for qwen3 dense,set VLLM_USE_OPT_RESHAPE_AND_CACHE=1

vllm/model_executor/model_loader/utils.py

@@ -212,6 +212,10 @@ 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"):
@@ -241,6 +245,10 @@ 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/v1/attention/backends/flash_attn.py

@@ -556,7 +556,7 @@ 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,