support w4 for qw_vl and disk offload for qw_2511

[gushiqiao]

No description provided.

[gemini-code-assist]

Summary of Changes

Hello @gushiqiao, 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 memory efficiency and operational flexibility of the system, particularly for large language and vision models like Qwen-VL and Qwen2.5. It introduces 4-bit quantization for Qwen-VL, enabling a substantial reduction in GPU memory footprint by offloading specific model components to the CPU. Concurrently, a sophisticated disk offloading and lazy loading mechanism has been implemented for transformer blocks, allowing model weights to be loaded asynchronously from disk only when required during inference. These changes, coupled with refactored device management, aim to optimize resource utilization and broaden the applicability of these models in memory-constrained environments.

Highlights

• Qwen-VL 4-bit Quantization: Implemented support for int4 quantization for the Qwen2.5-VL model, including specific device mapping configurations to offload visual and language model components to the CPU, significantly reducing GPU memory usage.
• Disk Offloading and Lazy Loading: Introduced a comprehensive disk offloading and lazy loading system for transformer blocks, enabling asynchronous prefetching of model weights from disk to CPU buffers and then to CUDA, optimizing memory management during inference.
• Dynamic Module Loading/Unloading: Enhanced memory management by conditionally loading and unloading text encoders, VAE, and transformer models on demand, particularly when lazy loading or module unloading features are enabled, to minimize peak memory consumption.
• Refactored Device Operations: Centralized device-specific operations, such as empty_cache(), by introducing a generic torch_device_module, which improves flexibility and paves the way for broader AI device support beyond CUDA.
• Granular Offloading Control: Refined the offloading manager to provide more granular control over block-level weight prefetching and initialization, ensuring efficient and precise management of model components.

🧠 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.

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 introduces support for 4-bit quantized models for qwen25_vl and enables disk offloading (lazy loading) for model weights, which are great additions for improving performance and reducing memory usage. The implementation is mostly solid, with good handling of model loading based on quantization and offloading configurations. I've identified a couple of areas for improvement: one is a code duplication in model.py that could be refactored for better maintainability, and the other is a potential bug in qwen_image_runner.py related to accessing vae_scale_factor during lazy loading which could lead to an AttributeError. Please see the detailed comments below.

[gemini-code-assist]

There is significant code duplication between this block in _load_ckpt and a similar block in _load_quant_ckpt (lines 186-199) for handling lazy_load logic. This duplication can make future maintenance more difficult and error-prone.

To improve maintainability, consider extracting this logic into a private helper method, for example _get_safetensors_files(self, safetensors_path), which would determine the correct files to load based on self.lazy_load and whether the path is a directory or a file. Both _load_ckpt and _load_quant_ckpt could then call this helper method to get the list of safetensor files.