A few small quality of life changes

[warner-benjamin]

A few small quality of life changes:

• added a python .gitignore preset to ignore the default image save location and model symbolic link
• allow users to set the device to run stable diffusion with, defaults to 'cuda'
• autocast will use the user set device so should run on CPU with bfloat16 accelerated ops if device='cpu'
• cleaned up the stacked context manager lines

Changes tested and working with text2img.

[warner-benjamin]

I see #43 also adds a gitignore with a few differences. Can harmonize this PR with 43, or you can choose 43's gitignore.

[artmen1516]

I tested this PR since I cannot run original code from my Macbook pro because of cuda device, i tested with the option you added -d cpu and it successfully launched the dream.py script, but when I write something to generate an image I get the same error I was getting with original code AssertionError: Torch not compiled with CUDA enabled .
Here's the error I'm getting:

note: I have configured a virtual environment with pipenv, not using conda.

[warner-benjamin]

I missed some hardcoded cuda sections. Fixed those, but the k_diffusion package still appears to have them. This should allow users with multiple GPUs to pick which GPU to use by setting device if they are not using k_lms.

This doesn't appear to allow CPU generation, though, I get a "LayerNormKernelImpl" not implemented for 'Half' error for half precision, and in full precision a expected scalar type BFloat16 but found Float error. It's possible upgrading to PyTorch 1.12 would resolve either of these issues, but I'm not going to spend more time trying to get CPU generation to run.

[lstein]

Thanks for the PR!

IMO there's more community need for folding in the basunjindal VRAM optimizations than getting the thing to run on CPU (tho' that would be nice). If someone out there wants to work on an optimization port in a way that lets the user select whether or not to use them, I'd be totally grateful.

Also, as a heads up, I'm about to do two things:

1. Refactor simplet2i.py to move all the PNG filenaming and manipulation into a separate module. This will change the API so that instead of returning an image name, txt2img() and img2img() return Image objects. I hope this doesn't break anyone's scripts.
2. I'm going to attempt to merge shared code among txt2img() and img2img() and remove a few levels of indentation. Please hold off on additional PRs until tomorrow (Thursday) to avoid massive merge conflicts.

Lastly, I was playing around with the hugging face Diffusion modules last night and they're really elegant! I am strongly tempted to base a new version of dream.py on the HF libraries. Do they have (or plan to have) the features that the community has been pasting into the various CompViz derivatives (including this one) such as textual inversion and support for all the k_* samplers?

[warner-benjamin]

IMO there's more community need for folding in the basunjindal VRAM optimizations than getting the thing to run on CPU (tho' that would be nice).

I agree. Setting the device was more about allowing users to pick which GPU to run it on if they have two, and I was hoping it would also just work on CPU since PyTorch-CPU is more mature than ROCm or MIPs.

[lstein]

Issue #69 reports img2img reporting dying with an error message about not knowing which GPU device to use. When I check img2img myself, I'm getting "sqrt" "_vml_cpu" not implemented for 'Half' . I suspect these problems are related to your PR? I appreciate your looking into it.

[lstein]

I was able to fix the sqrt error message by backing out the changes to ddim.py, lines 20-23. Still do not have confirmation on issue #69.

[warner-benjamin]

I thought I had chased down all the orphan device settings that were not respecting the Lightning model's device. But I clearly missed some. Looking into it.

[warner-benjamin]

I see you beat me to resolving it using a similar method to my approach.