Fix shared experts fusion error (#6289)

python/sglang/srt/models/deepseek_v2.py

@@ -1903,7 +1903,8 @@ class DeepseekV2ForCausalLM(nn.Module):
         if self.n_share_experts_fusion > 0:
             weights_list = list(weights)
             weights_dict = dict(weights_list)
-            if self.quant_config is None or self.quant_config.get_name() == "w8a8_int8":
+            if self.quant_config is not None:
+                if self.quant_config.get_name() == "w8a8_int8":
                     suffix_list = [
                         "down_proj.weight",
                         "down_proj.weight_scale",
@@ -1912,7 +1913,10 @@ class DeepseekV2ForCausalLM(nn.Module):
                         "up_proj.weight",
                         "up_proj.weight_scale",
                     ]
-            else:
+                elif (
+                    self.quant_config.get_name() == "fp8"
+                    or self.quant_config.get_name() == "blockwise_int8"
+                ):
                     suffix_list = [
                         "down_proj.weight",
                         "down_proj.weight_scale_inv",
@@ -1921,6 +1925,28 @@ class DeepseekV2ForCausalLM(nn.Module):
                         "up_proj.weight",
                         "up_proj.weight_scale_inv",
                     ]
+                elif self.quant_config.get_name() == "awq":
+                    suffix_list = [
+                        "down_proj.qweight",
+                        "down_proj.qzeros",
+                        "down_proj.scales",
+                        "gate_proj.qweight",
+                        "gate_proj.qzeros",
+                        "gate_proj.scales",
+                        "up_proj.qweight",
+                        "up_proj.qzeros",
+                        "up_proj.scales",
+                    ]
+                else:
+                    raise ValueError(
+                        f"Unsupported shared expert fusion for quantization: {self.quant_config.get_name()}."
+                    )
+            else:
+                suffix_list = [
+                    "down_proj.weight",
+                    "gate_proj.weight",
+                    "up_proj.weight",
+                ]
             names_to_remove = []
 
             moe_layers = (