From 1ee7f78e806b5e9b7476c74edf453b58fe108157 Mon Sep 17 00:00:00 2001 From: Misha Goin Date: Tue, 29 Sep 2026 17:23:27 -0400 Subject: [PATCH] [Perf] Reduce redundant Triton sampler warmup specializations (#58605) Signed-off-by: mgoin Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com> --- vllm/v1/sample/ops/topk_topp_triton.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/vllm/v1/sample/ops/topk_topp_triton.py b/vllm/v1/sample/ops/topk_topp_triton.py index 20b636e745..151f641d21 100755 --- a/vllm/v1/sample/ops/topk_topp_triton.py +++ b/vllm/v1/sample/ops/topk_topp_triton.py @@ -103,7 +103,7 @@ def _update_min_larger_stats(data, above_mask, min_larger, num_min_larger, senti return min_larger, num_min_larger -@triton.jit +@triton.jit(do_not_specialize_on_alignment=["BATCH_SIZE"]) def _topk_topp_kernel( LOGITS, LOGITS_STRIDE_0, @@ -1225,7 +1225,7 @@ def _topp_sb_combine( ) -@triton.jit +@triton.jit(do_not_specialize_on_alignment=["ROUND"]) def _topp_sb_step_kernel( LOGITS, LOGITS_STRIDE_0,