EngineX-Ascend/enginex-ascend-910-llama.cpp
10
0
Fork 0
Code Issues Pull Requests Actions 4 Projects Releases Wiki Activity

CUDA: fix overflow in FA, tune performance (#14840)

Browse Source
This commit is contained in:
Johannes Gäßler
2025-07-23 21:43:25 +02:00
committed by GitHub
parent b284197df4
commit a86f52b285
8 changed files with 98 additions and 246 deletions
Download Patch File Download Diff File Expand all files Collapse all files

51
ggml/src/ggml-cuda/fattn-vec-f32.cuh
View File

@@ -18,33 +18,13 @@ static __global__ void flash_attn_vec_ext_f32(
const float m1,
const uint32_t n_head_log2,
const float logit_softcap,
const int ne00,
const int ne01,
const int ne02,
const int ne03,
const int ne10,
const int ne11,
const int ne12,
const int ne13,
const int ne31,
const int ne32,
const int ne33,
const int nb31,
const int nb32,
const int nb33,
const int nb01,
const int nb02,
const int nb03,
const int nb11,
const int nb12,
const int nb13,
const int nb21,
const int nb22,
const int nb23,
const int ne0,
const int ne1,
const int ne2,
const int ne3) {
const int32_t ne00, const int32_t 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) {
#ifdef FLASH_ATTN_AVAILABLE
// Skip unused kernel variants for faster compilation:
@@ -59,8 +39,7 @@ static __global__ void flash_attn_vec_ext_f32(
GGML_UNUSED(nb31); GGML_UNUSED(nb32); GGML_UNUSED(nb33); GGML_UNUSED(nb01); GGML_UNUSED(nb02);
GGML_UNUSED(nb03); GGML_UNUSED(nb11); GGML_UNUSED(nb12);
GGML_UNUSED(nb13); GGML_UNUSED(nb21); GGML_UNUSED(nb22);
GGML_UNUSED(nb23); GGML_UNUSED(ne0); GGML_UNUSED(ne1);
GGML_UNUSED(ne2); GGML_UNUSED(ne3);
GGML_UNUSED(nb23);
NO_DEVICE_CODE;
return;
}
@@ -198,13 +177,16 @@ static __global__ void flash_attn_vec_ext_f32(
float VKQ[ncols] = {0.0f};
K += blockIdx.y*D * nb11;
V += blockIdx.y*D * nb21;
maskh += blockIdx.y*D;
for (int k_VKQ_0 = blockIdx.y*D; k_VKQ_0 < ne11; k_VKQ_0 += gridDim.y*D) {
// Calculate KQ tile and keep track of new maximum KQ values:
if (mask) {
#pragma unroll
for (int j = 0; j < ncols; ++j) {
maskf_shared[j*D + tid] = slope*__half2float(maskh[j*ne11 + k_VKQ_0 + tid]);
maskf_shared[j*D + tid] = slope*__half2float(maskh[j*ne11 + tid]);
}
__syncthreads();
@@ -246,7 +228,7 @@ static __global__ void flash_attn_vec_ext_f32(
#pragma unroll
for (int j = 0; j < ncols; ++j) {
float sum = vec_dot_KQ(K + (k_VKQ_0 + i_KQ)*nb11, Q_f2[j], Q_i32[j], Q_ds[j]);
float sum = vec_dot_KQ(K + i_KQ*nb11, Q_f2[j], Q_i32[j], Q_ds[j]);
sum = warp_reduce_sum(sum);
if (use_logit_softcap) {
@@ -297,13 +279,17 @@ static __global__ void flash_attn_vec_ext_f32(
break;
}
const float V_ki = dequantize_1_v(V + (k_VKQ_0 + k)*nb21, tid);
const float V_ki = dequantize_1_v(V + k*nb21, tid);
#pragma unroll
for (int j = 0; j < ncols; ++j) {
VKQ[j] += V_ki*KQ[j*D + k];
}
}
K += gridDim.y*D * nb11;
V += gridDim.y*D * nb21;
maskh += gridDim.y*D;
__syncthreads();
}
@@ -348,7 +334,6 @@ static __global__ void flash_attn_vec_ext_f32(
GGML_UNUSED(nb01); GGML_UNUSED(nb02); GGML_UNUSED(nb03);
GGML_UNUSED(nb11); GGML_UNUSED(nb12); GGML_UNUSED(nb13);
GGML_UNUSED(nb21); GGML_UNUSED(nb22); GGML_UNUSED(nb23);
GGML_UNUSED(ne0); GGML_UNUSED(ne1); GGML_UNUSED(ne2); GGML_UNUSED(ne3);
NO_DEVICE_CODE;
#endif // FLASH_ATTN_AVAILABLE
}