[Bugfix] Disable FlashInfer TRTLLM BF16 path for non-gated MoE (#36146)

Signed-off-by: Tomer Asida <57313761+tomeras91@users.noreply.github.com>
2026-03-05 19:49:05 +02:00
parent b93a9e6f6d
commit 86483ca774
1 changed files with 3 additions and 3 deletions
--- a/vllm/model_executor/layers/fused_moe/flashinfer_trtllm_moe.py
+++ b/vllm/model_executor/layers/fused_moe/flashinfer_trtllm_moe.py
@@ -25,12 +25,12 @@ def _supports_current_device() -> bool:


 def _supports_no_act_and_mul() -> bool:
-    """Supports non-gated MoE."""
-    return True
+    """BF16 kernels do not support non-gated MoE"""
+    return False


 def _supports_activation(activation: MoEActivation) -> bool:
-    return activation in [MoEActivation.SILU, MoEActivation.RELU2_NO_MUL]
+    return activation in [MoEActivation.SILU]


 def _supports_routing_method_bf16(