Add option to restrict media domains (vllm-project#25783)

Signed-off-by: Chenheli Hua <huachenheli@outlook.com>
Signed-off-by: Russell Bryant <rbryant@redhat.com>
Co-authored-by: Chenheli Hua <huachenheli@outlook.com>
Signed-off-by: simon-mo <simon.mo@hey.com>

Author: 2 people

docs/features/multimodal_inputs.md

@@ -6,6 +6,10 @@ This page teaches you how to pass multi-modal inputs to [multi-modal models][sup
     We are actively iterating on multi-modal support. See [this RFC](gh-issue:4194) for upcoming changes,
     and [open an issue on GitHub](https://github.com/vllm-project/vllm/issues/new/choose) if you have any feedback or feature requests.
 
+!!! tip
+    When serving multi-modal models, consider setting `--allowed-media-domains` to restrict domain that vLLM can access to prevent it from accessing arbitrary endpoints that can potentially be vulnerable to Server-Side Request Forgery (SSRF) attacks. You can provide a list of domains for this arg. For example: `--allowed-media-domains upload.wikimedia.org github.com www.bogotobogo.com`
+    This restriction is especially important if you run vLLM in a containerized environment where the vLLM pods may have unrestricted access to internal networks.
+
 ## Offline Inference
 
 To input multi-modal data, follow this schema in [vllm.inputs.PromptType][]:

docs/usage/security.md

@@ -60,6 +60,12 @@ Key points from the PyTorch security guide:
 - Implement proper authentication and authorization for management interfaces
 - Follow the principle of least privilege for all system components
 
+### 4. **Restrict Domains Access for Media URLs:**
+
+Restrict domains that vLLM can access for media URLs by setting
+`--allowed-media-domains` to prevent Server-Side Request Forgery (SSRF) attacks.
+(e.g. `--allowed-media-domains upload.wikimedia.org github.com www.bogotobogo.com`)
+
 ## Security and Firewalls: Protecting Exposed vLLM Systems
 
 While vLLM is designed to allow unsafe network services to be isolated to

tests/entrypoints/openai/test_lora_resolvers.py

@@ -45,6 +45,7 @@ class MockModelConfig:
     logits_processor_pattern: Optional[str] = None
     diff_sampling_param: Optional[dict] = None
     allowed_local_media_path: str = ""
+    allowed_media_domains: Optional[list[str]] = None
     encoder_config = None
     generation_config: str = "auto"
     skip_tokenizer_init: bool = False

tests/entrypoints/openai/test_serving_chat.py

@@ -240,6 +240,7 @@ class MockModelConfig:
     logits_processor_pattern = None
     diff_sampling_param: Optional[dict] = None
     allowed_local_media_path: str = ""
+    allowed_media_domains: Optional[list[str]] = None
     encoder_config = None
     generation_config: str = "auto"
     media_io_kwargs: dict[str, dict[str, Any]] = field(default_factory=dict)

tests/multimodal/test_utils.py

@@ -66,7 +66,12 @@ async def test_fetch_image_http(image_url: str):
 @pytest.mark.parametrize("suffix", get_supported_suffixes())
 async def test_fetch_image_base64(url_images: dict[str, Image.Image],
                                   raw_image_url: str, suffix: str):
-    connector = MediaConnector()
+    connector = MediaConnector(
+        # Domain restriction should not apply to data URLs.
+        allowed_media_domains=[
+            "www.bogotobogo.com",
+            "github.com",
+        ])
     url_image = url_images[raw_image_url]
 
     try:
@@ -387,3 +392,29 @@ def test_argsort_mm_positions(case):
     modality_idxs = argsort_mm_positions(mm_positions)
 
     assert modality_idxs == expected_modality_idxs
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("video_url", TEST_VIDEO_URLS)
+@pytest.mark.parametrize("num_frames", [-1, 32, 1800])
+async def test_allowed_media_domains(video_url: str, num_frames: int):
+    connector = MediaConnector(
+        media_io_kwargs={"video": {
+            "num_frames": num_frames,
+        }},
+        allowed_media_domains=[
+            "www.bogotobogo.com",
+            "github.com",
+        ])
+
+    video_sync, metadata_sync = connector.fetch_video(video_url)
+    video_async, metadata_async = await connector.fetch_video_async(video_url)
+    assert np.array_equal(video_sync, video_async)
+    assert metadata_sync == metadata_async
+
+    disallowed_url = "https://upload.wikimedia.org/wikipedia/commons/4/47/PNG_transparency_demonstration_1.png"
+    with pytest.raises(ValueError):
+        _, _ = connector.fetch_video(disallowed_url)
+
+    with pytest.raises(ValueError):
+        _, _ = await connector.fetch_video_async(disallowed_url)

vllm/config/model.py

@@ -137,6 +137,9 @@ class ModelConfig:
     """Allowing API requests to read local images or videos from directories
     specified by the server file system. This is a security risk. Should only
     be enabled in trusted environments."""
+    allowed_media_domains: Optional[list[str]] = None
+    """If set, only media URLs that belong to this domain can be used for 
+    multi-modal inputs. """
     revision: Optional[str] = None
     """The specific model version to use. It can be a branch name, a tag name,
     or a commit id. If unspecified, will use the default version."""

