Revert "[Bugfix][MoE] Tune FlashInfer experts to scheduler token limit" (#52989) (#53186)

Co-authored-by: vllm-ci-failure-analyzer <kevin@inferact.ai>
This commit is contained in:
vllm-agent
2026-08-21 12:39:28 -06:00
committed by GitHub
co-authored by vllm-ci-failure-analyzer
parent 7a2fdbaac4
commit ba53da60bb
6 changed files with 1 additions and 12 deletions
@@ -191,7 +191,6 @@ def flashinfer_trtllm_mxint4_moe(
e_score_correction_bias: torch.Tensor | None = None,
routing_method_type: int | None = None,
routing_replay_out: torch.Tensor | None = None,
tune_max_num_tokens: int = 8192,
) -> torch.Tensor:
"""
Apply FlashInfer TensorRT-LLM MxInt4 MoE kernel.
@@ -212,7 +211,6 @@ def flashinfer_trtllm_mxint4_moe(
topk_group: Top-k within groups (default: None -> 0)
e_score_correction_bias: Optional routing bias. dtype: bfloat16
routing_method_type: FlashInfer RoutingMethodType enum value
tune_max_num_tokens: Maximum token count covered by autotuning.
Returns:
Output tensor from MoE layer. dtype: same as x (bfloat16)
@@ -264,7 +262,7 @@ def flashinfer_trtllm_mxint4_moe(
enable_pdl=None,
do_finalize=True,
output=None,
tune_max_num_tokens=tune_max_num_tokens,
tune_max_num_tokens=8192,
routing_replay_out=routing_replay_out,
)
if isinstance(out, (tuple, list)):