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 15, 21:00 UTC
[Bugfix] Fix MoeWNA16Method activation (#14024)
Jee Jee Li committed
Feb 28, 2025 at 23:22 UTC
e0734387fb4a04067f557761a6e795ae3ab474d3
1 file changed
+2
-1
vllm/model_executor/layers/quantization/moe_wna16.py
+2
-1
@@ -293,9 +293,10 @@ class MoeWNA16Method(FusedMoEMethodBase):
293
custom_routing_function: Optional[Callable] = None,
294
scoring_func: str = "softmax",
295
e_score_correction_bias: Optional[torch.Tensor] = None,
296
+ activation: str = "silu",
297
) -> torch.Tensor:
298
from vllm.model_executor.layers.fused_moe import fused_experts
298
-
299
+ assert activation == "silu", "Only SiLU activation is supported."
300
topk_weights, topk_ids = FusedMoE.select_experts(
301
hidden_states=x,
302
router_logits=router_logits,