[3/n] chore: decouple AWQ implementation from vLLM dependency (#8113)

Co-authored-by: AniZpZ <zhuangsen.zp@antgroup.com>

python/sglang/srt/models/deepseek_v2.py

@@ -355,6 +355,7 @@ class DeepseekV2MoE(nn.Module):
                 self.shared_experts.gate_up_proj.quant_method, "quant_config"
             ) and self.shared_experts.gate_up_proj.quant_method.quant_config.get_name() in {
                 "awq",
+                "awq_marlin",
                 "moe_wna16",
             }
             self.shared_experts_is_int8 = (
@@ -929,7 +930,7 @@ class DeepseekV2AttentionMLA(nn.Module):
             has_fused_proj
             and hasattr(self.fused_qkv_a_proj_with_mqa.quant_method, "quant_config")
             and self.fused_qkv_a_proj_with_mqa.quant_method.quant_config.get_name()
-            in {"awq", "moe_wna16"}
+            in {"awq", "awq_marlin", "moe_wna16"}
         )
         self.use_min_latency_fused_a_gemm = (
             has_fused_proj
@@ -2551,6 +2552,7 @@ class DeepseekV2ForCausalLM(nn.Module):
                                 cat_dim = 0
                                 if self.quant_config is not None and (
                                     self.quant_config.get_name() == "awq"
+                                    or self.quant_config.get_name() == "awq_marlin"
                                     or self.quant_config.get_name() == "moe_wna16"
                                 ):
                                     cat_dim = 1