fix(deploy): pass trust_remote_code=True to load_tokenizer in in-framework deploy

Megatron-Bridge added a load_tokenizer() guard that raises ValueError when the
checkpoint tokenizer config has trust_remote_code=True baked in (e.g. Llama
tokenizers) but the caller does not opt in. The in-framework deploy path called
load_tokenizer(checkpoint_path) without it, breaking Ray and Triton deployments
of converted MBridge checkpoints (regression in container 26.06.01.rc0).

Adds a unit test covering the patched call (red-green verified).

Detected by: hermes PR-to-testcase pipeline
NVBug: 6393765

Signed-off-by: Pruthviraj Prakash <pruprakash@nvidia.com>

Author: pruprakash

nemo_deploy/llm/inference/inference_base.py

@@ -265,7 +265,11 @@ def setup_megatron_model_and_tokenizer_for_inference(
         megatron_args=mlm_args,
         use_cpu_init=False,
     )
-    tokenizer = load_tokenizer(checkpoint_path)
+    # In-framework deploy operates on a user-supplied checkpoint that is already trusted
+    # (locally converted and served by the same user), so opt in to the Megatron-Bridge
+    # load_tokenizer security guard that requires trust_remote_code=True when the checkpoint
+    # tokenizer config has it baked in (e.g. Llama tokenizers).
+    tokenizer = load_tokenizer(checkpoint_path, trust_remote_code=True)
     return model, tokenizer, mlm_args
 
 

tests/unit_tests/deploy/test_inference_base.py

@@ -434,6 +434,43 @@ def test_attention_backend_conversion_flash(
         # Verify that attention_backend was converted to enum
         self.assertEqual(mock_config.attention_backend, AttnBackend.flash)
 
+    @patch("nemo_deploy.llm.inference.inference_base.torch_distributed_init")
+    @patch("nemo_deploy.llm.inference.inference_base.load_model_config")
+    @patch("nemo_deploy.llm.inference.inference_base.initialize_megatron_for_inference")
+    @patch("nemo_deploy.llm.inference.inference_base.build_and_load_model")
+    @patch("nemo_deploy.llm.inference.inference_base.load_tokenizer")
+    def test_load_tokenizer_passes_trust_remote_code(
+        self, mock_load_tokenizer, mock_build_model, mock_init_megatron, mock_load_config, mock_torch_dist
+    ):
+        """Regression (NVBug 6393765): load_tokenizer must be called with trust_remote_code=True.
+
+        Megatron-Bridge's load_tokenizer raises ValueError when the checkpoint tokenizer config
+        has trust_remote_code=True baked in (e.g. Llama tokenizers) but the caller does not opt in.
+        The in-framework deploy path must pass trust_remote_code=True so Ray/Triton deployments of
+        converted MBridge checkpoints load the tokenizer instead of failing.
+        """
+        mock_config = MagicMock()
+        mock_config.attention_backend = None
+        mock_config.tensor_model_parallel_size = 1
+        mock_config.pipeline_model_parallel_size = 1
+        mock_config.context_parallel_size = 1
+        mock_config.expert_model_parallel_size = 1
+
+        mock_mlm_args = MagicMock()
+        mock_load_config.return_value = (mock_config, mock_mlm_args)
+
+        mock_build_model.return_value = [MagicMock()]
+        mock_load_tokenizer.return_value = MagicMock()
+
+        setup_megatron_model_and_tokenizer_for_inference(
+            checkpoint_path=self.mock_path,
+            tensor_model_parallel_size=1,
+            pipeline_model_parallel_size=1,
+        )
+
+        # Pre-fix this was load_tokenizer(checkpoint_path) (no kwarg) and the MB guard raised.
+        mock_load_tokenizer.assert_called_once_with(self.mock_path, trust_remote_code=True)
+
     @patch("nemo_deploy.llm.inference.inference_base.torch_distributed_init")
     @patch("nemo_deploy.llm.inference.inference_base.load_model_config")
     @patch("nemo_deploy.llm.inference.inference_base.initialize_megatron_for_inference")