sdgoij/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2026-05-15 13:34:06 +00:00
Code Issues Packages Projects Releases Wiki Activity

flush the gpu profile timestamp before the queryset is overflowed (#22995)

Browse Source
This commit is contained in:
Masashi Yoshimura
2026-05-14 02:22:44 +09:00
committed by GitHub
parent 2dfeca31cc
commit 527045bfb0
1 changed files with 10 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
ggml/src/ggml-webgpu/ggml-webgpu.cpp
View File

@@ -3148,6 +3148,16 @@ static ggml_status ggml_backend_webgpu_graph_compute(ggml_backend_t backend, str
}
ctx->param_arena.reset();
commands.clear();
#ifdef GGML_WEBGPU_GPU_PROFILE
// flush before the next batch can overflow the QuerySet
if (ctx->profile_timestamp_query_count + 2 * ctx->global_ctx->command_submit_batch_size >=
WEBGPU_MAX_PROFILE_QUERY_COUNT) {
ggml_backend_webgpu_collect_profile_results(ctx, profile_pipeline_names, num_inflight_batches);
// reset profile timestamp state
ctx->profile_timestamp_query_count = 0;
profile_pipeline_names.clear();
}
#endif
}
node_idx += num_encoded_ops;