-
Notifications
You must be signed in to change notification settings - Fork 2.8k
[fix] fix dsv3 weight loader tqdm and simplify shared experts fusion #7181
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Conversation
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Summary of Changes
Hello @Alcanderian, 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 issue with the DeepseekV3 weight loading process where the progress bar was inaccurate due to the weight iterator being prematurely consumed. It also simplifies the internal logic for handling and fusing shared experts by processing them directly within the loading loop.
Highlights
- Fix tqdm for weight loading: The progress bar for loading DeepseekV3 weights was misleading because the async iterator was converted to a list, causing the bar to complete before loading finished. This PR fixes that by processing weights while iterating.
- Simplify shared experts fusion: The logic for fusing shared experts has been simplified. Instead of collecting all weights into a list and then manipulating it, shared expert weights are now renamed directly during the iteration over the weight stream.
- Code Refactor: Removed a large block of code that handled the old list-based weight manipulation for shared experts.
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.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Code Review
This pull request effectively addresses the misleading tqdm
progress bar issue during dsv3 weight loading by simplifying the shared expert fusion logic. The previous approach of converting the weight iterator to a list caused the progress bar to complete prematurely. The new method of renaming shared expert weights on-the-fly during iteration is much cleaner, more efficient, and directly resolves the tqdm
problem.
The refactoring of deep_gemm_grouped_gemm_nt_f8f8bf16_masked
to use deep_gemm_wrapper
is a minor change and seems appropriate for centralizing JIT logic.
Overall, the changes improve code maintainability and correctness of the loading progress indication. I've made one minor suggestion regarding logging consistency.
@@ -2128,9 +2033,19 @@ def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]], is_nextn=Fal | |||
"hnorm", | |||
] | |||
|
|||
if self.num_fused_shared_experts > 0: | |||
assert self.num_fused_shared_experts == 1 |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
DeepSeek-Coder-V2-Lite-Instruct has 2 shared experts, which we also support—this assertion is unnecessary.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
DeepSeek-Coder-V2-Lite-Instruct has 2 shared experts, which we also support—this assertion is unnecessary.
DeepseekV2ForCausalLM will be disabled to use shared expert fusion. I will add or self.config.n_shared_expert != 1
at https://github.com/sgl-project/sglang/pull/7180/files#diff-5b9e34dd492bd8a14702a18b594721091092276fad1cf8736fba6ef1f33c1b04R1722
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Good job! Creating a new list is indeed unnecessary, and I believe this is partly why some users encounter hangs when cloning shared experts.
Motivation
Using
weights = [w for w in weights_list if w[0] not in names_to_remove]
to updateweights
will change it from async weight iterator to a list. Thats will make the tqdm looks very fast but the weight are still loading in the backround. Making the progress bar confusing.acc
after: from 11:54:08 to 11:54:59
before: from 12:00:27 to 12:01:19
Modifications
Checklist