pos =*/ ubatch.pos.data(), + /*.n_seq_id =*/ ubatch.n_seq_id.data(), + /*.seq_id =*/ ubatch.seq_id.data(), + /*.seq_id_unq =*/ ubatch.seq_id_unq.data(), + /*.seq_idx =*/ ubatch.seq_idx.data(), + /*.output =*/ ubatch.output.data(), + }; + + if (debug > 0) { + LLAMA_LOG_DEBUG("%s: added ubatch %d to split:\n", __func__, (int) ubatches.size() - 1); + + ubatch_print(res, debug); + } + + return res; +} + +void llama_batch_allocr::ubatch_print(const llama_ubatch & ubatch, int debug) { + if (debug > 0) { + LLAMA_LOG_DEBUG("%s: equal_seqs = %d\n", __func__, ubatch.equal_seqs); + LLAMA_LOG_DEBUG("%s: n_tokens = %d\n", __func__, ubatch.n_tokens); + LLAMA_LOG_DEBUG("%s: n_seq_tokens = %d\n", __func__, ubatch.n_seq_tokens); + LLAMA_LOG_DEBUG("%s: n_seqs = %d\n", __func__, ubatch.n_seqs); + LLAMA_LOG_DEBUG("%s: n_seqs_unq = %d\n", __func__, ubatch.n_seqs_unq); + + std::stringstream ss_seq_id_unq; + std::stringstream ss_seq_idx; + + ss_seq_id_unq << "[ "; + ss_seq_idx << "["; + + for (uint32_t s = 0; s < ubatch.n_seqs_unq; ++s) { + ss_seq_id_unq << ubatch.seq_id_unq[s] << " "; + } + + for (uint32_t s = 0; s < LLAMA_MAX_SEQ; ++s) { + if (ubatch.seq_idx[s] >= 0) { + ss_seq_idx << ubatch.seq_idx[s]%10; + } else { + ss_seq_idx << "."; + } + } + + ss_seq_id_unq << "]"; + ss_seq_idx << "]"; + + LLAMA_LOG_DEBUG("%s: token = %p\n", __func__, (void *) ubatch.token); + LLAMA_LOG_DEBUG("%s: embd = %p\n", __func__, (void *) ubatch.embd); + LLAMA_LOG_DEBUG("%s: pos = %p\n", __func__, (void *) ubatch.pos); + LLAMA_LOG_DEBUG("%s: n_seq_id = %p\n", __func__, (void *) ubatch.n_seq_id); + LLAMA_LOG_DEBUG("%s: seq_id = %p\n", __func__, (void *) ubatch.seq_id); + LLAMA_LOG_DEBUG("%s: seq_id_unq = %s\n", __func__, ss_seq_id_unq.str().c_str()); + LLAMA_LOG_DEBUG("%s: seq_idx = %s\n", __func__, ss_seq_idx.str().c_str()); + LLAMA_LOG_DEBUG("%s: output = %p\n", __func__, (void *) ubatch.output); + LLAMA_LOG_DEBUG("%s: n_outputs = %d\n", __func__, n_outputs); + + if (debug > 1) { + int seq_id_max = 0; + for (uint32_t i = 0; i < ubatch.n_tokens; ++i) { + for (int s = 0; s < ubatch.n_seq_id[i]; ++s) { + for (int s = 0; s < ubatch.n_seq_id[i]; ++s) { + seq_id_max = std::max(seq_id_max, ubatch.seq_id[i][s]); + } + } + } + ++seq_id_max; + + LLAMA_LOG_DEBUG("%s: token = [\n", __func__); + for (uint32_t i = 0; i < ubatch.n_tokens; ++i) { + std::vector seq_id(seq_id_max); + + for (int s = 0; s < ubatch.n_seq_id[i]; ++s) { + seq_id[ubatch.seq_id[i][s]] = 1; + } + + std::stringstream ss; + for (int s = 0; s < seq_id_max; ++s) { + if (seq_id[s]) { + ss << s%10; + } else { + ss << "."; + } + } + + if (ubatch.token) { + LLAMA_LOG_DEBUG("%s: %4d: id = %6d (%16s), pos = %4d, n_seq_id = %2d, seq_id = [%s], output = %d\n", + __func__, i, ubatch.token[i], vocab->token_to_piece(ubatch.token[i]).c_str(), + ubatch.pos[i], ubatch.n_seq_id[i], ss.str().c_str(), ubatch.output[i]); + } else { + LLAMA_LOG_DEBUG("%s: %4d: [embd], pos = %4d, n_seq_id = %2d, seq_id = [%s], output = %d\n", + __func__, i, ubatch.pos[i], ubatch.n_seq_id[i], ss.str().c_str(), ubatch.output[i]); + } + } + LLAMA_LOG_DEBUG("%s: ]\n", __func__); + } + } } // @@ -577,25 +775,25 @@ struct llama_batch llama_batch_get_one( llama_token * tokens, int32_t n_tokens) { return { - /*n_tokens =*/ n_tokens, - /*tokens =*/ tokens, - /*embd =*/ nullptr, - /*pos =*/ nullptr, - /*n_seq_id =*/ nullptr, - /*seq_id =*/ nullptr, - /*logits =*/ nullptr, + /*n_tokens =*/ n_tokens, + /*tokens =*/ tokens, + /*embd =*/ nullptr, + /*pos =*/ nullptr, + /*n_seq_id =*/ nullptr, + /*seq_id =*/ nullptr, + /*logits =*/ nullptr, }; } struct llama_batch llama_batch_init(int32_t n_tokens_alloc, int32_t embd, int32_t n_seq_max) { llama_batch batch = { - /*n_tokens =*/ 0, - /*tokens =*/ nullptr, - /*embd =*/ nullptr, - /*pos =*/ nullptr, - /*n_seq_id =*/ nullptr, - /*seq_id =*/ nullptr, - /*logits =*/ nullptr, + /*n_tokens =*/ 0, + /*tokens =*/ nullptr, + /*embd =*/ nullptr, + /*pos =*/ nullptr, + /*n_seq_id =*/ nullptr, + /*seq_id =*/ nullptr, + /*logits =*/ nullptr, }; if (embd) { diff --git a/src/llama-batch.h b/src/llama-batch.h index a555c1572..d2c537618 100644 --- a/src/llama-batch.h +++ b/src/llama-batch.h @@ -2,86 +2,44 @@ #include "llama.h" +#include "llama-cparams.h" + #include #include #include +#include +#include -// very similar to llama_batch, -// but has more metadata about sequences +// keep this struct lightweight +// it points to data in `llama_batch_allocr` struct llama_ubatch { bool equal_seqs; // TODO: whole_seqs for embeddings? uint32_t n_tokens; // total tokens (n_seq_tokens * n_seqs) - uint32_t n_seq_tokens; // tokens per sequence - uint32_t n_seqs; + uint32_t n_seq_tokens; // tokens per sequence set + uint32_t n_seqs; // sequence sets in the ubatch + uint32_t n_seqs_unq; // unique sequence ids in the ubatch - llama_token * token; // [n_tokens] - float * embd; // [n_embd, n_tokens] - llama_pos * pos; // [n_tokens] - int32_t * n_seq_id; // [n_seqs] - llama_seq_id ** seq_id; // [n_seqs] - int8_t * output; // [n_tokens] + // seq_id_unq: unique sequence ids in the ubatch + // seq_idx: indices of the unique sequence ids in the ubatch in [0, n_seqs_unq) + // used for extracting sequence pooled embeddings + + // // size | idx | val + llama_token * token; // [n_tokens] | i | id, token + float * embd; // [n_embd, n_tokens] | i | embd + llama_pos * pos; // [n_tokens] | i | pos + int32_t * n_seq_id; // [n_tokens] | i | - + llama_seq_id ** seq_id; // [n_tokens] | s | s0, s1, seq_id + llama_seq_id * seq_id_unq; // [n_seqs_unq] | s | seq_id + int32_t * seq_idx; // [LLAMA_MAX_SEQ] | - | seq_idx + int8_t * output; // [n_tokens] | i | - }; -struct llama_sbatch_seq { - int32_t n_seq_id; - - llama_seq_id * seq_id; - - size_t offset; - size_t length; -}; - -// sequence-length-aware batch splitting -struct llama_sbatch { - // tokens left in this batch - size_t n_tokens; - - size_t n_embd; - - // sorted indices into the batch - std::vector ids; - // batch indices of the output - std::vector out_ids; - std::vector seq; - - const llama_batch * batch = nullptr; - - // buffers for the ubatches - // TODO: very hacky, this needs a complete rework - struct ubatch_data { - std::vector token; - std::vector embd; - std::vector pos; - std::vector n_seq_id; - std::vector seq_id; - std::vector output; - }; - - std::vector udatas; - - llama_ubatch reserve_ubatch(size_t n_ubatch, bool has_embd = false); - - void add_seq_to_ubatch(llama_ubatch & ubatch, llama_sbatch_seq & seq, size_t length); - - // simple split, unknown number of sequences of unequal lengths - llama_ubatch split_simple(size_t n_ubatch); - - // make batches of equal-length sequences - llama_ubatch split_equal(size_t n_ubatch); - - // sequence-wise split - llama_ubatch split_seq(size_t n_ubatch); - - llama_sbatch() = default; - llama_sbatch(const llama_batch & batch, size_t n_embd, bool simple_split = false); -}; - -// a helper for sanitizing and fulfilling a batch +// a helper for sanitizing, fulfilling and splitting a batch class llama_batch_allocr { public: - llama_batch_allocr(); + llama_batch_allocr(uint32_t n_pos_per_embd); // sanitize and auto-gen missing data in the input batch // memory is optional. if provided will be used to check for sequence continuity and to determine the positions @@ -89,20 +47,57 @@ public: const llama_batch & batch_inp, const llama_vocab & vocab, const llama_memory_i * memory, - bool embd_all); + uint32_t n_embd, + bool output_all); const llama_batch & get_batch() const; + uint32_t get_n_tokens() const; uint32_t get_n_outputs() const; + // the array of output indices in the order they were encountered during the ubatch splitting + std::vector & get_out_ids(); + + // min/max positions of each sequence in the current ubatch llama_pos seq_pos_min(llama_seq_id seq_id) const; llama_pos seq_pos_max(llama_seq_id seq_id) const; + // call once before splitting the batch to reset the internal state + void split_reset(); + + // simple split, unknown number of sequence sets of unequal lengths + llama_ubatch split_simple(uint32_t n_ubatch); + + // make ubatches of equal-length sequences sets + llama_ubatch split_equal(uint32_t n_ubatch); + + // sequence-set-wise split - each ubatch contains a single sequence-set + llama_ubatch split_seq(uint32_t n_ubatch); + + // a helper method for creating a well-defined ubatch of tokens + // TODO: support embeddings if needed in the future + llama_ubatch ubatch_reserve(uint32_t n_seq_tokens, uint32_t n_seqs); + private: void clear(); + // create the next ubatch based on the provided batch indices (idxs) and the number of sequence sets (n_seqs) + // return llama_ubatch.n_tokens == 0 if the entire batch was consumed + llama_ubatch ubatch_add(const std::vector & idxs, uint32_t n_seqs, bool equal_seqs); + + // for debugging, start with LLAMA_BATCH_DEBUG=2 + void ubatch_print(const llama_ubatch & ubatch, int debug); + llama_batch batch; + // only for debugging purposes + const llama_vocab * vocab; + + // TODO: this is more of a temporary solution until we have a better way to handle multiple positions per token/embd + // ref: https://github.com/ggml-org/llama.cpp/issues/13694#issuecomment-2983871762 + const uint32_t n_pos_per_embd; + + uint32_t n_embd; uint32_t n_outputs; std::array seq_id_0 = { 0 }; // default sequence id @@ -110,10 +105,43 @@ private: std::vector pos; std::vector n_seq_id; std::vector seq_id; + std::vector seq_id_unq; + std::vector seq_idx; std::vector output; - std::vector> seq_pos; // seq_pos[s]: the set of positions in sequence s - std::vector> seq_cpl; // seq_cpl[s0][s1]: if sequence s0 is coupled to sequence s1 + using pos_set_t = std::set; + using seq_cpl_t = std::vector; + + std::vector seq_pos; // seq_pos[s]: the set of positions in sequence s + std::vector seq_cpl; // seq_cpl[s0][s1]: if sequence s0 is coupled to sequence s1 + + using idx_vec_t = std::vector; + using seq_set_t = std::bitset; + + std::vector seq_set; // seq_set[i]: the sequence set of token i + + std::unordered_map seq_set_map; // the indices at which the sequence set appears + + // batch indices of the output + std::vector out_ids; + + // used[i] indicates if token i has already been used in a previous ubatch + std::vector used; + + // llama_ubatch points to this data: + struct ubatch { + std::vector token; + std::vector embd; + std::vector pos; + std::vector n_seq_id; + std::vector seq_id; + std::vector seq_id_unq; + std::vector seq_idx; + std::vector output; + }; + + // current splitting state: + std::vector ubatches; int debug; }; diff --git a/src/llama-context.cpp b/src/llama-context.cpp index f56a58e9b..5a18a4fb3 100644 --- a/src/llama-context.cpp +++ b/src/llama-context.cpp @@ -20,7 +20,7 @@ llama_context::llama_context( const llama_model & model, llama_context_params params) : model(model), - batch_allocr(std::make_unique()) { + balloc(std::make_unique(model.hparams.n_pos_per_embd())) { LLAMA_LOG_INFO("%s: constructing llama_context\n", __func__); t_start_us = model.t_start_us; @@ -722,22 +722,26 @@ llm_graph_result_ptr llama_context::process_ubatch(const llama_ubatch & ubatch, } int llama_context::encode(const llama_batch & batch_inp) { + GGML_ASSERT((!batch_inp.token && batch_inp.embd) || (batch_inp.token && !batch_inp.embd)); // NOLINT + if (batch_inp.n_tokens == 0) { LLAMA_LOG_ERROR("%s: n_tokens == 0\n", __func__); return -1; } + const auto & hparams = model.hparams; + + const int64_t n_embd = hparams.n_embd; + // note: during encode, we always pass the full sequence starting from pos = 0 - if (!batch_allocr->init(batch_inp, model.vocab, nullptr, true)) { + if (!balloc->init(batch_inp, model.vocab, nullptr, n_embd, true)) { LLAMA_LOG_ERROR("%s: failed to initialize batch\n", __func__); return -1; } - const llama_batch & batch = batch_allocr->get_batch(); + const uint32_t n_tokens = balloc->get_n_tokens(); - const uint32_t n_tokens = batch.n_tokens; - - GGML_ASSERT((!batch.token && batch.embd) || (batch.token && !batch.embd)); // NOLINT + const llama_ubatch ubatch = balloc->split_simple(n_tokens); // micro-batching is not possible for non-causal encoding, so we process the batch in a single shot GGML_ASSERT(cparams.n_ubatch >= n_tokens && "encoder requires n_ubatch >= n_tokens"); @@ -751,14 +755,6 @@ int llama_context::encode(const llama_batch & batch_inp) { n_queued_tokens += n_tokens; - const auto & hparams = model.hparams; - - const int64_t n_embd = hparams.n_embd; - - llama_sbatch sbatch = llama_sbatch(batch, n_embd, /* simple_split */ true); - - const llama_ubatch ubatch = sbatch.split_simple(n_tokens); - // reserve output buffer if (output_reserve(n_tokens) < n_tokens) { LLAMA_LOG_ERROR("%s: could not reserve space for batch with %u outputs\n", __func__, n_tokens); @@ -817,34 +813,28 @@ int llama_context::encode(const llama_batch & batch_inp) { { // extract sequence embeddings auto & embd_seq_out = embd_seq; - embd_seq_out.clear(); - GGML_ASSERT(!ubatch.equal_seqs); // TODO: handle equal splits + for (uint32_t s = 0; s < ubatch.n_seqs_unq; ++s) { + const llama_seq_id seq_id = ubatch.seq_id_unq[s]; + const int32_t seq_idx = ubatch.seq_idx[seq_id]; - // TODO: fix indexing [UBATCH_IDX] - for (uint32_t i = 0; i < n_tokens; i++) { - const llama_seq_id seq_id = ubatch.seq_id[i][0]; - if (embd_seq_out.find(seq_id) != embd_seq_out.end()) { - continue; - } embd_seq_out[seq_id].resize(n_embd); - ggml_backend_tensor_get_async(backend_embd, t_embd, embd_seq_out[seq_id].data(), (n_embd*seq_id)*sizeof(float), n_embd*sizeof(float)); + ggml_backend_tensor_get_async(backend_embd, t_embd, embd_seq_out[seq_id].data(), (n_embd*seq_idx)*sizeof(float), n_embd*sizeof(float)); } } break; case LLAMA_POOLING_TYPE_RANK: { // extract the rerank score - n_cls_out floats per sequence auto & embd_seq_out = embd_seq; + const uint32_t n_cls_out = hparams.n_cls_out; - // TODO: fix indexing [UBATCH_IDX] - for (uint32_t s = 0; s < ubatch.n_seqs; ++s) { - const llama_seq_id seq_id = ubatch.seq_id[s][0]; - if (embd_seq_out.find(seq_id) != embd_seq_out.end()) { - continue; - } + for (uint32_t s = 0; s < ubatch.n_seqs_unq; ++s) { + const llama_seq_id seq_id = ubatch.seq_id_unq[s]; + const int32_t seq_idx = ubatch.seq_idx[seq_id]; + embd_seq_out[seq_id].resize(n_cls_out); - ggml_backend_tensor_get_async(backend_embd, t_embd, embd_seq_out[seq_id].data(), (n_cls_out*seq_id)*sizeof(float), n_cls_out*sizeof(float)); + ggml_backend_tensor_get_async(backend_embd, t_embd, embd_seq_out[seq_id].data(), (n_cls_out*seq_idx)*sizeof(float), n_cls_out*sizeof(float)); } } break; case LLAMA_POOLING_TYPE_UNSPECIFIED: @@ -869,12 +859,16 @@ int llama_context::encode(const llama_batch & batch_inp) { cross.v_embd.resize(cross.n_embd*cross.n_enc); memcpy(cross.v_embd.data(), embd, ggml_nbytes(t_embd)); + const auto & batch = balloc->get_batch(); + // remember the sequence ids used during the encoding - needed for cross attention later cross.seq_ids_enc.resize(n_tokens); for (uint32_t i = 0; i < n_tokens; i++) { cross.seq_ids_enc[i].clear(); + for (int s = 0; s < batch.n_seq_id[i]; s++) { - llama_seq_id seq_id = batch.seq_id[i][s]; + const llama_seq_id seq_id = batch.seq_id[i][s]; + cross.seq_ids_enc[i].insert(seq_id); } } @@ -884,6 +878,8 @@ int llama_context::encode(const llama_batch & batch_inp) { } int llama_context::decode(const llama_batch & batch_inp) { + GGML_ASSERT((!batch_inp.token && batch_inp.embd) || (batch_inp.token && !batch_inp.embd)); // NOLINT + if (!memory) { LLAMA_LOG_DEBUG("%s: cannot decode batches with this context (calling encode() instead)\n", __func__); return encode(batch_inp); @@ -894,29 +890,24 @@ int llama_context::decode(const llama_batch & batch_inp) { return -1; } - // when computing embeddings, all tokens are output - const bool embd_all = cparams.embeddings; - - if (!batch_allocr->init(batch_inp, model.vocab, memory.get(), embd_all)) { - LLAMA_LOG_ERROR("%s: failed to initialize batch\n", __func__); - return -1; - } - - const llama_batch & batch = batch_allocr->get_batch(); - const auto & vocab = model.vocab; const auto & hparams = model.hparams; const int32_t n_vocab = vocab.n_tokens(); const int64_t n_embd = hparams.n_embd; - const uint32_t n_tokens_all = batch.n_tokens; + // when computing embeddings, all tokens are output + const bool output_all = cparams.embeddings; - GGML_ASSERT((!batch.token && batch.embd) || (batch.token && !batch.embd)); // NOLINT + if (!balloc->init(batch_inp, vocab, memory.get(), n_embd, output_all)) { + LLAMA_LOG_ERROR("%s: failed to initialize batch\n", __func__); + return -1; + } - const uint32_t n_outputs_all = batch_allocr->get_n_outputs(); + const uint32_t n_tokens_all = balloc->get_n_tokens(); + const uint32_t n_outputs_all = balloc->get_n_outputs(); - if (embd_all) { + if (output_all) { // require that all tokens are output if (n_outputs_all != n_tokens_all) { LLAMA_LOG_ERROR("%s: pooled embedding requires that all tokens are output (n_outputs_all = %d, n_tokens_all = %d)\n", @@ -945,7 +936,7 @@ int llama_context::decode(const llama_batch & batch_inp) { llama_memory_state_ptr mstate; while (true) { - mstate = memory->init_batch(batch, cparams.n_ubatch, embd_all); + mstate = memory->init_batch(*balloc, cparams.n_ubatch, output_all); if (!mstate) { return -2; } @@ -966,19 +957,19 @@ int llama_context::decode(const llama_batch & batch_inp) { did_optimize = true; if (kv_self_update(true)) { - LLAMA_LOG_DEBUG("%s: retrying batch size %d after cache optimization\n", __func__, batch.n_tokens); + LLAMA_LOG_DEBUG("%s: retrying batch size %d after cache optimization\n", __func__, balloc->get_n_tokens()); continue; } } - LLAMA_LOG_WARN("%s: failed to find a memory slot for batch of size %d\n", __func__, batch.n_tokens); + LLAMA_LOG_WARN("%s: failed to find a memory slot for batch of size %d\n", __func__, balloc->get_n_tokens()); return 1; } case LLAMA_MEMORY_STATUS_FAILED_COMPUTE: { - LLAMA_LOG_ERROR("%s: compute failed while preparing batch of size %d\n", __func__, batch.n_tokens); + LLAMA_LOG_ERROR("%s: compute failed while preparing batch of size %d\n", __func__, balloc->get_n_tokens()); return -2; } @@ -1005,7 +996,6 @@ int llama_context::decode(const llama_batch & batch_inp) { if (n_outputs_all == n_tokens_all) { n_outputs_new = ubatch.n_tokens; } else { - GGML_ASSERT(ubatch.output); for (uint32_t i = 0; i < ubatch.n_tokens; i++) { n_outputs_new += (int32_t) (ubatch.output[i] != 0); } @@ -1105,27 +1095,27 @@ int llama_context::decode(const llama_batch & batch_inp) { // extract sequence embeddings (cleared before processing each batch) auto & embd_seq_out = embd_seq; - for (uint32_t s = 0; s < ubatch.n_seqs; ++s) { - const llama_seq_id seq_id = ubatch.seq_id[s][0]; - if (embd_seq_out.find(seq_id) != embd_seq_out.end()) { - continue; - } + for (uint32_t s = 0; s < ubatch.n_seqs_unq; ++s) { + const llama_seq_id seq_id = ubatch.seq_id_unq[s]; + const int32_t seq_idx = ubatch.seq_idx[seq_id]; + embd_seq_out[seq_id].resize(n_embd); - ggml_backend_tensor_get_async(backend_embd, t_embd, embd_seq_out[seq_id].data(), (n_embd*seq_id)*sizeof(float), n_embd*sizeof(float)); + ggml_backend_tensor_get_async(backend_embd, t_embd, embd_seq_out[seq_id].data(), (n_embd*seq_idx)*sizeof(float), n_embd*sizeof(float)); } } break; case LLAMA_POOLING_TYPE_RANK: { - // extract the rerank score - a single float per sequence + // extract the rerank score - n_cls_out floats per sequence auto & embd_seq_out = embd_seq; - for (uint32_t s = 0; s < ubatch.n_seqs; ++s) { - const llama_seq_id seq_id = ubatch.seq_id[s][0]; - if (embd_seq_out.find(seq_id) != embd_seq_out.end()) { - continue; - } - embd_seq_out[seq_id].resize(1); - ggml_backend_tensor_get_async(backend_embd, t_embd, embd_seq_out[seq_id].data(), (seq_id)*sizeof(float), sizeof(float)); + const uint32_t n_cls_out = hparams.n_cls_out; + + for (uint32_t s = 0; s < ubatch.n_seqs_unq; ++s) { + const llama_seq_id seq_id = ubatch.seq_id_unq[s]; + const int32_t seq_idx = ubatch.seq_idx[seq_id]; + + embd_seq_out[seq_id].resize(n_cls_out); + ggml_backend_tensor_get_async(backend_embd, t_embd, embd_seq_out[seq_id].data(), (n_cls_out*seq_idx)*sizeof(float), n_cls_out*sizeof(float)); } } break; case LLAMA_POOLING_TYPE_UNSPECIFIED: @@ -1145,7 +1135,7 @@ int llama_context::decode(const llama_batch & batch_inp) { if (n_outputs > 0) { bool sorted_output = true; - auto & out_ids = mstate->out_ids(); + auto & out_ids = balloc->get_out_ids(); GGML_ASSERT(out_ids.size() == (size_t) n_outputs); @@ -1318,8 +1308,8 @@ ggml_cgraph * llama_context::graph_reserve(uint32_t n_tokens, uint32_t n_seqs, u this->n_outputs = n_outputs; - llama_token token = model.vocab.token_bos(); // not actually used by llama_build_graph, but required to choose between token and embedding inputs graph - llama_ubatch ubatch = { true, n_tokens, n_tokens / n_seqs, n_seqs, &token, nullptr, nullptr, nullptr, nullptr, nullptr}; + llama_batch_allocr balloc(model.hparams.n_pos_per_embd()); + llama_ubatch ubatch = balloc.ubatch_reserve(n_tokens/n_seqs, n_seqs); auto * gf = graph_init(); auto res = graph_build(ctx_compute.get(), gf, ubatch, LLM_GRAPH_TYPE_DEFAULT, mstate); @@ -2039,7 +2029,12 @@ void llama_context::opt_epoch_iter( batch.logits [pos_batch] = true; } - const auto n_tokens_all = batch.n_tokens; + if (!balloc->init(batch, model.vocab, nullptr, model.hparams.n_embd, true)) { + LLAMA_LOG_ERROR("%s: failed to initialize batch\n", __func__); + return; + } + + const uint32_t n_tokens_all = balloc->get_n_tokens(); n_queued_tokens += n_tokens_all; @@ -2047,7 +2042,7 @@ void llama_context::opt_epoch_iter( uint32_t n_outputs_all = n_tokens_all; - auto mstate = memory->init_batch(batch, cparams.n_ubatch, true); + auto mstate = memory->init_batch(*balloc, cparams.n_ubatch, true); if (!mstate || mstate->get_status() != LLAMA_MEMORY_STATUS_SUCCESS) { LLAMA_LOG_ERROR("%s: could not initialize batch\n", __func__); break; diff --git a/src/llama-context.h b/src/llama-context.h index 040f03ae4..7d300c145 100644 --- a/src/llama-context.h +++ b/src/llama-context.h @@ -247,7 +247,7 @@ private: std::map> embd_seq; // reuse the batch_allocr to avoid unnecessary memory allocations - std::unique_ptr batch_allocr; + std::unique_ptr balloc; uint32_t n_outputs = 0; // number of actually-used outputs in the current ubatch or last logical batch diff --git a/src/llama-graph.cpp b/src/llama-graph.cpp index 65d98cbbb..083366fd6 100644 --- a/src/llama-graph.cpp +++ b/src/llama-graph.cpp @@ -130,110 +130,97 @@ void llm_graph_input_mean::set_input(const llama_ubatch * ubatch) { if (cparams.embeddings && cparams.pooling_type == LLAMA_POOLING_TYPE_MEAN) { const int64_t n_tokens = ubatch->n_tokens; const int64_t n_seq_tokens = ubatch->n_seq_tokens; - const int64_t n_seqs = ubatch->n_seqs; + const int64_t n_seqs_unq = ubatch->n_seqs_unq; GGML_ASSERT(mean); GGML_ASSERT(ggml_backend_buffer_is_host(mean->buffer)); float * data = (float *) mean->data; - memset(mean->data, 0, n_tokens * n_tokens * ggml_element_size(mean)); + memset(mean->data, 0, n_tokens*n_seqs_unq*ggml_element_size(mean)); - std::vector sum(n_tokens, 0); + std::vector sums(n_seqs_unq, 0); + for (int i = 0; i < n_tokens; i += n_seq_tokens) { + for (int s = 0; s < ubatch->n_seq_id[i]; ++s) { + const llama_seq_id seq_id = ubatch->seq_id[i][s]; + const int32_t seq_idx = ubatch->seq_idx[seq_id]; - // TODO: fix indexing [UBATCH_IDX] - for (int s = 0; s < n_seqs; ++s) { - const llama_seq_id seq_id = ubatch->seq_id[s][0]; - - // TODO: adapt limits to n_seqs when ubatch->equal_seqs is true - GGML_ASSERT(seq_id < n_tokens && "seq_id cannot be larger than n_tokens with pooling_type == MEAN"); - - sum[seq_id] += ubatch->n_seq_tokens; - } - - std::vector div(n_tokens, 0.0f); - for (int i = 0; i < n_tokens; ++i) { - const uint64_t s = sum[i]; - if (s > 0) { - div[i] = 1.0f/float(s); + sums[seq_idx] += ubatch->n_seq_tokens; } } - // TODO: fix indexing [UBATCH_IDX] - for (int s = 0; s < n_seqs; ++s) { - const llama_seq_id seq_id = ubatch->seq_id[s][0]; + std::vector div(n_seqs_unq, 0.0f); + for (int s = 0; s < n_seqs_unq; ++s) { + const uint64_t sum = sums[s]; + if (sum > 0) { + div[s] = 1.0f/float(sum); + } + } - for (int i = 0; i < n_seq_tokens; ++i) { - data[seq_id*n_tokens + s*n_seq_tokens + i] = div[seq_id]; + for (int i = 0; i < n_tokens; i += n_seq_tokens) { + for (int s = 0; s < ubatch->n_seq_id[i]; ++s) { + const llama_seq_id seq_id = ubatch->seq_id[i][s]; + const int32_t seq_idx = ubatch->seq_idx[seq_id]; + + for (int j = 0; j < n_seq_tokens; ++j) { + data[seq_idx*n_tokens + i + j] = div[seq_idx]; + } } } } } void llm_graph_input_cls::set_input(const llama_ubatch * ubatch) { - if (cparams.embeddings && ( - cparams.pooling_type == LLAMA_POOLING_TYPE_CLS || - cparams.pooling_type == LLAMA_POOLING_TYPE_RANK)) { - const int64_t n_tokens = ubatch->n_tokens; - const int64_t n_seq_tokens = ubatch->n_seq_tokens; - const int64_t n_seqs = ubatch->n_seqs; + const int64_t n_tokens = ubatch->n_tokens; + const int64_t n_seq_tokens = ubatch->n_seq_tokens; + const int64_t n_seqs_unq = ubatch->n_seqs_unq; + if (cparams.embeddings && ( + cparams.pooling_type == LLAMA_POOLING_TYPE_CLS || + cparams.pooling_type == LLAMA_POOLING_TYPE_RANK + )) { GGML_ASSERT(cls); GGML_ASSERT(ggml_backend_buffer_is_host(cls->buffer)); uint32_t * data = (uint32_t *) cls->data; - memset(cls->data, 0, n_tokens * ggml_element_size(cls)); + memset(cls->data, 0, n_seqs_unq*ggml_element_size(cls)); - // TODO: fix indexing [UBATCH_IDX] - for (int s = 0; s < n_seqs; ++s) { - const llama_seq_id seq_id = ubatch->seq_id[s][0]; + for (int i = 0; i < n_tokens; i += n_seq_tokens) { + for (int s = 0; s < ubatch->n_seq_id[i]; ++s) { + const llama_seq_id seq_id = ubatch->seq_id[i][s]; + const int32_t seq_idx = ubatch->seq_idx[seq_id]; - // TODO: adapt limits to n_seqs when ubatch->equal_seqs is true - GGML_ASSERT(seq_id < n_tokens && "seq_id cannot be larger than n_tokens with pooling_type == CLS or RANK"); - - for (int i = 0; i < n_seq_tokens; ++i) { - const llama_pos pos = ubatch->pos[s*n_seq_tokens + i]; - - if (pos == 0) { - data[seq_id] = s*n_seq_tokens + i; - } + data[seq_idx] = i; } } } if (cparams.embeddings && cparams.pooling_type == LLAMA_POOLING_TYPE_LAST) { - const int64_t n_tokens = ubatch->n_tokens; - const int64_t n_seq_tokens = ubatch->n_seq_tokens; - const int64_t n_seqs = ubatch->n_seqs; - GGML_ASSERT(cls); GGML_ASSERT(ggml_backend_buffer_is_host(cls->buffer)); uint32_t * data = (uint32_t *) cls->data; - memset(cls->data, 0, n_tokens * ggml_element_size(cls)); + memset(cls->data, 0, n_seqs_unq*ggml_element_size(cls)); - std::vector last_pos(n_tokens, -1); - std::vector last_row(n_tokens, -1); + std::vector last_pos(n_seqs_unq, -1); + std::vector last_row(n_seqs_unq, -1); - // TODO: fix indexing [UBATCH_IDX] - for (int s = 0; s < n_seqs; ++s) { - const llama_seq_id seq_id = ubatch->seq_id[s][0]; + for (int i = 0; i < n_tokens; ++i) { + const llama_pos pos = ubatch->pos[i]; - // TODO: adapt limits to n_seqs when ubatch->equal_seqs is true - GGML_ASSERT(seq_id < n_tokens && "seq_id cannot be larger than n_tokens with pooling_type == LAST"); + for (int s = 0; s < ubatch->n_seq_id[i]; ++s) { + const llama_seq_id seq_id = ubatch->seq_id[i][s]; + const int32_t seq_idx = ubatch->seq_idx[seq_id]; - for (int i = 0; i < n_seq_tokens; ++i) { - const llama_pos pos = ubatch->pos[s*n_seq_tokens + i]; - - if (pos >= last_pos[seq_id]) { - last_pos[seq_id] = pos; - last_row[seq_id] = s*n_seq_tokens + i; + if (pos >= last_pos[seq_idx]) { + last_pos[seq_idx] = pos; + last_row[seq_idx] = i; } } } - for (int i = 0; i < n_tokens; ++i) { - if (last_row[i] >= 0) { - data[i] = last_row[i]; + for (int s = 0; s < n_seqs_unq; ++s) { + if (last_row[s] >= 0) { + data[s] = last_row[s]; } } } @@ -266,89 +253,36 @@ void llm_graph_input_cross_embd::set_input(const llama_ubatch * ubatch) { } void llm_graph_input_attn_no_cache::set_input(const llama_ubatch * ubatch) { - if (kq_mask) { - if (cparams.causal_attn) { - const int64_t n_kv = ubatch->n_tokens; - const int64_t n_tokens = ubatch->n_tokens; - const int64_t n_seq_tokens = ubatch->n_seq_tokens; - const int64_t n_seqs = ubatch->n_seqs; + const int64_t n_kv = ubatch->n_tokens; + const int64_t n_tokens = ubatch->n_tokens; - GGML_ASSERT(ggml_backend_buffer_is_host(kq_mask->buffer)); - float * data = (float *) kq_mask->data; + GGML_ASSERT(kq_mask); + GGML_ASSERT(ggml_backend_buffer_is_host(kq_mask->buffer)); - for (int h = 0; h < 1; ++h) { - for (int s1 = 0; s1 < n_seqs; ++s1) { - const llama_seq_id seq_id = ubatch->seq_id[s1][0]; + float * data = (float *) kq_mask->data; - for (int j = 0; j < n_seq_tokens; ++j) { - const int32_t tj = s1*n_seq_tokens + j; + for (int h = 0; h < 1; ++h) { + for (int i1 = 0; i1 < n_tokens; ++i1) { + const llama_seq_id s1 = ubatch->seq_id[i1][0]; - for (int s0 = 0; s0 < n_seqs; ++s0) { - for (int i = 0; i < n_seq_tokens; ++i) { - const int32_t ti = s0*n_seq_tokens + i; - float f = -INFINITY; + for (int i0 = 0; i0 < n_tokens; ++i0) { + float f = -INFINITY; - // TODO: fix indexing [UBATCH_IDX] - for (int s = 0; s < ubatch->n_seq_id[s0]; ++s) { - if (ubatch->seq_id[s0][s] == seq_id && ubatch->pos[ti] <= ubatch->pos[tj]) { - if (hparams.use_alibi) { - f = -std::abs(ubatch->pos[ti] - ubatch->pos[tj]); - } else { - f = 0.0f; - } - break; - } - } + for (int s = 0; s < ubatch->n_seq_id[i0]; ++s) { + const llama_seq_id s0 = ubatch->seq_id[i0][0]; - data[h*(n_kv*n_tokens) + tj*n_kv + ti] = f; - } + // TODO: reimplement this like in llama_kv_cache_unified + if (s0 == s1 && (!cparams.causal_attn || ubatch->pos[i0] <= ubatch->pos[i1])) { + if (hparams.use_alibi) { + f = -std::abs(ubatch->pos[i0] - ubatch->pos[i1]); + } else { + f = 0.0f; } + break; } } - } - } else { - const int64_t n_tokens = ubatch->n_tokens; - const int64_t n_seq_tokens = ubatch->n_seq_tokens; - const int64_t n_seqs = ubatch->n_seqs; - const int64_t n_stride = ubatch->n_tokens; - GGML_ASSERT(ggml_backend_buffer_is_host(kq_mask->buffer)); - - float * data = (float *) kq_mask->data; - - for (int h = 0; h < 1; ++h) { - for (int s1 = 0; s1 < n_seqs; ++s1) { - const llama_seq_id seq_id = ubatch->seq_id[s1][0]; - - for (int j = 0; j < n_seq_tokens; ++j) { - const int32_t tj = s1*n_seq_tokens + j; - - for (int s0 = 0; s0 < n_seqs; ++s0) { - for (int i = 0; i < n_seq_tokens; ++i) { - const int32_t ti = s0*n_seq_tokens + i; - float f = -INFINITY; - - // TODO: fix indexing [UBATCH_IDX] - for (int s = 0; s < ubatch->n_seq_id[s0]; ++s) { - if (ubatch->seq_id[s0][s] == seq_id) { - if (hparams.use_alibi) { - f = -std::abs(ubatch->pos[ti] - ubatch->pos[tj]); - } else { - f = 0.0f; - } - break; - } - } - - data[h*(n_tokens*n_tokens) + tj*n_stride + ti] = f; - } - } - - for (int i = n_tokens; i < n_stride; ++i) { - data[h*(n_tokens*n_tokens) + tj*n_stride + i] = -INFINITY; - } - } - } + data[h*(n_kv*n_tokens) + i1*n_kv + i0] = f; } } } @@ -371,34 +305,36 @@ void llm_graph_input_attn_kv_unified_iswa::set_input(const llama_ubatch * ubatch } void llm_graph_input_attn_cross::set_input(const llama_ubatch * ubatch) { - if (cross_kq_mask) { - const int64_t n_enc = cross_kq_mask->ne[0]; - const int64_t n_tokens = ubatch->n_tokens; + GGML_ASSERT(cross_kq_mask); - GGML_ASSERT(ggml_backend_buffer_is_host(cross_kq_mask->buffer)); - GGML_ASSERT(!ubatch->equal_seqs); // TODO: use ubatch->n_seqs instead of failing + const int64_t n_enc = cross_kq_mask->ne[0]; + const int64_t n_tokens = ubatch->n_tokens; - float * data = (float *) cross_kq_mask->data; + GGML_ASSERT(ggml_backend_buffer_is_host(cross_kq_mask->buffer)); + GGML_ASSERT(!ubatch->equal_seqs); // TODO: use ubatch->n_seqs instead of failing - for (int h = 0; h < 1; ++h) { - for (int j = 0; j < n_tokens; ++j) { - for (int i = 0; i < n_enc; ++i) { - float f = -INFINITY; - // TODO: fix indexing [UBATCH_IDX] - for (int s = 0; s < ubatch->n_seq_id[j]; ++s) { - const llama_seq_id seq_id = ubatch->seq_id[j][s]; - if (cross->seq_ids_enc[i].find(seq_id) != cross->seq_ids_enc[i].end()) { - f = 0.0f; - } + float * data = (float *) cross_kq_mask->data; + + for (int h = 0; h < 1; ++h) { + for (int i = 0; i < n_tokens; ++i) { + for (int j = 0; j < n_enc; ++j) { + float f = -INFINITY; + + for (int s = 0; s < ubatch->n_seq_id[i]; ++s) { + const llama_seq_id seq_id = ubatch->seq_id[i][s]; + + if (cross->seq_ids_enc[j].find(seq_id) != cross->seq_ids_enc[j].end()) { + f = 0.0f; } - data[h*(n_enc*n_tokens) + j*n_enc + i] = f; } - } - for (int i = n_tokens; i < GGML_PAD(n_tokens, GGML_KQ_MASK_PAD); ++i) { - for (int j = 0; j < n_enc; ++j) { - data[h*(n_enc*n_tokens) + i*n_enc + j] = -INFINITY; - } + data[h*(n_enc*n_tokens) + i*n_enc + j] = f; + } + } + + for (int i = n_tokens; i < GGML_PAD(n_tokens, GGML_KQ_MASK_PAD); ++i) { + for (int j = 0; j < n_enc; ++j) { + data[h*(n_enc*n_tokens) + i*n_enc + j] = -INFINITY; } } } @@ -467,10 +403,6 @@ llm_graph_context::llm_graph_context(const llm_graph_params & params) : res (std::make_unique()) { } -int64_t llm_graph_context::n_pos_per_embd() const { - return hparams.rope_type == LLAMA_ROPE_TYPE_MROPE ? 4 : 1; -} - void llm_graph_context::cb(ggml_tensor * cur, const char * name, int il) const { if (cb_func) { cb_func(ubatch, cur, name, il); @@ -915,11 +847,11 @@ ggml_tensor * llm_graph_context::build_inp_embd(ggml_tensor * tok_embd) const { } ggml_tensor * llm_graph_context::build_inp_pos() const { - auto inp = std::make_unique(n_pos_per_embd()); + auto inp = std::make_unique(hparams.n_pos_per_embd()); auto & cur = inp->pos; - cur = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens*n_pos_per_embd()); + cur = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, (int64_t)n_tokens*hparams.n_pos_per_embd()); ggml_set_input(cur); res->add_input(std::move(inp)); @@ -959,7 +891,7 @@ ggml_tensor * llm_graph_context::build_inp_mean() const { auto & cur = inp->mean; - cur = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_tokens, n_tokens); + cur = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_tokens, ubatch.n_seqs_unq); ggml_set_input(cur); res->add_input(std::move(inp)); @@ -972,7 +904,7 @@ ggml_tensor * llm_graph_context::build_inp_cls() const { auto & cur = inp->cls; - cur = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); + cur = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, ubatch.n_seqs_unq); ggml_set_input(cur); res->add_input(std::move(inp)); diff --git a/src/llama-graph.h b/src/llama-graph.h index 58845e284..9e62fa607 100644 --- a/src/llama-graph.h +++ b/src/llama-graph.h @@ -95,14 +95,14 @@ public: class llm_graph_input_pos : public llm_graph_input_i { public: - llm_graph_input_pos(int64_t n_pos_per_embd) : n_pos_per_embd(n_pos_per_embd) {} + llm_graph_input_pos(uint32_t n_pos_per_embd) : n_pos_per_embd(n_pos_per_embd) {} virtual ~llm_graph_input_pos() = default; void set_input(const llama_ubatch * ubatch) override; ggml_tensor * pos = nullptr; // I32 [n_batch] - const int64_t n_pos_per_embd = 1; + const uint32_t n_pos_per_embd = 1; }; // temperature tuning, used by llama4 @@ -464,8 +464,6 @@ struct llm_graph_context { llm_graph_context(const llm_graph_params & params); - int64_t n_pos_per_embd() const; - void cb(ggml_tensor * cur, const char * name, int il) const; // diff --git a/src/llama-hparams.cpp b/src/llama-hparams.cpp index b40566ced..bba7a12dc 100644 --- a/src/llama-hparams.cpp +++ b/src/llama-hparams.cpp @@ -90,6 +90,10 @@ bool llama_hparams::is_recurrent(uint32_t il) const { return recurrent_layer_arr[il]; } +uint32_t llama_hparams::n_pos_per_embd() const { + return rope_type == LLAMA_ROPE_TYPE_MROPE ? 4 : 1; +} + bool llama_hparams::is_swa(uint32_t il) const { if (il < n_layer) { return swa_layers[il]; diff --git a/src/llama-hparams.h b/src/llama-hparams.h index 82bb5b608..7b315a9a7 100644 --- a/src/llama-hparams.h +++ b/src/llama-hparams.h @@ -192,6 +192,8 @@ struct llama_hparams { // whether or not the given layer is recurrent (for hybrid models) bool is_recurrent(uint32_t il) const; + uint32_t n_pos_per_embd() const; + bool is_swa(uint32_t il) const; }; diff --git a/src/llama-kv-cache-unified-iswa.cpp b/src/llama-kv-cache-unified-iswa.cpp index a869b1de8..0ced340de 100644 --- a/src/llama-kv-cache-unified-iswa.cpp +++ b/src/llama-kv-cache-unified-iswa.cpp @@ -95,19 +95,22 @@ llama_pos llama_kv_cache_unified_iswa::seq_pos_max(llama_seq_id seq_id) const { return kv_swa->seq_pos_max(seq_id); } -llama_memory_state_ptr llama_kv_cache_unified_iswa::init_batch(const llama_batch & batch, uint32_t n_ubatch, bool embd_all) { +llama_memory_state_ptr llama_kv_cache_unified_iswa::init_batch(llama_batch_allocr & balloc, uint32_t n_ubatch, bool embd_all) { GGML_UNUSED(embd_all); // first try simple split do { - auto sbatch = llama_sbatch(batch, hparams.n_embd, true); + balloc.split_reset(); std::vector ubatches; + while (true) { + auto ubatch = balloc.split_simple(n_ubatch); - while (sbatch.n_tokens > 0) { - auto ubatch = sbatch.split_simple(n_ubatch); + if (ubatch.n_tokens == 0) { + break; + } - ubatches.push_back(ubatch); + ubatches.push_back(std::move(ubatch)); // NOLINT } auto heads_base = kv_base->prepare(ubatches); @@ -123,19 +126,22 @@ llama_memory_state_ptr llama_kv_cache_unified_iswa::init_batch(const llama_batch assert(heads_base.size() == heads_swa.size()); return std::make_unique( - this, std::move(sbatch), std::move(heads_base), std::move(heads_swa), std::move(ubatches)); + this, std::move(heads_base), std::move(heads_swa), std::move(ubatches)); } while (false); // if it fails, try equal split do { - auto sbatch = llama_sbatch(batch, hparams.n_embd, false); + balloc.split_reset(); std::vector ubatches; + while (true) { + auto ubatch = balloc.split_equal(n_ubatch); - while (sbatch.n_tokens > 0) { - auto ubatch = sbatch.split_equal(n_ubatch); + if (ubatch.n_tokens == 0) { + break; + } - ubatches.push_back(ubatch); + ubatches.push_back(std::move(ubatch)); // NOLINT } auto heads_base = kv_base->prepare(ubatches); @@ -151,7 +157,7 @@ llama_memory_state_ptr llama_kv_cache_unified_iswa::init_batch(const llama_batch assert(heads_base.size() == heads_swa.size()); return std::make_unique( - this, std::move(sbatch), std::move(heads_base), std::move(heads_swa), std::move(ubatches)); + this, std::move(heads_base), std::move(heads_swa), std::move(ubatches)); } while (false); // TODO: if we fail again, we should attempt different splitting strategies @@ -214,15 +220,13 @@ llama_kv_cache_unified_iswa_state::llama_kv_cache_unified_iswa_state( llama_kv_cache_unified_iswa_state::llama_kv_cache_unified_iswa_state( llama_kv_cache_unified_iswa * kv, - llama_sbatch sbatch, std::vector heads_base, std::vector heads_swa, std::vector ubatches) : - sbatch(std::move(sbatch)), ubatches(std::move(ubatches)), // note: here we copy the ubatches. not sure if this is ideal - state_base(new llama_kv_cache_unified_state(kv->get_base(), {}, std::move(heads_base), this->ubatches)), - state_swa (new llama_kv_cache_unified_state(kv->get_swa (), {}, std::move(heads_swa), this->ubatches)), + state_base(new llama_kv_cache_unified_state(kv->get_base(), std::move(heads_base), this->ubatches)), + state_swa (new llama_kv_cache_unified_state(kv->get_swa (), std::move(heads_swa), this->ubatches)), status(llama_memory_status_combine(state_base->get_status(), state_swa->get_status())) { } @@ -252,12 +256,6 @@ bool llama_kv_cache_unified_iswa_state::apply() { return res; } -std::vector & llama_kv_cache_unified_iswa_state::out_ids() { - assert(status == LLAMA_MEMORY_STATUS_SUCCESS); - - return sbatch.out_ids; -} - llama_memory_status llama_kv_cache_unified_iswa_state::get_status() const { return status; } diff --git a/src/llama-kv-cache-unified-iswa.h b/src/llama-kv-cache-unified-iswa.h index 813eaf39b..071041585 100644 --- a/src/llama-kv-cache-unified-iswa.h +++ b/src/llama-kv-cache-unified-iswa.h @@ -32,7 +32,7 @@ public: // llama_memory_state_ptr init_batch( - const llama_batch & batch, + llama_batch_allocr & balloc, uint32_t n_ubatch, bool embd_all) override; @@ -90,7 +90,6 @@ public: // used to create a state from a batch llama_kv_cache_unified_iswa_state( llama_kv_cache_unified_iswa * kv, - llama_sbatch sbatch, std::vector heads_base, std::vector heads_swa, std::vector ubatches); @@ -104,8 +103,6 @@ public: bool next() override; bool apply() override; - std::vector & out_ids() override; - llama_memory_status get_status() const override; const llama_ubatch & get_ubatch() const override; @@ -119,8 +116,6 @@ public: private: //llama_kv_cache_unified_iswa * kv; - llama_sbatch sbatch; - // the index of the next ubatch to process size_t i_next = 0; diff --git a/src/llama-kv-cache-unified.cpp b/src/llama-kv-cache-unified.cpp index d44122889..6897b7971 100644 --- a/src/llama-kv-cache-unified.cpp +++ b/src/llama-kv-cache-unified.cpp @@ -308,17 +308,23 @@ llama_pos llama_kv_cache_unified::seq_pos_max(llama_seq_id seq_id) const { } llama_memory_state_ptr llama_kv_cache_unified::init_batch( - const llama_batch & batch, + llama_batch_allocr & balloc, uint32_t n_ubatch, bool embd_all) { GGML_UNUSED(embd_all); do { - auto sbatch = llama_sbatch(batch, hparams.n_embd, true); + balloc.split_reset(); std::vector ubatches; - while (sbatch.n_tokens > 0) { - ubatches.push_back(sbatch.split_simple(n_ubatch)); + while (true) { + auto ubatch = balloc.split_simple(n_ubatch); + + if (ubatch.n_tokens == 0) { + break; + } + + ubatches.push_back(std::move(ubatch)); // NOLINT } auto heads = prepare(ubatches); @@ -327,7 +333,7 @@ llama_memory_state_ptr llama_kv_cache_unified::init_batch( } return std::make_unique( - this, std::move(sbatch), std::move(heads), std::move(ubatches)); + this, std::move(heads), std::move(ubatches)); } while (false); return std::make_unique(LLAMA_MEMORY_STATUS_FAILED_PREPARE); @@ -644,12 +650,6 @@ int32_t llama_kv_cache_unified::find_slot(const llama_ubatch & ubatch) const { } void llama_kv_cache_unified::apply_ubatch(uint32_t head_cur, const llama_ubatch & ubatch) { - if (debug > 0) { - LLAMA_LOG_DEBUG("%s: ubatch info:\n", __func__); - LLAMA_LOG_DEBUG("%s: n_tokens = %d, equal_seqs = %d\n", __func__, ubatch.n_tokens, ubatch.equal_seqs); - LLAMA_LOG_DEBUG("%s: n_seq_tokens = %d, n_seqs = %d\n", __func__, ubatch.n_seq_tokens, ubatch.n_seqs); - } - // keep track of the max sequence position that we would overwrite with this ubatch // for non-SWA cache, this would be always empty llama_seq_id seq_pos_max_rm[LLAMA_MAX_SEQ]; @@ -657,27 +657,22 @@ void llama_kv_cache_unified::apply_ubatch(uint32_t head_cur, const llama_ubatch seq_pos_max_rm[s] = -1; } - for (uint32_t s = 0; s < ubatch.n_seqs; ++s) { - for (uint32_t j = 0; j < ubatch.n_seq_tokens; ++j) { - const uint32_t idx = s*ubatch.n_seq_tokens + j; + for (uint32_t i = 0; i < ubatch.n_tokens; ++i) { + if (!cells.is_empty(head_cur + i)) { + assert(cells.seq_count(head_cur + i) == 1); - if (!cells.is_empty(head_cur + idx)) { - assert(cells.seq_count(head_cur + idx) == 1); + const llama_seq_id seq_id = cells.seq_get(head_cur + i); + const llama_pos pos = cells.pos_get(head_cur + i); - const llama_seq_id seq_id = cells.seq_get(head_cur + idx); - const llama_pos pos = cells.pos_get(head_cur + idx); + seq_pos_max_rm[seq_id] = std::max(seq_pos_max_rm[seq_id], pos); - seq_pos_max_rm[seq_id] = std::max(seq_pos_max_rm[seq_id], pos); + cells.rm(head_cur + i); + } - cells.rm(head_cur + idx); - } + cells.pos_set(head_cur + i, ubatch.pos[i]); - cells.pos_set(head_cur + idx, ubatch.pos[idx]); - - // TODO: fix indexing [UBATCH_IDX] - for (int32_t i = 0; i < ubatch.n_seq_id[s]; i++) { - cells.seq_add(head_cur + idx, ubatch.seq_id[s][i]); - } + for (int32_t s = 0; s < ubatch.n_seq_id[i]; s++) { + cells.seq_add(head_cur + i, ubatch.seq_id[i][s]); } } @@ -696,6 +691,7 @@ void llama_kv_cache_unified::apply_ubatch(uint32_t head_cur, const llama_ubatch seq_rm(s, cells.seq_pos_min(s), seq_pos_max_rm[s] + 1); } } + // move the head at the end of the slot head = head_cur + ubatch.n_tokens; } @@ -792,9 +788,7 @@ ggml_tensor * llama_kv_cache_unified::cpy_v(ggml_context * ctx, ggml_tensor * v_ } void llama_kv_cache_unified::set_input_kq_mask(ggml_tensor * dst, const llama_ubatch * ubatch, bool causal_attn) const { - const uint32_t n_tokens = ubatch->n_tokens; - const uint32_t n_seq_tokens = ubatch->n_seq_tokens; - const uint32_t n_seqs = ubatch->n_seqs; + const uint32_t n_tokens = ubatch->n_tokens; GGML_ASSERT(ggml_backend_buffer_is_host(dst->buffer)); float * data = (float *) dst->data; @@ -814,52 +808,48 @@ void llama_kv_cache_unified::set_input_kq_mask(ggml_tensor * dst, const llama_ub // xxxxx----- // To visualize the mask, see https://github.com/ggml-org/llama.cpp/pull/12615 for (uint32_t h = 0; h < 1; ++h) { - for (uint32_t s = 0; s < n_seqs; ++s) { - const llama_seq_id seq_id = ubatch->seq_id[s][0]; + for (uint32_t i = 0; i < n_tokens; ++i) { + const llama_seq_id seq_id = ubatch->seq_id[i][0]; - for (uint32_t j = 0; j < n_seq_tokens; ++j) { - const uint32_t idx = s*n_seq_tokens + j; + const llama_pos p1 = ubatch->pos[i]; - const llama_pos p1 = ubatch->pos[idx]; + for (uint32_t j = 0; j < n_kv; ++j) { + float f = 0.0f; - for (uint32_t i = 0; i < n_kv; ++i) { - float f = 0.0f; + bool masked = false; - bool masked = false; + if (cells.is_empty(j)) { + masked = true; + } else { + const llama_pos p0 = cells.pos_get(j); - if (cells.is_empty(i)) { - masked = true; - } else { - const llama_pos p0 = cells.pos_get(i); + // mask the token if not the same sequence + masked = masked || (!cells.seq_has(j, seq_id)); - // mask the token if not the same sequence - masked = masked || (!cells.seq_has(i, seq_id)); + // mask future tokens + masked = masked || (causal_attn && p0 > p1); - // mask future tokens - masked = masked || (causal_attn && p0 > p1); + // apply SWA if any + masked = masked || (is_masked_swa(p0, p1)); - // apply SWA if any - masked = masked || (is_masked_swa(p0, p1)); - - if (!masked && hparams.use_alibi) { - f = -std::abs(p0 - p1); - } + if (!masked && hparams.use_alibi) { + f = -std::abs(p0 - p1); } - - if (masked) { - f = -INFINITY; - } - - data[h*(n_kv*n_tokens) + idx*n_kv + i] = f; } + + if (masked) { + f = -INFINITY; + } + + data[h*(n_kv*n_tokens) + i*n_kv + j] = f; } } // mask padded tokens if (data) { - for (uint32_t j = n_tokens; j < GGML_PAD(n_tokens, GGML_KQ_MASK_PAD); ++j) { - for (uint32_t i = 0; i < n_kv; ++i) { - data[h*(n_kv*n_tokens) + j*n_kv + i] = -INFINITY; + for (uint32_t i = n_tokens; i < GGML_PAD(n_tokens, GGML_KQ_MASK_PAD); ++i) { + for (uint32_t j = 0; j < n_kv; ++j) { + data[h*(n_kv*n_tokens) + i*n_kv + j] = -INFINITY; } } } @@ -887,12 +877,12 @@ void llama_kv_cache_unified::set_input_pos_bucket(ggml_tensor * dst, const llama const int32_t n_kv = dst->ne[0]; for (int h = 0; h < 1; ++h) { - for (int j = 0; j < n_tokens; ++j) { - for (int i = 0; i < n_kv; ++i) { + for (int i = 0; i < n_tokens; ++i) { + for (int j = 0; j < n_kv; ++j) { // the position when the cells is empty is irrelevant - it will be masked out later in the attention - const llama_pos p0 = cells.is_empty(i) ? -1 : cells.pos_get(i); + const llama_pos p0 = cells.is_empty(j) ? -1 : cells.pos_get(j); - data[h*(n_kv*n_tokens) + j*n_kv + i] = llama_relative_position_bucket(p0, ubatch->pos[j], hparams.n_rel_attn_bkts, false); + data[h*(n_kv*n_tokens) + i*n_kv + j] = llama_relative_position_bucket(p0, ubatch->pos[i], hparams.n_rel_attn_bkts, false); } } } @@ -1509,12 +1499,9 @@ bool llama_kv_cache_unified::state_read_meta(llama_io_read_i & io, uint32_t cell seq_rm(dest_seq_id, -1, -1); - llama_sbatch sbatch; - llama_ubatch ubatch = sbatch.reserve_ubatch(cell_count, /* has_embd */ false); + llama_batch_allocr balloc(hparams.n_pos_per_embd()); - ubatch.n_tokens = cell_count; - ubatch.n_seq_tokens = cell_count; - ubatch.n_seqs = 1; + llama_ubatch ubatch = balloc.ubatch_reserve(cell_count, 1); for (uint32_t i = 0; i < cell_count; ++i) { llama_pos pos; @@ -1746,9 +1733,8 @@ llama_kv_cache_unified_state::llama_kv_cache_unified_state( llama_kv_cache_unified_state::llama_kv_cache_unified_state( llama_kv_cache_unified * kv, - llama_sbatch sbatch, llama_kv_cache_unified::ubatch_heads heads, - std::vector ubatches) : status(LLAMA_MEMORY_STATUS_SUCCESS), kv(kv), sbatch(std::move(sbatch)), heads(std::move(heads)), ubatches(std::move(ubatches)) { + std::vector ubatches) : status(LLAMA_MEMORY_STATUS_SUCCESS), kv(kv), heads(std::move(heads)), ubatches(std::move(ubatches)) { } llama_kv_cache_unified_state::~llama_kv_cache_unified_state() = default; @@ -1781,12 +1767,6 @@ bool llama_kv_cache_unified_state::apply() { return true; } -std::vector & llama_kv_cache_unified_state::out_ids() { - assert(status == LLAMA_MEMORY_STATUS_SUCCESS); - - return sbatch.out_ids; -} - llama_memory_status llama_kv_cache_unified_state::get_status() const { return status; } diff --git a/src/llama-kv-cache-unified.h b/src/llama-kv-cache-unified.h index d96571d95..156064004 100644 --- a/src/llama-kv-cache-unified.h +++ b/src/llama-kv-cache-unified.h @@ -57,7 +57,7 @@ public: // llama_memory_state_ptr init_batch( - const llama_batch & batch, + llama_batch_allocr & balloc, uint32_t n_ubatch, bool embd_all) override; @@ -231,7 +231,6 @@ public: // used to create a decode state from a batch llama_kv_cache_unified_state( llama_kv_cache_unified * kv, - llama_sbatch sbatch, ubatch_heads heads, std::vector ubatches); @@ -244,8 +243,6 @@ public: bool next() override; bool apply() override; - std::vector & out_ids() override; - llama_memory_status get_status() const override; const llama_ubatch & get_ubatch() const override; @@ -286,8 +283,6 @@ private: // batch processing state // - llama_sbatch sbatch; - // the index of the next ubatch to process size_t i_next = 0; diff --git a/src/llama-kv-cells.h b/src/llama-kv-cells.h index 1d4e70f4d..349e9032e 100644 --- a/src/llama-kv-cells.h +++ b/src/llama-kv-cells.h @@ -384,10 +384,10 @@ private: // std::vector shift; - using bits_t = std::bitset; + using seq_set_t = std::bitset; // the bitset seq[i] tells us which sequences are currently occupying the i-th cell - std::vector seq; + std::vector seq; // the set seq_pos[s] tells us which positions are currently present for sequence s // this way seq_pos[s].begin() and seq_pos[s].rbegin() give us the min/max positions currently in the cache diff --git a/src/llama-memory-hybrid.cpp b/src/llama-memory-hybrid.cpp index d4b260db4..1b1668681 100644 --- a/src/llama-memory-hybrid.cpp +++ b/src/llama-memory-hybrid.cpp @@ -32,7 +32,7 @@ llama_memory_hybrid::llama_memory_hybrid( mem_attn(new llama_kv_cache_unified( model, filter_attn == nullptr ? - [&](int32_t il) { return !model.hparams.is_recurrent(il); } + [&](int32_t il) { return !hparams.is_recurrent(il); } : filter_attn, type_k, type_v, @@ -47,7 +47,7 @@ llama_memory_hybrid::llama_memory_hybrid( mem_recr(new llama_memory_recurrent( model, filter_recr == nullptr ? - [&](int32_t il) { return model.hparams.is_recurrent(il); } + [&](int32_t il) { return hparams.is_recurrent(il); } : filter_recr, type_r, type_s, @@ -56,42 +56,49 @@ llama_memory_hybrid::llama_memory_hybrid( n_seq_max )) {} -llama_memory_state_ptr llama_memory_hybrid::init_batch(const llama_batch & batch, uint32_t n_ubatch, bool embd_pooled) { +llama_memory_state_ptr llama_memory_hybrid::init_batch(llama_batch_allocr & balloc, uint32_t n_ubatch, bool embd_all) { + do { + balloc.split_reset(); - // since this includes a recurrent cache, we cannot use split_simple - auto sbatch = llama_sbatch(batch, hparams.n_embd, false); + // follow the recurrent pattern for creating the ubatch splits + std::vector ubatches; - // follow the recurrent pattern for creating the ubatch splits - std::vector ubatches; - while (sbatch.n_tokens > 0) { - llama_ubatch ubatch; + while (true) { + llama_ubatch ubatch; - if (embd_pooled) { - // Pooled embeddings cannot be split across ubatches (yet) - ubatch = sbatch.split_seq(n_ubatch); - } else { - ubatch = sbatch.split_equal(n_ubatch); + if (embd_all) { + // if all tokens are output, split by sequence + ubatch = balloc.split_seq(n_ubatch); + } else { + ubatch = balloc.split_equal(n_ubatch); + } + + if (ubatch.n_tokens == 0) { + break; + } + + ubatches.push_back(std::move(ubatch)); // NOLINT } - ubatches.push_back(ubatch); - } + // prepare the recurrent batches first + if (!mem_recr->prepare(ubatches)) { + // TODO: will the recurrent cache be in an undefined state at this point? + LLAMA_LOG_ERROR("%s: failed to prepare recurrent ubatches\n", __func__); + return std::make_unique(LLAMA_MEMORY_STATUS_FAILED_PREPARE); + } - // prepare the recurrent batches first - if (!mem_recr->prepare(ubatches)) { - // TODO: will the recurrent cache be in an undefined state at this point? - LLAMA_LOG_ERROR("%s: failed to prepare recurrent ubatches\n", __func__); - return std::make_unique(LLAMA_MEMORY_STATUS_FAILED_PREPARE); - } + // prepare the attention cache + auto heads_attn = mem_attn->prepare(ubatches); + if (heads_attn.empty()) { + LLAMA_LOG_ERROR("%s: failed to prepare attention ubatches\n", __func__); + return std::make_unique(LLAMA_MEMORY_STATUS_FAILED_PREPARE); + } - // prepare