- Add DeepSpeed ZeRO-3 configuration support
- Optimize memory usage during training
- Rename training scripts to reflect ZeRO usage
- Update related configuration files and trainers
- Add SFT (Supervised Fine-Tuning) trainers for all model variants:
- CogVideoX I2V and T2V
- CogVideoX-1.5 I2V and T2V
- Add DeepSpeed ZeRO configuration files:
- ZeRO-2 with and without CPU offload
- ZeRO-3 with and without CPU offload
- Add base accelerate config for distributed training
- Update trainer.py to support SFT training mode
This enables full-parameter fine-tuning with memory-efficient distributed training using DeepSpeed ZeRO optimization.
When loading videos with fewer frames than max_num_frames, repeat the last
frame to reach the required length instead of failing. This ensures consistent
tensor dimensions across the dataset while preserving as much original video
content as possible.
Add a table in README files showing hardware requirements for training
different CogVideoX models, including:
- Memory requirements for each model variant
- Supported training types (LoRA)
- Training resolutions
- Mixed precision settings
Updated in all language versions (EN/ZH/JA).
Before precomputing the latent cache and text embeddings, cast the VAE and
text encoder to the target training dtype (fp16/bf16) instead of keeping them
in fp32. This reduces memory usage during the precomputation phase.
The change occurs in prepare_dataset() where the models are moved to device
and cast to weight_dtype before being used to generate the cache.
When training i2v models without specifying image_column, automatically extract
and use first frames from training videos as conditioning images. This includes:
- Add load_images_from_videos() utility function to extract and cache first frames
- Update BaseI2VDataset to support auto-extraction when image_column is None
- Add validation and warning message in Args schema for i2v without image_column
The first frames are extracted once and cached to avoid repeated video loading.
- Add pipe.remove_all_hooks() after validation to prevent memory leaks
- Clean up validation pipeline properly to avoid potential issues in subsequent training steps
This change enables caching of prompt embeddings in the CogVideoX text-to-video
LoRA trainer, which can improve training efficiency by avoiding redundant text
encoding operations.
- Add validation to ensure (frames - 1) is multiple of 8
- Add specific resolution check (480x720) for cogvideox-5b models
- Add error handling for invalid resolution format
- Add caching for prompt embeddings
- Store cached files using safetensors format
- Add cache directory structure under data_root/cache
- Optimize memory usage by moving tensors to CPU after caching
- Add debug logging for cache hits
- Add info logging for cache writes
The caching system helps reduce redundant computation and memory usage during training by:
1. Caching prompt embeddings based on prompt text hash
2. Caching encoded video latents based on video filename
3. Moving tensors to CPU after caching to free GPU memory
