search

mlc-ai / web-stable-diffusion

Bringing stable diffusion models to web browsers. Everything runs inside the browser with no server support.
https://mlc.ai/web-stable-diffusion
Apache License 2.0
3.56k stars 227 forks source link

Demo fails on latest tvm/unity branch #17

Open infil00p opened 1 year ago

infil00p commented 1 year ago

I'm currently going through the web-stable-diffusion demo, and I'm finding that the clip_to_text_embeddings function isn't adding functions to the exported IRModule, causing the assert to fail. This issue seems to have started appearing after March 20th, and I'm working to track down the isuse in the repo:

This seems to happen on MacOS and Linux. I haven't tried in Windows because PyTorch doesn't support compiled models on Windows yet.

I'm not sure if there's been an API change on TVM, or if this is a TVM bug or an issue with the demo.

MasterJH5574 commented 1 year ago

Hi @infil00p, thanks for your feedback! Would you mind providing some more detailed information so that we can help check?

We are using the mlc-ai/relax repository (https://github.com/mlc-ai/relax) as codebase for the running the demo and build.py. This repository contains a few more commits than tvm/unity, and these changes will be eventually brought back to the unity branch in the future. This repository is also the built and packed as the mlc-ai-nightly pip package. We hope the latest mlc-ai/relax repository can address your issue.