[Bugfix][Kernel] Correct FlashInfer CUTLASS MoE tuning token bound (#46838)
Signed-off-by: Haobin Guo <haobing@nvidia.com>
Hiki committed
Jul 2, 2026 at 13:11 UTC
2665ed704b04219dd67a6bb82636cd51bbe98183
1 file changed
-2
vllm/model_executor/layers/fused_moe/experts/flashinfer_cutlass_moe.py
-2
@@ -93,7 +93,6 @@ class FlashInferExperts(mk.FusedMoEExpertsModular):
93
# - pass per-block weight scales to the kernel
94
# - skip input activation quantization (kernel applies scaling)
95
self.use_deepseek_fp8_block_scale = quant_config.is_block_quantized
96
- self.max_capture_size = moe_config.max_capture_size
96
self.gemm1_clamp_limit: torch.Tensor | None = None
97
if quant_config.gemm1_clamp_limit is not None:
98
self.gemm1_clamp_limit = torch.tensor(
@@ -398,7 +397,6 @@ class FlashInferExperts(mk.FusedMoEExpertsModular):
397
use_deepseek_fp8_block_scale=self.use_deepseek_fp8_block_scale,
398
use_mxfp8_act_scaling=use_mxfp8_act_scaling,
399
use_w4_group_scaling=use_w4_group_scaling,
401
- tune_max_num_tokens=max(self.max_capture_size, 1),
400
)
401
402
def moe_sum(self, input: torch.Tensor, output: torch.Tensor) -> None: