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 9, 01:15 UTC
[Bugfix] Disable RoutingMethodType.[Renormalize,RenormalizeNaive] TRTLLM per-tensor FP8 MoE (#33620)
Signed-off-by: mgoin <mgoin64@gmail.com>
Michael Goin committed
Feb 3, 2026 at 05:37 UTC
e346e2d056a66bb84287e4fea049bde9a37bd72b
1 file changed
+4
-2
vllm/model_executor/layers/fused_moe/flashinfer_trtllm_moe.py
+4
-2
@@ -72,8 +72,10 @@ def _supports_routing_method(
72
# NOTE(dbari): as above, potentially allow others here.
73
return routing_method in [
74
RoutingMethodType.Llama4,
75
- RoutingMethodType.Renormalize,
76
- RoutingMethodType.RenormalizeNaive,
75
+ # NOTE(mgoin): Disabled to investigate accuracy issues.
76
+ # See https://github.com/vllm-project/vllm/issues/33532
77
+ # RoutingMethodType.Renormalize,
78
+ # RoutingMethodType.RenormalizeNaive,
79
]
80
else:
81
raise ValueError("Unsupported quantization scheme.")