[Core] Remove lora additional vocabulary

[ahengljh]

Purpose

This PR removes the ability for each LoRA adapter to define its own additional
vocabulary, as proposed in issue #23474. All LoRA adapters for a given model now
share the same vocabulary as the base model.

Test Plan

1. Test basic LoRA initialization:

import vllm
llm = vllm.LLM(model="Qwen3-8B", enable_lora=True)

2. Test LoRA adapter loading (with existing adapters):

from vllm.lora.request import LoRARequest
lora_request = LoRARequest("adapter1", 1, "/path/to/lora/adapter")
outputs = llm.generate(["Test prompt"], lora_request=lora_request)

4. Run LoRA-related tests:

pytest tests/lora/test_lora_manager.py -v
pytest tests/lora/test_layers.py -v 

Test Result

• ✅ Successfully initialized vLLM with enable_lora=True without AttributeError
• ✅ LoRAConfig no longer has lora_extra_vocab_size attribute
• ✅ All model files updated to remove additional vocabulary references

(Optional) Documentation Update

No documentation updates required as this removes an undocumented internal feature.

Essential Elements of an Effective PR Description Checklist

- [X] The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)". - [X] The test plan, such as providing test command. - [X] 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.

BEFORE SUBMITTING, PLEASE READ https://docs.vllm.ai/en/latest/contributing

[gemini-code-assist]

Code Review

This pull request removes the functionality for LoRA adapters to have their own additional vocabulary, which simplifies the LoRA implementation across the codebase. The changes are extensive and touch many files, primarily removing lora_extra_vocab_size and related logic.

The changes are generally correct and align with the goal of the PR. I've identified a few areas where the code can be cleaned up for better maintainability, such as removing redundant code blocks and expressions. These patterns are repeated across several model definition files, and I've provided examples in the comments. Addressing them would improve the overall code quality.

[gemini-code-assist]

The expression self.base_layer.org_vocab_size + 0 is redundant. It can be simplified to self.base_layer.org_vocab_size for better code clarity.

Suggested change

	self.base_layer.org_vocab_size +
	lora_config.lora_extra_vocab_size,
	0, # No extra vocab size
	self.base_layer.org_vocab_size,

[gemini-code-assist]

The lora_vocab variable is now always 0 and is unused. It can be removed for clarity. This pattern of an unused lora_vocab variable is present in several other model files and should be addressed there as well.

[gemini-code-assist]

This if lora_config: block is now empty and can be removed to improve code clarity. This pattern is repeated in several other model files (e.g., minicpm_eagle.py, phi4flash.py, phi4mm.py, step3_text.py) and should be addressed there as well.

[jeejeelee]

As a start, maybe it would be better to just provide a deprecation warning for lora_extra_vocab_size

[ahengljh]

As a start, maybe it would be better to just provide a deprecation warning for lora_extra_vocab_size

Oh I see. So we should hold this PR and just add warning info to users at this moment right?

[jeejeelee]

As a start, maybe it would be better to just provide a deprecation warning for lora_extra_vocab_size

Oh I see. So we should hold this PR and just add warning info to users at this moment right?

I think so

[ahengljh]

As a start, maybe it would be better to just provide a deprecation warning for lora_extra_vocab_size

Oh I see. So we should hold this PR and just add warning info to users at this moment right?

I think so

I just made another one-line PR #23635 for adding the warning, and we may hold this PR until you believe it's time to deprecate this feature completely. What's your opinion on this?

[jeejeelee]

I have trigged the lora testing to verify these changes

[jeejeelee]

All LoRA tests are failing. Could you take a closer look into this? @ahengljh

[ahengljh]

All LoRA tests are failing. Could you take a closer look into this? @ahengljh

Yes, I noticed that and will try to fix them

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @ahengljh.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

[jeejeelee]

@ahengljh could we speed up a bit? Thank you.

[ahengljh]

@ahengljh could we speed up a bit? Thank you.

I have been trying run the lora tests on my local machine but I found that even I ran them on the main branch, they still fail, which seems difficult for me to debug. I need to investigate and try more.

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @ahengljh.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @ahengljh.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

[jeejeelee]

@ahengljh Would you like to continue working on this PR? If so, could you resolve the conflicts first?

[ahengljh]

@ahengljh Would you like to continue working on this PR? If so, could you resolve the conflicts first?

Yes, of course, and sorry for being late, I can solve conflicts first. But there has been some problems on my local gpu env to run tests to debug. So I may need a little help on passing tests.

[jeejeelee]

Okay, please feel free to ping me on slack

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @ahengljh.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

[WoosukKwon]

@ahengljh @jeejeelee Do we have any updates? 😅

[jeejeelee]

Sorry for missing this PR, @ahengljh would you like to continue working on this PR?