From 836aac92ffdaa337083934181cb6d00b64b2a1a6 Mon Sep 17 00:00:00 2001 From: Chauncey Date: Sun, 16 Aug 2026 21:15:47 +0800 Subject: [PATCH] [Perf][DSV4] Optimize sparse top-k metadata kernels for higher prefill throughput (#52084) Signed-off-by: chaunceyjiang --- vllm/models/deepseek_v4/common/ops/cache_utils.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vllm/models/deepseek_v4/common/ops/cache_utils.py b/vllm/models/deepseek_v4/common/ops/cache_utils.py index 811bf21d49..2d085f88fa 100644 --- a/vllm/models/deepseek_v4/common/ops/cache_utils.py +++ b/vllm/models/deepseek_v4/common/ops/cache_utils.py @@ -580,7 +580,7 @@ def combine_topk_swa_indices( return combined_indices, combined_lens -_COMBINE_TOPK_SWA_NUM_WORKERS = 128 +_COMBINE_TOPK_SWA_NUM_WORKERS = 256 # Representative pointer alignment variants for Triton pointer specialization.