From 69cf055936066779cbd54c237739e9ac657ed86b Mon Sep 17 00:00:00 2001 From: aoshen02 Date: Fri, 4 Sep 2026 11:06:34 +0800 Subject: [PATCH] [Performance][DSv4] Size dequant gather launch grid by rows (#55061) Signed-off-by: aoshen02 Co-authored-by: aoshen02 --- .../deepseek_v4/nvidia/ops/dequant_gather_k_cutedsl.py | 7 ++++++- 1 file changed, 6 insertions(+), 1 deletion(-) diff --git a/vllm/models/deepseek_v4/nvidia/ops/dequant_gather_k_cutedsl.py b/vllm/models/deepseek_v4/nvidia/ops/dequant_gather_k_cutedsl.py index 6c44324645..a4514293c9 100644 --- a/vllm/models/deepseek_v4/nvidia/ops/dequant_gather_k_cutedsl.py +++ b/vllm/models/deepseek_v4/nvidia/ops/dequant_gather_k_cutedsl.py @@ -73,7 +73,12 @@ class DequantGatherKCacheKernel: ), ) - grid = (out.shape[0], 1024, 1) + num_reqs = out.shape[0] + max_rows = out.shape[1] - offset + num_workers = cutlass.max( + 1, cutlass.min(cute.ceil_div(max_rows, 4), cute.ceil_div(8192, num_reqs)) + ) + grid = (num_reqs, num_workers, 1) self.kernel( out, k_data,