Add support for Qwen3 MoE+GPTQ

2025-11-15 20:14:45 +08:00
parent b296c44ae0
commit 8152e24cb2
35 changed files with 6468 additions and 574 deletions
--- a/vllm/model_executor/layers/quantization/moe_wna16.py
+++ b/vllm/model_executor/layers/quantization/moe_wna16.py
@@ -298,6 +298,10 @@ class MoeWNA16Method(FusedMoEMethodBase):
        e_score_correction_bias: Optional[torch.Tensor] = None,
        apply_router_weight_on_input: bool = False,
        activation: str = "silu",
+        enable_eplb: bool = False,
+        expert_load_view: torch.Tensor | None = None,
+        logical_to_physical_map: torch.Tensor | None = None,
+        logical_replica_count: torch.Tensor | None = None,
    ) -> torch.Tensor:
        from vllm.model_executor.layers.fused_moe import fused_experts
        assert activation == "silu", "Only SiLU activation is supported."