the attention cache - auto heads_attn = mem_attn->prepare(ubatches); - if (heads_attn.empty()) { - LLAMA_LOG_ERROR("%s: failed to prepare attention ubatches\n", __func__); - return std::make_unique(LLAMA_MEMORY_STATUS_FAILED_PREPARE); - } + return std::make_unique( + this, std::move(heads_attn), std::move(ubatches)); + } while(false); - return std::make_unique( - this, std::move(sbatch), std::move(heads_attn), std::move(ubatches)); + return std::make_unique(LLAMA_MEMORY_STATUS_FAILED_PREPARE); } llama_memory_state_ptr llama_memory_hybrid::init_full() { @@ -188,15 +195,13 @@ llama_memory_hybrid_state::llama_memory_hybrid_state( llama_memory_hybrid_state::llama_memory_hybrid_state( llama_memory_hybrid * mem, - llama_sbatch sbatch, std::vector heads_attn, std::vector ubatches) : - sbatch(std::move(sbatch)), ubatches(std::move(ubatches)), // note: here we copy the ubatches. not sure if this is ideal - state_attn(new llama_kv_cache_unified_state(mem->get_mem_attn(), {}, std::move(heads_attn), this->ubatches)), - state_recr(new llama_memory_recurrent_state(mem->get_mem_recr(), {}, this->ubatches)), - status(LLAMA_MEMORY_STATUS_SUCCESS) { + state_attn(new llama_kv_cache_unified_state(mem->get_mem_attn(), std::move(heads_attn), this->ubatches)), + state_recr(new llama_memory_recurrent_state(mem->get_mem_recr(), this->ubatches)), + status(llama_memory_status_combine(state_attn->get_status(), state_recr->get_status())) { } bool llama_memory_hybrid_state::next() { @@ -223,12 +228,6 @@ bool llama_memory_hybrid_state::apply() { return res; } -std::vector & llama_memory_hybrid_state::out_ids() { - assert(status == LLAMA_MEMORY_STATUS_SUCCESS); - - return sbatch.out_ids; -} - llama_memory_status llama_memory_hybrid_state::get_status() const { return status; } diff --git a/src/llama-memory-hybrid.h b/src/llama-memory-hybrid.h index b5700c522..4d27ab896 100644 --- a/src/llama-memory-hybrid.h +++ b/src/llama-memory-hybrid.h @@ -50,9 +50,9 @@ public: // llama_memory_state_ptr init_batch( - const llama_batch & batch, + llama_batch_allocr & balloc, uint32_t n_ubatch, - bool embd_pooled) override; + bool embd_all) override; llama_memory_state_ptr init_full() override; @@ -107,7 +107,6 @@ public: // init success llama_memory_hybrid_state( llama_memory_hybrid * mem, - llama_sbatch sbatch, std::vector heads_attn, std::vector ubatches); @@ -116,8 +115,6 @@ public: bool next() override; bool apply() override; - std::vector & out_ids() override; - llama_memory_status get_status() const override; const llama_ubatch & get_ubatch() const override; @@ -129,8 +126,6 @@ public: const llama_memory_recurrent_state * get_state_recr() const; private: - llama_sbatch sbatch; - // the index of the next ubatch to process size_t i_next = 0; diff --git a/src/llama-memory-recurrent.cpp b/src/llama-memory-recurrent.cpp index c4f9a6f1d..b064da008 100644 --- a/src/llama-memory-recurrent.cpp +++ b/src/llama-memory-recurrent.cpp @@ -362,29 +362,31 @@ llama_pos llama_memory_recurrent::seq_pos_max(llama_seq_id seq_id) const { return result; } -llama_memory_state_ptr llama_memory_recurrent::init_batch(const llama_batch & batch, uint32_t n_ubatch, bool embd_all) { - auto sbatch = llama_sbatch(batch, hparams.n_embd, false); - +llama_memory_state_ptr llama_memory_recurrent::init_batch(llama_batch_allocr & balloc, uint32_t n_ubatch, bool embd_all) { std::vector ubatches; - while (sbatch.n_tokens > 0) { + while (true) { llama_ubatch ubatch; if (embd_all) { // if all tokens are output, split by sequence - ubatch = sbatch.split_seq(n_ubatch); + ubatch = balloc.split_seq(n_ubatch); } else { - ubatch = sbatch.split_equal(n_ubatch); + ubatch = balloc.split_equal(n_ubatch); } - ubatches.push_back(ubatch); + if (ubatch.n_tokens == 0) { + break; + } + + ubatches.push_back(std::move(ubatch)); // NOLINT } if (!prepare(ubatches)) { return std::make_unique(LLAMA_MEMORY_STATUS_FAILED_PREPARE); } - return std::make_unique(this, std::move(sbatch), std::move(ubatches)); + return std::make_unique(this, std::move(ubatches)); } llama_memory_state_ptr llama_memory_recurrent::init_full() { @@ -423,9 +425,8 @@ bool llama_memory_recurrent::prepare(const std::vector & ubatches) } bool llama_memory_recurrent::find_slot(const llama_ubatch & ubatch) { - const uint32_t n_seqs = ubatch.n_seqs; - const uint32_t n_seq_tokens = ubatch.n_seq_tokens; + const uint32_t n_seqs = ubatch.n_seqs; // if we have enough unused cells before the current head -> // better to start searching from the beginning of the cache, hoping to fill it @@ -445,9 +446,11 @@ bool llama_memory_recurrent::find_slot(const llama_ubatch & ubatch) { // everything should fit if all seq_ids are smaller than the max for (uint32_t s = 0; s < n_seqs; ++s) { - const uint32_t n_seq_id = ubatch.n_seq_id[s]; + const uint32_t i = s*n_seq_tokens; // first token of sequence set s + const uint32_t n_seq_id = ubatch.n_seq_id[i]; + for (uint32_t j = 0; j < n_seq_id; ++j) { - const llama_seq_id seq_id = ubatch.seq_id[s][j]; + const llama_seq_id seq_id = ubatch.seq_id[i][j]; if (seq_id < 0 || (uint32_t) seq_id >= size) { // too big seq_id @@ -506,7 +509,8 @@ bool llama_memory_recurrent::find_slot(const llama_ubatch & ubatch) { // find usable cell range for (uint32_t s = 0; s < n_seqs; ++s) { - const llama_seq_id seq_id = ubatch.seq_id[s][0]; + const uint32_t i = s*n_seq_tokens; + const llama_seq_id seq_id = ubatch.seq_id[i][0]; auto & seq_meta = cells[seq_id]; bool has_cell = false; if (seq_meta.tail >= 0) { @@ -530,7 +534,7 @@ bool llama_memory_recurrent::find_slot(const llama_ubatch & ubatch) { seq_meta.tail = next_empty_cell; // find next empty cell if (s + 1 < n_seqs) { - for (uint32_t i = 0; i < size; ++i) { + for (uint32_t j = 0; j < size; ++j) { next_empty_cell += 1; if (next_empty_cell >= size) { next_empty_cell -= size; } auto & cell = cells[next_empty_cell]; @@ -544,8 +548,9 @@ bool llama_memory_recurrent::find_slot(const llama_ubatch & ubatch) { // gather and re-order for (uint32_t s = 0; s < n_seqs; ++s) { + const uint32_t i = s*n_seq_tokens; const int32_t dst_id = s + min; - const int32_t src_id = cells[ubatch.seq_id[s][0]].tail; + const int32_t src_id = cells[ubatch.seq_id[i][0]].tail; if (dst_id != src_id) { auto & dst_cell = cells[dst_id]; auto & src_cell = cells[src_id]; @@ -555,8 +560,8 @@ bool llama_memory_recurrent::find_slot(const llama_ubatch & ubatch) { std::swap(dst_cell.seq_id, src_cell.seq_id); // swap tails - for (uint32_t i = 0; i < size; ++i) { - int32_t & tail = cells[i].tail; + for (uint32_t j = 0; j < size; ++j) { + int32_t & tail = cells[j].tail; if (tail == src_id) { tail = dst_id; } else if (tail == dst_id) { @@ -568,7 +573,8 @@ bool llama_memory_recurrent::find_slot(const llama_ubatch & ubatch) { // update the pos of the used seqs for (uint32_t s = 0; s < n_seqs; ++s) { - const llama_pos last_pos = ubatch.pos[n_seq_tokens * s + n_seq_tokens - 1]; + const uint32_t i = s*n_seq_tokens; + const llama_pos last_pos = ubatch.pos[i + n_seq_tokens - 1]; const int32_t cell_id = s + min; auto & cell = cells[cell_id]; @@ -576,12 +582,12 @@ bool llama_memory_recurrent::find_slot(const llama_ubatch & ubatch) { // What should happen when the pos backtracks or skips a value? // Clearing the state mid-batch would require special-casing which isn't done. LLAMA_LOG_WARN("%s: non-consecutive token position %d after %d for sequence %d with %u new tokens\n", - __func__, last_pos, cell.pos, ubatch.seq_id[s][0], n_seq_tokens); + __func__, last_pos, cell.pos, ubatch.seq_id[i][0], n_seq_tokens); } cell.pos = last_pos; cell.seq_id.clear(); - for (int32_t j = 0; j < ubatch.n_seq_id[s]; ++j) { - const llama_seq_id seq_id = ubatch.seq_id[s][j]; + for (int32_t j = 0; j < ubatch.n_seq_id[i]; ++j) { + const llama_seq_id seq_id = ubatch.seq_id[i][j]; cell.seq_id.insert(seq_id); cells[seq_id].tail = cell_id; } @@ -827,12 +833,9 @@ bool llama_memory_recurrent::state_read_meta(llama_io_read_i & io, uint32_t cell seq_rm(dest_seq_id, -1, -1); - llama_sbatch sbatch; - llama_ubatch batch = sbatch.reserve_ubatch(cell_count, /* has_embd */ false); + llama_batch_allocr balloc(hparams.n_pos_per_embd()); - batch.n_tokens = cell_count; - batch.n_seq_tokens = cell_count; - batch.n_seqs = 1; + llama_ubatch ubatch = balloc.ubatch_reserve(cell_count, 1); for (uint32_t i = 0; i < cell_count; ++i) { llama_pos pos; @@ -846,12 +849,12 @@ bool llama_memory_recurrent::state_read_meta(llama_io_read_i & io, uint32_t cell return false; } - batch.pos[i] = pos; + ubatch.pos[i] = pos; } - batch.n_seq_id[0] = 1; - batch.seq_id[0] = &dest_seq_id; + ubatch.n_seq_id[0] = 1; + ubatch.seq_id[0] = &dest_seq_id; - if (!find_slot(batch)) { + if (!find_slot(ubatch)) { LLAMA_LOG_ERROR("%s: failed to find available cells in kv cache\n", __func__); return false; } @@ -859,8 +862,8 @@ bool llama_memory_recurrent::state_read_meta(llama_io_read_i & io, uint32_t cell // DEBUG CHECK: kv.head should be our first cell, kv.head + cell_count - 1 should be our last cell (verify seq_id and pos values) // Assume that this is one contiguous block of cells GGML_ASSERT(head + cell_count <= size); - GGML_ASSERT(cells[head].pos == batch.pos[0]); - GGML_ASSERT(cells[head + cell_count - 1].pos == batch.pos[cell_count - 1]); + GGML_ASSERT(cells[head].pos == ubatch.pos[0]); + GGML_ASSERT(cells[head + cell_count - 1].pos == ubatch.pos[cell_count - 1]); GGML_ASSERT(cells[head].has_seq_id(dest_seq_id)); GGML_ASSERT(cells[head + cell_count - 1].has_seq_id(dest_seq_id)); } else { @@ -1048,8 +1051,7 @@ llama_memory_recurrent_state::llama_memory_recurrent_state( llama_memory_recurrent_state::llama_memory_recurrent_state( llama_memory_recurrent * mem, - llama_sbatch sbatch, - std::vector ubatches) : status(LLAMA_MEMORY_STATUS_SUCCESS), mem(mem), sbatch(std::move(sbatch)), ubatches(std::move(ubatches)) {} + std::vector ubatches) : status(LLAMA_MEMORY_STATUS_SUCCESS), mem(mem), ubatches(std::move(ubatches)) {} llama_memory_recurrent_state::~llama_memory_recurrent_state() = default; @@ -1071,12 +1073,6 @@ bool llama_memory_recurrent_state::apply() { return true; } -std::vector & llama_memory_recurrent_state::out_ids() { - assert(status == LLAMA_MEMORY_STATUS_SUCCESS); - - return sbatch.out_ids; -} - llama_memory_status llama_memory_recurrent_state::get_status() const { return status; } diff --git a/src/llama-memory-recurrent.h b/src/llama-memory-recurrent.h index 290cc84ab..be58dae7c 100644 --- a/src/llama-memory-recurrent.h +++ b/src/llama-memory-recurrent.h @@ -35,7 +35,7 @@ public: // llama_memory_state_ptr init_batch( - const llama_batch & batch, + llama_batch_allocr & balloc, uint32_t n_ubatch, bool embd_all) override; @@ -137,7 +137,6 @@ public: // used to create a state from a batch llama_memory_recurrent_state( llama_memory_recurrent * mem, - llama_sbatch sbatch, std::vector ubatches); virtual ~llama_memory_recurrent_state(); @@ -149,8 +148,6 @@ public: bool next() override; bool apply() override; - std::vector & out_ids() override; - llama_memory_status get_status() const override; const llama_ubatch & get_ubatch() const override; @@ -173,8 +170,6 @@ private: llama_memory_recurrent * mem; - llama_sbatch sbatch; - size_t i_next = 0; std::vector ubatches; diff --git a/src/llama-memory.h b/src/llama-memory.h index 24668f861..d2ef0c2a3 100644 --- a/src/llama-memory.h +++ b/src/llama-memory.h @@ -7,6 +7,8 @@ struct llama_ubatch; +class llama_batch_allocr; + class llama_io_write_i; class llama_io_read_i; @@ -50,9 +52,6 @@ struct llama_memory_state_i { // return false on failure virtual bool apply() = 0; - // TODO: this might get reworked in the future when refactoring llama_batch - virtual std::vector & out_ids() = 0; - // get the current ubatch virtual const llama_ubatch & get_ubatch() const = 0; @@ -71,7 +70,7 @@ struct llama_memory_i { // return a state object containing the ubatches and KV cache state required to process them // check the llama_memory_state_i::get_status() for the result virtual llama_memory_state_ptr init_batch( - const llama_batch & batch, + llama_batch_allocr & balloc, uint32_t n_ubatch, bool embd_all) = 0; diff --git a/tools/server/server.cpp b/tools/server/server.cpp index 9d55b3338..aa18513e3 100644 --- a/tools/server/server.cpp +++ b/tools/server/server.cpp @@ -3385,38 +3385,6 @@ struct server_context { llama_set_embeddings(ctx, slot_batched->need_embd()); } - // pad the batch so that batch.n_tokens >= n_slots - // TODO: temporary workaround for https://github.com/ggml-org/llama.cpp/issues/13689 - if (slot_batched->need_embd()) { - const int n_slots = slots.size(); - - if (batch.n_tokens < n_slots) { - std::set seq_ids; - for (int j = 0; j < batch.n_tokens; ++j) { - seq_ids.insert(batch.seq_id[j][0]); - } - - // find unused sequence id - llama_seq_id seq_id = -1; - for (int i = 0; i < n_slots; ++i) { - if (seq_ids.find(i) == seq_ids.end()) { - seq_id = i; - } - } - - const int n_add = n_slots - batch.n_tokens; - - SRV_WRN("adding %d dummy tokens to the batch, seq_id = %d\n", n_add, seq_id); - - for (int j = 0; j < n_add; ++j) { - common_batch_add(batch, 0, j, { seq_id }, true); - } - - slots[seq_id].cache_tokens.clear(); - llama_memory_seq_rm(llama_get_memory(ctx), seq_id, -1, -1); - } - } - int32_t i_next = 0; // process the created batch of tokens From 812939a9e90f99d1bd5bb1bc6b99d12600671d50 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Fri, 20 Jun 2025 10:50:27 +0300 Subject: [PATCH 12/43] model : more uniform output id handling (#14275) * model : more uniform output id handling ggml-ci * cont : revert n_outputs < n_tokens optimization ggml-ci * cont : fix out_ids initialization ggml-ci --- src/llama-graph.cpp | 54 +-- src/llama-model.cpp | 847 ++++++++++++++++++++++---------------------- 2 files changed, 459 insertions(+), 442 deletions(-) diff --git a/src/llama-graph.cpp b/src/llama-graph.cpp index 083366fd6..7e162c555 100644 --- a/src/llama-graph.cpp +++ b/src/llama-graph.cpp @@ -92,36 +92,28 @@ void llm_graph_input_pos_bucket_kv::set_input(const llama_ubatch * ubatch) { } void llm_graph_input_out_ids::set_input(const llama_ubatch * ubatch) { - if (hparams.causal_attn || cparams.pooling_type == LLAMA_POOLING_TYPE_NONE) { - //GGML_ASSERT(out_ids && "every model that can must skip unused outputs"); + GGML_ASSERT(out_ids); - if (!out_ids) { - LLAMA_LOG_WARN("%s: 'out_ids' is not created\n", __func__); - } else { - const int64_t n_tokens = ubatch->n_tokens; + const int64_t n_tokens = ubatch->n_tokens; - GGML_ASSERT(ggml_backend_buffer_is_host(out_ids->buffer)); - int32_t * data = (int32_t *) out_ids->data; + GGML_ASSERT(ggml_backend_buffer_is_host(out_ids->buffer)); + int32_t * data = (int32_t *) out_ids->data; - if (n_outputs == n_tokens) { - for (int i = 0; i < n_tokens; ++i) { - data[i] = i; - } - } else if (ubatch->output) { - int32_t n_outputs = 0; - for (int i = 0; i < n_tokens; ++i) { - if (ubatch->output[i]) { - data[n_outputs++] = i; - } - } - // the graph needs to have been passed the correct number of outputs - GGML_ASSERT(n_outputs == n_outputs); - } else if (n_outputs == 1) { - // only keep last output - data[0] = n_tokens - 1; - } else { - GGML_ASSERT(n_outputs == 0); - } + if (n_outputs == n_tokens) { + for (int i = 0; i < n_tokens; ++i) { + data[i] = i; + } + + return; + } + + GGML_ASSERT(ubatch->output); + + int n_outputs = 0; + + for (int i = 0; i < n_tokens; ++i) { + if (ubatch->output[i]) { + data[n_outputs++] = i; } } } @@ -874,6 +866,14 @@ ggml_tensor * llm_graph_context::build_inp_attn_scale() const { } ggml_tensor * llm_graph_context::build_inp_out_ids() const { + // note: when all tokens are output, we could skip this optimization to spare the ggml_get_rows() calls, + // but this would make the graph topology depend on the number of output tokens, which can interere with + // features that require constant topology such as pipline parallelism + // ref: https://github.com/ggml-org/llama.cpp/pull/14275#issuecomment-2987424471 + //if (n_outputs < n_tokens) { + // return nullptr; + //} + auto inp = std::make_unique(hparams, cparams, n_outputs); auto & cur = inp->out_ids; diff --git a/src/llama-model.cpp b/src/llama-model.cpp index a5853f8b1..e2c82017f 100644 --- a/src/llama-model.cpp +++ b/src/llama-model.cpp @@ -4707,6 +4707,8 @@ struct llm_build_llama : public llm_graph_context { const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale; + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -4769,9 +4771,7 @@ struct llm_build_llama : public llm_graph_context { cb(cur, "attn_out", il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } @@ -4867,6 +4867,8 @@ struct llm_build_llama_iswa : public llm_graph_context { const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale; + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -4943,9 +4945,7 @@ struct llm_build_llama_iswa : public llm_graph_context { cb(cur, "attn_out", il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } @@ -5045,6 +5045,9 @@ struct llm_build_deci : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale; + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; const int64_t n_head_kv = hparams.n_head_kv(il); @@ -5118,9 +5121,7 @@ struct llm_build_deci : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } @@ -5199,6 +5200,8 @@ struct llm_build_baichuan : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -5250,9 +5253,7 @@ struct llm_build_baichuan : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } @@ -5321,6 +5322,8 @@ struct llm_build_xverse : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -5365,9 +5368,7 @@ struct llm_build_xverse : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } @@ -5435,6 +5436,8 @@ struct llm_build_falcon : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * attn_norm; @@ -5490,9 +5493,7 @@ struct llm_build_falcon : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); attn_norm = ggml_get_rows(ctx0, attn_norm, inp_out_ids); @@ -5561,6 +5562,8 @@ struct llm_build_grok : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -5620,9 +5623,7 @@ struct llm_build_grok : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f, il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } @@ -5721,6 +5722,8 @@ struct llm_build_dbrx : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -5771,9 +5774,7 @@ struct llm_build_dbrx : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } @@ -5853,6 +5854,8 @@ struct llm_build_starcoder : public llm_graph_context { inpL = ggml_add(ctx0, inpL, pos); cb(inpL, "inpL", -1); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { cur = build_norm(inpL, model.layers[il].attn_norm, @@ -5885,9 +5888,7 @@ struct llm_build_starcoder : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); } @@ -5952,6 +5953,8 @@ struct llm_build_refact : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -5984,9 +5987,7 @@ struct llm_build_refact : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } @@ -6072,78 +6073,79 @@ struct llm_build_bert : public llm_graph_context { auto * inp_attn = build_attn_inp_no_cache(); - // iterate layers + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * cur = inpL; - ggml_tensor * Qcur; - ggml_tensor * Kcur; - ggml_tensor * Vcur; + { + ggml_tensor * Qcur; + ggml_tensor * Kcur; + ggml_tensor * Vcur; - // self-attention - if (model.layers[il].wqkv) { - cur = build_lora_mm(model.layers[il].wqkv, cur); - cb(cur, "wqkv", il); + // self-attention + if (model.layers[il].wqkv) { + cur = build_lora_mm(model.layers[il].wqkv, cur); + cb(cur, "wqkv", il); - if (model.layers[il].bqkv) { - cur = ggml_add(ctx0, cur, model.layers[il].bqkv); - cb(cur, "bqkv", il); + if (model.layers[il].bqkv) { + cur = ggml_add(ctx0, cur, model.layers[il].bqkv); + cb(cur, "bqkv", il); + } + + Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd))); + Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd))); + Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa))); + } else { + Qcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wq, cur), model.layers[il].bq); + Kcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wk, cur), model.layers[il].bk); + Vcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wv, cur), model.layers[il].bv); } - Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd))); - Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd))); - Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa))); - } else { - Qcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wq, cur), model.layers[il].bq); - Kcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wk, cur), model.layers[il].bk); - Vcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wv, cur), model.layers[il].bv); + if (model.layers[il].attn_q_norm) { + Qcur = build_norm(Qcur, + model.layers[il].attn_q_norm, + model.layers[il].attn_q_norm_b, + LLM_NORM, il); + } + + if (model.layers[il].attn_k_norm) { + Kcur = build_norm(Kcur, + model.layers[il].attn_k_norm, + model.layers[il].attn_k_norm_b, + LLM_NORM, il); + } + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + // RoPE + if (model.arch == LLM_ARCH_NOMIC_BERT || model.arch == LLM_ARCH_NOMIC_BERT_MOE) { + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + } + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + cb(cur, "kqv_out", il); } - if (model.layers[il].attn_q_norm) { - Qcur = build_norm(Qcur, - model.layers[il].attn_q_norm, - model.layers[il].attn_q_norm_b, - LLM_NORM, il); - } - - if (model.layers[il].attn_k_norm) { - Kcur = build_norm(Kcur, - model.layers[il].attn_k_norm, - model.layers[il].attn_k_norm_b, - LLM_NORM, il); - } - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); - - // RoPE - if (model.arch == LLM_ARCH_NOMIC_BERT || model.arch == LLM_ARCH_NOMIC_BERT_MOE) { - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - Kcur = ggml_rope_ext( - ctx0, Kcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - } - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - cur = build_attn(inp_attn, gf, - model.layers[il].wo, model.layers[il].bo, - Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); - cb(cur, "kqv_out", il); - - if (il == n_layer - 1 && pooling_type == LLAMA_POOLING_TYPE_NONE) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); } @@ -6240,56 +6242,57 @@ struct llm_build_neo_bert : public llm_graph_context { auto * inp_attn = build_attn_inp_no_cache(); - // iterate layers + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * cur = inpL; - ggml_tensor * Qcur; - ggml_tensor * Kcur; - ggml_tensor * Vcur; - // pre-norm cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il); - // self-attention - cur = build_lora_mm(model.layers[il].wqkv, cur); - cb(cur, "wqkv", il); + { + ggml_tensor * Qcur; + ggml_tensor * Kcur; + ggml_tensor * Vcur; - Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd))); - Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd))); - Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa))); + // self-attention + cur = build_lora_mm(model.layers[il].wqkv, cur); + cb(cur, "wqkv", il); - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd))); + Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd))); + Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa))); - // RoPE - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); - Kcur = ggml_rope_ext( - ctx0, Kcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); + // RoPE + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); - cur = build_attn(inp_attn, gf, - model.layers[il].wo, nullptr, - Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); - cb(cur, "kqv_out", il); + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); - if (il == n_layer - 1 && pooling_type == LLAMA_POOLING_TYPE_NONE) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = build_attn(inp_attn, gf, + model.layers[il].wo, nullptr, + Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + cb(cur, "kqv_out", il); + } + + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); } @@ -6354,6 +6357,8 @@ struct llm_build_bloom : public llm_graph_context { LLM_NORM, -1); cb(inpL, "inp_norm", -1); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { cur = build_norm(inpL, model.layers[il].attn_norm, @@ -6386,9 +6391,7 @@ struct llm_build_bloom : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); } @@ -6465,6 +6468,8 @@ struct llm_build_mpt : public llm_graph_context { cb(inpL, "inpL", -1); } + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * attn_norm; @@ -6527,9 +6532,7 @@ struct llm_build_mpt : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); } @@ -6598,6 +6601,8 @@ struct llm_build_stablelm : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { // norm cur = build_norm(inpL, @@ -6673,9 +6678,7 @@ struct llm_build_stablelm : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); @@ -6750,6 +6753,8 @@ struct llm_build_qwen : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -6796,9 +6801,7 @@ struct llm_build_qwen : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } @@ -6867,6 +6870,8 @@ struct llm_build_qwen2 : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -6916,9 +6921,7 @@ struct llm_build_qwen2 : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } @@ -6988,6 +6991,8 @@ struct llm_build_qwen2vl : public llm_graph_context { int sections[4]; std::copy(std::begin(hparams.rope_sections), std::begin(hparams.rope_sections) + 4, sections); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -7037,9 +7042,7 @@ struct llm_build_qwen2vl : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } @@ -7106,6 +7109,8 @@ struct llm_build_qwen2moe : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -7164,9 +7169,7 @@ struct llm_build_qwen2moe : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } @@ -7265,6 +7268,8 @@ struct llm_build_qwen3 : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -7317,9 +7322,7 @@ struct llm_build_qwen3 : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } @@ -7386,6 +7389,8 @@ struct llm_build_qwen3moe : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -7438,9 +7443,7 @@ struct llm_build_qwen3moe : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } @@ -7516,6 +7519,8 @@ struct llm_build_phi2 : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { attn_norm_output = build_norm(inpL, model.layers[il].attn_norm, @@ -7578,9 +7583,7 @@ struct llm_build_phi2 : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f, il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); attn_norm_output = ggml_get_rows(ctx0, attn_norm_output, inp_out_ids); @@ -7652,6 +7655,8 @@ struct llm_build_phi3 : public llm_graph_context { inp_attn = build_attn_inp_kv_unified(); } + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { auto * residual = inpL; @@ -7715,9 +7720,7 @@ struct llm_build_phi3 : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f, il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor* inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); residual = ggml_get_rows(ctx0, residual, inp_out_ids); } @@ -7803,15 +7806,16 @@ struct llm_build_plamo : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); - for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { // norm cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il); cb(cur, "attn_norm", il); - ggml_tensor * attention_norm = cur; + ggml_tensor * sa_inp = cur; // self-attention { @@ -7849,18 +7853,17 @@ struct llm_build_plamo : public llm_graph_context { model.layers[il].wo, NULL, Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - ggml_tensor * sa_out = cur; - cur = attention_norm; - - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); - sa_out = ggml_get_rows(ctx0, sa_out, inp_out_ids); + sa_inp = ggml_get_rows(ctx0, sa_inp, inp_out_ids); inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); } + ggml_tensor * sa_out = cur; + + cur = sa_inp; + // feed-forward network { cur = build_ffn(cur, @@ -7925,6 +7928,8 @@ struct llm_build_gpt2 : public llm_graph_context { inpL = ggml_add(ctx0, inpL, pos); cb(inpL, "inpL", -1); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { cur = build_norm(inpL, model.layers[il].attn_norm, @@ -7957,9 +7962,7 @@ struct llm_build_gpt2 : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); } @@ -8029,6 +8032,8 @@ struct llm_build_codeshell : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { cur = build_norm(inpL, model.layers[il].attn_norm, @@ -8073,9 +8078,7 @@ struct llm_build_codeshell : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); } @@ -8129,128 +8132,128 @@ struct llm_build_codeshell : public llm_graph_context { struct llm_build_orion : public llm_graph_context { llm_build_orion(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { - const int64_t n_embd_head = hparams.n_embd_head_v; + const int64_t n_embd_head = hparams.n_embd_head_v; - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - GGML_ASSERT(n_embd_head == hparams.n_rot); + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); - ggml_tensor * cur; - ggml_tensor * inpL; + ggml_tensor * cur; + ggml_tensor * inpL; - inpL = build_inp_embd(model.tok_embd); + inpL = build_inp_embd(model.tok_embd); - // inp_pos - contains the positions - ggml_tensor * inp_pos = build_inp_pos(); + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); - auto * inp_attn = build_attn_inp_kv_unified(); + auto * inp_attn = build_attn_inp_kv_unified(); - for (int il = 0; il < n_layer; ++il) { - ggml_tensor * inpSA = inpL; + ggml_tensor * inp_out_ids = build_inp_out_ids(); - // norm - cur = build_norm(inpL, - model.layers[il].attn_norm, model.layers[il].attn_norm_b, - LLM_NORM, il); - cb(cur, "attn_norm", il); + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; - // self-attention - { - // compute Q and K and RoPE them - ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - // if (model.layers[il].bq) { - // Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); - // cb(Qcur, "Qcur", il); - // } + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, model.layers[il].attn_norm_b, + LLM_NORM, il); + cb(cur, "attn_norm", il); - ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - // if (model.layers[il].bk) { - // Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); - // cb(Kcur, "Kcur", il); - // } + // self-attention + { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + // if (model.layers[il].bq) { + // Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + // cb(Qcur, "Qcur", il); + // } - ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - // if (model.layers[il].bv) { - // Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); - // cb(Vcur, "Vcur", il); - // } + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + // if (model.layers[il].bk) { + // Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + // cb(Kcur, "Kcur", il); + // } - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + // if (model.layers[il].bv) { + // Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + // cb(Vcur, "Vcur", il); + // } - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); - Kcur = ggml_rope_ext( - ctx0, Kcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); - cur = build_attn(inp_attn, gf, - model.layers[il].wo, NULL, - Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, NULL, + Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, model.layers[il].ffn_norm_b, + LLM_NORM, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } + cur = inpL; - ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); + cur = build_norm(cur, + model.output_norm, model.output_norm_b, + LLM_NORM, -1); - // feed-forward network - cur = build_norm(ffn_inp, - model.layers[il].ffn_norm, model.layers[il].ffn_norm_b, - LLM_NORM, il); - cb(cur, "ffn_norm", il); + cb(cur, "result_norm", -1); + res->t_embd = cur; - cur = build_ffn(cur, - model.layers[il].ffn_up, NULL, NULL, - model.layers[il].ffn_gate, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, il); - cb(cur, "ffn_out", il); + // lm_head + cur = build_lora_mm(model.output, cur); - cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "result_output", -1); + res->t_logits = cur; - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = build_norm(cur, - model.output_norm, model.output_norm_b, - LLM_NORM, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - // lm_head - cur = build_lora_mm(model.output, cur); - - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); + ggml_build_forward_expand(gf, cur); } }; @@ -8271,6 +8274,8 @@ struct llm_build_internlm2 : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -8329,9 +8334,7 @@ struct llm_build_internlm2 : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } @@ -8407,6 +8410,8 @@ struct llm_build_minicpm3 : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -8526,15 +8531,13 @@ struct llm_build_minicpm3 : public llm_graph_context { q_states, k_states, v_states, nullptr, nullptr, kq_scale, il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } // scale_res - scale the hidden states for residual connection - const float scale_res = scale_depth/sqrtf(float(n_layer)); + const float scale_res = scale_depth/sqrtf(float(n_layer)); // TODO: is this correct? cur = ggml_scale(ctx0, cur, scale_res); cb(cur, "hidden_scaled", il); @@ -8611,6 +8614,8 @@ struct llm_build_gemma : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { // norm cur = build_norm(inpL, @@ -8656,9 +8661,7 @@ struct llm_build_gemma : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f, il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); } @@ -8727,6 +8730,8 @@ struct llm_build_gemma2_iswa : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified_iswa(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { // norm cur = build_norm(inpL, @@ -8771,18 +8776,16 @@ struct llm_build_gemma2_iswa : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f, il); } + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); + } + cur = build_norm(cur, model.layers[il].attn_post_norm, NULL, LLM_NORM_RMS, il); cb(cur, "attn_post_norm", il); - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); - } - ggml_tensor * sa_out = ggml_add(ctx0, cur, inpL); cb(sa_out, "sa_out", il); @@ -8861,6 +8864,8 @@ struct llm_build_gemma3_iswa : public llm_graph_context { // TODO: is causal == true correct? might need some changes auto * inp_attn = build_attn_inp_kv_unified_iswa(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { const float freq_base_l = model.get_rope_freq_base (cparams, il); const float freq_scale_l = model.get_rope_freq_scale(cparams, il); @@ -8913,18 +8918,16 @@ struct llm_build_gemma3_iswa : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f, il); } + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); + } + cur = build_norm(cur, model.layers[il].attn_post_norm, NULL, LLM_NORM_RMS, il); cb(cur, "attn_post_norm", il); - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); - } - ggml_tensor * sa_out = ggml_add(ctx0, cur, inpL); cb(sa_out, "sa_out", il); @@ -8995,6 +8998,8 @@ struct llm_build_starcoder2 : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -9053,9 +9058,7 @@ struct llm_build_starcoder2 : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } @@ -9118,6 +9121,8 @@ struct llm_build_mamba : public llm_graph_context { auto * rs_inp = build_rs_inp(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { // norm cur = build_norm(inpL, @@ -9127,9 +9132,7 @@ struct llm_build_mamba : public llm_graph_context { cur = build_mamba_layer(rs_inp, gf, cur, ubatch, il); - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); } @@ -9311,13 +9314,15 @@ struct llm_build_command_r : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); - for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { // norm cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM, il); cb(cur, "attn_norm", il); + ggml_tensor * ffn_inp = cur; // self-attention @@ -9385,9 +9390,7 @@ struct llm_build_command_r : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); ffn_inp = ggml_get_rows(ctx0, ffn_inp, inp_out_ids); @@ -9458,6 +9461,8 @@ struct llm_build_cohere2_iswa : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified_iswa(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { const bool is_swa = hparams.is_swa(il); @@ -9520,9 +9525,7 @@ struct llm_build_cohere2_iswa : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); ffn_inp = ggml_get_rows(ctx0, ffn_inp, inp_out_ids); @@ -9593,6 +9596,8 @@ struct llm_build_olmo : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -9651,9 +9656,7 @@ struct llm_build_olmo : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } @@ -9721,6 +9724,8 @@ struct llm_build_olmo2 : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -9771,18 +9776,16 @@ struct llm_build_olmo2 : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + cur = build_norm(cur, model.layers[il].attn_post_norm, NULL, LLM_NORM_RMS, il); cb(cur, "attn_post_norm", il); - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); cb(ffn_inp, "ffn_inp", il); @@ -9850,6 +9853,8 @@ struct llm_build_olmoe : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -9904,9 +9909,7 @@ struct llm_build_olmoe : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } @@ -9976,6 +9979,8 @@ struct llm_build_openelm : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { const int64_t n_head = hparams.n_head(il); const int64_t n_head_kv = hparams.n_head_kv(il); @@ -10037,11 +10042,9 @@ struct llm_build_openelm : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { residual = ggml_get_rows(ctx0, residual, inp_out_ids); - cur = ggml_get_rows(ctx0, cur, inp_out_ids); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); } ggml_tensor * ffn_inp = ggml_add(ctx0, residual, cur); @@ -10107,6 +10110,8 @@ struct llm_build_gptneox : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { cur = build_norm(inpL, model.layers[il].attn_norm, @@ -10151,9 +10156,7 @@ struct llm_build_gptneox : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); } @@ -10255,6 +10258,8 @@ struct llm_build_arctic : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -10301,9 +10306,7 @@ struct llm_build_arctic : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } @@ -10395,6 +10398,8 @@ struct llm_build_deepseek : public llm_graph_context { const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale; + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -10456,14 +10461,11 @@ struct llm_build_deepseek : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } - ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); cb(ffn_inp, "ffn_inp", il); @@ -10571,6 +10573,8 @@ struct llm_build_deepseek2 : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -10720,9 +10724,7 @@ struct llm_build_deepseek2 : public llm_graph_context { } } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } @@ -10818,6 +10820,8 @@ struct llm_build_bitnet : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -10900,9 +10904,7 @@ struct llm_build_bitnet : public llm_graph_context { cb(cur, "attn_o_out", il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } @@ -10977,6 +10979,8 @@ struct llm_build_t5_enc : public llm_graph_context { auto * inp_attn = build_attn_inp_no_cache(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -11010,9 +11014,7 @@ struct llm_build_t5_enc : public llm_graph_context { cb(cur, "kqv_out", il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } @@ -11083,6 +11085,8 @@ struct llm_build_t5_dec : public llm_graph_context { auto * inp_attn_self = build_attn_inp_kv_unified(); auto * inp_attn_cross = build_attn_inp_cross(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -11174,11 +11178,8 @@ struct llm_build_t5_dec : public llm_graph_context { //cb(cur, "kqv_out", il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); inpCA = ggml_get_rows(ctx0, inpCA, inp_out_ids); } @@ -11248,6 +11249,8 @@ struct llm_build_jais : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { cur = build_norm(inpL, model.layers[il].attn_norm, @@ -11280,9 +11283,7 @@ struct llm_build_jais : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/float(n_embd_head), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); } @@ -11346,6 +11347,8 @@ struct llm_build_chatglm : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -11412,9 +11415,7 @@ struct llm_build_chatglm : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } @@ -11479,6 +11480,8 @@ struct llm_build_glm4 : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -11545,9 +11548,7 @@ struct llm_build_glm4 : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } @@ -11630,6 +11631,8 @@ struct llm_build_nemotron : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -11689,9 +11692,7 @@ struct llm_build_nemotron : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } @@ -11759,6 +11760,8 @@ struct llm_build_exaone : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -11820,9 +11823,7 @@ struct llm_build_exaone : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } @@ -12098,6 +12099,8 @@ struct llm_build_rwkv6 : public llm_build_rwkv6_base { const auto n_seq_tokens = ubatch.n_seq_tokens; const auto n_seqs = ubatch.n_seqs; + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { const llama_layer * layer = &model.layers[il]; inpL = ggml_reshape_3d(ctx0, inpL, n_embd, n_seq_tokens, n_seqs); @@ -12139,13 +12142,16 @@ struct llm_build_rwkv6 : public llm_build_rwkv6_base { ); ggml_build_forward_expand(gf, build_rwkv_token_shift_store(token_shift, ubatch, il)); - if (il == n_layer - 1) { - // skip computing output for unused tokens - struct ggml_tensor * inp_out_ids = build_inp_out_ids(); - ffn_inp = ggml_get_rows(ctx0, ggml_reshape_2d(ctx0, ffn_inp, n_embd, n_tokens), inp_out_ids); - ffn_norm = ggml_get_rows(ctx0, ggml_reshape_2d(ctx0, ffn_norm, n_embd, n_tokens), inp_out_ids); - x_prev = ggml_get_rows(ctx0, ggml_reshape_2d(ctx0, x_prev, n_embd, n_tokens), inp_out_ids); - cur = ggml_get_rows(ctx0, ggml_reshape_2d(ctx0, cur, n_embd, n_tokens), inp_out_ids); + ffn_inp = ggml_reshape_2d(ctx0, ffn_inp, n_embd, n_tokens); + ffn_norm = ggml_reshape_2d(ctx0, ffn_norm, n_embd, n_tokens); + x_prev = ggml_reshape_2d(ctx0, x_prev, n_embd, n_tokens); + cur = ggml_reshape_2d(ctx0, cur, n_embd, n_tokens); + + if (il == n_layer - 1 && inp_out_ids) { + ffn_inp = ggml_get_rows(ctx0, ffn_inp, inp_out_ids); + ffn_norm = ggml_get_rows(ctx0, ffn_norm, inp_out_ids); + x_prev = ggml_get_rows(ctx0, x_prev, inp_out_ids); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); } cur = build_rwkv6_channel_mix(layer, ffn_norm, x_prev, LLM_ARCH_RWKV6); @@ -12193,6 +12199,8 @@ struct llm_build_rwkv6qwen2 : public llm_build_rwkv6_base { const auto n_seq_tokens = ubatch.n_seq_tokens; const auto n_seqs = ubatch.n_seqs; + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { const llama_layer * layer = &model.layers[il]; inpL = ggml_reshape_3d(ctx0, inpL, n_embd, n_seq_tokens, n_seqs); @@ -12217,11 +12225,12 @@ struct llm_build_rwkv6qwen2 : public llm_build_rwkv6_base { ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL); cb(ffn_inp, "ffn_inp", il); - if (il == n_layer - 1) { - // skip computing output for unused tokens - struct ggml_tensor * inp_out_ids = build_inp_out_ids(); - cur = ggml_get_rows(ctx0, ggml_reshape_2d(ctx0, cur, n_embd, n_tokens), inp_out_ids); - ffn_inp = ggml_get_rows(ctx0, ggml_reshape_2d(ctx0, ffn_inp, n_embd, n_tokens), inp_out_ids); + cur = ggml_reshape_2d(ctx0, cur, n_embd, n_tokens); + ffn_inp = ggml_reshape_2d(ctx0, ffn_inp, n_embd, n_tokens); + + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + ffn_inp = ggml_get_rows(ctx0, ffn_inp, inp_out_ids); } // feed-forward network @@ -12447,6 +12456,8 @@ struct llm_build_rwkv7 : public llm_build_rwkv7_base { const auto n_seq_tokens = ubatch.n_seq_tokens; const auto n_seqs = ubatch.n_seqs; + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { const llama_layer * layer = &model.layers[il]; inpL = ggml_reshape_3d(ctx0, inpL, n_embd, n_seq_tokens, n_seqs); @@ -12488,12 +12499,14 @@ struct llm_build_rwkv7 : public llm_build_rwkv7_base { ); ggml_build_forward_expand(gf, build_rwkv_token_shift_store(token_shift, ubatch, il)); - if (il == n_layer - 1) { - // skip computing output for unused tokens - struct ggml_tensor * inp_out_ids = build_inp_out_ids(); - ffn_inp = ggml_get_rows(ctx0, ggml_reshape_2d(ctx0, ffn_inp, n_embd, n_tokens), inp_out_ids); - ffn_norm = ggml_get_rows(ctx0, ggml_reshape_2d(ctx0, ffn_norm, n_embd, n_tokens), inp_out_ids); - x_prev = ggml_get_rows(ctx0, ggml_reshape_2d(ctx0, x_prev, n_embd, n_tokens), inp_out_ids); + ffn_inp = ggml_reshape_2d(ctx0, ffn_inp, n_embd, n_tokens); + ffn_norm = ggml_reshape_2d(ctx0, ffn_norm, n_embd, n_tokens); + x_prev = ggml_reshape_2d(ctx0, x_prev, n_embd, n_tokens); + + if (il == n_layer - 1 && inp_out_ids) { + ffn_inp = ggml_get_rows(ctx0, ffn_inp, inp_out_ids); + ffn_norm = ggml_get_rows(ctx0, ffn_norm, inp_out_ids); + x_prev = ggml_get_rows(ctx0, x_prev, inp_out_ids); } cur = build_rwkv7_channel_mix(layer, ffn_norm, x_prev, LLM_ARCH_RWKV7); @@ -12538,6 +12551,8 @@ struct llm_build_arwkv7 : public llm_build_rwkv7_base { const auto n_seq_tokens = ubatch.n_seq_tokens; const auto n_seqs = ubatch.n_seqs; + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { const llama_layer * layer = &model.layers[il]; inpL = ggml_reshape_3d(ctx0, inpL, n_embd, n_seq_tokens, n_seqs); @@ -12562,11 +12577,12 @@ struct llm_build_arwkv7 : public llm_build_rwkv7_base { ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL); cb(ffn_inp, "ffn_inp", il); - if (il == n_layer - 1) { - // skip computing output for unused tokens - struct ggml_tensor * inp_out_ids = build_inp_out_ids(); - cur = ggml_get_rows(ctx0, ggml_reshape_2d(ctx0, cur, n_embd, n_tokens), inp_out_ids); - ffn_inp = ggml_get_rows(ctx0, ggml_reshape_2d(ctx0, ffn_inp, n_embd, n_tokens), inp_out_ids); + cur = ggml_reshape_2d(ctx0, cur, n_embd, n_tokens); + ffn_inp = ggml_reshape_2d(ctx0, ffn_inp, n_embd, n_tokens); + + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + ffn_inp = ggml_get_rows(ctx0, ffn_inp, inp_out_ids); } // feed-forward network @@ -12635,6 +12651,9 @@ struct llm_build_granite : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale; + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -12697,9 +12716,7 @@ struct llm_build_granite : public llm_graph_context { cb(cur, "attn_out", il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } @@ -12818,6 +12835,8 @@ struct llm_build_chameleon : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -12894,21 +12913,19 @@ struct llm_build_chameleon : public llm_graph_context { cur = build_attn(inp_attn, gf, model.layers[il].wo, nullptr, Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); - - if (hparams.swin_norm) { - cur = build_norm(cur, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, il); - } } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } + if (hparams.swin_norm) { + cur = build_norm(cur, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + } + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); cb(ffn_inp, "ffn_inp", il); @@ -13149,6 +13166,8 @@ struct llm_build_plm : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -13252,9 +13271,7 @@ struct llm_build_plm : public llm_graph_context { q_states, k_states, v_states, nullptr, nullptr, kq_scale, il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } @@ -13314,6 +13331,8 @@ struct llm_build_bailingmoe : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -13375,9 +13394,7 @@ struct llm_build_bailingmoe : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_rot)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } @@ -13463,6 +13480,8 @@ struct llm_build_dots1 : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -13515,9 +13534,7 @@ struct llm_build_dots1 : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } @@ -13615,6 +13632,8 @@ struct llm_build_arcee : public llm_graph_context { const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale; + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -13677,9 +13696,7 @@ struct llm_build_arcee : public llm_graph_context { cb(cur, "attn_out", il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } From 9230dbe2c757e2d5071329095727d0fa9d4b85c4 Mon Sep 17 00:00:00 2001 From: Charles Xu Date: Fri, 20 Jun 2025 09:51:01 +0200 Subject: [PATCH 13/43] ggml: Update KleidiAI to v1.9.0 (#14277) --- ggml/src/ggml-cpu/CMakeLists.txt | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/ggml/src/ggml-cpu/CMakeLists.txt b/ggml/src/ggml-cpu/CMakeLists.txt index df0034057..52cae778c 100644 --- a/ggml/src/ggml-cpu/CMakeLists.txt +++ b/ggml/src/ggml-cpu/CMakeLists.txt @@ -465,9 +465,9 @@ function(ggml_add_cpu_backend_variant_impl tag_name) # Fetch KleidiAI sources: include(FetchContent) - set(KLEIDIAI_COMMIT_TAG "v1.6.0") + set(KLEIDIAI_COMMIT_TAG "v1.9.0") set(KLEIDIAI_DOWNLOAD_URL "https://github.com/ARM-software/kleidiai/archive/refs/tags/${KLEIDIAI_COMMIT_TAG}.tar.gz") - set(KLEIDIAI_ARCHIVE_MD5 "75b4ad68f25ab673dcc01065e5a0b05f") + set(KLEIDIAI_ARCHIVE_MD5 "2a8e1bb55d201557553545536489a017") if (POLICY CMP0135) cmake_policy(SET CMP0135 NEW) From d27b3ca1758dfb1718e333d497ef4b68ad109bc2 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Fri, 20 Jun 2025 11:19:15 +0300 Subject: [PATCH 14/43] ggml : fix repack work size for mul_mat_id (#14292) ggml-ci --- ggml/src/ggml-cpu/repack.cpp | 38 ++++++++++++++++++++++++------------ 1 file changed, 26 insertions(+), 12 deletions(-) diff --git a/ggml/src/ggml-cpu/repack.cpp b/ggml/src/ggml-cpu/repack.cpp index 5c6715d5c..290719290 100644 --- a/ggml/src/ggml-cpu/repack.cpp +++ b/ggml/src/ggml-cpu/repack.cpp @@ -1163,13 +1163,24 @@ template op) { case GGML_OP_MUL_MAT: - size = ggml_row_size(PARAM_TYPE, ggml_nelements(op->src[1])); - return true; + { + size = ggml_row_size(PARAM_TYPE, ggml_nelements(op->src[1])); + return true; + } case GGML_OP_MUL_MAT_ID: - size = ggml_row_size(PARAM_TYPE, ggml_nelements(op->src[1])); - size = GGML_PAD(size, sizeof(int64_t)); // + padding for next bloc. - size += sizeof(int64_t) * (1+op->src[0]->ne[2]) * op->src[1]->ne[2]; - return true; + { + size = ggml_row_size(PARAM_TYPE, ggml_nelements(op->src[1])); + size = GGML_PAD(size, sizeof(int64_t)); // + padding for next bloc. + + const int64_t ne02 = op->src[0]->ne[2]; // n_as, n_expert + const int64_t ne12 = op->src[1]->ne[2]; // n_tokens + + const size_t sizeof_mmid_row_mapping = sizeof(int64_t); + + size += sizeof_mmid_row_mapping*ne02*(ne12 + 1); + + return true; + } default: // GGML_ABORT("fatal error"); break; @@ -1305,14 +1316,17 @@ template wsize >= (GGML_PAD(nbw3, sizeof(int64_t)) + n_as * sizeof(int64_t) + - n_as * ne12 * sizeof(mmid_row_mapping))); + GGML_ASSERT(params->wsize >= + (GGML_PAD(nbw3, sizeof(int64_t)) + + n_as*(ne12 + 1)*sizeof(mmid_row_mapping)) + ); - auto * wdata = (char *) params->wdata; - auto * wdata_src1_end = (char *) wdata + GGML_PAD(nbw3, sizeof(int64_t)); - auto * matrix_row_counts = (int64_t *) (wdata_src1_end); // [n_as] + auto * wdata = (char *)params->wdata; + auto * wdata_src1_end = (char *)wdata + GGML_PAD(nbw3, sizeof(int64_t)); - struct mmid_row_mapping * matrix_rows = (struct mmid_row_mapping *) (matrix_row_counts + n_as); // [n_as][ne12] + // total of [n_as][ne12 + 1] elemets of type mmid_row_mapping (2*int32_t = int64_t) + auto * matrix_row_counts = (int64_t *) (wdata_src1_end); // [n_as] + struct mmid_row_mapping * matrix_rows = (struct mmid_row_mapping *) (matrix_row_counts + n_as); // [n_as][ne12] // src1: float32 => param type for (int64_t i12 = 0; i12 < ne12; ++i12) { From e28c1b93fd7d3f8faf9551d962e8a65fe2122e38 Mon Sep 17 00:00:00 2001 From: Diego Devesa Date: Fri, 20 Jun 2025 04:57:36 -0700 Subject: [PATCH 15/43] cuda : synchronize graph capture and cublas handle destruction (#14288) Workarounds an issue that may cause CUDA graph capture to fail when a cuBLAS handle is destroyed in a different thread --- ggml/src/ggml-cuda/common.cuh | 18 ++------------ ggml/src/ggml-cuda/ggml-cuda.cu | 44 ++++++++++++++++++++++++++++++--- 2 files changed, 43 insertions(+), 19 deletions(-) diff --git a/ggml/src/ggml-cuda/common.cuh b/ggml/src/ggml-cuda/common.cuh index c14a12f54..364efcaec 100644 --- a/ggml/src/ggml-cuda/common.cuh +++ b/ggml/src/ggml-cuda/common.cuh @@ -19,10 +19,10 @@ #endif #include "ggml-common.h" -#include #include #include #include +#include #include #include @@ -767,21 +767,7 @@ struct ggml_backend_cuda_context { name(GGML_CUDA_NAME + std::to_string(device)) { } - ~ggml_backend_cuda_context() { - if (copy_event != nullptr) { - CUDA_CHECK(cudaEventDestroy(copy_event)); - } - for (int i = 0; i < GGML_CUDA_MAX_DEVICES; ++i) { - for (int j = 0; j < GGML_CUDA_MAX_STREAMS; ++j) { - if (streams[i][j] != nullptr) { - CUDA_CHECK(cudaStreamDestroy(streams[i][j])); - } - } - if (cublas_handles[i] != nullptr) { - CUBLAS_CHECK(cublasDestroy(cublas_handles[i])); - } - } - } + ~ggml_backend_cuda_context(); cudaStream_t stream(int device, int stream) { if (streams[device][stream] == nullptr) { diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu index 80fe05073..530f541f9 100644 --- a/ggml/src/ggml-cuda/ggml-cuda.cu +++ b/ggml/src/ggml-cuda/ggml-cuda.cu @@ -48,6 +48,7 @@ #include #include #include +#include #include #include #include @@ -55,9 +56,8 @@ #include #include #include -#include -#include #include +#include #include #include #include @@ -515,6 +515,33 @@ std::unique_ptr ggml_backend_cuda_context::new_pool_for_device(i return std::unique_ptr(new ggml_cuda_pool_leg(device)); } +// destroying a cuBLAS handle while a graph is being captured in a different thread can result in a CUDA error +// this lock is used to ensure that no cuBLAS handle is destroyed while a graph is being captured + +static std::mutex ggml_cuda_lock; +static std::condition_variable ggml_cuda_lock_cv; +static std::atomic ggml_cuda_lock_counter; + +ggml_backend_cuda_context::~ggml_backend_cuda_context() { + std::unique_lock lock(ggml_cuda_lock); + ggml_cuda_lock_cv.wait(lock, []{ return ggml_cuda_lock_counter.load(std::memory_order_relaxed) == 0; }); + + if (copy_event != nullptr) { + CUDA_CHECK(cudaEventDestroy(copy_event)); + } + for (int i = 0; i < GGML_CUDA_MAX_DEVICES; ++i) { + for (int j = 0; j < GGML_CUDA_MAX_STREAMS; ++j) { + if (streams[i][j] != nullptr) { + CUDA_CHECK(cudaStreamDestroy(streams[i][j])); + } + } + if (cublas_handles[i] != nullptr) { + CUBLAS_CHECK(cublasDestroy(cublas_handles[i])); + } + } +} + + // cuda buffer struct ggml_backend_cuda_buffer_context { @@ -2689,6 +2716,11 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx CUDA_CHECK(cudaStreamEndCapture(cuda_ctx->stream(), &cuda_ctx->cuda_graph->graph)); graph_evaluated_or_captured = true; // CUDA graph has been captured + + std::lock_guard lock(ggml_cuda_lock); + if (ggml_cuda_lock_counter.fetch_sub(1, std::memory_order_relaxed) == 1) { + ggml_cuda_lock_cv.notify_all(); + } } else { graph_evaluated_or_captured = true; // ggml graph has been directly evaluated } @@ -2764,7 +2796,13 @@ static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend, } } - if (use_cuda_graph && cuda_graph_update_required) { // Start CUDA graph capture + if (use_cuda_graph && cuda_graph_update_required) { + // Start CUDA graph capture + { + std::lock_guard lock(ggml_cuda_lock); + ggml_cuda_lock_counter.fetch_add(1, std::memory_order_relaxed); + } + CUDA_CHECK(cudaStreamBeginCapture(cuda_ctx->stream(), cudaStreamCaptureModeRelaxed)); } From 88fc854b4bd2e3caf10e705e6afcbbca136f0a3c Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Sigbj=C3=B8rn=20Skj=C3=A6ret?