[vllm] dynamically switch lora request by Datta0 · Pull Request #3933 · unslothai/unsloth

Datta0 · 2026-01-26T04:08:50Z

for more information, see https://pre-commit.ci

gemini-code-assist · 2026-01-26T04:09:05Z

Summary of Changes

Hello @Datta0, I'm Gemini Code Assist¹! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed!

This pull request significantly enhances the vLLM integration by introducing dynamic LoRA (Low-Rank Adaptation) request handling. It modifies the model's generation functions to automatically detect updates to LoRA parameters and reload the corresponding LoRA weights on the fly. This allows for seamless switching between different fine-tuned LoRA adapters during inference without requiring explicit manual configuration, improving flexibility and ease of use for models utilizing LoRA with vLLM.

Highlights

Dynamic LoRA Request Switching: Implemented a mechanism to dynamically switch LoRA (Low-Rank Adaptation) requests for vLLM generation methods, allowing for on-the-fly updates to LoRA configurations.
Patched Generation Methods: The fast_generate and fast_generate_batches methods are now patched to automatically retrieve and apply the correct LoRA configuration before initiating generation.
Parameter Change Detection: Introduced logic within _get_lora_request to detect changes in trainable LoRA parameters (via _version attribute) and trigger a reload of the LoRA weights when an update is identified.
Temporary LoRA Storage: Utilizes temporary directories to manage the storage of dynamically loaded LoRA configurations, ensuring clean handling of resources.

🧠 New Feature in Public Preview: You can now enable Memory to help Gemini Code Assist learn from your team's feedback. This makes future code reviews more consistent and personalized to your project's style. Click here to enable Memory in your admin console.

Using Gemini Code Assist

The full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips.

Invoking Gemini

You can request assistance from Gemini at any point by creating a comment using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands on the current page.

Feature	Command	Description
Code Review	`/gemini review`	Performs a code review for the current pull request in its current state.
Pull Request Summary	`/gemini summary`	Provides a summary of the current pull request in its current state.
Comment	@gemini-code-assist	Responds in comments when explicitly tagged, both in pull request comments and review comments.
Help	`/gemini help`	Displays a list of available commands.

Customization

To customize Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a .gemini/ folder in the base of the repository. Detailed instructions can be found here.

Limitations & Feedback

Gemini Code Assist may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counter productive. You can react with 👍 and 👎 on @gemini-code-assist comments. If you're interested in giving your feedback about your experience with Gemini Code Assist for Github and other Google products, sign up here.

You can also get AI-powered code generation, chat, as well as code reviews directly in the IDE at no cost with the Gemini Code Assist IDE Extension.

Review the Privacy Notices, Generative AI Prohibited Use Policy, Terms of Service, and learn how to configure Gemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution. ↩

gemini-code-assist

Code Review

This pull request introduces dynamic LoRA request switching for vLLM-based fast inference. The implementation correctly detects changes in LoRA weights and reloads them by patching the fast_generate and fast_generate_batches methods. My review includes suggestions to improve robustness by adding logging to exception handling, and to improve code style and performance by moving inline imports and removing unnecessary pass statements.

gemini-code-assist · 2026-01-26T04:10:50Z

unsloth/models/llama.py

+
+                        from unsloth_zoo.vllm_utils import load_lora
+
+                        self._unsloth_cached_lora_request = load_lora(


The try...except Exception as e: block is too broad and silently fails by returning None. This can hide important errors during LoRA loading, making it difficult to debug why adapters might not be applying correctly. It's recommended to at least log the exception to provide visibility into potential issues.

Suggested change

self._unsloth_cached_lora_request = load_lora(

print(f"Unsloth: Failed to get LoRA request: {e}")

return None

gemini-code-assist · 2026-01-26T04:10:50Z

unsloth/models/llama.py

+                    )
+
+                    if not hasattr(self, "_unsloth_lora_id"):
+                        self._unsloth_lora_id = 0


The tempfile module is part of the standard library and should be imported at the top of the file for better code readability and to follow standard Python conventions. Inline imports are generally discouraged for non-optional, standard library modules.

gemini-code-assist · 2026-01-26T04:10:50Z

unsloth/models/llama.py

+                        self._unsloth_lora_temp_dir = tempfile.TemporaryDirectory()
+
+                    # Update if changed
+                    if (


