About new diffusers package in the dev 2a85f894 commit.

[mart-hill]

I've noticed a diffusers-0.31.0.dev0 package with the newest (as of the moment) 2a85f89 dev commit being installed. Is is maybe because of issue like this? It's from previous commit, e6ff17d.

17:26:27-928825 INFO     Base: class=StableDiffusionXLPipeline
Progress  4.05it/s ████▋                               14% 5/36 00:01 00:07 Base17:26:29-563824 WARNING  HiDiffusion: The feature size is not divisible by 2
17:26:29-568824 WARNING  HiDiffusion: The feature size is not divisible by 2 <- it's because I didn't take the image pixel sizes into HiDiffusion's consideration
17:26:29-673321 WARNING  HiDiffusion: The feature size is not divisible by 2
17:26:29-677321 WARNING  HiDiffusion: The feature size is not divisible by 2
17:26:29-682821 WARNING  HiDiffusion: The feature size is not divisible by 2
17:26:29-686824 WARNING  HiDiffusion: The feature size is not divisible by 2

...

17:26:31-159326 WARNING  HiDiffusion: The feature size is not divisible by 2
17:26:31-293821 WARNING  HiDiffusion: The feature size is not divisible by 2
17:26:31-298823 WARNING  HiDiffusion: The feature size is not divisible by 2
17:26:31-304826 WARNING  HiDiffusion: The feature size is not divisible by 2
17:26:31-308822 WARNING  HiDiffusion: The feature size is not divisible by 2
Progress  4.32it/s ███████████████████████████▌       83% 30/36 00:06 00:01 Base
17:26:35-335822 INFO     Interrupted... <- I did it, cause I've got just huge colorful dots at this point.
17:26:35-395323 INFO     Processed: images=0 its=0.00 time=7.47 timers={'init': 0.03, 'encode': 0.4, 'args': 0.42, 'move': 0.03, 'callback':
                         1.45, 'process': 6.98, 'post': 0.03} memory={'ram': {'used': 5.7, 'total': 63.68}, 'gpu': {'used': 8.99, 'total':
                         24.0}, 'retries': 0, 'oom': 0}
17:26:52-343822 INFO     Base: class=StableDiffusionPipeline
C:\actions-runner\_work\pytorch\pytorch\builder\windows\pytorch\aten\src\ATen\native\cuda\Indexing.cu:1284: block: [112,0,0], thread: [96,0,0] Assertion `srcIndex < srcSelectDimSize` failed.
C:\actions-runner\_work\pytorch\pytorch\builder\windows\pytorch\aten\src\ATen\native\cuda\Indexing.cu:1284: block: [112,0,0], thread: [97,0,0] Assertion `srcIndex < srcSelectDimSize` failed.
C:\actions-runner\_work\pytorch\pytorch\builder\windows\pytorch\aten\src\ATen\native\cuda\Indexing.cu:1284: block: [112,0,0], thread: [98,0,0] Assertion `srcIndex < srcSelectDimSize` failed.
C:\actions-runner\_work\pytorch\pytorch\builder\windows\pytorch\aten\src\ATen\native\cuda\Indexing.cu:1284: block: [112,0,0], thread: [99,0,0] Assertion `srcIndex < srcSelectDimSize` failed.
C:\actions-runner\_work\pytorch\pytorch\builder\windows\pytorch\aten\src\ATen\native\cuda\Indexing.cu:1284: block: [112,0,0], thread: [100,0,0] Assertion `srcIndex < srcSelectDimSize` failed.

...

17:26:53-902822 ERROR    Prompt parser encode: CUDA error: device-side assert triggered
                         CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be
                         incorrect.
                         For debugging consider passing CUDA_LAUNCH_BLOCKING=1
                         Compile with `TORCH_USE_CUDA_DSA` to enable device-side assertions.

17:26:53-946822 ERROR    Torch generator: seeds=[211] device=cuda CUDA error: device-side assert triggered
                         CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be
                         incorrect.
                         For debugging consider passing CUDA_LAUNCH_BLOCKING=1
                         Compile with `TORCH_USE_CUDA_DSA` to enable device-side assertions.

I mean, the UI suddenly went SD1.5 on SDXL model, and 'contaminated' the CUDA 'memory space', if I could name it that way, because no amount of Unload/Reload would fix it, until I would restart the UI. 😅

Well, I updated it to the 2a85f89 commit, while at it and noticed the diffusers package change. 🙂
Ah, Ctrl-X also seemed to misbehave sometimes (when I tested it firsthand), and didn't 'let go' of the pipeline, when I disabled CTRL-X in the scripts and that also seemed to irritate the UI a lot. :)

[vladmandic]

diffusers upgrade is a REGULAR op, stop calling it out. I explained reasoning behind it in dev-branch channel on discord.

regarding error, yes, once any aten error occurs, it's unrecoverable. one of such causes was fixed yestersay, but each possible path leading to it is different and error itself is not useful - need to know full trace leading to it.

[mart-hill]

I didn't bear any ill will with 'calling it out', but OK, I'll stop.

[vladmandic]

no ill will assumed. i'm just saying that those upgrades are normal and regular. most likely 1-2 per week at the moment. and typically not related to issues at hand.

[mart-hill]

I thought these errors I was getting were the push to change diffusers package, hence I put this in the discussions, not issue report - I'm always trying to solve such things with recreating the whole venv first, before I report anything . 😅