vllm/config/speculative.py

@@ -281,6 +281,8 @@ def __post_init__(self):
                     trust_remote_code,
                     allowed_local_media_path=self.target_model_config.
                     allowed_local_media_path,
+                    allowed_media_domains=self.target_model_config.
+                    allowed_media_domains,
                     dtype=self.target_model_config.dtype,
                     seed=self.target_model_config.seed,
                     revision=self.revision,

vllm/engine/arg_utils.py

@@ -297,6 +297,8 @@ class EngineArgs:
     tokenizer_mode: TokenizerMode = ModelConfig.tokenizer_mode
     trust_remote_code: bool = ModelConfig.trust_remote_code
     allowed_local_media_path: str = ModelConfig.allowed_local_media_path
+    allowed_media_domains: Optional[
+        list[str]] = ModelConfig.allowed_media_domains
     download_dir: Optional[str] = LoadConfig.download_dir
     safetensors_load_strategy: str = LoadConfig.safetensors_load_strategy
     load_format: Union[str, LoadFormats] = LoadConfig.load_format
@@ -531,6 +533,8 @@ def add_cli_args(parser: FlexibleArgumentParser) -> FlexibleArgumentParser:
                                  **model_kwargs["hf_config_path"])
         model_group.add_argument("--allowed-local-media-path",
                                  **model_kwargs["allowed_local_media_path"])
+        model_group.add_argument("--allowed-media-domains",
+                                 **model_kwargs["allowed_media_domains"])
         model_group.add_argument("--revision", **model_kwargs["revision"])
         model_group.add_argument("--code-revision",
                                  **model_kwargs["code_revision"])
@@ -997,6 +1001,7 @@ def create_model_config(self) -> ModelConfig:
             tokenizer_mode=self.tokenizer_mode,
             trust_remote_code=self.trust_remote_code,
             allowed_local_media_path=self.allowed_local_media_path,
+            allowed_media_domains=self.allowed_media_domains,
             dtype=self.dtype,
             seed=self.seed,
             revision=self.revision,

vllm/entrypoints/chat_utils.py

@@ -637,6 +637,10 @@ def model_cls(self) -> type[SupportsMultiModal]:
     def allowed_local_media_path(self):
         return self._model_config.allowed_local_media_path
 
+    @property
+    def allowed_media_domains(self):
+        return self._model_config.allowed_media_domains
+
     @property
     def mm_registry(self):
         return MULTIMODAL_REGISTRY
@@ -837,6 +841,7 @@ def __init__(self, tracker: MultiModalItemTracker) -> None:
         self._connector = MediaConnector(
             media_io_kwargs=media_io_kwargs,
             allowed_local_media_path=tracker.allowed_local_media_path,
+            allowed_media_domains=tracker.allowed_media_domains,
         )
 
     def parse_image(
@@ -921,6 +926,7 @@ def __init__(self, tracker: AsyncMultiModalItemTracker) -> None:
         self._connector = MediaConnector(
             media_io_kwargs=media_io_kwargs,
             allowed_local_media_path=tracker.allowed_local_media_path,
+            allowed_media_domains=tracker.allowed_media_domains,
         )
 
     def parse_image(

vllm/entrypoints/llm.py

@@ -86,6 +86,8 @@ class LLM:
             or videos from directories specified by the server file system.
             This is a security risk. Should only be enabled in trusted
             environments.
+        allowed_media_domains: If set, only media URLs that belong to this 
+            domain can be used for multi-modal inputs.
         tensor_parallel_size: The number of GPUs to use for distributed
             execution with tensor parallelism.
         dtype: The data type for the model weights and activations. Currently,
@@ -169,6 +171,7 @@ def __init__(
         skip_tokenizer_init: bool = False,
         trust_remote_code: bool = False,
         allowed_local_media_path: str = "",
+        allowed_media_domains: Optional[list[str]] = None,
         tensor_parallel_size: int = 1,
         dtype: ModelDType = "auto",
         quantization: Optional[QuantizationMethods] = None,
@@ -264,6 +267,7 @@ def __init__(
             skip_tokenizer_init=skip_tokenizer_init,
             trust_remote_code=trust_remote_code,
             allowed_local_media_path=allowed_local_media_path,
+            allowed_media_domains=allowed_media_domains,
             tensor_parallel_size=tensor_parallel_size,
             dtype=dtype,
             quantization=quantization,