[torch.compile] fix sym_tensor_indices (vllm-project#12191)

Signed-off-by: youkaichao <[email protected]>

Author: youkaichao

vllm/compilation/backends.py

@@ -624,9 +624,13 @@ def __call__(self, graph: fx.GraphModule, example_inputs) -> Callable:
         ]
 
         # index of tensors that have symbolic shapes (batch size)
+        # for weights and static buffers, they will have concrete shapes.
+        # symbolic shape only happens for input tensors.
+        from torch.fx.experimental.symbolic_shapes import is_symbolic
         self.sym_tensor_indices = [
             i for i, x in enumerate(fake_args)
-            if isinstance(x, torch._subclasses.fake_tensor.FakeTensor)
+            if isinstance(x, torch._subclasses.fake_tensor.FakeTensor) and \
+                any(is_symbolic(d) for d in x.size())
         ]
 
         # compiler managed cudagraph input buffers