Add vLLM fallback and GRPO completion normalization by danielhanchen · Pull Request #3958 · unslothai/unsloth

danielhanchen · 2026-02-02T02:52:32Z

Summary

add safe vLLM import guard for guided decoding patch
normalize GRPO completions and retry reward funcs on string vs dict mismatch
wrap vLLM fast_generate with HF fallback and disable vLLM inference for FP8 models
attach HF model and tokenizer to vLLM engine for fallback
add DeepSeek v2 MoE alias and Qwen VL compatibility helpers
guard RL trainer compilation to fall back to original trainer on failure

Testing

python -m py_compile unsloth/import_fixes.py unsloth/init.py unsloth/models/rl_replacements.py unsloth/models/_utils.py unsloth/models/llama.py unsloth/models/vision.py unsloth/models/rl.py

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

gemini-code-assist · 2026-02-02T02:52:51Z

Summary of Changes

Hello @danielhanchen, 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 stability and compatibility of the Unsloth library, particularly concerning vLLM integration and Reinforcement Learning (RL) trainers. It introduces mechanisms to gracefully handle potential failures in vLLM inference and improves the flexibility of reward function processing in GRPO. Additionally, it addresses specific model compatibility issues, ensuring a smoother user experience across a wider range of models and configurations.

Highlights

vLLM Fallback Mechanism: Implemented a robust fallback system for vLLM's fast_generate to seamlessly switch to Hugging Face's generate if vLLM encounters an error or is explicitly disabled (e.g., for FP8 models). This ensures continued inference capability even when vLLM is not optimal or fails.
GRPO Completion Normalization: Enhanced GRPO reward functions to normalize completions, converting non-string outputs (like dictionaries or lists) into a consistent string format. This includes a retry mechanism to handle reward functions that might expect dictionary-based completions, improving compatibility and robustness.
Improved vLLM Import Guarding: Strengthened the import guards for vLLM's GuidedDecodingParams patch, making it more resilient to various import errors and providing clearer warnings when the patch cannot be applied.
Model Compatibility Fixes: Added compatibility fixes for specific models, including an alias for DeepSeek v2 MoE and adjustments for Qwen VL models regarding max_pixels and tie_word_embeddings.
RL Trainer Compilation Safeguard: Introduced a safeguard for RL trainer compilation, allowing the system to fall back to the original trainer if the custom compilation process fails, preventing crashes and improving stability.

🧠 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. ↩

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 697b5005c1

ℹ️ About Codex in GitHub

Your team has set up Codex to 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 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

chatgpt-codex-connector · 2026-02-02T02:54:50Z

unsloth/models/_utils.py

+def _vllm_fallback_to_hf_generate(model, args, kwargs):
+    tokenizer = getattr(model, "_saved_temp_tokenizer", None)
+    hf_kwargs = dict(kwargs)
+    sampling_params = hf_kwargs.pop("sampling_params", None)
+    hf_kwargs.pop("lora_request", None)
+


Map/remove vLLM prompts kw in HF fallback

When vLLM falls back to HF, _vllm_fallback_to_hf_generate copies kwargs verbatim and only removes sampling_params/lora_request. If the caller used the common vLLM signature fast_generate(prompts=...), the fallback passes an unexpected prompts kw into model.generate, which raises a TypeError because HF generate doesn’t accept that argument. This only appears when vLLM fails or inference is disabled, but it will reliably break those fallback calls unless prompts is converted into tokenized input_ids or removed.

Useful? React with 👍 / 👎.

gemini-code-assist

Code Review

This pull request introduces several valuable improvements, notably a robust fallback mechanism for vLLM inference and normalization for GRPO completions. The added compatibility fixes for various models and the defensive programming approach to imports and dynamic code generation enhance the library's stability.

I've identified a couple of areas for improvement:

The vLLM fallback logic has a hardcoded CUDA device, which should be generalized for other hardware.
A logging statement in the RL trainer patch could be updated to use the project's logger for consistency.

Overall, these are great changes that make Unsloth more resilient and user-friendly.

gemini-code-assist · 2026-02-02T02:55:09Z

unsloth/models/_utils.py

+                return_tensors = "pt",
+                padding = True,
+            )
+            device = getattr(model, "device", None) or "cuda"


The fallback device is hardcoded to "cuda". This might cause issues on non-NVIDIA hardware like AMD (ROCm/hip) or Intel (XPU) GPUs. It's better to use the dynamically determined DEVICE_TYPE_TORCH for broader compatibility.

