Bug fix: eval OOM due to deepcopy of torch model

[hellopahe]

PR type

• Bug Fix

Issues and fixes

• When evalscope dumps the taskConfig, it deep copies the config, which triggers cloning large model tensors, causing OOM.
• By passing model_ref instead of torch model, this PR prevent evalscope's deepcopying behavior.

Reproduce

import os, json
from swift.llm.train.sft import sft_main

def main():
    os.environ['CUDA_VISIBLE_DEVICES'] = '0'
    args = [
        '--model', 'Qwen/Qwen2.5-0.5B-Instruct',
        '--model_type', 'qwen2_5',
        '--train_type', 'lora',
        '--dataset', 'swift/self-cognition#10',
        '--torch_dtype', 'float32',
        '--num_train_epochs', '3',
        '--per_device_train_batch_size', '1',
        '--learning_rate', '5e-4',
        '--lora_rank', '4',
        '--lora_alpha', '8',
        '--target_modules', 'q_proj',
        '--gradient_accumulation_steps', '1',
        '--eval_steps', '6',
        '--save_steps', '30',
        '--logging_steps', '1',
        '--max_length', '2048',
        '--output_dir', 'dev/output',
        '--warmup_ratio', '0.05',
        '--dataloader_num_workers', '0',
        '--model_author', 'author',
        '--model_name', 'name',
        '--eval_use_evalscope',
        '--per_device_eval_batch_size', '1',
        '--eval_dataset', 'gsm8k',
        '--eval_dataset_args', '{"gsm8k":{"few_shot_num":0}}',
        '--eval_limit', '2',
        '--report_to', 'none',
    ]
    sft_main(args)

if __name__ == '__main__':
    main()

[gemini-code-assist]

Summary of Changes

Hello @hellopahe, 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 addresses an Out-Of-Memory (OOM) issue that occurs when evalscope attempts to deep-copy large torch model tensors during evaluation. The solution involves implementing a lightweight in-process registry to store model references, allowing evalscope to work with a small identifier instead of the entire model object, thereby preventing excessive memory consumption and improving evaluation stability.

Highlights

• Memory Optimization: Introduced a lightweight in-process registry to store torch model objects and templates, preventing Out-Of-Memory (OOM) errors caused by evalscope deep-copying large model tensors.
• Registry API: Added register_eval_model, get_eval_model, and unregister_eval_model functions in swift/llm/eval/utils.py to manage the lifecycle of registered models.
• EvalScope Integration: Modified the _evalscope_eval method to register the current model and template with a unique key, passing only this lightweight model_ref to evalscope's TaskConfig instead of the full torch model object.
• Safe Cleanup: Ensured that registered models are unregistered using a try...finally block, guaranteeing cleanup even if evaluation fails.

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 in your pull request via creating an issue comment (i.e. comment on the pull request page) using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands.

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 issue 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.

Footnotes

1. 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 addresses an Out-Of-Memory error during evaluation by introducing an in-process registry to avoid deep-copying large model tensors. The approach of passing a model reference instead of the model object is sound. My review focuses on improving the robustness and thread-safety of this new registry mechanism. I've suggested making the registry thread-safe to prevent potential race conditions, ensuring the model keys are unique across different trainer instances, and adding a check to handle cases where a model is not found in the registry to prevent crashes.