[Perf][DSV4] Optimize sparse top-k metadata kernels for higher prefill throughput (#52084)
Signed-off-by: chaunceyjiang <chaunceyjiang@gmail.com>
Chauncey committed
Aug 16, 2026 at 21:15 UTC
836aac92ffdaa337083934181cb6d00b64b2a1a6
1 file changed
+1
-1
vllm/models/deepseek_v4/common/ops/cache_utils.py
+1
-1
@@ -580,7 +580,7 @@ def combine_topk_swa_indices(
580
return combined_indices, combined_lens
581
582
583
-_COMBINE_TOPK_SWA_NUM_WORKERS = 128
583
+_COMBINE_TOPK_SWA_NUM_WORKERS = 256
584
585
586
# Representative pointer alignment variants for Triton pointer specialization.