[V0 Deprecation] Remove V0 logic from get_input_embeddings interface
#25242
Conversation
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
DarkLight1337
added
the
ready
|
|
||
| def merge_multimodal_embeddings_from_map( | ||
| inputs_embeds: torch.Tensor, multimodal_embeddings: NestedTensors, | ||
| placeholder_map: MultiModalPlaceholderMap.IndexMap) -> torch.Tensor: |
There was a problem hiding this comment.
MultiModalPlaceholderMap itself can be removed once the V0 model runners are gone
|
I have also decided not to touch |
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
There was a problem hiding this comment.
Code Review
This pull request refactors the get_input_embeddings method across several multimodal models (hyperclovax_vision, ultravox) to remove V0 compatibility logic. The changes simplify the code by consistently using the merge_multimodal_embeddings utility and removing deprecated code paths and functions like merge_multimodal_embeddings_from_map. The SupportsMultiModal interface in interfaces.py is also cleaned up by removing overloads and parameters related to V0. These changes improve code clarity and maintainability by centralizing the embedding merge logic and removing obsolete code. The implementation appears correct and is a positive step towards simplifying the codebase.
vllm-project#25242) Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
vllm-project#25242) Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
vllm-project#25242) Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk> Signed-off-by: charlifu <charlifu@amd.com>
Purpose
Remove V0-related code from
get_input_embeddingsin order to facilitate #16229Test Plan
Test Result
Essential Elements of an Effective PR Description Checklist
supported_models.mdandexamplesfor a new model.