= Date: Fri, 20 Jun 2025 14:04:09 +0200 Subject: [PATCH 16/43] llama : improve sep token handling (#14272) --- ci/run.sh | 2 +- common/arg.cpp | 7 +++ common/common.h | 1 + convert_hf_to_gguf.py | 14 ------ examples/embedding/embedding.cpp | 34 +++++++++++-- gguf-py/gguf/constants.py | 1 + gguf-py/gguf/gguf_writer.py | 3 ++ gguf-py/gguf/vocab.py | 83 ++++++++++++++++++++++++++++++-- include/llama.h | 1 + src/llama-arch.cpp | 1 + src/llama-arch.h | 1 + src/llama-model-saver.cpp | 1 + src/llama-vocab.cpp | 24 +++++++-- src/llama-vocab.h | 1 + tools/server/utils.hpp | 16 ++++-- 15 files changed, 161 insertions(+), 29 deletions(-) diff --git a/ci/run.sh b/ci/run.sh index 940055705..e1b777c30 100755 --- a/ci/run.sh +++ b/ci/run.sh @@ -779,7 +779,7 @@ function gg_run_rerank_tiny { model_f16="${path_models}/ggml-model-f16.gguf" # for this model, the SEP token is "" - (time ./bin/llama-embedding --model ${model_f16} -p "what is panda?hi\nwhat is panda?it's a bear\nwhat is panda?The giant panda (Ailuropoda melanoleuca), sometimes called a panda bear or simply panda, is a bear species endemic to China." -ngl 99 -c 0 --pooling rank --embd-normalize -1 --verbose-prompt) 2>&1 | tee -a $OUT/${ci}-rk-f16.log + (time ./bin/llama-embedding --model ${model_f16} -p "what is panda?\thi\nwhat is panda?\tit's a bear\nwhat is panda?\tThe giant panda (Ailuropoda melanoleuca), sometimes called a panda bear or simply panda, is a bear species endemic to China." -ngl 99 -c 0 --pooling rank --embd-normalize -1 --verbose-prompt) 2>&1 | tee -a $OUT/${ci}-rk-f16.log # sample output # rerank score 0: 0.029 diff --git a/common/arg.cpp b/common/arg.cpp index 3dfaa71ef..c4ad85c47 100644 --- a/common/arg.cpp +++ b/common/arg.cpp @@ -2706,6 +2706,13 @@ common_params_context common_params_parser_init(common_params & params, llama_ex params.embd_sep = value; } ).set_examples({LLAMA_EXAMPLE_EMBEDDING})); + add_opt(common_arg( + {"--cls-separator"}, "STRING", + "separator of classification sequences (default \\t) for example \"<#seq#>\"", + [](common_params & params, const std::string & value) { + params.cls_sep = value; + } + ).set_examples({LLAMA_EXAMPLE_EMBEDDING})); add_opt(common_arg( {"--host"}, "HOST", string_format("ip address to listen, or bind to an UNIX socket if the address ends with .sock (default: %s)", params.hostname.c_str()), diff --git a/common/common.h b/common/common.h index 5710c4e97..e08a59eae 100644 --- a/common/common.h +++ b/common/common.h @@ -358,6 +358,7 @@ struct common_params { int32_t embd_normalize = 2; // normalisation for embeddings (-1=none, 0=max absolute int16, 1=taxicab, 2=euclidean, >2=p-norm) std::string embd_out = ""; // empty = default, "array" = [[],[]...], "json" = openai style, "json+" = same "json" + cosine similarity matrix std::string embd_sep = "\n"; // separator of embeddings + std::string cls_sep = "\t"; // separator of classification sequences // server params int32_t port = 8080; // server listens on this network port diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py index 2e08db345..2fe76589e 100755 --- a/convert_hf_to_gguf.py +++ b/convert_hf_to_gguf.py @@ -2145,7 +2145,6 @@ class Llama4Model(LlamaModel): def set_vocab(self): self._set_vocab_gpt2() - self.gguf_writer.add_add_bos_token(True) def set_gguf_parameters(self): super().set_gguf_parameters() @@ -3918,9 +3917,6 @@ class BertModel(TextModel): special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens)) special_vocab.add_to_gguf(self.gguf_writer) - self.gguf_writer.add_add_bos_token(True) - self.gguf_writer.add_add_eos_token(True) - @ModelBase.register("DistilBertModel", "DistilBertForMaskedLM", "DistilBertForSequenceClassification") class DistilBertModel(BertModel): @@ -3962,8 +3958,6 @@ class RobertaModel(BertModel): bpe_tok_path = self.dir_model / "tokenizer.json" if bpe_tok_path.exists(): self._set_vocab_gpt2() - self.gguf_writer.add_add_bos_token(True) - self.gguf_writer.add_add_eos_token(True) # we need this to validate the size of the token_type embeddings # though currently we are passing all zeros to the token_type embeddings @@ -4848,8 +4842,6 @@ class JinaBertV2Model(BertModel): self.gguf_writer.add_token_type_count(2) else: raise NotImplementedError(f'Tokenizer {tokenizer_class} is not supported for JinaBertModel') - self.gguf_writer.add_add_bos_token(True) - self.gguf_writer.add_add_eos_token(True) @ModelBase.register("OpenELMForCausalLM") @@ -5451,9 +5443,6 @@ class T5Model(TextModel): special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens)) special_vocab.add_to_gguf(self.gguf_writer) - self.gguf_writer.add_add_bos_token(False) - self.gguf_writer.add_add_eos_token(True) - def set_gguf_parameters(self): if (n_ctx := self.find_hparam(["n_positions"], optional=True)) is None: logger.warning("Couldn't find context length in config.json, assuming default value of 512") @@ -5591,9 +5580,6 @@ class T5EncoderModel(TextModel): special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens)) special_vocab.add_to_gguf(self.gguf_writer) - self.gguf_writer.add_add_bos_token(False) - self.gguf_writer.add_add_eos_token(True) - def set_gguf_parameters(self): if (n_ctx := self.find_hparam(["n_positions"], optional=True)) is None: logger.warning("Couldn't find context length in config.json, assuming default value of 512") diff --git a/examples/embedding/embedding.cpp b/examples/embedding/embedding.cpp index 681929d27..0ec2999a0 100644 --- a/examples/embedding/embedding.cpp +++ b/examples/embedding/embedding.cpp @@ -133,10 +133,36 @@ int main(int argc, char ** argv) { // max batch size const uint64_t n_batch = params.n_batch; + // get added sep and eos token, if any + const std::string added_sep_token = llama_vocab_get_add_sep(vocab) ? llama_vocab_get_text(vocab, llama_vocab_sep(vocab)) : ""; + const std::string added_eos_token = llama_vocab_get_add_eos(vocab) ? llama_vocab_get_text(vocab, llama_vocab_eos(vocab)) : ""; + // tokenize the prompts and trim std::vector> inputs; for (const auto & prompt : prompts) { - auto inp = common_tokenize(ctx, prompt, true, true); + std::vector inp; + + // split classification pairs and insert expected separator tokens + if (pooling_type == LLAMA_POOLING_TYPE_RANK && prompt.find(params.cls_sep) != std::string::npos) { + std::vector pairs = split_lines(prompt, params.cls_sep); + std::string final_prompt; + + for (size_t i = 0; i < pairs.size(); i++) { + final_prompt += pairs[i]; + if (i != pairs.size() - 1) { + if (!added_eos_token.empty()) { + final_prompt += added_eos_token; + } + if (!added_sep_token.empty()) { + final_prompt += added_sep_token; + } + } + } + + inp = common_tokenize(ctx, final_prompt, true, true); + } else { + inp = common_tokenize(ctx, prompt, true, true); + } if (inp.size() > n_batch) { LOG_ERR("%s: number of tokens in input line (%lld) exceeds batch size (%lld), increase batch size and re-run\n", __func__, (long long int) inp.size(), (long long int) n_batch); @@ -145,11 +171,11 @@ int main(int argc, char ** argv) { inputs.push_back(inp); } - // check if the last token is SEP + // check if the last token is SEP/EOS // it should be automatically added by the tokenizer when 'tokenizer.ggml.add_eos_token' is set to 'true' for (auto & inp : inputs) { - if (inp.empty() || inp.back() != llama_vocab_sep(vocab)) { - LOG_WRN("%s: last token in the prompt is not SEP\n", __func__); + if (inp.empty() || (inp.back() != llama_vocab_sep(vocab) && inp.back() != llama_vocab_eos(vocab))) { + LOG_WRN("%s: last token in the prompt is not SEP or EOS\n", __func__); LOG_WRN("%s: 'tokenizer.ggml.add_eos_token' should be set to 'true' in the GGUF header\n", __func__); } } diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py index 834a1d5e1..0429b0aaf 100644 --- a/gguf-py/gguf/constants.py +++ b/gguf-py/gguf/constants.py @@ -198,6 +198,7 @@ class Keys: MASK_ID = "tokenizer.ggml.mask_token_id" ADD_BOS = "tokenizer.ggml.add_bos_token" ADD_EOS = "tokenizer.ggml.add_eos_token" + ADD_SEP = "tokenizer.ggml.add_sep_token" ADD_PREFIX = "tokenizer.ggml.add_space_prefix" REMOVE_EXTRA_WS = "tokenizer.ggml.remove_extra_whitespaces" PRECOMPILED_CHARSMAP = "tokenizer.ggml.precompiled_charsmap" diff --git a/gguf-py/gguf/gguf_writer.py b/gguf-py/gguf/gguf_writer.py index 54ca0c33f..b9b63d052 100644 --- a/gguf-py/gguf/gguf_writer.py +++ b/gguf-py/gguf/gguf_writer.py @@ -891,6 +891,9 @@ class GGUFWriter: def add_add_eos_token(self, value: bool) -> None: self.add_bool(Keys.Tokenizer.ADD_EOS, value) + def add_add_sep_token(self, value: bool) -> None: + self.add_bool(Keys.Tokenizer.ADD_SEP, value) + def add_add_space_prefix(self, value: bool) -> None: self.add_bool(Keys.Tokenizer.ADD_PREFIX, value) diff --git a/gguf-py/gguf/vocab.py b/gguf-py/gguf/vocab.py index 44d066ee7..6c4d3a422 100644 --- a/gguf-py/gguf/vocab.py +++ b/gguf-py/gguf/vocab.py @@ -119,6 +119,7 @@ class SpecialVocab: logger.warning(f'Special token type {typ}, id {tid} out of range, must be under {self.n_vocab} - skipping') def _try_load_from_tokenizer_json(self, path: Path) -> bool: + tokenizer = None tokenizer_file = path / 'tokenizer.json' if tokenizer_file.is_file(): with open(tokenizer_file, encoding = 'utf-8') as f: @@ -152,11 +153,87 @@ class SpecialVocab: added_tokens = tokenizer.get('added_tokens', {}) else: added_tokens = {} + tokenizer_config = None tokenizer_config_file = path / 'tokenizer_config.json' - if not tokenizer_config_file.is_file(): + if tokenizer_config_file.is_file(): + with open(tokenizer_config_file, encoding = 'utf-8') as f: + tokenizer_config = json.load(f) + if tokenizer: + special_bos = (tokenizer_config or {}).get('bos_token') + special_cls = (tokenizer_config or {}).get('cls_token') + special_eos = (tokenizer_config or {}).get('eos_token') + special_sep = (tokenizer_config or {}).get('sep_token') + if not special_bos and special_cls and tokenizer_config: + tokenizer_config['bos_token'] = special_bos = special_cls + if not special_eos and special_sep and tokenizer_config: + tokenizer_config['eos_token'] = special_eos = special_sep + post_processor = tokenizer.get('post_processor', {}) + for processor in post_processor.get('processors', [post_processor]): + if processor.get('type') == 'RobertaProcessing': + self.add_special_token['bos'] = True + self.add_special_token['eos'] = True + self.add_special_token['sep'] = True + if not special_cls and tokenizer_config: + special_cls = processor.get('cls', [special_bos])[0] + tokenizer_config['cls_token'] = special_cls + if not special_sep and tokenizer_config: + special_sep = processor.get('sep', [special_eos])[0] + tokenizer_config['sep_token'] = special_sep + continue + # Crude parsing of TemplateProcessing to determine if BOS/SEP/EOS should be added + # Only works with simple templates, **will** get it wrong on unusual sequences + if processor.get('type') == 'TemplateProcessing': + tmpl_single = processor.get('single', []) + tmpl_pair = processor.get('pair', []) + special_first = None + special_last = None + if len(tmpl_single) > 1: + if special_first := tmpl_single[0].get('SpecialToken', {}).get('id'): + if not tokenizer_config: + special_bos = special_first + self.add_special_token['bos'] = True if special_first in (special_bos, special_cls) else False + if special_first not in (special_bos, special_cls): + logger.warning(f'Unknown leading special token {special_first!r} in TemplateProcessing') + if special_last := tmpl_single[-1].get('SpecialToken', {}).get('id'): + if not tokenizer_config: + special_eos = special_last + self.add_special_token['eos'] = True if special_last == special_eos else False + if special_last != special_eos: + logger.warning(f'Unknown trailing special token {special_last!r} in TemplateProcessing') + if tmpl_pair: + seq_start = 1 if tmpl_pair[0].get('SpecialToken', {}).get('id') == special_first else 0 + seq_stop = -1 if tmpl_pair[-1].get('SpecialToken', {}).get('id') == special_last else None + if seq_start == 0 or seq_stop is None: + logger.warning('TemplateProcessing leading/trailing special tokens do not match TemplateProcessing') + if tmpl_pair := tmpl_pair[slice(seq_start, seq_stop)]: + tmpl_a = tmpl_pair[0].get('Sequence', {}).get('id') + tmpl_b = tmpl_pair[-1].get('Sequence', {}).get('id') + if tmpl_a != 'A' or tmpl_b != 'B': + logger.warning(f'Unknown sequence {tmpl_a}...{tmpl_b} in TemplateProcessing') + # A [sep] [eos] B + if tmpl_a == 'A' and tmpl_b == 'B' and (tmpl_pair := tmpl_pair[1:-1]): + add_sep = False + if special_entry := tmpl_pair[0].get('SpecialToken', {}).get('id'): + if special_entry in (special_sep, special_eos) and not special_last: + add_sep = True + if special_entry not in (special_sep, special_eos): + logger.warning(f'Unknown separator token {special_entry!r} in TemplateProcessing') + else: + logger.warning(f'Unknown middle sequence {tmpl_pair[0]!r} in TemplateProcessing') + if len(tmpl_pair) == 2: + if special_entry := tmpl_pair[1].get('SpecialToken', {}).get('id'): + if special_entry in (special_sep, special_eos): + add_sep = True + if special_entry not in (special_sep, special_eos): + logger.warning(f'Unknown second separator token {special_entry!r} in TemplateProcessing') + else: + logger.warning(f'Unknown second middle sequence {tmpl_pair[1]!r} in TemplateProcessing') + self.add_special_token['sep'] = add_sep + if add_sep and not special_sep and tokenizer_config: + tokenizer_config['sep_token'] = special_eos + continue + if not tokenizer_config: return True - with open(tokenizer_config_file, encoding = 'utf-8') as f: - tokenizer_config = json.load(f) chat_template_alt = None chat_template_file = path / 'chat_template.json' if chat_template_file.is_file(): diff --git a/include/llama.h b/include/llama.h index 635508b10..3475d5965 100644 --- a/include/llama.h +++ b/include/llama.h @@ -1044,6 +1044,7 @@ extern "C" { LLAMA_API bool llama_vocab_get_add_bos(const struct llama_vocab * vocab); LLAMA_API bool llama_vocab_get_add_eos(const struct llama_vocab * vocab); + LLAMA_API bool llama_vocab_get_add_sep(const struct llama_vocab * vocab); LLAMA_API llama_token llama_vocab_fim_pre(const struct llama_vocab * vocab); LLAMA_API llama_token llama_vocab_fim_suf(const struct llama_vocab * vocab); diff --git a/src/llama-arch.cpp b/src/llama-arch.cpp index 0bc60565d..8dadef204 100644 --- a/src/llama-arch.cpp +++ b/src/llama-arch.cpp @@ -198,6 +198,7 @@ static const std::map LLM_KV_NAMES = { { LLM_KV_TOKENIZER_MASK_ID, "tokenizer.ggml.mask_token_id" }, { LLM_KV_TOKENIZER_ADD_BOS, "tokenizer.ggml.add_bos_token" }, { LLM_KV_TOKENIZER_ADD_EOS, "tokenizer.ggml.add_eos_token" }, + { LLM_KV_TOKENIZER_ADD_SEP, "tokenizer.ggml.add_sep_token" }, { LLM_KV_TOKENIZER_ADD_PREFIX, "tokenizer.ggml.add_space_prefix" }, { LLM_KV_TOKENIZER_REMOVE_EXTRA_WS, "tokenizer.ggml.remove_extra_whitespaces" }, { LLM_KV_TOKENIZER_PRECOMPILED_CHARSMAP, "tokenizer.ggml.precompiled_charsmap" }, diff --git a/src/llama-arch.h b/src/llama-arch.h index 51b242c66..5b0230c15 100644 --- a/src/llama-arch.h +++ b/src/llama-arch.h @@ -194,6 +194,7 @@ enum llm_kv { LLM_KV_TOKENIZER_MASK_ID, LLM_KV_TOKENIZER_ADD_BOS, LLM_KV_TOKENIZER_ADD_EOS, + LLM_KV_TOKENIZER_ADD_SEP, LLM_KV_TOKENIZER_ADD_PREFIX, LLM_KV_TOKENIZER_REMOVE_EXTRA_WS, LLM_KV_TOKENIZER_PRECOMPILED_CHARSMAP, diff --git a/src/llama-model-saver.cpp b/src/llama-model-saver.cpp index a70b98923..563823dc3 100644 --- a/src/llama-model-saver.cpp +++ b/src/llama-model-saver.cpp @@ -228,6 +228,7 @@ void llama_model_saver::add_kv_from_model() { // add_kv(LLM_KV_TOKENIZER_MASK_ID, ???); add_kv(LLM_KV_TOKENIZER_ADD_BOS, vocab.get_add_bos()); add_kv(LLM_KV_TOKENIZER_ADD_EOS, vocab.get_add_eos()); + add_kv(LLM_KV_TOKENIZER_ADD_SEP, vocab.get_add_sep()); add_kv(LLM_KV_TOKENIZER_ADD_PREFIX, vocab.get_add_space_prefix()); add_kv(LLM_KV_TOKENIZER_REMOVE_EXTRA_WS, vocab.get_remove_extra_whitespaces()); add_kv(LLM_KV_TOKENIZER_PRECOMPILED_CHARSMAP, vocab.get_precompiled_charsmap()); diff --git a/src/llama-vocab.cpp b/src/llama-vocab.cpp index d90f1d6b1..4ab120d9b 100644 --- a/src/llama-vocab.cpp +++ b/src/llama-vocab.cpp @@ -1269,6 +1269,7 @@ struct llama_vocab::impl { bool add_space_prefix = false; bool add_bos = false; bool add_eos = false; + bool add_sep = false; bool ignore_merges = false; bool clean_spaces = false; // clean_up_tokenization_spaces bool remove_extra_whitespaces = false; @@ -1421,6 +1422,8 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) { special_sep_id = 102; special_pad_id = 0; special_mask_id = 103; + + add_sep = true; } else if (tokenizer_model == "gpt2") { type = LLAMA_VOCAB_TYPE_BPE; @@ -1550,12 +1553,15 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) { tokenizer_pre == "jina-es" || tokenizer_pre == "jina-de" || tokenizer_pre == "gigachat" || - tokenizer_pre == "jina-v1-en" || tokenizer_pre == "jina-v2-es" || - tokenizer_pre == "jina-v2-de" || + tokenizer_pre == "jina-v2-de") { + pre_type = LLAMA_VOCAB_PRE_TYPE_GPT2; + } else if ( + tokenizer_pre == "jina-v1-en" || tokenizer_pre == "jina-v2-code" || tokenizer_pre == "roberta-bpe") { pre_type = LLAMA_VOCAB_PRE_TYPE_GPT2; + add_sep = true; } else if ( tokenizer_pre == "refact") { pre_type = LLAMA_VOCAB_PRE_TYPE_REFACT; @@ -1665,6 +1671,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) { clean_spaces = true; add_bos = true; add_eos = false; + add_sep = true; } else if (type == LLAMA_VOCAB_TYPE_UGM) { pre_type = LLAMA_VOCAB_PRE_TYPE_DEFAULT; add_bos = false; @@ -1801,7 +1808,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) { } } - // Handle add_bos and add_eos + // Handle add_bos, add_eos and add_sep { bool temp = true; @@ -1811,6 +1818,9 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) { if (ml.get_key(LLM_KV_TOKENIZER_ADD_EOS, temp, false)) { add_eos = temp; } + if (ml.get_key(LLM_KV_TOKENIZER_ADD_SEP, temp, false)) { + add_sep = temp; + } } // auto-detect special tokens by text @@ -3000,6 +3010,10 @@ bool llama_vocab::get_add_eos() const { return pimpl->add_eos; } +bool llama_vocab::get_add_sep() const { + return pimpl->add_sep; +} + bool llama_vocab::get_ignore_merges() const { return pimpl->ignore_merges; } @@ -3191,6 +3205,10 @@ bool llama_vocab_get_add_eos(const struct llama_vocab * vocab) { return vocab->get_add_eos(); } +bool llama_vocab_get_add_sep(const struct llama_vocab * vocab) { + return vocab->get_add_sep(); +} + llama_token llama_vocab_fim_pre(const struct llama_vocab * vocab) { return vocab->token_fim_pre(); } diff --git a/src/llama-vocab.h b/src/llama-vocab.h index daa6cf308..40e4d1c05 100644 --- a/src/llama-vocab.h +++ b/src/llama-vocab.h @@ -74,6 +74,7 @@ struct llama_vocab { bool get_add_space_prefix () const; bool get_add_bos () const; bool get_add_eos () const; + bool get_add_sep () const; bool get_ignore_merges () const; bool get_clean_spaces () const; bool get_remove_extra_whitespaces () const; diff --git a/tools/server/utils.hpp b/tools/server/utils.hpp index f3e0392a4..f8fab2c86 100644 --- a/tools/server/utils.hpp +++ b/tools/server/utils.hpp @@ -271,12 +271,20 @@ static llama_tokens format_rerank(const struct llama_vocab * vocab, const llama_ } result.reserve(doc.size() + query.size() + 4); - result.push_back(llama_vocab_bos(vocab)); + if (llama_vocab_get_add_bos(vocab)) { + result.push_back(llama_vocab_bos(vocab)); + } result.insert(result.end(), query.begin(), query.end()); - result.push_back(eos_token); - result.push_back(llama_vocab_sep(vocab)); + if (llama_vocab_get_add_eos(vocab)) { + result.push_back(eos_token); + } + if (llama_vocab_get_add_sep(vocab)) { + result.push_back(llama_vocab_sep(vocab)); + } result.insert(result.end(), doc.begin(), doc.end()); - result.push_back(eos_token); + if (llama_vocab_get_add_eos(vocab)) { + result.push_back(eos_token); + } return result; } From 6369be07359d03723f38c0a4a014ff0f698a0738 Mon Sep 17 00:00:00 2001 From: Christian Kastner Date: Fri, 20 Jun 2025 12:17:32 +0000 Subject: [PATCH 17/43] Implement GGML_CPU_ALL_VARIANTS for PowerPC (#14286) * Add PowerPC feature detection and scoring * ggml-cpu: Implement GGML_CPU_ALL_VARIANTS for PowerPC * ggml-cpu: Delay some initializations until function is called When using GGML_BACKEND_DL=ON, these initializations might use instructions that are not supported by the current CPU. --------- Co-authored-by: Diego Devesa --- ggml/src/CMakeLists.txt | 17 ++++ ggml/src/ggml-cpu/CMakeLists.txt | 21 +++++ ggml/src/ggml-cpu/arch/powerpc/cpu-feats.cpp | 82 ++++++++++++++++++++ ggml/src/ggml-cpu/repack.cpp | 29 +++---- 4 files changed, 135 insertions(+), 14 deletions(-) create mode 100644 ggml/src/ggml-cpu/arch/powerpc/cpu-feats.cpp diff --git a/ggml/src/CMakeLists.txt b/ggml/src/CMakeLists.txt index 0c453741b..9cb2c228d 100644 --- a/ggml/src/CMakeLists.txt +++ b/ggml/src/CMakeLists.txt @@ -286,6 +286,10 @@ function(ggml_add_cpu_backend_variant tag_name) foreach (feat ${ARGN}) set(GGML_INTERNAL_${feat} ON) endforeach() + elseif (GGML_SYSTEM_ARCH STREQUAL "PowerPC") + foreach (feat ${ARGN}) + set(GGML_INTERNAL_${feat} ON) + endforeach() endif() ggml_add_cpu_backend_variant_impl(${tag_name}) @@ -337,6 +341,19 @@ if (GGML_CPU_ALL_VARIANTS) else() message(FATAL_ERROR "Unsupported ARM target OS: ${CMAKE_SYSTEM_NAME}") endif() + elseif (GGML_SYSTEM_ARCH STREQUAL "PowerPC") + if (CMAKE_SYSTEM_NAME MATCHES "Linux") + ggml_add_cpu_backend_variant(power0) + ggml_add_cpu_backend_variant(power7_1 POWER7) + ggml_add_cpu_backend_variant(power7_2 POWER7 VSX) + ggml_add_cpu_backend_variant(power8_1 POWER8) + ggml_add_cpu_backend_variant(power8_2 POWER8 VSX) + ggml_add_cpu_backend_variant(power9 POWER9 VSX) + ggml_add_cpu_backend_variant(power10 POWER10 VSX) + ggml_add_cpu_backend_variant(power11 POWER11 VSX) + else() + message(FATAL_ERROR "Unsupported PowerPC target OS: ${CMAKE_SYSTEM_NAME}") + endif() else() message(FATAL_ERROR "GGML_CPU_ALL_VARIANTS not yet supported with ${GGML_SYSTEM_ARCH} on ${CMAKE_SYSTEM_NAME}") endif() diff --git a/ggml/src/ggml-cpu/CMakeLists.txt b/ggml/src/ggml-cpu/CMakeLists.txt index 52cae778c..71b1d67b8 100644 --- a/ggml/src/ggml-cpu/CMakeLists.txt +++ b/ggml/src/ggml-cpu/CMakeLists.txt @@ -388,6 +388,27 @@ function(ggml_add_cpu_backend_variant_impl tag_name) else() list(APPEND ARCH_FLAGS -mcpu=native -mtune=native -mpowerpc64) endif() + elseif(GGML_CPU_ALL_VARIANTS) + # Begin with the lowest baseline + set(ARCH_DEFINITIONS "") + + # When a feature is selected, bump the MCPU to the first + # version that supported it + foreach(PVER RANGE 7 11) + if(DEFINED GGML_INTERNAL_POWER${PVER}) + set(POWERPC_MCPU "power${PVER}") + list(APPEND ARCH_DEFINITIONS GGML_USE_POWER${PVER}) + endif() + endforeach() + if (GGML_INTERNAL_VSX) + list(APPEND ARCH_DEFINITIONS GGML_USE_VSX) + list(APPEND ARCH_FLAGS -mvsx) + endif() + + if (DEFINED POWERPC_MCPU) + list(APPEND ARCH_FLAGS -mcpu=${POWERPC_MCPU}) + endif() + ggml_add_cpu_backend_features(${GGML_CPU_NAME} powerpc ${ARCH_DEFINITIONS}) else() if (GGML_CPU_POWERPC_CPUTYPE) list(APPEND ARCH_FLAGS -mcpu=${GGML_CPU_POWERPC_CPUTYPE}) diff --git a/ggml/src/ggml-cpu/arch/powerpc/cpu-feats.cpp b/ggml/src/ggml-cpu/arch/powerpc/cpu-feats.cpp new file mode 100644 index 000000000..fedd64302 --- /dev/null +++ b/ggml/src/ggml-cpu/arch/powerpc/cpu-feats.cpp @@ -0,0 +1,82 @@ +# include "ggml-backend-impl.h" + +#if defined(__powerpc64__) || defined(__ppc64__) || defined(__PPC64__) + +#if defined(__linux__) +#include +#endif + +#include + +struct powerpc_features { + std::string platform = ""; + int power_version = -1; + + bool has_vsx = false; + + powerpc_features() { +#if defined(__linux__) + unsigned long auxval = getauxval(AT_PLATFORM); + if (auxval) { + platform = std::string(reinterpret_cast(auxval)); + // TBD: Do systems exist that return this in uppercase? + if (platform.substr(0, 5) == "power") { + // Extractt a numeric suffix, if one exists + int vpos = -1; + for (int i = platform.length() - 1; i >= 0; i--) { + if (std::isdigit(platform[i])) { + vpos = i; + } else { + break; + } + } + if (vpos > -1) { + power_version = std::stoi(platform.substr(vpos)); + } + } + } +#endif + if (power_version >= 9) { + has_vsx = true; + } + } +}; + +static int ggml_backend_cpu_powerpc_score() { + int score = 1; + powerpc_features pf; + +// Platform scores +#if defined(GGML_USE_POWER7) + if (pf.power_version < 7) { return 0; } + score += 1<<1; +#endif +#if defined(GGML_USE_POWER8) + if (pf.power_version < 8) { return 0; } + score += 1<<2; +#endif +#if defined(GGML_USE_POWER9) + if (pf.power_version < 9) { return 0; } + score += 1<<3; +#endif +#if defined(GGML_USE_POWER10) + if (pf.power_version < 10) { return 0; } + score += 1<<4; +#endif +#if defined(GGML_USE_POWER11) + if (pf.power_version < 11) { return 0; } + score += 1<<5; +#endif + +// Feature scores +#if defined(GGML_USE_VSX) + if (!pf.has_vsx) { return 0; } + score += 1<<6; +#endif + + return score; +} + +GGML_BACKEND_DL_SCORE_IMPL(ggml_backend_cpu_powerpc_score) + +#endif // defined(__powerpc64__) || defined(__ppc64__) || defined(__PPC64__) diff --git a/ggml/src/ggml-cpu/repack.cpp b/ggml/src/ggml-cpu/repack.cpp index 290719290..692c53e01 100644 --- a/ggml/src/ggml-cpu/repack.cpp +++ b/ggml/src/ggml-cpu/repack.cpp @@ -1411,44 +1411,45 @@ template q4_0_4x4_q8_0; -static const tensor_traits q4_0_4x8_q8_0; -static const tensor_traits q4_0_8x8_q8_0; -static const tensor_traits q4_K_8x8_q8_K; - -// instance for IQ4 -static const tensor_traits iq4_nl_4x4_q8_0; - } // namespace ggml::cpu::repack static const ggml::cpu::tensor_traits * ggml_repack_get_optimal_repack_type(const struct ggml_tensor * cur) { + + // instance for Q4 + static const ggml::cpu::repack::tensor_traits q4_0_4x4_q8_0; + static const ggml::cpu::repack::tensor_traits q4_0_4x8_q8_0; + static const ggml::cpu::repack::tensor_traits q4_0_8x8_q8_0; + static const ggml::cpu::repack::tensor_traits q4_K_8x8_q8_K; + + // instance for IQ4 + static const ggml::cpu::repack::tensor_traits iq4_nl_4x4_q8_0; + if (cur->type == GGML_TYPE_Q4_0) { if (ggml_cpu_has_avx2() || (ggml_cpu_has_sve() && ggml_cpu_has_matmul_int8() && ggml_cpu_get_sve_cnt() == QK8_0)) { if (cur->ne[1] % 8 == 0) { - return &ggml::cpu::repack::q4_0_8x8_q8_0; + return &q4_0_8x8_q8_0; } } if (ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) { if (cur->ne[1] % 4 == 0) { - return &ggml::cpu::repack::q4_0_4x8_q8_0; + return &q4_0_4x8_q8_0; } } if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) { if (cur->ne[1] % 4 == 0) { - return &ggml::cpu::repack::q4_0_4x4_q8_0; + return &q4_0_4x4_q8_0; } } } else if (cur->type == GGML_TYPE_Q4_K) { if (ggml_cpu_has_avx2()) { if (cur->ne[1] % 8 == 0) { - return &ggml::cpu::repack::q4_K_8x8_q8_K; + return &q4_K_8x8_q8_K; } } } else if (cur->type == GGML_TYPE_IQ4_NL) { if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) { if (cur->ne[1] % 4 == 0) { - return &ggml::cpu::repack::iq4_nl_4x4_q8_0; + return &iq4_nl_4x4_q8_0; } } } From 8308f98c7fb778e54bf75538f5234d8bd20915e9 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Nicol=C3=B2=20Scipione?= Date: Fri, 20 Jun 2025 15:07:21 +0200 Subject: [PATCH 18/43] sycl: add usage of enqueue_functions extension (#14244) * Add header and namespace to use enqueue_functions extension * Convert submit and parallel_for to use new extension in convert.cpp * Convert submit and parallel_for to use extension in ggml-sycl.cpp * Convert submit and parallel_for to use extension in gla.cpp * Convert submit and parallel_for in mmq.cpp * Convert submit and parallel_for in mmvq.cpp * Convert submit and parallel_for in remaining files * Convert all simple parallel_for to nd_launch from enqueue_functions extension * Wrapping extension in general function Create a general function that enable the enqueue_functions extension if it is enable in the compiler, otherwise call the general SYCL function to launch kernels. --------- Signed-off-by: nscipione --- ggml/src/ggml-sycl/binbcast.cpp | 11 +- ggml/src/ggml-sycl/concat.cpp | 69 +++--- ggml/src/ggml-sycl/conv.cpp | 14 +- ggml/src/ggml-sycl/convert.cpp | 265 +++++++++------------- ggml/src/ggml-sycl/cpy.cpp | 166 ++++++++------ ggml/src/ggml-sycl/dmmv.cpp | 116 ++++------ ggml/src/ggml-sycl/dpct/helper.hpp | 32 ++- ggml/src/ggml-sycl/element_wise.cpp | 258 +++++++++------------ ggml/src/ggml-sycl/getrows.cpp | 15 +- ggml/src/ggml-sycl/ggml-sycl.cpp | 93 ++++---- ggml/src/ggml-sycl/gla.cpp | 4 +- ggml/src/ggml-sycl/im2col.cpp | 2 +- ggml/src/ggml-sycl/mmq.cpp | 140 +++++------- ggml/src/ggml-sycl/mmvq.cpp | 333 +++++++++++----------------- ggml/src/ggml-sycl/norm.cpp | 129 +++++------ ggml/src/ggml-sycl/rope.cpp | 44 ++-- ggml/src/ggml-sycl/softmax.cpp | 6 +- ggml/src/ggml-sycl/tsembd.cpp | 11 +- ggml/src/ggml-sycl/wkv.cpp | 28 +-- 19 files changed, 750 insertions(+), 986 deletions(-) diff --git a/ggml/src/ggml-sycl/binbcast.cpp b/ggml/src/ggml-sycl/binbcast.cpp index 0a3883ae1..741630dba 100644 --- a/ggml/src/ggml-sycl/binbcast.cpp +++ b/ggml/src/ggml-sycl/binbcast.cpp @@ -225,9 +225,9 @@ struct bin_bcast_sycl { dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, block_num) * - sycl::range<3>(1, 1, block_size), + sycl_parallel_for( + stream, + sycl::nd_range<3>(sycl::range<3>(1, 1, block_num) * sycl::range<3>(1, 1, block_size), sycl::range<3>(1, 1, block_size)), [=](sycl::nd_item<3> item_ct1) { k_bin_bcast_unravel( @@ -246,9 +246,8 @@ struct bin_bcast_sycl { dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for( + stream, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { k_bin_bcast(src0_dd, src1_dd, dst_dd, ne0, ne1, ne2, ne3, ne10, ne11, ne12, ne13, s1, s2, s3, s01, s02, s03, s11, s12, s13, diff --git a/ggml/src/ggml-sycl/concat.cpp b/ggml/src/ggml-sycl/concat.cpp index 7aa91c861..3501484a1 100644 --- a/ggml/src/ggml-sycl/concat.cpp +++ b/ggml/src/ggml-sycl/concat.cpp @@ -89,33 +89,24 @@ static void concat_f32_sycl(const float *x, const float *y, float *dst, sycl::range<3> gridDim(ne2, ne1, num_blocks); switch (dim) { case 0: - stream->parallel_for( - sycl::nd_range<3>(gridDim * - sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE), - sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE)), - [=](sycl::nd_item<3> item_ct1) { - concat_f32_dim0(x, y, dst, ne0, ne00, item_ct1); - }); - break; + sycl_parallel_for(stream, + sycl::nd_range<3>(gridDim * sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { concat_f32_dim0(x, y, dst, ne0, ne00, item_ct1); }); + break; case 1: - stream->parallel_for( - sycl::nd_range<3>(gridDim * - sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE), - sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE)), - [=](sycl::nd_item<3> item_ct1) { - concat_f32_dim1(x, y, dst, ne0, ne01, item_ct1); - }); - break; + sycl_parallel_for(stream, + sycl::nd_range<3>(gridDim * sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { concat_f32_dim1(x, y, dst, ne0, ne01, item_ct1); }); + break; // dim >=2 will be dispatched to the default path default: - stream->parallel_for( - sycl::nd_range<3>(gridDim * - sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE), - sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE)), - [=](sycl::nd_item<3> item_ct1) { - concat_f32_dim2(x, y, dst, ne0, ne02, item_ct1); - }); - break; + sycl_parallel_for(stream, + sycl::nd_range<3>(gridDim * sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { concat_f32_dim2(x, y, dst, ne0, ne02, item_ct1); }); + break; } } @@ -129,33 +120,29 @@ static void concat_f32_sycl_non_cont( int64_t ne2, int64_t ne3, uint64_t nb0, uint64_t nb1, uint64_t nb2, uint64_t nb3, int32_t dim) { sycl::range<3> gridDim(ne3, ne2, ne1); - stream->parallel_for( - sycl::nd_range<3>(gridDim, sycl::range<3>(1, 1, 1)), - [=](sycl::nd_item<3> item_ct1) { - int64_t i3 = item_ct1.get_group(0); - int64_t i2 = item_ct1.get_group(1); - int64_t i1 = item_ct1.get_group(2); + sycl_parallel_for(stream, sycl::nd_range<3>(gridDim, sycl::range<3>(1, 1, 1)), [=](sycl::nd_item<3> item_ct1) { + int64_t i3 = item_ct1.get_group(0); + int64_t i2 = item_ct1.get_group(1); + int64_t i1 = item_ct1.get_group(2); - int64_t o[4] = {0, 0, 0, 0}; - o[dim] = dim == 0 ? ne00 : (dim == 1 ? ne01 : (dim == 2 ? ne02 : ne03)); + int64_t o[4] = { 0, 0, 0, 0 }; + o[dim] = dim == 0 ? ne00 : (dim == 1 ? ne01 : (dim == 2 ? ne02 : ne03)); - const float *x; + const float * x; - for (int i0 = item_ct1.get_local_id(2); i0 < ne0; - i0 += item_ct1.get_local_range(2)) { + for (int i0 = item_ct1.get_local_id(2); i0 < ne0; i0 += item_ct1.get_local_range(2)) { if (i0 < ne00 && i1 < ne01 && i2 < ne02 && i3 < ne03) { - x = (const float *)(src0 + (i3)*nb03 + (i2)*nb02 + (i1)*nb01 + - (i0)*nb00); + x = (const float *) (src0 + (i3) *nb03 + (i2) *nb02 + (i1) *nb01 + (i0) *nb00); } else { - x = (const float *)(src1 + (i3 - o[3]) * nb13 + (i2 - o[2]) * nb12 + - (i1 - o[1]) * nb11 + (i0 - o[0]) * nb10); + x = (const float *) (src1 + (i3 - o[3]) * nb13 + (i2 - o[2]) * nb12 + (i1 - o[1]) * nb11 + + (i0 - o[0]) * nb10); } float *y = (float *)(dst + i3 * nb3 + i2 * nb2 + i1 * nb1 + i0 * nb0); *y = *x; - } - }); + } + }); } void ggml_sycl_op_concat(ggml_backend_sycl_context & ctx, ggml_tensor *dst) { diff --git a/ggml/src/ggml-sycl/conv.cpp b/ggml/src/ggml-sycl/conv.cpp index 475bd34a2..c2f991e8d 100644 --- a/ggml/src/ggml-sycl/conv.cpp +++ b/ggml/src/ggml-sycl/conv.cpp @@ -59,16 +59,10 @@ static void conv_transpose_1d_f32_f32_sycl( const int num_blocks = (output_size + SYCL_CONV_TRANPOSE_1D_BLOCK_SIZE - 1) / SYCL_CONV_TRANPOSE_1D_BLOCK_SIZE; const sycl::range<3> block_dims(1, 1, SYCL_CONV_TRANPOSE_1D_BLOCK_SIZE); const sycl::range<3> block_nums(1, 1, num_blocks); - stream->parallel_for( - sycl::nd_range<3>( - block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { - conv_transpose_1d_kernel( - s0, output_size, - src0_ne0, src0_ne1, src0_ne2, - src1_ne0, dst_ne0, - src0, src1, dst, item_ct1); - }); + sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { + conv_transpose_1d_kernel(s0, output_size, src0_ne0, src0_ne1, src0_ne2, src1_ne0, dst_ne0, src0, src1, dst, + item_ct1); + }); } void ggml_sycl_op_conv_transpose_1d(ggml_backend_sycl_context & ctx, ggml_tensor *dst) { diff --git a/ggml/src/ggml-sycl/convert.cpp b/ggml/src/ggml-sycl/convert.cpp index 96d2583b1..0ef567122 100644 --- a/ggml/src/ggml-sycl/convert.cpp +++ b/ggml/src/ggml-sycl/convert.cpp @@ -33,14 +33,11 @@ static void dequantize_block_sycl(const void *__restrict__ vx, { dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->parallel_for( - sycl::nd_range<3>( - sycl::range<3>(1, 1, num_blocks) * - sycl::range<3>(1, 1, SYCL_DEQUANTIZE_BLOCK_SIZE), - sycl::range<3>(1, 1, SYCL_DEQUANTIZE_BLOCK_SIZE)), - [=](sycl::nd_item<3> item_ct1) { - dequantize_block(vx, y, k, item_ct1); - }); + sycl_parallel_for( + stream, + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_DEQUANTIZE_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_DEQUANTIZE_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { dequantize_block(vx, y, k, item_ct1); }); } } @@ -53,24 +50,18 @@ static void dequantize_row_q2_K_sycl(const void *vx, dst_t *y, const int64_t k, dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * - sycl::range<3>(1, 1, 64), - sycl::range<3>(1, 1, 64)), - [=](sycl::nd_item<3> item_ct1) { - dequantize_block_q2_K(vx, y, item_ct1); - }); + sycl_parallel_for( + stream, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 64), sycl::range<3>(1, 1, 64)), + [=](sycl::nd_item<3> item_ct1) { dequantize_block_q2_K(vx, y, item_ct1); }); } #else { dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * - sycl::range<3>(1, 1, 32), - sycl::range<3>(1, 1, 32)), - [=](sycl::nd_item<3> item_ct1) { - dequantize_block_q2_K(vx, y, item_ct1); - }); + sycl_parallel_for( + stream, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)), + [=](sycl::nd_item<3> item_ct1) { dequantize_block_q2_K(vx, y, item_ct1); }); } #endif @@ -85,24 +76,18 @@ static void dequantize_row_q3_K_sycl(const void *vx, dst_t *y, const int64_t k, dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * - sycl::range<3>(1, 1, 64), - sycl::range<3>(1, 1, 64)), - [=](sycl::nd_item<3> item_ct1) { - dequantize_block_q3_K(vx, y, item_ct1); - }); + sycl_parallel_for( + stream, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 64), sycl::range<3>(1, 1, 64)), + [=](sycl::nd_item<3> item_ct1) { dequantize_block_q3_K(vx, y, item_ct1); }); } #else { dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * - sycl::range<3>(1, 1, 32), - sycl::range<3>(1, 1, 32)), - [=](sycl::nd_item<3> item_ct1) { - dequantize_block_q3_K(vx, y, item_ct1); - }); + sycl_parallel_for( + stream, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)), + [=](sycl::nd_item<3> item_ct1) { dequantize_block_q3_K(vx, y, item_ct1); }); } #endif } @@ -116,12 +101,9 @@ static void dequantize_row_q4_0_sycl(const void *vx, dst_t *y, const int64_t k, dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * - sycl::range<3>(1, 1, 32), - sycl::range<3>(1, 1, 32)), - [=](sycl::nd_item<3> item_ct1) { - dequantize_block_q4_0(vx, y, nb32, item_ct1); - }); + sycl_parallel_for( + stream, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)), + [=](sycl::nd_item<3> item_ct1) { dequantize_block_q4_0(vx, y, nb32, item_ct1); }); } } @@ -135,13 +117,12 @@ static void dequantize_row_q4_0_sycl_reorder(const void *vx, dst_t *y, const int int constexpr WARP_K = WARP_SIZE * QK4_0; const int n_warp = (k + WARP_K - 1) / WARP_K; GGML_ASSERT(k % 2 == 0); - stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, n_warp) * - sycl::range<3>(1, 1, WARP_SIZE), - sycl::range<3>(1, 1, WARP_SIZE)), - [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]]{ - dequantize_block_q4_0_reorder(vx, y, k, item_ct1); - }); - + sycl_parallel_for(stream, + sycl::nd_range<3>(sycl::range<3>(1, 1, n_warp) * sycl::range<3>(1, 1, WARP_SIZE), + sycl::range<3>(1, 1, WARP_SIZE)), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + dequantize_block_q4_0_reorder(vx, y, k, item_ct1); + }); } template @@ -153,12 +134,9 @@ static void dequantize_row_q4_1_sycl(const void *vx, dst_t *y, const int64_t k, dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * - sycl::range<3>(1, 1, 32), - sycl::range<3>(1, 1, 32)), - [=](sycl::nd_item<3> item_ct1) { - dequantize_block_q4_1(vx, y, nb32, item_ct1); - }); + sycl_parallel_for( + stream, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)), + [=](sycl::nd_item<3> item_ct1) { dequantize_block_q4_1(vx, y, nb32, item_ct1); }); } } @@ -171,14 +149,13 @@ static void dequantize_row_q4_K_sycl(const void *vx, dst_t *y, const int64_t k, dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->submit([&](sycl::handler &cgh) { + sycl_launch(stream, [&](sycl::handler & cgh) { sycl::local_accessor scale_local_acc(sycl::range<1>(12), cgh); - cgh.parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * - sycl::range<3>(1, 1, 32), - sycl::range<3>(1, 1, 32)), - [=](sycl::nd_item<3> item_ct1) { - dequantize_block_q4_K(vx, y, get_pointer(scale_local_acc), item_ct1); - }); + sycl_parallel_for( + cgh, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)), + [=](sycl::nd_item<3> item_ct1) { + dequantize_block_q4_K(vx, y, get_pointer(scale_local_acc), item_ct1); + }); }); } } @@ -191,13 +168,13 @@ static void dequantize_row_q4_K_sycl_reorder(const void * vx, dst_t * y, const i dpct::has_capability_or_fail(stream->get_device(), { sycl::aspect::fp16 }); - stream->submit([&](sycl::handler & cgh) { + sycl_launch(stream, [&](sycl::handler & cgh) { sycl::local_accessor scale_local_acc(sycl::range<1>(12), cgh); - cgh.parallel_for(sycl::nd_range<1>(sycl::range<1>(global_size), sycl::range<1>(local_size)), - [=](sycl::nd_item<1> item_ct1) { - dequantize_block_q4_K_reorder(vx, y, get_pointer(scale_local_acc), item_ct1, nb); - }); + sycl_parallel_for<1>(cgh, sycl::nd_range<1>(sycl::range<1>(global_size), sycl::range<1>(local_size)), + [=](sycl::nd_item<1> item_ct1) { + dequantize_block_q4_K_reorder(vx, y, get_pointer(scale_local_acc), item_ct1, nb); + }); }); } @@ -210,24 +187,18 @@ static void dequantize_row_q5_K_sycl(const void *vx, dst_t *y, const int64_t k, dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * - sycl::range<3>(1, 1, 64), - sycl::range<3>(1, 1, 64)), - [=](sycl::nd_item<3> item_ct1) { - dequantize_block_q5_K(vx, y, item_ct1); - }); + sycl_parallel_for( + stream, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 64), sycl::range<3>(1, 1, 64)), + [=](sycl::nd_item<3> item_ct1) { dequantize_block_q5_K(vx, y, item_ct1); }); } #else { dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * - sycl::range<3>(1, 1, 32), - sycl::range<3>(1, 1, 32)), - [=](sycl::nd_item<3> item_ct1) { - dequantize_block_q5_K(vx, y, item_ct1); - }); + sycl_parallel_for( + stream, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)), + [=](sycl::nd_item<3> item_ct1) { dequantize_block_q5_K(vx, y, item_ct1); }); } #endif @@ -242,24 +213,18 @@ static void dequantize_row_q6_K_sycl(const void *vx, dst_t *y, const int64_t k, dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * - sycl::range<3>(1, 1, 64), - sycl::range<3>(1, 1, 64)), - [=](sycl::nd_item<3> item_ct1) { - dequantize_block_q6_K(vx, y, item_ct1); - }); + sycl_parallel_for( + stream, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 64), sycl::range<3>(1, 1, 64)), + [=](sycl::nd_item<3> item_ct1) { dequantize_block_q6_K(vx, y, item_ct1); }); } #else { dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * - sycl::range<3>(1, 1, 32), - sycl::range<3>(1, 1, 32)), - [=](sycl::nd_item<3> item_ct1) { - dequantize_block_q6_K(vx, y, item_ct1); - }); + sycl_parallel_for( + stream, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)), + [=](sycl::nd_item<3> item_ct1) { dequantize_block_q6_K(vx, y, item_ct1); }); } #endif @@ -271,9 +236,9 @@ static void dequantize_row_q6_K_sycl_reorder(const void * vx, dst_t * y, const i dpct::has_capability_or_fail(stream->get_device(), { sycl::aspect::fp16 }); - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 64), sycl::range<3>(1, 1, 64)), - [=](sycl::nd_item<3> item_ct1) { dequantize_block_q6_K_reorder(vx, y, item_ct1, nb); }); + sycl_parallel_for(stream, + sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 64), sycl::range<3>(1, 1, 64)), + [=](sycl::nd_item<3> item_ct1) { dequantize_block_q6_K_reorder(vx, y, item_ct1, nb); }); } template @@ -284,15 +249,10 @@ static void dequantize_row_iq1_s_sycl(const void *vx, dst_t *y, const int64_t k, dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->submit([&](sycl::handler &cgh) { - cgh.parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * - sycl::range<3>(1, 1, 32), - sycl::range<3>(1, 1, 32)), - [=](sycl::nd_item<3> item_ct1) { - dequantize_block_iq1_s( - vx, y, item_ct1, iq1s_grid_gpu - ); - }); + sycl_launch(stream, [&](sycl::handler & cgh) { + sycl_parallel_for( + cgh, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)), + [=](sycl::nd_item<3> item_ct1) { dequantize_block_iq1_s(vx, y, item_ct1, iq1s_grid_gpu); }); }); } } @@ -305,15 +265,10 @@ static void dequantize_row_iq1_m_sycl(const void *vx, dst_t *y, const int64_t k, dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->submit([&](sycl::handler &cgh) { - cgh.parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * - sycl::range<3>(1, 1, 32), - sycl::range<3>(1, 1, 32)), - [=](sycl::nd_item<3> item_ct1) { - dequantize_block_iq1_m( - vx, y, item_ct1, iq1s_grid_gpu - ); - }); + sycl_launch(stream, [&](sycl::handler & cgh) { + sycl_parallel_for( + cgh, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)), + [=](sycl::nd_item<3> item_ct1) { dequantize_block_iq1_m(vx, y, item_ct1, iq1s_grid_gpu); }); }); } } @@ -326,15 +281,12 @@ static void dequantize_row_iq2_xxs_sycl(const void *vx, dst_t *y, const int64_t dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->submit([&](sycl::handler &cgh) { - cgh.parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * - sycl::range<3>(1, 1, 32), - sycl::range<3>(1, 1, 32)), - [=](sycl::nd_item<3> item_ct1) { - dequantize_block_iq2_xxs( - vx, y, item_ct1, iq2xxs_grid, - ksigns_iq2xs, kmask_iq2xs); - }); + sycl_launch(stream, [&](sycl::handler & cgh) { + sycl_parallel_for( + cgh, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)), + [=](sycl::nd_item<3> item_ct1) { + dequantize_block_iq2_xxs(vx, y, item_ct1, iq2xxs_grid, ksigns_iq2xs, kmask_iq2xs); + }); }); } } @@ -347,15 +299,12 @@ static void dequantize_row_iq2_xs_sycl(const void *vx, dst_t *y, const int64_t k dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->submit([&](sycl::handler &cgh) { - cgh.parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * - sycl::range<3>(1, 1, 32), - sycl::range<3>(1, 1, 32)), - [=](sycl::nd_item<3> item_ct1) { - dequantize_block_iq2_xs( - vx, y, item_ct1, iq2xs_grid, - ksigns_iq2xs, kmask_iq2xs); - }); + sycl_launch(stream, [&](sycl::handler & cgh) { + sycl_parallel_for( + cgh, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)), + [=](sycl::nd_item<3> item_ct1) { + dequantize_block_iq2_xs(vx, y, item_ct1, iq2xs_grid, ksigns_iq2xs, kmask_iq2xs); + }); }); } } @@ -368,13 +317,10 @@ static void dequantize_row_iq2_s_sycl(const void *vx, dst_t *y, const int64_t k, dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->submit([&](sycl::handler &cgh) { - cgh.parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * - sycl::range<3>(1, 1, 32), - sycl::range<3>(1, 1, 32)), - [=](sycl::nd_item<3> item_ct1) { - dequantize_block_iq2_s(vx, y, item_ct1); - }); + sycl_launch(stream, [&](sycl::handler & cgh) { + sycl_parallel_for( + cgh, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)), + [=](sycl::nd_item<3> item_ct1) { dequantize_block_iq2_s(vx, y, item_ct1); }); }); } } @@ -388,15 +334,12 @@ static void dequantize_row_iq3_xxs_sycl(const void *vx, dst_t *y, const int64_t dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->submit([&](sycl::handler &cgh) { - cgh.parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * - sycl::range<3>(1, 1, 32), - sycl::range<3>(1, 1, 32)), - [=](sycl::nd_item<3> item_ct1) { - dequantize_block_iq3_xxs( - vx, y, item_ct1, iq3xxs_grid, - ksigns_iq2xs, kmask_iq2xs); - }); + sycl_launch(stream, [&](sycl::handler & cgh) { + sycl_parallel_for( + cgh, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)), + [=](sycl::nd_item<3> item_ct1) { + dequantize_block_iq3_xxs(vx, y, item_ct1, iq3xxs_grid, ksigns_iq2xs, kmask_iq2xs); + }); }); } } @@ -409,14 +352,10 @@ static void dequantize_row_iq3_s_sycl(const void *vx, dst_t *y, const int64_t k, dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->submit([&](sycl::handler &cgh) { - cgh.parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * - sycl::range<3>(1, 1, 32), - sycl::range<3>(1, 1, 32)), - [=](sycl::nd_item<3> item_ct1) { - dequantize_block_iq3_s( - vx, y, item_ct1, kmask_iq2xs, iq3s_grid); - }); + sycl_launch(stream, [&](sycl::handler & cgh) { + sycl_parallel_for( + cgh, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)), + [=](sycl::nd_item<3> item_ct1) { dequantize_block_iq3_s(vx, y, item_ct1, kmask_iq2xs, iq3s_grid); }); }); } } @@ -432,14 +371,11 @@ static void dequantize_row_iq4_xs_sycl(const void *vx, dst_t *y, const int64_t k dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->submit([&](sycl::handler &cgh) { - cgh.parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * - sycl::range<3>(1, 1, 32), - sycl::range<3>(1, 1, 32)), - [=](sycl::nd_item<3> item_ct1) { - dequantize_block_iq4_xs(vx, y, item_ct1); - }); + sycl_launch(stream, [&](sycl::handler & cgh) { + sycl_parallel_for( + cgh, + sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)), + [=](sycl::nd_item<3> item_ct1) { dequantize_block_iq4_xs(vx, y, item_ct1); }); }); } #endif @@ -453,14 +389,11 @@ static void dequantize_row_iq4_nl_sycl(const void *vx, dst_t *y, const int64_t k dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->submit([&](sycl::handler &cgh) { - cgh.parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * - sycl::range<3>(1, 1, 32), - sycl::range<3>(1, 1, 32)), - [=](sycl::nd_item<3> item_ct1) { - dequantize_block_iq4_nl(vx, y, item_ct1); - }); + sycl_launch(stream, [&](sycl::handler & cgh) { + sycl_parallel_for( + cgh, + sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)), + [=](sycl::nd_item<3> item_ct1) { dequantize_block_iq4_nl(vx, y, item_ct1); }); }); } } diff --git a/ggml/src/ggml-sycl/cpy.cpp b/ggml/src/ggml-sycl/cpy.cpp index bec137140..1ffd7f122 100644 --- a/ggml/src/ggml-sycl/cpy.cpp +++ b/ggml/src/ggml-sycl/cpy.cpp @@ -413,7 +413,8 @@ static void ggml_cpy_f16_f32_sycl(const char * cx, char * cdst, const int ne, co { dpct::has_capability_or_fail(stream->get_device(), { sycl::aspect::fp16 }); - stream->parallel_for( + sycl_parallel_for( + stream, sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) { @@ -431,7 +432,8 @@ static void ggml_cpy_f32_f32_sycl(const char * cx, char * cdst, const int ne, co { dpct::has_capability_or_fail(stream->get_device(), { sycl::aspect::fp16 }); - stream->parallel_for( + sycl_parallel_for( + stream, sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) { @@ -449,7 +451,8 @@ static void ggml_cpy_f32_f16_sycl(const char * cx, char * cdst, const int ne, co { dpct::has_capability_or_fail(stream->get_device(), { sycl::aspect::fp16 }); - stream->parallel_for( + sycl_parallel_for( + stream, sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) { @@ -465,11 +468,11 @@ static void ggml_cpy_f32_q8_0_sycl(const char * cx, char * cdst, const int ne, c const int nb12, const int nb13, queue_ptr stream) { GGML_ASSERT(ne % QK8_0 == 0); const int num_blocks = ne / QK8_0; - stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), - [=](sycl::nd_item<3> item_ct1) { - cpy_f32_q(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, - ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1); - }); + sycl_parallel_for(stream, sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), + [=](sycl::nd_item<3> item_ct1) { + cpy_f32_q(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, + ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1); + }); } static void ggml_cpy_q8_0_f32_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01, @@ -477,11 +480,11 @@ static void ggml_cpy_q8_0_f32_sycl(const char * cx, char * cdst, const int ne, c const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, queue_ptr stream) { const int num_blocks = ne; - stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), - [=](sycl::nd_item<3> item_ct1) { - cpy_q_f32(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, - ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1); - }); + sycl_parallel_for(stream, sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), + [=](sycl::nd_item<3> item_ct1) { + cpy_q_f32(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, + ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1); + }); } static void ggml_cpy_f32_q4_0_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01, @@ -490,11 +493,11 @@ static void ggml_cpy_f32_q4_0_sycl(const char * cx, char * cdst, const int ne, c const int nb12, const int nb13, queue_ptr stream) { GGML_ASSERT(ne % QK4_0 == 0); const int num_blocks = ne / QK4_0; - stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), - [=](sycl::nd_item<3> item_ct1) { - cpy_f32_q(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, - ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1); - }); + sycl_parallel_for(stream, sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), + [=](sycl::nd_item<3> item_ct1) { + cpy_f32_q(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, + ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1); + }); } static void ggml_cpy_q4_0_f32_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01, @@ -502,8 +505,9 @@ static void ggml_cpy_q4_0_f32_sycl(const char * cx, char * cdst, const int ne, c const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, queue_ptr stream) { const int num_blocks = ne; - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for( + stream, sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), + [=](sycl::nd_item<3> item_ct1) { cpy_q_f32, QK4_0>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1); @@ -516,11 +520,11 @@ static void ggml_cpy_f32_q4_1_sycl(const char * cx, char * cdst, const int ne, c const int nb12, const int nb13, queue_ptr stream) { GGML_ASSERT(ne % QK4_1 == 0); const int num_blocks = ne / QK4_1; - stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), - [=](sycl::nd_item<3> item_ct1) { - cpy_f32_q(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, - ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1); - }); + sycl_parallel_for(stream, sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), + [=](sycl::nd_item<3> item_ct1) { + cpy_f32_q(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, + ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1); + }); } static void ggml_cpy_q4_1_f32_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01, @@ -528,8 +532,9 @@ static void ggml_cpy_q4_1_f32_sycl(const char * cx, char * cdst, const int ne, c const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, queue_ptr stream) { const int num_blocks = ne; - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for( + stream, sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), + [=](sycl::nd_item<3> item_ct1) { cpy_q_f32, QK4_1>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1); @@ -542,11 +547,11 @@ static void ggml_cpy_f32_q5_0_sycl(const char * cx, char * cdst, const int ne, c const int nb12, const int nb13, queue_ptr stream) { GGML_ASSERT(ne % QK5_0 == 0); const int num_blocks = ne / QK5_0; - stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), - [=](sycl::nd_item<3> item_ct1) { - cpy_f32_q(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, - ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1); - }); + sycl_parallel_for(stream, sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), + [=](sycl::nd_item<3> item_ct1) { + cpy_f32_q(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, + ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1); + }); } static void ggml_cpy_q5_0_f32_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01, @@ -554,8 +559,9 @@ static void ggml_cpy_q5_0_f32_sycl(const char * cx, char * cdst, const int ne, c const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, queue_ptr stream) { const int num_blocks = ne; - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for( + stream, sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), + [=](sycl::nd_item<3> item_ct1) { cpy_q_f32, QK5_0>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1); @@ -568,11 +574,11 @@ static void ggml_cpy_f32_q5_1_sycl(const char * cx, char * cdst, const int ne, c const int nb12, const int nb13, queue_ptr stream) { GGML_ASSERT(ne % QK5_1 == 0); const int num_blocks = ne / QK5_1; - stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), - [=](sycl::nd_item<3> item_ct1) { - cpy_f32_q(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, - ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1); - }); + sycl_parallel_for(stream, sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), + [=](sycl::nd_item<3> item_ct1) { + cpy_f32_q(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, + ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1); + }); } static void ggml_cpy_q5_1_f32_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01, @@ -580,8 +586,9 @@ static void ggml_cpy_q5_1_f32_sycl(const char * cx, char * cdst, const int ne, c const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, queue_ptr stream) { const int num_blocks = ne; - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for( + stream, sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), + [=](sycl::nd_item<3> item_ct1) { cpy_q_f32, QK5_1>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1); @@ -594,11 +601,11 @@ static void ggml_cpy_f32_iq4_nl_sycl(const char * cx, char * cdst, const int ne, const int nb12, const int nb13, queue_ptr stream) { GGML_ASSERT(ne % QK4_NL == 0); const int num_blocks = ne / QK4_NL; - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), [=](sycl::nd_item<3> item_ct1) { - cpy_f32_q(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, - ne12, nb10, nb11, nb12, nb13, item_ct1); - }); + sycl_parallel_for(stream, sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), + [=](sycl::nd_item<3> item_ct1) { + cpy_f32_q(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, + ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1); + }); } static void ggml_cpy_f16_f16_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01, @@ -609,7 +616,8 @@ static void ggml_cpy_f16_f16_sycl(const char * cx, char * cdst, const int ne, co { dpct::has_capability_or_fail(stream->get_device(), { sycl::aspect::fp16 }); - stream->parallel_for( + sycl_parallel_for( + stream, sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) { @@ -628,7 +636,8 @@ static void ggml_cpy_i16_i16_sycl(const char * cx, char * cdst, const int ne, co // dpct::has_capability_or_fail(stream->get_device(), // {sycl::aspect::fp16}); - stream->parallel_for( + sycl_parallel_for( + stream, sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) { @@ -647,7 +656,8 @@ static void ggml_cpy_i32_i32_sycl(const char * cx, char * cdst, const int ne, co // dpct::has_capability_or_fail(stream->get_device(), // {sycl::aspect::fp16}); - stream->parallel_for( + sycl_parallel_for( + stream, sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) { @@ -662,11 +672,13 @@ static void ggml_cpy_q8_0_q8_0(const char * cx, char * cdst, const int ne, const const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, queue_ptr stream) { const int num_blocks = ceil_div(ne, SYCL_CPY_BLOCK_SIZE); - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), - sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) { - cpy_q_q(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1); - }); + sycl_parallel_for(stream, + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { + cpy_q_q(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, + ne12, nb10, nb11, nb12, nb13, item_ct1); + }); } @@ -675,11 +687,13 @@ static void ggml_cpy_q5_0_q5_0(const char * cx, char * cdst, const int ne, const const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, queue_ptr stream) { const int num_blocks = ceil_div(ne, SYCL_CPY_BLOCK_SIZE); - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), - sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) { - cpy_q_q(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1); - }); + sycl_parallel_for(stream, + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { + cpy_q_q(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, + ne12, nb10, nb11, nb12, nb13, item_ct1); + }); } @@ -689,11 +703,13 @@ static void ggml_cpy_q5_1_q5_1(const char * cx, char * cdst, const int ne, const const int nb12, const int nb13, queue_ptr stream) { const int num_blocks = ceil_div(ne, SYCL_CPY_BLOCK_SIZE); - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), - sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) { - cpy_q_q(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1); - }); + sycl_parallel_for(stream, + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { + cpy_q_q(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, + ne12, nb10, nb11, nb12, nb13, item_ct1); + }); } @@ -702,10 +718,13 @@ static void ggml_cpy_q4_0_q4_0(const char * cx, char * cdst, const int ne, const const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, queue_ptr stream) { const int num_blocks = ceil_div(ne, SYCL_CPY_BLOCK_SIZE); - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) { - cpy_q_q(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1); - }); + sycl_parallel_for(stream, + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { + cpy_q_q(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, + ne12, nb10, nb11, nb12, nb13, item_ct1); + }); } @@ -715,10 +734,13 @@ static void ggml_cpy_q4_1_q4_1(const char * cx, char * cdst, const int ne, const const int nb12, const int nb13, queue_ptr stream) { const int num_blocks = ceil_div(ne, SYCL_CPY_BLOCK_SIZE); - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) { - cpy_q_q(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1); - }); + sycl_parallel_for(stream, + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { + cpy_q_q(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, + ne12, nb10, nb11, nb12, nb13, item_ct1); + }); } void ggml_sycl_cpy(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1) try { diff --git a/ggml/src/ggml-sycl/dmmv.cpp b/ggml/src/ggml-sycl/dmmv.cpp index 4f2760110..70579c0c3 100644 --- a/ggml/src/ggml-sycl/dmmv.cpp +++ b/ggml/src/ggml-sycl/dmmv.cpp @@ -208,12 +208,10 @@ static void convert_mul_mat_vec_f16_sycl(const void *vx, const dfloat *y, dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { - dequantize_mul_mat_vec<1, 1, convert_f16>(vx, y, dst, ncols, - nrows, item_ct1); - }); + sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + dequantize_mul_mat_vec<1, 1, convert_f16>(vx, y, dst, ncols, nrows, item_ct1); + }); } } @@ -877,12 +875,11 @@ static void dequantize_mul_mat_vec_q4_0_sycl_reorder(const void *vx, const dfloa dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { - dequantize_mul_mat_vec_reorder( - vx, y, dst, ncols, nrows, item_ct1); - }); + sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + dequantize_mul_mat_vec_reorder(vx, y, dst, ncols, + nrows, item_ct1); + }); } } @@ -900,12 +897,10 @@ static void dequantize_mul_mat_vec_q4_0_sycl(const void *vx, const dfloat *y, dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { - dequantize_mul_mat_vec( - vx, y, dst, ncols, nrows, item_ct1); - }); + sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + dequantize_mul_mat_vec(vx, y, dst, ncols, nrows, item_ct1); + }); } } @@ -921,12 +916,10 @@ static void dequantize_mul_mat_vec_q4_1_sycl(const void *vx, const dfloat *y, dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { - dequantize_mul_mat_vec( - vx, y, dst, ncols, nrows, item_ct1); - }); + sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + dequantize_mul_mat_vec(vx, y, dst, ncols, nrows, item_ct1); + }); } } @@ -942,12 +935,10 @@ static void dequantize_mul_mat_vec_q5_0_sycl(const void *vx, const dfloat *y, dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { - dequantize_mul_mat_vec( - vx, y, dst, ncols, nrows, item_ct1); - }); + sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + dequantize_mul_mat_vec(vx, y, dst, ncols, nrows, item_ct1); + }); } } @@ -963,12 +954,10 @@ static void dequantize_mul_mat_vec_q5_1_sycl(const void *vx, const dfloat *y, dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { - dequantize_mul_mat_vec( - vx, y, dst, ncols, nrows, item_ct1); - }); + sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + dequantize_mul_mat_vec(vx, y, dst, ncols, nrows, item_ct1); + }); } } @@ -984,12 +973,10 @@ static void dequantize_mul_mat_vec_q8_0_sycl(const void *vx, const dfloat *y, dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { - dequantize_mul_mat_vec( - vx, y, dst, ncols, nrows, item_ct1); - }); + sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + dequantize_mul_mat_vec(vx, y, dst, ncols, nrows, item_ct1); + }); } } @@ -1002,11 +989,10 @@ static void dequantize_mul_mat_vec_q2_K_sycl(const void *vx, const float *y, const int block_num_y = (nrows + ny - 1) / ny; const sycl::range<3> block_nums(1, 1, block_num_y); const sycl::range<3> block_dims(1, ny, QK_WARP_SIZE); - stream->parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(QK_WARP_SIZE)]] { - dequantize_mul_mat_vec_q2_k(vx, y, dst, ncols, nrows, item_ct1); - }); + sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(QK_WARP_SIZE)]] { + dequantize_mul_mat_vec_q2_k(vx, y, dst, ncols, nrows, item_ct1); + }); } static void dequantize_mul_mat_vec_q3_K_sycl(const void *vx, const float *y, @@ -1018,11 +1004,10 @@ static void dequantize_mul_mat_vec_q3_K_sycl(const void *vx, const float *y, const int block_num_y = (nrows + ny - 1) / ny; const sycl::range<3> block_nums(1, 1, block_num_y); const sycl::range<3> block_dims(1, ny, QK_WARP_SIZE); - stream->parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(QK_WARP_SIZE)]] { - dequantize_mul_mat_vec_q3_k(vx, y, dst, ncols, nrows, item_ct1); - }); + sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(QK_WARP_SIZE)]] { + dequantize_mul_mat_vec_q3_k(vx, y, dst, ncols, nrows, item_ct1); + }); } static void dequantize_mul_mat_vec_q4_K_sycl(const void *vx, const float *y, @@ -1034,11 +1019,10 @@ static void dequantize_mul_mat_vec_q4_K_sycl(const void *vx, const float *y, const int block_num_y = (nrows + ny - 1) / ny; const sycl::range<3> block_nums(1, 1, block_num_y); const sycl::range<3> block_dims(1, ny, QK_WARP_SIZE); - stream->parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(QK_WARP_SIZE)]] { - dequantize_mul_mat_vec_q4_k(vx, y, dst, ncols, nrows, item_ct1); - }); + sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(QK_WARP_SIZE)]] { + dequantize_mul_mat_vec_q4_k(vx, y, dst, ncols, nrows, item_ct1); + }); } static void dequantize_mul_mat_vec_q5_K_sycl(const void *vx, const float *y, @@ -1047,11 +1031,10 @@ static void dequantize_mul_mat_vec_q5_K_sycl(const void *vx, const float *y, dpct::queue_ptr stream) { GGML_ASSERT(ncols % QK_K == 0); const sycl::range<3> block_dims(1, 1, QK_WARP_SIZE); - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(QK_WARP_SIZE)]] { - dequantize_mul_mat_vec_q5_k(vx, y, dst, ncols, item_ct1); - }); + sycl_parallel_for(stream, sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(QK_WARP_SIZE)]] { + dequantize_mul_mat_vec_q5_k(vx, y, dst, ncols, item_ct1); + }); } static void dequantize_mul_mat_vec_q6_K_sycl(const void *vx, const float *y, @@ -1063,11 +1046,10 @@ static void dequantize_mul_mat_vec_q6_K_sycl(const void *vx, const float *y, const int block_num_y = (nrows + ny - 1) / ny; const sycl::range<3> block_nums(1, 1, block_num_y); const sycl::range<3> block_dims(1, ny, QK_WARP_SIZE); - stream->parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(QK_WARP_SIZE)]] { - dequantize_mul_mat_vec_q6_k(vx, y, dst, ncols, nrows, item_ct1); - }); + sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(QK_WARP_SIZE)]] { + dequantize_mul_mat_vec_q6_k(vx, y, dst, ncols, nrows, item_ct1); + }); } void ggml_sycl_op_dequantize_mul_mat_vec( diff --git a/ggml/src/ggml-sycl/dpct/helper.hpp b/ggml/src/ggml-sycl/dpct/helper.hpp index d538965b0..27c727860 100644 --- a/ggml/src/ggml-sycl/dpct/helper.hpp +++ b/ggml/src/ggml-sycl/dpct/helper.hpp @@ -13,10 +13,10 @@ #ifndef GGML_SYCL_DPCT_HELPER_HPP #define GGML_SYCL_DPCT_HELPER_HPP +#include #include #include #include -#include #ifdef GGML_SYCL_USE_INTEL_ONEMKL #include @@ -118,6 +118,36 @@ inline auto get_onemath_backend(sycl::queue& queue) #endif } +#ifdef SYCL_EXT_ONEAPI_ENQUEUE_FUNCTIONS + namespace syclex = sycl::ext::oneapi::experimental; +#endif + +template +__dpct_inline__ void sycl_parallel_for(sycl::handler & cgh, sycl::nd_range nd_range, Func && func) { +#ifdef SYCL_EXT_ONEAPI_ENQUEUE_FUNCTIONS + syclex::nd_launch(cgh, nd_range, func); +#else + cgh.parallel_for(nd_range, func); +#endif +} + +template +__dpct_inline__ void sycl_parallel_for(sycl::queue * q, sycl::nd_range nd_range, Func && func) { +#ifdef SYCL_EXT_ONEAPI_ENQUEUE_FUNCTIONS + syclex::nd_launch(*q, nd_range, func); +#else + q->parallel_for(nd_range, func); +#endif +} + +template __dpct_inline__ void sycl_launch(sycl::queue * stream, Func && func) { +#ifdef SYCL_EXT_ONEAPI_ENQUEUE_FUNCTIONS + syclex::submit(*stream, func); +#else + stream->submit(func); +#endif +} + namespace dpct { typedef sycl::queue *queue_ptr; diff --git a/ggml/src/ggml-sycl/element_wise.cpp b/ggml/src/ggml-sycl/element_wise.cpp index 5b7c4f0b4..c56924ce8 100644 --- a/ggml/src/ggml-sycl/element_wise.cpp +++ b/ggml/src/ggml-sycl/element_wise.cpp @@ -329,60 +329,51 @@ static void acc_f32_sycl(const float *x, const float *y, float *dst, const int ne12, const int nb1, const int nb2, const int offset, queue_ptr stream) { int num_blocks = (n_elements + SYCL_ACC_BLOCK_SIZE - 1) / SYCL_ACC_BLOCK_SIZE; - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * - sycl::range<3>(1, 1, SYCL_ACC_BLOCK_SIZE), - sycl::range<3>(1, 1, SYCL_ACC_BLOCK_SIZE)), - [=](sycl::nd_item<3> item_ct1) { - acc_f32(x, y, dst, n_elements, ne10, ne11, ne12, nb1, nb2, offset, - item_ct1); - }); + sycl_parallel_for(stream, + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_ACC_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_ACC_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { + acc_f32(x, y, dst, n_elements, ne10, ne11, ne12, nb1, nb2, offset, item_ct1); + }); } template static void gelu_sycl(const T *x, T *dst, const int k, queue_ptr stream) { const int num_blocks = (k + SYCL_GELU_BLOCK_SIZE - 1) / SYCL_GELU_BLOCK_SIZE; - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * - sycl::range<3>(1, 1, SYCL_GELU_BLOCK_SIZE), - sycl::range<3>(1, 1, SYCL_GELU_BLOCK_SIZE)), - [=](sycl::nd_item<3> item_ct1) { - gelu(x, dst, k, item_ct1); - }); + sycl_parallel_for(stream, + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_GELU_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_GELU_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { gelu(x, dst, k, item_ct1); }); } template static void silu_sycl(const T *x, T *dst, const int k, queue_ptr stream) { const int num_blocks = (k + SYCL_SILU_BLOCK_SIZE - 1) / SYCL_SILU_BLOCK_SIZE; - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * - sycl::range<3>(1, 1, SYCL_SILU_BLOCK_SIZE), - sycl::range<3>(1, 1, SYCL_SILU_BLOCK_SIZE)), - [=](sycl::nd_item<3> item_ct1) { - silu(x, dst, k, item_ct1); - }); + sycl_parallel_for(stream, + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_SILU_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_SILU_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { silu(x, dst, k, item_ct1); }); } template static void sgn_sycl(const T * x, T * dst, const int k, queue_ptr stream) { // hard code for now const int num_blocks = ceil_div(k, 256); - stream->parallel_for( - sycl::nd_range<3>((sycl::range<3>(1, 1, num_blocks) * sycl::range(1, 1, 256)), sycl::range(1, 1, 256)), [=](sycl::nd_item<3> item_ct1) { - sgn(x, dst, k, item_ct1); - }); + sycl_parallel_for( + stream, sycl::nd_range<3>((sycl::range<3>(1, 1, num_blocks) * sycl::range(1, 1, 256)), sycl::range(1, 1, 256)), + [=](sycl::nd_item<3> item_ct1) { sgn(x, dst, k, item_ct1); }); } template static void abs_sycl(const T * x, T * dst, const int k, queue_ptr stream) { // hard code for now const int num_blocks = ceil_div(k, 256); - stream->parallel_for( - sycl::nd_range<3>((sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, 256)), sycl::range<3>(1, 1, 256)), [=](sycl::nd_item<3> item_ct1) { - abs_op(x, dst, k, item_ct1); - }); + sycl_parallel_for( + stream, + sycl::nd_range<3>((sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, 256)), sycl::range<3>(1, 1, 256)), + [=](sycl::nd_item<3> item_ct1) { abs_op(x, dst, k, item_ct1); }); } @@ -390,23 +381,20 @@ template static void elu_sycl(const T * x, T * dst, const int k, queue_ptr stream) { // hard code for now const int num_blocks = ceil_div(k, 256); - stream->parallel_for( - sycl::nd_range<3>((sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, 256)), sycl::range<3>(1, 1, 256)), [=](sycl::nd_item<3> item_ct1) { - elu_op(x, dst, k, item_ct1); - }); + sycl_parallel_for( + stream, + sycl::nd_range<3>((sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, 256)), sycl::range<3>(1, 1, 256)), + [=](sycl::nd_item<3> item_ct1) { elu_op(x, dst, k, item_ct1); }); } template static void gelu_quick_sycl(const T *x, T *dst, const int k, queue_ptr stream) { const int num_blocks = (k + SYCL_GELU_BLOCK_SIZE - 1) / SYCL_GELU_BLOCK_SIZE; - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * - sycl::range<3>(1, 1, SYCL_GELU_BLOCK_SIZE), - sycl::range<3>(1, 1, SYCL_GELU_BLOCK_SIZE)), - [=](sycl::nd_item<3> item_ct1) { - gelu_quick(x, dst, k, item_ct1); - }); + sycl_parallel_for(stream, + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_GELU_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_GELU_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { gelu_quick(x, dst, k, item_ct1); }); } @@ -414,169 +402,133 @@ template static void gelu_erf_sycl(const T *x, T *dst, const int k, queue_ptr stream) { const int num_blocks = ceil_div(k, SYCL_GELU_BLOCK_SIZE); - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * - sycl::range<3>(1, 1, SYCL_GELU_BLOCK_SIZE), - sycl::range<3>(1, 1, SYCL_GELU_BLOCK_SIZE)), - [=](sycl::nd_item<3> item_ct1) { - gelu_erf(x, dst, k, item_ct1); - }); + sycl_parallel_for(stream, + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_GELU_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_GELU_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { gelu_erf(x, dst, k, item_ct1); }); } template static void tanh_sycl(const T *x, T *dst, const int k, queue_ptr stream) { const int num_blocks = (k + SYCL_TANH_BLOCK_SIZE - 1) / SYCL_TANH_BLOCK_SIZE; - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * - sycl::range<3>(1, 1, SYCL_TANH_BLOCK_SIZE), - sycl::range<3>(1, 1, SYCL_TANH_BLOCK_SIZE)), - [=](sycl::nd_item<3> item_ct1) { - tanh(x, dst, k, item_ct1); - }); + sycl_parallel_for(stream, + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_TANH_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_TANH_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { tanh(x, dst, k, item_ct1); }); } template static void relu_sycl(const T *x, T *dst, const int k, queue_ptr stream) { const int num_blocks = (k + SYCL_RELU_BLOCK_SIZE - 1) / SYCL_RELU_BLOCK_SIZE; - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * - sycl::range<3>(1, 1, SYCL_RELU_BLOCK_SIZE), - sycl::range<3>(1, 1, SYCL_RELU_BLOCK_SIZE)), - [=](sycl::nd_item<3> item_ct1) { - relu(x, dst, k, item_ct1); - }); + sycl_parallel_for(stream, + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_RELU_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_RELU_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { relu(x, dst, k, item_ct1); }); } template static void hardsigmoid_sycl(const T *x, T *dst, const int k, queue_ptr stream) { const int num_blocks = (k + SYCL_HARDSIGMOID_BLOCK_SIZE - 1) / SYCL_HARDSIGMOID_BLOCK_SIZE; - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * - sycl::range<3>(1, 1, SYCL_HARDSIGMOID_BLOCK_SIZE), + sycl_parallel_for( + stream, + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_HARDSIGMOID_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_HARDSIGMOID_BLOCK_SIZE)), - [=](sycl::nd_item<3> item_ct1) { - hardsigmoid(x, dst, k, item_ct1); - }); + [=](sycl::nd_item<3> item_ct1) { hardsigmoid(x, dst, k, item_ct1); }); } template static void hardswish_sycl(const T *x, T *dst, const int k, queue_ptr stream) { const int num_blocks = (k + SYCL_HARDSWISH_BLOCK_SIZE - 1) / SYCL_HARDSWISH_BLOCK_SIZE; - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * - sycl::range<3>(1, 1, SYCL_HARDSWISH_BLOCK_SIZE), + sycl_parallel_for( + stream, + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_HARDSWISH_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_HARDSWISH_BLOCK_SIZE)), - [=](sycl::nd_item<3> item_ct1) { - hardswish(x, dst, k, item_ct1); - }); + [=](sycl::nd_item<3> item_ct1) { hardswish(x, dst, k, item_ct1); }); } template static void exp_sycl(const T *x, T *dst, const int k, queue_ptr stream) { const int num_blocks = (k + SYCL_EXP_BLOCK_SIZE - 1) / SYCL_EXP_BLOCK_SIZE; - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * - sycl::range<3>(1, 1, SYCL_EXP_BLOCK_SIZE), - sycl::range<3>(1, 1, SYCL_EXP_BLOCK_SIZE)), - [=](sycl::nd_item<3> item_ct1) { - exp(x, dst, k, item_ct1); - }); + sycl_parallel_for(stream, + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_EXP_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_EXP_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { exp(x, dst, k, item_ct1); }); } template static void log_sycl(const T *x, T *dst, const int k, queue_ptr stream) { const int num_blocks = (k + SYCL_EXP_BLOCK_SIZE - 1) / SYCL_EXP_BLOCK_SIZE; - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * - sycl::range<3>(1, 1, SYCL_EXP_BLOCK_SIZE), - sycl::range<3>(1, 1, SYCL_EXP_BLOCK_SIZE)), - [=](sycl::nd_item<3> item_ct1) { - log(x, dst, k, item_ct1); - }); + sycl_parallel_for(stream, + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_EXP_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_EXP_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { log(x, dst, k, item_ct1); }); } template static void neg_sycl(const T *x, T *dst, const int k, queue_ptr stream) { const int num_blocks = (k + SYCL_NEG_BLOCK_SIZE - 1) / SYCL_NEG_BLOCK_SIZE; - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * - sycl::range<3>(1, 1, SYCL_NEG_BLOCK_SIZE), - sycl::range<3>(1, 1, SYCL_NEG_BLOCK_SIZE)), - [=](sycl::nd_item<3> item_ct1) { - neg(x, dst, k, item_ct1); - }); + sycl_parallel_for(stream, + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_NEG_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_NEG_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { neg(x, dst, k, item_ct1); }); } template static void step_sycl(const T *x, T *dst, const int k, queue_ptr stream) { const int num_blocks = (k + SYCL_NEG_BLOCK_SIZE - 1) / SYCL_NEG_BLOCK_SIZE; - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * - sycl::range<3>(1, 1, SYCL_NEG_BLOCK_SIZE), - sycl::range<3>(1, 1, SYCL_NEG_BLOCK_SIZE)), - [=](sycl::nd_item<3> item_ct1) { - step(x, dst, k, item_ct1); - }); + sycl_parallel_for(stream, + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_NEG_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_NEG_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { step(x, dst, k, item_ct1); }); } template static void sigmoid_sycl(const T *x, T *dst, const int k, queue_ptr stream) { const int num_blocks = (k + SYCL_SIGMOID_BLOCK_SIZE - 1) / SYCL_SIGMOID_BLOCK_SIZE; - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * - sycl::range<3>(1, 1, SYCL_SIGMOID_BLOCK_SIZE), + sycl_parallel_for( + stream, + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_SIGMOID_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_SIGMOID_BLOCK_SIZE)), - [=](sycl::nd_item<3> item_ct1) { - sigmoid(x, dst, k, item_ct1); - }); + [=](sycl::nd_item<3> item_ct1) { sigmoid(x, dst, k, item_ct1); }); } template static void sqrt_sycl(const T *x, T *dst, const int k, queue_ptr stream) { const int num_blocks = (k + SYCL_SQRT_BLOCK_SIZE - 1) / SYCL_SQRT_BLOCK_SIZE; - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * - sycl::range<3>(1, 1, SYCL_SQRT_BLOCK_SIZE), - sycl::range<3>(1, 1, SYCL_SQRT_BLOCK_SIZE)), - [=](sycl::nd_item<3> item_ct1) { - sqrt(x, dst, k, item_ct1); - }); + sycl_parallel_for(stream, + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_SQRT_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_SQRT_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { sqrt(x, dst, k, item_ct1); }); } template static void sin_sycl(const T *x, T *dst, const int k, queue_ptr stream) { const int num_blocks = (k + SYCL_SIN_BLOCK_SIZE - 1) / SYCL_SIN_BLOCK_SIZE; - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * - sycl::range<3>(1, 1, SYCL_SIN_BLOCK_SIZE), - sycl::range<3>(1, 1, SYCL_SIN_BLOCK_SIZE)), - [=](sycl::nd_item<3> item_ct1) { - sin(x, dst, k, item_ct1); - }); + sycl_parallel_for(stream, + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_SIN_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_SIN_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { sin(x, dst, k, item_ct1); }); } template static void cos_sycl(const T *x, T *dst, const int k, queue_ptr stream) { const int num_blocks = (k + SYCL_SIN_BLOCK_SIZE - 1) / SYCL_SIN_BLOCK_SIZE; - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * - sycl::range<3>(1, 1, SYCL_SIN_BLOCK_SIZE), - sycl::range<3>(1, 1, SYCL_SIN_BLOCK_SIZE)), - [=](sycl::nd_item<3> item_ct1) { - cos(x, dst, k, item_ct1); - }); + sycl_parallel_for(stream, + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_SIN_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_SIN_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { cos(x, dst, k, item_ct1); }); } template @@ -584,26 +536,20 @@ static void leaky_relu_sycl(const T *x, T *dst, const int k, const float negative_slope, queue_ptr stream) { const int num_blocks = (k + SYCL_RELU_BLOCK_SIZE - 1) / SYCL_RELU_BLOCK_SIZE; - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * - sycl::range<3>(1, 1, SYCL_RELU_BLOCK_SIZE), - sycl::range<3>(1, 1, SYCL_RELU_BLOCK_SIZE)), - [=](sycl::nd_item<3> item_ct1) { - leaky_relu(x, dst, k, negative_slope, item_ct1); - }); + sycl_parallel_for(stream, + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_RELU_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_RELU_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { leaky_relu(x, dst, k, negative_slope, item_ct1); }); } template static void sqr_sycl(const T *x, T *dst, const int k, queue_ptr stream) { const int num_blocks = (k + SYCL_SQR_BLOCK_SIZE - 1) / SYCL_SQR_BLOCK_SIZE; - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * - sycl::range<3>(1, 1, SYCL_SQR_BLOCK_SIZE), - sycl::range<3>(1, 1, SYCL_SQR_BLOCK_SIZE)), - [=](sycl::nd_item<3> item_ct1) { - sqr(x, dst, k, item_ct1); - }); + sycl_parallel_for(stream, + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_SQR_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_SQR_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { sqr(x, dst, k, item_ct1); }); } template @@ -614,9 +560,8 @@ static void upscale_sycl(const T *x, T *dst, const int nb00, const int nb01, int dst_size = ne10 * ne11 * ne12 * ne13; int num_blocks = (dst_size + SYCL_UPSCALE_BLOCK_SIZE - 1) / SYCL_UPSCALE_BLOCK_SIZE; sycl::range<1> gridDim(num_blocks * SYCL_UPSCALE_BLOCK_SIZE); - stream->parallel_for( - sycl::nd_range<1>(gridDim, sycl::range<1>(SYCL_UPSCALE_BLOCK_SIZE)), - [=](sycl::nd_item<1> item_ct1) { + sycl_parallel_for<1>( + stream, sycl::nd_range<1>(gridDim, sycl::range<1>(SYCL_UPSCALE_BLOCK_SIZE)), [=](sycl::nd_item<1> item_ct1) { upscale(x, dst, nb00, nb01, nb02, nb03, ne10, ne11, ne12, ne13, sf0, sf1, sf2, sf3, item_ct1); }); } @@ -627,12 +572,10 @@ static void pad_sycl(const T *x, T *dst, const int ne00, const int ne1, const int ne2, queue_ptr stream) { int num_blocks = (ne0 + SYCL_PAD_BLOCK_SIZE - 1) / SYCL_PAD_BLOCK_SIZE; sycl::range<3> gridDim(ne2, ne1, num_blocks); - stream->parallel_for( - sycl::nd_range<3>(gridDim * sycl::range<3>(1, 1, SYCL_PAD_BLOCK_SIZE), - sycl::range<3>(1, 1, SYCL_PAD_BLOCK_SIZE)), - [=](sycl::nd_item<3> item_ct1) { - pad(x, dst, ne0, ne00, ne01, ne02, item_ct1); - }); + sycl_parallel_for(stream, + sycl::nd_range<3>(gridDim * sycl::range<3>(1, 1, SYCL_PAD_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_PAD_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { pad(x, dst, ne0, ne00, ne01, ne02, item_ct1); }); } template @@ -640,13 +583,10 @@ static void clamp_sycl(const T *x, T *dst, const float min, const float max, const int k, queue_ptr stream) { const int num_blocks = (k + SYCL_CLAMP_BLOCK_SIZE - 1) / SYCL_CLAMP_BLOCK_SIZE; - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * - sycl::range<3>(1, 1, SYCL_CLAMP_BLOCK_SIZE), - sycl::range<3>(1, 1, SYCL_CLAMP_BLOCK_SIZE)), - [=](sycl::nd_item<3> item_ct1) { - clamp(x, dst, min, max, k, item_ct1); - }); + sycl_parallel_for(stream, + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CLAMP_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_CLAMP_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { clamp(x, dst, min, max, k, item_ct1); }); } inline void ggml_sycl_op_sgn(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { diff --git a/ggml/src/ggml-sycl/getrows.cpp b/ggml/src/ggml-sycl/getrows.cpp index 03f8dd907..9c76ffeb9 100644 --- a/ggml/src/ggml-sycl/getrows.cpp +++ b/ggml/src/ggml-sycl/getrows.cpp @@ -118,12 +118,10 @@ static void get_rows_sycl(ggml_backend_sycl_context & ctx, const ggml_tensor *sr GGML_ASSERT(ne00 % 2 == 0); - stream->parallel_for(sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { - k_get_rows( - src0_dd, src1_dd, dst_dd, ne00, ne12, s1, s2, - s3, nb01, nb02, nb03, s10, s11, s12, item_ct1); - }); + sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { + k_get_rows(src0_dd, src1_dd, dst_dd, ne00, ne12, s1, s2, s3, nb01, nb02, nb03, s10, s11, s12, + item_ct1); + }); GGML_UNUSED(dst); GGML_UNUSED(ctx); @@ -156,9 +154,8 @@ static void get_rows_sycl_float(ggml_backend_sycl_context & ctx, const ggml_tens dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for( + stream, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { k_get_rows_float(src0_dd, src1_dd, dst_dd, ne00, ne12, s1, s2, s3, nb01, nb02, nb03, s10, s11, s12, item_ct1); }); diff --git a/ggml/src/ggml-sycl/ggml-sycl.cpp b/ggml/src/ggml-sycl/ggml-sycl.cpp index 4b7610362..f25a96a62 100644 --- a/ggml/src/ggml-sycl/ggml-sycl.cpp +++ b/ggml/src/ggml-sycl/ggml-sycl.cpp @@ -1887,13 +1887,12 @@ static void argsort_f32_i32_sycl(const float *x, int *dst, const int ncols, const size_t shared_mem = ncols_pad * sizeof(int); if (order == GGML_SORT_ORDER_ASC) { - stream->submit([&](sycl::handler &cgh) { + sycl_launch(stream, [&](sycl::handler & cgh) { sycl::local_accessor dpct_local_acc_ct1( sycl::range<1>(shared_mem), cgh); - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for( + cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { k_argsort_f32_i32( x, dst, ncols, ncols_pad, item_ct1, dpct_local_acc_ct1.get_multi_ptr() @@ -1901,13 +1900,12 @@ static void argsort_f32_i32_sycl(const float *x, int *dst, const int ncols, }); }); } else if (order == GGML_SORT_ORDER_DESC) { - stream->submit([&](sycl::handler &cgh) { + sycl_launch(stream, [&](sycl::handler & cgh) { sycl::local_accessor dpct_local_acc_ct1( sycl::range<1>(shared_mem), cgh); - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for( + cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { k_argsort_f32_i32( x, dst, ncols, ncols_pad, item_ct1, dpct_local_acc_ct1.get_multi_ptr() @@ -1925,50 +1923,47 @@ static void argmax_f32_i32_sycl(const float *x, int *dst, const int ncols, const sycl::range<3> block_nums(1, nrows, 1); const size_t shared_mem = 256 * sizeof(float); - stream->submit([&](sycl::handler &cgh) { + sycl_launch(stream, [&](sycl::handler & cgh) { sycl::local_accessor shared_data( sycl::range<1>(shared_mem/sizeof(float)), cgh); sycl::local_accessor shared_indices( sycl::range<1>(shared_mem/sizeof(float)), cgh); - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { - const int tid = item_ct1.get_local_id(2); - const int row = item_ct1.get_global_id(1); + sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { + const int tid = item_ct1.get_local_id(2); + const int row = item_ct1.get_global_id(1); - float max_val = -INFINITY; - int max_idx = -1; + float max_val = -INFINITY; + int max_idx = -1; - for (int col = tid; col < ncols; col += 256) { - float val = x[row * ncols + col]; - if (val > max_val) { - max_val = val; - max_idx = col; + for (int col = tid; col < ncols; col += 256) { + float val = x[row * ncols + col]; + if (val > max_val) { + max_val = val; + max_idx = col; + } + } + + shared_data[tid] = max_val; + shared_indices[tid] = max_idx; + item_ct1.barrier(sycl::access::fence_space::local_space); + + for (int stride = 256 / 2; stride > 0; stride >>= 1) { + if (tid < stride) { + float val1 = shared_data[tid]; + float val2 = shared_data[tid + stride]; + if (val2 > val1) { + shared_data[tid] = val2; + shared_indices[tid] = shared_indices[tid + stride]; } } - - shared_data[tid] = max_val; - shared_indices[tid] = max_idx; item_ct1.barrier(sycl::access::fence_space::local_space); + } - for (int stride = 256/2; stride > 0; stride >>= 1) { - if (tid < stride) { - float val1 = shared_data[tid]; - float val2 = shared_data[tid + stride]; - if (val2 > val1) { - shared_data[tid] = val2; - shared_indices[tid] = shared_indices[tid + stride]; - } - } - item_ct1.barrier(sycl::access::fence_space::local_space); - } - - - if (tid == 0) { - dst[row] = shared_indices[0]; - } - }); + if (tid == 0) { + dst[row] = shared_indices[0]; + } + }); }); } static void diag_mask_inf_f32_sycl(const float *x, float *dst, @@ -2952,7 +2947,7 @@ static void ggml_sycl_mul_mat_batched_sycl(ggml_backend_sycl_context & ctx, cons void ** ptrs_dst_get = ptrs_dst.get(); size_t nb12_scaled = src1->type == GGML_TYPE_F16 ? nb12 : s12 * sizeof(sycl::half); size_t nb13_scaled = src1->type == GGML_TYPE_F16 ? nb13 : s13 * sizeof(sycl::half); - cgh.parallel_for(sycl::nd_range<3>(block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for(cgh, sycl::nd_range<3>(block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { k_compute_batched_ptrs(src0_f16, src1_f16, dst_ddf, ptrs_src_get, ptrs_dst_get, ne12, ne13, ne23, nb02, nb03, nb12_scaled, nb13_scaled, nbd2, nbd3, r2, r3, item_ct1); }); @@ -3456,7 +3451,7 @@ static void ggml_sycl_mul_mat_id(ggml_backend_sycl_context & ctx, { sycl::range<3> block_dims(1, 1, std::min((unsigned int)ne10, 768u)); sycl::range<3> grid_dims(1, n_ids, ids->ne[1]); - stream->submit([&](sycl::handler &cgh) { + sycl_launch(stream, [&](sycl::handler & cgh) { sycl::local_accessor src1_row_acc(cgh); char *__restrict src1_contiguous_get = @@ -3468,9 +3463,8 @@ static void ggml_sycl_mul_mat_id(ggml_backend_sycl_context & ctx, size_t ids_nb_ct6 = ids->nb[1]; size_t ids_nb_ct7 = ids->nb[0]; - cgh.parallel_for( - sycl::nd_range<3>(grid_dims * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for( + cgh, sycl::nd_range<3>(grid_dims * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { k_copy_src1_to_contiguous( src1_original, src1_contiguous_get, dev_cur_src1_row_get, @@ -3501,15 +3495,14 @@ static void ggml_sycl_mul_mat_id(ggml_backend_sycl_context & ctx, { sycl::range<3> block_dims(1, 1, std::min((unsigned int)ne0, 768u)); sycl::range<3> grid_dims(1, 1, num_src1_rows); - stream->submit([&](sycl::handler &cgh) { + sycl_launch(stream, [&](sycl::handler & cgh) { const char *__restrict dst_contiguous_get = dst_contiguous.get(); const mmid_row_mapping *__restrict dev_row_mapping_get = dev_row_mapping.get(); - cgh.parallel_for( - sycl::nd_range<3>(grid_dims * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for( + cgh, sycl::nd_range<3>(grid_dims * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { k_copy_dst_from_contiguous(dst_original, dst_contiguous_get, dev_row_mapping_get, diff --git a/ggml/src/ggml-sycl/gla.cpp b/ggml/src/ggml-sycl/gla.cpp index 879184fdd..b40cbf1f1 100644 --- a/ggml/src/ggml-sycl/gla.cpp +++ b/ggml/src/ggml-sycl/gla.cpp @@ -11,13 +11,13 @@ static void gated_linear_attn_f32_kernel(const dpct::queue_ptr stream, u_int B, const u_int n_seq_tokens = T / B; sycl::range<1> block_dims((C / H)); sycl::range<1> grid_dims((B * H)); - stream->submit([&](sycl::handler & cgh) { + sycl_launch(stream, [&](sycl::handler & cgh) { /* local memory accessors*/ auto _k = sycl::local_accessor(sycl::range<1>(head_size), cgh); auto _r = sycl::local_accessor(sycl::range<1>(head_size), cgh); auto _td = sycl::local_accessor(sycl::range<1>(head_size), cgh); - cgh.parallel_for(sycl::nd_range<1>(grid_dims * block_dims, block_dims), [=](sycl::nd_item<1> item) { + sycl_parallel_for<1>(cgh, sycl::nd_range<1>(grid_dims * block_dims, block_dims), [=](sycl::nd_item<1> item) { u_int tid = item.get_local_id(0); u_int bid = item.get_group(0); diff --git a/ggml/src/ggml-sycl/im2col.cpp b/ggml/src/ggml-sycl/im2col.cpp index aa19c2527..52737cc74 100644 --- a/ggml/src/ggml-sycl/im2col.cpp +++ b/ggml/src/ggml-sycl/im2col.cpp @@ -70,7 +70,7 @@ static void im2col_sycl_internal(const float * x, T * dst, int64_t IW, int64_t I const int64_t CHW = IC * KH * KW; - stream->parallel_for(sycl::nd_range<3>(block_nums * local_range, local_range), [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * local_range, local_range), [=](sycl::nd_item<3> item_ct1) { im2col_kernel(x, dst, batch_offset, offset_delta, IC, IW, IH, OH, OW, KW, KH, parallel_elements, CHW, s0, s1, p0, p1, d0, d1, item_ct1); }); diff --git a/ggml/src/ggml-sycl/mmq.cpp b/ggml/src/ggml-sycl/mmq.cpp index ffb272aa2..c72fcd38e 100644 --- a/ggml/src/ggml-sycl/mmq.cpp +++ b/ggml/src/ggml-sycl/mmq.cpp @@ -1818,7 +1818,7 @@ static void ggml_mul_mat_q4_0_q8_1_sycl(const void *vx, const void *vy, dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->submit([&](sycl::handler &cgh) { + sycl_launch(stream, [&](sycl::handler & cgh) { sycl::local_accessor tile_x_qs_q4_0_acc_ct1( sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh); sycl::local_accessor tile_x_d_q4_0_acc_ct1( @@ -1829,9 +1829,8 @@ static void ggml_mul_mat_q4_0_q8_1_sycl(const void *vx, const void *vy, sycl::local_accessor tile_y_ds_acc_ct1( sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh); - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for( + cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { mul_mat_q4_0( vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, item_ct1, @@ -1853,7 +1852,7 @@ static void ggml_mul_mat_q4_0_q8_1_sycl(const void *vx, const void *vy, dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->submit([&](sycl::handler &cgh) { + sycl_launch(stream, [&](sycl::handler & cgh) { sycl::local_accessor tile_x_qs_q4_0_acc_ct1( sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh); sycl::local_accessor tile_x_d_q4_0_acc_ct1( @@ -1864,9 +1863,8 @@ static void ggml_mul_mat_q4_0_q8_1_sycl(const void *vx, const void *vy, sycl::local_accessor tile_y_ds_acc_ct1( sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh); - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for( + cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { mul_mat_q4_0( vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, item_ct1, @@ -1933,7 +1931,7 @@ static void ggml_mul_mat_q4_1_q8_1_sycl(const void *vx, const void *vy, dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->submit([&](sycl::handler &cgh) { + sycl_launch(stream, [&](sycl::handler & cgh) { sycl::local_accessor tile_x_qs_q4_1_acc_ct1( sycl::range<1>(mmq_y * (WARP_SIZE) + +mmq_y), cgh); sycl::local_accessor tile_x_dm_q4_1_acc_ct1( @@ -1944,9 +1942,8 @@ static void ggml_mul_mat_q4_1_q8_1_sycl(const void *vx, const void *vy, sycl::local_accessor tile_y_ds_acc_ct1( sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh); - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for( + cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { mul_mat_q4_1( vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, item_ct1, @@ -1968,7 +1965,7 @@ static void ggml_mul_mat_q4_1_q8_1_sycl(const void *vx, const void *vy, dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->submit([&](sycl::handler &cgh) { + sycl_launch(stream, [&](sycl::handler & cgh) { sycl::local_accessor tile_x_qs_q4_1_acc_ct1( sycl::range<1>(mmq_y * (WARP_SIZE) + +mmq_y), cgh); sycl::local_accessor tile_x_dm_q4_1_acc_ct1( @@ -1979,9 +1976,8 @@ static void ggml_mul_mat_q4_1_q8_1_sycl(const void *vx, const void *vy, sycl::local_accessor tile_y_ds_acc_ct1( sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh); - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for( + cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { mul_mat_q4_1( vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, item_ct1, @@ -2048,7 +2044,7 @@ static void ggml_mul_mat_q5_0_q8_1_sycl(const void *vx, const void *vy, dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->submit([&](sycl::handler &cgh) { + sycl_launch(stream, [&](sycl::handler & cgh) { sycl::local_accessor tile_x_ql_q5_0_acc_ct1( sycl::range<1>(mmq_y * (2 * WARP_SIZE) + mmq_y), cgh); sycl::local_accessor tile_x_d_q5_0_acc_ct1( @@ -2059,9 +2055,8 @@ static void ggml_mul_mat_q5_0_q8_1_sycl(const void *vx, const void *vy, sycl::local_accessor tile_y_ds_acc_ct1( sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh); - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for( + cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { mul_mat_q5_0( vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, item_ct1, @@ -2083,7 +2078,7 @@ static void ggml_mul_mat_q5_0_q8_1_sycl(const void *vx, const void *vy, dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->submit([&](sycl::handler &cgh) { + sycl_launch(stream, [&](sycl::handler & cgh) { sycl::local_accessor tile_x_ql_q5_0_acc_ct1( sycl::range<1>(mmq_y * (2 * WARP_SIZE) + mmq_y), cgh); sycl::local_accessor tile_x_d_q5_0_acc_ct1( @@ -2094,9 +2089,8 @@ static void ggml_mul_mat_q5_0_q8_1_sycl(const void *vx, const void *vy, sycl::local_accessor tile_y_ds_acc_ct1( sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh); - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for( + cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { mul_mat_q5_0( vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, item_ct1, @@ -2163,7 +2157,7 @@ static void ggml_mul_mat_q5_1_q8_1_sycl(const void *vx, const void *vy, dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->submit([&](sycl::handler &cgh) { + sycl_launch(stream, [&](sycl::handler & cgh) { sycl::local_accessor tile_x_ql_q5_1_acc_ct1( sycl::range<1>(mmq_y * (2 * WARP_SIZE) + mmq_y), cgh); sycl::local_accessor tile_x_dm_q5_1_acc_ct1( @@ -2174,9 +2168,8 @@ static void ggml_mul_mat_q5_1_q8_1_sycl(const void *vx, const void *vy, sycl::local_accessor tile_y_ds_acc_ct1( sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh); - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for( + cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { mul_mat_q5_1( vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, item_ct1, @@ -2198,7 +2191,7 @@ static void ggml_mul_mat_q5_1_q8_1_sycl(const void *vx, const void *vy, dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->submit([&](sycl::handler &cgh) { + sycl_launch(stream, [&](sycl::handler & cgh) { sycl::local_accessor tile_x_ql_q5_1_acc_ct1( sycl::range<1>(mmq_y * (2 * WARP_SIZE) + mmq_y), cgh); sycl::local_accessor tile_x_dm_q5_1_acc_ct1( @@ -2209,9 +2202,8 @@ static void ggml_mul_mat_q5_1_q8_1_sycl(const void *vx, const void *vy, sycl::local_accessor tile_y_ds_acc_ct1( sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh); - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for( + cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { mul_mat_q5_1( vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, item_ct1, @@ -2278,7 +2270,7 @@ static void ggml_mul_mat_q8_0_q8_1_sycl(const void *vx, const void *vy, dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->submit([&](sycl::handler &cgh) { + sycl_launch(stream, [&](sycl::handler & cgh) { sycl::local_accessor tile_x_qs_q8_0_acc_ct1( sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh); sycl::local_accessor tile_x_d_q8_0_acc_ct1( @@ -2289,9 +2281,8 @@ static void ggml_mul_mat_q8_0_q8_1_sycl(const void *vx, const void *vy, sycl::local_accessor tile_y_ds_acc_ct1( sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh); - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for( + cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { mul_mat_q8_0( vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, item_ct1, @@ -2313,7 +2304,7 @@ static void ggml_mul_mat_q8_0_q8_1_sycl(const void *vx, const void *vy, dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->submit([&](sycl::handler &cgh) { + sycl_launch(stream, [&](sycl::handler & cgh) { sycl::local_accessor tile_x_qs_q8_0_acc_ct1( sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh); sycl::local_accessor tile_x_d_q8_0_acc_ct1( @@ -2324,9 +2315,8 @@ static void ggml_mul_mat_q8_0_q8_1_sycl(const void *vx, const void *vy, sycl::local_accessor tile_y_ds_acc_ct1( sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh); - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for( + cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { mul_mat_q8_0( vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, item_ct1, @@ -2393,7 +2383,7 @@ static void ggml_mul_mat_q2_K_q8_1_sycl(const void *vx, const void *vy, dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->submit([&](sycl::handler &cgh) { + sycl_launch(stream, [&](sycl::handler & cgh) { sycl::local_accessor tile_x_ql_q2_K_acc_ct1( sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh); sycl::local_accessor tile_x_dm_q2_K_acc_ct1( @@ -2406,9 +2396,8 @@ static void ggml_mul_mat_q2_K_q8_1_sycl(const void *vx, const void *vy, sycl::local_accessor tile_y_ds_acc_ct1( sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh); - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for( + cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { mul_mat_q2_K( vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, item_ct1, @@ -2431,7 +2420,7 @@ static void ggml_mul_mat_q2_K_q8_1_sycl(const void *vx, const void *vy, dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->submit([&](sycl::handler &cgh) { + sycl_launch(stream, [&](sycl::handler & cgh) { sycl::local_accessor tile_x_ql_q2_K_acc_ct1( sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh); sycl::local_accessor tile_x_dm_q2_K_acc_ct1( @@ -2444,9 +2433,8 @@ static void ggml_mul_mat_q2_K_q8_1_sycl(const void *vx, const void *vy, sycl::local_accessor tile_y_ds_acc_ct1( sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh); - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for( + cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { mul_mat_q2_K( vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, item_ct1, @@ -2516,7 +2504,7 @@ static void ggml_mul_mat_q3_K_q8_1_sycl(const void *vx, const void *vy, dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->submit([&](sycl::handler &cgh) { + sycl_launch(stream, [&](sycl::handler & cgh) { sycl::local_accessor tile_x_ql_q3_K_acc_ct1( sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh); sycl::local_accessor tile_x_dm_q3_K_acc_ct1( @@ -2531,9 +2519,8 @@ static void ggml_mul_mat_q3_K_q8_1_sycl(const void *vx, const void *vy, sycl::local_accessor tile_y_ds_acc_ct1( sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh); - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for( + cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { mul_mat_q3_K( vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, item_ct1, @@ -2557,7 +2544,7 @@ static void ggml_mul_mat_q3_K_q8_1_sycl(const void *vx, const void *vy, dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->submit([&](sycl::handler &cgh) { + sycl_launch(stream, [&](sycl::handler & cgh) { sycl::local_accessor tile_x_ql_q3_K_acc_ct1( sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh); sycl::local_accessor tile_x_dm_q3_K_acc_ct1( @@ -2572,9 +2559,8 @@ static void ggml_mul_mat_q3_K_q8_1_sycl(const void *vx, const void *vy, sycl::local_accessor tile_y_ds_acc_ct1( sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh); - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for( + cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { mul_mat_q3_K( vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, item_ct1, @@ -2644,7 +2630,7 @@ static void ggml_mul_mat_q4_K_q8_1_sycl(const void *vx, const void *vy, dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->submit([&](sycl::handler &cgh) { + sycl_launch(stream, [&](sycl::handler & cgh) { sycl::local_accessor tile_x_ql_q4_K_acc_ct1( sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh); sycl::local_accessor tile_x_dm_q4_K_acc_ct1( @@ -2657,9 +2643,8 @@ static void ggml_mul_mat_q4_K_q8_1_sycl(const