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, 05:15 UTC
[Bugfix] Disable FlashInfer TRTLLM BF16 path for non-gated MoE (#36146)
Signed-off-by: Tomer Asida <57313761+tomeras91@users.noreply.github.com>
tomeras91 committed
Mar 5, 2026 at 19:49 UTC
86483ca7749b3d7a2ae16283a7896c203983f1ef
1 file changed
+3
-3
vllm/model_executor/layers/fused_moe/flashinfer_trtllm_moe.py
+3
-3
@@ -25,12 +25,12 @@ def _supports_current_device() -> bool:
25
26
27
def _supports_no_act_and_mul() -> bool:
28
- """Supports non-gated MoE."""
29
- return True
28
+ """BF16 kernels do not support non-gated MoE"""
29
+ return False
30
31
32
def _supports_activation(activation: MoEActivation) -> bool:
33
- return activation in [MoEActivation.SILU, MoEActivation.RELU2_NO_MUL]
33
+ return activation in [MoEActivation.SILU]
34
35
36
def _supports_routing_method_bf16(