[Core][Multimodal] Allow passing multi_modal_uuids as multimodal identifiers. (#23394)

Signed-off-by: Roger Wang <hey@rogerw.io>
This commit is contained in:
Roger Wang
2025-08-30 18:01:22 -07:00
committed by GitHub
parent 5b8077b8ac
commit 749be00a98
10 changed files with 455 additions and 54 deletions
+13 -2
View File
@@ -22,7 +22,8 @@ if TYPE_CHECKING:
from PIL.Image import Image
from transformers.feature_extraction_utils import BatchFeature
from .hasher import MultiModalHashDict
from .processing import MultiModalHashes
else:
torch = LazyLoader("torch", globals(), "torch")
@@ -115,6 +116,16 @@ The built-in modalities are defined by
[`MultiModalDataBuiltins`][vllm.multimodal.inputs.MultiModalDataBuiltins].
"""
MultiModalUUIDDict: TypeAlias = Mapping[str, Union[list[Optional[str]], str]]
"""
A dictionary containing user-provided UUIDs for items in each modality.
If a UUID for an item is not provided, its entry will be `None` and
MultiModalHasher will compute a hash for the item.
The UUID will be used to identify the item for all caching purposes
(input processing caching, embedding caching, prefix caching, etc).
"""
@dataclass(frozen=True)
class PlaceholderRange:
@@ -939,7 +950,7 @@ class MultiModalInputs(TypedDict):
mm_kwargs: MultiModalKwargsOptionalItems
"""Keyword arguments to be directly passed to the model after batching."""
mm_hashes: "MultiModalHashDict"
mm_hashes: "MultiModalHashes"
"""The hashes of the multi-modal data."""
mm_placeholders: "MultiModalPlaceholderDict"