A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 6, 23:00 UTC
[Bug] Fix TRTLLM Block FP8 MoE Monolithic (#36296)
Signed-off-by: wzhao18 <wzhao18.sz@gmail.com> Co-authored-by: Robert Shaw <114415538+robertgshaw2-redhat@users.noreply.github.com>
Wei Zhao committed
Mar 10, 2026 at 22:04 UTC
84e436ed1c94b1b94f809927b5d6bff45f7af919
1 file changed
-3
vllm/model_executor/layers/fused_moe/experts/trtllm_fp8_moe.py
-3
@@ -176,9 +176,6 @@ class TrtLlmFp8Experts(mk.FusedMoEExpertsMonolithic):
176
assert not apply_router_weight_on_input
177
assert activation == MoEActivation.SILU
178
179
- if e_score_correction_bias is not None:
180
- e_score_correction_bias = e_score_correction_bias.to(hidden_states.dtype)
181
-
179
if self.routing_method_type == RoutingMethodType.DeepSeekV3:
180
router_logits = router_logits.to(torch.float32)
181