Create README.md
Browse files
README.md
ADDED
@@ -0,0 +1,85 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
license: other
|
3 |
+
tags:
|
4 |
+
- text-to-image
|
5 |
+
- flux
|
6 |
+
---
|
7 |
+
|
8 |
+
# Flux Dev Quant
|
9 |
+
|
10 |
+
## Setup
|
11 |
+
|
12 |
+
```
|
13 |
+
pip install accelerate diffusers optimum-quanto transformers sentencepiece
|
14 |
+
pip install --upgrade git+https://github.com/huggingface/diffusers.git@main
|
15 |
+
```
|
16 |
+
|
17 |
+
There are places where the pre-trained weights in fp16 **overflow**, resulting in a blank image. Wait for the updated diffusers library.
|
18 |
+
|
19 |
+
## Inference
|
20 |
+
|
21 |
+
```python
|
22 |
+
from diffusers import AutoencoderKL, FluxPipeline, FlowMatchEulerDiscreteScheduler, FluxTransformer2DModel
|
23 |
+
import gc
|
24 |
+
from optimum.quanto.models import QuantizedDiffusersModel, QuantizedTransformersModel
|
25 |
+
import sys
|
26 |
+
import torch
|
27 |
+
from transformers import CLIPTextModel, CLIPTokenizer, T5EncoderModel, T5TokenizerFast
|
28 |
+
|
29 |
+
class Flux2DModel(QuantizedDiffusersModel):
|
30 |
+
base_class = FluxTransformer2DModel
|
31 |
+
|
32 |
+
class T5Model(QuantizedTransformersModel):
|
33 |
+
auto_class = T5EncoderModel
|
34 |
+
|
35 |
+
FLUX_DEV = sys.argv[1] if len(sys.argv) > 1 else 'black-forest-labs/FLUX.1-dev'
|
36 |
+
FLUX_INT = sys.argv[2] if len(sys.argv) > 2 else './flux-int4'
|
37 |
+
T5_INT = sys.argv[3] if len(sys.argv) > 3 else './flux-t5'
|
38 |
+
SAMPLER_STEP = 2
|
39 |
+
PROMPT_CLIP = ''
|
40 |
+
PROMPT_T5 = sys.argv[4] if len(sys.argv) > 4 else 'cat playing piano'
|
41 |
+
|
42 |
+
if __name__ == '__main__':
|
43 |
+
torch.set_default_dtype(torch.float16)
|
44 |
+
print('Step 1/5')
|
45 |
+
T5EncoderModel.from_config = lambda c: T5EncoderModel(c) # Duck and tape for Quanto support.
|
46 |
+
wrapped_t5 = T5Model.from_pretrained(T5_INT)
|
47 |
+
print('Step 2/5')
|
48 |
+
wrapped_model = Flux2DModel.from_pretrained(FLUX_INT)
|
49 |
+
print('Step 3/5')
|
50 |
+
pipe = FluxPipeline.from_pretrained(FLUX_DEV,
|
51 |
+
scheduler=FlowMatchEulerDiscreteScheduler.from_pretrained(FLUX_DEV, subfolder='scheduler'),
|
52 |
+
text_encoder=CLIPTextModel.from_pretrained(FLUX_DEV, subfolder='text_encoder'),
|
53 |
+
text_encoder_2=wrapped_t5._wrapped,
|
54 |
+
tokenizer=CLIPTokenizer.from_pretrained(FLUX_DEV, subfolder='tokenizer'),
|
55 |
+
tokenizer_2=T5TokenizerFast.from_pretrained(FLUX_DEV, subfolder='tokenizer_2'),
|
56 |
+
transformer=wrapped_model._wrapped,
|
57 |
+
vae=AutoencoderKL.from_pretrained(FLUX_DEV, subfolder='vae'),
|
58 |
+
torch_dtype=torch.float16).to('cuda')
|
59 |
+
latents = pipe('cat playing piano', num_inference_steps=SAMPLER_STEP, output_type='latent').images
|
60 |
+
print('Step 4/5')
|
61 |
+
transformer = pipe.transformer.to('cpu')
|
62 |
+
te_2 = pipe.text_encoder_2.to('cpu')
|
63 |
+
pipe.transformer = None
|
64 |
+
pipe.text_encoder_2 = None
|
65 |
+
del transformer
|
66 |
+
del te_2
|
67 |
+
gc.collect()
|
68 |
+
torch.cuda.empty_cache()
|
69 |
+
print('Step 5/5')
|
70 |
+
latents = FluxPipeline._unpack_latents(latents, 1024, 1024, pipe.vae_scale_factor)
|
71 |
+
latents = (latents / pipe.vae.config.scaling_factor) + pipe.vae.config.shift_factor
|
72 |
+
# Either use fp16 or move vae to cpu and keep it in full precision.
|
73 |
+
vae: AutoencoderKL = pipe.vae.to(dtype=torch.float16)
|
74 |
+
image, = vae.decode(latents.to(dtype=vae.dtype), return_dict=False)
|
75 |
+
image = pipe.image_processor.postprocess(image.detach(), output_type='pil')[0]
|
76 |
+
image.save('./cat.png')
|
77 |
+
```
|
78 |
+
|
79 |
+
## Disclaimer
|
80 |
+
|
81 |
+
Use of this code and the copy of documentation requires citation and attribution to the author via a link to their Hugging Face profile in all resulting work.
|
82 |
+
|
83 |
+
## License
|
84 |
+
|
85 |
+
[FLUX.1 Dev Non-Commercial License](http://huggingface.co/black-forest-labs/FLUX.1-dev/blob/main/LICENSE.md)
|