Dp balancer #991
Dp balancer #991
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 @shihaobai, 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 Data Parallelism (DP) request handling by introducing a more intelligent load balancing mechanism. The core change is the implementation of a new DP balancer designed to optimize GPU utilization by distributing requests more evenly across DP ranks, thereby preventing idle resources. This refactoring centralizes the request assignment logic, making the system more robust and efficient in managing concurrent requests in DP mode.
Highlights
- Enhanced Data Parallelism (DP) Load Balancing: Introduced a new
DpBalancerForPdclass to intelligently distribute incoming requests across Data Parallel (DP) ranks, aiming to balance batch sizes and prevent GPU resource waste by minimizing idle ranks. - Configurable DP Balancer: Added a
--dp_balancercommand-line argument, allowing users to specify the DP balancing strategy. The defaultround_robinoption now utilizes the newDpBalancerForPdimplementation. - Centralized Request Assignment Logic: Refactored the
DpQueueto delegate the assignment of requests to specific DP ranks to the newdp_balancer. Requests are now temporarily queued and then assigned by the balancer based on the current load across DP ranks during batch generation. - Increased Default Concurrency: The default value for the
--running_max_req_sizeargument was increased from 1000 to 2048, potentially accommodating higher throughput and larger concurrent request loads with the improved balancing.
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 or fill out our survey 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 configure Gemini 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
This pull request introduces a data parallelism (DP) balancer to manage request distribution across DP ranks. The existing round-robin logic is refactored into a RoundRobinDpBalancer class, and a new DpBalancerForPd is added for prefill/decode modes. The main issue is the incorrect balancer selection in the factory function. Additionally, there are some minor issues related to code style and exception handling.
| if args.dp_balancer == "round_robin": | ||
| return DpBalancerForPd(dp_size_in_node, inner_queues) | ||
| if args.run_mode in ["prefill", "decode"]: | ||
| return DpBalancerForPd(dp_size_in_node, inner_queues) |
There was a problem hiding this comment.
The get_dp_balancer function returns DpBalancerForPd when args.dp_balancer is "round_robin". This seems incorrect as it should return RoundRobinDpBalancer in this case. This could lead to unexpected behavior. Consider swapping the return values for the round_robin case to ensure the correct balancer is used.
| if args.dp_balancer == "round_robin": | |
| return DpBalancerForPd(dp_size_in_node, inner_queues) | |
| if args.run_mode in ["prefill", "decode"]: | |
| return DpBalancerForPd(dp_size_in_node, inner_queues) | |
| def get_dp_balancer(args, dp_size_in_node: int, inner_queues: List[BaseQueue]): | |
| if args.dp_balancer == "round_robin": | |
| return RoundRobinDpBalancer(dp_size_in_node, inner_queues) | |
| if args.run_mode in ["prefill", "decode"]: | |
| return DpBalancerForPd(dp_size_in_node, inner_queues) |
| suggested_dp_index = select_dp_indexes[0] | ||
| else: | ||
| # if multiple dp ranks have the same minimum load, randomly select one | ||
| import random |
There was a problem hiding this comment.
The import random statement is located inside the assign_reqs_to_dp method. It's better to move this import statement to the top of the file for better code organization and readability.
| logger.error(f"generate new batch failed: {e}") | ||
| raise e |
There was a problem hiding this comment.
When an exception occurs during batch generation, using raise e can obscure the original stack trace. Using a bare raise will preserve the original traceback, making debugging easier.
| logger.error(f"generate new batch failed: {e}") | |
| raise e | |
| except Exception as e: | |
| logger.error(f"generate new batch failed: {e}") | |
| raise |
No description provided.