[MoE] Align TRTLLM MXFP4 autotune buckets (#50942)

Signed-off-by: Thien Tran <gau.nernst@yahoo.com.sg>
Co-authored-by: OpenAI Codex <noreply@openai.com>
This commit is contained in:
Thien Tran
2026-08-05 18:17:53 -07:00
committed by GitHub
co-authored by OpenAI Codex
parent b50fdebce0
commit 9f3169960a
@@ -14,7 +14,10 @@ from vllm.model_executor.layers.fused_moe.config import (
from vllm.model_executor.layers.fused_moe.topk_weight_and_reduce import (
TopKWeightAndReduceNoOP,
)
from vllm.model_executor.layers.fused_moe.utils import trtllm_moe_pack_topk_ids_weights
from vllm.model_executor.layers.fused_moe.utils import (
fi_moe_largest_bucket,
trtllm_moe_pack_topk_ids_weights,
)
from vllm.model_executor.layers.quantization.utils.flashinfer_utils import (
activation_to_flashinfer_int,
has_flashinfer_situ_activation,
@@ -113,8 +116,6 @@ class TrtLlmMxfp4ExpertsBase:
)
self.gemm1_clamp_limit = None
self.max_capture_size = moe_config.max_capture_size
@staticmethod
def _supports_current_device() -> bool:
p = current_platform
@@ -263,7 +264,7 @@ class TrtLlmMxfp4ExpertsMonolithic(
routing_method_type=self.routing_method_type,
do_finalize=True,
activation_type=self._flashinfer_activation_type(activation),
tune_max_num_tokens=max(self.max_capture_size, 1),
tune_max_num_tokens=fi_moe_largest_bucket(self.moe_config),
output=output,
routing_replay_out=routing_replay_out,
)
@@ -382,7 +383,7 @@ class TrtLlmMxfp4ExpertsModular(TrtLlmMxfp4ExpertsBase, mk.FusedMoEExpertsModula
enable_pdl=True,
activation_type=self._flashinfer_activation_type(activation),
output=output,
tune_max_num_tokens=max(self.max_capture_size, 1),
tune_max_num_tokens=fi_moe_largest_bucket(self.moe_config),
)
def apply(