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 1, 20:00 UTC
[Cohere] fix RoutingMethodType (#44021)
Signed-off-by: Terrencezzj <terrence@cohere.ai>
Terrence Zhao committed
Jun 5, 2026 at 19:25 UTC
a50e675b0d864660b10bfdc40f3b0224695d0a92
3 files changed
+7
-3
vllm/model_executor/layers/fused_moe/experts/trtllm_bf16_moe.py
+2
@@ -80,6 +80,8 @@ class TrtLlmBf16Experts(mk.FusedMoEExpertsMonolithic):
80
RoutingMethodType.Llama4,
81
RoutingMethodType.Renormalize,
82
RoutingMethodType.RenormalizeNaive,
83
+ RoutingMethodType.SigmoidRenorm,
84
+ RoutingMethodType.Sigmoid,
85
]
86
87
@staticmethod
vllm/model_executor/layers/fused_moe/experts/trtllm_nvfp4_moe.py
+1
-1
@@ -350,9 +350,9 @@ class TrtLlmNvFp4ExpertsMonolithic(
350
RoutingMethodType.RenormalizeNaive,
351
RoutingMethodType.Llama4,
352
RoutingMethodType.SigmoidRenorm,
353
+ RoutingMethodType.Sigmoid,
354
RoutingMethodType.MiniMax2,
355
RoutingMethodType.Simulated,
355
- RoutingMethodType.SigmoidRenorm,
356
]
357
358
@staticmethod
vllm/model_executor/layers/fused_moe/router/custom_routing_router.py
+4
-2
@@ -38,9 +38,11 @@ class CustomRoutingRouter(BaseRouter):
38
# NOTE: FLASHINFER_TRTLLM support the Llama4 router.
39
if self.custom_routing_function == Llama4MoE.custom_routing_function:
40
return RoutingMethodType.Llama4
41
- # Cohere MoE uses a sigmoid -> top-k -> renormalize routing function.
41
+ # Cohere MoE uses sigmoid -> top-k, optionally followed by renormalize.
42
if self.custom_routing_function == token_choice_with_bias:
43
- return RoutingMethodType.SigmoidRenorm
43
+ if self.renormalize:
44
+ return RoutingMethodType.SigmoidRenorm
45
+ return RoutingMethodType.Sigmoid
46
return RoutingMethodType.Custom
47
48
def _compute_routing(