bridge.diffusion.recipes.wan.wan#
Module Contents#
Functions#
Return a pre-training configuration for WAN 1.3B model. |
|
Return a pre-training configuration for WAN 14B model. |
|
Return a fine-tuning configuration for WAN 1.3B model. |
|
Return a fine-tuning configuration for WAN 14B model. |
|
Return a Wan 1.3B pretraining configuration tuned for text-to-image data. |
|
Return a Wan 1.3B pretraining configuration tuned for text-to-video data. |
API#
- bridge.diffusion.recipes.wan.wan.wan_1_3b_pretrain_config() megatron.bridge.training.config.ConfigContainer#
Return a pre-training configuration for WAN 1.3B model.
Default parallelism: TP=1, PP=1, CP=8. Uses mock/synthetic data when dataset.path is not set. To use real data, override via CLI: dataset.path=/path/to/wds
- bridge.diffusion.recipes.wan.wan.wan_14b_pretrain_config() megatron.bridge.training.config.ConfigContainer#
Return a pre-training configuration for WAN 14B model.
Default parallelism: TP=2, PP=1, CP=4, SP=True. Uses mock/synthetic data when dataset.path is not set. To use real data, override via CLI: dataset.path=/path/to/wds
- bridge.diffusion.recipes.wan.wan.wan_1_3b_sft_config(
- pretrained_checkpoint: str | None = None,
Return a fine-tuning configuration for WAN 1.3B model.
Uses the same defaults as wan_1_3b_pretrain_config() and overrides checkpoint to load from pretrained_checkpoint when provided.
- bridge.diffusion.recipes.wan.wan.wan_14b_sft_config(
- pretrained_checkpoint: str | None = None,
Return a fine-tuning configuration for WAN 14B model.
Uses the same defaults as wan_14b_pretrain_config() and overrides checkpoint to load from pretrained_checkpoint when provided.
- bridge.diffusion.recipes.wan.wan.wan_1_3b_text2image_pretrain_config() megatron.bridge.training.config.ConfigContainer#
Return a Wan 1.3B pretraining configuration tuned for text-to-image data.
Wraps wan_1_3b_pretrain_config and overrides sequence length on both the model and the dataset for spatial-only inputs.
- bridge.diffusion.recipes.wan.wan.wan_1_3b_text2video_pretrain_config() megatron.bridge.training.config.ConfigContainer#
Return a Wan 1.3B pretraining configuration tuned for text-to-video data.
Wraps wan_1_3b_pretrain_config and overrides sequence length on both the model and the dataset for spatio-temporal inputs, with context parallelism reduced to 4 to fit the longer sequence.