-
Notifications
You must be signed in to change notification settings - Fork 25
Add HummingbirdXT backend integration #533
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.
Pull request overview
This PR integrates HummingbirdXT, a DiT-based model distilled from Wan2.2, into the Primus framework for video generation with Text-To-Video (T2V) capabilities. HummingbirdXT achieves high-quality video generation with just 3 inference steps through distillation, significantly improving performance.
Changes:
- Added HummingbirdXT as a git submodule in
third_party/ - Implemented backend adapter, trainers, and argument builder for HummingbirdXT integration
- Created configuration files for post-training with Wan2.2 model
- Added installation scripts and requirements for HummingbirdXT dependencies
Reviewed changes
Copilot reviewed 13 out of 13 changed files in this pull request and generated 1 comment.
Show a summary per file
| File | Description |
|---|---|
| third_party/HummingbirdXT | Adds HummingbirdXT submodule reference |
| .gitmodules | Registers HummingbirdXT submodule with repository URL |
| runner/helpers/hooks/train/posttrain/hummingbirdxt/requirements-hummingbirdxt.txt | Lists Python dependencies for HummingbirdXT |
| runner/helpers/hooks/train/posttrain/hummingbirdxt/00_install_requirements.sh | Shell script to install HummingbirdXT dependencies |
| primus/backends/hummingbirdxt/init.py | Registers HummingbirdXT backend and trainer with Primus |
| primus/backends/hummingbirdxt/hummingbirdxt_adapter.py | Implements backend adapter for configuration conversion and trainer loading |
| primus/backends/hummingbirdxt/argument_builder.py | Converts Primus configs to HummingbirdXT argument format |
| primus/backends/hummingbirdxt/hummingbirdxt_base_trainer.py | Base trainer class with version detection methods |
| primus/backends/hummingbirdxt/hummingbirdxt_posttrain_trainer.py | Post-training trainer implementation for Wan22 score distillation |
| primus/configs/modules/hummingbirdxt/post_trainer.yaml | Module configuration marking HummingbirdXT as trainable |
| primus/configs/models/hummingbirdxt/wan22.yaml | Complete model configuration with training and inference parameters |
| examples/hummingbirdxt/configs/wan22_posttrain.yaml | Example configuration for Wan2.2 post-training |
| primus/backends/megatron_bridge/argument_builder.py | Updates comments to reference MegatronBridge instead of TorchTitan |
69d19a2 to
ceb736d
Compare
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.
Pull request overview
Copilot reviewed 13 out of 13 changed files in this pull request and generated 1 comment.
| batch_size: 1 | ||
| ema_weight: 0.99 | ||
| ema_start_step: 200 | ||
| log_iters: 1000 |
Copilot
AI
Feb 3, 2026
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.
The negative_prompt field contains Chinese text without any accompanying English translation or explanation. Consider adding a comment explaining what this negative prompt controls for better maintainability by English-speaking developers.
| log_iters: 1000 | |
| log_iters: 1000 | |
| # Negative prompt in Chinese listing undesirable visual qualities (e.g., oversaturated colors, overexposed/static/blurred frames, subtitles/painting-like style, gray/low-quality images, JPEG artifacts, ugly or deformed/maimed bodies, extra or fused fingers/limbs, crowded or messy backgrounds, and people walking backwards). |
Add HummingbirdXT backend integration into Primus.
HummingbirdXT is a DiT-based model distilled from Wan2.2, which can be used for video generation and currently has the Text-To-Video (T2V) function integrated to Primus.
After distillation, HummingbirdXT can complete inference and generate high-quality videos with just 3 steps, significantly improving inference speed.
The accuracy has already been verified; please see this DOCUMENT for details.
sample launch command: