From 2665ed704b04219dd67a6bb82636cd51bbe98183 Mon Sep 17 00:00:00 2001 From: Hiki Date: Thu, 2 Jul 2026 13:11:00 +0800 Subject: [PATCH] [Bugfix][Kernel] Correct FlashInfer CUTLASS MoE tuning token bound (#46838) Signed-off-by: Haobin Guo --- .../layers/fused_moe/experts/flashinfer_cutlass_moe.py | 2 -- 1 file changed, 2 deletions(-) diff --git a/vllm/model_executor/layers/fused_moe/experts/flashinfer_cutlass_moe.py b/vllm/model_executor/layers/fused_moe/experts/flashinfer_cutlass_moe.py index 76cd15ff5a0..2b8657723bd 100644 --- a/vllm/model_executor/layers/fused_moe/experts/flashinfer_cutlass_moe.py +++ b/vllm/model_executor/layers/fused_moe/experts/flashinfer_cutlass_moe.py @@ -93,7 +93,6 @@ class FlashInferExperts(mk.FusedMoEExpertsModular): # - pass per-block weight scales to the kernel # - skip input activation quantization (kernel applies scaling) self.use_deepseek_fp8_block_scale = quant_config.is_block_quantized - self.max_capture_size = moe_config.max_capture_size self.gemm1_clamp_limit: torch.Tensor | None = None if quant_config.gemm1_clamp_limit is not None: self.gemm1_clamp_limit = torch.tensor( @@ -398,7 +397,6 @@ class FlashInferExperts(mk.FusedMoEExpertsModular): use_deepseek_fp8_block_scale=self.use_deepseek_fp8_block_scale, use_mxfp8_act_scaling=use_mxfp8_act_scaling, use_w4_group_scaling=use_w4_group_scaling, - tune_max_num_tokens=max(self.max_capture_size, 1), ) def moe_sum(self, input: torch.Tensor, output: torch.Tensor) -> None: