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-vec.cuh
+9 -9
View File
@@ -33,7 +33,7 @@ static __global__ void flash_attn_ext_vec(
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,
@@ -150,7 +150,7 @@ static __global__ void flash_attn_ext_vec(
float2 * tmp_q_ds = (float2 *) (tmp_q_i32 + D/sizeof(int));
// Set memory to zero if out of bounds:
if (ncols > 1 && ic0 + j >= ne01) {
if (ncols > 1 && ic0 + j >= int(ne01.z)) {
#pragma unroll
for (int i0 = 0; i0 < int(D/sizeof(int)); i0 += WARP_SIZE) {
const int i = i0 + threadIdx.x;
@@ -201,7 +201,7 @@ static __global__ void flash_attn_ext_vec(
const int i = i0 + (nthreads_KQ == WARP_SIZE ? threadIdx.x : threadIdx.x % nthreads_KQ)*cpy_ne;
float2 tmp[cpy_ne] = {{0.0f, 0.0f}};
if (ncols == 1 || ic0 + j < ne01) {
if (ncols == 1 || ic0 + j < int(ne01.z)) {
ggml_cuda_memcpy_1<cpy_nb>(tmp, &Q_j[i]);
ggml_cuda_memcpy_1<cpy_nb>(tmp + cpy_ne/2, &Q_j[i + cpy_ne/2]);
}
@@ -222,7 +222,7 @@ static __global__ void flash_attn_ext_vec(
#pragma unroll
for (int i0 = 0; i0 < D/2; i0 += nthreads_KQ*cpy_ne) {
const int i = i0 + (nthreads_KQ == WARP_SIZE ? threadIdx.x : threadIdx.x % nthreads_KQ)*cpy_ne;
if (ncols == 1 || ic0 + j < ne01) {
if (ncols == 1 || ic0 + j < int(ne01.z)) {
ggml_cuda_memcpy_1<cpy_nb>(&Q_reg[j][i0/nthreads_KQ], &Q_j[i]);
ggml_cuda_memcpy_1<cpy_nb>(&Q_reg[j][i0/nthreads_KQ + cpy_ne/2], &Q_j[i + cpy_ne/2]);
}
@@ -266,7 +266,7 @@ static __global__ void flash_attn_ext_vec(
sum = logit_softcap*tanhf(sum);
}
if (mask) {
if (mask && (ncols == 1 || ic0 + j < int(ne01.z))) {
sum += slope*__half2float(maskh[j*ne11 + i_KQ]);
}
@@ -412,7 +412,7 @@ static __global__ void flash_attn_ext_vec(
#pragma unroll
for (int j_VKQ = 0; j_VKQ < ncols; ++j_VKQ) {
if (ncols > 1 && ic0 + j_VKQ >= ne01) {
if (ncols > 1 && ic0 + j_VKQ >= int(ne01.z)) {
break;
}
@@ -479,7 +479,7 @@ static __global__ void flash_attn_ext_vec(
if (gridDim.y == 1) {
dst_val /= KQ_sum[j_VKQ];
}
dst[(((sequence*ne01 + ic0 + j_VKQ)*ne02 + head)*gridDim.y + blockIdx.y)*D + i0 + tid] = dst_val;
dst[(((sequence*int(ne01.z) + ic0 + j_VKQ)*ne02 + head)*gridDim.y + blockIdx.y)*D + i0 + tid] = dst_val;
}
}
@@ -489,8 +489,8 @@ static __global__ void flash_attn_ext_vec(
}
if (gridDim.y != 1 && tid < ncols && (ncols == 1 || ic0 + tid < ne01)) {
dst_meta[((sequence*ne01 + ic0 + tid)*ne02 + head)*gridDim.y + blockIdx.y] = make_float2(KQ_max[tid], KQ_sum[tid]);
if (gridDim.y != 1 && tid < ncols && (ncols == 1 || ic0 + tid < int(ne01.z))) {
dst_meta[((sequence*int(ne01.z) + ic0 + tid)*ne02 + head)*gridDim.y + blockIdx.y] = make_float2(KQ_max[tid], KQ_sum[tid]);
}
#else
GGML_UNUSED_VARS(Q, K, V, mask, sinks, KV_max, dst, dst_meta, scale,