fix: 修复ep的变量未定义

set VLLM_USE_FUSED_QA_KVA_GEMM=1
feat:w4a8Linear调用apply_int8_linear，以支持blaslt

setup.py

@@ -559,10 +559,10 @@ def get_version_add(sha: Optional[str] = None) -> str:
             if sha is None:
                 sha = get_sha(vllm_root)
             if (major, minor) >= ('2', '5'):
-                version = 'das.opt5.' + sha[:7]
+                version = 'das.opt6.' + sha[:7]
     else:
         if (major, minor) >= ('2', '5'):
-            version = 'das.opt5'
+            version = 'das.opt6'
 
 
     # dtk version

vllm/envs.py

@@ -1365,7 +1365,7 @@ environment_variables: dict[str, Callable[[], Any]] = {
     # Only quantized DeepSeek models supported.
     # Unquantized versions are not supported.
     "VLLM_USE_FUSED_QA_KVA_GEMM":
-        lambda: (os.environ.get("VLLM_USE_FUSED_QA_KVA_GEMM", "False").lower() in
+        lambda: (os.environ.get("VLLM_USE_FUSED_QA_KVA_GEMM", "True").lower() in
                 ("true", "1")),
     "VLLM_ZERO_OVERHEAD_ENHANCE":
         lambda: (os.getenv('VLLM_ZERO_OVERHEAD_ENHANCE', '0').lower() in

vllm/model_executor/layers/quantization/slimquant_w4a8.py

@@ -5,6 +5,7 @@ from vllm.model_executor.utils import set_weight_attrs
 from vllm.distributed import get_tensor_model_parallel_world_size
 from torch.nn.parameter import Parameter
 from vllm.model_executor.layers.linear import (LinearBase,LinearMethodBase)
+from vllm.model_executor.layers.quantization.utils.w8a8_utils import apply_int8_linear
 from vllm.model_executor.layers.quantization.base_config import (
     QuantizationConfig, QuantizeMethodBase)
 from vllm.model_executor.layers.fused_moe import (FusedMoE, FusedMoEMethodBase,
@@ -111,7 +112,9 @@ class SlimQuantW4A8Int8LinearMethod(LinearMethodBase):
                     for key, value in configs_dict.items():
                         m=int(key.split('_')[0])
                         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,device=layer.weight.device,best_config=value)
-        else: 
+        elif self.w8a8_strategy==3:
+            layer.weight.data = layer.weight.data.T
+        else:
             weight_data=layer.weight.data
             _weight=weight_data.T.contiguous().reshape(n,-1)
             layer.weight.data=_weight
@@ -158,81 +161,13 @@ class SlimQuantW4A8Int8LinearMethod(LinearMethodBase):
         input_quant_args: Optional[list[torch.Tensor]] = None,
         silu_quant_args: Optional[list[torch.Tensor]] = None, **_
     ):
-        if envs.USE_FUSED_RMS_QUANT and input_quant_args is not None:
-            assert len(input_quant_args) == 2
-            x_q, x_scale = input_quant_args
-        elif envs.USE_FUSED_SILU_MUL_QUANT and silu_quant_args is not None:
-            assert len(silu_quant_args) == 2
-            x_q, x_scale = silu_quant_args
-        elif envs.USE_FUSED_CUSTOM_ALL_REDUCE_RMS_QUANT and input_quant_args is not None:
-            assert len(input_quant_args) == 2
-            x_q, x_scale = input_quant_args
-        elif envs.USE_FUSED_CUSTOM_ALL_REDUCE_RMS_QUANT and silu_quant_args is not None:
-            assert len(silu_quant_args) == 2
-            x_q, x_scale = silu_quant_args
-        else:
-            x_q, x_scale = per_token_quant_int8(x)
-
-        if self.w8a8_strategy==1:
-            m=x_q.shape[0]
-            k=x_q.shape[1]
-            n=layer.weight.shape[1]
-            
-            if len(W8A8_TRITONJSON.triton_json_dict)==0:
-                best_config=None
-                
-            elif f"1_{n}_{k}" in  W8A8_TRITONJSON.triton_json_dict:
-                if m<=16:
-                    m_=m
-                elif m<=64:
-                    m_ = ((m + 3) // 4) * 4 #取值到最近的4的倍数
-                elif m<=160:
-                    m_ = (m // 8) * 8
-                    
-                elif m<200: #256
-                    m_=160
-                elif m<480: #512
-                    m_=256
-                elif m<960: #1024
-                    m_=512
-                elif m<2048:
-                    m_=1024
-                elif m<4096:
-                    m_=2048
-                elif m<6000:
-                    m_=4096
-                else:
-                    m_=8192  
-
-                best_config=W8A8_TRITONJSON.triton_json_dict[f"{m_}_{n}_{k}"]
-                
-            else: 
-                best_config=None
-                
-            #if best_config==None:
-            #    print("m:{},n:{},k:{}".format(m,n,k))
-            #    print("config not found!")
-
-            return ops.triton_scaled_mm(x_q,
-                                        layer.weight,
-                                        scale_a=x_scale,
-                                        scale_b=layer.weight_scale,
-                                        out_dtype=x.dtype,
-                                        bias=bias,best_config=best_config)
-        elif self.w8a8_strategy==2:
-            return ops.cutlass_scaled_mm(x_q,
-                                        layer.weight,
-                                        scale_a=x_scale,
-                                        scale_b=layer.weight_scale,
-                                        out_dtype=x.dtype,
-                                        bias=bias)
-        else:
-            return ops.rocblas_scaled_mm(x_q,
-                                        layer.weight,
-                                        scale_a=x_scale,
-                                        scale_b=layer.weight_scale,
-                                        out_dtype=x.dtype,
-                                        bias=bias) 
+        return apply_int8_linear(input=x,
+                                 weight=layer.weight,
+                                 weight_scale=layer.weight_scale,
+                                 bias=bias,
+                                 w8a8_strategy=self.w8a8_strategy,
+                                 input_quant_args=input_quant_args,
+                                 silu_quant_args=silu_quant_args)
 
