[Perf] Update TRTLLM supported MoE routing methods (#39141)

Signed-off-by: wzhao18 <wzhao18.sz@gmail.com>
Signed-off-by: Wei Zhao <51183510+wzhao18@users.noreply.github.com>
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
Co-authored-by: root <root@bia0030.bia.clusters.nvidia.com>
Co-authored-by: root <root@bia0036.bia.clusters.nvidia.com>
This commit is contained in:
Wei Zhao
2026-04-27 14:16:22 -04:00
committed by GitHub
co-authored by gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com> root root
parent 6232fb4b66
commit c8bbe05189
5 changed files with 26 additions and 92 deletions
-12
View File
@@ -50,7 +50,6 @@ from vllm.model_executor.layers.attention_layer_base import AttentionLayerBase
from vllm.model_executor.layers.fused_moe import (
FusedMoE,
GateLinear,
RoutingMethodType,
fused_moe_make_expert_params_mapping,
)
from vllm.model_executor.layers.layernorm import LayerNorm, RMSNorm
@@ -338,17 +337,6 @@ class DeepseekV2MoE(nn.Module):
else None,
)
# NOTE(rob): this is a hack until we finish off the PR for
# merging TRTLLM kernels into the MK framework. Then we can
# query the MonolithicMK for the expected router logits.
# NOTE(dbari): Use BF16 if routing is not Deepseek, e.g. Mistral Large 3
self.gate.set_out_dtype(
torch.float32
if self.experts.quant_method.is_monolithic
and self.experts.routing_method_type == RoutingMethodType.DeepSeekV3
else torch.bfloat16
)
# Pre-cast the bias to match the gate output dtype so the
# conversion is not repeated on every forward pass. All
# downstream references (FusedMoE, router) share the same