"User specified autocast device_type must be 'cuda' or 'cpu'" on M1 Mac

[bfirsh]

We're trying to convert our guide which currently on HN right now (!) to use your fork but running into trouble.

We're installing with plain pip rather than Conda using these requirements.

However we're getting this error:

ben@MacBook-Pro ~/p/tmp/stable-diffusion [lstein]
± % python scripts/dream.py --full_precision                                                                                                                                                                   !10211
* 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 5.78s

* Initialization done! Awaiting your command (-h for help, 'q' to quit)
dream> a red juicy apple floating in outer space, like a planet
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:

More context: https://news.ycombinator.com/item?id=32680293

Any ideas on how to fix @lstein et al?

[tonglil]

I was able to get prompt generation to run with this PR.

These samplers resulted in black images:

• klms
• kdpm_2a
• keuler_a
• kheun

Use another sampler to get images.

However img2img fails with the same error. fixed by @mabrowning's changes.

[mabrowning]

This works for me on top of current main 1b3c7ac:

$ git diff
diff --git a/scripts/dream.py b/scripts/dream.py
index 2e3f070..bdb4efd 100755
--- a/scripts/dream.py
+++ b/scripts/dream.py
@@ -54,12 +54,14 @@ def main():
         height=height,
         sampler_name=opt.sampler_name,
         weights=weights,
+        precision="full" if opt.full_precision else "autocast",
         full_precision=opt.full_precision,
         config=config,
         grid  = opt.grid,
         # this is solely for recreating the prompt
         latent_diffusion_weights=opt.laion400m,
         embedding_path=opt.embedding_path,
     )
 
     # make sure the output directory exists

[stefanvaduva]

the error is gone after that change, but the generated image is black on an M1 Max

[mabrowning]

the error is gone after that change, but the generated image is black on an M1 Max

Use --pmls, or --keuler samplers.

[essyd]

This works for me on top of current main 1b3c7ac:

$ git diff
diff --git a/scripts/dream.py b/scripts/dream.py
index 2e3f070..bdb4efd 100755
--- a/scripts/dream.py
+++ b/scripts/dream.py
@@ -54,12 +54,14 @@ def main():
         height=height,
         sampler_name=opt.sampler_name,
         weights=weights,
+        precision="full" if opt.full_precision else "autocast",
         full_precision=opt.full_precision,
         config=config,
         grid  = opt.grid,
         # this is solely for recreating the prompt
         latent_diffusion_weights=opt.laion400m,
         embedding_path=opt.embedding_path,
     )
 
     # make sure the output directory exists

Unfortunately this didn't seem to work for me. Added that line to dream.py as shown (unless I'm missing something obvious it's been a long day lol)

[grantbarrett]

This also did not work for me.

[GlitchedGoat]

Mine wasn't working either, even with the above change, but it turns out I'd misread the README and I needed the --full_precision flag:
python scripts/dream.py --full_precision

(This also works with the --web flag for a nicer UI.)

[aphillipo]

This worked for me! It did complain about not using the GPU but I heard my laptop fans for the first time and got a black image. Trying -A pmls to see something from the dreams...

[i3oc9i]

I have applyed the PR #268. ==> gh pr checkout 268

and I got these results on a Mac Studio M1 Ultra

-Addim is OK - 50/50 [00:21<00:00, 2.32it/s]
-Ak_dpm_2_a got BLACK-IMAGE - 50/50 [00:41<00:00, 1.20it/s]
-Ak_dpm_2 is OK - 50/50 [00:42<00:00, 1.18it/s]
-Ak_euler_a got BLACK-IMAGE - 50/50 [00:20<00:00, 2.39it/s]
-Ak_euler is OK - 50/50 [00:21<00:00, 2.32it/s]
-Ak_heun got BLACK-IMAGE - 50/50 [00:42<00:00, 1.19it/s]
-Ak_lms got BLACK-IMAGE - 50/50 [00:21<00:00, 2.33it/s]
-Aplms is OK - 50/50 [00:21<00:00, 2.28it/s]

[adamjgrant]

I first applied the code change #293 (comment)

Then pulled from main, now all images are black.

Removed the code change (Appears it's no longer necessary) and tried each with the following results:

(On commit 3ee82d8)

Model	Result
KHEUN	Black
KHEULER_A	Black
KHEULER	✅
KDPM_2A	Black
KDPM_2	✅
KLMS	Black
PLMS	✅
DDIM	✅

Appears to be the same results as @i3oc9i

[gabrielrotbart]

I added this PR that also fixes this issue, it's a bit more verbose than @mabrowning 's elegant solution but makes sure that autocast is not set regardless of what precision is passed in

[mabrowning]

After 3ee82d8 is merged to main I no longer see this particular error and my patch isn't needed.

[essyd]

This worked for me! It did complain about not using the GPU but I heard my laptop fans for the first time and got a black image. Trying -A pmls to see something from the dreams...

Perfect! I did the same, thank you.

[lstein]

1. Which PR is best to merge? 2. Are some samplers working and others not? If so, need to detect and issue a warning if tried.

On Fri, Sep 2, 2022 at 8:43 AM Mark Browning ***@***.***> wrote: After 3ee82d8 <3ee82d8> I no longer see this particular error and my patch isn't needed. — Reply to this email directly, view it on GitHub <#293 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAA3EVL3D2IZ5JWRFF4AJZLV4HY7BANCNFSM6AAAAAAQCUPS5E> . You are receiving this because you were mentioned.Message ID: ***@***.***>

-- Written on my cell phone. Anything that seems odd is the fault of auto-correct.