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 5, 13:30 UTC
[Bugfix] Re-enable Renormalize routing for TRT-LLM MoE experts (#38859)
Signed-off-by: Yifan Zong <yzong@redhat.com>
yzong-rh committed
Apr 3, 2026 at 13:48 UTC
a5a623d9616337ed53e8fa0e1376e2d17b505e38
2 files changed
+7
-12
vllm/model_executor/layers/fused_moe/experts/trtllm_bf16_moe.py
+2
-5
@@ -79,11 +79,8 @@ class TrtLlmBf16Experts(mk.FusedMoEExpertsMonolithic):
79
RoutingMethodType.Default,
80
RoutingMethodType.DeepSeekV3,
81
RoutingMethodType.Llama4,
82
- # NOTE: TRTLLM Kernel has issue with Qwen3.5 router.
83
- # Re-enable once the issue is resolved.
84
- # https://github.com/vllm-project/vllm/issues/37591
85
- # RoutingMethodType.Renormalize,
86
- # RoutingMethodType.RenormalizeNaive
82
+ RoutingMethodType.Renormalize,
83
+ RoutingMethodType.RenormalizeNaive,
84
]
85
86
@staticmethod
vllm/model_executor/layers/fused_moe/experts/trtllm_fp8_moe.py
+5
-7
@@ -277,13 +277,7 @@ class TrtLlmFp8ExpertsMonolithic(TrtLlmFp8ExpertsBase, mk.FusedMoEExpertsMonolit
277
weight_key: QuantKey | None,
278
activation_key: QuantKey | None,
279
) -> bool:
280
- """Monolithic kernels need to express router support.
281
- Renormalize/RenormalizeNaive are excluded: the monolithic kernel's
282
- internal routing for these methods produces output uncorrelated
283
- with the modular kernel's output and with Triton kernel's output
284
- for Qwen3.5-35B-A3B-FP8.
285
- See: https://github.com/vllm-project/vllm/issues/37591
286
- """
280
+ """Monolithic kernels need to express router support."""
281
# NOTE(dbari): TopK routing could also be enabled, but need to validate models
282
# NOTE(dbari): Default is not implemented and should not be enabled until it is
283
@@ -295,6 +289,8 @@ class TrtLlmFp8ExpertsMonolithic(TrtLlmFp8ExpertsBase, mk.FusedMoEExpertsMonolit
289
return routing_method in [
290
RoutingMethodType.DeepSeekV3,
291
RoutingMethodType.Simulated,
292
+ RoutingMethodType.Renormalize,
293
+ RoutingMethodType.RenormalizeNaive,
294
]
295
elif (weight_key, activation_key) == (kFp8StaticTensorSym, kFp8StaticTensorSym):
296
# NOTE(dbari): as above, potentially allow others here.
@@ -302,6 +298,8 @@ class TrtLlmFp8ExpertsMonolithic(TrtLlmFp8ExpertsBase, mk.FusedMoEExpertsMonolit
298
RoutingMethodType.DeepSeekV3,
299
RoutingMethodType.Llama4,
300
RoutingMethodType.Simulated,
301
+ RoutingMethodType.Renormalize,
302
+ RoutingMethodType.RenormalizeNaive,
303
]
304
else:
305
raise ValueError("Unsupported quantization scheme.")