karylab_agents/vllm
1
0
Fork 0
forked from Karylab-cklius/vllm
Code Pull Requests 2 Actions 1 Packages Activity

[Bugfix][Kernel] Correct FlashInfer CUTLASS MoE tuning token bound (#46838)

Browse Source 
Signed-off-by: Haobin Guo <haobing@nvidia.com>
This commit is contained in:
Hiki
2026-07-02 05:11:00 +00:00
committed by GitHub
parent 09663abde0
commit 2665ed704b
1 changed files with 0 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
vllm/model_executor/layers/fused_moe/experts/flashinfer_cutlass_moe.py
-2
View File
@@ -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: