change tbo about cudagrah

vllm/model_executor/layers/activation.py

@@ -75,8 +75,8 @@ class SiluAndMul(CustomOp):
 
     def forward_native(self, x: torch.Tensor) -> torch.Tensor:
         """PyTorch-native implementation equivalent to forward()."""
-        if not torch.compiler.is_compiling():  # 非 capture 阶段
-            return self.forward_cuda(x)  # 强制走 fused kernel
+        if not torch.compiler.is_compiling() and envs.VLLM_ENABLE_TBO:
+            return self.forward_cuda(x)
         else:
             d = x.shape[-1] // 2
             return F.silu(x[..., :d]) * x[..., d:]

vllm/model_executor/layers/layernorm.py

@@ -166,10 +166,9 @@ class RMSNorm(CustomOp):
         residual: Optional[torch.Tensor] = None,
     ) -> Union[torch.Tensor, tuple[torch.Tensor, torch.Tensor]]:
         
-        if not torch.compiler.is_compiling():  # 非 capture 阶段
-            return self.forward_cuda(x, residual)  # 强制走 fused kernel
+        if not torch.compiler.is_compiling() and envs.VLLM_ENABLE_TBO:  
+            return self.forward_cuda(x, residual)  
         else:
-            # 否则fallback到原始实现
             orig_dtype = x.dtype
             x = x.to(torch.float32)
             if residual is not None: