tqcq/llama.cpp
0
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-08-15 12:42:40 -04:00
Code Issues Packages Projects Releases Wiki Activity

llama : add gpt-oss (#15091)

Browse Source 
* oai moe

* compat with new checkpoint

* add attn sink impl

* add rope scaling yarn

* logits match with latest transformers code

* wip chat template

* rm trailing space

* use ggml_scale_bias

* rm redundant is_swa_all

* convert interleaved gate_up

* graph : fix activation function to match reference (#7)

* vocab : handle o200k_harmony special tokens

* ggml : add attention sinks support (#1)

* llama : add attn sinks

* ggml : add attn sinks

* cuda : add attn sinks

* vulkan : add support for sinks in softmax

remove unnecessary return

* ggml : add fused swiglu_oai op (#11)

* ggml : add fused swiglu_oai op

* Update ggml/src/ggml-cpu/ops.cpp

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

* update CUDA impl

* cont : metal impl

* add vulkan impl

* test-backend-ops : more test cases, clean up

* llama : remove unfused impl

* remove extra lines

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

---------

Co-authored-by: slaren <slarengh@gmail.com>

* repack mxfp4 upon conversion

* clean up a bit

* enable thinking

* add quick hack to render only some special tokens

* fix bf16 conversion

* remove vocab hack

* webui ok

* support chat parsing for gpt-oss

* fix webui

* direct mapping mxfp4, FINALLY

* force using mxfp4

* properly use lazy tensor

* ggml : add mxfp4

ggml : use e8m0 conversion instead of powf

Co-authored-by: Diego Devesa <slarengh@gmail.com>

change kvalues_mxfp4 table to match e2m1 (#6)

metal : remove quantization for now (not used)

cuda : fix disabled CUDA graphs due to ffn moe bias

vulkan : add support for mxfp4

cont : add cm2 dequant

* ggml : add ggml_add_id (#13)

* ggml : add ggml_add_id

* add cuda impl

* llama : add weight support check for add_id

* perf opt

* add vulkan impl

* rename cuda files

* add metal impl

* allow in-place ggml_add_id

* llama : keep biases on CPU with --cpu-moe

* llama : fix compile error

ggml-ci

* cuda : add fallback for __nv_cvt_e8m0_to_bf16raw

ggml-ci

* cleanup

ggml-ci

* sycl : fix supports_op for MXFP4

ggml-ci

* fix Unknown reasoning format

* ggml-cpu : fix AVX build

ggml-ci

* fix hip build

ggml-ci

* cuda : add mxfp4 dequantization support for cuBLAS

ggml-ci

* ggml-cpu : fix mxfp4 fallback definitions for some architectures

ggml-ci

* cuda : fix version required for __nv_cvt_e8m0_to_bf16raw

---------

Co-authored-by: Xuan Son Nguyen <son@huggingface.co>
Co-authored-by: slaren <slarengh@gmail.com>
This commit is contained in:
Georgi Gerganov
2025-08-05 22:10:36 +03:00
committed by GitHub
parent f324a3b715
commit fd1234cb46
83 changed files with 2942 additions and 227 deletions
Download Patch File Download Diff File Expand all files Collapse all files

38
src/llama-quant.cpp
View File

@@ -211,7 +211,10 @@ static ggml_type llama_tensor_get_type(quantize_state_impl & qs, ggml_type new_t
const int64_t nx = tensor->ne[0];
const int64_t qk_k = ggml_blck_size(new_type);
if (arch == LLM_ARCH_FALCON || nx % qk_k != 0) {
if (ftype == LLAMA_FTYPE_MOSTLY_MXFP4_MOE) {
new_type = GGML_TYPE_Q8_0;
}
else if (arch == LLM_ARCH_FALCON || nx % qk_k != 0) {
new_type = GGML_TYPE_Q8_0;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS ||
@@ -223,6 +226,14 @@ static ggml_type llama_tensor_get_type(quantize_state_impl & qs, ggml_type new_t
new_type = GGML_TYPE_Q6_K;
}
}
} else if (ftype == LLAMA_FTYPE_MOSTLY_MXFP4_MOE) {
// MoE tensors -> MXFP4
// other tensors -> Q8_0
if (tensor->ne[2] > 1) {
new_type = GGML_TYPE_MXFP4;
} else {
new_type = GGML_TYPE_Q8_0;
}
} else if (name == "token_embd.weight" || name == "per_layer_token_embd.weight") {
if (qs.params->token_embedding_type < GGML_TYPE_COUNT) {
new_type = qs.params->token_embedding_type;
@@ -533,6 +544,8 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
case LLAMA_FTYPE_MOSTLY_BF16: default_type = GGML_TYPE_BF16; break;
case LLAMA_FTYPE_ALL_F32: default_type = GGML_TYPE_F32; break;
case LLAMA_FTYPE_MOSTLY_MXFP4_MOE: default_type = GGML_TYPE_MXFP4; break;
// K-quants
case LLAMA_FTYPE_MOSTLY_Q2_K_S:
case LLAMA_FTYPE_MOSTLY_Q2_K: default_type = GGML_TYPE_Q2_K; break;
@@ -984,6 +997,29 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
const float * imatrix_03 = imatrix ? imatrix + i03 * n_per_row : nullptr;
new_size += llama_tensor_quantize_impl(new_type, f32_data_03, new_data_03, chunk_size, nrows, n_per_row, imatrix_03, workers, nthread_use);
// TODO: temporary sanity check that the F16 -> MXFP4 is lossless
#if 1
if (new_type == GGML_TYPE_MXFP4) {
auto * x = f32_data_03;
//LLAMA_LOG_INFO("nrows = %d, n_per_row = %d\n", nrows, n_per_row);
std::vector<float> deq(nrows*n_per_row);
const ggml_type_traits * qtype = ggml_get_type_traits(new_type);
qtype->to_float(new_data_03, deq.data(), deq.size());
double err = 0.0f;
for (int i = 0; i < (int) deq.size(); ++i) {
err += fabsf(deq[i] - x[i]);
//if (fabsf(deq[i] - x[i]) > 0.00001 && i < 256) {
if (deq[i] != x[i]) {
LLAMA_LOG_INFO("deq[%d] = %f, x[%d] = %f\n", i, deq[i], i, x[i]);
}
}
//LLAMA_LOG_INFO("err = %f\n", err);
GGML_ASSERT(err == 0.00000);
}
#endif
}
LLAMA_LOG_INFO("size = %8.2f MiB -> %8.2f MiB\n", ggml_nbytes(tensor)/1024.0/1024.0, new_size/1024.0/1024.0);
}