Model erroring on every generation

#2
by peteredm - opened

I've been unable to try this model as errors with a nondescript "Error" view upon any generation (drawing or upload).

Natural Language Learning Group org

Thanks for the report @peteredm ! Could you give us more information on your workflow? I just tested it again using the public space and it is working fine for me (test image: https://w.wiki/A7Cc). When does "Error" appear for you? Only during compilation or already before the model starts generating something? Do you use a private or public instance?

Hey @potamides ! Yesterday it would error immediately without even placing me in a queue. Today, I've get into the queue and see a loading spinner for hundreds of seconds before it errors or i abort. Here's what I've tried:

  • Using your reference image on the public space
  • Using different figure images on the public space
  • Using sketches on the public space

Default settings used throughout.

Screenshot 2024-06-06 at 8.48.13 AM.png

I was eventually able to get some output by firing up a private A10G, all features appear to work there.

Natural Language Learning Group org
edited Jun 6, 2024

I think I found the source of the problem. We currently run in a memory leak with the way we create threads for each request. This also explains why it works for your private instance. I will try to fix this asap. Thanks for bringing this to my attention!

Natural Language Learning Group org
edited Jun 7, 2024

After digging deeper there also seems to be an issue with how gradio cancels events. I have reported it upstream (#8503).

Natural Language Learning Group org

To wrap this up, the issue had two root causes. One was a problem with Gradio, which I reported in #8503 and worked around in commit 186ec56. The other issue was a memory leak, which I fixed in commit a028088. With that, the problem should be fixed.

potamides changed discussion status to closed

Sign up or log in to comment