This repository was archived by the owner on May 20, 2026. It is now read-only.

Adding DiT training and Inference#27

Merged

abhinavg4 merged 30 commits into

mainfrom

dit_debug

Nov 11, 2025

Commits on Nov 10, 2025

first dit commit.
sajadn
committed
inference working with old checkpoint.
sajadn
committed
add on_validation_start step to generate video.
sajadn
committed
EDM loss bug fix.
sajadn
committed
Refactor DIT components: add comments for potential removals and clarify usage of common elements in the codebase.

abhinavg4
authored and
sajadn
committed
Add utility functions for tensor operations and dynamic imports
sajadn
committed
remove unused layers from spec.
sajadn
committed
Remove unused AdaLN and STDiTLayerWithAdaLN classes from dit_layer_spec.py to streamline the codebase.
sajadn
committed
Refactor DIT model files: removed unused functions from `dit_provider.py`, updated import paths in `dit.py`, and adjusted training parameters for `task_encoder_seq_length` and `save_interval` to op…
sajadn
committed
Add README for DiT model setup: include instructions for cloning repositories, dataset location, Docker setup, installation, and running the model on the butterfly dataset.
sajadn
committed
Remove unused files and refactor DIT model components: deleted obsolete files including `decode_cosmos_latent.py`, `attention_wan.py`, `diffusers_vae.py`, `dit_embeddings.py`, `dit_inference.py`, a…
sajadn
committed
Add new components for DiT model:
sajadn
committed
revert THD to SBHD and use remove prints.
sajadn
committed
add support sequence for sequence packing.
sajadn
committed
move files in proper location, uncomment cp related code, cleaning.
sajadn
committed
fix lint errors.
sajadn
committed
fix README lint error.
sajadn
committed
add the missing copyright.
sajadn
committed
removed dependency on nemo CausalVideoTokenizer by implementing a minimal Tokenizer version.
sajadn
committed
remove nemo vae modules.
sajadn
committed
update readme.md
sajadn
committed
remove unused temporal_self_attention attribute from DiTWithAdaLNSubmodules class
sajadn
committed
create an abstract DiffusionTaskEncoder to encompass the sequence packing logic that is shared across models.
sajadn
committed
fix use seq_length instead of packing_buffer_size to set max number of tokens.
sajadn
committed
Add the missing copy write. Minor update to the config.
sajadn
committed
lint still complains about the nemo_vfm file.
sajadn
committed
move data modules to common, move cosmos to common/tokenizer instad of common/models. Make inference compatible with the new data loading.
sajadn
committed
uncomment omega config overwrite.
sajadn
committed
resolve conflict for dit_debug merge with mainlie
sajadn
committed
change import (remote lint complains).
sajadn
committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Adding DiT training and Inference#27

Adding DiT training and Inference#27
abhinavg4 merged 30 commits into
mainfrom
dit_debug

Commits on Nov 10, 2025

first dit commit.

inference working with old checkpoint.

add on_validation_start step to generate video.

EDM loss bug fix.

Refactor DIT components: add comments for potential removals and clarify usage of common elements in the codebase.

Add utility functions for tensor operations and dynamic imports

remove unused layers from spec.

Remove unused AdaLN and STDiTLayerWithAdaLN classes from dit_layer_spec.py to streamline the codebase.

Refactor DIT model files: removed unused functions from `dit_provider.py`, updated import paths in `dit.py`, and adjusted training parameters for `task_encoder_seq_length` and `save_interval` to op…

Add README for DiT model setup: include instructions for cloning repositories, dataset location, Docker setup, installation, and running the model on the butterfly dataset.

Remove unused files and refactor DIT model components: deleted obsolete files including `decode_cosmos_latent.py`, `attention_wan.py`, `diffusers_vae.py`, `dit_embeddings.py`, `dit_inference.py`, a…

Add new components for DiT model:

revert THD to SBHD and use remove prints.

add support sequence for sequence packing.

move files in proper location, uncomment cp related code, cleaning.

fix lint errors.

fix README lint error.

add the missing copyright.

removed dependency on nemo CausalVideoTokenizer by implementing a minimal Tokenizer version.

remove nemo vae modules.

update readme.md

remove unused temporal_self_attention attribute from DiTWithAdaLNSubmodules class

create an abstract DiffusionTaskEncoder to encompass the sequence packing logic that is shared across models.

fix use seq_length instead of packing_buffer_size to set max number of tokens.

Add the missing copy write. Minor update to the config.

lint still complains about the nemo_vfm file.

move data modules to common, move cosmos to common/tokenizer instad of common/models. Make inference compatible with the new data loading.

uncomment omega config overwrite.

resolve conflict for dit_debug merge with mainlie

change import (remote lint complains).