LLM-Inference/whisper.cpp
1
0
Fork 0
mirror of https://github.com/ggml-org/whisper.cpp.git synced 2026-08-12 22:31:22 +04:00
Code Issues Packages Projects Releases Wiki Activity

vulkan: improve topk perf for large k, fix overflow in unit tests (llama/17582)

Browse Source
This commit is contained in:
Jeff Bolz
2025-12-12 17:53:12 +02:00
committed by Georgi Gerganov
parent 463003e76c
commit dbf8766ffa
1 changed files with 3 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-vulkan/ggml-vulkan.cpp
+3 -1
View File
@@ -10239,7 +10239,9 @@ static void ggml_vk_topk(ggml_backend_vk_context * ctx, vk_context& subctx, cons
// Prefer going as small as num_topk_pipelines - 3 for perf reasons.
// But if K is larger, then we need a larger workgroup
uint32_t max_pipeline = num_topk_pipelines - 3;
uint32_t max_pipeline = num_topk_pipelines - 1;
uint32_t preferred_pipeline = std::max(num_topk_pipelines - 3, (uint32_t)log2f(float(k)) + 2);
max_pipeline = std::min(preferred_pipeline, max_pipeline);
uint32_t min_pipeline = (uint32_t)log2f(float(k)) + 1;
// require full subgroup
min_pipeline = std::max(min_pipeline, ctx->device->subgroup_size_log2);