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 7, 06:30 UTC
[Bugfix] Fix passing of activation_type to trtllm fused MoE NVFP4 and FP8 (#36017)
Signed-off-by: amitz-nv <203509407+amitz-nv@users.noreply.github.com>
amitz-nv committed
Mar 5, 2026 at 00:23 UTC
d7adcadb9bf4c7ea240fcc6cc668192bc2260ec0
2 files changed
+2
-2
vllm/model_executor/layers/fused_moe/experts/trtllm_fp8_moe.py
+1
-2
@@ -240,12 +240,11 @@ class TrtLlmFp8Experts(mk.FusedMoEExpertsMonolithic):
240
) -> torch.Tensor:
241
# Delay import for non-CUDA.
242
import flashinfer
243
- from flashinfer.fused_moe.core import ActivationType
243
244
# Confirm supported activation function.
245
assert activation in [MoEActivation.SILU, MoEActivation.RELU2_NO_MUL]
246
248
- activation_type = ActivationType(activation_to_flashinfer_int(activation))
247
+ activation_type = activation_to_flashinfer_int(activation)
248
249
# Confirm Llama-4 routing is proper.
250
if self.routing_method_type == RoutingMethodType.Llama4:
vllm/model_executor/layers/fused_moe/experts/trtllm_nvfp4_moe.py
+1
@@ -323,4 +323,5 @@ class TrtLlmNvFp4ExpertsMonolithic(
323
routed_scaling_factor=routed_scaling_factor,
324
routing_method_type=self.routing_method_type,
325
do_finalize=True,
326
+ activation_type=activation_to_flashinfer_int(activation),
327
)[0]