[VLM] replace deprecated image processor use_fast (#34175)

test/registered/unit/server_args/test_server_args.py

@@ -142,6 +142,41 @@ class TestMmEncoderDataParallelLogging(CustomTestCase):
         self.assertIn("high-resolution or multi-image", logs.output[0])
 
 
+class TestImageProcessorBackend(CustomTestCase):
+    def test_new_backend_does_not_set_legacy_flag(self):
+        server_args = ServerArgs(model_path="dummy", image_processor_backend="pil")
+
+        server_args._handle_deprecated_args()
+
+        self.assertEqual(server_args.image_processor_backend, "pil")
+        self.assertFalse(server_args.disable_fast_image_processor)
+
+    def test_legacy_flag_maps_to_pil_with_one_warning(self):
+        server_args = ServerArgs(model_path="dummy", disable_fast_image_processor=True)
+
+        with self.assertLogs(server_args_module.logger, level="WARNING") as logs:
+            server_args._handle_deprecated_args()
+
+        self.assertEqual(server_args.image_processor_backend, "pil")
+        self.assertTrue(server_args.disable_fast_image_processor)
+        self.assertEqual(
+            sum(
+                "--disable-fast-image-processor is deprecated" in x for x in logs.output
+            ),
+            1,
+        )
+
+    def test_legacy_flag_rejects_torchvision_backend(self):
+        server_args = ServerArgs(
+            model_path="dummy",
+            image_processor_backend="torchvision",
+            disable_fast_image_processor=True,
+        )
+
+        with self.assertRaisesRegex(ValueError, "conflicts.*torchvision"):
+            server_args._handle_deprecated_args()
+
+
 class TestMultimodalFeatureTransport(CustomTestCase):
     @staticmethod
     def _set_model_type(server_args, *, is_multimodal):

test/registered/unit/test_server_args_migration.py

@@ -87,6 +87,12 @@ class TestServerArgsAnnotatedCli(CustomTestCase):
         self.assertEqual(sa.deepep_mode, "low_latency")
         self.assertEqual(sa.elastic_ep_backend, "none")
 
+    def test_image_processor_backend_choices(self):
+        for backend in ("auto", "torchvision", "pil"):
+            with self.subTest(backend=backend):
+                sa = self._parse(["--image-processor-backend", backend])
+                self.assertEqual(sa.image_processor_backend, backend)
+
     def test_deprecated_flags_still_work(self):
         """Deprecated flags set the correct dest field."""
         sa = self._parse(["--stream-output"])

test/registered/unit/utils/test_hf_transformers.py

@@ -39,6 +39,72 @@ register_cpu_ci(est_time=6, suite="base-a-test-cpu")
 
 
 class TestGetProcessor(unittest.TestCase):
+    def test_does_not_forward_backend_to_auto_processor(self):
+        config = SimpleNamespace(model_type="test_vlm", auto_map={})
+        loaded_processor = MagicMock()
+        loaded_processor.image_processor.backend = "torchvision"
+        loaded_processor.tokenizer.chat_template = "template"
+        auto_config = MagicMock()
+        auto_config.from_pretrained.return_value = config
+        auto_processor = MagicMock()
+        auto_processor.from_pretrained.return_value = loaded_processor
+        auto_image_processor = MagicMock()
+
+        with patch.multiple(
+            processor_utils,
+            AutoConfig=auto_config,
+            AutoProcessor=auto_processor,
+            AutoImageProcessor=auto_image_processor,
+        ):
+            processor_utils.get_processor(
+                "test-model", image_processor_backend="torchvision"
+            )
+
+        call_kwargs = auto_processor.from_pretrained.call_args.kwargs
+        self.assertNotIn("backend", call_kwargs)
+        self.assertNotIn("use_fast", call_kwargs)
+        auto_image_processor.from_pretrained.assert_not_called()
+
+    def test_applies_pil_backend_only_to_image_processor(self):
+        config = SimpleNamespace(model_type="test_vlm", auto_map={})
+
+        for processor_kwargs in (
+            {"image_processor_backend": "pil"},
+            {"use_fast": False},
+        ):
+            with self.subTest(processor_kwargs=processor_kwargs):
+                loaded_processor = MagicMock()
+                loaded_processor.image_processor.backend = "torchvision"
+                loaded_processor.tokenizer.chat_template = "template"
+                pil_processor = MagicMock(backend="pil")
+                auto_config = MagicMock()
+                auto_config.from_pretrained.return_value = config
+                auto_processor = MagicMock()
+                auto_processor.from_pretrained.return_value = loaded_processor
+                auto_image_processor = MagicMock()
+                auto_image_processor.from_pretrained.return_value = pil_processor
+
+                with patch.multiple(
+                    processor_utils,
+                    AutoConfig=auto_config,
+                    AutoProcessor=auto_processor,
+                    AutoImageProcessor=auto_image_processor,
+                ):
+                    processor = processor_utils.get_processor(
+                        "test-model", **processor_kwargs
+                    )
+
+                call_kwargs = auto_processor.from_pretrained.call_args.kwargs
+                self.assertNotIn("backend", call_kwargs)
+                self.assertNotIn("use_fast", call_kwargs)
+                auto_image_processor.from_pretrained.assert_called_once_with(
+                    "test-model",
+                    trust_remote_code=False,
+                    revision=None,
+                    backend="pil",
+                )
+                self.assertIs(processor.image_processor, pil_processor)
+
     def test_resolves_model_name_before_loading_config(self):
         remote_model = "s3://bucket/model"
         local_model = "/cache/model"

test/registered/vlm/test_token_id_retokenize_e2e.py

@@ -53,9 +53,7 @@ def _build_drift_prompt(model, image_token):
     token), followed by one image placeholder. drift_delta is how many extra
     tokens the non-canonical form carries vs. the canonical re-tokenization.
     """
-    tok = AutoProcessor.from_pretrained(
-        model, trust_remote_code=True, use_fast=True
-    ).tokenizer
+    tok = AutoProcessor.from_pretrained(model, trust_remote_code=True).tokenizer
 
     def enc(text):
         return tok.encode(text, add_special_tokens=False)

test/registered/vlm/test_vision_openai_server_extra.py

@@ -14,7 +14,7 @@ class TestQwen3OmniServer(OmniOpenAITestMixin):
     extra_args = [  # workaround to fit into H100
         "--mem-fraction-static=0.90",
         "--disable-cuda-graph",
-        "--disable-fast-image-processor",
+        "--image-processor-backend=pil",
         "--grammar-backend=none",
     ]
 

test/registered/vlm/test_vlm_input_format.py

@@ -74,7 +74,7 @@ class VLMInputTestBase:
             cls.main_image.append(Image.open(BytesIO(response.content)))
 
         cls.processor = AutoProcessor.from_pretrained(
-            cls.model_path, trust_remote_code=True, use_fast=True
+            cls.model_path, trust_remote_code=True
         )
         _fix_added_tokens_encoding(cls.processor.tokenizer)
         cls._init_visual()