ver4a/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

HIP: implement FlashAttention via rocWMMA for CDNA and RDNA3+ (#12032)

Browse source 
Adds GGML_HIP_ROCWMMA_FATTN and rocwmma header check
Adds rocWMMA support to fattn-wmma-f16

---

Signed-off-by: Carl Klemm <carl@uvos.xyz>
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
Co-authored-by: Ben Jackson <ben@ben.com>
This commit is contained in:
David Huang 2025-03-04 05:10:54 +08:00 • committed by GitHub
parent dfd6b2c0be
commit becade5de7
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
6 changed files with 145 additions and 95 deletions
Download patch file Download diff file Expand all files Collapse all files

15
ggml/src/ggml-cuda/common.cuh
View file

@ -62,6 +62,7 @@
#define GGML_CUDA_CC_RDNA2 (GGML_CUDA_CC_OFFSET_AMD + 0x1030) // RX 6000, minimum for dp4a
#define GGML_CUDA_CC_RDNA3 (GGML_CUDA_CC_OFFSET_AMD + 0x1100) // RX 7000, minimum for WMMA
#define GGML_CUDA_CC_IS_AMD(cc) (cc >= GGML_CUDA_CC_OFFSET_AMD)
#define GGML_CUDA_CC_IS_RDNA(cc) (cc >= GGML_CUDA_CC_RDNA1)
#define GGML_CUDA_CC_IS_RDNA1(cc) (cc >= GGML_CUDA_CC_RDNA1 && cc < GGML_CUDA_CC_RDNA2)
#define GGML_CUDA_CC_IS_RDNA2(cc) (cc >= GGML_CUDA_CC_RDNA2 && cc < GGML_CUDA_CC_RDNA3)
@ -196,6 +197,10 @@ typedef float2 dfloat2;
#define FP16_MMA_AVAILABLE
#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA
#if defined(GGML_HIP_ROCWMMA_FATTN) && (defined(CDNA) || defined(RDNA3))
#define FP16_MMA_AVAILABLE
#endif // defined(GGML_HIP_ROCWMMA_FATTN) && (defined(CDNA) || defined(RDNA3))
#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_TURING
#define NEW_MMA_AVAILABLE
#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_TURING
@ -223,12 +228,18 @@ static bool fast_fp16_hardware_available(const int cc) {
// Any FP16 tensor core instructions are available for ggml code.
static bool fp16_mma_available(const int cc) {
return cc < GGML_CUDA_CC_OFFSET_AMD && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA;
#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && !defined(GGML_HIP_ROCWMMA_FATTN)
return false;
#else
return cc < GGML_CUDA_CC_OFFSET_AMD && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA ||
GGML_CUDA_CC_IS_CDNA(cc) || cc >= GGML_CUDA_CC_RDNA3;
#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && !defined(GGML_HIP_ROCWMMA_FATTN)
}
// To be used for feature selection of external libraries, e.g. cuBLAS.
static bool fp16_mma_hardware_available(const int cc) {
return cc < GGML_CUDA_CC_OFFSET_AMD && cc >= GGML_CUDA_CC_VOLTA;
return cc < GGML_CUDA_CC_OFFSET_AMD && cc >= GGML_CUDA_CC_VOLTA ||
GGML_CUDA_CC_IS_CDNA(cc) || cc >= GGML_CUDA_CC_RDNA3;
}
// Volta technically had FP16 tensor cores but they work very differently compared to Turing and later.