public_git/whisper.cpp
1
0
Fork 0
mirror of https://github.com/ggml-org/whisper.cpp.git synced 2026-10-11 16:55:39 +02:00
Code Issues Packages Projects Releases Wiki Activity

vulkan: MOE aware mat_mul_id tile selection (llama/29182)

Browse Source 
mut_mul_id selected its matmul tile with total token count.
For MoE dispatch grid the true N per workgroup is per-expert rows.
At pp128 on Sarvam 30B that is 6, not 128, so the picker took the l-tile for ~6 live rows.
Most workers in each group had nothing to do.
This wasted time. The slow part was 55% of the whole job.
This commit is contained in:
Ankit Khandelwal
2026-10-06 10:38:09 +03:00
committed by Georgi Gerganov
parent deae6f3439
commit 8cc3ca48c3
1 changed files with 4 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-vulkan/ggml-vulkan.cpp
+4 -3
View File
@@ -7407,10 +7407,11 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context&
GGML_ASSERT(mmp_map != nullptr);
const uint32_t kpad = quantize_y ? 0 : ggml_vk_align_size(ne10, ggml_vk_guess_matmul_pipeline_align_map(ctx, *mmp_map, ne01, nei1, true));
const bool aligned = !quantize_y && ne10 == kpad && ne01 > 8 && nei1 > 8;
const uint32_t n_per_expert = (uint32_t)CEIL_DIV(nei0 * nei1, n_as);
const uint32_t kpad = quantize_y ? 0 : ggml_vk_align_size(ne10, ggml_vk_guess_matmul_pipeline_align_map(ctx, *mmp_map, ne01, n_per_expert, true));
const bool aligned = !quantize_y && ne10 == kpad && ne01 > 8 && n_per_expert > 8;
vk_pipeline pipeline = ggml_vk_guess_matmul_pipeline_map(ctx, *mmp_map, ne01, nei1, aligned, true);
vk_pipeline pipeline = ggml_vk_guess_matmul_pipeline_map(ctx, *mmp_map, ne01, n_per_expert, aligned, true);
if (ggml_nbytes(src0) > ctx->device->properties.limits.maxStorageBufferRange) {
pipeline = ggml_vk_get_64b_indexing_pipeline(ctx, pipeline);