This inline import of load_lora from unsloth_zoo.vllm_utils is inside a conditional block that is executed frequently if LoRA weights change. To avoid the overhead of re-importing the module repeatedly, it's better to move this import statement to a higher scope, for example, at the beginning of the _get_lora_request function, before the conditional check.

gemini-code-assist · 2026-01-26T04:10:50Z

unsloth/models/llama.py

+                        from unsloth_zoo.vllm_utils import load_lora
+
+                        self._unsloth_cached_lora_request = load_lora(
+                            self,


This pass statement is unnecessary and can be removed for cleaner code.

gemini-code-assist · 2026-01-26T04:10:50Z

unsloth/models/llama.py

+                        )
+                    return self._unsloth_cached_lora_request
+                except Exception as e:
+                    return None


This pass statement is unnecessary and can be removed for cleaner code.

gemini-code-assist · 2026-01-26T04:10:50Z

unsloth/models/llama.py

+            def _fast_generate_patched(self, *args, **kwargs):
+                if "lora_request" not in kwargs:
+                    req = _get_lora_request(self)
+                    if req is not None:


This pass statement is unnecessary and can be removed for cleaner code.

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: a346034817

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 · 2026-01-26T04:11:35Z

unsloth/models/llama.py

+                    # Update if changed
+                    if (
+                        current_version != self._unsloth_lora_version
+                        or self._unsloth_cached_lora_request is None
+                    ):


Invalidate LoRA cache when active adapter changes

The cached lora_request is only refreshed when the sum of trainable parameter versions changes. Switching adapters via PEFT (e.g., model.set_adapter(...) / active_adapters updates) typically does not mutate parameter tensors, so current_version stays the same and the cached request is reused. In that scenario, fast_generate keeps sending the old adapter’s LoRA to vLLM, producing outputs for the wrong adapter. Consider including the active adapter name(s) (or a hash of relevant PEFT config) in the cache key so adapter switches invalidate the request.

Useful? React with 👍 / 👎.

dynamically switch lora request

46a821d

Datta0 mentioned this pull request Jan 26, 2026

[vllm] refactor lora request handling unslothai/unsloth-zoo#445

Closed

[pre-commit.ci] auto fixes from pre-commit.com hooks

a346034

for more information, see https://pre-commit.ci

gemini-code-assist bot reviewed Jan 26, 2026

View reviewed changes

chatgpt-codex-connector bot reviewed Jan 26, 2026

View reviewed changes

Merge branch 'nightly' into vllm_lora_req

d36f482

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Comments

[vllm] dynamically switch lora request#3933

[vllm] dynamically switch lora request#3933
Datta0 wants to merge 3 commits intounslothai:nightlyfrom
Datta0:vllm_lora_req

Datta0 commented Jan 26, 2026

Uh oh!

gemini-code-assist bot commented Jan 26, 2026

Uh oh!

gemini-code-assist bot left a comment

Uh oh!

gemini-code-assist bot Jan 26, 2026

Uh oh!

gemini-code-assist bot Jan 26, 2026

Uh oh!

gemini-code-assist bot Jan 26, 2026

Uh oh!

gemini-code-assist bot Jan 26, 2026

Uh oh!

gemini-code-assist bot Jan 26, 2026

Uh oh!

gemini-code-assist bot Jan 26, 2026

Uh oh!

chatgpt-codex-connector bot left a comment

Uh oh!

chatgpt-codex-connector bot Jan 26, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant


		from unsloth_zoo.vllm_utils import load_lora

		self._unsloth_cached_lora_request = load_lora(

	self._unsloth_cached_lora_request = load_lora(
	print(f"Unsloth: Failed to get LoRA request: {e}")
	return None

Uh oh!

Comments

Conversation

Datta0 commented Jan 26, 2026

Uh oh!

gemini-code-assist bot commented Jan 26, 2026

Summary of Changes

Highlights

Footnotes

Uh oh!

gemini-code-assist bot left a comment

Choose a reason for hiding this comment

Code Review

Uh oh!

gemini-code-assist bot Jan 26, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist bot Jan 26, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist bot Jan 26, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist bot Jan 26, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist bot Jan 26, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist bot Jan 26, 2026

Choose a reason for hiding this comment

Uh oh!

chatgpt-codex-connector bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

chatgpt-codex-connector bot Jan 26, 2026

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant