fix AttributeError crash when running on non-CUDA systems

[lstein]

No description provided.

[lstein]

This is a very minor update, but gets dream.py running on Mac MPS systems.

[tildebyte]

Squash, please 😁

Otherwise, LGTM 👍

[tiems90]

Just checked out this branch on an M1 Mac, but still getting an issue here.

Traceback (most recent call last):
  File "/Users/timo/Code/images/stable-diffusion/scripts/dream.py", line 535, in <module>
    main()
  File "/Users/timo/Code/images/stable-diffusion/scripts/dream.py", line 100, in main
    main_loop(t2i, opt.outdir, opt.prompt_as_dir, cmd_parser, infile)
  File "/Users/timo/Code/images/stable-diffusion/scripts/dream.py", line 227, in main_loop
    t2i.prompt2image(image_callback=image_writer, **vars(opt))
  File "/Users/timo/Code/images/stable-diffusion/ldm/simplet2i.py", line 289, in prompt2image
    torch.cuda.torch.cuda.reset_peak_memory_stats()
  File "/opt/homebrew/Caskroom/miniforge/base/envs/ldm/lib/python3.10/site-packages/torch/cuda/memory.py", line 260, in reset_peak_memory_stats
    return torch._C._cuda_resetPeakMemoryStats(device)
AttributeError: module 'torch._C' has no attribute '_cuda_resetPeakMemoryStats'

[eldridgegreg]

Just checked out this branch on an M1 Mac, but still getting an issue here.

Traceback (most recent call last):
  File "/Users/timo/Code/images/stable-diffusion/scripts/dream.py", line 535, in <module>
    main()
  File "/Users/timo/Code/images/stable-diffusion/scripts/dream.py", line 100, in main
    main_loop(t2i, opt.outdir, opt.prompt_as_dir, cmd_parser, infile)
  File "/Users/timo/Code/images/stable-diffusion/scripts/dream.py", line 227, in main_loop
    t2i.prompt2image(image_callback=image_writer, **vars(opt))
  File "/Users/timo/Code/images/stable-diffusion/ldm/simplet2i.py", line 289, in prompt2image
    torch.cuda.torch.cuda.reset_peak_memory_stats()
  File "/opt/homebrew/Caskroom/miniforge/base/envs/ldm/lib/python3.10/site-packages/torch/cuda/memory.py", line 260, in reset_peak_memory_stats
    return torch._C._cuda_resetPeakMemoryStats(device)
AttributeError: module 'torch._C' has no attribute '_cuda_resetPeakMemoryStats'

-        torch.cuda.torch.cuda.reset_peak_memory_stats()
+        if self.device == 'cuda':
+            torch.cuda.torch.cuda.reset_peak_memory_stats()

Resolves that for me but script is still non-functional with:

dream> monkey eating a doughnut
User specified autocast device_type must be 'cuda' or 'cpu'
Are you sure your system has an adequate NVIDIA GPU?
Usage stats:
   0 image(s) generated in 0.00s
   Max VRAM used for this generation: 0.00G
Outputs:

[tildebyte]

AttributeError: module 'torch._C' has no attribute '_cuda_resetPeakMemoryStats'

As mentioned elsewhere, Mx/MPS support is very much W.I.P.; there's a PR open ti fix this exact issue.

[tiems90]

Yeah, I believed this is exactly that PR based on the branch name fix-crash-on-MPS. I read there's no access to Mx devices for you two, so I thought it might be useful to know the fix in this PR does not resolve the issue.

If it's not helpful, apologies.

[tildebyte]

Yeah, I believed this is exactly that PR based on the branch name fix-crash-on-MPS. I read there's no access to Mx devices for you two, so I thought it might be useful to know the fix in this PR does not resolve the issue.

If it's not helpful, apologies.

No worries! Feedback is always helpful. I wanted to set expectations for Mac users that things are in flux.

[lstein]

Squash, please grin

Otherwise, LGTM +1

I dunno why I get the excess commit messages. I think it's because I create a branch, make my edits, commit, and then get a warning that the branch is out of date with main and needs a merge. The funny thing is that I try to do a pull from origin before branching. Maybe it is the mixture of merging pull requests from the github interface and working locally that is tripping me up.

Anyway, I've got the github interface set to doing a --squash each time.

[tildebyte]

mixture of merging pull requests from the github interface and working locally that is tripping me up

Probably

Doing a rebase or fast-forward against main is the best practice.

The only problem I have with always squashing everything in a PR is that there are plenty of times when you DO want multiple commits in history, to decouple things which are related (and thus in the same PR), but which don't depend on each other. It's not hard to imagine a scenario where you need to revert a single commit with a regression, but instead of doing that, you have to back out several "commits" worth of work from a single PR, then re-commit the changes which aren't involved with the regression.

[yudhanjaya]

The issue is still here. Testing on M1 Mac Pro yields same error:

dream> a cat pharaoh dreaming of empire
Traceback (most recent call last):
  File "/Users/yudhanjaya/stable-diffusion/scripts/dream.py", line 543, in <module>
    main()
  File "/Users/yudhanjaya/stable-diffusion/scripts/dream.py", line 103, in main
    main_loop(t2i, opt.outdir, opt.prompt_as_dir, cmd_parser, infile)
  File "/Users/yudhanjaya/stable-diffusion/scripts/dream.py", line 230, in main_loop
    t2i.prompt2image(image_callback=image_writer, **vars(opt))
  File "/Users/yudhanjaya/stable-diffusion/ldm/simplet2i.py", line 282, in prompt2image
    torch.cuda.torch.cuda.reset_peak_memory_stats()
  File "/Users/yudhanjaya/miniconda3/envs/ldm/lib/python3.10/site-packages/torch/cuda/memory.py", line 260, in reset_peak_memory_stats
    return torch._C._cuda_resetPeakMemoryStats(device)
AttributeError: module 'torch._C' has no attribute '_cuda_resetPeakMemoryStats'

commenting out the torch.cuda.torch.cuda.reset_peak_memory_stats() line gets it to run (same as before), but the output is again the same:

* Initialization done! Awaiting your command (-h for help, 'q' to quit)
dream> a cat pharaoh dreaming of empire
Generating:   0%|                                         | 0/1 [00:00<?, ?it/s]
"LayerNormKernelImpl" not implemented for 'Half'
Are you sure your system has an adequate NVIDIA GPU?
Usage stats:
   0 image(s) generated in 0.07s
   Max VRAM used for this generation: 0.00G
Outputs:

Changing the device default to cpu doesn't make a difference. It looks like an error is generated with LayerNormKernelImpl, then the error handler catches it and prints the general message.

[lstein]

Doing a rebase or fast-forward against main is the best practice.

It definitely takes a lot more discipline to manage a fast-moving multi-collaborator project like this one than any of the projects I've been involved in before. I guess rebase -i and I are going to get to be good friends.

[tildebyte]

Doing a rebase or fast-forward against main is the best practice.

It definitely takes a lot more discipline to manage a fast-moving multi-collaborator project like this one than any of the projects I've been involved in before. I guess rebase -i and I are going to get to be good friends.

May I humbly offer my personal cheatsheet? https://github.com/ben-alkov/_git_-Beyond-The-Basics/blob/master/git%20-%20beyond%20the%20basics.md#commit-magic