[support] vit fa support cu_seqlens and max_seqlens #953
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 significantly enhances the performance and efficiency of attention mechanisms within the Qwen2-VL models. By introducing explicit support for variable-length sequences via cu_seqlens and max_seqlens in the Flash Attention implementation, it optimizes memory usage and computation, particularly beneficial for batch processing in multimodal contexts.
Highlights
- Flash Attention Enhancement: Implemented support for
cu_seqlens(cumulative sequence lengths) andmax_seqlens(maximum sequence length) within the custom Triton Flash Attention kernel. This enables efficient processing of variable-length sequences in attention computations. - Attention Module Refactoring: The
Qwen2_5_VLVisionAttentionandQwen2VLVisionAttentionmodules have been updated to exclusively utilize the enhancedflash_attention_fwdfunction. This change removes older, less optimized attention implementations (eager and SDPA) and hardcodes the use of Flash Attention. - Codebase Integration: The updated Flash Attention implementation has been integrated across the
qwen2_5_vlandqwen2_vlvisual model components, streamlining the attention computation pipeline and improving performance for multimodal models.
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 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 refactors the visual attention mechanism in qwen2_5_vl and qwen2_vl models to use a unified flash attention implementation, enhancing performance. It introduces changes to a Triton kernel in flashattention_nopad.py to support variable sequence lengths via cu_seqlens and max_seqlens. The changes simplify the model code by removing alternative attention implementations. The implementation looks solid, with suggestions to improve code clarity.
| v = v.transpose(0, 1) | ||
| attn_output = F.scaled_dot_product_attention(q, k, v, attention_mask, dropout_p=0.0) | ||
| attn_output = attn_output.transpose(0, 1) | ||
| attn_output = torch.empty_like(q, dtype=q.dtype, device=q.device) |
There was a problem hiding this comment.
The arguments dtype=q.dtype and device=q.device are redundant when using torch.empty_like(q). The torch.empty_like function already creates a tensor with the same dtype and device as the input tensor q. Simplifying this line improves readability.
attn_output = torch.empty_like(q)| seq_start = tl.load(cu_seqlens + cur_batch).to(tl.int32) | ||
| seq_end = tl.load(cu_seqlens + cur_batch + 1).to(tl.int32) | ||
| seq_len = seq_end - seq_start |
There was a problem hiding this comment.
Consider using a different variable name for the per-batch-item sequence length to avoid shadowing the kernel parameter seq_len. This improves code clarity.
| attn_output = F.scaled_dot_product_attention(q, k, v, attention_mask, dropout_p=0.0) | ||
| attn_output = attn_output.transpose(0, 1) | ||
| max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max().item() | ||
| attn_output = torch.empty_like(q) |
There was a problem hiding this comment.
The arguments dtype=q.dtype and device=q.device are redundant when using torch.empty_like(q). The torch.empty_like function already creates a tensor with the same dtype and device as the input tensor q. Simplifying this line improves readability.
attn_output = torch.empty_like(q)
hiworldwzj
changed the title
hiworldwzj
changed the title
No description provided.