In #6490 we enabled non-blocking torch device transfers throughout the model manager's memory management code. When using this torch feature, torch attempts to wait until the tensor transfer has completed before allowing any access to the tensor. Theoretically, that should make this a safe feature to use.
This provides a small performance improvement but causes race conditions in some situations. Specific platforms/systems are affected, and complicated data dependencies can make this unsafe.
- Intermittent black images on MPS devices - reported on discord and #6545, fixed with special handling in #6549.
- Intermittent OOMs and black images on a P4000 GPU on Windows - reported in #6613, fixed in this commit.
On my system, I haven't experience any issues with generation, but targeted testing of non-blocking ops did expose a race condition when moving tensors from CUDA to CPU.
One workaround is to use torch streams with manual sync points. Our application logic is complicated enough that this would be a lot of work and feels ripe for edge cases and missed spots.
Much safer is to fully revert non-locking - which is what this change does.
This issue is caused by a race condition. When a large image is served to the client, it is done using a streaming `FileResponse`. This concurrently serves the image straight from disk. The file is kept open by FastAPI until the image is fully served.
When a user deletes an image before the file is done serving, the delete fails because the file is still held by FastAPI.
To reproduce the issue:
- Create a very large image (8k reliably creates the issue).
- Create a smaller image, so that the first image in the gallery is not the large image.
- Refresh the app. The small image should be selected.
- Select the large image and immediately delete it. You have to be fast, to delete it before it finishes loading.
- In the terminal, we expect to see an error saying `Failed to delete image file`, and the image does not disappear from the UI.
- After a short wait, once the image has fully loaded, try deleting it again. We expect this to work.
The workaround is to instead serve the image from memory.
Loading the image to memory is very fast, so there is only a tiny window in which we could create the race condition, but it technically could still occur, because FastAPI is asynchronous and handles requests concurrently.
Once we load the image into memory, deletions of that image will work. Then we return a normal `Response` object with the image bytes. This is essentially what `FileResponse` does - except it uses `anyio.open_file`, which is async.
The tradeoff is that the server thread is blocked while opening the file. I think this is a fair tradeoff.
A future enhancement could be to implement soft deletion of images (db is already set up for this), and then clean up deleted image files on startup/shutdown. We could move back to using the async `FileResponse` for best responsiveness in the server without any risk of race conditions.
For some reason, I started getting this indefinite hang when the app checks if port 9090 is available. After some fiddling around, I found that adding a timeout resolves the issue.
I confirmed that the util still works by starting the app on 9090, then starting a second instance. The second instance correctly saw 9090 in use and moved to 9091.
