From 3f611f6106f11cde5ff48b0f8a6dd317241a16f4 Mon Sep 17 00:00:00 2001 From: Wentao Ye <44945378+yewentao256@users.noreply.github.com> Date: Wed, 13 May 2026 15:37:26 -0400 Subject: [PATCH] [CI] Fix pre-commit issue (#42563) Signed-off-by: yewentao256 --- .../layers/quantization/quark/quark_moe.py | 19 ++++++++++--------- 1 file changed, 10 insertions(+), 9 deletions(-) diff --git a/vllm/model_executor/layers/quantization/quark/quark_moe.py b/vllm/model_executor/layers/quantization/quark/quark_moe.py index 212ed5e83d..2dd38c07fe 100644 --- a/vllm/model_executor/layers/quantization/quark/quark_moe.py +++ b/vllm/model_executor/layers/quantization/quark/quark_moe.py @@ -1644,7 +1644,7 @@ class QuarkNvfp4MoEMethod(QuarkMoEMethod): layer.register_parameter("w2_input_scale_2", w2_input_scale_2) set_weight_attrs(w2_input_scale_2, extra_weight_attrs) - def process_weights_after_loading(self, layer: FusedMoE) -> None: + def process_weights_after_loading(self, layer: RoutedExperts) -> None: """ Convert NVFP4 MoE weights into kernel format and setup the kernel. """ @@ -1698,12 +1698,11 @@ class QuarkNvfp4MoEMethod(QuarkMoEMethod): self.moe_quant_config = self.get_fused_moe_quant_config(layer) if self.moe_quant_config: assert self.experts_cls is not None - self.moe_mk = make_nvfp4_moe_kernel( + self.moe_kernel = make_nvfp4_moe_kernel( moe_quant_config=self.moe_quant_config, moe_config=self.moe, experts_cls=self.experts_cls, - shared_experts=layer.shared_experts, - routing_tables=layer._maybe_init_expert_routing_tables(), + routing_tables=layer._expert_routing_tables(), ) def get_fused_moe_quant_config( @@ -1721,14 +1720,15 @@ class QuarkNvfp4MoEMethod(QuarkMoEMethod): def apply( self, - layer: FusedMoE, + layer: RoutedExperts, x: torch.Tensor, topk_weights: torch.Tensor, topk_ids: torch.Tensor, - shared_experts_input: Any | None, - ) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]: - assert self.moe_mk is not None - return self.moe_mk.apply( + shared_experts: SharedExperts | None, + shared_experts_input: torch.Tensor | None, + ) -> torch.Tensor: + assert self.moe_kernel is not None + return self.moe_kernel.apply( x, layer.w13_weight, layer.w2_weight, @@ -1738,5 +1738,6 @@ class QuarkNvfp4MoEMethod(QuarkMoEMethod): global_num_experts=layer.global_num_experts, expert_map=layer.expert_map, apply_router_weight_on_input=layer.apply_router_weight_on_input, + shared_experts=shared_experts, shared_experts_input=shared_experts_input, )