[LoRA][2/2]Remove LoRA extra vocab

[jeejeelee]

Purpose

In this PR:

• Clean up all code related to lora_extra_vocab_size and lora_vocab_padding_size FIX [RFC]: Disallow extra vocab for LoRA #23474
• Train the LoRA for meta-llama/Llama-3.2-3B-Instruct to replace the previous llama2-7B LoRA model, to test adding LoRA to the lm_head and embedding layers and reduce the CI test pressure
• Using the Qwen/Qwen3-0.6B model for related LoRA testing can also reduce the testing pressure on CI

Will continue to clean up the related code subsequently

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.

[gemini-code-assist]

Code Review

This pull request is part of an effort to remove the LoRA extra vocabulary feature. The changes to the model implementations in granite.py and teleflm.py are consistent with this goal, and tests related to the extra vocabulary feature have been correctly removed. However, I've found a critical issue in tests/lora/test_lora_manager.py where a change breaks a test, which will need to be addressed.

[gemini-code-assist]

load_file(os.path.join(sql_lora_files, "")) will attempt to load from a directory path, which will raise an IsADirectoryError and cause the test to fail.

Since new embeddings are being removed, new_embeddings should likely be an empty dictionary.

Please note that changing this line to new_embeddings = {} will reveal another issue in this test: a KeyError will be raised on line 83. The test logic from line 77 onwards needs to be updated to reflect that lora.embeddings_tensor is now always None. The if/else block can be simplified to assert lora.embeddings_tensor is None.

Suggested change

	new_embeddings = load_file(os.path.join(sql_lora_files, ""))
	new_embeddings = {}

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[chatgpt-codex-connector]

Avoid loading deleted new_embeddings file

The updated test still calls load_file(os.path.join(sql_lora_files, "")), which resolves to the LoRA directory itself. safetensors.torch.load_file only accepts paths to .safetensors files and will raise IsADirectoryError, so test_from_lora_tensors now crashes before exercising any behaviour. If extra vocab embeddings are no longer used, this load should be dropped or replaced with a stub so the test can run.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

💡 Codex Review

vllm/vllm/v1/worker/tpu_model_runner.py

Lines 219 to 223 in 808b6e0

	self.hidden_size = model_config.get_hidden_size()
	self.vocab_size = model_config.get_vocab_size()
	if self.lora_config is not None:
	self.vocab_size += self.lora_config.lora_extra_vocab_size

TPU runner still references removed LoRA extra vocab attr

The LoRAConfig dataclass no longer exposes lora_extra_vocab_size, but the TPU model runner continues to access self.lora_config.lora_extra_vocab_size when computing the model vocabulary. Any LoRA-enabled run on TPU will now raise an AttributeError during initialization, preventing TPU LoRA serving altogether. Consider removing this addition or guarding it behind a compatibility shim so TPU paths remain functional.

---

vllm/vllm/lora/punica_wrapper/punica_tpu.py

Lines 289 to 296 in 808b6e0

	def _update_base_metadata(
	self,
	mapping: "LoRAMapping",
	lora_index_to_id: list[int | None],
	max_loras: int,
	vocab_size: int,
	extra_vocab_size: int,
	):

Punica TPU metadata hook requires extra arg no longer passed

PunicaWrapperBase.update_metadata now calls _update_base_metadata with four positional parameters (mapping, lora_index_to_id, max_loras, vocab_size) and unconditionally sets extra_vocab_size to 0 internally. The TPU implementation still overrides _update_base_metadata with a five‑argument signature, so the call from the base class will raise TypeError: _update_base_metadata() missing 1 required positional argument whenever TPU LoRA metadata is updated. The override should drop the extra_vocab_size parameter or accept a default to keep the method compatible.

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".