assert num_tokens_after_padding bounds

Signed-off-by: Varun Sundar Rabindranath <vsundarr@redhat.com>

Author: Varun Sundar Rabindranath

vllm/v1/worker/gpu_model_runner.py

@@ -3401,6 +3401,8 @@ def _dummy_run(
         with self.maybe_dummy_run_with_lora(
             self.lora_config, num_scheduled_tokens, remove_lora
         ):
+        # Make sure padding doesn't exceed max_num_tokens
+        assert num_tokens_after_padding <= self.max_num_tokens
             model_kwargs = self._init_model_kwargs(num_tokens_after_padding)
             if self.supports_mm_inputs and not self.model_config.is_encoder_decoder:
                 input_ids = None