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

Add flash attention MMA / Tiles to support MiMo-V2.5 (#22812)

Browse Source 
* mimo-v2.5: add flash attention mma/tiles for for d_kq=192 d_v=128

* mimo-v2.5: follow (256, 256) fattn templates

* mimo-v2.5: cleanup comments

* mimo-v2.5: further comment cleanup

* mimo-v2.5: address PR feedback
fix GQA handling
check for other dangling 320/576 carveouts and mirror them for 192
Add to backend ops test so new paths are covered
This commit is contained in:
AesSedai
2026-05-09 11:28:29 +08:00
committed by GitHub
parent 66001722aa
commit 046e284437
14 changed files with 105 additions and 10 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-cuda/fattn.cu
+29 -4
View File
@@ -139,6 +139,22 @@ static void ggml_cuda_flash_attn_ext_mma_f16(ggml_backend_cuda_context & ctx, gg
GGML_ASSERT(V->ne[0] == 128);
ggml_cuda_flash_attn_ext_mma_f16_switch_ncols2<128, 128>(ctx, dst);
break;
case 192: {
// MiMo-V2.5 / V2.5-Pro / V2-Flash: gqa_ratio is 8 (SWA) or 16 (full attn)
GGML_ASSERT(V->ne[0] == 128);
float max_bias = 0.0f;
memcpy(&max_bias, (const float *) KQV->op_params + 1, sizeof(float));
const bool use_gqa_opt = mask && max_bias == 0.0f;
GGML_ASSERT(use_gqa_opt);
GGML_ASSERT(Q->ne[2] % K->ne[2] == 0);
const int gqa_ratio = Q->ne[2] / K->ne[2];
if (gqa_ratio % 16 == 0) {
ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<192, 128, 16>(ctx, dst);
} else {
GGML_ASSERT(gqa_ratio % 8 == 0);
ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<192, 128, 8>(ctx, dst);
}
} break;
case 256:
GGML_ASSERT(V->ne[0] == 256);
ggml_cuda_flash_attn_ext_mma_f16_switch_ncols2<256, 256>(ctx, dst);
@@ -368,6 +384,14 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const
return BEST_FATTN_KERNEL_NONE;
}
break;
case 192:
if (V->ne[0] != 128 || !gqa_opt_applies) {
return BEST_FATTN_KERNEL_NONE;
}
if (gqa_ratio % 8 != 0) {
return BEST_FATTN_KERNEL_NONE;
}
break;
case 320:
if (V->ne[0] != 256 || !gqa_opt_applies) {
return BEST_FATTN_KERNEL_NONE;
@@ -425,7 +449,8 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const
}
// For small batch sizes the vector kernel may be preferable over the kernels optimized for large batch sizes:
const bool can_use_vector_kernel = Q->ne[0] <= 256 && Q->ne[0] % 64 == 0 && K->ne[1] % FATTN_KQ_STRIDE == 0;
// 192 satisfies % 64 == 0 but has no vec instance (DKQ != DV); force it onto the MMA path.
const bool can_use_vector_kernel = Q->ne[0] <= 256 && Q->ne[0] % 64 == 0 && Q->ne[0] != 192 && K->ne[1] % FATTN_KQ_STRIDE == 0;
// If Turing tensor cores are available, use them:
if (turing_mma_available(cc) && Q->ne[0] != 40 && Q->ne[0] != 72) {
@@ -454,7 +479,7 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const
if (volta_mma_available(cc) && Q->ne[0] != 40 && Q->ne[0] != 72) {
int gqa_ratio_eff = 1;
const int ncols2_max = Q->ne[0] == 576 ? 16 : 8;
const int ncols2_max = (Q->ne[0] == 576 || Q->ne[0] == 192) ? 16 : 8;
while (gqa_ratio % (2*gqa_ratio_eff) == 0 && gqa_ratio_eff < ncols2_max) {
gqa_ratio_eff *= 2;
}
@@ -468,7 +493,7 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const
}
// Use the WMMA kernel if possible:
if (ggml_cuda_should_use_wmma_fattn(cc) && K->ne[1] % FATTN_KQ_STRIDE == 0 && Q->ne[0] != 40 && Q->ne[0] != 72 && Q->ne[0] != 512 && Q->ne[0] != 576) {
if (ggml_cuda_should_use_wmma_fattn(cc) && K->ne[1] % FATTN_KQ_STRIDE == 0 && Q->ne[0] != 40 && Q->ne[0] != 72 && Q->ne[0] != 192 && Q->ne[0] != 512 && Q->ne[0] != 576) {
if (can_use_vector_kernel && Q->ne[1] <= 2) {
return BEST_FATTN_KERNEL_VEC;
}
@@ -501,7 +526,7 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const
}
// Use MFMA flash attention for CDNA (MI100+):
if (amd_mfma_available(cc) && Q->ne[0] != 40 && Q->ne[0] != 72 && Q->ne[0] != 256 && Q->ne[0] != 512 && Q->ne[0] != 576) {
if (amd_mfma_available(cc) && Q->ne[0] != 40 && Q->ne[0] != 72 && Q->ne[0] != 192 && Q->ne[0] != 256 && Q->ne[0] != 512 && Q->ne[0] != 576) {
const int64_t eff_nq = Q->ne[1] * (gqa_opt_applies ? gqa_ratio : 1);
// MMA vs tile crossover benchmarked on MI300X @ d32768:
// hsk=64 (gqa=4): MMA wins at eff >= 128 (+11%)