set VLLM_USE_TC_PAGED_ATTN=1

vllm/envs.py

@@ -206,7 +206,7 @@ environment_variables: Dict[str, Callable[[], Any]] = {
     
     # flag to control vllm to use optimized tc paged attn kernels
     "VLLM_USE_TC_PAGED_ATTN":
-    lambda: (os.environ.get("VLLM_USE_TC_PAGED_ATTN", "False").lower() in
+    lambda: (os.environ.get("VLLM_USE_TC_PAGED_ATTN", "True").lower() in
              ("true", "1")),
     
     # flag to control if vllm print pa parameters