void *vx, const void *vy, sycl::local_accessor tile_y_ds_acc_ct1( sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh); - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for( + cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { mul_mat_q4_K( vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, item_ct1, @@ -2682,7 +2667,7 @@ static void ggml_mul_mat_q4_K_q8_1_sycl(const void *vx, const void *vy, dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->submit([&](sycl::handler &cgh) { + sycl_launch(stream, [&](sycl::handler & cgh) { sycl::local_accessor tile_x_ql_q4_K_acc_ct1( sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh); sycl::local_accessor tile_x_dm_q4_K_acc_ct1( @@ -2695,9 +2680,8 @@ static void ggml_mul_mat_q4_K_q8_1_sycl(const void *vx, const void *vy, sycl::local_accessor tile_y_ds_acc_ct1( sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh); - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for( + cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { mul_mat_q4_K( vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, item_ct1, @@ -2765,7 +2749,7 @@ static void ggml_mul_mat_q5_K_q8_1_sycl(const void *vx, const void *vy, dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->submit([&](sycl::handler &cgh) { + sycl_launch(stream, [&](sycl::handler & cgh) { sycl::local_accessor tile_x_ql_q5_K_acc_ct1( sycl::range<1>(mmq_y * (2 * WARP_SIZE) + mmq_y), cgh); sycl::local_accessor tile_x_dm_q5_K_acc_ct1( @@ -2778,9 +2762,8 @@ static void ggml_mul_mat_q5_K_q8_1_sycl(const void *vx, const void *vy, sycl::local_accessor tile_y_ds_acc_ct1( sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh); - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for( + cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { mul_mat_q5_K( vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, item_ct1, @@ -2803,7 +2786,7 @@ static void ggml_mul_mat_q5_K_q8_1_sycl(const void *vx, const void *vy, dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->submit([&](sycl::handler &cgh) { + sycl_launch(stream, [&](sycl::handler & cgh) { sycl::local_accessor tile_x_ql_q5_K_acc_ct1( sycl::range<1>(mmq_y * (2 * WARP_SIZE) + mmq_y), cgh); sycl::local_accessor tile_x_dm_q5_K_acc_ct1( @@ -2816,9 +2799,8 @@ static void ggml_mul_mat_q5_K_q8_1_sycl(const void *vx, const void *vy, sycl::local_accessor tile_y_ds_acc_ct1( sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh); - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for( + cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { mul_mat_q5_K( vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, item_ct1, @@ -2886,7 +2868,7 @@ static void ggml_mul_mat_q6_K_q8_1_sycl(const void *vx, const void *vy, dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->submit([&](sycl::handler &cgh) { + sycl_launch(stream, [&](sycl::handler & cgh) { sycl::local_accessor tile_x_ql_acc_ct1( sycl::range<1>(mmq_y * (2 * WARP_SIZE) + mmq_y), cgh); sycl::local_accessor tile_x_dm_acc_ct1( @@ -2899,9 +2881,8 @@ static void ggml_mul_mat_q6_K_q8_1_sycl(const void *vx, const void *vy, sycl::local_accessor tile_y_ds_acc_ct1( sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh); - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for( + cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { mul_mat_q6_K( vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, item_ct1, @@ -2924,7 +2905,7 @@ static void ggml_mul_mat_q6_K_q8_1_sycl(const void *vx, const void *vy, dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->submit([&](sycl::handler &cgh) { + sycl_launch(stream, [&](sycl::handler & cgh) { sycl::local_accessor tile_x_ql_acc_ct1( sycl::range<1>(mmq_y * (2 * WARP_SIZE) + mmq_y), cgh); sycl::local_accessor tile_x_dm_acc_ct1( @@ -2937,9 +2918,8 @@ static void ggml_mul_mat_q6_K_q8_1_sycl(const void *vx, const void *vy, sycl::local_accessor tile_y_ds_acc_ct1( sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh); - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for( + cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { mul_mat_q6_K( vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, item_ct1, diff --git a/ggml/src/ggml-sycl/mmvq.cpp b/ggml/src/ggml-sycl/mmvq.cpp index 5b7f06407..c21929d51 100644 --- a/ggml/src/ggml-sycl/mmvq.cpp +++ b/ggml/src/ggml-sycl/mmvq.cpp @@ -544,12 +544,12 @@ static void reorder_mul_mat_vec_q4_0_q8_1_sycl(const void * vx, const void * vy, const sycl::range<3> global_size(1, GGML_SYCL_MMV_Y, (block_num_y * WARP_SIZE)); const sycl::range<3> workgroup_size(1, GGML_SYCL_MMV_Y, num_subgroups * WARP_SIZE); - stream->submit([&](sycl::handler & cgh) { - cgh.parallel_for(sycl::nd_range<3>(global_size, workgroup_size), - [=](sycl::nd_item<3> nd_item) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { - mul_mat_vec_q_reorder>(vx, vy, dst, ncols, nrows, - nd_item); - }); + sycl_launch(stream, [&](sycl::handler & cgh) { + sycl_parallel_for(cgh, sycl::nd_range<3>(global_size, workgroup_size), + [=](sycl::nd_item<3> nd_item) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + mul_mat_vec_q_reorder>(vx, vy, dst, ncols, nrows, + nd_item); + }); }); } @@ -561,12 +561,12 @@ static void mul_mat_vec_q4_0_q8_1_sycl(const void * vx, const void * vy, float * const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); { - stream->submit([&](sycl::handler & cgh) { - cgh.parallel_for(sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { - mul_mat_vec_q( - vx, vy, dst, ncols, nrows, item_ct1); - }); + sycl_launch(stream, [&](sycl::handler & cgh) { + sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + mul_mat_vec_q( + vx, vy, dst, ncols, nrows, item_ct1); + }); }); } } @@ -580,17 +580,12 @@ static void mul_mat_vec_q4_1_q8_1_sycl(const void *vx, const void *vy, const sycl::range<3> block_nums(1, 1, block_num_y); const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); { - - stream->submit([&](sycl::handler &cgh) { - - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) - [[sycl::reqd_sub_group_size(WARP_SIZE)]] { - mul_mat_vec_q( - vx, vy, dst, ncols, nrows, item_ct1); - }); + sycl_launch(stream, [&](sycl::handler & cgh) { + sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + mul_mat_vec_q( + vx, vy, dst, ncols, nrows, item_ct1); + }); }); } } @@ -604,17 +599,12 @@ static void mul_mat_vec_q5_0_q8_1_sycl(const void *vx, const void *vy, const sycl::range<3> block_nums(1, 1, block_num_y); const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); { - - stream->submit([&](sycl::handler &cgh) { - - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) - [[sycl::reqd_sub_group_size(WARP_SIZE)]] { - mul_mat_vec_q( - vx, vy, dst, ncols, nrows, item_ct1); - }); + sycl_launch(stream, [&](sycl::handler & cgh) { + sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + mul_mat_vec_q( + vx, vy, dst, ncols, nrows, item_ct1); + }); }); } } @@ -628,17 +618,12 @@ static void mul_mat_vec_q5_1_q8_1_sycl(const void *vx, const void *vy, const sycl::range<3> block_nums(1, 1, block_num_y); const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); { - - stream->submit([&](sycl::handler &cgh) { - - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) - [[sycl::reqd_sub_group_size(WARP_SIZE)]] { - mul_mat_vec_q( - vx, vy, dst, ncols, nrows, item_ct1); - }); + sycl_launch(stream, [&](sycl::handler & cgh) { + sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + mul_mat_vec_q( + vx, vy, dst, ncols, nrows, item_ct1); + }); }); } } @@ -652,17 +637,12 @@ static void mul_mat_vec_q8_0_q8_1_sycl(const void *vx, const void *vy, const sycl::range<3> block_nums(1, 1, block_num_y); const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); { - - stream->submit([&](sycl::handler &cgh) { - - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) - [[sycl::reqd_sub_group_size(WARP_SIZE)]] { - mul_mat_vec_q( - vx, vy, dst, ncols, nrows, item_ct1); - }); + sycl_launch(stream, [&](sycl::handler & cgh) { + sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + mul_mat_vec_q( + vx, vy, dst, ncols, nrows, item_ct1); + }); }); } } @@ -676,17 +656,12 @@ static void mul_mat_vec_q2_K_q8_1_sycl(const void *vx, const void *vy, const sycl::range<3> block_nums(1, 1, block_num_y); const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); { - - stream->submit([&](sycl::handler &cgh) { - - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) - [[sycl::reqd_sub_group_size(WARP_SIZE)]] { - mul_mat_vec_q( - vx, vy, dst, ncols, nrows, item_ct1); - }); + sycl_launch(stream, [&](sycl::handler & cgh) { + sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + mul_mat_vec_q( + vx, vy, dst, ncols, nrows, item_ct1); + }); }); } } @@ -700,17 +675,12 @@ static void mul_mat_vec_q3_K_q8_1_sycl(const void *vx, const void *vy, const sycl::range<3> block_nums(1, 1, block_num_y); const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); { - - stream->submit([&](sycl::handler &cgh) { - - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) - [[sycl::reqd_sub_group_size(WARP_SIZE)]] { - mul_mat_vec_q( - vx, vy, dst, ncols, nrows, item_ct1); - }); + sycl_launch(stream, [&](sycl::handler & cgh) { + sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + mul_mat_vec_q( + vx, vy, dst, ncols, nrows, item_ct1); + }); }); } } @@ -724,17 +694,12 @@ static void mul_mat_vec_q4_K_q8_1_sycl(const void *vx, const void *vy, const sycl::range<3> block_nums(1, 1, block_num_y); const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); { - - stream->submit([&](sycl::handler &cgh) { - - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) - [[sycl::reqd_sub_group_size(WARP_SIZE)]] { - mul_mat_vec_q( - vx, vy, dst, ncols, nrows, item_ct1); - }); + sycl_launch(stream, [&](sycl::handler & cgh) { + sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + mul_mat_vec_q( + vx, vy, dst, ncols, nrows, item_ct1); + }); }); } } @@ -750,12 +715,12 @@ static void reorder_mul_mat_vec_q4_k_q8_1_sycl(const void * vx, const void * vy, const sycl::range<3> global_size(1, GGML_SYCL_MMV_Y, block_num_y * WARP_SIZE); const sycl::range<3> workgroup_size(1, GGML_SYCL_MMV_Y, num_subgroups * WARP_SIZE); - stream->submit([&](sycl::handler & cgh) { - cgh.parallel_for(sycl::nd_range<3>(global_size, workgroup_size), - [=](sycl::nd_item<3> nd_item) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { - mul_mat_vec_q_reorder>(vx, vy, dst, ncols, - nrows, nd_item); - }); + sycl_launch(stream, [&](sycl::handler & cgh) { + sycl_parallel_for(cgh, sycl::nd_range<3>(global_size, workgroup_size), + [=](sycl::nd_item<3> nd_item) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + mul_mat_vec_q_reorder>(vx, vy, dst, ncols, nrows, + nd_item); + }); }); } @@ -769,17 +734,12 @@ static void mul_mat_vec_q5_K_q8_1_sycl(const void *vx, const void *vy, const sycl::range<3> block_nums(1, 1, block_num_y); const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); { - - stream->submit([&](sycl::handler &cgh) { - - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) - [[sycl::reqd_sub_group_size(WARP_SIZE)]] { - mul_mat_vec_q( - vx, vy, dst, ncols, nrows, item_ct1); - }); + sycl_launch(stream, [&](sycl::handler & cgh) { + sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + mul_mat_vec_q( + vx, vy, dst, ncols, nrows, item_ct1); + }); }); } } @@ -794,12 +754,12 @@ static void reorder_mul_mat_vec_q6_k_q8_1_sycl(const void * vx, const void * vy, const sycl::range<3> global_size(1, GGML_SYCL_MMV_Y, block_num_y * WARP_SIZE); const sycl::range<3> workgroup_size(1, GGML_SYCL_MMV_Y, num_subgroups * WARP_SIZE); - stream->submit([&](sycl::handler & cgh) { - cgh.parallel_for(sycl::nd_range<3>(global_size, workgroup_size), - [=](sycl::nd_item<3> nd_item) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { - mul_mat_vec_q_reorder>(vx, vy, dst, ncols, nrows, - nd_item); - }); + sycl_launch(stream, [&](sycl::handler & cgh) { + sycl_parallel_for(cgh, sycl::nd_range<3>(global_size, workgroup_size), + [=](sycl::nd_item<3> nd_item) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + mul_mat_vec_q_reorder>(vx, vy, dst, ncols, nrows, + nd_item); + }); }); } static void mul_mat_vec_q6_K_q8_1_sycl(const void *vx, const void *vy, @@ -811,17 +771,12 @@ static void mul_mat_vec_q6_K_q8_1_sycl(const void *vx, const void *vy, const sycl::range<3> block_nums(1, 1, block_num_y); const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); { - - stream->submit([&](sycl::handler &cgh) { - - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) - [[sycl::reqd_sub_group_size(WARP_SIZE)]] { - mul_mat_vec_q( - vx, vy, dst, ncols, nrows, item_ct1); - }); + sycl_launch(stream, [&](sycl::handler & cgh) { + sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + mul_mat_vec_q( + vx, vy, dst, ncols, nrows, item_ct1); + }); }); } } @@ -836,14 +791,12 @@ static void mul_mat_vec_iq2_xxs_q8_1_sycl(const void *vx, const void *vy, const sycl::range<3> block_nums(1, 1, block_num_y); const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); { - stream->submit([&](sycl::handler &cgh) { - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) - [[sycl::reqd_sub_group_size(WARP_SIZE)]] { - mul_mat_vec_q_iq2_xxs_q8_1( - vx, vy, dst, ncols, nrows, item_ct1); - }); + sycl_launch(stream, [&](sycl::handler & cgh) { + sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + mul_mat_vec_q_iq2_xxs_q8_1(vx, vy, dst, ncols, + nrows, item_ct1); + }); }); } } @@ -857,14 +810,12 @@ static void mul_mat_vec_iq2_xs_q8_1_sycl(const void *vx, const void *vy, const sycl::range<3> block_nums(1, 1, block_num_y); const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); { - stream->submit([&](sycl::handler & cgh) { - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) - [[sycl::reqd_sub_group_size(WARP_SIZE)]] { - mul_mat_vec_q_iq2_xs_q8_1( - vx, vy, dst, ncols, nrows, item_ct1); - }); + sycl_launch(stream, [&](sycl::handler & cgh) { + sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + mul_mat_vec_q_iq2_xs_q8_1(vx, vy, dst, ncols, + nrows, item_ct1); + }); }); } } @@ -878,15 +829,12 @@ static void mul_mat_vec_iq2_s_q8_1_sycl(const void *vx, const void *vy, const sycl::range<3> block_nums(1, 1, block_num_y); const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); { - - stream->submit([&](sycl::handler &cgh) { - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) - [[sycl::reqd_sub_group_size(WARP_SIZE)]] { - mul_mat_vec_q_iq2_s_q8_1( - vx, vy, dst, ncols, nrows, item_ct1); - }); + sycl_launch(stream, [&](sycl::handler & cgh) { + sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + mul_mat_vec_q_iq2_s_q8_1(vx, vy, dst, ncols, nrows, + item_ct1); + }); }); } } @@ -900,15 +848,12 @@ static void mul_mat_vec_iq3_xxs_q8_1_sycl(const void *vx, const void *vy, const sycl::range<3> block_nums(1, 1, block_num_y); const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); { - - stream->submit([&](sycl::handler &cgh) { - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) - [[sycl::reqd_sub_group_size(WARP_SIZE)]] { - mul_mat_vec_q_iq3_xxs_q8_1( - vx, vy, dst, ncols, nrows, item_ct1); - }); + sycl_launch(stream, [&](sycl::handler & cgh) { + sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + mul_mat_vec_q_iq3_xxs_q8_1(vx, vy, dst, ncols, + nrows, item_ct1); + }); }); } } @@ -922,15 +867,12 @@ static void mul_mat_vec_iq3_s_q8_1_sycl(const void *vx, const void *vy, const sycl::range<3> block_nums(1, 1, block_num_y); const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); { - - stream->submit([&](sycl::handler &cgh) { - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) - [[sycl::reqd_sub_group_size(WARP_SIZE)]] { - mul_mat_vec_q_iq3_s_q8_1( - vx, vy, dst, ncols, nrows, item_ct1); - }); + sycl_launch(stream, [&](sycl::handler & cgh) { + sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + mul_mat_vec_q_iq3_s_q8_1(vx, vy, dst, ncols, nrows, + item_ct1); + }); }); } } @@ -944,15 +886,12 @@ static void mul_mat_vec_iq1_s_q8_1_sycl(const void *vx, const void *vy, const sycl::range<3> block_nums(1, 1, block_num_y); const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); { - - stream->submit([&](sycl::handler &cgh) { - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) - [[sycl::reqd_sub_group_size(WARP_SIZE)]] { - mul_mat_vec_q_iq1_s_q8_1( - vx, vy, dst, ncols, nrows, item_ct1); - }); + sycl_launch(stream, [&](sycl::handler & cgh) { + sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + mul_mat_vec_q_iq1_s_q8_1(vx, vy, dst, ncols, nrows, + item_ct1); + }); }); } } @@ -966,14 +905,12 @@ static void mul_mat_vec_iq1_m_q8_1_sycl(const void *vx, const void *vy, const sycl::range<3> block_nums(1, 1, block_num_y); const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); { - stream->submit([&](sycl::handler &cgh) { - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) - [[sycl::reqd_sub_group_size(WARP_SIZE)]] { - mul_mat_vec_q_iq1_m_q8_1( - vx, vy, dst, ncols, nrows, item_ct1); - }); + sycl_launch(stream, [&](sycl::handler & cgh) { + sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + mul_mat_vec_q_iq1_m_q8_1(vx, vy, dst, ncols, nrows, + item_ct1); + }); }); } } @@ -987,15 +924,12 @@ static void mul_mat_vec_iq4_nl_q8_1_sycl(const void *vx, const void *vy, const sycl::range<3> block_nums(1, 1, block_num_y); const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); { - - stream->submit([&](sycl::handler &cgh) { - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) - [[sycl::reqd_sub_group_size(WARP_SIZE)]] { - mul_mat_vec_q_iq4_nl_q8_1( - vx, vy, dst, ncols, nrows, item_ct1); - }); + sycl_launch(stream, [&](sycl::handler & cgh) { + sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + mul_mat_vec_q_iq4_nl_q8_1(vx, vy, dst, ncols, nrows, + item_ct1); + }); }); } } @@ -1009,15 +943,12 @@ static void mul_mat_vec_iq4_xs_q8_1_sycl(const void *vx, const void *vy, const sycl::range<3> block_nums(1, 1, block_num_y); const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); { - - stream->submit([&](sycl::handler &cgh) { - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) - [[sycl::reqd_sub_group_size(WARP_SIZE)]] { - mul_mat_vec_q_iq4_xs_q8_1( - vx, vy, dst, ncols, nrows, item_ct1); - }); + sycl_launch(stream, [&](sycl::handler & cgh) { + sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + mul_mat_vec_q_iq4_xs_q8_1(vx, vy, dst, ncols, + nrows, item_ct1); + }); }); } } diff --git a/ggml/src/ggml-sycl/norm.cpp b/ggml/src/ggml-sycl/norm.cpp index 4ec141684..79d846b41 100644 --- a/ggml/src/ggml-sycl/norm.cpp +++ b/ggml/src/ggml-sycl/norm.cpp @@ -254,14 +254,13 @@ static void norm_f32_sycl(const float * x, float * dst, const int ncols, const i GGML_ASSERT(ncols % WARP_SIZE == 0); if (ncols < 1024) { const sycl::range<3> block_dims(1, 1, WARP_SIZE); - stream->submit([&](sycl::handler& cgh) { - cgh.parallel_for( - sycl::nd_range<3>(global_dims * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) - [[sycl::reqd_sub_group_size(WARP_SIZE)]] { - norm_f32(x, dst, ncols, stride_row, stride_channel, stride_sample, eps, item_ct1, nullptr, WARP_SIZE); - }); - }); + sycl_launch(stream, [&](sycl::handler & cgh) { + sycl_parallel_for(cgh, sycl::nd_range<3>(global_dims * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + norm_f32(x, dst, ncols, stride_row, stride_channel, stride_sample, eps, item_ct1, + nullptr, WARP_SIZE); + }); + }); } else { const int work_group_size = ggml_sycl_info().max_work_group_sizes[device]; @@ -272,16 +271,15 @@ static void norm_f32_sycl(const float * x, float * dst, const int ncols, const i the limit. To get the device limit, query info::device::max_work_group_size. Adjust the work-group size if needed. */ - stream->submit([&](sycl::handler& cgh) { + sycl_launch(stream, [&](sycl::handler & cgh) { sycl::local_accessor s_sum_acc_ct1( sycl::range<1>(work_group_size / WARP_SIZE), cgh); - cgh.parallel_for( - sycl::nd_range<3>(global_dims * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) - [[sycl::reqd_sub_group_size(WARP_SIZE)]] { - norm_f32(x, dst, ncols, stride_row, stride_channel, stride_sample, eps, item_ct1, get_pointer(s_sum_acc_ct1), work_group_size); - }); - }); + sycl_parallel_for(cgh, sycl::nd_range<3>(global_dims * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + norm_f32(x, dst, ncols, stride_row, stride_channel, stride_sample, eps, item_ct1, + get_pointer(s_sum_acc_ct1), work_group_size); + }); + }); } } @@ -290,18 +288,14 @@ static void group_norm_f32_sycl(const float* x, float* dst, const int ne_elements, queue_ptr stream, int device) { if (group_size < 1024) { const sycl::range<3> block_dims(1, 1, WARP_SIZE); - stream->submit([&](sycl::handler& cgh) { + sycl_launch(stream, [&](sycl::handler & cgh) { const float eps_ct4 = eps; - cgh.parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_groups) * block_dims, - block_dims), - [=](sycl::nd_item<3> item_ct1) - [[sycl::reqd_sub_group_size(WARP_SIZE)]] { - group_norm_f32( - x, dst, group_size, ne_elements, eps_ct4, item_ct1, - nullptr, WARP_SIZE); - }); - }); + sycl_parallel_for(cgh, sycl::nd_range<3>(sycl::range<3>(1, 1, num_groups) * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + group_norm_f32(x, dst, group_size, ne_elements, eps_ct4, item_ct1, nullptr, + WARP_SIZE); + }); + }); } else { const int work_group_size = ggml_sycl_info().max_work_group_sizes[device]; @@ -313,22 +307,18 @@ static void group_norm_f32_sycl(const float* x, float* dst, info::device::max_work_group_size. Adjust the work-group size if needed. */ - stream->submit([&](sycl::handler& cgh) { + sycl_launch(stream, [&](sycl::handler & cgh) { sycl::local_accessor s_sum_acc_ct1(sycl::range<1>(work_group_size / WARP_SIZE), cgh); const float eps_ct4 = eps; - cgh.parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_groups) * block_dims, - block_dims), - [=](sycl::nd_item<3> item_ct1) - [[sycl::reqd_sub_group_size(WARP_SIZE)]] { - group_norm_f32(x, dst, group_size, ne_elements, - eps_ct4, item_ct1, - get_pointer(s_sum_acc_ct1), work_group_size); - }); - }); + sycl_parallel_for(cgh, sycl::nd_range<3>(sycl::range<3>(1, 1, num_groups) * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + group_norm_f32(x, dst, group_size, ne_elements, eps_ct4, item_ct1, + get_pointer(s_sum_acc_ct1), work_group_size); + }); + }); } } @@ -340,14 +330,13 @@ static void rms_norm_f32_sycl(const float* x, float* dst, const int ncols, const const sycl::range<3> global_dims(nsamples, nchannels, nrows); if (ncols < 1024) { const sycl::range<3> block_dims(1, 1, WARP_SIZE); - stream->submit([&](sycl::handler& cgh) { - cgh.parallel_for( - sycl::nd_range<3>(global_dims * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) - [[sycl::reqd_sub_group_size(WARP_SIZE)]] { - rms_norm_f32(x, dst, ncols, stride_row, stride_channel, stride_sample, eps, item_ct1, nullptr, WARP_SIZE); - }); - }); + sycl_launch(stream, [&](sycl::handler & cgh) { + sycl_parallel_for(cgh, sycl::nd_range<3>(global_dims * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + rms_norm_f32(x, dst, ncols, stride_row, stride_channel, stride_sample, eps, item_ct1, + nullptr, WARP_SIZE); + }); + }); } else { const int work_group_size = ggml_sycl_info().max_work_group_sizes[device]; @@ -358,16 +347,15 @@ static void rms_norm_f32_sycl(const float* x, float* dst, const int ncols, const the limit. To get the device limit, query info::device::max_work_group_size. Adjust the work-group size if needed. */ - stream->submit([&](sycl::handler& cgh) { + sycl_launch(stream, [&](sycl::handler & cgh) { sycl::local_accessor s_sum_acc_ct1(sycl::range<1>(work_group_size / WARP_SIZE), cgh); - cgh.parallel_for( - sycl::nd_range<3>(global_dims * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) - [[sycl::reqd_sub_group_size(WARP_SIZE)]] { - rms_norm_f32(x, dst, ncols, stride_row, stride_channel, stride_sample, eps, item_ct1, get_pointer(s_sum_acc_ct1), work_group_size); - }); - }); + sycl_parallel_for(cgh, sycl::nd_range<3>(global_dims * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + rms_norm_f32(x, dst, ncols, stride_row, stride_channel, stride_sample, eps, item_ct1, + get_pointer(s_sum_acc_ct1), work_group_size); + }); + }); } } @@ -378,16 +366,12 @@ static void l2_norm_f32_sycl(const float* x, float* dst, const int ncols, // printf("%s ncols=%d, nrows=%d, WARP_SIZE=%d\n", __func__, ncols, nrows, WARP_SIZE); if (ncols < 1024) { const sycl::range<3> block_dims(1, 1, WARP_SIZE); - stream->submit([&](sycl::handler& cgh) { - cgh.parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims, - block_dims), - [=](sycl::nd_item<3> item_ct1) - [[sycl::reqd_sub_group_size(WARP_SIZE)]] { - l2_norm_f32(x, dst, ncols, eps, item_ct1, - nullptr, WARP_SIZE); - }); - }); + sycl_launch(stream, [&](sycl::handler & cgh) { + sycl_parallel_for(cgh, sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + l2_norm_f32(x, dst, ncols, eps, item_ct1, nullptr, WARP_SIZE); + }); + }); } else { const int work_group_size = ggml_sycl_info().max_work_group_sizes[device]; @@ -398,18 +382,15 @@ static void l2_norm_f32_sycl(const float* x, float* dst, const int ncols, the limit. To get the device limit, query info::device::max_work_group_size. Adjust the work-group size if needed. */ - stream->submit([&](sycl::handler& cgh) { + sycl_launch(stream, [&](sycl::handler & cgh) { sycl::local_accessor s_sum_acc_ct1(sycl::range<1>(work_group_size / WARP_SIZE), cgh); - cgh.parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims, - block_dims), - [=](sycl::nd_item<3> item_ct1) - [[sycl::reqd_sub_group_size(WARP_SIZE)]] { - l2_norm_f32(x, dst, ncols, eps, item_ct1, - get_pointer(s_sum_acc_ct1), work_group_size); - }); - }); + sycl_parallel_for(cgh, sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + l2_norm_f32(x, dst, ncols, eps, item_ct1, get_pointer(s_sum_acc_ct1), + work_group_size); + }); + }); } } diff --git a/ggml/src/ggml-sycl/rope.cpp b/ggml/src/ggml-sycl/rope.cpp index 44473e1e5..e44c6b6ef 100644 --- a/ggml/src/ggml-sycl/rope.cpp +++ b/ggml/src/ggml-sycl/rope.cpp @@ -235,20 +235,22 @@ static void rope_norm_sycl(const T * x, T * dst, const int ne0, const int ne1, c the limit. To get the device limit, query info::device::max_work_group_size. Adjust the work-group size if needed. */ - stream->parallel_for(sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { - rope_norm(x, dst, ne0, ne1, s1, s2, n_dims, pos, freq_scale, ext_factor, attn_factor, corr_dims, - theta_scale, freq_factors, item_ct1); - }); + sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) { + rope_norm(x, dst, ne0, ne1, s1, s2, n_dims, pos, freq_scale, ext_factor, + attn_factor, corr_dims, theta_scale, freq_factors, item_ct1); + }); } else { /* DPCT1049:41: The work-group size passed to the SYCL kernel may exceed the limit. To get the device limit, query info::device::max_work_group_size. Adjust the work-group size if needed. */ - stream->parallel_for(sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { - rope_norm(x, dst, ne0, ne1, s1, s2, n_dims, pos, freq_scale, ext_factor, attn_factor, corr_dims, - theta_scale, freq_factors, item_ct1); - }); + sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) { + rope_norm(x, dst, ne0, ne1, s1, s2, n_dims, pos, freq_scale, ext_factor, + attn_factor, corr_dims, theta_scale, freq_factors, item_ct1); + }); } } @@ -267,15 +269,17 @@ static void rope_neox_sycl(const T * x, T * dst, const int ne0, const int ne1, c dpct::has_capability_or_fail(stream->get_device(), { sycl::aspect::fp16 }); if (freq_factors == nullptr) { - stream->parallel_for(sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { - rope_neox(x, dst, ne0, ne1, s1, s2, n_dims, pos, freq_scale, ext_factor, attn_factor, corr_dims, - theta_scale, freq_factors, item_ct1); - }); + sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) { + rope_neox(x, dst, ne0, ne1, s1, s2, n_dims, pos, freq_scale, ext_factor, + attn_factor, corr_dims, theta_scale, freq_factors, item_ct1); + }); } else { - stream->parallel_for(sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { - rope_neox(x, dst, ne0, ne1, s1, s2, n_dims, pos, freq_scale, ext_factor, attn_factor, corr_dims, - theta_scale, freq_factors, item_ct1); - }); + sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) { + rope_neox(x, dst, ne0, ne1, s1, s2, n_dims, pos, freq_scale, ext_factor, + attn_factor, corr_dims, theta_scale, freq_factors, item_ct1); + }); } } @@ -298,12 +302,12 @@ static void rope_multi_sycl(const T * x, T * dst, const int ne0, const int ne1, } // launch kernel if (freq_factors == nullptr) { - stream->parallel_for(nd_range, [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for(stream, nd_range, [=](sycl::nd_item<3> item_ct1) { rope_multi(x, dst, ne0, ne1, ne2, s1, s2, n_dims, pos, freq_scale, ext_factor, attn_factor, corr_dims, theta_scale, freq_factors, sections, item_ct1); }); } else { - stream->parallel_for(nd_range, [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for(stream, nd_range, [=](sycl::nd_item<3> item_ct1) { rope_multi(x, dst, ne0, ne1, ne2, s1, s2, n_dims, pos, freq_scale, ext_factor, attn_factor, corr_dims, theta_scale, freq_factors, sections, item_ct1); }); @@ -333,12 +337,12 @@ static void rope_vision_sycl(const T * x, T * dst, const int ne0, const int ne1, } // launch kernel if (freq_factors == nullptr) { - stream->parallel_for(nd_range, [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for(stream, nd_range, [=](sycl::nd_item<3> item_ct1) { rope_vision(x, dst, ne0, ne1, ne2, s1, s2, n_dims, pos, freq_scale, ext_factor, attn_factor, corr_dims, theta_scale, freq_factors, sections, item_ct1); }); } else { - stream->parallel_for(nd_range, [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for(stream, nd_range, [=](sycl::nd_item<3> item_ct1) { rope_vision(x, dst, ne0, ne1, ne2, s1, s2, n_dims, pos, freq_scale, ext_factor, attn_factor, corr_dims, theta_scale, freq_factors, sections, item_ct1); }); diff --git a/ggml/src/ggml-sycl/softmax.cpp b/ggml/src/ggml-sycl/softmax.cpp index 52fcf4b3d..7b60c292e 100644 --- a/ggml/src/ggml-sycl/softmax.cpp +++ b/ggml/src/ggml-sycl/softmax.cpp @@ -127,11 +127,11 @@ static void soft_max_f32_submitter(const float * x, const T * mask, float * dst, const int nrows_y, const float scale, const float max_bias, const float m0, const float m1, uint32_t n_head_log2, sycl::range<3> block_nums, sycl::range<3> block_dims, const size_t n_local_scratch, queue_ptr stream) { - stream->submit([&](sycl::handler &cgh) { + sycl_launch(stream, [&](sycl::handler & cgh) { sycl::local_accessor local_buf_acc(n_local_scratch, cgh); - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), + sycl_parallel_for( + cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { soft_max_f32(x, mask, dst, ncols_par, nrows_y, scale, max_bias, m0, diff --git a/ggml/src/ggml-sycl/tsembd.cpp b/ggml/src/ggml-sycl/tsembd.cpp index f6ca626ea..721c8fa6f 100644 --- a/ggml/src/ggml-sycl/tsembd.cpp +++ b/ggml/src/ggml-sycl/tsembd.cpp @@ -45,14 +45,9 @@ static void timestep_embedding_f32_sycl( int num_blocks = (half_ceil + SYCL_TIMESTEP_EMBEDDING_BLOCK_SIZE - 1) / SYCL_TIMESTEP_EMBEDDING_BLOCK_SIZE; sycl::range<3> block_dims(1, 1, SYCL_TIMESTEP_EMBEDDING_BLOCK_SIZE); sycl::range<3> gridDim(1, ne00, num_blocks); - stream->parallel_for( - sycl::nd_range<3>( - gridDim * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { - timestep_embedding_f32( - x, dst, nb1, dim, max_period, item_ct1 - ); - }); + sycl_parallel_for(stream, sycl::nd_range<3>(gridDim * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { + timestep_embedding_f32(x, dst, nb1, dim, max_period, item_ct1); + }); } void ggml_sycl_op_timestep_embedding(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { diff --git a/ggml/src/ggml-sycl/wkv.cpp b/ggml/src/ggml-sycl/wkv.cpp index c10e2f764..3ed5bbf35 100644 --- a/ggml/src/ggml-sycl/wkv.cpp +++ b/ggml/src/ggml-sycl/wkv.cpp @@ -207,12 +207,11 @@ void ggml_sycl_op_rwkv_wkv6(ggml_backend_sycl_context& ctx, ggml_tensor* dst) { // Submit kernel if (C / H == WKV_BLOCK_SIZE) { - stream->submit([&](sycl::handler& cgh) { + sycl_launch(stream, [&](sycl::handler & cgh) { sycl::local_accessor shared_mem_acc(shared_mem_size, cgh); - cgh.parallel_for( - sycl::nd_range<3>(grid_dims * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for( + cgh, sycl::nd_range<3>(grid_dims * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { rwkv_wkv6_f32_kernel( B, T, C, H, k_d, v_d, r_d, tf_d, td_d, s_d, dst_d, item_ct1, (float*)shared_mem_acc.get_multi_ptr().get() @@ -220,12 +219,11 @@ void ggml_sycl_op_rwkv_wkv6(ggml_backend_sycl_context& ctx, ggml_tensor* dst) { }); }); } else { - stream->submit([&](sycl::handler& cgh) { + sycl_launch(stream, [&](sycl::handler & cgh) { sycl::local_accessor shared_mem_acc(shared_mem_size, cgh); - cgh.parallel_for( - sycl::nd_range<3>(grid_dims * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for( + cgh, sycl::nd_range<3>(grid_dims * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { rwkv_wkv6_f32_kernel( B, T, C, H, k_d, v_d, r_d, tf_d, td_d, s_d, dst_d, item_ct1, (float*)shared_mem_acc.get_multi_ptr().get() @@ -264,12 +262,11 @@ void ggml_sycl_op_rwkv_wkv7(ggml_backend_sycl_context& ctx, ggml_tensor* dst) { // Submit kernel if (C / H == WKV_BLOCK_SIZE) { - stream->submit([&](sycl::handler& cgh) { + sycl_launch(stream, [&](sycl::handler & cgh) { sycl::local_accessor shared_mem_acc(shared_mem_size, cgh); - cgh.parallel_for( - sycl::nd_range<3>(grid_dims * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for( + cgh, sycl::nd_range<3>(grid_dims * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { rwkv_wkv7_f32_kernel( B, T, C, H, r_d, w_d, k_d, v_d, a_d, b_d, s_d, dst_d, item_ct1, (float*)shared_mem_acc.get_multi_ptr().get() @@ -277,12 +274,11 @@ void ggml_sycl_op_rwkv_wkv7(ggml_backend_sycl_context& ctx, ggml_tensor* dst) { }); }); } else { - stream->submit([&](sycl::handler& cgh) { + sycl_launch(stream, [&](sycl::handler & cgh) { sycl::local_accessor shared_mem_acc(shared_mem_size, cgh); - cgh.parallel_for( - sycl::nd_range<3>(grid_dims * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for( + cgh, sycl::nd_range<3>(grid_dims * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { rwkv_wkv7_f32_kernel( B, T, C, H, r_d, w_d, k_d, v_d, a_d, b_d, s_d, dst_d, item_ct1, (float*)shared_mem_acc.get_multi_ptr().get() From dd6e6d0b6a4bbe3ebfc931d1eb14db2f2b1d70af Mon Sep 17 00:00:00 2001 From: Ruikai Peng Date: Fri, 20 Jun 2025 22:13:06 +0800 Subject: [PATCH 19/43] vocab : prevent tokenizer overflow (#14301) * vocab : prevent stack overflow in tokenize * vocab : return error instead of aborting on oversized token count * vocab : INT32_MIN from llama_tokenize on overflow --- common/common.cpp | 3 +++ include/llama.h | 1 + src/llama-vocab.cpp | 5 +++++ 3 files changed, 9 insertions(+) diff --git a/common/common.cpp b/common/common.cpp index c2c94e7ae..e4e71ad13 100644 --- a/common/common.cpp +++ b/common/common.cpp @@ -1290,6 +1290,9 @@ std::vector common_tokenize( int n_tokens = text.length() + 2 * add_special; std::vector result(n_tokens); n_tokens = llama_tokenize(vocab, text.data(), text.length(), result.data(), result.size(), add_special, parse_special); + if (n_tokens == std::numeric_limits::min()) { + throw std::runtime_error("Tokenization failed: input text too large, tokenization result exceeds int32_t limit"); + } if (n_tokens < 0) { result.resize(-n_tokens); int check = llama_tokenize(vocab, text.data(), text.length(), result.data(), result.size(), add_special, parse_special); diff --git a/include/llama.h b/include/llama.h index 3475d5965..b04720bee 100644 --- a/include/llama.h +++ b/include/llama.h @@ -1088,6 +1088,7 @@ extern "C" { /// @param tokens The tokens pointer must be large enough to hold the resulting tokens. /// @return Returns the number of tokens on success, no more than n_tokens_max /// @return Returns a negative number on failure - the number of tokens that would have been returned + /// @return Returns INT32_MIN on overflow (e.g., tokenization result size exceeds int32_t limit) /// @param add_special Allow to add BOS and EOS tokens if model is configured to do so. /// @param parse_special Allow tokenizing special and/or control tokens which otherwise are not exposed and treated /// as plaintext. Does not insert a leading space. diff --git a/src/llama-vocab.cpp b/src/llama-vocab.cpp index 4ab120d9b..4aaf4c825 100644 --- a/src/llama-vocab.cpp +++ b/src/llama-vocab.cpp @@ -3074,6 +3074,11 @@ int32_t llama_vocab::tokenize( bool add_special, bool parse_special) const { auto res = tokenize(std::string(text, text_len), add_special, parse_special); + if (res.size() >= static_cast(std::numeric_limits::max())) { + LLAMA_LOG_ERROR("%s: tokenization result size %zu exceeds int32_t limit\n", __func__, res.size()); + return std::numeric_limits::min(); + } + if (n_tokens_max < (int) res.size()) { // LLAMA_LOG_ERROR("%s: too many tokens\n", __func__); return -((int) res.size()); From 22015b2092e291022ea3cfedc6aeb8f2643807da Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Sigbj=C3=B8rn=20Skj=C3=A6ret?= Date: Fri, 20 Jun 2025 16:37:44 +0200 Subject: [PATCH 20/43] lint : remove trailing whitepace (#14304) --- src/llama-vocab.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/src/llama-vocab.cpp b/src/llama-vocab.cpp index 4aaf4c825..5c9eb8756 100644 --- a/src/llama-vocab.cpp +++ b/src/llama-vocab.cpp @@ -3078,7 +3078,7 @@ int32_t llama_vocab::tokenize( LLAMA_LOG_ERROR("%s: tokenization result size %zu exceeds int32_t limit\n", __func__, res.size()); return std::numeric_limits::min(); } - + if (n_tokens_max < (int) res.size()) { // LLAMA_LOG_ERROR("%s: too many tokens\n", __func__); return -((int) res.size()); From c959f462a0b4d42eaf930ffd72df0e435c97d5d5 Mon Sep 17 00:00:00 2001 From: Aman Gupta Date: Fri, 20 Jun 2025 22:48:24 +0800 Subject: [PATCH 21/43] CUDA: add conv_2d_transpose (#14287) * CUDA: add conv_2d_transpose * remove direct include of cuda_fp16 * Review: add brackets for readability, remove ggml_set_param and add asserts --- ggml/src/ggml-cuda/conv2d-transpose.cu | 91 +++++++++++++++++++++++++ ggml/src/ggml-cuda/conv2d-transpose.cuh | 4 ++ ggml/src/ggml-cuda/ggml-cuda.cu | 5 ++ tests/test-backend-ops.cpp | 34 +++++++++ 4 files changed, 134 insertions(+) create mode 100644 ggml/src/ggml-cuda/conv2d-transpose.cu create mode 100644 ggml/src/ggml-cuda/conv2d-transpose.cuh diff --git a/ggml/src/ggml-cuda/conv2d-transpose.cu b/ggml/src/ggml-cuda/conv2d-transpose.cu new file mode 100644 index 000000000..03224e404 --- /dev/null +++ b/ggml/src/ggml-cuda/conv2d-transpose.cu @@ -0,0 +1,91 @@ +#include + +#include "conv2d-transpose.cuh" +#include "ggml.h" + +__global__ void conv2d_transpose_kernel(const float * __restrict__ input, const half * __restrict__ kernel, + float * __restrict__ output, const int in_w, const int in_h, const int out_w, + const int out_h, const int kernel_w, const int kernel_h, const int stride, + const int c_in, const int c_out, const int batches) { + const int global_idx = blockIdx.x * blockDim.x + threadIdx.x; + + const int total_elements = out_w * out_h * c_out * batches; + + if (global_idx >= total_elements) { + return; + } + + const int out_x_idx = global_idx % out_w; + const int out_y_idx = (global_idx / out_w) % out_h; + const int c_idx = (global_idx / (out_w * out_h)) % c_out; + const int n_idx = global_idx / (out_w * out_h * c_out); + + float accumulator = 0; + // For each output idx, find the inputs that contribute to it by checking stride alignment and bounds + + for (int c_in_idx = 0; c_in_idx < c_in; c_in_idx++) { + for (int kh = 0; kh < kernel_h; ++kh) { + int in_y = out_y_idx - kh; + if (in_y < 0 || in_y % stride) continue; + in_y /= stride; + if (in_y >= in_h) continue; + + for (int kw = 0; kw < kernel_w; ++kw) { + int in_x = out_x_idx - kw; + if (in_x < 0 || in_x % stride) continue; + in_x /= stride; + if (in_x >= in_w) continue; + + const int input_idx = (in_w * in_h * c_in) * n_idx + (in_w * in_h) * c_in_idx + (in_w) *in_y + in_x; + const int kernel_idx = + (kernel_h * kernel_w * c_out) * c_in_idx + (kernel_h * kernel_w) * c_idx + (kernel_w) *kh + kw; + + float input_val = input[input_idx]; + half kern_val = kernel[kernel_idx]; + + accumulator += input_val * (float) kern_val; + } + } + } + + output[(out_w * out_h * c_out) * n_idx + (out_w * out_h) * c_idx + (out_w) *out_y_idx + out_x_idx] = accumulator; +} + +//input is (W, H, C_in, N), Kernel is (W, H, C_out, C_in) +void ggml_cuda_conv_2d_transpose_p0(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { + const ggml_tensor * kernel = dst->src[0]; + const ggml_tensor * input = dst->src[1]; + + GGML_ASSERT(kernel->type == GGML_TYPE_F16 && input->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32); + + const float * input_data = (const float *) input->data; + float * output_data = (float *) dst->data; + const half * kernel_data = (const half *) kernel->data; + + const int input_w = input->ne[0]; + const int input_h = input->ne[1]; + const int output_w = dst->ne[0]; + const int output_h = dst->ne[1]; + const int channels_in = input->ne[2]; + const int channels_out = kernel->ne[2]; + const int kernel_w = kernel->ne[0]; + const int kernel_h = kernel->ne[1]; + const int stride = dst->op_params[0]; + const int batches = input->ne[3]; + + GGML_ASSERT(channels_in == kernel->ne[3]); + GGML_ASSERT(stride > 0); + + cudaStream_t st = ctx.stream(); + + GGML_ASSERT(ggml_is_contiguous(input)); + GGML_ASSERT(ggml_is_contiguous(kernel)); + GGML_ASSERT(ggml_is_contiguous(dst)); + + const int total = (output_w * output_h * channels_out * batches); + const int blocks = (total + CUDA_CONV2D_TRANSPOSE_BLOCK_SIZE - 1) / CUDA_CONV2D_TRANSPOSE_BLOCK_SIZE; + + conv2d_transpose_kernel<<>>( + input_data, kernel_data, output_data, input_w, input_h, output_w, output_h, kernel_w, kernel_h, stride, + channels_in, channels_out, batches); +} diff --git a/ggml/src/ggml-cuda/conv2d-transpose.cuh b/ggml/src/ggml-cuda/conv2d-transpose.cuh new file mode 100644 index 000000000..c9430b248 --- /dev/null +++ b/ggml/src/ggml-cuda/conv2d-transpose.cuh @@ -0,0 +1,4 @@ +#include "common.cuh" + +#define CUDA_CONV2D_TRANSPOSE_BLOCK_SIZE 256 +void ggml_cuda_conv_2d_transpose_p0(ggml_backend_cuda_context & ctx, ggml_tensor * dst); diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu index 530f541f9..5bab92e34 100644 --- a/ggml/src/ggml-cuda/ggml-cuda.cu +++ b/ggml/src/ggml-cuda/ggml-cuda.cu @@ -12,6 +12,7 @@ #include "ggml-cuda/concat.cuh" #include "ggml-cuda/conv-transpose-1d.cuh" #include "ggml-cuda/conv2d-dw.cuh" +#include "ggml-cuda/conv2d-transpose.cuh" #include "ggml-cuda/convert.cuh" #include "ggml-cuda/count-equal.cuh" #include "ggml-cuda/cpy.cuh" @@ -2341,6 +2342,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg case GGML_OP_CONV_2D_DW: ggml_cuda_op_conv2d_dw(ctx, dst); break; + case GGML_OP_CONV_TRANSPOSE_2D: + ggml_cuda_conv_2d_transpose_p0(ctx, dst); + break; case GGML_OP_CONV_TRANSPOSE_1D: ggml_cuda_op_conv_transpose_1d(ctx,dst); break; @@ -3252,6 +3256,7 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g } case GGML_OP_IM2COL: case GGML_OP_CONV_2D_DW: + case GGML_OP_CONV_TRANSPOSE_2D: case GGML_OP_POOL_2D: case GGML_OP_SUM: case GGML_OP_SUM_ROWS: diff --git a/tests/test-backend-ops.cpp b/tests/test-backend-ops.cpp index 509a4b35f..772bee346 100644 --- a/tests/test-backend-ops.cpp +++ b/tests/test-backend-ops.cpp @@ -2725,6 +2725,35 @@ struct test_conv_transpose_1d : public test_case { } }; +// GGML_OP_CONV_TRANSPOSE_2D +struct test_conv_transpose_2d : public test_case { + const std::array ne_input; + const std::array ne_kernel; + const int stride; + + std::string vars() override { + return VARS_TO_STR3(ne_input, ne_kernel, stride); + } + + test_conv_transpose_2d(std::array ne_input = {10, 10, 3, 1}, // [input_width, input_height, input_channels, 1] + std::array ne_kernel = {3, 3, 3, 1}, // [kernel_width, kernel_height, input_channels, 1] + int stride = 1) + : ne_input(ne_input), ne_kernel(ne_kernel), stride(stride){} + + ggml_tensor * build_graph(ggml_context * ctx) override { + ggml_tensor * input = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne_input.data()); + ggml_set_name(input, "input"); + + ggml_tensor * kernel = ggml_new_tensor(ctx, GGML_TYPE_F16, 4, ne_kernel.data()); + ggml_set_name(kernel, "kernel"); + + ggml_tensor * out = ggml_conv_transpose_2d_p0(ctx, kernel, input, stride); + ggml_set_name(out, "out"); + + return out; + } +}; + // GGML_OP_IM2COL struct test_im2col : public test_case { const ggml_type type_input; @@ -4050,6 +4079,9 @@ static std::vector> make_test_cases_eval() { test_cases.emplace_back(new test_conv_transpose_1d({3,2,1,1}, {3,1,2,1}, 1, 0, 1)); test_cases.emplace_back(new test_conv_transpose_1d({2,1,1,1}, {3,1,1,1}, 1, 0, 1)); + test_cases.emplace_back(new test_conv_transpose_2d({3, 2, 3, 1}, {2, 2, 1, 3}, 1)); + test_cases.emplace_back(new test_conv_transpose_2d({10, 10, 9, 1}, {3, 3, 1, 9}, 2)); + test_cases.emplace_back(new test_count_equal(GGML_TYPE_F32, {4, 500, 1, 1})); test_cases.emplace_back(new test_count_equal(GGML_TYPE_F32, {4, 5000, 1, 1})); @@ -4618,6 +4650,8 @@ static std::vector> make_test_cases_perf() { test_cases.emplace_back(new test_conv_2d_dw({512, 512, 256, 1}, {3, 3, 1, 256}, 1, 1, 1, false)); test_cases.emplace_back(new test_conv_2d_dw({512, 512, 256, 1}, {3, 3, 1, 256}, 1, 1, 1, true)); + test_cases.emplace_back(new test_conv_transpose_2d({256, 256, 256, 1}, {3, 3, 16, 256}, 1)); + return test_cases; } From d860dd99a4178f58d1d1fa64eebc2aabc95392a7 Mon Sep 17 00:00:00 2001 From: David Chiu Date: Sat, 21 Jun 2025 01:43:35 +0800 Subject: [PATCH 22/43] docs : fix the link to llama.h (#14293) --- docs/build.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/docs/build.md b/docs/build.md index 680b0d839..20a6f606e 100644 --- a/docs/build.md +++ b/docs/build.md @@ -1,6 +1,6 @@ # Build llama.cpp locally -The main product of this project is the `llama` library. Its C-style interface can be found in [include/llama.h](include/llama.h). +The main product of this project is the `llama` library. Its C-style interface can be found in [include/llama.h](../include/llama.h). The project also includes many example programs and tools using the `llama` library. The examples range from simple, minimal code snippets to sophisticated sub-projects such as an OpenAI-compatible HTTP server. From b7147673f26c7ceb926a43c4734002bba291bcb8 Mon Sep 17 00:00:00 2001 From: Acly Date: Wed, 18 Jun 2025 13:34:50 +0200 Subject: [PATCH 23/43] Add `ggml_roll` (ggml/1274) * ggml : add ggml_roll * use set/get_op_params & std::min --- ggml/include/ggml.h | 12 +++++++ ggml/src/ggml-cpu/ggml-cpu.c | 5 +++ ggml/src/ggml-cpu/ops.cpp | 67 ++++++++++++++++++++++++++++++++++++ ggml/src/ggml-cpu/ops.h | 1 + ggml/src/ggml.c | 34 ++++++++++++++++-- 5 files changed, 117 insertions(+), 2 deletions(-) diff --git a/ggml/include/ggml.h b/ggml/include/ggml.h index 1a57f1cd7..9c4e24023 100644 --- a/ggml/include/ggml.h +++ b/ggml/include/ggml.h @@ -489,6 +489,7 @@ extern "C" { GGML_OP_UPSCALE, // nearest interpolate GGML_OP_PAD, GGML_OP_PAD_REFLECT_1D, + GGML_OP_ROLL, GGML_OP_ARANGE, GGML_OP_TIMESTEP_EMBEDDING, GGML_OP_ARGSORT, @@ -1801,6 +1802,17 @@ extern "C" { int p0, int p1); + // Move tensor elements by an offset given for each dimension. Elements that + // are shifted beyond the last position are wrapped around to the beginning. + GGML_API struct ggml_tensor * ggml_roll( + struct ggml_context * ctx, + struct ggml_tensor * a, + int shift0, + int shift1, + int shift2, + int shift3); + + // Ref: https://github.com/CompVis/stable-diffusion/blob/main/ldm/modules/diffusionmodules/util.py#L151 // timesteps: [N,] // return: [N, dim] diff --git a/ggml/src/ggml-cpu/ggml-cpu.c b/ggml/src/ggml-cpu/ggml-cpu.c index 1bb9c4e36..1d3cd009a 100644 --- a/ggml/src/ggml-cpu/ggml-cpu.c +++ b/ggml/src/ggml-cpu/ggml-cpu.c @@ -1890,6 +1890,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm { ggml_compute_forward_pad_reflect_1d(params, tensor); } break; + case GGML_OP_ROLL: + { + ggml_compute_forward_roll(params, tensor); + } break; case GGML_OP_ARANGE: { ggml_compute_forward_arange(params, tensor); @@ -2214,6 +2218,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) { case GGML_OP_UPSCALE: case GGML_OP_PAD: case GGML_OP_PAD_REFLECT_1D: + case GGML_OP_ROLL: case GGML_OP_ARANGE: case GGML_OP_TIMESTEP_EMBEDDING: case GGML_OP_ARGSORT: diff --git a/ggml/src/ggml-cpu/ops.cpp b/ggml/src/ggml-cpu/ops.cpp index 08facb6d0..eff4a53e3 100644 --- a/ggml/src/ggml-cpu/ops.cpp +++ b/ggml/src/ggml-cpu/ops.cpp @@ -6793,6 +6793,73 @@ void ggml_compute_forward_pad_reflect_1d( } } +// ggml_compute_forward_roll + +static int64_t ggml_wrap_index(int64_t i, int64_t ne) { + if (i < 0) { + return i + ne; + } else if (i >= ne) { + return i - ne; + } + return i; +} + +static void ggml_compute_forward_roll_f32( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + const float * src_data = (const float *) src0->data; + float * dst_data = (float *) dst->data; + + GGML_TENSOR_UNARY_OP_LOCALS + + const int s0 = ggml_get_op_params_i32(dst, 0); + const int s1 = ggml_get_op_params_i32(dst, 1); + const int s2 = ggml_get_op_params_i32(dst, 2); + const int s3 = ggml_get_op_params_i32(dst, 3); + + const int64_t total = ne1 * ne2 * ne3; + const int64_t per_thread = (total + params->nth) / params->nth; + const int64_t start = params->ith * per_thread; + const int64_t end = std::min(start + per_thread, total); + + for (int64_t i = start; i < end; ++i) { + const int64_t i1 = i % ne1; + const int64_t i2 = (i / ne1) % ne2; + const int64_t i3 = i / (ne2 * ne1); + float * dst_row = dst_data + (i3*nb3 + i2*nb2 + i1*nb1) / sizeof(float); + + const int64_t i01 = ggml_wrap_index(i1 - s1, ne01); + const int64_t i02 = ggml_wrap_index(i2 - s2, ne02); + const int64_t i03 = ggml_wrap_index(i3 - s3, ne03); + const float * src_row = src_data + (i03*nb03 + i02*nb02 + i01*nb01) / sizeof(float); + + const int64_t s = ggml_wrap_index(-s0, ne00); + const int64_t n = ne00 - s; + ggml_vec_cpy_f32(n, dst_row, src_row + s); + ggml_vec_cpy_f32(s, dst_row + n, src_row); + } +} + +void ggml_compute_forward_roll( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_roll_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + // ggml_compute_forward_arange static void ggml_compute_forward_arange_f32( diff --git a/ggml/src/ggml-cpu/ops.h b/ggml/src/ggml-cpu/ops.h index dc081b9e6..2d8544d7d 100644 --- a/ggml/src/ggml-cpu/ops.h +++ b/ggml/src/ggml-cpu/ops.h @@ -72,6 +72,7 @@ void ggml_compute_forward_pool_2d_back(const struct ggml_compute_params * params void ggml_compute_forward_upscale(const struct ggml_compute_params * params, struct ggml_tensor * dst); void ggml_compute_forward_pad(const struct ggml_compute_params * params, struct ggml_tensor * dst); void ggml_compute_forward_pad_reflect_1d(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_roll(const struct ggml_compute_params * params, struct ggml_tensor * dst); void ggml_compute_forward_arange(const struct ggml_compute_params * params, struct ggml_tensor * dst); void ggml_compute_forward_timestep_embedding(const struct ggml_compute_params * params, struct ggml_tensor * dst); void ggml_compute_forward_argsort(const struct ggml_compute_params * params, struct ggml_tensor * dst); diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c index a8edad377..f8e7c595b 100644 --- a/ggml/src/ggml.c +++ b/ggml/src/ggml.c @@ -955,6 +955,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = { "UPSCALE", "PAD", "PAD_REFLECT_1D", + "ROLL", "ARANGE", "TIMESTEP_EMBEDDING", "ARGSORT", @@ -985,7 +986,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = { "OPT_STEP_ADAMW", }; -static_assert(GGML_OP_COUNT == 82, "GGML_OP_COUNT != 82"); +static_assert(GGML_OP_COUNT == 83, "GGML_OP_COUNT != 83"); static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = { "none", @@ -1050,6 +1051,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = { "upscale(x)", "pad(x)", "pad_reflect_1d(x)", + "roll(x)", "arange(start, stop, step)", "timestep_embedding(timesteps, dim, max_period)", "argsort(x)", @@ -1080,7 +1082,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = { "adamw(x)", }; -static_assert(GGML_OP_COUNT == 82, "GGML_OP_COUNT != 82"); +static_assert(GGML_OP_COUNT == 83, "GGML_OP_COUNT != 83"); static_assert(GGML_OP_POOL_COUNT == 2, "GGML_OP_POOL_COUNT != 2"); @@ -4341,6 +4343,34 @@ struct ggml_tensor * ggml_pad_reflect_1d( return result; } +// ggml_roll + +struct ggml_tensor * ggml_roll( + struct ggml_context * ctx, + struct ggml_tensor * a, + int shift0, + int shift1, + int shift2, + int shift3) { + GGML_ASSERT(a->nb[0] == ggml_type_size(a->type)); + GGML_ASSERT(abs(shift0) < a->ne[0]); + GGML_ASSERT(abs(shift1) < a->ne[1]); + GGML_ASSERT(abs(shift2) < a->ne[2]); + GGML_ASSERT(abs(shift3) < a->ne[3]); + + struct ggml_tensor * result = ggml_dup_tensor(ctx, a); + + ggml_set_op_params_i32(result, 0, shift0); + ggml_set_op_params_i32(result, 1, shift1); + ggml_set_op_params_i32(result, 2, shift2); + ggml_set_op_params_i32(result, 3, shift3); + + result->op = GGML_OP_ROLL; + result->src[0] = a; + + return result; +} + // ggml_arange struct ggml_tensor * ggml_arange( From 06cbedfca1587473df9b537f1dd4d6bfa2e3de13 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Fri, 20 Jun 2025 20:50:24 +0300 Subject: [PATCH 24/43] sync : ggml ggml-ci --- scripts/sync-ggml.last | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/scripts/sync-ggml.last b/scripts/sync-ggml.last index bb5d56a0e..bd1e04ed0 100644 --- a/scripts/sync-ggml.last +++ b/scripts/sync-ggml.last @@ -1 +1 @@ -8cda0a3c19f2c7dc493887353c42f6956bc268b1 +9e4bee1c5afc2d677a5b32ecb90cbdb483e81fff From b23fa0b3f40165ca3aae8ad4ee756e72f9a130dd Mon Sep 17 00:00:00 2001 From: Daniel Han Date: Fri, 20 Jun 2025 21:32:01 -0700 Subject: [PATCH 25/43] convert : fix Llama 4 conversion (#14311) --- convert_hf_to_gguf.