Ruurd/LAD-training
Viewer
• Updated
• 971k • 14
Ruurd Kuiper PRO
Ruurd
AI & ML interests
None yet
Recent Activity
updated
a dataset about 22 hours ago
Ruurd/LAD-training published
a dataset about 22 hours ago
Ruurd/LAD-training upvoted a paper 5 days ago
dLLM: Simple Diffusion Language Modeling Organizations
updated a dataset about 22 hours ago
published a dataset about 22 hours ago
upvoted a paper 5 days ago
commented a paper 5 days ago
liked a Space about 2 months ago
liked a Space 2 months ago
liked a model 3 months ago
updated 2 Spaces 3 months ago
published a model 4 months ago
upvoted an article 4 months ago
Article
Diffusion Language Models: The New Paradigm
•
45
reacted to sergiopaniego's post with 🔥 4 months ago
Post
5415
fine-tuning a 14B model with TRL + SFT on a free Colab (T4 GPU)?
thanks to the latest TRL optimizations, you actually can!
sharing a new notebook showing how to do it 😎
colab: https://colab.research.google.com/github/huggingface/trl/blob/main/examples/notebooks/sft_trl_lora_qlora.ipynb
notebooks in TRL: https://github.com/huggingface/trl/tree/main/examples/notebooks
thanks to the latest TRL optimizations, you actually can!
sharing a new notebook showing how to do it 😎
colab: https://colab.research.google.com/github/huggingface/trl/blob/main/examples/notebooks/sft_trl_lora_qlora.ipynb
notebooks in TRL: https://github.com/huggingface/trl/tree/main/examples/notebooks
- 2 replies
upvoted an article 9 months ago
Article
LAD: LoRA-Adapted Denoiser
•
6
upvoted a paper 9 months ago
liked a Space 9 months ago
reacted to their post with 🔥👀❤️ 9 months ago
Post
2342
The past year I have been trying to get diffusion models to work for language generation, without having to retrain a LLM from scratch. And recently, we finally succeeded:
We introduce "LAD: LoRA-Adapted Denoiser", a method to convert a LLaMA model into a text diffusion model using LoRA finetuning and structured input corruption.
🎯 Try the demo and read the write-up here!
https://ruurdkuiper.github.io/tini-lad/
Unlike autoregressive (word-for-word) models like ChatGPT, diffusion models iteratively refine a noised sequence. However, most current diffusion approaches rely on all-parameter retraining and repeatedly remasking tokens, which is costly and slow during both training and inference!
🧠 With LAD:
- We can finetune an autoregressive model for diffusive generation in just 10 hours on a single GPU.
- Test-time compute is fully adjustable: fewer steps means faster outputs while more steps improve output quality.
- Due to our unique noising schedule, remasking is not always needed during inference. All tokens are attended to in each iteration!
🔍 LAD is built using:
– A frozen LLaMA-8B backbone
– Structured noising: token swaps, duplications, replacements, span shifts
– Modified attention masks for bidirectional decoding
💡 We show that even small, fast-trained models can perform diffusive generation — with competitive benchmark performance, perplexity and more flexible test-time behavior than traditional transformers.
We introduce "LAD: LoRA-Adapted Denoiser", a method to convert a LLaMA model into a text diffusion model using LoRA finetuning and structured input corruption.
🎯 Try the demo and read the write-up here!
https://ruurdkuiper.github.io/tini-lad/
Unlike autoregressive (word-for-word) models like ChatGPT, diffusion models iteratively refine a noised sequence. However, most current diffusion approaches rely on all-parameter retraining and repeatedly remasking tokens, which is costly and slow during both training and inference!
🧠 With LAD:
- We can finetune an autoregressive model for diffusive generation in just 10 hours on a single GPU.
- Test-time compute is fully adjustable: fewer steps means faster outputs while more steps improve output quality.
- Due to our unique noising schedule, remasking is not always needed during inference. All tokens are attended to in each iteration!
🔍 LAD is built using:
– A frozen LLaMA-8B backbone
– Structured noising: token swaps, duplications, replacements, span shifts
– Modified attention masks for bidirectional decoding
💡 We show that even small, fast-trained models can perform diffusive generation — with competitive benchmark performance, perplexity and more flexible test-time behavior than traditional transformers.
-
- 2 replies
replied to their post 9 months ago
Thanks! I’m trying to get it under attention as I think the leap from pretraining (10000+ hours) diffusion models to mere finetuning (10 hours) for adaptation is a big one, and could really help this method gain some traction.
published an
article 9 months ago
Article
LAD: LoRA-Adapted Denoiser
•
6