fix n-nodes inconsistencies when gpu G-MEM size differ#931
Conversation
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
There was a problem hiding this comment.
Summary of Changes
Hello @SangChengC, 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 inconsistencies in memory allocation across multiple nodes or GPUs that may have different amounts of available memory. By using the minimum available memory size across all devices in a distributed setup, the change ensures stable operation and prevents errors that could arise from uneven memory capacities.
Highlights
- Distributed Memory Consistency: Ensures that when running across multiple GPUs or nodes with differing memory sizes, the allocated KV cache size is determined by the GPU with the minimum available memory. This prevents potential out-of-memory issues on devices with less VRAM.
- Memory Profiling Adjustment: Modifies the
profile_sizemethod inMemoryManagerto incorporate a distributedall_reduceoperation (usingMIN) after calculating the initial available memory size based on the local GPU.
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 is currently in preview and 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 to provide feedback.
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
-
Review the Privacy Notices, Generative AI Prohibited Use Policy, Terms of Service, and learn how to configureGemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution. ↩
There was a problem hiding this comment.
Code Review
The pull request addresses potential inconsistencies in memory allocation size across nodes with differing GPU memory by ensuring that the minimum calculated memory size is used across all ranks. The core logic introduced correctly uses dist.all_reduce with ReduceOp.MIN for this purpose. However, the get_available_gpu_memory function already performs a similar reduction, making the added all_reduce on self.size redundant. Removing this redundant step would slightly improve efficiency without affecting correctness.
| if world_size > 1: | ||
| tensor = torch.tensor(self.size, dtype=torch.int64, device=f"cuda:{get_current_device_id()}") | ||
| dist.all_reduce(tensor, op=dist.ReduceOp.MIN) | ||
| self.size = tensor.item() |
There was a problem hiding this comment.
The get_available_gpu_memory function already performs an all_reduce with ReduceOp.MIN to determine the minimum available memory across all ranks. Since self.size is calculated directly from this available_memory, it will already represent the minimum possible size across all ranks. Therefore, the subsequent all_reduce on self.size is redundant.
2 participants
No description provided.