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py index 2fe76589e..bbf8b30ff 100755 --- a/convert_hf_to_gguf.py +++ b/convert_hf_to_gguf.py @@ -2193,7 +2193,7 @@ class Llama4VisionModel(MmprojModel): name += ".weight" if "multi_modal_projector.linear_1" in name: # despite the name with number postfix, this is a single fully connected layer - return [(gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.V_MMPROJ_FC], data_torch)] + return [(gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.V_MMPROJ_FC] + '.weight', data_torch)] return [(self.map_tensor_name(name), data_torch)] return [] From 692e3cdd0a069ab56411b64506a67537d767683e Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Sat, 21 Jun 2025 08:03:46 +0300 Subject: [PATCH 26/43] memory : rename interface to llama_memory_context_i (#14296) * memory : rename interface to llama_memory_context_i ggml-ci * cont : fix comments * cont : use "mctx" for referencing a memory context ggml-ci --- src/llama-context.cpp | 78 ++++++++++----------- src/llama-context.h | 24 +++---- src/llama-graph.cpp | 104 ++++++++++++++-------------- src/llama-graph.h | 53 +++++++------- src/llama-kv-cache-unified-iswa.cpp | 70 +++++++++---------- src/llama-kv-cache-unified-iswa.h | 36 +++++----- src/llama-kv-cache-unified.cpp | 50 ++++++------- src/llama-kv-cache-unified.h | 32 ++++----- src/llama-memory-hybrid.cpp | 70 +++++++++---------- src/llama-memory-hybrid.h | 28 ++++---- src/llama-memory-recurrent.cpp | 46 ++++++------ src/llama-memory-recurrent.h | 28 ++++---- src/llama-memory.h | 33 ++++----- src/llama-model.cpp | 28 ++++---- 14 files changed, 339 insertions(+), 341 deletions(-) diff --git a/src/llama-context.cpp b/src/llama-context.cpp index 5a18a4fb3..e352d81e4 100644 --- a/src/llama-context.cpp +++ b/src/llama-context.cpp @@ -280,8 +280,8 @@ llama_context::llama_context( // simulate full KV cache - const auto mstate = memory->init_full(); - if (!mstate) { + const auto mctx = memory->init_full(); + if (!mctx) { throw std::runtime_error("failed to initialize KV cache"); } @@ -289,7 +289,7 @@ llama_context::llama_context( // reserve pp graph first so that buffers are only allocated once { - auto * gf = graph_reserve(n_tokens, n_seqs, n_tokens, mstate.get()); + auto * gf = graph_reserve(n_tokens, n_seqs, n_tokens, mctx.get()); if (!gf) { throw std::runtime_error("failed to allocate compute pp buffers"); } @@ -300,7 +300,7 @@ llama_context::llama_context( // reserve with tg graph to get the number of splits and nodes { - auto * gf = graph_reserve(1, 1, 1, mstate.get()); + auto * gf = graph_reserve(1, 1, 1, mctx.get()); if (!gf) { throw std::runtime_error("failed to allocate compute tg buffers"); } @@ -311,7 +311,7 @@ llama_context::llama_context( // reserve again with pp graph to avoid ggml-alloc reallocations during inference { - auto * gf = graph_reserve(n_tokens, n_seqs, n_tokens, mstate.get()); + auto * gf = graph_reserve(n_tokens, n_seqs, n_tokens, mctx.get()); if (!gf) { throw std::runtime_error("failed to allocate compute pp buffers"); } @@ -444,8 +444,8 @@ bool llama_context::kv_self_update(bool optimize) { optimize |= memory_force_optimize; memory_force_optimize = false; - const auto mstate = memory->init_update(this, optimize); - switch (mstate->get_status()) { + const auto mctx = memory->init_update(this, optimize); + switch (mctx->get_status()) { case LLAMA_MEMORY_STATUS_SUCCESS: { // noop @@ -463,22 +463,22 @@ bool llama_context::kv_self_update(bool optimize) { } } - if (!mstate->apply()) { + if (!mctx->apply()) { LLAMA_LOG_ERROR("%s: failed to apply memory update\n", __func__); } } // if the memory module did any computation, we have to reserve a new worst-case graph { - const auto mstate = memory->init_full(); - if (!mstate) { - throw std::runtime_error("failed to initialize memory state"); + const auto mctx = memory->init_full(); + if (!mctx) { + throw std::runtime_error("failed to initialize memory context"); } const uint32_t n_seqs = cparams.n_seq_max; const uint32_t n_tokens = std::min(cparams.n_ctx, cparams.n_ubatch); - auto * gf = graph_reserve(n_tokens, n_seqs, n_tokens, mstate.get()); + auto * gf = graph_reserve(n_tokens, n_seqs, n_tokens, mctx.get()); if (!gf) { LLAMA_LOG_ERROR("%s: failed to reserve graph after the memory update\n", __func__); } @@ -678,9 +678,9 @@ bool llama_context::apply_adapter_cvec( return cvec.apply(model, data, len, n_embd, il_start, il_end); } -llm_graph_result_ptr llama_context::process_ubatch(const llama_ubatch & ubatch, llm_graph_type gtype, llama_memory_state_i * mstate, ggml_status & ret) { - if (mstate && !mstate->apply()) { - LLAMA_LOG_ERROR("%s: failed to apply memory state\n", __func__); +llm_graph_result_ptr llama_context::process_ubatch(const llama_ubatch & ubatch, llm_graph_type gtype, llama_memory_context_i * mctx, ggml_status & ret) { + if (mctx && !mctx->apply()) { + LLAMA_LOG_ERROR("%s: failed to apply memory context\n", __func__); ret = GGML_STATUS_FAILED; return nullptr; } @@ -692,7 +692,7 @@ llm_graph_result_ptr llama_context::process_ubatch(const llama_ubatch & ubatch, return nullptr; } - auto res = graph_build(ctx_compute.get(), gf, ubatch, gtype, mstate); + auto res = graph_build(ctx_compute.get(), gf, ubatch, gtype, mctx); if (!res) { LLAMA_LOG_ERROR("%s: failed to build graph\n", __func__); ret = GGML_STATUS_FAILED; @@ -933,21 +933,21 @@ int llama_context::decode(const llama_batch & batch_inp) { // handle any pending defrags/shifts kv_self_update(false); - llama_memory_state_ptr mstate; + llama_memory_context_ptr mctx; while (true) { - mstate = memory->init_batch(*balloc, cparams.n_ubatch, output_all); - if (!mstate) { + mctx = memory->init_batch(*balloc, cparams.n_ubatch, output_all); + if (!mctx) { return -2; } - switch (mstate->get_status()) { + switch (mctx->get_status()) { case LLAMA_MEMORY_STATUS_SUCCESS: { } break; case LLAMA_MEMORY_STATUS_NO_UPDATE: { - LLAMA_LOG_ERROR("%s: unexpected memory state status: %d\n", __func__, mstate->get_status()); + LLAMA_LOG_ERROR("%s: unexpected memory context status: %d\n", __func__, mctx->get_status()); return -2; } @@ -987,7 +987,7 @@ int llama_context::decode(const llama_batch & batch_inp) { int64_t n_outputs_prev = 0; do { - const auto & ubatch = mstate->get_ubatch(); + const auto & ubatch = mctx->get_ubatch(); // count the outputs in this ubatch { @@ -1009,7 +1009,7 @@ int llama_context::decode(const llama_batch & batch_inp) { ggml_backend_sched_set_eval_callback(sched.get(), cparams.cb_eval, cparams.cb_eval_user_data); ggml_status status; - const auto res = process_ubatch(ubatch, LLM_GRAPH_TYPE_DECODER, mstate.get(), status); + const auto res = process_ubatch(ubatch, LLM_GRAPH_TYPE_DECODER, mctx.get(), status); if (!res) { // the last ubatch failed or was aborted -> remove all positions of that ubatch from the KV cache @@ -1126,7 +1126,7 @@ int llama_context::decode(const llama_batch & batch_inp) { } n_outputs_prev += n_outputs; - } while (mstate->next()); + } while (mctx->next()); // set to total number of outputs in the batch, for use in llama_get_logits_ith n_outputs = n_outputs_all; @@ -1292,7 +1292,7 @@ ggml_cgraph * llama_context::graph_init() { return ggml_new_graph_custom(ctx_compute.get(), graph_max_nodes(), false); } -ggml_cgraph * llama_context::graph_reserve(uint32_t n_tokens, uint32_t n_seqs, uint32_t n_outputs, const llama_memory_state_i * mstate) { +ggml_cgraph * llama_context::graph_reserve(uint32_t n_tokens, uint32_t n_seqs, uint32_t n_outputs, const llama_memory_context_i * mctx) { LLAMA_LOG_DEBUG("%s: reserving a graph for ubatch with n_tokens = %4u, n_seqs = %2u, n_outputs = %4u\n", __func__, n_tokens, n_seqs, n_outputs); if (n_tokens % n_seqs != 0) { @@ -1312,7 +1312,7 @@ ggml_cgraph * llama_context::graph_reserve(uint32_t n_tokens, uint32_t n_seqs, u llama_ubatch ubatch = balloc.ubatch_reserve(n_tokens/n_seqs, n_seqs); auto * gf = graph_init(); - auto res = graph_build(ctx_compute.get(), gf, ubatch, LLM_GRAPH_TYPE_DEFAULT, mstate); + auto res = graph_build(ctx_compute.get(), gf, ubatch, LLM_GRAPH_TYPE_DEFAULT, mctx); this->n_outputs = save_n_outputs; @@ -1333,11 +1333,11 @@ ggml_cgraph * llama_context::graph_reserve(uint32_t n_tokens, uint32_t n_seqs, u } llm_graph_result_ptr llama_context::graph_build( - ggml_context * ctx, - ggml_cgraph * gf, - const llama_ubatch & ubatch, - llm_graph_type gtype, - const llama_memory_state_i * mstate) { + ggml_context * ctx, + ggml_cgraph * gf, + const llama_ubatch & ubatch, + llm_graph_type gtype, + const llama_memory_context_i * mctx) { return model.build_graph( { /*.ctx =*/ ctx, @@ -1349,7 +1349,7 @@ llm_graph_result_ptr llama_context::graph_build( /*.backend_cpu =*/ backend_cpu, /*.cvec =*/ &cvec, /*.loras =*/ &loras, - /*.mstate =*/ mstate, + /*.mctx =*/ mctx, /*.cross =*/ &cross, /*.n_outputs =*/ n_outputs, /*.cb =*/ graph_get_cb(), @@ -2042,8 +2042,8 @@ void llama_context::opt_epoch_iter( uint32_t n_outputs_all = n_tokens_all; - auto mstate = memory->init_batch(*balloc, cparams.n_ubatch, true); - if (!mstate || mstate->get_status() != LLAMA_MEMORY_STATUS_SUCCESS) { + auto mctx = memory->init_batch(*balloc, cparams.n_ubatch, true); + if (!mctx || mctx->get_status() != LLAMA_MEMORY_STATUS_SUCCESS) { LLAMA_LOG_ERROR("%s: could not initialize batch\n", __func__); break; } @@ -2056,17 +2056,17 @@ void llama_context::opt_epoch_iter( uint32_t pos_batch = 0; do { - const auto & ubatch = mstate->get_ubatch(); + const auto & ubatch = mctx->get_ubatch(); n_outputs = ubatch.n_tokens; - if (!mstate->apply()) { - LLAMA_LOG_ERROR("%s: failed to update the memory state\n", __func__); + if (!mctx->apply()) { + LLAMA_LOG_ERROR("%s: failed to update the memory context\n", __func__); break; } auto * gf = graph_init(); - auto res = graph_build(ctx_compute.get(), gf, ubatch, LLM_GRAPH_TYPE_DEFAULT, mstate.get()); + auto res = graph_build(ctx_compute.get(), gf, ubatch, LLM_GRAPH_TYPE_DEFAULT, mctx.get()); struct ggml_context * ctx_compute_opt; { @@ -2101,7 +2101,7 @@ void llama_context::opt_epoch_iter( ggml_free(ctx_compute_opt); pos_batch += ubatch.n_tokens; - } while (mstate->next()); + } while (mctx->next()); } } diff --git a/src/llama-context.h b/src/llama-context.h index 7d300c145..9ce05715a 100644 --- a/src/llama-context.h +++ b/src/llama-context.h @@ -18,7 +18,7 @@ class llama_io_read_i; class llama_io_write_i; struct llama_memory_i; -struct llama_memory_state_i; +struct llama_memory_context_i; struct llama_context { // init scheduler and compute buffers, reserve worst-case graphs @@ -93,14 +93,14 @@ struct llama_context { int32_t il_end); // process a single ubatch with a specific graph type - // if memory_state is provided, it will be applied first to the context's memory + // if memory_context is provided, it will be applied first to the context's memory // ret contains the status of the graph computation // returns nullptr only if ret != GGML_STATUS_SUCCESS llm_graph_result_ptr process_ubatch( - const llama_ubatch & ubatch, - llm_graph_type gtype, - llama_memory_state_i * mstate, - ggml_status & ret); + const llama_ubatch & ubatch, + llm_graph_type gtype, + llama_memory_context_i * mctx, + ggml_status & ret); int encode(const llama_batch & batch_inp); int decode(const llama_batch & batch_inp); @@ -197,15 +197,15 @@ public: ggml_status graph_compute(ggml_cgraph * gf, bool batched); // reserve a graph with a dummy ubatch of the specified size - ggml_cgraph * graph_reserve(uint32_t n_tokens, uint32_t n_seqs, uint32_t n_outputs, const llama_memory_state_i * mstate); + ggml_cgraph * graph_reserve(uint32_t n_tokens, uint32_t n_seqs, uint32_t n_outputs, const llama_memory_context_i * mctx); private: llm_graph_result_ptr graph_build( - ggml_context * ctx, - ggml_cgraph * gf, - const llama_ubatch & ubatch, - llm_graph_type gtype, - const llama_memory_state_i * mstate); + ggml_context * ctx, + ggml_cgraph * gf, + const llama_ubatch & ubatch, + llm_graph_type gtype, + const llama_memory_context_i * mctx); llm_graph_cb graph_get_cb() const; diff --git a/src/llama-graph.cpp b/src/llama-graph.cpp index 7e162c555..48589a50a 100644 --- a/src/llama-graph.cpp +++ b/src/llama-graph.cpp @@ -87,7 +87,7 @@ void llm_graph_input_pos_bucket::set_input(const llama_ubatch * ubatch) { void llm_graph_input_pos_bucket_kv::set_input(const llama_ubatch * ubatch) { if (pos_bucket) { - kv_state->set_input_pos_bucket(pos_bucket, ubatch); + mctx->set_input_pos_bucket(pos_bucket, ubatch); } } @@ -221,7 +221,7 @@ void llm_graph_input_cls::set_input(const llama_ubatch * ubatch) { void llm_graph_input_rs::set_input(const llama_ubatch * ubatch) { GGML_UNUSED(ubatch); - const int64_t n_rs = mem_state->get_n_rs(); + const int64_t n_rs = mctx->get_n_rs(); if (s_copy) { GGML_ASSERT(ggml_backend_buffer_is_host(s_copy->buffer)); @@ -229,7 +229,7 @@ void llm_graph_input_rs::set_input(const llama_ubatch * ubatch) { // assuming copy destinations ALWAYS happen ONLY on the cells between head and head+n for (uint32_t i = 0; i < n_rs; ++i) { - data[i] = mem_state->s_copy(i); + data[i] = mctx->s_copy(i); } } } @@ -282,17 +282,17 @@ void llm_graph_input_attn_no_cache::set_input(const llama_ubatch * ubatch) { void llm_graph_input_attn_kv_unified::set_input(const llama_ubatch * ubatch) { if (self_kq_mask) { - kv_state->set_input_kq_mask(self_kq_mask, ubatch, cparams.causal_attn); + mctx->set_input_kq_mask(self_kq_mask, ubatch, cparams.causal_attn); } } void llm_graph_input_attn_kv_unified_iswa::set_input(const llama_ubatch * ubatch) { if (self_kq_mask) { - kv_state->get_base()->set_input_kq_mask(self_kq_mask, ubatch, cparams.causal_attn); + mctx->get_base()->set_input_kq_mask(self_kq_mask, ubatch, cparams.causal_attn); } if (self_kq_mask_swa) { - kv_state->get_swa()->set_input_kq_mask(self_kq_mask_swa, ubatch, cparams.causal_attn); + mctx->get_swa()->set_input_kq_mask(self_kq_mask_swa, ubatch, cparams.causal_attn); } } @@ -334,10 +334,10 @@ void llm_graph_input_attn_cross::set_input(const llama_ubatch * ubatch) { void llm_graph_input_mem_hybrid::set_input(const llama_ubatch * ubatch) { if (self_kq_mask) { - mem_state->get_state_attn()->set_input_kq_mask(self_kq_mask, ubatch, cparams.causal_attn); + mctx->get_attn()->set_input_kq_mask(self_kq_mask, ubatch, cparams.causal_attn); } - const int64_t n_rs = mem_state->get_state_recr()->get_n_rs(); + const int64_t n_rs = mctx->get_recr()->get_n_rs(); if (s_copy) { GGML_ASSERT(ggml_backend_buffer_is_host(s_copy->buffer)); @@ -345,7 +345,7 @@ void llm_graph_input_mem_hybrid::set_input(const llama_ubatch * ubatch) { // assuming copy destinations ALWAYS happen ONLY on the cells between head and head+n for (uint32_t i = 0; i < n_rs; ++i) { - data[i] = mem_state->get_state_recr()->s_copy(i); + data[i] = mctx->get_recr()->s_copy(i); } } } @@ -389,7 +389,7 @@ llm_graph_context::llm_graph_context(const llm_graph_params & params) : backend_cpu (params.backend_cpu), cvec (params.cvec), loras (params.loras), - mstate (params.mstate), + mctx (params.mctx), cross (params.cross), cb_func (params.cb), res (std::make_unique()) { @@ -950,11 +950,11 @@ ggml_tensor * llm_graph_context::build_inp_pos_bucket_enc() const { } ggml_tensor * llm_graph_context::build_inp_pos_bucket_dec() const { - const auto * kv_state = static_cast(mstate); + const auto * mctx_cur = static_cast(mctx); - auto inp = std::make_unique(hparams, kv_state); + auto inp = std::make_unique(hparams, mctx_cur); - const auto n_kv = kv_state->get_n_kv(); + const auto n_kv = mctx_cur->get_n_kv(); auto & cur = inp->pos_bucket; @@ -982,14 +982,14 @@ ggml_tensor * llm_graph_context::build_pos_bias(ggml_tensor * pos_bucket, ggml_t } llm_graph_input_mem_hybrid * llm_graph_context::build_inp_mem_hybrid() const { - const auto * mem_state = static_cast(mstate); + const auto * mctx_cur = static_cast(mctx); - auto inp = std::make_unique(hparams, cparams, mem_state); + auto inp = std::make_unique(hparams, cparams, mctx_cur); { GGML_ASSERT(hparams.swa_type == LLAMA_SWA_TYPE_NONE && "Hybrid recurrent is not supported with SWA attention layers"); - const auto n_kv = inp->mem_state->get_state_attn()->get_n_kv(); + const auto n_kv = inp->mctx->get_attn()->get_n_kv(); inp->self_kq_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD)); //cb(inp->self_kq_mask, "KQ_mask", -1); @@ -999,7 +999,7 @@ llm_graph_input_mem_hybrid * llm_graph_context::build_inp_mem_hybrid() const { } { - const auto n_rs = mem_state->get_state_recr()->get_n_rs(); + const auto n_rs = mctx_cur->get_recr()->get_n_rs(); inp->s_copy = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_rs); ggml_set_input(inp->s_copy); @@ -1183,14 +1183,14 @@ ggml_tensor * llm_graph_context::build_attn( } llm_graph_input_attn_kv_unified * llm_graph_context::build_attn_inp_kv_unified() const { - const auto * kv_state = static_cast(mstate); + const auto * mctx_cur = static_cast(mctx); - auto inp = std::make_unique(hparams, cparams, kv_state); + auto inp = std::make_unique(hparams, cparams, mctx_cur); { GGML_ASSERT(hparams.swa_type == LLAMA_SWA_TYPE_NONE && "Use llama_kv_cache_unified_iswa for SWA"); - const auto n_kv = kv_state->get_n_kv(); + const auto n_kv = mctx_cur->get_n_kv(); inp->self_kq_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD)); //cb(inp->self_kq_mask, "KQ_mask", -1); @@ -1220,19 +1220,19 @@ ggml_tensor * llm_graph_context::build_attn( ggml_build_forward_expand(gf, k_cur); ggml_build_forward_expand(gf, v_cur); - const auto * kv_state = static_cast(mstate); + const auto * mctx_cur = static_cast(mctx); // store to KV cache { - ggml_build_forward_expand(gf, kv_state->cpy_k(ctx0, k_cur, il)); - ggml_build_forward_expand(gf, kv_state->cpy_v(ctx0, v_cur, il)); + ggml_build_forward_expand(gf, mctx_cur->cpy_k(ctx0, k_cur, il)); + ggml_build_forward_expand(gf, mctx_cur->cpy_v(ctx0, v_cur, il)); } const auto & kq_mask = inp->get_kq_mask(); ggml_tensor * q = q_cur; - ggml_tensor * k = kv_state->get_k(ctx0, il); - ggml_tensor * v = kv_state->get_v(ctx0, il); + ggml_tensor * k = mctx_cur->get_k(ctx0, il); + ggml_tensor * v = mctx_cur->get_v(ctx0, il); ggml_tensor * cur = build_attn_mha(gf, q, k, v, kq_b, kq_mask, v_mla, kq_scale); cb(cur, "kqv_out", il); @@ -1270,23 +1270,23 @@ ggml_tensor * llm_graph_context::build_attn( ggml_build_forward_expand(gf, k_cur); ggml_build_forward_expand(gf, v_cur); - const auto * kv_state_iswa = static_cast(mstate); + const auto * mctx_iswa = static_cast(mctx); const bool is_swa = hparams.is_swa(il); - const auto * kv_state = is_swa ? kv_state_iswa->get_swa() : kv_state_iswa->get_base(); + const auto * mctx_cur = is_swa ? mctx_iswa->get_swa() : mctx_iswa->get_base(); // store to KV cache { - ggml_build_forward_expand(gf, kv_state->cpy_k(ctx0, k_cur, il)); - ggml_build_forward_expand(gf, kv_state->cpy_v(ctx0, v_cur, il)); + ggml_build_forward_expand(gf, mctx_cur->cpy_k(ctx0, k_cur, il)); + ggml_build_forward_expand(gf, mctx_cur->cpy_v(ctx0, v_cur, il)); } const auto & kq_mask = is_swa ? inp->get_kq_mask_swa() : inp->get_kq_mask(); ggml_tensor * q = q_cur; - ggml_tensor * k = kv_state->get_k(ctx0, il); - ggml_tensor * v = kv_state->get_v(ctx0, il); + ggml_tensor * k = mctx_cur->get_k(ctx0, il); + ggml_tensor * v = mctx_cur->get_v(ctx0, il); ggml_tensor * cur = build_attn_mha(gf, q, k, v, kq_b, kq_mask, v_mla, kq_scale); cb(cur, "kqv_out", il); @@ -1379,19 +1379,19 @@ ggml_tensor * llm_graph_context::build_attn( ggml_build_forward_expand(gf, k_cur); ggml_build_forward_expand(gf, v_cur); - const auto * kv_state = static_cast(mstate)->get_state_attn(); + const auto * mctx_cur = static_cast(mctx)->get_attn(); // store to KV cache { - ggml_build_forward_expand(gf, kv_state->cpy_k(ctx0, k_cur, il)); - ggml_build_forward_expand(gf, kv_state->cpy_v(ctx0, v_cur, il)); + ggml_build_forward_expand(gf, mctx_cur->cpy_k(ctx0, k_cur, il)); + ggml_build_forward_expand(gf, mctx_cur->cpy_v(ctx0, v_cur, il)); } const auto & kq_mask = inp->get_kq_mask(); ggml_tensor * q = q_cur; - ggml_tensor * k = kv_state->get_k(ctx0, il); - ggml_tensor * v = kv_state->get_v(ctx0, il); + ggml_tensor * k = mctx_cur->get_k(ctx0, il); + ggml_tensor * v = mctx_cur->get_v(ctx0, il); ggml_tensor * cur = build_attn_mha(gf, q, k, v, kq_b, kq_mask, v_mla, kq_scale); cb(cur, "kqv_out", il); @@ -1412,12 +1412,12 @@ ggml_tensor * llm_graph_context::build_attn( } llm_graph_input_attn_kv_unified_iswa * llm_graph_context::build_attn_inp_kv_unified_iswa() const { - const auto * kv_state = static_cast(mstate); + const auto * mctx_cur = static_cast(mctx); - auto inp = std::make_unique(hparams, cparams, kv_state); + auto inp = std::make_unique(hparams, cparams, mctx_cur); { - const auto n_kv = kv_state->get_base()->get_n_kv(); + const auto n_kv = mctx_cur->get_base()->get_n_kv(); inp->self_kq_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD)); //cb(inp->self_kq_mask, "KQ_mask", -1); @@ -1429,7 +1429,7 @@ llm_graph_input_attn_kv_unified_iswa * llm_graph_context::build_attn_inp_kv_unif { GGML_ASSERT(hparams.swa_type != LLAMA_SWA_TYPE_NONE && "Use llama_kv_cache_unified for non-SWA"); - const auto n_kv = kv_state->get_swa()->get_n_kv(); + const auto n_kv = mctx_cur->get_swa()->get_n_kv(); inp->self_kq_mask_swa = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD)); //cb(inp->self_kq_mask_swa, "KQ_mask_swa", -1); @@ -1485,11 +1485,11 @@ ggml_tensor * llm_graph_context::build_rs( } llm_graph_input_rs * llm_graph_context::build_rs_inp() const { - const auto * kv_state = static_cast(mstate); + const auto * mctx_cur = static_cast(mctx); - auto inp = std::make_unique(kv_state); + auto inp = std::make_unique(mctx_cur); - const auto n_rs = kv_state->get_n_rs(); + const auto n_rs = mctx_cur->get_n_rs(); inp->s_copy = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_rs); ggml_set_input(inp->s_copy); @@ -1504,9 +1504,9 @@ ggml_tensor * llm_graph_context::build_rs( int32_t state_size, int32_t n_seqs, bool avoid_copies) const { - const auto * kv_state = static_cast(mstate); + const auto * mctx_cur = static_cast(mctx); - return build_rs(gf, s, inp->s_copy, state_size, n_seqs, kv_state->get_n_rs(), kv_state->get_head(), kv_state->get_size(), kv_state->get_rs_z(), avoid_copies); + return build_rs(gf, s, inp->s_copy, state_size, n_seqs, mctx_cur->get_n_rs(), mctx_cur->get_head(), mctx_cur->get_size(), mctx_cur->get_rs_z(), avoid_copies); } ggml_tensor * llm_graph_context::build_rs( @@ -1516,9 +1516,9 @@ ggml_tensor * llm_graph_context::build_rs( int32_t state_size, int32_t n_seqs, bool avoid_copies) const { - const auto * kv_state = static_cast(mstate)->get_state_recr(); + const auto * mctx_cur = static_cast(mctx)->get_recr(); - return build_rs(gf, s, inp->s_copy, state_size, n_seqs, kv_state->get_n_rs(), kv_state->get_head(), kv_state->get_size(), kv_state->get_rs_z(), avoid_copies); + return build_rs(gf, s, inp->s_copy, state_size, n_seqs, mctx_cur->get_n_rs(), mctx_cur->get_head(), mctx_cur->get_size(), mctx_cur->get_rs_z(), avoid_copies); } ggml_tensor * llm_graph_context::build_rwkv_token_shift_load( @@ -1526,13 +1526,13 @@ ggml_tensor * llm_graph_context::build_rwkv_token_shift_load( ggml_cgraph * gf, const llama_ubatch & ubatch, int il) const { - const auto * kv_state = static_cast(mstate); + const auto * mctx_cur = static_cast(mctx); const auto token_shift_count = hparams.token_shift_count; const int64_t n_seqs = ubatch.n_seqs; - ggml_tensor * token_shift_all = kv_state->get_r_l(il); + ggml_tensor * token_shift_all = mctx_cur->get_r_l(il); ggml_tensor * token_shift = build_rs( inp, gf, token_shift_all, @@ -1547,19 +1547,19 @@ ggml_tensor * llm_graph_context::build_rwkv_token_shift_store( ggml_tensor * token_shift, const llama_ubatch & ubatch, int il) const { - const auto * kv_state = static_cast(mstate); + const auto * mctx_cur = static_cast(mctx); const auto token_shift_count = hparams.token_shift_count; const auto n_embd = hparams.n_embd; const int64_t n_seqs = ubatch.n_seqs; - const auto kv_head = kv_state->get_head(); + const auto kv_head = mctx_cur->get_head(); return ggml_cpy( ctx0, ggml_view_1d(ctx0, token_shift, n_embd * n_seqs * token_shift_count, 0), - ggml_view_1d(ctx0, kv_state->get_r_l(il), hparams.n_embd_r()*n_seqs, hparams.n_embd_r()*kv_head*ggml_element_size(kv_state->get_r_l(il))) + ggml_view_1d(ctx0, mctx_cur->get_r_l(il), hparams.n_embd_r()*n_seqs, hparams.n_embd_r()*kv_head*ggml_element_size(mctx_cur->get_r_l(il))) ); } diff --git a/src/llama-graph.h b/src/llama-graph.h index 9e62fa607..b433f266d 100644 --- a/src/llama-graph.h +++ b/src/llama-graph.h @@ -17,12 +17,12 @@ struct ggml_tensor; struct llama_ubatch; struct llama_cparams; -struct llama_memory_state_i; +struct llama_memory_context_i; -class llama_kv_cache_unified_state; -class llama_kv_cache_unified_iswa_state; -class llama_memory_recurrent_state; -class llama_memory_hybrid_state; +class llama_kv_cache_unified_context; +class llama_kv_cache_unified_iswa_context; +class llama_memory_recurrent_context; +class llama_memory_hybrid_context; // certain models (typically multi-modal) can produce different types of graphs enum llm_graph_type { @@ -136,7 +136,7 @@ class llm_graph_input_pos_bucket_kv : public llm_graph_input_i { public: llm_graph_input_pos_bucket_kv( const llama_hparams & hparams, - const llama_kv_cache_unified_state * kv_state) : hparams(hparams), kv_state(kv_state) {} + const llama_kv_cache_unified_context * mctx) : hparams(hparams), mctx(mctx) {} virtual ~llm_graph_input_pos_bucket_kv() = default; void set_input(const llama_ubatch * ubatch) override; @@ -144,7 +144,8 @@ public: ggml_tensor * pos_bucket = nullptr; // I32 [n_kv, n_batch] const llama_hparams & hparams; - const llama_kv_cache_unified_state * kv_state; + + const llama_kv_cache_unified_context * mctx; }; class llm_graph_input_out_ids : public llm_graph_input_i { @@ -191,14 +192,14 @@ public: class llm_graph_input_rs : public llm_graph_input_i { public: - llm_graph_input_rs(const llama_memory_recurrent_state * mem_state) : mem_state(mem_state) {} + llm_graph_input_rs(const llama_memory_recurrent_context * mctx) : mctx(mctx) {} virtual ~llm_graph_input_rs() = default; void set_input(const llama_ubatch * ubatch) override; ggml_tensor * s_copy; // I32 [kv_size] - const llama_memory_recurrent_state * mem_state; + const llama_memory_recurrent_context * mctx; }; class llm_graph_input_cross_embd : public llm_graph_input_i { @@ -238,10 +239,10 @@ public: llm_graph_input_attn_kv_unified( const llama_hparams & hparams, const llama_cparams & cparams, - const llama_kv_cache_unified_state * kv_state) : + const llama_kv_cache_unified_context * mctx) : hparams(hparams), cparams(cparams), - kv_state(kv_state) { + mctx(mctx) { } ~llm_graph_input_attn_kv_unified() = default; @@ -255,7 +256,7 @@ public: const llama_hparams & hparams; const llama_cparams & cparams; - const llama_kv_cache_unified_state * kv_state; + const llama_kv_cache_unified_context * mctx; }; class llm_graph_input_attn_kv_unified_iswa : public llm_graph_input_i { @@ -263,10 +264,10 @@ public: llm_graph_input_attn_kv_unified_iswa( const llama_hparams & hparams, const llama_cparams & cparams, - const llama_kv_cache_unified_iswa_state * kv_state) : + const llama_kv_cache_unified_iswa_context * mctx) : hparams(hparams), cparams(cparams), - kv_state(kv_state) { + mctx(mctx) { } ~llm_graph_input_attn_kv_unified_iswa() = default; @@ -283,7 +284,7 @@ public: const llama_hparams & hparams; const llama_cparams & cparams; - const llama_kv_cache_unified_iswa_state * kv_state; + const llama_kv_cache_unified_iswa_context * mctx; }; class llm_graph_input_attn_cross : public llm_graph_input_i { @@ -306,10 +307,10 @@ public: llm_graph_input_mem_hybrid( const llama_hparams & hparams, const llama_cparams & cparams, - const llama_memory_hybrid_state * mem_state) : + const llama_memory_hybrid_context * mctx) : hparams(hparams), cparams(cparams), - mem_state(mem_state) { + mctx(mctx) { } virtual ~llm_graph_input_mem_hybrid() = default; @@ -325,7 +326,7 @@ public: const llama_hparams & hparams; const llama_cparams & cparams; - const llama_memory_hybrid_state * mem_state; + const llama_memory_hybrid_context * mctx; }; // @@ -401,10 +402,10 @@ struct llm_graph_params { ggml_backend_sched_t sched; ggml_backend_t backend_cpu; - const llama_adapter_cvec * cvec; - const llama_adapter_loras * loras; - const llama_memory_state_i * mstate; - const llama_cross * cross; + const llama_adapter_cvec * cvec; + const llama_adapter_loras * loras; + const llama_memory_context_i * mctx; + const llama_cross * cross; uint32_t n_outputs; @@ -453,10 +454,10 @@ struct llm_graph_context { ggml_backend_t backend_cpu; // TODO: needed by build_attn_mha, figure out a way to remove? - const llama_adapter_cvec * cvec; - const llama_adapter_loras * loras; - const llama_memory_state_i * mstate; - const llama_cross * cross; + const llama_adapter_cvec * cvec; + const llama_adapter_loras * loras; + const llama_memory_context_i * mctx; + const llama_cross * cross; const llm_graph_cb & cb_func; diff --git a/src/llama-kv-cache-unified-iswa.cpp b/src/llama-kv-cache-unified-iswa.cpp index 0ced340de..b9169299c 100644 --- a/src/llama-kv-cache-unified-iswa.cpp +++ b/src/llama-kv-cache-unified-iswa.cpp @@ -95,7 +95,7 @@ llama_pos llama_kv_cache_unified_iswa::seq_pos_max(llama_seq_id seq_id) const { return kv_swa->seq_pos_max(seq_id); } -llama_memory_state_ptr llama_kv_cache_unified_iswa::init_batch(llama_batch_allocr & balloc, uint32_t n_ubatch, bool embd_all) { +llama_memory_context_ptr llama_kv_cache_unified_iswa::init_batch(llama_batch_allocr & balloc, uint32_t n_ubatch, bool embd_all) { GGML_UNUSED(embd_all); // first try simple split @@ -125,7 +125,7 @@ llama_memory_state_ptr llama_kv_cache_unified_iswa::init_batch(llama_batch_alloc assert(heads_base.size() == heads_swa.size()); - return std::make_unique( + return std::make_unique( this, std::move(heads_base), std::move(heads_swa), std::move(ubatches)); } while (false); @@ -156,22 +156,22 @@ llama_memory_state_ptr llama_kv_cache_unified_iswa::init_batch(llama_batch_alloc assert(heads_base.size() == heads_swa.size()); - return std::make_unique( + return std::make_unique( this, std::move(heads_base), std::move(heads_swa), std::move(ubatches)); } while (false); // TODO: if we fail again, we should attempt different splitting strategies // but to do that properly, we first have to refactor the batches to be more flexible - return std::make_unique(LLAMA_MEMORY_STATUS_FAILED_PREPARE); + return std::make_unique(LLAMA_MEMORY_STATUS_FAILED_PREPARE); } -llama_memory_state_ptr llama_kv_cache_unified_iswa::init_full() { - return std::make_unique(this); +llama_memory_context_ptr llama_kv_cache_unified_iswa::init_full() { + return std::make_unique(this); } -llama_memory_state_ptr llama_kv_cache_unified_iswa::init_update(llama_context * lctx, bool optimize) { - return std::make_unique(this, lctx, optimize); +llama_memory_context_ptr llama_kv_cache_unified_iswa::init_update(llama_context * lctx, bool optimize) { + return std::make_unique(this, lctx, optimize); } bool llama_kv_cache_unified_iswa::get_can_shift() const { @@ -197,46 +197,46 @@ llama_kv_cache_unified * llama_kv_cache_unified_iswa::get_swa() const { } // -// llama_kv_cache_unified_iswa_state +// llama_kv_cache_unified_iswa_context // -llama_kv_cache_unified_iswa_state::llama_kv_cache_unified_iswa_state(llama_memory_status status) : status(status) {} +llama_kv_cache_unified_iswa_context::llama_kv_cache_unified_iswa_context(llama_memory_status status) : status(status) {} -llama_kv_cache_unified_iswa_state::llama_kv_cache_unified_iswa_state( +llama_kv_cache_unified_iswa_context::llama_kv_cache_unified_iswa_context( llama_kv_cache_unified_iswa * kv) : - state_base(kv->get_base()->init_full()), - state_swa (kv->get_swa ()->init_full()), - status(llama_memory_status_combine(state_base->get_status(), state_swa->get_status())) { + ctx_base(kv->get_base()->init_full()), + ctx_swa (kv->get_swa ()->init_full()), + status(llama_memory_status_combine(ctx_base->get_status(), ctx_swa->get_status())) { } -llama_kv_cache_unified_iswa_state::llama_kv_cache_unified_iswa_state( +llama_kv_cache_unified_iswa_context::llama_kv_cache_unified_iswa_context( llama_kv_cache_unified_iswa * kv, llama_context * lctx, bool optimize) : - state_base(kv->get_base()->init_update(lctx, optimize)), - state_swa (kv->get_swa ()->init_update(lctx, optimize)), - status(llama_memory_status_combine(state_base->get_status(), state_swa->get_status())) { + ctx_base(kv->get_base()->init_update(lctx, optimize)), + ctx_swa (kv->get_swa ()->init_update(lctx, optimize)), + status(llama_memory_status_combine(ctx_base->get_status(), ctx_swa->get_status())) { } -llama_kv_cache_unified_iswa_state::llama_kv_cache_unified_iswa_state( +llama_kv_cache_unified_iswa_context::llama_kv_cache_unified_iswa_context( llama_kv_cache_unified_iswa * kv, std::vector heads_base, std::vector heads_swa, std::vector ubatches) : ubatches(std::move(ubatches)), // note: here we copy the ubatches. not sure if this is ideal - state_base(new llama_kv_cache_unified_state(kv->get_base(), std::move(heads_base), this->ubatches)), - state_swa (new llama_kv_cache_unified_state(kv->get_swa (), std::move(heads_swa), this->ubatches)), - status(llama_memory_status_combine(state_base->get_status(), state_swa->get_status())) { + ctx_base(new llama_kv_cache_unified_context(kv->get_base(), std::move(heads_base), this->ubatches)), + ctx_swa (new llama_kv_cache_unified_context(kv->get_swa (), std::move(heads_swa), this->ubatches)), + status(llama_memory_status_combine(ctx_base->get_status(), ctx_swa->get_status())) { } -llama_kv_cache_unified_iswa_state:: ~llama_kv_cache_unified_iswa_state() = default; +llama_kv_cache_unified_iswa_context:: ~llama_kv_cache_unified_iswa_context() = default; -bool llama_kv_cache_unified_iswa_state::next() { +bool llama_kv_cache_unified_iswa_context::next() { assert(status == LLAMA_MEMORY_STATUS_SUCCESS); - state_base->next(); - state_swa ->next(); + ctx_base->next(); + ctx_swa ->next(); if (++i_next >= ubatches.size()) { return false; @@ -245,35 +245,35 @@ bool llama_kv_cache_unified_iswa_state::next() { return true; } -bool llama_kv_cache_unified_iswa_state::apply() { +bool llama_kv_cache_unified_iswa_context::apply() { assert(status == LLAMA_MEMORY_STATUS_SUCCESS); bool res = true; - res = res & state_base->apply(); - res = res & state_swa ->apply(); + res = res & ctx_base->apply(); + res = res & ctx_swa ->apply(); return res; } -llama_memory_status llama_kv_cache_unified_iswa_state::get_status() const { +llama_memory_status llama_kv_cache_unified_iswa_context::get_status() const { return status; } -const llama_ubatch & llama_kv_cache_unified_iswa_state::get_ubatch() const { +const llama_ubatch & llama_kv_cache_unified_iswa_context::get_ubatch() const { assert(status == LLAMA_MEMORY_STATUS_SUCCESS); return ubatches[i_next]; } -const llama_kv_cache_unified_state * llama_kv_cache_unified_iswa_state::get_base() const { +const llama_kv_cache_unified_context * llama_kv_cache_unified_iswa_context::get_base() const { assert(status == LLAMA_MEMORY_STATUS_SUCCESS); - return static_cast(state_base.get()); + return static_cast(ctx_base.get()); } -const llama_kv_cache_unified_state * llama_kv_cache_unified_iswa_state::get_swa() const { +const llama_kv_cache_unified_context * llama_kv_cache_unified_iswa_context::get_swa() const { assert(status == LLAMA_MEMORY_STATUS_SUCCESS); - return static_cast(state_swa.get()); + return static_cast(ctx_swa.get()); } diff --git a/src/llama-kv-cache-unified-iswa.h b/src/llama-kv-cache-unified-iswa.h index 071041585..46c1ed614 100644 --- a/src/llama-kv-cache-unified-iswa.h +++ b/src/llama-kv-cache-unified-iswa.h @@ -31,14 +31,14 @@ public: // llama_memory_i // - llama_memory_state_ptr init_batch( + llama_memory_context_ptr init_batch( llama_batch_allocr & balloc, uint32_t n_ubatch, bool embd_all) override; - llama_memory_state_ptr init_full() override; + llama_memory_context_ptr init_full() override; - llama_memory_state_ptr init_update(llama_context * lctx, bool optimize) override; + llama_memory_context_ptr init_update(llama_context * lctx, bool optimize) override; bool get_can_shift() const override; @@ -72,32 +72,32 @@ private: std::unique_ptr kv_swa; }; -class llama_kv_cache_unified_iswa_state : public llama_memory_state_i { +class llama_kv_cache_unified_iswa_context : public llama_memory_context_i { public: // used for errors - llama_kv_cache_unified_iswa_state(llama_memory_status status); + llama_kv_cache_unified_iswa_context(llama_memory_status status); - // used to create a full-cache state - llama_kv_cache_unified_iswa_state( + // used to create a full-cache context + llama_kv_cache_unified_iswa_context( llama_kv_cache_unified_iswa * kv); - // used to create an update state - llama_kv_cache_unified_iswa_state( + // used to create an update context + llama_kv_cache_unified_iswa_context( llama_kv_cache_unified_iswa * kv, llama_context * lctx, bool optimize); - // used to create a state from a batch - llama_kv_cache_unified_iswa_state( + // used to create a batch processing context from a batch + llama_kv_cache_unified_iswa_context( llama_kv_cache_unified_iswa * kv, std::vector heads_base, std::vector heads_swa, std::vector ubatches); - virtual ~llama_kv_cache_unified_iswa_state(); + virtual ~llama_kv_cache_unified_iswa_context(); // - // llama_memory_state_i + // llama_memory_context_i // bool next() override; @@ -107,11 +107,11 @@ public: const llama_ubatch & get_ubatch() const override; // - // llama_kv_cache_unified_iswa_state specific API + // llama_kv_cache_unified_iswa_context specific API // - const llama_kv_cache_unified_state * get_base() const; - const llama_kv_cache_unified_state * get_swa() const; + const llama_kv_cache_unified_context * get_base() const; + const llama_kv_cache_unified_context * get_swa() const; private: //llama_kv_cache_unified_iswa * kv; @@ -121,8 +121,8 @@ private: std::vector ubatches; - const llama_memory_state_ptr state_base; - const llama_memory_state_ptr state_swa; + const llama_memory_context_ptr ctx_base; + const llama_memory_context_ptr ctx_swa; const llama_memory_status status; }; diff --git a/src/llama-kv-cache-unified.cpp b/src/llama-kv-cache-unified.cpp index 6897b7971..b506d32ed 100644 --- a/src/llama-kv-cache-unified.cpp +++ b/src/llama-kv-cache-unified.cpp @@ -307,7 +307,7 @@ llama_pos llama_kv_cache_unified::seq_pos_max(llama_seq_id seq_id) const { return cells.seq_pos_max(seq_id); } -llama_memory_state_ptr llama_kv_cache_unified::init_batch( +llama_memory_context_ptr llama_kv_cache_unified::init_batch( llama_batch_allocr & balloc, uint32_t n_ubatch, bool embd_all) { @@ -332,18 +332,18 @@ llama_memory_state_ptr llama_kv_cache_unified::init_batch( break; } - return std::make_unique( + return std::make_unique( this, std::move(heads), std::move(ubatches)); } while (false); - return std::make_unique(LLAMA_MEMORY_STATUS_FAILED_PREPARE); + return std::make_unique(LLAMA_MEMORY_STATUS_FAILED_PREPARE); } -llama_memory_state_ptr llama_kv_cache_unified::init_full() { - return std::make_unique(this); +llama_memory_context_ptr llama_kv_cache_unified::init_full() { + return std::make_unique(this); } -llama_memory_state_ptr llama_kv_cache_unified::init_update(llama_context * lctx, bool optimize) { +llama_memory_context_ptr llama_kv_cache_unified::init_update(llama_context * lctx, bool optimize) { bool do_shift = get_has_shift(); defrag_info dinfo; @@ -373,7 +373,7 @@ llama_memory_state_ptr llama_kv_cache_unified::init_update(llama_context * lctx, } } - return std::make_unique(this, lctx, do_shift, std::move(dinfo)); + return std::make_unique(this, lctx, do_shift, std::move(dinfo)); } llama_kv_cache_unified::ubatch_heads llama_kv_cache_unified::prepare(const std::vector & ubatches) { @@ -1710,18 +1710,18 @@ bool llama_kv_cache_unified::state_read_data(llama_io_read_i & io, uint32_t cell } // -// llama_kv_cache_unified_state +// llama_kv_cache_unified_context // -llama_kv_cache_unified_state::llama_kv_cache_unified_state(llama_memory_status status) : status(status) {} +llama_kv_cache_unified_context::llama_kv_cache_unified_context(llama_memory_status status) : status(status) {} -llama_kv_cache_unified_state::llama_kv_cache_unified_state( +llama_kv_cache_unified_context::llama_kv_cache_unified_context( llama_kv_cache_unified * kv) : status(LLAMA_MEMORY_STATUS_SUCCESS), kv(kv) { n_kv = kv->get_size(); head = 0; } -llama_kv_cache_unified_state::llama_kv_cache_unified_state( +llama_kv_cache_unified_context::llama_kv_cache_unified_context( llama_kv_cache_unified * kv, llama_context * lctx, bool do_shift, @@ -1731,15 +1731,15 @@ llama_kv_cache_unified_state::llama_kv_cache_unified_state( } } -llama_kv_cache_unified_state::llama_kv_cache_unified_state( +llama_kv_cache_unified_context::llama_kv_cache_unified_context( llama_kv_cache_unified * kv, llama_kv_cache_unified::ubatch_heads heads, std::vector ubatches) : status(LLAMA_MEMORY_STATUS_SUCCESS), kv(kv), heads(std::move(heads)), ubatches(std::move(ubatches)) { } -llama_kv_cache_unified_state::~llama_kv_cache_unified_state() = default; +llama_kv_cache_unified_context::~llama_kv_cache_unified_context() = default; -bool llama_kv_cache_unified_state::next() { +bool llama_kv_cache_unified_context::next() { assert(status == LLAMA_MEMORY_STATUS_SUCCESS); if (++i_next >= ubatches.size()) { @@ -1749,7 +1749,7 @@ bool llama_kv_cache_unified_state::next() { return true; } -bool llama_kv_cache_unified_state::apply() { +bool llama_kv_cache_unified_context::apply() { assert(status == LLAMA_MEMORY_STATUS_SUCCESS); // no ubatches -> this is a KV cache update @@ -1767,45 +1767,45 @@ bool llama_kv_cache_unified_state::apply() { return true; } -llama_memory_status llama_kv_cache_unified_state::get_status() const { +llama_memory_status llama_kv_cache_unified_context::get_status() const { return status; } -const llama_ubatch & llama_kv_cache_unified_state::get_ubatch() const { +const llama_ubatch & llama_kv_cache_unified_context::get_ubatch() const { assert(status == LLAMA_MEMORY_STATUS_SUCCESS); return ubatches[i_next]; } -uint32_t llama_kv_cache_unified_state::get_n_kv() const { +uint32_t llama_kv_cache_unified_context::get_n_kv() const { return n_kv; } -ggml_tensor * llama_kv_cache_unified_state::get_k(ggml_context * ctx, int32_t il) const { +ggml_tensor * llama_kv_cache_unified_context::get_k(ggml_context * ctx, int32_t il) const { return kv->get_k(ctx, il, n_kv); } -ggml_tensor * llama_kv_cache_unified_state::get_v(ggml_context * ctx, int32_t il) const { +ggml_tensor * llama_kv_cache_unified_context::get_v(ggml_context * ctx, int32_t il) const { return kv->get_v(ctx, il, n_kv); } -ggml_tensor * llama_kv_cache_unified_state::cpy_k(ggml_context * ctx, ggml_tensor * k_cur, int32_t il) const { +ggml_tensor * llama_kv_cache_unified_context::cpy_k(ggml_context * ctx, ggml_tensor * k_cur, int32_t il) const { return kv->cpy_k(ctx, k_cur, il, head); } -ggml_tensor * llama_kv_cache_unified_state::cpy_v(ggml_context * ctx, ggml_tensor * v_cur, int32_t il) const { +ggml_tensor * llama_kv_cache_unified_context::cpy_v(ggml_context * ctx, ggml_tensor * v_cur, int32_t il) const { return kv->cpy_v(ctx, v_cur, il, head); } -void llama_kv_cache_unified_state::set_input_k_shift(ggml_tensor * dst) const { +void llama_kv_cache_unified_context::set_input_k_shift(ggml_tensor * dst) const { kv->set_input_k_shift(dst); } -void llama_kv_cache_unified_state::set_input_kq_mask(ggml_tensor * dst, const llama_ubatch * ubatch, bool causal_attn) const { +void llama_kv_cache_unified_context::set_input_kq_mask(ggml_tensor * dst, const llama_ubatch * ubatch, bool causal_attn) const { kv->set_input_kq_mask(dst, ubatch, causal_attn); } -void llama_kv_cache_unified_state::set_input_pos_bucket(ggml_tensor * dst, const llama_ubatch * ubatch) const { +void llama_kv_cache_unified_context::set_input_pos_bucket(ggml_tensor * dst, const llama_ubatch * ubatch) const { kv->set_input_pos_bucket(dst, ubatch); } diff --git a/src/llama-kv-cache-unified.h b/src/llama-kv-cache-unified.h index 156064004..4c53f1273 100644 --- a/src/llama-kv-cache-unified.h +++ b/src/llama-kv-cache-unified.h @@ -56,14 +56,14 @@ public: // llama_memory_i // - llama_memory_state_ptr init_batch( + llama_memory_context_ptr init_batch( llama_batch_allocr & balloc, uint32_t n_ubatch, bool embd_all) override; - llama_memory_state_ptr init_full() override; + llama_memory_context_ptr init_full() override; - llama_memory_state_ptr init_update(llama_context * lctx, bool optimize) override; + llama_memory_context_ptr init_update(llama_context * lctx, bool optimize) override; bool get_can_shift() const override; @@ -208,36 +208,36 @@ private: bool state_read_data(llama_io_read_i & io, uint32_t cell_count); }; -class llama_kv_cache_unified_state : public llama_memory_state_i { +class llama_kv_cache_unified_context : public llama_memory_context_i { public: // some shorthands using ubatch_heads = llama_kv_cache_unified::ubatch_heads; using defrag_info = llama_kv_cache_unified::defrag_info; // used for errors - llama_kv_cache_unified_state(llama_memory_status status); + llama_kv_cache_unified_context(llama_memory_status status); - // used to create a full-cache state - llama_kv_cache_unified_state( + // used to create a full-cache context + llama_kv_cache_unified_context( llama_kv_cache_unified * kv); - // used to create an update state - llama_kv_cache_unified_state( + // used to create an update context + llama_kv_cache_unified_context( llama_kv_cache_unified * kv, llama_context * lctx, bool do_shift, defrag_info dinfo); - // used to create a decode state from a batch - llama_kv_cache_unified_state( + // used to create a batch procesing context from a batch + llama_kv_cache_unified_context( llama_kv_cache_unified * kv, ubatch_heads heads, std::vector ubatches); - virtual ~llama_kv_cache_unified_state(); + virtual ~llama_kv_cache_unified_context(); // - // llama_memory_state_i + // llama_memory_context_i // bool next() override; @@ -247,7 +247,7 @@ public: const llama_ubatch & get_ubatch() const override; // - // llama_kv_cache_unified_state specific API + // llama_kv_cache_unified_context specific API // uint32_t get_n_kv() const; @@ -272,7 +272,7 @@ private: llama_context * lctx; // - // update state + // update context // bool do_shift = false; @@ -280,7 +280,7 @@ private: defrag_info dinfo; // - // batch processing state + // batch processing context // // the index of the next ubatch to process diff --git a/src/llama-memory-hybrid.cpp b/src/llama-memory-hybrid.cpp index 1b1668681..15cde98d1 100644 --- a/src/llama-memory-hybrid.cpp +++ b/src/llama-memory-hybrid.cpp @@ -56,7 +56,7 @@ llama_memory_hybrid::llama_memory_hybrid( n_seq_max )) {} -llama_memory_state_ptr llama_memory_hybrid::init_batch(llama_batch_allocr & balloc, uint32_t n_ubatch, bool embd_all) { +llama_memory_context_ptr llama_memory_hybrid::init_batch(llama_batch_allocr & balloc, uint32_t n_ubatch, bool embd_all) { do { balloc.split_reset(); @@ -82,31 +82,31 @@ llama_memory_state_ptr llama_memory_hybrid::init_batch(llama_batch_allocr & ball // prepare the recurrent batches first if (!mem_recr->prepare(ubatches)) { - // TODO: will the recurrent cache be in an undefined state at this point? + // TODO: will the recurrent cache be in an undefined context at this point? LLAMA_LOG_ERROR("%s: failed to prepare recurrent ubatches\n", __func__); - return std::make_unique(LLAMA_MEMORY_STATUS_FAILED_PREPARE); + return std::make_unique(LLAMA_MEMORY_STATUS_FAILED_PREPARE); } // prepare the attention cache auto heads_attn = mem_attn->prepare(ubatches); if (heads_attn.empty()) { LLAMA_LOG_ERROR("%s: failed to prepare attention ubatches\n", __func__); - return std::make_unique(LLAMA_MEMORY_STATUS_FAILED_PREPARE); + return std::make_unique(LLAMA_MEMORY_STATUS_FAILED_PREPARE); } - return std::make_unique( + return std::make_unique( this, std::move(heads_attn), std::move(ubatches)); } while(false); - return std::make_unique(LLAMA_MEMORY_STATUS_FAILED_PREPARE); + return std::make_unique(LLAMA_MEMORY_STATUS_FAILED_PREPARE); } -llama_memory_state_ptr llama_memory_hybrid::init_full() { - return std::make_unique(this); +llama_memory_context_ptr llama_memory_hybrid::init_full() { + return std::make_unique(this); } -llama_memory_state_ptr llama_memory_hybrid::init_update(llama_context * lctx, bool optimize) { - return std::make_unique(this, lctx, optimize); +llama_memory_context_ptr llama_memory_hybrid::init_update(llama_context * lctx, bool optimize) { + return std::make_unique(this, lctx, optimize); } bool llama_memory_hybrid::get_can_shift() const { @@ -176,39 +176,39 @@ llama_memory_recurrent * llama_memory_hybrid::get_mem_recr() const { return mem_recr.get(); } -llama_memory_hybrid_state::llama_memory_hybrid_state(llama_memory_status status) : status(status) {} +llama_memory_hybrid_context::llama_memory_hybrid_context(llama_memory_status status) : status(status) {} -llama_memory_hybrid_state::llama_memory_hybrid_state(llama_memory_hybrid * mem) : - state_attn(mem->get_mem_attn()->init_full()), - state_recr(mem->get_mem_recr()->init_full()), - status(llama_memory_status_combine(state_attn->get_status(), state_recr->get_status())) { +llama_memory_hybrid_context::llama_memory_hybrid_context(llama_memory_hybrid * mem) : + ctx_attn(mem->get_mem_attn()->init_full()), + ctx_recr(mem->get_mem_recr()->init_full()), + status(llama_memory_status_combine(ctx_attn->get_status(), ctx_recr->get_status())) { } -llama_memory_hybrid_state::llama_memory_hybrid_state( +llama_memory_hybrid_context::llama_memory_hybrid_context( llama_memory_hybrid * mem, llama_context * lctx, bool optimize) : - state_attn(mem->get_mem_attn()->init_update(lctx, optimize)), - state_recr(mem->get_mem_recr()->init_update(lctx, optimize)), - status(llama_memory_status_combine(state_attn->get_status(), state_recr->get_status())) { + ctx_attn(mem->get_mem_attn()->init_update(lctx, optimize)), + ctx_recr(mem->get_mem_recr()->init_update(lctx, optimize)), + status(llama_memory_status_combine(ctx_attn->get_status(), ctx_recr->get_status())) { } -llama_memory_hybrid_state::llama_memory_hybrid_state( +llama_memory_hybrid_context::llama_memory_hybrid_context( llama_memory_hybrid * mem, std::vector heads_attn, std::vector ubatches) : ubatches(std::move(ubatches)), // note: here we copy the ubatches. not sure if this is ideal - state_attn(new llama_kv_cache_unified_state(mem->get_mem_attn(), std::move(heads_attn), this->ubatches)), - state_recr(new llama_memory_recurrent_state(mem->get_mem_recr(), this->ubatches)), - status(llama_memory_status_combine(state_attn->get_status(), state_recr->get_status())) { + ctx_attn(new llama_kv_cache_unified_context(mem->get_mem_attn(), std::move(heads_attn), this->ubatches)), + ctx_recr(new llama_memory_recurrent_context(mem->get_mem_recr(), this->ubatches)), + status(llama_memory_status_combine(ctx_attn->get_status(), ctx_recr->get_status())) { } -bool llama_memory_hybrid_state::next() { +bool llama_memory_hybrid_context::next() { assert(status == LLAMA_MEMORY_STATUS_SUCCESS); - state_attn->next(); - state_recr->next(); + ctx_attn->next(); + ctx_recr->next(); if (++i_next >= ubatches.size()) { return false; @@ -217,30 +217,30 @@ bool llama_memory_hybrid_state::next() { return true; } -bool llama_memory_hybrid_state::apply() { +bool llama_memory_hybrid_context::apply() { assert(status == LLAMA_MEMORY_STATUS_SUCCESS); bool res = true; - res = res & state_attn->apply(); - res = res & state_recr->apply(); + res = res & ctx_attn->apply(); + res = res & ctx_recr->apply(); return res; } -llama_memory_status llama_memory_hybrid_state::get_status() const { +llama_memory_status llama_memory_hybrid_context::get_status() const { return status; } -const llama_ubatch & llama_memory_hybrid_state::get_ubatch() const { +const llama_ubatch & llama_memory_hybrid_context::get_ubatch() const { assert(status == LLAMA_MEMORY_STATUS_SUCCESS); return ubatches[i_next]; } -const llama_kv_cache_unified_state * llama_memory_hybrid_state::get_state_attn() const { - return static_cast(state_attn.get()); +const llama_kv_cache_unified_context * llama_memory_hybrid_context::get_attn() const { + return static_cast(ctx_attn.get()); } -const llama_memory_recurrent_state * llama_memory_hybrid_state::get_state_recr() const { - return static_cast(state_recr.get()); +const llama_memory_recurrent_context * llama_memory_hybrid_context::get_recr() const { + return static_cast(ctx_recr.get()); } diff --git a/src/llama-memory-hybrid.h b/src/llama-memory-hybrid.h index 4d27ab896..f0c2420e9 100644 --- a/src/llama-memory-hybrid.h +++ b/src/llama-memory-hybrid.h @@ -49,14 +49,14 @@ public: // llama_memory_i // - llama_memory_state_ptr init_batch( + llama_memory_context_ptr init_batch( llama_batch_allocr & balloc, uint32_t n_ubatch, bool embd_all) override; - llama_memory_state_ptr init_full() override; + llama_memory_context_ptr init_full() override; - llama_memory_state_ptr init_update(llama_context * lctx, bool optimize) override; + llama_memory_context_ptr init_update(llama_context * lctx, bool optimize) override; bool get_can_shift() const override; @@ -90,27 +90,27 @@ private: const std::unique_ptr mem_recr; }; -class llama_memory_hybrid_state : public llama_memory_state_i { +class llama_memory_hybrid_context : public llama_memory_context_i { public: // init failure - explicit llama_memory_hybrid_state(llama_memory_status status); + explicit llama_memory_hybrid_context(llama_memory_status status); // init full - explicit llama_memory_hybrid_state(llama_memory_hybrid * mem); + explicit llama_memory_hybrid_context(llama_memory_hybrid * mem); // init update - explicit llama_memory_hybrid_state( + explicit llama_memory_hybrid_context( llama_memory_hybrid * mem, llama_context * lctx, bool optimize); // init success - llama_memory_hybrid_state( + llama_memory_hybrid_context( llama_memory_hybrid * mem, std::vector heads_attn, std::vector ubatches); - ~llama_memory_hybrid_state() = default; + ~llama_memory_hybrid_context() = default; bool next() override; bool apply() override; @@ -119,11 +119,11 @@ public: const llama_ubatch & get_ubatch() const override; // - // llama_memory_hybrid_state + // llama_memory_hybrid_context // - const llama_kv_cache_unified_state * get_state_attn() const; - const llama_memory_recurrent_state * get_state_recr() const; + const llama_kv_cache_unified_context * get_attn() const; + const llama_memory_recurrent_context * get_recr() const; private: // the index of the next ubatch to process @@ -131,8 +131,8 @@ private: std::vector ubatches; - const llama_memory_state_ptr state_attn; - const llama_memory_state_ptr state_recr; + const llama_memory_context_ptr ctx_attn; + const llama_memory_context_ptr ctx_recr; const llama_memory_status status; }; diff --git a/src/llama-memory-recurrent.cpp b/src/llama-memory-recurrent.cpp index b064da008..1b1e95d56 100644 --- a/src/llama-memory-recurrent.cpp +++ b/src/llama-memory-recurrent.cpp @@ -362,7 +362,7 @@ llama_pos llama_memory_recurrent::seq_pos_max(llama_seq_id seq_id) const { return result; } -llama_memory_state_ptr llama_memory_recurrent::init_batch(llama_batch_allocr & balloc, uint32_t n_ubatch, bool embd_all) { +llama_memory_context_ptr llama_memory_recurrent::init_batch(llama_batch_allocr & balloc, uint32_t n_ubatch, bool embd_all) { std::vector ubatches; while (true) { @@ -383,21 +383,21 @@ llama_memory_state_ptr llama_memory_recurrent::init_batch(llama_batch_allocr & b } if (!prepare(ubatches)) { - return std::make_unique(LLAMA_MEMORY_STATUS_FAILED_PREPARE); + return std::make_unique(LLAMA_MEMORY_STATUS_FAILED_PREPARE); } - return std::make_unique(this, std::move(ubatches)); + return std::make_unique(this, std::move(ubatches)); } -llama_memory_state_ptr llama_memory_recurrent::init_full() { - return std::make_unique(this); +llama_memory_context_ptr llama_memory_recurrent::init_full() { + return std::make_unique(this); } -llama_memory_state_ptr llama_memory_recurrent::init_update(llama_context * lctx, bool optimize) { +llama_memory_context_ptr llama_memory_recurrent::init_update(llama_context * lctx, bool optimize) { GGML_UNUSED(lctx); GGML_UNUSED(optimize); - return std::make_unique(LLAMA_MEMORY_STATUS_NO_UPDATE); + return std::make_unique(LLAMA_MEMORY_STATUS_NO_UPDATE); } bool llama_memory_recurrent::prepare(const std::vector & ubatches) { @@ -1040,22 +1040,22 @@ bool llama_memory_recurrent::state_read_data(llama_io_read_i & io, uint32_t cell } // -// llama_memory_recurrent_state +// llama_memory_recurrent_context // -llama_memory_recurrent_state::llama_memory_recurrent_state(llama_memory_status status) : status(status) {} +llama_memory_recurrent_context::llama_memory_recurrent_context(llama_memory_status status) : status(status) {} -llama_memory_recurrent_state::llama_memory_recurrent_state( +llama_memory_recurrent_context::llama_memory_recurrent_context( llama_memory_recurrent * mem) : status(LLAMA_MEMORY_STATUS_SUCCESS), mem(mem), is_full(true) { } -llama_memory_recurrent_state::llama_memory_recurrent_state( +llama_memory_recurrent_context::llama_memory_recurrent_context( llama_memory_recurrent * mem, std::vector ubatches) : status(LLAMA_MEMORY_STATUS_SUCCESS), mem(mem), ubatches(std::move(ubatches)) {} -llama_memory_recurrent_state::~llama_memory_recurrent_state() = default; +llama_memory_recurrent_context::~llama_memory_recurrent_context() = default; -bool llama_memory_recurrent_state::next() { +bool llama_memory_recurrent_context::next() { assert(status == LLAMA_MEMORY_STATUS_SUCCESS); if (++i_next >= ubatches.size()) { @@ -1065,7 +1065,7 @@ bool llama_memory_recurrent_state::next() { return true; } -bool llama_memory_recurrent_state::apply() { +bool llama_memory_recurrent_context::apply() { assert(status == LLAMA_MEMORY_STATUS_SUCCESS); mem->find_slot(ubatches[i_next]); @@ -1073,40 +1073,40 @@ bool llama_memory_recurrent_state::apply() { return true; } -llama_memory_status llama_memory_recurrent_state::get_status() const { +llama_memory_status llama_memory_recurrent_context::get_status() const { return status; } -const llama_ubatch & llama_memory_recurrent_state::get_ubatch() const { +const llama_ubatch & llama_memory_recurrent_context::get_ubatch() const { assert(status == LLAMA_MEMORY_STATUS_SUCCESS); return ubatches[i_next]; } -uint32_t llama_memory_recurrent_state::get_n_rs() const { +uint32_t llama_memory_recurrent_context::get_n_rs() const { return is_full ? mem->size : mem->n; } -uint32_t llama_memory_recurrent_state::get_head() const { +uint32_t llama_memory_recurrent_context::get_head() const { return is_full ? 