Suggested change

device = getattr(model, "device", None) or "cuda"

device = getattr(model, "device", None) or DEVICE_TYPE_TORCH

gemini-code-assist · 2026-02-02T02:55:09Z

unsloth/models/rl.py

+        if os.environ.get("UNSLOTH_LOGGING_ENABLED", "0") == "1":
+            print(


For consistency with the rest of the codebase, it's better to use the configured logger for logging messages instead of print. This allows users to control log levels and formatting centrally.

Suggested change

if os.environ.get("UNSLOTH_LOGGING_ENABLED", "0") == "1":

print(

if os.environ.get("UNSLOTH_LOGGING_ENABLED", "0") == "1":

logger.warning(f"Unsloth: Failed to compile {RLTrainer_name} ({exc}), falling back to original trainer.")

Datta0 · 2026-02-02T04:03:22Z

unsloth/models/_utils.py

    return _fast_generate_wrapper
+
+
+def make_vllm_fast_generate_wrapper(model, vllm_generate):


In what cases is the fallback triggered ? What case did you observe vllm inference failing?

Datta0 · 2026-02-02T04:11:51Z

unsloth/models/_utils.py

+                    "Unsloth: vLLM fast_generate failed and no tokenizer was cached for HF fallback."
+                )
+            inputs = tokenizer(
+                first_arg,


This can go wrong very easily. if we're passing something to tokenizer, we better check what arg it is at least?

Datta0 · 2026-02-02T04:12:55Z

unsloth/models/rl_replacements.py

+    )
+
+    function = function.replace(
+        '        reward_kwargs["trainer_state"] = self.state\n',


Why are we tracking this?

Datta0 · 2026-02-02T04:14:08Z

unsloth/models/rl_replacements.py

+        "                return ''.join(_unsloth_completion_to_text(item) for item in completion)\n"
+        "            return str(completion)\n"
+        "        completion_texts = [_unsloth_completion_to_text(c) for c in completions]\n"
+        "        completions_are_text = all(isinstance(c, str) for c in completions)\n",


What else is a possibility? When are tokens not text for LLM/VLM GRPO?

danielhanchen and others added 2 commits February 2, 2026 02:51

Add safe fallbacks for vllm and GRPO edge cases

0eb6a7f

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

697b500

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

chatgpt-codex-connector bot reviewed Feb 2, 2026

View reviewed changes

gemini-code-assist bot reviewed Feb 2, 2026

View reviewed changes

Datta0 reviewed Feb 2, 2026

View reviewed changes

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Comments

Add vLLM fallback and GRPO completion normalization#3958

Add vLLM fallback and GRPO completion normalization#3958
danielhanchen wants to merge 2 commits intomainfrom
fix/compat-guards

danielhanchen commented Feb 2, 2026

Uh oh!

gemini-code-assist bot commented Feb 2, 2026

Uh oh!

chatgpt-codex-connector bot left a comment

Uh oh!

chatgpt-codex-connector bot Feb 2, 2026

Uh oh!

gemini-code-assist bot left a comment

Uh oh!

gemini-code-assist bot Feb 2, 2026

Uh oh!

gemini-code-assist bot Feb 2, 2026

Uh oh!

Datta0 Feb 2, 2026

Uh oh!

Datta0 Feb 2, 2026

Uh oh!

Datta0 Feb 2, 2026

Uh oh!

Datta0 Feb 2, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

	device = getattr(model, "device", None) or "cuda"
	device = getattr(model, "device", None) or DEVICE_TYPE_TORCH

		if os.environ.get("UNSLOTH_LOGGING_ENABLED", "0") == "1":
		print(

		return _fast_generate_wrapper


		def make_vllm_fast_generate_wrapper(model, vllm_generate):

Uh oh!

Comments

Conversation

danielhanchen commented Feb 2, 2026

Summary

Testing

Uh oh!

gemini-code-assist bot commented Feb 2, 2026

Summary of Changes

Highlights

Footnotes

Uh oh!

chatgpt-codex-connector bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

chatgpt-codex-connector bot Feb 2, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist bot left a comment

Choose a reason for hiding this comment

Code Review

Uh oh!

gemini-code-assist bot Feb 2, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist bot Feb 2, 2026

Choose a reason for hiding this comment

Uh oh!

Datta0 Feb 2, 2026

Choose a reason for hiding this comment

Uh oh!

Datta0 Feb 2, 2026

Choose a reason for hiding this comment

Uh oh!

Datta0 Feb 2, 2026

Choose a reason for hiding this comment

Uh oh!

Datta0 Feb 2, 2026

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants