[VLM] Support caching in merged multi-modal processor #11341
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Signed-off-by: DarkLight1337 <[email protected]>
|
👋 Hi! Thank you for contributing to the vLLM project. Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging. To run CI, PR reviewers can do one of these:
🚀 |
DarkLight1337
changed the title
DarkLight1337
changed the title
Signed-off-by: DarkLight1337 <[email protected]>
This was referenced Dec 19, 2024
Signed-off-by: DarkLight1337 <[email protected]>
Signed-off-by: DarkLight1337 <[email protected]>
Signed-off-by: DarkLight1337 <[email protected]>
Signed-off-by: DarkLight1337 <[email protected]>
Signed-off-by: DarkLight1337 <[email protected]>
Signed-off-by: DarkLight1337 <[email protected]>
Signed-off-by: DarkLight1337 <[email protected]>
DarkLight1337
commented
Dec 19, 2024
vllm/multimodal/processing.py
Outdated
Comment on lines
616
to
661
| def _iter_bytes_to_hash(self, key: str, obj: object) -> Iterable[bytes]: | ||
| # Recursive cases | ||
| if isinstance(obj, (list, tuple)): | ||
| for elem in obj: | ||
| yield from self._iter_bytes_to_hash(key, elem) | ||
| return | ||
| if isinstance(obj, dict): | ||
| for k, v in obj.items(): | ||
| yield from self._iter_bytes_to_hash(f"{key}.{k}", v) | ||
| return | ||
|
|
||
| # Simple cases | ||
| if isinstance(obj, str): | ||
| yield key.encode("utf-8") | ||
| yield obj.encode("utf-8") | ||
| return | ||
| if isinstance(obj, bytes): | ||
| yield key.encode("utf-8") | ||
| yield obj | ||
| return | ||
| if isinstance(obj, Image): | ||
| yield key.encode("utf-8") | ||
| yield obj.tobytes() | ||
| return | ||
|
|
||
| # Convertible to NumPy arrays | ||
| if isinstance(obj, torch.Tensor): | ||
| obj = obj.numpy() | ||
| if isinstance(obj, (int, float)): | ||
| obj = np.array(obj) | ||
| if isinstance(obj, np.ndarray): | ||
| yield key.encode("utf-8") | ||
| yield obj.tobytes() | ||
| return | ||
|
|
||
| msg = f"Unable to hash object of type {type(obj)}" | ||
| raise NotImplementedError(msg) | ||
|
|
||
| def _hash_kwargs(self, **kwargs: object) -> str: | ||
| hasher = blake3() | ||
|
|
||
| for k, v in kwargs.items(): | ||
| for item_bytes in self._iter_bytes_to_hash(k, v): | ||
| hasher.update(item_bytes) | ||
|
|
||
| return hasher.hexdigest() |
There was a problem hiding this comment.
I'm a bit worried about unintentional hash collisions. Is there a better way to do this?
Signed-off-by: DarkLight1337 <[email protected]>
Signed-off-by: DarkLight1337 <[email protected]>
Signed-off-by: DarkLight1337 <[email protected]>
Signed-off-by: DarkLight1337 <[email protected]>
Signed-off-by: DarkLight1337 <[email protected]>
Signed-off-by: DarkLight1337 <[email protected]>
DarkLight1337
force-pushed
the
mm-processor-cache
branch
from
b2dac49 to
5dde7d0
Compare
Signed-off-by: DarkLight1337 <[email protected]>
Signed-off-by: DarkLight1337 <[email protected]>
Signed-off-by: DarkLight1337 <[email protected]>
Signed-off-by: DarkLight1337 <[email protected]>
Signed-off-by: DarkLight1337 <[email protected]>
Signed-off-by: DarkLight1337 <[email protected]>
Signed-off-by: DarkLight1337 <[email protected]>
Signed-off-by: DarkLight1337 <[email protected]>
Signed-off-by: DarkLight1337 <[email protected]>
Signed-off-by: DarkLight1337 <[email protected]>
Signed-off-by: DarkLight1337 <[email protected]>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
V1 multi-modal cache is currently incompatible with the merged multi-modal processor. To mitigate the performance hit, this PR adds a cache inside the merged multi-modal processor.
Note: Even with this PR, none of the models that currently use merged multi-modal processor actually support fine-grained caching because their HF processors all require text inputs.Now supported by using the inner modality-specific processor.