Lumpiasty/llama.cpp
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

[MUSA] enable fp16/fast_fp16/bf16_mma on PH1 (#17551)

Browse Source 
* [MUSA] enable fp16/fast_fp16/bf16_mma on PH1

Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com>

* Update ggml/src/ggml-cuda/fattn-vec.cuh

Co-authored-by: Johannes Gäßler <johannesg@5d6.de>

* Update ggml/src/ggml-cuda/fattn-vec.cuh

Co-authored-by: Johannes Gäßler <johannesg@5d6.de>

* Update ggml/src/ggml-cuda/fattn-tile.cuh

Co-authored-by: Johannes Gäßler <johannesg@5d6.de>

* Address review comments

Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com>

---------

Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com>
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
This commit is contained in:
R0CKSTAR
2025-11-28 14:08:29 +01:00
committed by GitHub
co-authored by Johannes Gäßler
parent 2e7ef98f18
commit c6f7a423c8
5 changed files with 21 additions and 14 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-cuda/common.cuh
+12 -8
View File
@@ -84,12 +84,12 @@
#define GGML_CUDA_CC_QY1 (GGML_CUDA_CC_OFFSET_MTHREADS + 0x210) // MTT S80, MTT S3000
#define GGML_CUDA_CC_QY2 (GGML_CUDA_CC_OFFSET_MTHREADS + 0x220) // MTT S4000
#define GGML_CUDA_CC_NG (GGML_CUDA_CC_OFFSET_MTHREADS + 0x310) // TBD
#define GGML_CUDA_CC_PH1 (GGML_CUDA_CC_OFFSET_MTHREADS + 0x310) // MTT S5000
#define GGML_CUDA_CC_IS_MTHREADS(cc) (cc >= GGML_CUDA_CC_OFFSET_MTHREADS && cc < GGML_CUDA_CC_OFFSET_AMD)
#define GGML_CUDA_CC_IS_QY1(cc) (cc >= GGML_CUDA_CC_QY1 && cc < GGML_CUDA_CC_QY2)
#define GGML_CUDA_CC_IS_QY2(cc) (cc >= GGML_CUDA_CC_QY2 && cc < GGML_CUDA_CC_NG)
#define GGML_CUDA_CC_IS_NG(cc) (cc >= GGML_CUDA_CC_NG)
#define GGML_CUDA_CC_IS_QY2(cc) (cc >= GGML_CUDA_CC_QY2 && cc < GGML_CUDA_CC_PH1)
#define GGML_CUDA_CC_IS_PH1(cc) (cc >= GGML_CUDA_CC_PH1)
#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA) && CUDART_VERSION >= 11070
# define GGML_CUDA_USE_CUB
@@ -212,9 +212,9 @@ static const char * cu_get_error_str(CUresult err) {
#define GGML_USE_VMM
#endif // (!defined(GGML_USE_HIP) && !defined(GGML_CUDA_NO_VMM)) || (defined(GGML_USE_HIP) && !defined(GGML_HIP_NO_VMM))
#if defined(GGML_USE_HIP) || __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL
#if defined(GGML_USE_HIP) || defined(GGML_USE_MUSA) || __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL
#define FP16_AVAILABLE
#endif // defined(GGML_USE_HIP) || __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL
#endif // defined(GGML_USE_HIP) || defined(GGML_USE_MUSA) || __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL
#if defined(FP16_AVAILABLE) && __CUDA_ARCH__ != 610
#define FAST_FP16_AVAILABLE
@@ -250,12 +250,14 @@ static const char * cu_get_error_str(CUresult err) {
#endif // !defined(GGML_CUDA_NO_FA) && !(defined(GGML_USE_MUSA) && __MUSA_ARCH__ < 220)
static bool fp16_available(const int cc) {
return ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_PASCAL;
return ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_PASCAL ||
(GGML_CUDA_CC_IS_MTHREADS(cc) && cc >= GGML_CUDA_CC_PH1);
}
static bool fast_fp16_available(const int cc) {
return GGML_CUDA_CC_IS_AMD(cc) ||
(GGML_CUDA_CC_IS_NVIDIA(cc) && fp16_available(cc) && ggml_cuda_highest_compiled_arch(cc) != 610);
(GGML_CUDA_CC_IS_NVIDIA(cc) && fp16_available(cc) && ggml_cuda_highest_compiled_arch(cc) != 610) ||
(GGML_CUDA_CC_IS_MTHREADS(cc) && fp16_available(cc));
}
// To be used for feature selection of external libraries, e.g. cuBLAS.
@@ -272,7 +274,9 @@ static bool fp16_mma_hardware_available(const int cc) {
}
static bool bf16_mma_hardware_available(const int cc) {
return (GGML_CUDA_CC_IS_NVIDIA(cc) && cc >= GGML_CUDA_CC_AMPERE) || GGML_CUDA_CC_IS_CDNA(cc) || cc >= GGML_CUDA_CC_RDNA3;
return (GGML_CUDA_CC_IS_NVIDIA(cc) && cc >= GGML_CUDA_CC_AMPERE) ||
GGML_CUDA_CC_IS_CDNA(cc) || cc >= GGML_CUDA_CC_RDNA3 ||
(GGML_CUDA_CC_IS_MTHREADS(cc) && cc >= GGML_CUDA_CC_PH1);
}
static bool fp32_mma_hardware_available(const int cc) {