[Perf][DSV4] Optimize sparse top-k metadata kernels for higher prefill throughput (vllm-project#52084)

Signed-off-by: chaunceyjiang <chaunceyjiang@gmail.com>
Signed-off-by: Wyett <wyettzeng@gmail.com>

Author: chaunceyjiang

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.