skip is_rocm_aiter_moe_enabled and remove qwen unused int8 code

vllm/model_executor/layers/fused_moe/layer.py

@@ -25,8 +25,8 @@ from vllm.distributed import (get_dp_group, get_ep_group,
 from vllm.forward_context import ForwardContext, get_forward_context
 from vllm.logger import init_logger
 from vllm.model_executor.custom_op import CustomOp
-from vllm.model_executor.layers.fused_moe.rocm_aiter_fused_moe import (
-    is_rocm_aiter_moe_enabled)
+# from vllm.model_executor.layers.fused_moe.rocm_aiter_fused_moe import (
+#     is_rocm_aiter_moe_enabled)
 from vllm.model_executor.layers.quantization.base_config import (
     QuantizationConfig, QuantizeMethodBase)
 from vllm.model_executor.layers.fused_moe.fused_moe import (
@@ -55,11 +55,11 @@ else:
     fused_experts = None  # type: ignore
     FusedMoEPermuteExpertsUnpermute = None  # type: ignore
     FusedMoEPrepareAndFinalize = None  # type: ignore
-if is_rocm_aiter_moe_enabled():
-    from vllm.model_executor.layers.fused_moe.rocm_aiter_fused_moe import (  # noqa: E501
-        rocm_aiter_grouped_topk as grouped_topk)
-else:
-    from vllm.model_executor.layers.fused_moe.fused_moe import grouped_topk
+# if is_rocm_aiter_moe_enabled():
+#     from vllm.model_executor.layers.fused_moe.rocm_aiter_fused_moe import (  # noqa: E501
+#         rocm_aiter_grouped_topk as grouped_topk)
+# else:
+from vllm.model_executor.layers.fused_moe.fused_moe import grouped_topk
 if current_platform.is_tpu():
     from .moe_pallas import fused_moe as fused_moe_pallas
 else:
@@ -443,7 +443,7 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, CustomOp):
         self.topk_indices_dtype = None
         self.moe = moe
 
-        self.rocm_aiter_moe_enabled = is_rocm_aiter_moe_enabled()
+        self.rocm_aiter_moe_enabled = False # is_rocm_aiter_moe_enabled()
         if self.rocm_aiter_moe_enabled:
             from .rocm_aiter_fused_moe import rocm_aiter_fused_experts
             self.rocm_aiter_fused_experts = rocm_aiter_fused_experts

vllm/model_executor/layers/fused_moe/rocm_aiter_fused_moe.py

@@ -35,10 +35,9 @@ class ActivationMethod(IntEnum):
 
 @cache
 def is_rocm_aiter_moe_enabled() -> bool:
-    return False
-    # return current_platform.is_rocm() \
-    #     and envs.VLLM_ROCM_USE_AITER_MOE \
-    #     and envs.VLLM_ROCM_USE_AITER
+    return current_platform.is_rocm() \
+        and envs.VLLM_ROCM_USE_AITER_MOE \
+        and envs.VLLM_ROCM_USE_AITER
 
 
 def rocm_aiter_asm_moe_tkw1_impl(

vllm/model_executor/models/qwen.py

@@ -383,50 +383,6 @@ class QWenBaseModel(nn.Module):
                     weight.data.copy_(_weight)
                     
                     weight.data=weight.data.reshape(ori_shape[1],-1)
-                                        
-            os.environ['LM_NN'] = '0'
-            lay_key_words = [
-                "attn.c_attn.weight",
-                "attn.c_proj.weight",
-                "mlp.gate_up_proj.weight",
-                "mlp.c_proj.weight",
-            ]
-            combined_words = "|".join(lay_key_words)
-            weight_shapes=[]
-            all_json={}
-            matched_key_words=set()
-            
-            for layername in loaded_params:
-                weight = params_dict[layername] 
-                matches = re.findall(combined_words, layername)
-                if matches and "scale" not in layername:
-                    weight_data =params_dict[layername]
-                    n=weight_data.shape[0]
-                    
-                    #rocblas和cutlass目前都需要weight做处理，但是triton不用
-                    if self.w8a8_strategy!=1:
-                        _weight=weight_data.T.contiguous().reshape(n,-1)
-                        weight_data.data.copy_(_weight)  
-                    
-                    #下面是针对模型记录模型出现k和n值 
-                    elif len(matched_key_words) < 4 and matches[0] not in matched_key_words:
-                        matched_key_words.add(matches[0]) 
-                        k=weight_data.shape[1]
-                        weight_shapes.append({n,k})
-                
-                        json_file=self.tritonsingleton.get_w8a8json_name(n,k)
-                        configs_dict=self.tritonsingleton.get_triton_cache(json_file,n,k)
-                        if configs_dict:
-                            all_json.update(configs_dict)
-                                              
-            if self.w8a8_strategy==1:
-                self.tritonsingleton.triton_json_dict.append(all_json)
-                #找到的所有config都进行一次warmup
-                for key, value in all_json.items():
-                    m=int(key.split('_')[0])
-                    n=int(key.split('_')[1])
-                    k=int(key.split('_')[2])
-                    ops.triton_int8_gemm_helper(m=m,n=n,k=k,per_token_act_quant=True,per_out_channel_weight_quant=True,use_bias=False,best_config=value)
             
         return loaded_params