I'm using stable-diffusion on a 2022 Macbook M2 Air with 24 GB unified memory.
I see this taking about 2.0s/it.
I've moved many deps from pip to conda-forge, to take advantage of the
precompiled binaries. Some notes for Mac users, since I've seen a lot of
confusion about this:
One doesn't need the `apple` channel to run this on a Mac-- that's only
used by `tensorflow-deps`, required for running tensorflow-metal. For
that, I have an example environment.yml here:
https://developer.apple.com/forums/thread/711792?answerId=723276022#723276022
However, the `CONDA_ENV=osx-arm64` environment variable *is* needed to
ensure that you do not run any Intel-specific packages such as `mkl`,
which will fail with [cryptic errors](https://github.com/CompVis/stable-diffusion/issues/25#issuecomment-1226702274)
on the ARM architecture and cause the environment to break.
I've also added a comment in the env file about 3.10 not working yet.
When it becomes possible to update, those commands run on an osx-arm64
machine should work to determine the new version set.
Here's what a successful run of dream.py should look like:
```
$ python scripts/dream.py --full_precision SIGABRT(6) ↵ 08:42:59
* Initializing, be patient...
Loading model from models/ldm/stable-diffusion-v1/model.ckpt
LatentDiffusion: Running in eps-prediction mode
DiffusionWrapper has 859.52 M params.
making attention of type 'vanilla' with 512 in_channels
Working with z of shape (1, 4, 32, 32) = 4096 dimensions.
making attention of type 'vanilla' with 512 in_channels
Using slower but more accurate full-precision math (--full_precision)
>> Setting Sampler to k_lms
model loaded in 6.12s
* Initialization done! Awaiting your command (-h for help, 'q' to quit)
dream> "an astronaut riding a horse"
Generating: 0%| | 0/1 [00:00<?, ?it/s]/Users/corajr/Documents/lstein/ldm/modules/embedding_manager.py:152: UserWarning: The operator 'aten::nonzero' is not currently supported on the MPS backend and will fall back to run on the CPU. This may have performance implications. (Triggered internally at /Users/runner/work/_temp/anaconda/conda-bld/pytorch_1662016319283/work/aten/src/ATen/mps/MPSFallback.mm:11.)
placeholder_idx = torch.where(
100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 50/50 [01:37<00:00, 1.95s/it]
Generating: 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [01:38<00:00, 98.55s/it]
Usage stats:
1 image(s) generated in 98.60s
Max VRAM used for this generation: 0.00G
Outputs:
outputs/img-samples/000001.1525943180.png: "an astronaut riding a horse" -s50 -W512 -H512 -C7.5 -Ak_lms -F -S1525943180
```
- move all device init logic to T2I.__init__
- handle m1 specific edge case with autocast device type
- check torch.cuda.is_available before using cuda
* This functionality is triggered by the --fit option in the CLI (default
false), and by the "fit" checkbox in the WebGUI (default True)
* In addition, this commit contains a number of whitespace changes to
make the code more readable, as well as an attempt to unify the visual
appearance of info and warning messages.
* fix AttributeError crash when running on non-CUDA systems; closes issue #234 and issue #250
* although this prevents dream.py script from crashing immediately on MPS systems, MPS support still very much a work in progress.
This adds an option -t argument that will print out color-coded tokenization, SD has a maximum of 77 tokens, it silently discards tokens over the limit if your prompt is too long.
By using -t you can see how your prompt is being tokenized which helps prompt crafting.
