replace triton_ of rms and act_and_mul

vllm/model_executor/layers/activation.py

@@ -77,6 +77,8 @@ class SiluAndMul(CustomOp):
         """PyTorch-native implementation equivalent to forward()."""
         if not torch.compiler.is_compiling() and envs.VLLM_ENABLE_TBO:
             return self.forward_cuda(x)
+        elif not torch.compiler.is_compiling() and envs.VLLM_USE_OPT_OP:
+            return self.forward_cuda(x)
         else:
             d = x.shape[-1] // 2
             return F.silu(x[..., :d]) * x[..., d:]

vllm/model_executor/layers/layernorm.py

@@ -168,6 +168,8 @@ class RMSNorm(CustomOp):
         
         if not torch.compiler.is_compiling() and envs.VLLM_ENABLE_TBO:  
             return self.forward_cuda(x, residual)  
+        elif not torch.compiler.is_compiling() and envs.VLLM_USE_OPT_OP:
+            return self.forward_cuda(x, residual)  
         else:
             orig_dtype = x.dtype
             x = x.to(torch.float32)