fix CPU backend

Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com>

Author: LucasWilkinson

vllm/v1/cudagraph_dispatcher.py

@@ -161,9 +161,9 @@ def dispatch(
         A new batch descriptor is returned as we might dispatch a uniform batch
         to a graph that supports a more general batch (uniform to non-uniform).
         """
-        # if not initialized, just skip dispatching.
         if (
             not self.keys_initialized
+            or self.cudagraph_mode == CUDAGraphMode.NONE
             or num_tokens > self.compilation_config.max_cudagraph_capture_size
         ):
             return CUDAGraphMode.NONE, BatchDescriptor(num_tokens)