How to run huggingface model?
I was able to get the model working with the github code after commenting out "@torch.compile" over the fused flash attention (triton problems dunno?), but I was wondering if there's a better way to run this with diffusers or what?
Can you share more information about your environment to reproduce this issue? (what version of pytorch are you using?)
I followed the instruction to conda create -f requirements.yaml but ended up upgrading the python3.10 after an issue with torchvision throwing errors on import.
name: bd3lm
channels:
- pytorch
- anaconda
- nvidia
- defaults
dependencies:
- cuda-nvcc=12.4.99
- jupyter=1.0.0
- pip=23.3.1
- python=3.10
- pytorch=2.6
- pip:
- datasets==2.18.0
- einops==0.7.0
- fsspec==2024.2.0
- git-lfs==1.6
- h5py==3.10.0
- hydra-core==1.3.2
- ipdb==0.13.13
- lightning==2.2.1
- notebook==7.1.1
- nvitop==1.3.2
- omegaconf==2.3.0
- packaging==23.2
- pandas==2.2.1
- rich==13.7.1
- seaborn==0.13.2
- scikit-learn==1.4.0
- timm==0.9.16
- transformers==4.38.2
- triton==2.2.0
- wandb==0.13.5
Think the torchvision error might have something to do with deprecation of pytorch channel, but not sure.
I was able to get the huggingface model to work! Super exciting stuff. I was really just more curious about roadmap to implement this in a more traditional, HF-flavored route by using diffusers or something, because obviously there's no way to use the built in .generate() with a MaskedLM.
Great to hear! To make the environment setup easier for others, I'll update the setup to install all dependencies through pip instead of using conda channels.
I think it's a great idea to support .generate()in our Block Diffusion models; depending on my bandwidth in the next couple months, my plan is to implement this and release a public notebook with example usage. Will update you if/when that happens. We will definitely support this if we end up releasing bigger models in the future