tqcq/llama.cpp
0
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-08-17 21:51:27 -04:00
Code Issues Packages Projects Releases Wiki Activity

CUDA: add FP32 FlashAttention vector kernel (#7188)

Browse Source 
* CUDA: add FP32 FlashAttention vector kernel

* fixup! CUDA: add FP32 FlashAttention vector kernel

* fixup! fixup! CUDA: add FP32 FlashAttention vector kernel

* fixup! fixup! fixup! CUDA: add FP32 FlashAttention vector kernel
This commit is contained in:
Johannes Gäßler
2024-05-12 19:40:45 +02:00
committed by GitHub
parent 6f1b63606f
commit dc685be466
9 changed files with 899 additions and 458 deletions
Download Patch File Download Diff File Expand all files Collapse all files

11
ggml-cuda.cu
View File

@@ -2713,6 +2713,7 @@ GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t
}
GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, const ggml_tensor * op) {
ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *) backend->context;
switch (op->op) {
case GGML_OP_UNARY:
switch (ggml_get_unary_op(op)) {
@@ -2840,8 +2841,16 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons
case GGML_OP_ARANGE:
case GGML_OP_TIMESTEP_EMBEDDING:
case GGML_OP_LEAKY_RELU:
case GGML_OP_FLASH_ATTN_EXT:
return true;
case GGML_OP_FLASH_ATTN_EXT:
#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
return op->src[0]->ne[0] == 64 || op->src[0]->ne[0] == 128;
#else
if (op->src[0]->ne[0] == 64 || op->src[0]->ne[0] == 128) {
return true;
}
return ggml_cuda_info().devices[cuda_ctx->device].cc >= CC_VOLTA;
#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
default:
return false;
}