Skip to content

[New Model] support new model ovis2.6#34426

Open
myselvess wants to merge 2 commits intovllm-project:mainfrom
alibaba:ovis2.6
Open

[New Model] support new model ovis2.6#34426
myselvess wants to merge 2 commits intovllm-project:mainfrom
alibaba:ovis2.6

Conversation

@myselvess
Copy link
Contributor

@myselvess myselvess commented Feb 12, 2026

Purpose

  1. support new model ovis2.6 (https://huggingface.co/AIDC-AI/Ovis2.6-30B-A3B).
  2. fixed a bug in Ovis2_5MultiModalProcessor._get_prompt_updates to support input prompt_token_ids directly.
  3. fixed a bug where post_layernorm was not needed for the last_hidden_state in Siglip2VisionTransformer.

ps: Ovis2.6-2B is coming soon

Test Plan

Test Result


Essential Elements of an Effective PR Description Checklist
  • The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
  • The test plan, such as providing test command.
  • The test results, such as pasting the results comparison before and after, or e2e results
  • (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.
  • (Optional) Release notes update. If your change is user facing, please update the release notes draft in the Google Doc.

@mergify
Copy link

mergify bot commented Feb 12, 2026

Documentation preview: https://vllm--34426.org.readthedocs.build/en/34426/

@mergify mergify bot added documentation Improvements or additions to documentation new-model Requests to new models labels Feb 12, 2026
Copy link
Contributor

@gemini-code-assist gemini-code-assist bot left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Code Review

This pull request introduces support for the new ovis2.6 model by adapting and reusing the existing ovis2.5 implementation. It also includes two bug fixes: one in Ovis2_5MultiModalProcessor to correctly handle prompt_token_ids, and another in Siglip2VisionTransformer where an unnecessary post_layernorm is removed. The changes are well-structured. My review includes one suggestion to replace a magic number with a constant-derived expression for better code clarity and maintainability.

Copy link
Member

@Isotr0py Isotr0py left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Hmmm, I'm a bit worried that these changes to support Ovis2.6 will also break existing Ovis2.5 support...

Comment on lines +428 to +429
"Ovis2_6ForCausalLM": ("ovis2_5", "Ovis2_5"),
"Ovis2_6_MoeForCausalLM": ("ovis2_5", "Ovis2_5"),
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Can you also update tests/models/registry.py? If Ovis-2.6-2B is unavailable yet, we can set is_available_online=False for it.

Comment on lines -348 to -350
vte_vocab_size - len(INDICATOR_IDS) + abs(x + 300) - 1
vte_vocab_size - len(INDICATOR_IDS) + (x - INDICATOR_IDS[0])
for x in visual_indicators
if x < -300
Copy link
Member

@Isotr0py Isotr0py Feb 12, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Will Ovis2.5's tokenizer also be updated? I'm a bit worried that this change will break Ovis2.5 support. Especially Ovis2.5 has various backbones with different tokenizer.

@DarkLight1337 DarkLight1337 added the multi-modality Related to multi-modality (#4194) label Feb 12, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

documentation Improvements or additions to documentation multi-modality Related to multi-modality (#4194) new-model Requests to new models

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants