fix ulysses #5501
fix ulysses #5501
Conversation
There was a problem hiding this comment.
Summary of Changes
Hello @tastelikefeet, 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 a bug within the Ulysses sequence parallel training module. The core change involves modifying how causal masks are generated and applied. Specifically, it ensures that input embeddings and cache positions are correctly managed for sequence parallel operations, which is crucial for the proper functioning of the training process.
Highlights
- Causal Mask Adjustment: A new create_causal_mask function has been implemented within ulysses.py to correctly reshape input_embeds and adjust cache_position for sequence parallel processing.
- Masking Utility Override: The original masking_utils.create_causal_mask is now temporarily stored and then replaced with the new, modified version to ensure the correct mask generation logic is applied during Ulysses training.
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. 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
-
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 patch to create_causal_mask in ulysses.py to better handle sequence parallelism. While the core logic of adjusting input_embeds and cache_position is sound, I've identified a critical issue with the monkey-patching implementation that could lead to infinite recursion. Additionally, I've suggested a performance optimization to use torch.empty instead of torch.ones for creating a dummy tensor, as its values are not used. The suggested code change addresses both of these points.
| def create_causal_mask(config, input_embeds, attention_mask, cache_position, *args, **kwargs): | ||
| input_embeds = torch.ones( | ||
| (input_embeds.shape[0], input_embeds.shape[1] * self.sp_world_size, input_embeds.shape[2]), | ||
| dtype=input_embeds.dtype, | ||
| device=input_embeds.device) | ||
| cache_position = torch.arange(0, input_embeds.shape[1], device=input_embeds.device) | ||
| return masking_utils.origin_create_causal_mask(config, input_embeds, attention_mask, cache_position, | ||
| *args, **kwargs) | ||
|
|
||
| masking_utils.origin_create_causal_mask = masking_utils.create_causal_mask | ||
| masking_utils.create_causal_mask = create_causal_mask |
There was a problem hiding this comment.
This monkey-patch has two potential issues:
- Idempotency: The patch is not idempotent. If
init_sequence_parallelis called multiple times (e.g., from differentUlyssesinstances), it will wrapcreate_causal_maskrepeatedly, leading tomasking_utils.origin_create_causal_maskpointing to an already-patched function. This would cause an infinite recursion, which is a critical bug. - Efficiency:
torch.onesis used to create a dummyinput_embedstensor. However, the underlyingcreate_causal_maskfunction only needs the tensor's metadata (shape, dtype, device), not its values. Usingtorch.emptyis more efficient as it avoids unnecessary memory initialization.
I've combined fixes for both issues in the suggestion below.
| def create_causal_mask(config, input_embeds, attention_mask, cache_position, *args, **kwargs): | |
| input_embeds = torch.ones( | |
| (input_embeds.shape[0], input_embeds.shape[1] * self.sp_world_size, input_embeds.shape[2]), | |
| dtype=input_embeds.dtype, | |
| device=input_embeds.device) | |
| cache_position = torch.arange(0, input_embeds.shape[1], device=input_embeds.device) | |
| return masking_utils.origin_create_causal_mask(config, input_embeds, attention_mask, cache_position, | |
| *args, **kwargs) | |
| masking_utils.origin_create_causal_mask = masking_utils.create_causal_mask | |
| masking_utils.create_causal_mask = create_causal_mask | |
| if not hasattr(masking_utils, 'origin_create_causal_mask'): | |
| def create_causal_mask(config, input_embeds, attention_mask, cache_position, *args, **kwargs): | |
| input_embeds = torch.empty( | |
| (input_embeds.shape[0], input_embeds.shape[1] * self.sp_world_size, input_embeds.shape[2]), | |
| dtype=input_embeds.dtype, | |
| device=input_embeds.device) | |
| cache_position = torch.arange(0, input_embeds.shape[1], device=input_embeds.device) | |
| return masking_utils.origin_create_causal_mask(config, input_embeds, attention_mask, cache_position, | |
| *args, **kwargs) | |
| masking_utils.origin_create_causal_mask = masking_utils.create_causal_mask | |
| masking_utils.create_causal_mask = create_causal_mask |
PR type
PR information
Write the detail information belongs to this PR.
Experiment results
Paste your experiment result here(if needed).