使用groupgemm完成高吞吐模式适配

python/sglang/srt/layers/quantization/compressed_tensors/compressed_tensors_marlin.py

@@ -42,7 +42,7 @@ class SlimQuantCompressedTensorsMarlinConfig(CompressedTensorsConfig):
         sparsity_ignore_list: list[str],
         kv_cache_scheme: Optional[dict[str, Any]] = None,
         config: Optional[dict[str, Any]] = None,
-        packed_modules_mapping: Optional[dict[str, list[str]]] = None,   
+        packed_modules_mapping: Optional[dict[str, list[str]]] = None,
     ):
         super().__init__(
             target_scheme_map,
@@ -52,10 +52,10 @@ class SlimQuantCompressedTensorsMarlinConfig(CompressedTensorsConfig):
             sparsity_ignore_list,
             kv_cache_scheme,
             config,
-            packed_modules_mapping,  
+            packed_modules_mapping,
         )
 
-    
+
     @classmethod
     def override_quantization_method(
             cls, hf_quant_cfg, user_quant) -> Optional[str]:
@@ -73,7 +73,7 @@ class SlimQuantCompressedTensorsMarlinConfig(CompressedTensorsConfig):
             prefix: str,
     ) -> Optional["QuantizeMethodBase"]:
         from sglang.srt.layers.moe.fused_moe_triton.layer import FusedMoE  # Avoid circular import
-        # from sglang.srt.layers.radix_attention import RadixAttention
+        from sglang.srt.layers.radix_attention import RadixAttention
         # Check if the layer is skipped for quantization.
         if should_ignore_layer(prefix,
                                ignore=self.ignore,
@@ -85,8 +85,8 @@ class SlimQuantCompressedTensorsMarlinConfig(CompressedTensorsConfig):
                 return UnquantizedEmbeddingMethod()#UnquantizedLinearMethod()
             layer.scheme = scheme
             return CompressedTensorsLinearMethod(self)
-        # if isinstance(layer, RadixAttention):
-        #     return CompressedTensorsKVCacheMethod(self)
+        if isinstance(layer, RadixAttention):
+            return CompressedTensorsKVCacheMethod(self)
         if isinstance(layer, FusedMoE):
             return CompressedTensorsMarlinMoEMethod.get_moe_method(self, layer)
         return None