Merge branch 'main' into batch_invariant_b200

Author: yewentao256

.gitignore

@@ -221,3 +221,6 @@ csrc/moe/marlin_moe_wna16/kernel_*
 
 # Ignore ep_kernels_workspace folder
 ep_kernels_workspace/
+
+# Allow tracked library source folders under submodules (e.g., benchmarks/lib)
+!vllm/benchmarks/lib/

vllm/config/model.py

@@ -168,12 +168,6 @@ class ModelConfig:
     """The specific revision to use for the model code on the Hugging Face Hub.
     It can be a branch name, a tag name, or a commit id. If unspecified, will
     use the default version."""
-    rope_scaling: dict[str, Any] = field(default_factory=dict)
-    """RoPE scaling configuration. For example,
-    `{"rope_type":"dynamic","factor":2.0}`."""
-    rope_theta: float | None = None
-    """RoPE theta. Use with `rope_scaling`. In some cases, changing the RoPE
-    theta improves the performance of the scaled model."""
     tokenizer_revision: str | None = None
     """The specific revision to use for the tokenizer on the Hugging Face Hub.
     It can be a branch name, a tag name, or a commit id. If unspecified, will
@@ -338,8 +332,6 @@ def compute_hash(self) -> str:
         factors.append(self.generation_config)
         factors.append(self.model_impl)
         factors.append(self.override_generation_config)
-        factors.append(self.rope_scaling)
-        factors.append(self.rope_theta)
         factors.append(self.video_pruning_rate)
         factors.append(self.enable_prompt_embeds)
 
@@ -481,25 +473,6 @@ def __post_init__(
                     hf_overrides_kw[key] = value
             hf_overrides_fn = None
 
-        if self.rope_scaling:
-            hf_override: dict[str, Any] = {"rope_scaling": self.rope_scaling}
-            hf_overrides_kw.update(hf_override)
-            hf_overrides_str = json.dumps(hf_overrides_kw)
-            msg = (
-                "`--rope-scaling` will be removed in a future release. "
-                f"'Please instead use `--hf-overrides '{hf_overrides_str}'`"
-            )
-            warnings.warn(DeprecationWarning(msg), stacklevel=2)
-        if self.rope_theta is not None:
-            hf_override = {"rope_theta": self.rope_theta}
-            hf_overrides_kw.update(hf_override)
-            hf_overrides_str = json.dumps(hf_overrides_kw)
-            msg = (
-                "`--rope-theta` will be removed in a future release. "
-                f"'Please instead use `--hf-overrides '{hf_overrides_str}'`"
-            )
-            warnings.warn(DeprecationWarning(msg), stacklevel=2)
-
         self.maybe_pull_model_tokenizer_for_runai(self.model, self.tokenizer)
 
         if (

vllm/engine/arg_utils.py

@@ -438,8 +438,6 @@ class EngineArgs:
     aggregate_engine_logging: bool = False
     revision: str | None = ModelConfig.revision
     code_revision: str | None = ModelConfig.code_revision
-    rope_scaling: dict[str, Any] = get_field(ModelConfig, "rope_scaling")
-    rope_theta: float | None = ModelConfig.rope_theta
     hf_token: bool | str | None = ModelConfig.hf_token
     hf_overrides: HfOverrides = get_field(ModelConfig, "hf_overrides")
     tokenizer_revision: str | None = ModelConfig.tokenizer_revision
@@ -617,8 +615,6 @@ def add_cli_args(parser: FlexibleArgumentParser) -> FlexibleArgumentParser:
         )
         model_group.add_argument("--revision", **model_kwargs["revision"])
         model_group.add_argument("--code-revision", **model_kwargs["code_revision"])
-        model_group.add_argument("--rope-scaling", **model_kwargs["rope_scaling"])
-        model_group.add_argument("--rope-theta", **model_kwargs["rope_theta"])
         model_group.add_argument(
             "--tokenizer-revision", **model_kwargs["tokenizer_revision"]
         )
@@ -1184,8 +1180,6 @@ def create_model_config(self) -> ModelConfig:
             seed=self.seed,
             revision=self.revision,
             code_revision=self.code_revision,
-            rope_scaling=self.rope_scaling,
-            rope_theta=self.rope_theta,
             hf_token=self.hf_token,
             hf_overrides=self.hf_overrides,
             tokenizer_revision=self.tokenizer_revision,

vllm/entrypoints/chat_utils.py

@@ -43,11 +43,12 @@
 # pydantic needs the TypedDict from typing_extensions
 from typing_extensions import Required, TypedDict
 
+from vllm import envs
 from vllm.config import ModelConfig
 from vllm.logger import init_logger
 from vllm.model_executor.models import SupportsMultiModal
 from vllm.multimodal import MULTIMODAL_REGISTRY, MultiModalDataDict, MultiModalUUIDDict
-from vllm.multimodal.utils import MediaConnector
+from vllm.multimodal.utils import MEDIA_CONNECTOR_REGISTRY, MediaConnector
 from vllm.transformers_utils.chat_templates import get_chat_template_fallback_path
 from vllm.transformers_utils.processor import cached_get_processor
 from vllm.transformers_utils.tokenizer import AnyTokenizer, MistralTokenizer
@@ -806,7 +807,9 @@ def __init__(self, tracker: MultiModalItemTracker) -> None:
         self._tracker = tracker
         multimodal_config = self._tracker.model_config.multimodal_config
         media_io_kwargs = getattr(multimodal_config, "media_io_kwargs", None)
-        self._connector = MediaConnector(
+
+        self._connector: MediaConnector = MEDIA_CONNECTOR_REGISTRY.load(
+            envs.VLLM_MEDIA_CONNECTOR,
             media_io_kwargs=media_io_kwargs,
             allowed_local_media_path=tracker.allowed_local_media_path,
             allowed_media_domains=tracker.allowed_media_domains,
@@ -891,7 +894,8 @@ def __init__(self, tracker: AsyncMultiModalItemTracker) -> None:
         self._tracker = tracker
         multimodal_config = self._tracker.model_config.multimodal_config
         media_io_kwargs = getattr(multimodal_config, "media_io_kwargs", None)
-        self._connector = MediaConnector(
+        self._connector: MediaConnector = MEDIA_CONNECTOR_REGISTRY.load(
+            envs.VLLM_MEDIA_CONNECTOR,
             media_io_kwargs=media_io_kwargs,
             allowed_local_media_path=tracker.allowed_local_media_path,
             allowed_media_domains=tracker.allowed_media_domains,

vllm/envs.py

@@ -70,6 +70,7 @@
     VLLM_MEDIA_LOADING_THREAD_COUNT: int = 8
     VLLM_MAX_AUDIO_CLIP_FILESIZE_MB: int = 25
     VLLM_VIDEO_LOADER_BACKEND: str = "opencv"
+    VLLM_MEDIA_CONNECTOR: str = "http"
     VLLM_MM_INPUT_CACHE_GIB: int = 4
     VLLM_TARGET_DEVICE: str = "cuda"
     VLLM_MAIN_CUDA_VERSION: str = "12.8"
@@ -738,6 +739,14 @@ def get_vllm_port() -> int | None:
     "VLLM_VIDEO_LOADER_BACKEND": lambda: os.getenv(
         "VLLM_VIDEO_LOADER_BACKEND", "opencv"
     ),
+    # Media connector implementation.
+    # - "http": Default connector that supports fetching media via HTTP.
+    #
+    # Custom implementations can be registered
+    # via `@MEDIA_CONNECTOR_REGISTRY.register("my_custom_media_connector")` and
+    # imported at runtime.
+    # If a non-existing backend is used, an AssertionError will be thrown.
+    "VLLM_MEDIA_CONNECTOR": lambda: os.getenv("VLLM_MEDIA_CONNECTOR", "http"),
     # [DEPRECATED] Cache size (in GiB per process) for multimodal input cache
     # Default is 4 GiB per API process + 4 GiB per engine core process
     "VLLM_MM_INPUT_CACHE_GIB": lambda: int(os.getenv("VLLM_MM_INPUT_CACHE_GIB", "4")),

vllm/multimodal/utils.py

@@ -20,6 +20,7 @@
 from vllm.connections import HTTPConnection, global_http_connection
 from vllm.logger import init_logger
 from vllm.utils.jsontree import json_map_leaves
+from vllm.utils.registry import ExtensionManager
 
 from .audio import AudioMediaIO
 from .base import MediaIO
@@ -46,7 +47,10 @@
 
 _M = TypeVar("_M")
 
+MEDIA_CONNECTOR_REGISTRY = ExtensionManager()
 
+
+@MEDIA_CONNECTOR_REGISTRY.register("http")
 class MediaConnector:
     def __init__(
         self,

vllm/multimodal/video.py

@@ -14,6 +14,7 @@
 
 from vllm import envs
 from vllm.logger import init_logger
+from vllm.utils.registry import ExtensionManager
 
 from .base import MediaIO
 from .image import ImageMediaIO
@@ -63,25 +64,7 @@ def load_bytes(
         raise NotImplementedError
 
 
-class VideoLoaderRegistry:
-    def __init__(self) -> None:
-        self.name2class: dict[str, type] = {}
-
-    def register(self, name: str):
-        def wrap(cls_to_register):
-            self.name2class[name] = cls_to_register
-            return cls_to_register
-
-        return wrap
-
-    @staticmethod
-    def load(cls_name: str) -> VideoLoader:
-        cls = VIDEO_LOADER_REGISTRY.name2class.get(cls_name)
-        assert cls is not None, f"VideoLoader class {cls_name} not found"
-        return cls()
-
-
-VIDEO_LOADER_REGISTRY = VideoLoaderRegistry()
+VIDEO_LOADER_REGISTRY = ExtensionManager()
 
 
 @VIDEO_LOADER_REGISTRY.register("opencv")

vllm/utils/registry.py

@@ -0,0 +1,49 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+from typing import Any
+
+
+class ExtensionManager:
+    """
+    A registry for managing pluggable extension classes.
+
+    This class provides a simple mechanism to register and instantiate
+    extension classes by name. It is commonly used to implement plugin
+    systems where different implementations can be swapped at runtime.
+
+    Examples:
+        Basic usage with a registry instance:
+
+        >>> FOO_REGISTRY = ExtensionManager()
+        >>> @FOO_REGISTRY.register("my_foo_impl")
+        ... class MyFooImpl(Foo):
+        ...     def __init__(self, value):
+        ...         self.value = value
+        >>> foo_impl = FOO_REGISTRY.load("my_foo_impl", value=123)
+
+    """
+
+    def __init__(self) -> None:
+        """
+        Initialize an empty extension registry.
+        """
+        self.name2class: dict[str, type] = {}
+
+    def register(self, name: str):
+        """
+        Decorator to register a class with the given name.
+        """
+
+        def wrap(cls_to_register):
+            self.name2class[name] = cls_to_register
+            return cls_to_register
+
+        return wrap
+
+    def load(self, cls_name: str, *args, **kwargs) -> Any:
+        """
+        Instantiate and return a registered extension class by name.
+        """
+        cls = self.name2class.get(cls_name)
+        assert cls is not None, f"Extension class {cls_name} not found"
+        return cls(*args, **kwargs)