0 : mem->head; } -int32_t llama_memory_recurrent_state::get_rs_z() const { +int32_t llama_memory_recurrent_context::get_rs_z() const { return is_full ? 0 : mem->rs_z; } -uint32_t llama_memory_recurrent_state::get_size() const { +uint32_t llama_memory_recurrent_context::get_size() const { return mem->size; } -ggml_tensor * llama_memory_recurrent_state::get_r_l(int32_t il) const { +ggml_tensor * llama_memory_recurrent_context::get_r_l(int32_t il) const { return mem->r_l[il]; } -ggml_tensor * llama_memory_recurrent_state::get_s_l(int32_t il) const { +ggml_tensor * llama_memory_recurrent_context::get_s_l(int32_t il) const { return mem->s_l[il]; } -int32_t llama_memory_recurrent_state::s_copy(int i) const { +int32_t llama_memory_recurrent_context::s_copy(int i) const { return mem->cells[i + mem->head].src0; } diff --git a/src/llama-memory-recurrent.h b/src/llama-memory-recurrent.h index be58dae7c..4d094f9a0 100644 --- a/src/llama-memory-recurrent.h +++ b/src/llama-memory-recurrent.h @@ -11,8 +11,8 @@ // llama_memory_recurrent // -// TODO: extract the cache state used for graph computation into llama_memory_recurrent_state_i -// see the implementation of llama_kv_cache_unified_state_i for an example how to do it +// TODO: extract the cache state used for graph computation into llama_memory_recurrent_context_i +// see the implementation of llama_kv_cache_unified_context_i for an example how to do it class llama_memory_recurrent : public llama_memory_i { public: @@ -34,14 +34,14 @@ public: // llama_memory_i // - llama_memory_state_ptr init_batch( + llama_memory_context_ptr init_batch( llama_batch_allocr & balloc, uint32_t n_ubatch, bool embd_all) override; - llama_memory_state_ptr init_full() override; + llama_memory_context_ptr init_full() override; - llama_memory_state_ptr init_update(llama_context * lctx, bool optimize) override; + llama_memory_context_ptr init_update(llama_context * lctx, bool optimize) override; void clear(bool data) override; @@ -125,24 +125,24 @@ private: bool state_read_data(llama_io_read_i & io, uint32_t cell_count); }; -class llama_memory_recurrent_state : public llama_memory_state_i { +class llama_memory_recurrent_context : public llama_memory_context_i { public: // used for errors - llama_memory_recurrent_state(llama_memory_status status); + llama_memory_recurrent_context(llama_memory_status status); - // used to create a full-cache state - llama_memory_recurrent_state( + // used to create a full-cache or update context + llama_memory_recurrent_context( llama_memory_recurrent * mem); - // used to create a state from a batch - llama_memory_recurrent_state( + // used to create a batch processing context from a batch + llama_memory_recurrent_context( llama_memory_recurrent * mem, std::vector ubatches); - virtual ~llama_memory_recurrent_state(); + virtual ~llama_memory_recurrent_context(); // - // llama_memory_state_i + // llama_memory_context_i // bool next() override; @@ -152,7 +152,7 @@ public: const llama_ubatch & get_ubatch() const override; // - // llama_memory_recurrent_state specific API + // llama_memory_recurrent_context specific API // uint32_t get_n_rs() const; diff --git a/src/llama-memory.h b/src/llama-memory.h index d2ef0c2a3..16b7e5ee2 100644 --- a/src/llama-memory.h +++ b/src/llama-memory.h @@ -3,7 +3,6 @@ #include "llama.h" #include -#include struct llama_ubatch; @@ -28,23 +27,21 @@ enum llama_memory_status { LLAMA_MEMORY_STATUS_FAILED_COMPUTE, }; -// helper function for combining the status of two memory states +// helper function for combining the status of two memory contexts // useful for implementing hybrid memory types (e.g. iSWA) llama_memory_status llama_memory_status_combine(llama_memory_status s0, llama_memory_status s1); -// the interface for managing the memory state during batch processing +// the interface for managing the memory context during batch processing // this interface is implemented per memory type. see: -// - llama_kv_cache_unified_state -// - llama_kv_cache_unified_iswa_state +// - llama_kv_cache_unified_context +// - llama_kv_cache_unified_iswa_context // ... // -// the only method that can mutate the memory and the memory state is llama_memory_i::apply() -// -// TODO: rename to llama_memory_context_i ? -struct llama_memory_state_i { - virtual ~llama_memory_state_i() = default; +// the only method that should mutate the memory and the memory context is llama_memory_i::apply() +struct llama_memory_context_i { + virtual ~llama_memory_context_i() = default; - // consume the current ubatch from the state and proceed to the next one + // consume the current ubatch from the context and proceed to the next one // return false if we are done virtual bool next() = 0; @@ -55,11 +52,11 @@ struct llama_memory_state_i { // get the current ubatch virtual const llama_ubatch & get_ubatch() const = 0; - // get the status of the memory state - used for error handling and checking if any updates would be applied + // get the status of the memory context - used for error handling and checking if any updates would be applied virtual llama_memory_status get_status() const = 0; }; -using llama_memory_state_ptr = std::unique_ptr; +using llama_memory_context_ptr = std::unique_ptr; // general concept of LLM memory // the KV cache is a type of LLM memory, but there can be other types @@ -67,19 +64,19 @@ struct llama_memory_i { virtual ~llama_memory_i() = default; // split the input batch into a set of ubatches and verify that they can fit into the cache - // return a state object containing the ubatches and KV cache state required to process them - // check the llama_memory_state_i::get_status() for the result - virtual llama_memory_state_ptr init_batch( + // return a context object containing the ubatches and memory state required to process them + // check the llama_memory_context_i::get_status() for the result + virtual llama_memory_context_ptr init_batch( llama_batch_allocr & balloc, uint32_t n_ubatch, bool embd_all) = 0; // simulate full cache, used for allocating worst-case compute buffers - virtual llama_memory_state_ptr init_full() = 0; + virtual llama_memory_context_ptr init_full() = 0; // prepare for any pending memory updates, such as shifts, defrags, etc. // status == LLAMA_MEMORY_STATUS_NO_UPDATE if there is nothing to update - virtual llama_memory_state_ptr init_update(llama_context * lctx, bool optimize) = 0; + virtual llama_memory_context_ptr init_update(llama_context * lctx, bool optimize) = 0; // getters virtual bool get_can_shift() const = 0; diff --git a/src/llama-model.cpp b/src/llama-model.cpp index e2c82017f..9b19da984 100644 --- a/src/llama-model.cpp +++ b/src/llama-model.cpp @@ -9171,9 +9171,9 @@ struct llm_build_mamba : public llm_graph_context { ggml_tensor * cur, const llama_ubatch & ubatch, int il) const { - const auto * kv_state = static_cast(mstate); + const auto * mctx_cur = static_cast(mctx); - const auto kv_head = kv_state->get_head(); + const auto kv_head = mctx_cur->get_head(); const int64_t d_conv = hparams.ssm_d_conv; const int64_t d_inner = hparams.ssm_d_inner; @@ -9191,8 +9191,8 @@ struct llm_build_mamba : public llm_graph_context { GGML_ASSERT(ubatch.equal_seqs); GGML_ASSERT(ubatch.n_tokens == n_seq_tokens * n_seqs); - ggml_tensor * conv_states_all = kv_state->get_r_l(il); - ggml_tensor * ssm_states_all = kv_state->get_s_l(il); + ggml_tensor * conv_states_all = mctx_cur->get_r_l(il); + ggml_tensor * ssm_states_all = mctx_cur->get_s_l(il); // (ab)using the KV cache to store the states ggml_tensor * conv = build_rs( @@ -11916,7 +11916,7 @@ struct llm_build_rwkv6_base : public llm_graph_context { ggml_tensor * x_prev, const llama_ubatch & ubatch, int il) const { - const auto * kv_state = static_cast(mstate); + const auto * mctx_cur = static_cast(mctx); const auto n_tokens = ubatch.n_tokens; const auto n_seqs = ubatch.n_seqs; @@ -11926,7 +11926,7 @@ struct llm_build_rwkv6_base : public llm_graph_context { const auto n_head = n_embd / head_size; const auto n_head_kv = hparams.n_head_kv(il); - const auto kv_head = kv_state->get_head(); + const auto kv_head = mctx_cur->get_head(); const auto & layer = model.layers[il]; @@ -12038,7 +12038,7 @@ struct llm_build_rwkv6_base : public llm_graph_context { } ggml_tensor * wkv_state = build_rs( - inp, gf, kv_state->get_s_l(il), + inp, gf, mctx_cur->get_s_l(il), hparams.n_embd_s(), n_seqs); ggml_tensor * wkv_output; @@ -12057,9 +12057,9 @@ struct llm_build_rwkv6_base : public llm_graph_context { wkv_state, ggml_view_1d( ctx0, - kv_state->get_s_l(il), + mctx_cur->get_s_l(il), hparams.n_embd_s() * n_seqs, - hparams.n_embd_s() * kv_head * ggml_element_size(kv_state->get_s_l(il)) + hparams.n_embd_s() * kv_head * ggml_element_size(mctx_cur->get_s_l(il)) ) ) ); @@ -12313,7 +12313,7 @@ struct llm_build_rwkv7_base : public llm_graph_context { ggml_tensor *& first_layer_value, const llama_ubatch & ubatch, int il) const { - const auto * kv_state = static_cast(mstate); + const auto * mctx_cur = static_cast(mctx); const auto n_tokens = ubatch.n_tokens; const auto n_seqs = ubatch.n_seqs; @@ -12322,7 +12322,7 @@ struct llm_build_rwkv7_base : public llm_graph_context { const auto head_count = n_embd / head_size; const auto n_seq_tokens = ubatch.n_seq_tokens; - const auto kv_head = kv_state->get_head(); + const auto kv_head = mctx_cur->get_head(); const auto & layer = model.layers[il]; @@ -12393,7 +12393,7 @@ struct llm_build_rwkv7_base : public llm_graph_context { a = ggml_reshape_3d(ctx0, a, head_size, head_count, n_tokens); ggml_tensor * wkv_state = build_rs( - inp, gf, kv_state->get_s_l(il), + inp, gf, mctx_cur->get_s_l(il), hparams.n_embd_s(), n_seqs); ggml_tensor * wkv_output = ggml_rwkv_wkv7(ctx0, r, w, k, v, ggml_neg(ctx0, kk), ggml_mul(ctx0, kk, a), wkv_state); @@ -12407,9 +12407,9 @@ struct llm_build_rwkv7_base : public llm_graph_context { wkv_state, ggml_view_1d( ctx0, - kv_state->get_s_l(il), + mctx_cur->get_s_l(il), hparams.n_embd_s() * n_seqs, - hparams.n_embd_s() * kv_head * ggml_element_size(kv_state->get_s_l(il)) + hparams.n_embd_s() * kv_head * ggml_element_size(mctx_cur->get_s_l(il)) ) ) ); From 67ae5312e255ae97852a4a216e2245580bfafd72 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Sat, 21 Jun 2025 08:04:18 +0300 Subject: [PATCH 27/43] metal : fix thread-safety (#14300) ggml-ci --- ggml/src/ggml-metal/ggml-metal.m | 88 ++++++++++++++++++++++---------- 1 file changed, 60 insertions(+), 28 deletions(-) diff --git a/ggml/src/ggml-metal/ggml-metal.m b/ggml/src/ggml-metal/ggml-metal.m index 4e7f373cb..19f4d59e5 100644 --- a/ggml/src/ggml-metal/ggml-metal.m +++ b/ggml/src/ggml-metal/ggml-metal.m @@ -48,22 +48,28 @@ static struct ggml_backend_metal_device_context { int mtl_device_ref_count; id mtl_library; + NSLock * mtl_lock; + bool has_simdgroup_reduction; bool has_simdgroup_mm; bool has_residency_sets; bool has_bfloat; bool use_bfloat; + size_t max_size; + char name[128]; } g_ggml_ctx_dev_main = { /*.mtl_device =*/ nil, /*.mtl_device_ref_count =*/ 0, /*.mtl_library =*/ nil, + /*.mtl_lock =*/ nil, /*.has_simdgroup_reduction =*/ false, /*.has_simdgroup_mm =*/ false, /*.has_residency_sets =*/ false, /*.has_bfloat =*/ false, /*.use_bfloat =*/ false, + /*.max_size =*/ 0, /*.name =*/ "", }; @@ -71,6 +77,10 @@ static struct ggml_backend_metal_device_context { static id ggml_backend_metal_device_acq(struct ggml_backend_metal_device_context * ctx) { assert(ctx != NULL); + if (ctx->mtl_lock == nil) { + ctx->mtl_lock = [[NSLock alloc] init]; + } + if (ctx->mtl_device == nil) { ctx->mtl_device = MTLCreateSystemDefaultDevice(); } @@ -94,6 +104,8 @@ static id ggml_backend_metal_device_acq(struct ggml_backend_metal_dev ctx->use_bfloat = false; #endif + ctx->max_size = ctx->mtl_device.maxBufferLength; + strncpy(ctx->name, [[ctx->mtl_device name] UTF8String], sizeof(ctx->name) - 1); } @@ -110,6 +122,11 @@ static void ggml_backend_metal_device_rel(struct ggml_backend_metal_device_conte ctx->mtl_device_ref_count--; if (ctx->mtl_device_ref_count == 0) { + if (ctx->mtl_lock) { + [ctx->mtl_lock release]; + ctx->mtl_lock = nil; + } + if (ctx->mtl_library) { [ctx->mtl_library release]; ctx->mtl_library = nil; @@ -977,7 +994,7 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de struct ggml_backend_metal_context * ctx = calloc(1, sizeof(struct ggml_backend_metal_context)); struct ggml_backend_metal_device_context * ctx_dev = dev->context; - id device = ggml_backend_metal_device_acq(ctx_dev); + id device = ctx_dev->mtl_device; GGML_LOG_INFO("%s: picking default device: %s\n", __func__, [[device name] UTF8String]); @@ -991,9 +1008,16 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de ctx->d_queue = dispatch_queue_create("ggml-metal", DISPATCH_QUEUE_CONCURRENT); // load library - if (ctx_dev->mtl_library == nil) { - ctx_dev->mtl_library = ggml_metal_load_library(device, ctx_dev->use_bfloat); + { + [ctx_dev->mtl_lock lock]; + + if (ctx_dev->mtl_library == nil) { + ctx_dev->mtl_library = ggml_metal_load_library(device, ctx_dev->use_bfloat); + } + + [ctx_dev->mtl_lock unlock]; } + id metal_library = ctx_dev->mtl_library; if (metal_library == nil) { GGML_LOG_ERROR("%s: error: metal library is nil\n", __func__); @@ -5284,7 +5308,6 @@ static void ggml_backend_metal_buffer_free_buffer(ggml_backend_buffer_t buffer) } ggml_backend_metal_buffer_rset_free(ctx); - ggml_backend_metal_device_rel(buffer->buft->device->context); if (ctx->owned) { #if TARGET_OS_OSX @@ -5393,7 +5416,10 @@ static ggml_backend_buffer_t ggml_backend_metal_buffer_type_alloc_buffer(ggml_ba } struct ggml_backend_metal_device_context * ctx_dev = (struct ggml_backend_metal_device_context *)buft->device->context; - id device = ggml_backend_metal_device_acq(ctx_dev); + + GGML_ASSERT(ctx_dev->mtl_device != nil); + + id device = ctx_dev->mtl_device; ctx->all_data = ggml_metal_host_malloc(size_aligned); ctx->all_size = size_aligned; @@ -5416,14 +5442,12 @@ static ggml_backend_buffer_t ggml_backend_metal_buffer_type_alloc_buffer(ggml_ba if (size_aligned > 0 && (ctx->all_data == NULL || ctx->buffers[0].metal == nil)) { GGML_LOG_ERROR("%s: error: failed to allocate buffer, size = %8.2f MiB\n", __func__, size_aligned / 1024.0 / 1024.0); free(ctx); - ggml_backend_metal_device_rel(ctx_dev); return NULL; } if (!ggml_backend_metal_buffer_rset_init(ctx, ctx_dev, device)) { GGML_LOG_ERROR("%s: error: failed to initialize residency set\n", __func__); free(ctx); - ggml_backend_metal_device_rel(ctx_dev); return NULL; } @@ -5434,17 +5458,14 @@ static ggml_backend_buffer_t ggml_backend_metal_buffer_type_alloc_buffer(ggml_ba static size_t ggml_backend_metal_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) { return 32; + GGML_UNUSED(buft); } static size_t ggml_backend_metal_buffer_type_get_max_size(ggml_backend_buffer_type_t buft) { - id device = ggml_backend_metal_device_acq(buft->device->context); - const size_t max_size = device.maxBufferLength; - ggml_backend_metal_device_rel(buft->device->context); + const size_t max_size = ((struct ggml_backend_metal_device_context *)buft->device->context)->max_size; return max_size; - - GGML_UNUSED(buft); } static bool ggml_backend_metal_buffer_type_is_host(ggml_backend_buffer_type_t buft) { @@ -5517,7 +5538,10 @@ ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t siz } struct ggml_backend_metal_device_context * ctx_dev = &g_ggml_ctx_dev_main; - id device = ggml_backend_metal_device_acq(ctx_dev); + + GGML_ASSERT(ctx_dev->mtl_device != nil); + + id device = ctx_dev->mtl_device; // the buffer fits into the max buffer size allowed by the device if (size_aligned <= device.maxBufferLength) { @@ -5573,7 +5597,6 @@ ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t siz if (!ggml_backend_metal_buffer_rset_init(ctx, ctx_dev, device)) { GGML_LOG_ERROR("%s: error: failed to initialize residency set\n", __func__); free(ctx); - ggml_backend_metal_device_rel(ctx_dev); return NULL; } @@ -5589,10 +5612,8 @@ static const char * ggml_backend_metal_name(ggml_backend_t backend) { } static void ggml_backend_metal_free(ggml_backend_t backend) { - struct ggml_backend_metal_context * ctx = backend->context; - struct ggml_backend_metal_device_context * ctx_dev = backend->device->context; + struct ggml_backend_metal_context * ctx = backend->context; - ggml_backend_metal_device_rel(ctx_dev); ggml_metal_free(ctx); free(backend); @@ -5732,6 +5753,8 @@ bool ggml_backend_metal_supports_family(ggml_backend_t backend, int family) { struct ggml_backend_metal_device_context * ctx_dev = backend->device->context; + GGML_ASSERT(ctx_dev->mtl_device != nil); + return [ctx_dev->mtl_device supportsFamily:(MTLGPUFamilyApple1 + family - 1)]; } @@ -5751,10 +5774,7 @@ static const char * ggml_backend_metal_device_get_name(ggml_backend_dev_t dev) { } static const char * ggml_backend_metal_device_get_description(ggml_backend_dev_t dev) { - // acq/rel just to populate ctx->name in case it hasn't been done yet struct ggml_backend_metal_device_context * ctx_dev = (struct ggml_backend_metal_device_context *)dev->context; - ggml_backend_metal_device_acq(ctx_dev); - ggml_backend_metal_device_rel(ctx_dev); return ctx_dev->name; } @@ -5762,12 +5782,10 @@ static const char * ggml_backend_metal_device_get_description(ggml_backend_dev_t static void ggml_backend_metal_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) { if (@available(macOS 10.12, iOS 16.0, *)) { struct ggml_backend_metal_device_context * ctx_dev = (struct ggml_backend_metal_device_context *)dev->context; - id device = ggml_backend_metal_device_acq(ctx_dev); + id device = ctx_dev->mtl_device; *total = device.recommendedMaxWorkingSetSize; *free = *total - device.currentAllocatedSize; - - ggml_backend_metal_device_rel(ctx_dev); } else { *free = 1; *total = 1; @@ -5845,7 +5863,10 @@ static ggml_backend_buffer_t ggml_backend_metal_device_buffer_from_ptr(ggml_back } struct ggml_backend_metal_device_context * ctx_dev = (struct ggml_backend_metal_device_context *)dev->context; - id device = ggml_backend_metal_device_acq(ctx_dev); + + GGML_ASSERT(ctx_dev->mtl_device != nil); + + id device = ctx_dev->mtl_device; // the buffer fits into the max buffer size allowed by the device if (size_aligned <= device.maxBufferLength) { @@ -5901,7 +5922,6 @@ static ggml_backend_buffer_t ggml_backend_metal_device_buffer_from_ptr(ggml_back if (!ggml_backend_metal_buffer_rset_init(ctx, ctx_dev, device)) { GGML_LOG_ERROR("%s: error: failed to initialize residency set\n", __func__); free(ctx); - ggml_backend_metal_device_rel(ctx_dev); return NULL; } @@ -5915,8 +5935,9 @@ static bool ggml_backend_metal_device_supports_op(ggml_backend_dev_t dev, const } static bool ggml_backend_metal_device_supports_buft(ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft) { - return buft->iface.get_name == ggml_backend_metal_buffer_type_get_name || - buft->iface.get_name == ggml_backend_metal_buffer_from_ptr_type_get_name; + return + buft->iface.get_name == ggml_backend_metal_buffer_type_get_name || + buft->iface.get_name == ggml_backend_metal_buffer_from_ptr_type_get_name; GGML_UNUSED(dev); } @@ -6001,8 +6022,19 @@ static struct ggml_backend_reg_i ggml_backend_metal_reg_i = { /* .get_proc_address = */ ggml_backend_metal_get_proc_address, }; +// called upon program exit +static void ggml_metal_cleanup(void) { + ggml_backend_metal_device_rel(&g_ggml_ctx_dev_main); +} + +// TODO: make thread-safe ggml_backend_reg_t ggml_backend_metal_reg(void) { - // TODO: make this thread-safe somehow? + ggml_backend_metal_device_acq(&g_ggml_ctx_dev_main); + + // register cleanup callback + // TODO: not ideal, but not sure if there is a better way to do this in Objective-C + atexit(ggml_metal_cleanup); + { g_ggml_backend_metal_reg = (struct ggml_backend_reg) { /* .api_version = */ GGML_BACKEND_API_VERSION, From 58cba76a9aab728717509d62b67c13afd8dc227a Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Sigbj=C3=B8rn=20Skj=C3=A6ret?= Date: Sat, 21 Jun 2025 07:33:21 +0200 Subject: [PATCH 28/43] gguf-py : fix TemplateProcessing pair when bos/eos is missing (#14312) --- gguf-py/gguf/vocab.py | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/gguf-py/gguf/vocab.py b/gguf-py/gguf/vocab.py index 6c4d3a422..a792d56f0 100644 --- a/gguf-py/gguf/vocab.py +++ b/gguf-py/gguf/vocab.py @@ -201,9 +201,9 @@ class SpecialVocab: if special_last != special_eos: logger.warning(f'Unknown trailing special token {special_last!r} in TemplateProcessing') if tmpl_pair: - seq_start = 1 if tmpl_pair[0].get('SpecialToken', {}).get('id') == special_first else 0 - seq_stop = -1 if tmpl_pair[-1].get('SpecialToken', {}).get('id') == special_last else None - if seq_start == 0 or seq_stop is None: + seq_start = 1 if special_first and tmpl_pair[0].get('SpecialToken', {}).get('id') == special_first else 0 + seq_stop = -1 if special_last and tmpl_pair[-1].get('SpecialToken', {}).get('id') == special_last else None + if (special_first and seq_start == 0) or (special_last and seq_stop is None): logger.warning('TemplateProcessing leading/trailing special tokens do not match TemplateProcessing') if tmpl_pair := tmpl_pair[slice(seq_start, seq_stop)]: tmpl_a = tmpl_pair[0].get('Sequence', {}).get('id') From bb16041caef45cd4348cd6f84906b5dfec7a1f6a Mon Sep 17 00:00:00 2001 From: Markus Tavenrath Date: Sat, 21 Jun 2025 08:17:12 +0200 Subject: [PATCH 29/43] Add support for VK_EXT_debug_utils to add labels to Vulkan objects. (#13792) * Add support for VK_EXT_debug_utils to add labels to Vulkan objects. In step 1 compute pipelines are getting labeled. * remove #ifdef for debug utils and add queue marker. --- ggml/src/ggml-vulkan/ggml-vulkan.cpp | 58 +++++++++++++++++++++++++++- 1 file changed, 57 insertions(+), 1 deletion(-) diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp index 1375bfeb9..99be5e45b 100644 --- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp +++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp @@ -1041,6 +1041,14 @@ void vk_memory_logger::log_deallocation(vk_buffer_ref buf_ref) { struct vk_instance_t { vk::Instance instance; + bool debug_utils_support = false; // VK_EXT_debug_utils enabled + PFN_vkSetDebugUtilsObjectNameEXT pfn_vkSetDebugUtilsObjectNameEXT = {}; + PFN_vkQueueBeginDebugUtilsLabelEXT pfn_vkQueueBeginDebugUtilsLabelEXT = {}; + PFN_vkQueueEndDebugUtilsLabelEXT pfn_vkQueueEndDebugUtilsLabelEXT = {}; + PFN_vkCmdBeginDebugUtilsLabelEXT pfn_vkCmdBeginDebugUtilsLabelEXT = {}; + PFN_vkCmdEndDebugUtilsLabelEXT pfn_vkCmdEndDebugUtilsLabelEXT = {}; + PFN_vkCmdInsertDebugUtilsLabelEXT pfn_vkCmdInsertDebugUtilsLabelEXT = {}; + std::vector device_indices; vk_device devices[GGML_VK_MAX_DEVICES]; }; @@ -1180,6 +1188,14 @@ static void ggml_vk_create_pipeline_func(vk_device& device, vk_pipeline& pipelin } pipeline->compiled = true; + if (vk_instance.debug_utils_support) { + vk::DebugUtilsObjectNameInfoEXT duoni; + duoni.objectType = vk::ObjectType::ePipeline; + duoni.pObjectName = pipeline->name.c_str(); + duoni.objectHandle = reinterpret_cast(static_cast(pipeline->pipeline)); + vk_instance.pfn_vkSetDebugUtilsObjectNameEXT(device->device, &static_cast(duoni)); + } + { std::lock_guard guard(device->mutex); device->pipelines.insert({ pipeline->name, pipeline }); @@ -3561,6 +3577,8 @@ static void ggml_vk_print_gpu_info(size_t idx) { static bool ggml_vk_instance_validation_ext_available(const std::vector& instance_extensions); static bool ggml_vk_instance_portability_enumeration_ext_available(const std::vector& instance_extensions); +static bool ggml_vk_instance_debug_utils_ext_available(const std::vector & instance_extensions); + static void ggml_vk_instance_init() { if (vk_instance_initialized) { return; @@ -3581,7 +3599,7 @@ static void ggml_vk_instance_init() { #ifdef __APPLE__ const bool portability_enumeration_ext = ggml_vk_instance_portability_enumeration_ext_available(instance_extensions); #endif - + const bool debug_utils_ext = ggml_vk_instance_debug_utils_ext_available(instance_extensions) && getenv("GGML_VK_DEBUG_MARKERS") != nullptr; std::vector layers; if (validation_ext) { @@ -3596,6 +3614,9 @@ static void ggml_vk_instance_init() { extensions.push_back("VK_KHR_portability_enumeration"); } #endif + if (debug_utils_ext) { + extensions.push_back("VK_EXT_debug_utils"); + } vk::InstanceCreateInfo instance_create_info(vk::InstanceCreateFlags{}, &app_info, layers, extensions); #ifdef __APPLE__ if (portability_enumeration_ext) { @@ -3619,6 +3640,18 @@ static void ggml_vk_instance_init() { vk_instance.instance = vk::createInstance(instance_create_info); vk_instance_initialized = true; + if (debug_utils_ext) { + vk_instance.debug_utils_support = true; + vk_instance.pfn_vkSetDebugUtilsObjectNameEXT = (PFN_vkSetDebugUtilsObjectNameEXT) vkGetInstanceProcAddr(vk_instance.instance, "vkSetDebugUtilsObjectNameEXT"); + vk_instance.pfn_vkQueueBeginDebugUtilsLabelEXT = (PFN_vkQueueBeginDebugUtilsLabelEXT) vkGetInstanceProcAddr(vk_instance.instance, "vkQueueBeginDebugUtilsLabelEXT"); + vk_instance.pfn_vkQueueEndDebugUtilsLabelEXT = (PFN_vkQueueEndDebugUtilsLabelEXT) vkGetInstanceProcAddr(vk_instance.instance, "vkQueueEndDebugUtilsLabelEXT"); + vk_instance.pfn_vkCmdBeginDebugUtilsLabelEXT = (PFN_vkCmdBeginDebugUtilsLabelEXT) vkGetInstanceProcAddr(vk_instance.instance, "vkCmdBeginDebugUtilsLabelEXT"); + vk_instance.pfn_vkCmdEndDebugUtilsLabelEXT = (PFN_vkCmdEndDebugUtilsLabelEXT) vkGetInstanceProcAddr(vk_instance.instance, "vkCmdEndDebugUtilsLabelEXT"); + vk_instance.pfn_vkCmdInsertDebugUtilsLabelEXT = (PFN_vkCmdInsertDebugUtilsLabelEXT) vkGetInstanceProcAddr(vk_instance.instance, "vkCmdInsertDebugUtilsLabelEXT"); + + } + + size_t num_available_devices = vk_instance.instance.enumeratePhysicalDevices().size(); vk_perf_logger_enabled = getenv("GGML_VK_PERF_LOGGER") != nullptr; // Emulate behavior of CUDA_VISIBLE_DEVICES for Vulkan @@ -9656,6 +9689,13 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg VK_LOG_DEBUG("ggml_backend_vk_graph_compute(" << cgraph->n_nodes << " nodes)"); ggml_backend_vk_context * ctx = (ggml_backend_vk_context *)backend->context; + if (vk_instance.debug_utils_support) { + vk::DebugUtilsLabelEXT dul = {}; + dul.pLabelName = "ggml_backend_vk_graph_compute"; + dul.color = std::array{1.0f, 1.0f, 1.0f, 1.0f}; + vk_instance.pfn_vkQueueBeginDebugUtilsLabelEXT(ctx->device->compute_queue.queue, reinterpret_cast(&dul)); + } + uint64_t total_mat_mul_bytes = 0; for (int i = 0; i < cgraph->n_nodes; i++) { ggml_vk_build_graph(ctx, cgraph->nodes[i], i, nullptr, 0, true, false, false, false); @@ -10345,6 +10385,22 @@ static bool ggml_vk_instance_portability_enumeration_ext_available(const std::ve UNUSED(instance_extensions); } +// Extension availability +static bool ggml_vk_instance_debug_utils_ext_available( + const std::vector & instance_extensions) { + // Check for portability enumeration extension for MoltenVK support + for (const auto & properties : instance_extensions) { + if (strcmp("VK_EXT_debug_utils", properties.extensionName) == 0) { + return true; + } + } + + std::cerr << "ggml_vulkan: WARNING: Instance extension VK_EXT_debug_utils not found." << std::endl; + return false; + + UNUSED(instance_extensions); +} + static bool ggml_vk_khr_cooperative_matrix_support(const vk::PhysicalDeviceProperties& props, const vk::PhysicalDeviceDriverProperties& driver_props, vk_device_architecture arch) { switch (props.vendorID) { case VK_VENDOR_ID_INTEL: From aa0ef5c578eef4c2adc7be1282f21bab5f3e8d26 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Sigbj=C3=B8rn=20Skj=C3=A6ret?= Date: Sat, 21 Jun 2025 18:12:05 +0200 Subject: [PATCH 30/43] gguf-py : fix Qwen3-Embedding eos token (#14314) --- gguf-py/gguf/vocab.py | 10 ++++++++++ 1 file changed, 10 insertions(+) diff --git a/gguf-py/gguf/vocab.py b/gguf-py/gguf/vocab.py index a792d56f0..3b08f6134 100644 --- a/gguf-py/gguf/vocab.py +++ b/gguf-py/gguf/vocab.py @@ -197,6 +197,16 @@ class SpecialVocab: if special_last := tmpl_single[-1].get('SpecialToken', {}).get('id'): if not tokenizer_config: special_eos = special_last + elif special_last != special_eos: + if 'eot' not in self.special_token_types: + self.special_token_types = tuple(self.special_token_types) + ('eot', ) + tokenizer_config['eot_token'] = special_eos + elif 'eom' not in self.special_token_types: + self.special_token_types = tuple(self.special_token_types) + ('eom', ) + tokenizer_config['eom_token'] = special_eos + else: + logger.warning(f'Overriding EOS token {special_eos!r} with {special_last!r} without EOT/EOM fallback!') + tokenizer_config['eos_token'] = special_eos = special_last self.add_special_token['eos'] = True if special_last == special_eos else False if special_last != special_eos: logger.warning(f'Unknown trailing special token {special_last!r} in TemplateProcessing') From aa064b2eb7f7779db5e094a9d8d66d5033557de2 Mon Sep 17 00:00:00 2001 From: Aman Gupta Date: Sun, 22 Jun 2025 12:39:54 +0800 Subject: [PATCH 31/43] CUDA: add mean operation (#14313) * CUDA: add mean operation * add back sum_rows_f32_cuda * Review: early exit if col!=0 --- ggml/src/ggml-cuda/common.cuh | 20 ++++++++++++++++++++ ggml/src/ggml-cuda/ggml-cuda.cu | 5 +++++ ggml/src/ggml-cuda/mean.cu | 19 +++++++++++++++++++ ggml/src/ggml-cuda/mean.cuh | 3 +++ ggml/src/ggml-cuda/sumrows.cu | 23 +++++------------------ ggml/src/ggml-cuda/sumrows.cuh | 1 - tests/test-backend-ops.cpp | 2 ++ 7 files changed, 54 insertions(+), 19 deletions(-) create mode 100644 ggml/src/ggml-cuda/mean.cu create mode 100644 ggml/src/ggml-cuda/mean.cuh diff --git a/ggml/src/ggml-cuda/common.cuh b/ggml/src/ggml-cuda/common.cuh index 364efcaec..2f2fce067 100644 --- a/ggml/src/ggml-cuda/common.cuh +++ b/ggml/src/ggml-cuda/common.cuh @@ -362,6 +362,26 @@ static __device__ __forceinline__ half2 warp_reduce_sum(half2 a) { #endif // FP16_AVAILABLE } +// Row reduction kernel template - compute sum (norm=false) or mean (norm=true) +template +static __global__ void reduce_rows_f32(const float * x, float * dst, const int ncols) { + const int row = blockIdx.x; + const int col = threadIdx.x; + + float sum = 0.0f; + for (int i = col; i < ncols; i += blockDim.x) { + sum += x[row * ncols + i]; + } + + sum = warp_reduce_sum(sum); + + if (col != 0) { + return; + } + + dst[row] = norm ? sum / ncols : sum; +} + template static __device__ __forceinline__ float warp_reduce_max(float x) { #pragma unroll diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu index 5bab92e34..c6bdd4fb3 100644 --- a/ggml/src/ggml-cuda/ggml-cuda.cu +++ b/ggml/src/ggml-cuda/ggml-cuda.cu @@ -37,6 +37,7 @@ #include "ggml-cuda/ssm-scan.cuh" #include "ggml-cuda/sum.cuh" #include "ggml-cuda/sumrows.cuh" +#include "ggml-cuda/mean.cuh" #include "ggml-cuda/tsembd.cuh" #include "ggml-cuda/unary.cuh" #include "ggml-cuda/upscale.cuh" @@ -2357,6 +2358,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg case GGML_OP_SUM_ROWS: ggml_cuda_op_sum_rows(ctx, dst); break; + case GGML_OP_MEAN: + ggml_cuda_op_mean(ctx, dst); + break; case GGML_OP_SSM_CONV: ggml_cuda_op_ssm_conv(ctx, dst); break; @@ -3260,6 +3264,7 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g case GGML_OP_POOL_2D: case GGML_OP_SUM: case GGML_OP_SUM_ROWS: + case GGML_OP_MEAN: case GGML_OP_ARGSORT: case GGML_OP_ACC: return true; diff --git a/ggml/src/ggml-cuda/mean.cu b/ggml/src/ggml-cuda/mean.cu new file mode 100644 index 000000000..4b238a399 --- /dev/null +++ b/ggml/src/ggml-cuda/mean.cu @@ -0,0 +1,19 @@ +#include "mean.cuh" + +void ggml_cuda_op_mean(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { + const ggml_tensor * src0 = dst->src[0]; + const float * src0_d = (const float *) src0->data; + float * dst_d = (float *) dst->data; + cudaStream_t stream = ctx.stream(); + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT(dst->type == GGML_TYPE_F32); + GGML_ASSERT(ggml_is_contiguous(src0)); + + const int64_t ncols = src0->ne[0]; + const int64_t nrows = ggml_nrows(src0); + + const dim3 block_dims(WARP_SIZE, 1, 1); + const dim3 block_nums(nrows, 1, 1); + reduce_rows_f32<<>>(src0_d, dst_d, ncols); +} diff --git a/ggml/src/ggml-cuda/mean.cuh b/ggml/src/ggml-cuda/mean.cuh new file mode 100644 index 000000000..2b9b10433 --- /dev/null +++ b/ggml/src/ggml-cuda/mean.cuh @@ -0,0 +1,3 @@ +#include "common.cuh" + +void ggml_cuda_op_mean(ggml_backend_cuda_context & ctx, ggml_tensor * dst); diff --git a/ggml/src/ggml-cuda/sumrows.cu b/ggml/src/ggml-cuda/sumrows.cu index 38dbf1b5e..2eee08fa0 100644 --- a/ggml/src/ggml-cuda/sumrows.cu +++ b/ggml/src/ggml-cuda/sumrows.cu @@ -1,25 +1,9 @@ #include "sumrows.cuh" -static __global__ void k_sum_rows_f32(const float * x, float * dst, const int ncols) { - const int row = blockIdx.x; - const int col = threadIdx.x; - - float sum = 0.0f; - for (int i = col; i < ncols; i += blockDim.x) { - sum += x[row * ncols + i]; - } - - sum = warp_reduce_sum(sum); - - if (col == 0) { - dst[row] = sum; - } -} - void sum_rows_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, cudaStream_t stream) { const dim3 block_dims(WARP_SIZE, 1, 1); const dim3 block_nums(nrows, 1, 1); - k_sum_rows_f32<<>>(x, dst, ncols); + reduce_rows_f32<<>>(x, dst, ncols); } void ggml_cuda_op_sum_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { @@ -35,5 +19,8 @@ void ggml_cuda_op_sum_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { const int64_t ncols = src0->ne[0]; const int64_t nrows = ggml_nrows(src0); - sum_rows_f32_cuda(src0_d, dst_d, ncols, nrows, stream); + const dim3 block_dims(WARP_SIZE, 1, 1); + const dim3 block_nums(nrows, 1, 1); + + reduce_rows_f32<<>>(src0_d, dst_d, ncols); } diff --git a/ggml/src/ggml-cuda/sumrows.cuh b/ggml/src/ggml-cuda/sumrows.cuh index 191db1c13..3431c599b 100644 --- a/ggml/src/ggml-cuda/sumrows.cuh +++ b/ggml/src/ggml-cuda/sumrows.cuh @@ -1,5 +1,4 @@ #include "common.cuh" void sum_rows_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, cudaStream_t stream); - void ggml_cuda_op_sum_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst); diff --git a/tests/test-backend-ops.cpp b/tests/test-backend-ops.cpp index 772bee346..7be7f2205 100644 --- a/tests/test-backend-ops.cpp +++ b/tests/test-backend-ops.cpp @@ -4652,6 +4652,8 @@ static std::vector> make_test_cases_perf() { test_cases.emplace_back(new test_conv_transpose_2d({256, 256, 256, 1}, {3, 3, 16, 256}, 1)); + test_cases.emplace_back(new test_mean(GGML_TYPE_F32, {256, 256, 3, 1})); + return test_cases; } From 40bfa04c95c19fb42bafd4e21b5c2a7771846801 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Sigbj=C3=B8rn=20Skj=C3=A6ret?= Date: Sun, 22 Jun 2025 07:37:43 +0200 Subject: [PATCH 32/43] common : use std::string_view now that we target c++17 (#14319) --- common/json-schema-to-grammar.cpp | 49 ++----------------------------- 1 file changed, 3 insertions(+), 46 deletions(-) diff --git a/common/json-schema-to-grammar.cpp b/common/json-schema-to-grammar.cpp index d38a74f95..637891f50 100644 --- a/common/json-schema-to-grammar.cpp +++ b/common/json-schema-to-grammar.cpp @@ -41,49 +41,6 @@ static std::string build_repetition(const std::string & item_rule, int min_items return result; } -/* Minimalistic replacement for std::string_view, which is only available from C++17 onwards */ -class string_view { - const std::string & _str; - const size_t _start; - const size_t _end; -public: - string_view(const std::string & str, size_t start = 0, size_t end = std::string::npos) : _str(str), _start(start), _end(end == std::string::npos ? str.length() : end) {} - - size_t size() const { - return _end - _start; - } - - size_t length() const { - return size(); - } - - operator std::string() const { - return str(); - } - - std::string str() const { - return _str.substr(_start, _end - _start); - } - - string_view substr(size_t pos, size_t len = std::string::npos) const { - return string_view(_str, _start + pos, len == std::string::npos ? _end : _start + pos + len); - } - - char operator[](size_t pos) const { - auto index = _start + pos; - if (index >= _end) { - throw std::out_of_range("string_view index out of range"); - } - return _str[_start + pos]; - } - - bool operator==(const string_view & other) const { - std::string this_str = *this; - std::string other_str = other; - return this_str == other_str; - } -}; - static void _build_min_max_int(int min_value, int max_value, std::stringstream & out, int decimals_left = 16, bool top_level = true) { auto has_min = min_value != std::numeric_limits::min(); auto has_max = max_value != std::numeric_limits::max(); @@ -112,14 +69,14 @@ static void _build_min_max_int(int min_value, int max_value, std::stringstream & } out << "}"; }; - std::function uniform_range = - [&](const string_view & from, const string_view & to) { + std::function uniform_range = + [&](const std::string_view & from, const std::string_view & to) { size_t i = 0; while (i < from.length() && i < to.length() && from[i] == to[i]) { i++; } if (i > 0) { - out << "\"" << from.substr(0, i).str() << "\""; + out << "\"" << from.substr(0, i) << "\""; } if (i < from.length() && i < to.length()) { if (i > 0) { From 5d5c066de8a3d2cb32f04c4d5ad1560945f30bf3 Mon Sep 17 00:00:00 2001 From: yuiseki Date: Sun, 22 Jun 2025 21:44:57 +0900 Subject: [PATCH 33/43] mtmd : fix Pixtral OOM with large images by capping image_size to 1024 (#14326) Mistral Small 2506 models using Pixtral vision encoder were running out of GPU memory when processing images larger than 1024x1024 pixels due to exponential memory growth from unlimited image size. This fix applies the same 1024x1024 limit used by Qwen2VL models to prevent OOM issues while maintaining compatibility with existing models. --- tools/mtmd/clip.cpp | 3 +++ 1 file changed, 3 insertions(+) diff --git a/tools/mtmd/clip.cpp b/tools/mtmd/clip.cpp index 30283d6f1..a990520ed 100644 --- a/tools/mtmd/clip.cpp +++ b/tools/mtmd/clip.cpp @@ -2211,6 +2211,9 @@ struct clip_model_loader { { hparams.rope_theta = 10000.0f; hparams.warmup_image_size = hparams.patch_size * 8; + // Mistral Small 2506 needs 1024x1024 image size cap to prevent OOM + // ref: https://github.com/ggml-org/llama.cpp/issues/14310 + hparams.image_size = 1024; get_u32(KEY_SPATIAL_MERGE_SIZE, hparams.spatial_merge_size, false); } break; case PROJECTOR_TYPE_GEMMA3: From af3373f1adfca56119f3e4de0e6a0a8df8edf3d9 Mon Sep 17 00:00:00 2001 From: uvos Date: Sun, 22 Jun 2025 16:51:23 +0200 Subject: [PATCH 34/43] HIP: enable vec fattn on RDNA4 (#14323) --- ggml/src/ggml-cuda/common.cuh | 14 ++++++++++++-- ggml/src/ggml-cuda/ggml-cuda.cu | 7 ++----- 2 files changed, 14 insertions(+), 7 deletions(-) diff --git a/ggml/src/ggml-cuda/common.cuh b/ggml/src/ggml-cuda/common.cuh index 2f2fce067..86c4d29a5 100644 --- a/ggml/src/ggml-cuda/common.cuh +++ b/ggml/src/ggml-cuda/common.cuh @@ -241,8 +241,18 @@ static bool fp16_mma_available(const int cc) { #if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && !defined(GGML_HIP_ROCWMMA_FATTN) return false; #else - return (GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA) || - GGML_CUDA_CC_IS_CDNA(cc) || GGML_CUDA_CC_IS_RDNA3(cc) || GGML_CUDA_CC_IS_RDNA4(cc); + if ((GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA) || + GGML_CUDA_CC_IS_CDNA(cc) || GGML_CUDA_CC_IS_RDNA3(cc)) { + return true; + } else if (GGML_CUDA_CC_IS_RDNA4(cc)) { +#if defined(GGML_HIP_ROCWMMA_FATTN) && defined(GGML_HIP_ROCWMMA_FATTN_GFX12) + return true; +#else + return false; +#endif // defined(GGML_HIP_ROCWMMA_FATTN) && defined(GGML_HIP_ROCWMMA_FATTN_GFX12) + } else { + return false; + } #endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && !defined(GGML_HIP_ROCWMMA_FATTN) } diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu index c6bdd4fb3..462db71e1 100644 --- a/ggml/src/ggml-cuda/ggml-cuda.cu +++ b/ggml/src/ggml-cuda/ggml-cuda.cu @@ -100,8 +100,7 @@ int ggml_cuda_get_device() { static cudaError_t ggml_cuda_device_malloc(void ** ptr, size_t size, int device) { ggml_cuda_set_device(device); cudaError_t err; - if (getenv("GGML_CUDA_ENABLE_UNIFIED_MEMORY") != nullptr) - { + if (getenv("GGML_CUDA_ENABLE_UNIFIED_MEMORY") != nullptr) { err = cudaMallocManaged(ptr, size); #if defined(GGML_USE_HIP) if (err == hipSuccess) { @@ -119,9 +118,7 @@ static cudaError_t ggml_cuda_device_malloc(void ** ptr, size_t size, int device) err = cudaMalloc(ptr, size); } #endif // defined(GGML_USE_HIP) - } - else - { + } else { err = cudaMalloc(ptr, size); } return err; From f1f5e82df6222dcbaca6396c0de44df259a1694f Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Sun, 22 Jun 2025 20:10:07 +0300 Subject: [PATCH 35/43] examples : fix is_first logic for tokenization (#14329) ggml-ci --- examples/simple-chat/simple-chat.cpp | 2 +- tools/run/run.cpp | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/examples/simple-chat/simple-chat.cpp b/examples/simple-chat/simple-chat.cpp index 2aee0a919..cf1178043 100644 --- a/examples/simple-chat/simple-chat.cpp +++ b/examples/simple-chat/simple-chat.cpp @@ -98,7 +98,7 @@ int main(int argc, char ** argv) { auto generate = [&](const std::string & prompt) { std::string response; - const bool is_first = llama_memory_seq_pos_max(llama_get_memory(ctx), 0) == 0; + const bool is_first = llama_memory_seq_pos_max(llama_get_memory(ctx), 0) == -1; // tokenize the prompt const int n_prompt_tokens = -llama_tokenize(vocab, prompt.c_str(), prompt.size(), NULL, 0, is_first, true); diff --git a/tools/run/run.cpp b/tools/run/run.cpp index c65afd61e..b85565155 100644 --- a/tools/run/run.cpp +++ b/tools/run/run.cpp @@ -939,7 +939,7 @@ static int apply_chat_template(const struct common_chat_templates * tmpls, Llama // Function to tokenize the prompt static int tokenize_prompt(const llama_vocab * vocab, const std::string & prompt, std::vector & prompt_tokens, const LlamaData & llama_data) { - const bool is_first = llama_memory_seq_pos_max(llama_get_memory(llama_data.context.get()), 0) == 0; + const bool is_first = llama_memory_seq_pos_max(llama_get_memory(llama_data.context.get()), 0) == -1; const int n_prompt_tokens = -llama_tokenize(vocab, prompt.c_str(), prompt.size(), NULL, 0, is_first, true); prompt_tokens.resize(n_prompt_tokens); From 66aba7aca9a245d71f1ddf02c7a97223529752a8 Mon Sep 17 00:00:00 2001 From: Ruikai Peng Date: Mon, 23 Jun 2025 01:28:06 +0800 Subject: [PATCH 36/43] run : avoid double tokenization (#14327) * run : avoid double tokenization by adopting common_tokenize heuristic * build : fix windows gcc and clang warnings * lint : fixed trailing whitepace * run : fix is_first flag --- tools/run/run.cpp | 32 ++++++++++++++++++++++++-------- 1 file changed, 24 insertions(+), 8 deletions(-) diff --git a/tools/run/run.cpp b/tools/run/run.cpp index b85565155..6fe728c68 100644 --- a/tools/run/run.cpp +++ b/tools/run/run.cpp @@ -9,6 +9,9 @@ #include #if defined(_WIN32) +# ifndef NOMINMAX +# define NOMINMAX +# endif # include # include #else @@ -940,16 +943,29 @@ static int apply_chat_template(const struct common_chat_templates * tmpls, Llama static int tokenize_prompt(const llama_vocab * vocab, const std::string & prompt, std::vector & prompt_tokens, const LlamaData & llama_data) { const bool is_first = llama_memory_seq_pos_max(llama_get_memory(llama_data.context.get()), 0) == -1; - - const int n_prompt_tokens = -llama_tokenize(vocab, prompt.c_str(), prompt.size(), NULL, 0, is_first, true); - prompt_tokens.resize(n_prompt_tokens); - if (llama_tokenize(vocab, prompt.c_str(), prompt.size(), prompt_tokens.data(), prompt_tokens.size(), is_first, - true) < 0) { - printe("failed to tokenize the prompt\n"); + int n_tokens = prompt.size() + 2 * is_first; + prompt_tokens.resize(n_tokens); + n_tokens = llama_tokenize(vocab, prompt.c_str(), prompt.size(), + prompt_tokens.data(), prompt_tokens.size(), + is_first, /*parse_special =*/true); + if (n_tokens == std::numeric_limits::min()) { + printe("tokenization failed: input too large\n"); return -1; } - - return n_prompt_tokens; + if (n_tokens < 0) { + prompt_tokens.resize(-n_tokens); + int check = llama_tokenize(vocab, prompt.c_str(), prompt.size(), + prompt_tokens.data(), prompt_tokens.size(), + is_first, /*parse_special =*/true); + if (check != -n_tokens) { + printe("failed to tokenize the prompt (size mismatch)\n"); + return -1; + } + n_tokens = check; + } else { + prompt_tokens.resize(n_tokens); + } + return n_tokens; } // Check if we have enough space in the context to evaluate this batch From 238005c2dc67426cf678baa2d54c881701693288 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Sigbj=C3=B8rn=20Skj=C3=A6ret?= Date: Sun, 22 Jun 2025 19:46:17 +0200 Subject: [PATCH 37/43] gguf-py : fix SpecialVocab parsing when post_processor is null (#14330) --- gguf-py/gguf/vocab.py | 144 +++++++++++++++++++++--------------------- 1 file changed, 72 insertions(+), 72 deletions(-) diff --git a/gguf-py/gguf/vocab.py b/gguf-py/gguf/vocab.py index 3b08f6134..3f541b0c0 100644 --- a/gguf-py/gguf/vocab.py +++ b/gguf-py/gguf/vocab.py @@ -167,81 +167,81 @@ class SpecialVocab: tokenizer_config['bos_token'] = special_bos = special_cls if not special_eos and special_sep and tokenizer_config: tokenizer_config['eos_token'] = special_eos = special_sep - post_processor = tokenizer.get('post_processor', {}) - for processor in post_processor.get('processors', [post_processor]): - if processor.get('type') == 'RobertaProcessing': - self.add_special_token['bos'] = True - self.add_special_token['eos'] = True - self.add_special_token['sep'] = True - if not special_cls and tokenizer_config: - special_cls = processor.get('cls', [special_bos])[0] - tokenizer_config['cls_token'] = special_cls - if not special_sep and tokenizer_config: - special_sep = processor.get('sep', [special_eos])[0] - tokenizer_config['sep_token'] = special_sep - continue - # Crude parsing of TemplateProcessing to determine if BOS/SEP/EOS should be added - # Only works with simple templates, **will** get it wrong on unusual sequences - if processor.get('type') == 'TemplateProcessing': - tmpl_single = processor.get('single', []) - tmpl_pair = processor.get('pair', []) - special_first = None - special_last = None - if len(tmpl_single) > 1: - if special_first := tmpl_single[0].get('SpecialToken', {}).get('id'): - if not tokenizer_config: - special_bos = special_first - self.add_special_token['bos'] = True if special_first in (special_bos, special_cls) else False - if special_first not in (special_bos, special_cls): - logger.warning(f'Unknown leading special token {special_first!r} in TemplateProcessing') - if special_last := tmpl_single[-1].get('SpecialToken', {}).get('id'): - if not tokenizer_config: - special_eos = special_last - elif special_last != special_eos: - if 'eot' not in self.special_token_types: - self.special_token_types = tuple(self.special_token_types) + ('eot', ) - tokenizer_config['eot_token'] = special_eos - elif 'eom' not in self.special_token_types: - self.special_token_types = tuple(self.special_token_types) + ('eom', ) - tokenizer_config['eom_token'] = special_eos - else: - logger.warning(f'Overriding EOS token {special_eos!r} with {special_last!r} without EOT/EOM fallback!') - tokenizer_config['eos_token'] = special_eos = special_last - self.add_special_token['eos'] = True if special_last == special_eos else False - if special_last != special_eos: - logger.warning(f'Unknown trailing special token {special_last!r} in TemplateProcessing') - if tmpl_pair: - seq_start = 1 if special_first and tmpl_pair[0].get('SpecialToken', {}).get('id') == special_first else 0 - seq_stop = -1 if special_last and tmpl_pair[-1].get('SpecialToken', {}).get('id') == special_last else None - if (special_first and seq_start == 0) or (special_last and seq_stop is None): - logger.warning('TemplateProcessing leading/trailing special tokens do not match TemplateProcessing') - if tmpl_pair := tmpl_pair[slice(seq_start, seq_stop)]: - tmpl_a = tmpl_pair[0].get('Sequence', {}).get('id') - tmpl_b = tmpl_pair[-1].get('Sequence', {}).get('id') - if tmpl_a != 'A' or tmpl_b != 'B': - logger.warning(f'Unknown sequence {tmpl_a}...