[Docker] Add radixark/miles:dev, update SGLang version to v0.5.8

[yueming-yuan]

https://hub.docker.com/repository/docker/radixark/miles/tags/dev/sha256-2ec7d0b3e5d6184e5e925b9057ab04aa49619d3e336f87e4520ce3b2128ad938

[gemini-code-assist]

Summary of Changes

Hello @yueming-yuan, 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 updates the SGLang framework to version v0.5.8, bringing a suite of performance improvements, new features, and bug fixes. The changes span across the Docker build environment, core Megatron-LM components, and SGLang's specialized layers, particularly focusing on quantization, Mixture-of-Experts (MoE) routing, and speculative decoding mechanisms. The update aims to provide a more robust and efficient foundation for large language model inference and training.

Highlights

• SGLang Version Update: Updated SGLang to version v0.5.8, incorporating numerous upstream changes and optimizations for enhanced performance and functionality.
• Docker Build Environment: Introduced a new Docker build environment (Dockerfile.dev) for SGLang v0.5.8, including updated dependencies and build configurations for various AI/ML libraries.
• Megatron-LM Patches: Applied extensive patches to Megatron-LM, enhancing checkpoint loading flexibility, introducing fake INT4 quantization support, and refining Transformer layer configurations with new post-attention/MLP layernorms.
• SGLang Core Enhancements: Implemented significant SGLang-specific patches, improving Mixture-of-Experts (MoE) routing, refining quantization methods (especially for Marlin conversion), and optimizing speculative decoding with CUDA graph enhancements.
• Memory Management and CUDA Architecture Support: Enhanced memory management for disaggregated decode/prefill queues and KV cache in SGLang, and improved CUDA architecture detection and support in custom kernels.

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 updates the SGLang version to v0.5.8, introducing a new Dockerfile for development and extensive patches for Megatron-LM and SGLang, including dependency updates, new features, bug fixes, and refactoring. However, a critical Remote Code Execution (RCE) vulnerability has been identified in the Megatron-LM patch due to hardcoding trust_remote_code=True, and a high-severity Broken Access Control issue exists in the SGLang patch where a new API endpoint for weight post-processing lacks authentication. These security concerns must be addressed before merging.

[gemini-code-assist]

Hardcoding trust_remote_code=True in AutoTokenizer.from_pretrained introduces a critical Remote Code Execution (RCE) vulnerability. This allows arbitrary Python code execution from the model repository, enabling an attacker to achieve RCE if an untrusted model is loaded. This is a significant security regression. It is strongly recommended to revert this change or make trust_remote_code a configurable option that is disabled by default, as the previous implementation using a variable was safer.

            trust_remote_code=trust_remote_code,

[gemini-code-assist]

The new API endpoint /post_process_weights lacks the @auth_level decorator used by other sensitive endpoints in the same file (e.g., /update_weight_version). This allows unauthenticated users to trigger model weight post-processing logic, which could lead to Denial of Service (DoS) or unauthorized modification of the model's internal state. Given that other weight-related operations are protected, this endpoint should also require administrative privileges.

[email protected]("/post_process_weights")
+@auth_level(AuthLevel.ADMIN_OPTIONAL)

[gemini-code-assist]

For better Docker image layer caching and to reduce image size, it's recommended to combine apt update and apt install into a single RUN instruction. You should also clean up the apt cache in the same layer to prevent it from being stored in the image.

RUN apt update && apt install -y nvtop rsync dnsutils && rm -rf /var/lib/apt/lists/*