set VLLM_USE_FUSED_QA_KVA_GEMM=1

vllm/envs.py

@@ -1344,7 +1344,7 @@ 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