{tmpl_b} in TemplateProcessing') - # A [sep] [eos] B - if tmpl_a == 'A' and tmpl_b == 'B' and (tmpl_pair := tmpl_pair[1:-1]): - add_sep = False - if special_entry := tmpl_pair[0].get('SpecialToken', {}).get('id'): - if special_entry in (special_sep, special_eos) and not special_last: - add_sep = True - if special_entry not in (special_sep, special_eos): - logger.warning(f'Unknown separator token {special_entry!r} in TemplateProcessing') - else: - logger.warning(f'Unknown middle sequence {tmpl_pair[0]!r} in TemplateProcessing') - if len(tmpl_pair) == 2: - if special_entry := tmpl_pair[1].get('SpecialToken', {}).get('id'): - if special_entry in (special_sep, special_eos): + if post_processor := tokenizer.get('post_processor'): + for processor in post_processor.get('processors', [post_processor]): + if processor.get('type') == 'RobertaProcessing': + self.add_special_token['bos'] = True + self.add_special_token['eos'] = True + self.add_special_token['sep'] = True + if not special_cls and tokenizer_config: + special_cls = processor.get('cls', [special_bos])[0] + tokenizer_config['cls_token'] = special_cls + if not special_sep and tokenizer_config: + special_sep = processor.get('sep', [special_eos])[0] + tokenizer_config['sep_token'] = special_sep + continue + # Crude parsing of TemplateProcessing to determine if BOS/SEP/EOS should be added + # Only works with simple templates, **will** get it wrong on unusual sequences + if processor.get('type') == 'TemplateProcessing': + tmpl_single = processor.get('single', []) + tmpl_pair = processor.get('pair', []) + special_first = None + special_last = None + if len(tmpl_single) > 1: + if special_first := tmpl_single[0].get('SpecialToken', {}).get('id'): + if not tokenizer_config: + special_bos = special_first + self.add_special_token['bos'] = True if special_first in (special_bos, special_cls) else False + if special_first not in (special_bos, special_cls): + logger.warning(f'Unknown leading special token {special_first!r} in TemplateProcessing') + if special_last := tmpl_single[-1].get('SpecialToken', {}).get('id'): + if not tokenizer_config: + special_eos = special_last + elif special_last != special_eos: + if 'eot' not in self.special_token_types: + self.special_token_types = tuple(self.special_token_types) + ('eot', ) + tokenizer_config['eot_token'] = special_eos + elif 'eom' not in self.special_token_types: + self.special_token_types = tuple(self.special_token_types) + ('eom', ) + tokenizer_config['eom_token'] = special_eos + else: + logger.warning(f'Overriding EOS token {special_eos!r} with {special_last!r} without EOT/EOM fallback!') + tokenizer_config['eos_token'] = special_eos = special_last + self.add_special_token['eos'] = True if special_last == special_eos else False + if special_last != special_eos: + logger.warning(f'Unknown trailing special token {special_last!r} in TemplateProcessing') + if tmpl_pair: + seq_start = 1 if special_first and tmpl_pair[0].get('SpecialToken', {}).get('id') == special_first else 0 + seq_stop = -1 if special_last and tmpl_pair[-1].get('SpecialToken', {}).get('id') == special_last else None + if (special_first and seq_start == 0) or (special_last and seq_stop is None): + logger.warning('TemplateProcessing leading/trailing special tokens do not match TemplateProcessing') + if tmpl_pair := tmpl_pair[slice(seq_start, seq_stop)]: + tmpl_a = tmpl_pair[0].get('Sequence', {}).get('id') + tmpl_b = tmpl_pair[-1].get('Sequence', {}).get('id') + if tmpl_a != 'A' or tmpl_b != 'B': + logger.warning(f'Unknown sequence {tmpl_a}...{tmpl_b} in TemplateProcessing') + # A [sep] [eos] B + if tmpl_a == 'A' and tmpl_b == 'B' and (tmpl_pair := tmpl_pair[1:-1]): + add_sep = False + if special_entry := tmpl_pair[0].get('SpecialToken', {}).get('id'): + if special_entry in (special_sep, special_eos) and not special_last: add_sep = True if special_entry not in (special_sep, special_eos): - logger.warning(f'Unknown second separator token {special_entry!r} in TemplateProcessing') + logger.warning(f'Unknown separator token {special_entry!r} in TemplateProcessing') else: - logger.warning(f'Unknown second middle sequence {tmpl_pair[1]!r} in TemplateProcessing') - self.add_special_token['sep'] = add_sep - if add_sep and not special_sep and tokenizer_config: - tokenizer_config['sep_token'] = special_eos - continue + logger.warning(f'Unknown middle sequence {tmpl_pair[0]!r} in TemplateProcessing') + if len(tmpl_pair) == 2: + if special_entry := tmpl_pair[1].get('SpecialToken', {}).get('id'): + if special_entry in (special_sep, special_eos): + add_sep = True + if special_entry not in (special_sep, special_eos): + logger.warning(f'Unknown second separator token {special_entry!r} in TemplateProcessing') + else: + logger.warning(f'Unknown second middle sequence {tmpl_pair[1]!r} in TemplateProcessing') + self.add_special_token['sep'] = add_sep + if add_sep and not special_sep and tokenizer_config: + tokenizer_config['sep_token'] = special_eos + continue if not tokenizer_config: return True chat_template_alt = None From fa4a9f2a1ccda2573189a9d4995bdf0bceb41156 Mon Sep 17 00:00:00 2001 From: Ed Addario <29247825+EAddario@users.noreply.github.com> Date: Sun, 22 Jun 2025 22:16:26 +0100 Subject: [PATCH 38/43] quantize : handle user-defined pruning of whole layers (blocks) (#13037) --- include/llama.h | 1 + src/llama-quant.cpp | 83 +++++++++++++++++++++++++++++++++++-- tools/quantize/quantize.cpp | 44 +++++++++++++++++--- 3 files changed, 119 insertions(+), 9 deletions(-) diff --git a/include/llama.h b/include/llama.h index b04720bee..f4123d14a 100644 --- a/include/llama.h +++ b/include/llama.h @@ -390,6 +390,7 @@ extern "C" { void * imatrix; // pointer to importance matrix data void * kv_overrides; // pointer to vector containing overrides void * tensor_types; // pointer to vector containing tensor types + void * prune_layers; // pointer to vector containing layer indices to prune } llama_model_quantize_params; typedef struct llama_logit_bias { diff --git a/src/llama-quant.cpp b/src/llama-quant.cpp index 8cf45732f..43229e193 100644 --- a/src/llama-quant.cpp +++ b/src/llama-quant.cpp @@ -1,5 +1,4 @@ #include "llama-quant.h" - #include "llama-impl.h" #include "llama-model.h" #include "llama-model-loader.h" @@ -27,6 +26,56 @@ static void zeros(std::ofstream & file, size_t n) { } } +static std::string remap_layer(const std::string & orig_name, const std::vector & prune, std::map & mapped, int & next_id) { + if (prune.empty()) { + return orig_name; + } + + static const std::regex pattern(R"(blk\.(\d+)\.)"); + if (std::smatch match; std::regex_search(orig_name, match, pattern)) { + const int blk = std::stoi(match[1]); + std::string new_name = orig_name; + + if (mapped.count(blk)) { + // Already mapped, do nothing + } else if (std::find(prune.begin(), prune.end(), blk) != prune.end()) { + mapped[blk] = ""; + } else if (blk < prune.front()) { + mapped[blk] = std::to_string(blk); + next_id = blk + 1; + } else { + mapped[blk] = std::to_string(next_id); + ++next_id; + } + + return mapped[blk].empty() ? mapped[blk] : new_name.replace(match.position(1), match.length(1), mapped[blk]); + } + + return orig_name; +} + +static std::string remap_imatrix (const std::string & orig_name, const std::map & mapped) { + if (mapped.empty()) { + return orig_name; + } + + static const std::regex pattern(R"(blk\.(\d+)\.)"); + if (std::smatch match; std::regex_search(orig_name, match, pattern)) { + const std::string blk(match[1]); + std::string new_name = orig_name; + + for (const auto & p : mapped) { + if (p.second == blk) { + LLAMA_LOG_DEBUG("(blk.%d imatrix) ", p.first); + return new_name.replace(match.position(1), match.length(1), std::to_string(p.first)); + } + } + GGML_ABORT("\n%s: imatrix mapping error for %s\n", __func__, orig_name.c_str()); + } + + return orig_name; +} + struct quantize_state_impl { const llama_model & model; const llama_model_quantize_params * params; @@ -568,6 +617,11 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std:: const size_t align = GGUF_DEFAULT_ALIGNMENT; gguf_context_ptr ctx_out { gguf_init_empty() }; + std::vector prune_list = {}; + if (params->prune_layers) { + prune_list = *static_cast *>(params->prune_layers); + } + // copy the KV pairs from the input file gguf_set_kv (ctx_out.get(), ml.meta.get()); gguf_set_val_u32(ctx_out.get(), "general.quantization_version", GGML_QNT_VERSION); // TODO: use LLM_KV @@ -597,12 +651,32 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std:: } } + std::map mapped; + int blk_id = 0; + int pruned_attention_w = 0; + // make a list of weights std::vector tensors; tensors.reserve(ml.weights_map.size()); for (const auto & it : ml.weights_map) { + const std::string remapped_name(remap_layer(it.first, prune_list, mapped, blk_id)); + if (remapped_name.empty()) { + if (it.first.find("attn_v.weight") != std::string::npos || + it.first.find("attn_qkv.weight") != std::string::npos || + it.first.find("attn_kv_b.weight") != std::string::npos) { + pruned_attention_w++; + } + LLAMA_LOG_DEBUG("%s: pruning tensor %s\n", __func__, it.first.c_str()); + continue; + } else if (remapped_name != it.first) { + ggml_set_name(it.second.tensor, remapped_name.c_str()); + LLAMA_LOG_DEBUG("%s: tensor %s remapped to %s\n", __func__, it.first.c_str(), ggml_get_name(it.second.tensor)); + } tensors.push_back(&it.second); } + if (!prune_list.empty()) { + gguf_set_val_u32(ctx_out.get(), ml.llm_kv(LLM_KV_BLOCK_COUNT).c_str(), blk_id); + } // keep_split requires that the weights are sorted by split index if (params->keep_split) { @@ -640,7 +714,7 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std:: if (llama_model_has_encoder(&model)) { n_attn_layer *= 3; } - GGML_ASSERT((qs.n_attention_wv == n_attn_layer) && "n_attention_wv is unexpected"); + GGML_ASSERT((qs.n_attention_wv == n_attn_layer - pruned_attention_w) && "n_attention_wv is unexpected"); } size_t total_size_org = 0; @@ -681,7 +755,7 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std:: for (size_t i = 0; i < ctx_outs.size(); ++i) { gguf_set_val_u16(ctx_outs[i].get(), ml.llm_kv(LLM_KV_SPLIT_NO).c_str(), i); gguf_set_val_u16(ctx_outs[i].get(), ml.llm_kv(LLM_KV_SPLIT_COUNT).c_str(), n_split); - gguf_set_val_i32(ctx_outs[i].get(), ml.llm_kv(LLM_KV_SPLIT_TENSORS_COUNT).c_str(), ml.n_tensors); + gguf_set_val_i32(ctx_outs[i].get(), ml.llm_kv(LLM_KV_SPLIT_TENSORS_COUNT).c_str(), (int32_t)tensors.size()); } } @@ -832,7 +906,7 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std:: const float * imatrix = nullptr; if (imatrix_data) { - auto it = imatrix_data->find(tensor->name); + auto it = imatrix_data->find(remap_imatrix(tensor->name, mapped)); if (it == imatrix_data->end()) { LLAMA_LOG_INFO("\n====== %s: did not find weights for %s\n", __func__, tensor->name); } else { @@ -947,6 +1021,7 @@ llama_model_quantize_params llama_model_quantize_default_params() { /*.imatrix =*/ nullptr, /*.kv_overrides =*/ nullptr, /*.tensor_type =*/ nullptr, + /*.prune_layers =*/ nullptr }; return result; diff --git a/tools/quantize/quantize.cpp b/tools/quantize/quantize.cpp index 3f54af7c5..8acc76517 100644 --- a/tools/quantize/quantize.cpp +++ b/tools/quantize/quantize.cpp @@ -107,13 +107,11 @@ static bool try_parse_ftype(const std::string & ftype_str_in, llama_ftype & ftyp return false; } -// usage: -// ./llama-quantize [--allow-requantize] [--leave-output-tensor] [--pure] models/llama/ggml-model.gguf [models/llama/ggml-model-quant.gguf] type [nthreads] -// [[noreturn]] static void usage(const char * executable) { - printf("usage: %s [--help] [--allow-requantize] [--leave-output-tensor] [--pure] [--imatrix] [--include-weights] [--exclude-weights] [--output-tensor-type]\n", executable); - printf(" [--token-embedding-type] [--tensor-type] [--keep-split] [--override-kv] model-f32.gguf [model-quant.gguf] type [nthreads]\n\n"); + printf("usage: %s [--help] [--allow-requantize] [--leave-output-tensor] [--pure] [--imatrix] [--include-weights]\n", executable); + printf(" [--exclude-weights] [--output-tensor-type] [--token-embedding-type] [--tensor-type] [--prune-layers] [--keep-split] [--override-kv]\n"); + printf(" model-f32.gguf [model-quant.gguf] type [nthreads]\n\n"); printf(" --allow-requantize: Allows requantizing tensors that have already been quantized. Warning: This can severely reduce quality compared to quantizing from 16bit or 32bit\n"); printf(" --leave-output-tensor: Will leave output.weight un(re)quantized. Increases model size but may also increase quality, especially when requantizing\n"); printf(" --pure: Disable k-quant mixtures and quantize all tensors to the same type\n"); @@ -124,6 +122,8 @@ static void usage(const char * executable) { printf(" --token-embedding-type ggml_type: use this ggml_type for the token embeddings tensor\n"); printf(" --tensor-type TENSOR=TYPE: quantize this tensor to this ggml_type. example: --tensor-type attn_q=q8_0\n"); printf(" Advanced option to selectively quantize tensors. May be specified multiple times.\n"); + printf(" --prune-layers L0,L1,L2...comma-separated list of layer numbers to prune from the model\n"); + printf(" Advanced option to remove all tensors from the given layers\n"); printf(" --keep-split: will generate quantized model in the same shards as input\n"); printf(" --override-kv KEY=TYPE:VALUE\n"); printf(" Advanced option to override model metadata by key in the quantized model. May be specified multiple times.\n"); @@ -286,6 +286,32 @@ static bool parse_tensor_type(const char * data, std::vector & prune_layers) { + if (!data) { + printf("\n%s: no layer pruning ids provided\n\n", __func__); + return false; + } + + const auto block_ids = string_split(data, ','); + for (const auto & block_id : block_ids) { + int id; + try { + id = std::stoi(block_id); + } catch (...) { + id = -1; + } + if (id < 0) { + printf("\n%s: invalid layer id '%s'\n\n", __func__, block_id.c_str()); + return false; + } + prune_layers.emplace_back(id); + } + + sort(prune_layers.begin(), prune_layers.end()); + prune_layers.erase(std::unique(prune_layers.begin(), prune_layers.end()), prune_layers.end()); + return true; +} + int main(int argc, char ** argv) { if (argc < 3) { usage(argv[0]); @@ -298,6 +324,7 @@ int main(int argc, char ** argv) { std::vector included_weights, excluded_weights; std::vector kv_overrides; std::vector tensor_types; + std::vector prune_layers; for (; arg_idx < argc && strncmp(argv[arg_idx], "--", 2) == 0; arg_idx++) { if (strcmp(argv[arg_idx], "--leave-output-tensor") == 0) { @@ -324,6 +351,10 @@ int main(int argc, char ** argv) { if (arg_idx == argc-1 || !parse_tensor_type(argv[++arg_idx], tensor_types)) { usage(argv[0]); } + } else if (strcmp(argv[arg_idx], "--prune-layers") == 0) { + if (arg_idx == argc-1 || !parse_layer_prune(argv[++arg_idx], prune_layers)) { + usage(argv[0]); + } } else if (strcmp(argv[arg_idx], "--override-kv") == 0) { if (arg_idx == argc-1 || !string_parse_kv_override(argv[++arg_idx], kv_overrides)) { usage(argv[0]); @@ -411,6 +442,9 @@ int main(int argc, char ** argv) { if (!tensor_types.empty()) { params.tensor_types = &tensor_types; } + if (!prune_layers.empty()) { + params.prune_layers = &prune_layers; + } llama_backend_init(); From 3a9457df962b5883f2773f1c295e8c19df60d89f Mon Sep 17 00:00:00 2001 From: Jeff Bolz Date: Mon, 23 Jun 2025 03:19:24 -0500 Subject: [PATCH 39/43] vulkan: update windows SDK in CI (#14334) --- .github/workflows/build.yml | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/.github/workflows/build.yml b/.github/workflows/build.yml index c4783a6df..be2828973 100644 --- a/.github/workflows/build.yml +++ b/.github/workflows/build.yml @@ -683,7 +683,7 @@ jobs: env: OPENBLAS_VERSION: 0.3.23 SDE_VERSION: 9.33.0-2024-01-07 - VULKAN_VERSION: 1.4.309.0 + VULKAN_VERSION: 1.4.313.2 strategy: matrix: @@ -736,7 +736,7 @@ jobs: id: get_vulkan if: ${{ matrix.build == 'kompute-x64' || matrix.build == 'vulkan-x64' }} run: | - curl.exe -o $env:RUNNER_TEMP/VulkanSDK-Installer.exe -L "https://sdk.lunarg.com/sdk/download/${env:VULKAN_VERSION}/windows/VulkanSDK-${env:VULKAN_VERSION}-Installer.exe" + curl.exe -o $env:RUNNER_TEMP/VulkanSDK-Installer.exe -L "https://sdk.lunarg.com/sdk/download/${env:VULKAN_VERSION}/windows/vulkansdk-windows-X64-${env:VULKAN_VERSION}.exe" & "$env:RUNNER_TEMP\VulkanSDK-Installer.exe" --accept-licenses --default-answer --confirm-command install Add-Content $env:GITHUB_ENV "VULKAN_SDK=C:\VulkanSDK\${env:VULKAN_VERSION}" Add-Content $env:GITHUB_PATH "C:\VulkanSDK\${env:VULKAN_VERSION}\bin" From 7b50d589a863c7631135c1226f6eab65cb406212 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Mon, 23 Jun 2025 12:27:35 +0300 Subject: [PATCH 40/43] kv-cells : fix tracking of seq_pos (#14339) * kv-cells : fix tracking of seq_pos during cache reuse ggml-ci * cont : improve error message ggml-ci * cont : add more comments --- include/llama.h | 8 +++++--- src/llama-batch.cpp | 19 +++++++++++++++---- src/llama-context.cpp | 1 - src/llama-kv-cells.h | 42 ++++++++++++++++++++++++++++++++--------- tools/server/server.cpp | 3 +++ 5 files changed, 56 insertions(+), 17 deletions(-) diff --git a/include/llama.h b/include/llama.h index f4123d14a..3eda9bc68 100644 --- a/include/llama.h +++ b/include/llama.h @@ -944,12 +944,14 @@ extern "C" { // Requires the context to have a memory. // For encode-decoder contexts, processes the batch using the decoder. // Positive return values does not mean a fatal error, but rather a warning. - // Upon non-zero return values, the memory state is restored to the state before this call + // Upon fatal-error or abort, the ubatches that managed to be been processed will remain in the memory state of the context + // To handle this correctly, query the memory state using llama_memory_seq_pos_min() and llama_memory_seq_pos_max() + // Upon other return values, the memory state is restored to the state before this call // 0 - success // 1 - could not find a KV slot for the batch (try reducing the size of the batch or increase the context) - // 2 - aborted + // 2 - aborted (processed ubatches will remain in the context's memory) // -1 - invalid input batch - // < -1 - error + // < -1 - fatal error (processed ubatches will remain in the context's memory) LLAMA_API int32_t llama_decode( struct llama_context * ctx, struct llama_batch batch); diff --git a/src/llama-batch.cpp b/src/llama-batch.cpp index b3c996e18..401e11364 100644 --- a/src/llama-batch.cpp +++ b/src/llama-batch.cpp @@ -245,10 +245,11 @@ bool llama_batch_allocr::init( } if (memory) { + bool ok = true; + if (batch.token) { if (seq_pos_min(s) != memory->seq_pos_max(s) + 1) { - LLAMA_LOG_ERROR("%s: sequence %d does not start from the last position stored in the memory\n", __func__, s); - return false; + ok = false; } } else { assert(batch.embd); @@ -256,10 +257,20 @@ bool llama_batch_allocr::init( // for embeddings (typically used as vision input), we allow them to have repeating positions // ref: https://github.com/ggml-org/llama.cpp/issues/13694#issuecomment-2983871762 if (seq_pos_min(s) != memory->seq_pos_max(s) && seq_pos_min(s) != memory->seq_pos_max(s) + 1) { - LLAMA_LOG_ERROR("%s: sequence %d does not start from the last position stored in the memory\n", __func__, s); - return false; + ok = false; } } + + if (!ok) { + LLAMA_LOG_ERROR( + "%s: the tokens of sequence %d in the input batch have inconsistent sequence positions:\n" + " - the last position stored in the memory module of the context (i.e. the KV cache) for sequence %d is X = %d\n" + " - the tokens for sequence %d in the input batch have a starting position of Y = %d\n" + " it is required that the sequence positions remain consecutive: Y = X + 1\n", + __func__, s, s, memory->seq_pos_max(s), s, seq_pos_min(s)); + + return false; + } } if (seq_pos_max(s) - seq_pos_min(s) + 1 > (int) seq_pos[s].size()) { diff --git a/src/llama-context.cpp b/src/llama-context.cpp index e352d81e4..06e93b19c 100644 --- a/src/llama-context.cpp +++ b/src/llama-context.cpp @@ -1018,7 +1018,6 @@ int llama_context::decode(const llama_batch & batch_inp) { pos_min[s] = std::numeric_limits::max(); } - // TODO: fix sequence indexing for (uint32_t i = 0; i < ubatch.n_tokens; ++i) { const auto & seq_id = ubatch.seq_id[i][0]; diff --git a/src/llama-kv-cells.h b/src/llama-kv-cells.h index 349e9032e..c95d63594 100644 --- a/src/llama-kv-cells.h +++ b/src/llama-kv-cells.h @@ -7,6 +7,7 @@ #include #include #include +#include // meta information about KV cells that can be part of multiple sequences at the same time // TODO: add unit tests @@ -164,7 +165,7 @@ public: assert(seq_id >= 0); seq[i].reset(seq_id); - seq_pos[seq_id].erase(pos[i]); + seq_pos_dec(seq_id, pos[i]); if (seq[i].none()) { pos[i] = -1; @@ -187,7 +188,7 @@ public: seq[i].reset(); seq[i].set(seq_id); - seq_pos[seq_id].insert(pos[i]); + seq_pos_inc(seq_id, pos[i]); return false; } @@ -232,7 +233,7 @@ public: assert(!seq[i].test(seq_id)); seq[i].set(seq_id); - seq_pos[seq_id].insert(pos[i]); + seq_pos_inc(seq_id, pos[i]); } // return the sequence id of this cell @@ -259,7 +260,9 @@ public: return -1; } - return *seq_pos[seq_id].begin(); + assert(seq_pos[seq_id].begin()->second > 0); + + return seq_pos[seq_id].begin()->first; } // the maximum position of sequence seq_id currently present in any of the cells @@ -272,7 +275,9 @@ public: return -1; } - return *seq_pos[seq_id].rbegin(); + assert(seq_pos[seq_id].rbegin()->second > 0); + + return seq_pos[seq_id].rbegin()->first; } // note: call only if the cell is not empty @@ -389,17 +394,36 @@ private: // the bitset seq[i] tells us which sequences are currently occupying the i-th cell std::vector seq; - // the set seq_pos[s] tells us which positions are currently present for sequence s + // the set seq_pos[s][p] tells us how many times the position p is currently present for sequence s + // if the position p is not present, seq_pos[s][p] is not set // this way seq_pos[s].begin() and seq_pos[s].rbegin() give us the min/max positions currently in the cache - std::set seq_pos[LLAMA_MAX_SEQ]; + // + // note that we cannot a use an std::set because in some cases a position can occur more than once for the same seq: + // - during performing a cache reuse via (rm + add) + // - some vision models have input embeddings with repeating positions + // + std::map seq_pos[LLAMA_MAX_SEQ]; // helper functions for updating `seq_pos`, once cell at a time: + void seq_pos_dec(llama_seq_id s, llama_pos p) { + auto it = seq_pos[s].find(p); + assert(it != seq_pos[s].end()); + + if (--it->second == 0) { + seq_pos[s].erase(it); + } + } + + void seq_pos_inc(llama_seq_id s, llama_pos p) { + seq_pos[s][p]++; + } + // remove cell i void seq_pos_rm(uint32_t i) { for (int s = 0; s < LLAMA_MAX_SEQ; ++s) { if (seq[i].test(s)) { - seq_pos[s].erase(pos[i]); + seq_pos_dec(s, pos[i]); } } } @@ -408,7 +432,7 @@ private: void seq_pos_add(uint32_t i) { for (int s = 0; s < LLAMA_MAX_SEQ; ++s) { if (seq[i].test(s)) { - seq_pos[s].insert(pos[i]); + seq_pos_inc(s, pos[i]); } } } diff --git a/tools/server/server.cpp b/tools/server/server.cpp index aa18513e3..852352383 100644 --- a/tools/server/server.cpp +++ b/tools/server/server.cpp @@ -3418,9 +3418,12 @@ struct server_context { } if (ret < -1) { + // TODO: update slot state based on llama_memory_seq_pos_min() and llama_memory_seq_pos_max() err = "Compute error."; } + // TODO: handle ret == 2 (abort) when we start aborting + if (!err.empty()) { SRV_ERR("%s, i = %d, n_batch = %d, ret = %d\n", err.c_str(), i, n_batch, ret); for (auto & slot : slots) { From defe2158dd5e250b4ef53994057a4ec03131a263 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Johannes=20G=C3=A4=C3=9Fler?= Date: Mon, 23 Jun 2025 13:11:31 +0200 Subject: [PATCH 41/43] CUDA: mul_mat_v support for batch sizes > 1 (#14262) * CUDA: mul_mat_v support for batch sizes > 1 * use 64 bit math for initial offset calculation --- ggml/src/ggml-cuda/common.cuh | 4 + ggml/src/ggml-cuda/ggml-cuda.cu | 24 ++- ggml/src/ggml-cuda/mmv.cu | 326 +++++++++++++++++++++++--------- ggml/src/ggml-cuda/mmv.cuh | 5 +- 4 files changed, 256 insertions(+), 103 deletions(-) diff --git a/ggml/src/ggml-cuda/common.cuh b/ggml/src/ggml-cuda/common.cuh index 86c4d29a5..1369bc2d9 100644 --- a/ggml/src/ggml-cuda/common.cuh +++ b/ggml/src/ggml-cuda/common.cuh @@ -262,6 +262,10 @@ static bool fp16_mma_hardware_available(const int cc) { GGML_CUDA_CC_IS_CDNA(cc) || GGML_CUDA_CC_IS_RDNA3(cc) || GGML_CUDA_CC_IS_RDNA4(cc); } +static bool bf16_mma_hardware_available(const int cc) { + return GGML_CUDA_CC_IS_NVIDIA(cc) && cc >= GGML_CUDA_CC_AMPERE; +} + // Volta technically had FP16 tensor cores but they work very differently compared to Turing and later. static bool new_mma_available(const int cc) { return GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_TURING; diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu index 462db71e1..b3e6833c3 100644 --- a/ggml/src/ggml-cuda/ggml-cuda.cu +++ b/ggml/src/ggml-cuda/ggml-cuda.cu @@ -1943,16 +1943,14 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor && ggml_nbytes(src0) != ggml_backend_buffer_get_alloc_size(src0->buffer, src0) && src0->view_src; bool use_mul_mat_vec = (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || src0->type == GGML_TYPE_BF16) - && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32 - && src0->ne[0] % 2 == 0 && src1->ne[1] == 1; + && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32; bool use_mul_mat_vec_q = ggml_is_quantized(src0->type) && !bad_padding_clear && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32 && src1->ne[1] <= MMVQ_MAX_BATCH_SIZE; bool use_mul_mat_q = ggml_is_quantized(src0->type) && !bad_padding_clear && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32; - bool any_gpus_with_slow_fp16 = false; - bool any_gpus_without_fp16_mma = false; + bool any_gpus_with_slow_fp16 = false; if (split) { ggml_backend_cuda_split_buffer_type_context * buft_ctx = (ggml_backend_cuda_split_buffer_type_context *) src0->buffer->buft->context; @@ -1963,16 +1961,16 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor continue; } - const int cc = ggml_cuda_info().devices[id].cc; - use_mul_mat_q = use_mul_mat_q && ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[1]); - any_gpus_with_slow_fp16 = any_gpus_with_slow_fp16 || !fast_fp16_hardware_available(cc); - any_gpus_without_fp16_mma = any_gpus_without_fp16_mma || !fp16_mma_hardware_available(cc); + const int cc = ggml_cuda_info().devices[id].cc; + use_mul_mat_q = use_mul_mat_q && ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[1]); + use_mul_mat_vec = use_mul_mat_vec && ggml_cuda_should_use_mmv(src0->type, cc, src0->ne, src1->ne[1]); + any_gpus_with_slow_fp16 = any_gpus_with_slow_fp16 || !fast_fp16_hardware_available(cc); } } else { - const int cc = ggml_cuda_info().devices[ctx.device].cc; - use_mul_mat_q = use_mul_mat_q && ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[1]); - any_gpus_with_slow_fp16 = any_gpus_with_slow_fp16 || !fast_fp16_hardware_available(cc); - any_gpus_without_fp16_mma = any_gpus_without_fp16_mma || !fp16_mma_hardware_available(cc); + const int cc = ggml_cuda_info().devices[ctx.device].cc; + use_mul_mat_q = use_mul_mat_q && ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[1]); + use_mul_mat_vec = use_mul_mat_vec && ggml_cuda_should_use_mmv(src0->type, cc, src0->ne, src1->ne[1]); + any_gpus_with_slow_fp16 = any_gpus_with_slow_fp16 || !fast_fp16_hardware_available(cc); } // debug helpers @@ -1983,7 +1981,7 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor //printf("src0 is contiguous %d, transposed %d, type = %s, name = %s\n", ggml_is_contiguous(src0), ggml_is_transposed(src0), ggml_type_name(src0->type), src0->name); //printf("src1 is contiguous %d, transposed %d, type = %s, name = %s\n", ggml_is_contiguous(src1), ggml_is_transposed(src1), ggml_type_name(src1->type), src1->name); - if (!split && use_mul_mat_vec && (src0->ne[1] <= MMV_MAX_ROWS || any_gpus_without_fp16_mma)) { + if (!split && use_mul_mat_vec) { // the custom F16 vector kernel can be used over batched cuBLAS GEMM // but this is only faster for GPUs without tensor cores or with a thin src0 matrix (particularly KQV in attention) ggml_cuda_mul_mat_vec(ctx, src0, src1, nullptr, dst); diff --git a/ggml/src/ggml-cuda/mmv.cu b/ggml/src/ggml-cuda/mmv.cu index d8c385e23..1502e9d94 100644 --- a/ggml/src/ggml-cuda/mmv.cu +++ b/ggml/src/ggml-cuda/mmv.cu @@ -2,25 +2,26 @@ #include "common.cuh" #include "mmv.cuh" -template +template static __global__ void mul_mat_vec( const T * __restrict__ x, const float * __restrict__ y, const int32_t * __restrict__ ids, float * __restrict__ dst, - const int64_t ncols2, const int64_t nchannels_y, const int64_t stride_row, - const int64_t channel_ratio, const int64_t stride_channel_x, const int64_t stride_channel_y, const int64_t stride_channel_dst, - const int64_t sample_ratio, const int64_t stride_sample_x, const int64_t stride_sample_y, const int64_t stride_sample_dst) { - const int64_t row = blockIdx.x; - const int64_t channel_dst = blockIdx.y; - const int64_t channel_x = ids ? ids[channel_dst] : channel_dst / channel_ratio; - const int64_t channel_y = ids ? channel_dst % nchannels_y : channel_dst; - const int64_t sample_dst = blockIdx.z; - const int64_t sample_x = sample_dst / sample_ratio; - const int64_t sample_y = sample_dst; - const int tid = threadIdx.x; + const int ncols2, const int nchannels_y, const int stride_row, const int stride_col_y2, const int stride_col_dst, + const int channel_ratio, const int stride_channel_x, const int stride_channel_y, const int stride_channel_dst, + const int sample_ratio, const int stride_sample_x, const int stride_sample_y, const int stride_sample_dst) { + const int row = blockIdx.x; + const int channel_dst = blockIdx.y; + const int channel_x = ids ? ids[channel_dst] : channel_dst / channel_ratio; + const int channel_y = ids ? channel_dst % nchannels_y : channel_dst; + const int sample_dst = blockIdx.z; + const int sample_x = sample_dst / sample_ratio; + const int sample_y = sample_dst; + const int tid = threadIdx.x; + constexpr int warp_size = ggml_cuda_get_physical_warp_size(); - x += sample_x *stride_sample_x + channel_x *stride_channel_x + row*stride_row; - y += sample_y *stride_sample_y + channel_y *stride_channel_y; - dst += sample_dst*stride_sample_dst + channel_dst*stride_channel_dst; + x += int64_t(sample_x) *stride_sample_x + channel_x *stride_channel_x + row*stride_row; + y += int64_t(sample_y) *stride_sample_y + channel_y *stride_channel_y; + dst += int64_t(sample_dst)*stride_sample_dst + channel_dst*stride_channel_dst; const float2 * y2 = (const float2 *) y; @@ -34,81 +35,108 @@ static __global__ void mul_mat_vec( __syncthreads(); } - float sumf = 0.0f; + float sumf[ncols_dst] = {0.0f}; if constexpr (std::is_same::value) { const float2 * x2 = (const float2 *) x; - for (int64_t col2 = tid; col2 < ncols2; col2 += block_size) { + for (int col2 = tid; col2 < ncols2; col2 += block_size) { const float2 tmpx = x2[col2]; - const float2 tmpy = y2[col2]; - sumf += tmpx.x*tmpy.x; - sumf += tmpx.y*tmpy.y; + +#pragma unroll + for (int j = 0; j < ncols_dst; ++j) { + const float2 tmpy = y2[j*stride_col_y2 + col2]; + sumf[j] += tmpx.x*tmpy.x; + sumf[j] += tmpx.y*tmpy.y; + } } } else if constexpr (std::is_same::value) { const half2 * x2 = (const half2 *) x; if (std::is_same::value) { - for (int64_t col2 = tid; col2 < ncols2; col2 += block_size) { + for (int col2 = tid; col2 < ncols2; col2 += block_size) { const float2 tmpx = __half22float2(x2[col2]); - const float2 tmpy = y2[col2]; - sumf += tmpx.x * tmpy.x; - sumf += tmpx.y * tmpy.y; + +#pragma unroll + for (int j = 0; j < ncols_dst; ++j) { + const float2 tmpy = y2[j*stride_col_y2 + col2]; + sumf[j] += tmpx.x * tmpy.x; + sumf[j] += tmpx.y * tmpy.y; + } } } else { #ifdef FP16_AVAILABLE - half2 sumh2 = make_half2(0.0f, 0.0f); + half2 sumh2[ncols_dst] = {{0.0f, 0.0f}}; - for (int64_t col2 = tid; col2 < ncols2; col2 += block_size) { - const float2 tmp = y2[col2]; - sumh2 += x2[col2] * make_half2(tmp.x, tmp.y); + for (int col2 = tid; col2 < ncols2; col2 += block_size) { + const half2 tmpx = x2[col2]; + +#pragma unroll + for (int j = 0; j < ncols_dst; ++j) { + const float2 tmpy = y2[j*stride_col_y2 + col2]; + sumh2[j] += tmpx * make_half2(tmpy.x, tmpy.y); + } } - sumf = __low2float(sumh2) + __high2float(sumh2); +#pragma unroll + for (int j = 0; j < ncols_dst; ++j) { + sumf[j] = __low2float(sumh2[j]) + __high2float(sumh2[j]); + } #else NO_DEVICE_CODE; #endif // FP16_AVAILABLE } } else if constexpr (std::is_same::value) { const int * x2 = (const int *) x; - for (int64_t col2 = tid; col2 < ncols2; col2 += block_size) { - const int tmpx = x2[col2]; - const float2 tmpy = y2[col2]; - sumf += float(reinterpret_cast(&tmpx)[0]) * tmpy.x; - sumf += float(reinterpret_cast(&tmpx)[1]) * tmpy.y; + for (int col2 = tid; col2 < ncols2; col2 += block_size) { + const int tmpx = x2[col2]; +#pragma unroll + for (int j = 0; j < ncols_dst; ++j) { + const float2 tmpy = y2[j*stride_col_y2 + col2]; + sumf[j] += float(reinterpret_cast(&tmpx)[0]) * tmpy.x; + sumf[j] += float(reinterpret_cast(&tmpx)[1]) * tmpy.y; + } } } else { static_assert(std::is_same::value, "unsupported type"); } - sumf = warp_reduce_sum(sumf); +#pragma unroll + for (int j = 0; j < ncols_dst; ++j) { + sumf[j] = warp_reduce_sum(sumf[j]); - if (block_size > warp_size) { - buf_iw[tid/warp_size] = sumf; - __syncthreads(); - if (tid >= warp_size) { - return; + if (block_size > warp_size) { + buf_iw[tid/warp_size] = sumf[j]; + __syncthreads(); + if (tid < warp_size) { + sumf[j] = buf_iw[tid]; + sumf[j] = warp_reduce_sum(sumf[j]); + } + if (j < ncols_dst) { + __syncthreads(); + } } - sumf = buf_iw[tid]; - sumf = warp_reduce_sum(sumf); } - if (tid != 0) { + if (tid >= ncols_dst) { return; } - dst[row] = sumf; + dst[tid*stride_col_dst + row] = sumf[tid]; } -template +template static void launch_mul_mat_vec_cuda( const T * x, const float * y, const int32_t * ids, float * dst, - const int64_t ncols, const int64_t nrows, const int64_t stride_row, const int64_t nchannels_x, const int64_t nchannels_y, const int64_t nchannels_dst, + const int64_t ncols, const int64_t nrows, + const int64_t stride_row, const int64_t stride_col_y, const int64_t stride_col_dst, + const int64_t nchannels_x, const int64_t nchannels_y, const int64_t nchannels_dst, const int64_t stride_channel_x, const int64_t stride_channel_y, const int64_t stride_channel_dst, const int64_t nsamples_x, const int64_t nsamples_dst, const int64_t stride_sample_x, const int64_t stride_sample_y, const int64_t stride_sample_dst, cudaStream_t stream) { - GGML_ASSERT(ncols % 2 == 0); - GGML_ASSERT(stride_row % 2 == 0); + GGML_ASSERT(ncols % 2 == 0); + GGML_ASSERT(stride_row % 2 == 0); + GGML_ASSERT(stride_col_y % 2 == 0); GGML_ASSERT(ids || nchannels_dst % nchannels_x == 0); GGML_ASSERT( nsamples_dst % nsamples_x == 0); const int64_t channel_ratio = nchannels_dst / nchannels_x; @@ -138,44 +166,52 @@ static void launch_mul_mat_vec_cuda( const dim3 block_dims(block_size_best, 1, 1); switch (block_size_best) { case 32: { - mul_mat_vec<<>> - (x, y, ids, dst, ncols/2, nchannels_y, stride_row, channel_ratio, stride_channel_x, stride_channel_y, - stride_channel_dst, sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst); + mul_mat_vec<<>> + (x, y, ids, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst, + channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst, + sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst); } break; case 64: { - mul_mat_vec<<>> - (x, y, ids, dst, ncols/2, nchannels_y, stride_row, channel_ratio, stride_channel_x, stride_channel_y, - stride_channel_dst, sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst); + mul_mat_vec<<>> + (x, y, ids, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst, + channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst, + sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst); } break; case 96: { - mul_mat_vec<<>> - (x, y, ids, dst, ncols/2, nchannels_y, stride_row, channel_ratio, stride_channel_x, stride_channel_y, - stride_channel_dst, sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst); + mul_mat_vec<<>> + (x, y, ids, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst, + channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst, + sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst); } break; case 128: { - mul_mat_vec<<>> - (x, y, ids, dst, ncols/2, nchannels_y, stride_row, channel_ratio, stride_channel_x, stride_channel_y, - stride_channel_dst, sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst); + mul_mat_vec<<>> + (x, y, ids, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst, + channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst, + sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst); } break; case 160: { - mul_mat_vec<<>> - (x, y, ids, dst, ncols/2, nchannels_y, stride_row, channel_ratio, stride_channel_x, stride_channel_y, - stride_channel_dst, sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst); + mul_mat_vec<<>> + (x, y, ids, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst, + channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst, + sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst); } break; case 192: { - mul_mat_vec<<>> - (x, y, ids, dst, ncols/2, nchannels_y, stride_row, channel_ratio, stride_channel_x, stride_channel_y, - stride_channel_dst, sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst); + mul_mat_vec<<>> + (x, y, ids, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst, + channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst, + sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst); } break; case 224: { - mul_mat_vec<<>> - (x, y, ids, dst, ncols/2, nchannels_y, stride_row, channel_ratio, stride_channel_x, stride_channel_y, - stride_channel_dst, sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst); + mul_mat_vec<<>> + (x, y, ids, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst, + channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst, + sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst); } break; case 256: { - mul_mat_vec<<>> - (x, y, ids, dst, ncols/2, nchannels_y, stride_row, channel_ratio, stride_channel_x, stride_channel_y, - stride_channel_dst, sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst); + mul_mat_vec<<>> + (x, y, ids, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst, + channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst, + sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst); } break; default: { GGML_ABORT("fatal error"); @@ -183,23 +219,91 @@ static void launch_mul_mat_vec_cuda( } } +template +static void mul_mat_vec_cuda_switch_ncols_dst( + const T * x, const float * y, const int32_t * ids, float * dst, + const int64_t ncols, const int64_t nrows, const int64_t ncols_dst, + const int64_t stride_row, const int64_t stride_col_y, const int64_t stride_col_dst, + const int64_t nchannels_x, const int64_t nchannels_y, const int64_t nchannels_dst, + const int64_t stride_channel_x, const int64_t stride_channel_y, const int64_t stride_channel_dst, const int64_t nsamples_x, + const int64_t nsamples_dst, const int64_t stride_sample_x, const int64_t stride_sample_y, const int64_t stride_sample_dst, + cudaStream_t stream) { + switch (ncols_dst) { + case 1: + launch_mul_mat_vec_cuda + (x, y, ids, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst, + nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, + stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream); + break; + case 2: + launch_mul_mat_vec_cuda + (x, y, ids, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst, + nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, + stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream); + break; + case 3: + launch_mul_mat_vec_cuda + (x, y, ids, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst, + nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, + stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream); + break; + case 4: + launch_mul_mat_vec_cuda + (x, y, ids, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst, + nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, + stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream); + break; + case 5: + launch_mul_mat_vec_cuda + (x, y, ids, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst, + nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, + stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream); + break; + case 6: + launch_mul_mat_vec_cuda + (x, y, ids, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst, + nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, + stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream); + break; + case 7: + launch_mul_mat_vec_cuda + (x, y, ids, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst, + nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, + stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream); + break; + case 8: + launch_mul_mat_vec_cuda + (x, y, ids, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst, + nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, + stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream); + break; + default: + GGML_ABORT("fatal error"); + break; + } +} + template static void mul_mat_vec_cuda( const T * x, const float * y, const int32_t * ids, float * dst, - const int64_t ncols, const int64_t nrows, const int64_t stride_row, const int64_t nchannels_x, const int64_t nchannels_y, const int64_t nchannels_dst, + const int64_t ncols, const int64_t nrows, const int64_t ncols_dst, + const int64_t stride_row, const int64_t stride_col_y, const int stride_col_dst, + const int64_t nchannels_x, const int64_t nchannels_y, const int64_t nchannels_dst, const int64_t stride_channel_x, const int64_t stride_channel_y, const int64_t stride_channel_dst, const int64_t nsamples_x, const int64_t nsamples_dst, const int64_t stride_sample_x, const int64_t stride_sample_y, const int64_t stride_sample_dst, enum ggml_prec prec, cudaStream_t stream) { if constexpr(std::is_same::value) { if (prec == GGML_PREC_DEFAULT) { - launch_mul_mat_vec_cuda - (x, y, ids, dst, ncols, nrows, stride_row, nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, + mul_mat_vec_cuda_switch_ncols_dst + (x, y, ids, dst, ncols, nrows, ncols_dst, stride_row, stride_col_y, stride_col_dst, + nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream); return; } } - launch_mul_mat_vec_cuda - (x, y, ids, dst, ncols, nrows, stride_row, nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, + mul_mat_vec_cuda_switch_ncols_dst + (x, y, ids, dst, ncols, nrows, ncols_dst, stride_row, stride_col_y, stride_col_dst, + nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream); } @@ -246,24 +350,24 @@ void ggml_cuda_mul_mat_vec(ggml_backend_cuda_context & ctx, const ggml_tensor * const int64_t stride_channel_dst = ids ? s1 : s2; const int64_t stride_channel_y = ids ? s11 : s12; - GGML_ASSERT(ncols_dst == 1); + GGML_ASSERT(!ids || ncols_dst == 1); switch (src0->type) { case GGML_TYPE_F32: { const float * src0_d = (const float *) src0->data; - mul_mat_vec_cuda(src0_d, src1_d, ids_d, dst_d, ne00, ne01, s01, + mul_mat_vec_cuda(src0_d, src1_d, ids_d, dst_d, ne00, ne01, ncols_dst, s01, s11, s1, ne02, nchannels_y, nchannels_dst, s02, stride_channel_y, stride_channel_dst, ne03, ne3, s03, s13, s3, prec, ctx.stream()); } break; case GGML_TYPE_F16: { const half * src0_d = (const half *) src0->data; - mul_mat_vec_cuda(src0_d, src1_d, ids_d, dst_d, ne00, ne01, s01, + mul_mat_vec_cuda(src0_d, src1_d, ids_d, dst_d, ne00, ne01, ncols_dst, s01, s11, s1, ne02, nchannels_y, nchannels_dst, s02, stride_channel_y, stride_channel_dst, ne03, ne3, s03, s13, s3, prec, ctx.stream()); } break; case GGML_TYPE_BF16: { const nv_bfloat16 * src0_d = (const nv_bfloat16 *) src0->data; - mul_mat_vec_cuda(src0_d, src1_d, ids_d, dst_d, ne00, ne01, s01, + mul_mat_vec_cuda(src0_d, src1_d, ids_d, dst_d, ne00, ne01, ncols_dst, s01, s11, s1, ne02, nchannels_y, nchannels_dst, s02, stride_channel_y, stride_channel_dst, ne03, ne3, s03, s13, s3, prec, ctx.stream()); } break; @@ -282,16 +386,19 @@ void ggml_cuda_op_mul_mat_vec( GGML_ASSERT(dst->type == GGML_TYPE_F32); const int64_t ne00 = src0->ne[0]; + const int64_t ne10 = src1->ne[0]; + const int64_t ne0 = dst->ne[0]; const int64_t row_diff = row_high - row_low; - GGML_ASSERT(src1_ncols == 1); - - const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc; + const int id = ggml_cuda_get_device(); + const int cc = ggml_cuda_info().devices[id].cc; const enum ggml_prec prec = fast_fp16_available(cc) ? ggml_prec(dst->op_params[0]) : GGML_PREC_F32; // ggml_cuda_op provides single, contiguous matrices const int64_t stride_row = ne00; + const int64_t stride_col_y = ne10; + const int64_t stride_col_dst = id == ctx.device ? ne0 : row_diff; // main device has larger memory buffer const int64_t nchannels_x = 1; const int64_t nchannels_y = 1; const int64_t nchannels_dst = 1; @@ -307,19 +414,19 @@ void ggml_cuda_op_mul_mat_vec( switch (src0->type) { case GGML_TYPE_F32: { const float * src0_d = (const float *) src0_dd_i; - mul_mat_vec_cuda(src0_d, src1_ddf_i, nullptr, dst_dd_i, ne00, row_diff, stride_row, + mul_mat_vec_cuda(src0_d, src1_ddf_i, nullptr, dst_dd_i, ne00, row_diff, src1_ncols, stride_row, stride_col_y, stride_col_dst, nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, prec, stream); } break; case GGML_TYPE_F16: { const half * src0_d = (const half *) src0_dd_i; - mul_mat_vec_cuda(src0_d, src1_ddf_i, nullptr, dst_dd_i, ne00, row_diff, stride_row, + mul_mat_vec_cuda(src0_d, src1_ddf_i, nullptr, dst_dd_i, ne00, row_diff, src1_ncols, stride_row, stride_col_y, stride_col_dst, nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, prec, stream); } break; case GGML_TYPE_BF16: { const nv_bfloat16 * src0_d = (const nv_bfloat16 *) src0_dd_i; - mul_mat_vec_cuda(src0_d, src1_ddf_i, nullptr, dst_dd_i, ne00, row_diff, stride_row, + mul_mat_vec_cuda(src0_d, src1_ddf_i, nullptr, dst_dd_i, ne00, row_diff, src1_ncols, stride_row, stride_col_y, stride_col_dst, nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, prec, stream); } break; @@ -334,3 +441,48 @@ void ggml_cuda_op_mul_mat_vec( GGML_UNUSED(src1_ncols); GGML_UNUSED(src1_padded_row_size); } + +bool ggml_cuda_should_use_mmv(enum ggml_type type, int cc, const int64_t * src0_ne, int64_t ne11) { + if (src0_ne[0] % 2 != 0) { + return false; + } + switch (type) { + case GGML_TYPE_F32: + if (GGML_CUDA_CC_IS_NVIDIA(cc)) { + if (cc >= GGML_CUDA_CC_ADA_LOVELACE) { + return ne11 <= 8; + } + if (cc >= GGML_CUDA_CC_TURING) { + return ne11 <= 4; + } + return ne11 <= 3; + } + return ne11 <= 8; + case GGML_TYPE_F16: + if (GGML_CUDA_CC_IS_NVIDIA(cc)) { + const bool src0_small = (src0_ne[1] <= 512 || src0_ne[2]*src0_ne[3] == 1); + if (cc >= GGML_CUDA_CC_ADA_LOVELACE) { + return src0_small && ne11 <= 4; + } + if (fp16_mma_hardware_available(cc)) { + return src0_small && ne11 <= 3; + } + return ne11 <= 8; + } + return ne11 <= 8; + case GGML_TYPE_BF16: + if (GGML_CUDA_CC_IS_NVIDIA(cc)) { + const bool src0_small = (src0_ne[1] <= 512 || src0_ne[2]*src0_ne[3] == 1); + if (cc >= GGML_CUDA_CC_ADA_LOVELACE) { + return src0_small && ne11 <= 4; + } + if (bf16_mma_hardware_available(cc)) { + return src0_small && ne11 <= 3; + } + return ne11 <= 8; + } + return ne11 <= 8; + default: + return false; + } +} diff --git a/ggml/src/ggml-cuda/mmv.cuh b/ggml/src/ggml-cuda/mmv.cuh index 756e7e1cc..1330bcb6a 100644 --- a/ggml/src/ggml-cuda/mmv.cuh +++ b/ggml/src/ggml-cuda/mmv.cuh @@ -1,8 +1,5 @@ #include "common.cuh" -// maximum number of src0 rows with which to use mul_mat_vec over cuBLAS if FP16 tensor cores are available -#define MMV_MAX_ROWS 512 - void ggml_cuda_mul_mat_vec(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst); void ggml_cuda_op_mul_mat_vec( @@ -10,3 +7,5 @@ void ggml_cuda_op_mul_mat_vec( const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i, const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols, const int64_t src1_padded_row_size, cudaStream_t stream); + +bool ggml_cuda_should_use_mmv(enum ggml_type type, int cc, const int64_t * src0_ne, int64_t ne11); From 72c6bc3f3d0cf3bf160dddf1b803fed52bcbb0a3 Mon Sep 17 00:00:00 2001 From: Molly Sophia Date: Mon, 23 Jun 2025 19:56:19 +0800 Subject: [PATCH 42/43] llama : better rwkv chat template and add missing `inputs.use_jinja` setting (#14336) * llama-cli : add missing `inputs.use_jinja` setting Signed-off-by: Molly Sophia * llama : better legacy chat template for rwkv Signed-off-by: Molly Sophia --------- Signed-off-by: Molly Sophia --- src/llama-chat.cpp | 17 +++++++++++------ tools/main/main.cpp | 1 + 2 files changed, 12 insertions(+), 6 deletions(-) diff --git a/src/llama-chat.cpp b/src/llama-chat.cpp index 0839cad3e..5d317f4ee 100644 --- a/src/llama-chat.cpp +++ b/src/llama-chat.cpp @@ -528,12 +528,17 @@ int32_t llm_chat_apply_template( } } else if (tmpl == LLM_CHAT_TEMPLATE_RWKV_WORLD) { // this template requires the model to have "\n\n" as EOT token - for (auto message : chat) { - std::string role(message->role); - if (role == "user") { - ss << "User: " << message->content << "\n\nAssistant:"; - } else { - ss << message->content << "\n\n"; + for (size_t i = 0; i < chat.size(); i++) { + std::string role(chat[i]->role); + if (role == "system") { + ss << "System: " << trim(chat[i]->content) << "\n\n"; + } else if (role == "user") { + ss << "User: " << trim(chat[i]->content) << "\n\n"; + if (i == chat.size() - 1) { + ss << "Assistant:"; + } + } else if (role == "assistant") { + ss << "Assistant: " << trim(chat[i]->content) << "\n\n"; } } } else if (tmpl == LLM_CHAT_TEMPLATE_GRANITE) { diff --git a/tools/main/main.cpp b/tools/main/main.cpp index 19b247b0d..154b37cdb 100644 --- a/tools/main/main.cpp +++ b/tools/main/main.cpp @@ -292,6 +292,7 @@ int main(int argc, char ** argv) { if (!params.system_prompt.empty() || !params.prompt.empty()) { common_chat_templates_inputs inputs; + inputs.use_jinja = g_params->use_jinja; inputs.messages = chat_msgs; inputs.add_generation_prompt = !params.prompt.empty(); From bf2a99e3cb06a14dce2a586450d012dc31f922ae Mon Sep 17 00:00:00 2001 From: Jeff Bolz Date: Mon, 23 Jun 2025 08:44:48 -0500 Subject: [PATCH 43/43] vulkan: update windows SDK in release.yml (#14344) --- .github/workflows/release.yml | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/.github/workflows/release.yml b/.github/workflows/release.yml index 9874736cb..64fff175e 100644 --- a/.github/workflows/release.yml +++ b/.github/workflows/release.yml @@ -302,7 +302,7 @@ jobs: env: OPENBLAS_VERSION: 0.3.23 - VULKAN_VERSION: 1.4.309.0 + VULKAN_VERSION: 1.4.313.2 strategy: matrix: @@ -332,7 +332,7 @@ jobs: id: get_vulkan if: ${{ matrix.backend == 'vulkan' }} run: | - curl.exe -o $env:RUNNER_TEMP/VulkanSDK-Installer.exe -L "https://sdk.lunarg.com/sdk/download/${env:VULKAN_VERSION}/windows/VulkanSDK-${env:VULKAN_VERSION}-Installer.exe" + curl.exe -o $env:RUNNER_TEMP/VulkanSDK-Installer.exe -L "https://sdk.lunarg.com/sdk/download/${env:VULKAN_VERSION}/windows/vulkansdk-windows-X64-${env:VULKAN_VERSION}.exe" & "$env:RUNNER_TEMP\VulkanSDK-Installer.exe" --accept-licenses --default-answer --confirm-command install Add-Content $env:GITHUB_ENV "VULKAN_SDK=C:\VulkanSDK\${env:VULKAN_VERSION}" Add-Content $env:GITHUB_PATH "C:\VulkanSDK\${env:VULKAN_VERSION}\bin"