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

CUDA: generalized (mma) FA, add Volta support (#17505)

Browse Source 
* CUDA: generalized (mma) FA, add Volta support

* use struct for MMA FA kernel config

---------

Co-authored-by: Aman Gupta <aman>
This commit is contained in:
Johannes Gäßler
2025-12-03 16:57:05 +01:00
committed by GitHub
co-authored by Aman Gupta <aman>
parent 190c4838bd
commit 2e1c9cd814
10 changed files with 966 additions and 759 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-cuda/fattn-wmma-f16.cu
+9 -8
View File
@@ -38,14 +38,14 @@ static __global__ void flash_attn_ext_f16(
const float m1,
const uint32_t n_head_log2,
const float logit_softcap,
const int32_t ne00, const int32_t ne01, const int32_t ne02, const int32_t ne03,
const int32_t ne00, const uint3 ne01, const int32_t ne02, const int32_t ne03,
const int32_t nb01, const int32_t nb02, const int32_t nb03,
const int32_t ne10, const int32_t ne11, const int32_t ne12, const int32_t ne13,
const int32_t nb11, const int32_t nb12, const int64_t nb13,
const int32_t nb21, const int32_t nb22, const int64_t nb23,
const int32_t ne31, const int32_t ne32, const int32_t ne33,
const int32_t nb31, const int32_t nb32, const int64_t nb33) {
#if defined(FLASH_ATTN_AVAILABLE) && (__CUDA_ARCH__ == GGML_CUDA_CC_VOLTA || (defined(GGML_HIP_ROCWMMA_FATTN) && defined(GGML_USE_WMMA_FATTN)))
#if defined(FLASH_ATTN_AVAILABLE) && (defined(GGML_HIP_ROCWMMA_FATTN) && defined(GGML_USE_WMMA_FATTN))
// Skip unused kernel variants for faster compilation:
if (use_logit_softcap && !(D == 128 || D == 256)) {
NO_DEVICE_CODE;
@@ -149,7 +149,7 @@ static __global__ void flash_attn_ext_f16(
if (i0 + warp_size > D && i >= D) {
break;
}
KQ[j*D_padded + i] = ic0 + j < ne01 ? Q_f[j*stride_Q + i] * scale : 0.0f;
KQ[j*D_padded + i] = ic0 + j < int(ne01.z) ? Q_f[j*stride_Q + i] * scale : 0.0f;
}
}
@@ -218,7 +218,8 @@ static __global__ void flash_attn_ext_f16(
for (int k0 = 0; k0 < FATTN_KQ_STRIDE; k0 += warp_size) {
const int k = k0 + threadIdx.x;
KQ_f_tmp[k0/warp_size] += mask ? __half2float(slopeh*maskh[j*(nb31/sizeof(half)) + k_VKQ_0 + k]) : 0.0f;
KQ_f_tmp[k0/warp_size] += mask && ic0 + j < int(ne01.z) ?
__half2float(slopeh*maskh[j*(nb31/sizeof(half)) + k_VKQ_0 + k]) : 0.0f;
KQ_max_new = max(KQ_max_new, KQ_f_tmp[k0/warp_size]);
}
KQ_max_new = warp_reduce_max<warp_size>(KQ_max_new);
@@ -270,7 +271,7 @@ static __global__ void flash_attn_ext_f16(
for (int k0 = 0; k0 < FATTN_KQ_STRIDE/2; k0 += warp_size) {
const int k = k0 + threadIdx.x;
KQ2_tmp[k0/warp_size] += mask ? slope2*mask2[(j*ne11 + k_VKQ_0)/2 + k] : make_half2(0.0f, 0.0f);
KQ2_tmp[k0/warp_size] += mask && ic0 + j < int(ne01.z) ? slope2*mask2[(j*ne11 + k_VKQ_0)/2 + k] : make_half2(0.0f, 0.0f);
KQ_max_new = ggml_cuda_hmax2(KQ_max_new, KQ2_tmp[k0/warp_size]);
}
KQ_max_new = __half2half2(warp_reduce_max<warp_size>(ggml_cuda_hmax(__low2half(KQ_max_new), __high2half(KQ_max_new))));
@@ -431,7 +432,7 @@ static __global__ void flash_attn_ext_f16(
#pragma unroll
for (int j0 = 0; j0 < ncols; j0 += nwarps) {
const int j_VKQ = j0 + threadIdx.y;
if (ic0 + j_VKQ >= ne01) {
if (ic0 + j_VKQ >= int(ne01.z)) {
return;
}
@@ -442,7 +443,7 @@ static __global__ void flash_attn_ext_f16(
KQ_rowsum_j = __low2float(KQ_rowsum_h2[j0/nwarps]) + __high2float(KQ_rowsum_h2[j0/nwarps]);
}
const int j_dst_unrolled = ((sequence*ne01 + ic0 + j_VKQ)*ne02 + head)*gridDim.y + blockIdx.y;
const int j_dst_unrolled = ((sequence*int(ne01.z) + ic0 + j_VKQ)*ne02 + head)*gridDim.y + blockIdx.y;
#pragma unroll
for (int i0 = 0; i0 < D; i0 += warp_size) {
@@ -481,7 +482,7 @@ static __global__ void flash_attn_ext_f16(
ne31, ne32, ne33,
nb31, nb32, nb33);
NO_DEVICE_CODE;
#endif // defined(FLASH_ATTN_AVAILABLE) && (__CUDA_ARCH__ == GGML_CUDA_CC_VOLTA || (defined(GGML_HIP_ROCWMMA_FATTN) && defined(GGML_USE_WMMA_FATTN)))
#endif // defined(FLASH_ATTN_AVAILABLE) && (defined(GGML_HIP_ROCWMMA_FATTN) && defined(GGML_USE_WMMA_FATTN))
}
constexpr int get_max_power_of_2(int x) {