 
 class SlimQuantW4A8Int8MoEMethod:

vllm/model_executor/model_loader/utils.py

@@ -288,6 +288,9 @@ def get_model_architecture(
                         os.environ['VLLM_USE_OPT_RESHAPE_AND_CACHE'] = '1'
                     if not envs.is_set("VLLM_USE_FUSED_RMS_ROPE"):
                         os.environ['VLLM_USE_FUSED_RMS_ROPE'] = '1'
+                if architectures in [['Qwen3ForCausalLM']]:
+                    if not envs.is_set("VLLM_USE_OPT_RESHAPE_AND_CACHE"):
+                        os.environ['VLLM_USE_OPT_RESHAPE_AND_CACHE'] = '1'
                 
             if architectures in [['DeepseekV32ForCausalLM']]:
                 if not envs.is_set("VLLM_USE_V32_ENCODE"):
@@ -336,6 +339,9 @@ def get_model_architecture(
                         os.environ['VLLM_USE_OPT_RESHAPE_AND_CACHE'] = '1'
                     if not envs.is_set("VLLM_USE_FUSED_RMS_ROPE"):
                         os.environ['VLLM_USE_FUSED_RMS_ROPE'] = '1'
+                if architectures in [['Qwen3ForCausalLM']]:
+                    if not envs.is_set("VLLM_USE_OPT_RESHAPE_AND_CACHE"):
+                        os.environ['VLLM_USE_OPT_RESHAPE_AND_CACHE'] = '1'
                  
             if architectures in [['DeepseekV32ForCausalLM']]:
                 if not envs.is_set("VLLM_USE_V32_ENCODE"):

vllm/model_executor/models/deepseek_v2.py

@@ -422,9 +422,6 @@ class DeepseekV2MoE(nn.Module):
                             # Fix FP16 overflow
                             # See DeepseekV2DecoderLayer for more details.
                             # fp16 mode not fused quant
-                            if i_q is not None:
-                                i_q=iqis[0]
-                                i_s=iqis[1]
                             final_hidden_states = self.experts(hidden_states=hidden_states,
                                                             router_logits=router_logits,
                                                             i_q=i_q, i_s=i_s)
@@ -469,9 +466,8 @@ class DeepseekV2MoE(nn.Module):
                             assert shared_output is not None
                             final_hidden_states += (shared_output * (1. / self.routed_scaling_factor))
                     else:
-                        if i_q is not None:
-                            i_q=iqis[0]
-                            i_s=iqis[1]
+                        if iqis is not None:
+                            i_q, i_s = iqis
                         final_hidden_states = self.experts(hidden_states=hidden_states,
                                                            router_logits=router_logits,
                                                            i_q=i_q, i_s=i_s)

vllm/v1/attention/backends/flash_attn.py

@@ -577,7 +577,7 @@ class FlashAttentionImpl(AttentionImpl):
                     layer._v_scale,
                 )
             else:
-                if envs.VLLM_USE_OPT_RESHAPE_AND_CACHE and key.dtype == value.dtype == torch.float16:
+                if envs.VLLM_USE_OPT_RESHAPE_AND_CACHE:
                     from lightop import reshape_and_cache_cuda
                     reshape_and_cache_cuda(
                         key, value,