Skip to content

adobe-research/Sparse-LaViDa

Repository files navigation

[CVPR 2026] Sparse-LaViDa: Sparse Multimodal Discrete Diffusion Language Models

[Paper] [Huggingface]

Model Architecture

Installation

conda create --name lavida python=3.13
conda activate lavida
pip install -e .[lavida]
pip install wheel
MAX_JOBS=32 pip install flash-attn==2.7.4.post1 --no-build-isolation
pip install jupyter notebook
pip install -U huggingface_hub[hf_xet] --force-reinstall

Download Checkpoint

Please download checkpoints from [Huggingface]

Inference

Example inference script is provided as demo_sparse_lavida.py. This script will run both the standard decoding and sparse decoding with token truncation, and benchmark output latency.

Training

See Training Readme for general training of LaViDa-O.

Please run the training script to finetune the model for sparse parameterization. The script is just standard SFT script with the following key flags for sparsity:

    --block_causal True \
    --gen_enc_add_pos_emb True \
    --num_register_tokens 64,64 \
    --num_register_groups 25,25 \

LICENSE

Both the model and code are licensed with Adobe Research License, which is included here [License.pdf].

About

No description, website, or topics provided.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages