Create README.md
Browse files
README.md
ADDED
@@ -0,0 +1,145 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
license: openrail++
|
3 |
+
---
|
4 |
+
|
5 |
+
# FLAX Latent Consistency Model (LCM) LoRA: SDXL - UNet
|
6 |
+
|
7 |
+
Unet with merged LCM weights (lora_scale=0.7) and converted to work with FLAX.
|
8 |
+
|
9 |
+
## Setup
|
10 |
+
|
11 |
+
To use on TPUs:
|
12 |
+
|
13 |
+
```bash
|
14 |
+
git clone https://github.com/entrpn/diffusers
|
15 |
+
cd diffusers
|
16 |
+
git checkout lcm_flax
|
17 |
+
pip install jax[tpu] -f https://storage.googleapis.com/jax-releases/libtpu_releases.html
|
18 |
+
pip install transformers flax torch torchvision
|
19 |
+
pip install .
|
20 |
+
```
|
21 |
+
|
22 |
+
## Run
|
23 |
+
|
24 |
+
|
25 |
+
```python
|
26 |
+
import os
|
27 |
+
from diffusers import FlaxStableDiffusionXLPipeline
|
28 |
+
import torch
|
29 |
+
import time
|
30 |
+
import jax
|
31 |
+
import jax.numpy as jnp
|
32 |
+
from flax.jax_utils import replicate
|
33 |
+
import numpy as np
|
34 |
+
from jax.experimental.compilation_cache import compilation_cache as cc
|
35 |
+
cc.initialize_cache(os.path.expanduser("~/jax_cache"))
|
36 |
+
|
37 |
+
from diffusers import (
|
38 |
+
FlaxUNet2DConditionModel,
|
39 |
+
FlaxLCMScheduler
|
40 |
+
)
|
41 |
+
|
42 |
+
base_model = "stabilityai/stable-diffusion-xl-base-1.0"
|
43 |
+
lcm_model = "sd_lora_model"
|
44 |
+
weight_dtype = jnp.bfloat16
|
45 |
+
revision= 'refs/pr/95'
|
46 |
+
|
47 |
+
pipeline, params = FlaxStableDiffusionXLPipeline.from_pretrained(
|
48 |
+
base_model, revision=revision, dtype=weight_dtype
|
49 |
+
)
|
50 |
+
|
51 |
+
del params["unet"]
|
52 |
+
|
53 |
+
unet, unet_params = FlaxUNet2DConditionModel.from_pretrained(
|
54 |
+
"jffacevedo/flax_lcm_unet",
|
55 |
+
dtype=weight_dtype,
|
56 |
+
)
|
57 |
+
|
58 |
+
scheduler, scheduler_state = FlaxLCMScheduler.from_pretrained(
|
59 |
+
base_model,
|
60 |
+
subfolder="scheduler",
|
61 |
+
revision=revision,
|
62 |
+
dtype=jnp.float32
|
63 |
+
)
|
64 |
+
|
65 |
+
params["unet"] = unet_params
|
66 |
+
pipeline.unet = unet
|
67 |
+
|
68 |
+
pipeline.scheduler = scheduler
|
69 |
+
|
70 |
+
params = jax.tree_util.tree_map(lambda x: x.astype(weight_dtype), params)
|
71 |
+
params["scheduler"] = scheduler_state
|
72 |
+
|
73 |
+
default_prompt = "high-quality photo of a baby dolphin playing in a pool and wearing a party hat"
|
74 |
+
default_neg_prompt = ""
|
75 |
+
default_seed = 42
|
76 |
+
default_guidance_scale = 1.0
|
77 |
+
default_num_steps = 4
|
78 |
+
|
79 |
+
def tokenize_prompt(prompt, neg_prompt):
|
80 |
+
prompt_ids = pipeline.prepare_inputs(prompt)
|
81 |
+
neg_prompt_ids = pipeline.prepare_inputs(neg_prompt)
|
82 |
+
return prompt_ids, neg_prompt_ids
|
83 |
+
|
84 |
+
NUM_DEVICES = jax.device_count()
|
85 |
+
|
86 |
+
p_params = replicate(params)
|
87 |
+
|
88 |
+
def replicate_all(prompt_ids, neg_prompt_ids, seed):
|
89 |
+
p_prompt_ids = replicate(prompt_ids)
|
90 |
+
p_neg_prompt_ids = replicate(neg_prompt_ids)
|
91 |
+
rng = jax.random.PRNGKey(seed)
|
92 |
+
rng = jax.random.split(rng, NUM_DEVICES)
|
93 |
+
return p_prompt_ids, p_neg_prompt_ids, rng
|
94 |
+
|
95 |
+
def generate(
|
96 |
+
prompt,
|
97 |
+
negative_prompt,
|
98 |
+
seed=default_seed,
|
99 |
+
guidance_scale=default_guidance_scale,
|
100 |
+
num_inference_steps=default_num_steps,
|
101 |
+
):
|
102 |
+
prompt_ids, neg_prompt_ids = tokenize_prompt(prompt, negative_prompt)
|
103 |
+
prompt_ids, neg_prompt_ids, rng = replicate_all(prompt_ids, neg_prompt_ids, seed)
|
104 |
+
images = pipeline(
|
105 |
+
prompt_ids,
|
106 |
+
p_params,
|
107 |
+
rng,
|
108 |
+
num_inference_steps=num_inference_steps,
|
109 |
+
guidance_scale=guidance_scale,
|
110 |
+
do_classifier_free_guidance=False,
|
111 |
+
jit=True,
|
112 |
+
).images
|
113 |
+
print("images.shape: ", images.shape)
|
114 |
+
# convert the images to PIL
|
115 |
+
images = images.reshape((images.shape[0] * images.shape[1], ) + images.shape[-3:])
|
116 |
+
return pipeline.numpy_to_pil(np.array(images))
|
117 |
+
|
118 |
+
start = time.time()
|
119 |
+
print(f"Compiling ...")
|
120 |
+
generate(default_prompt, default_neg_prompt)
|
121 |
+
print(f"Compiled in {time.time() - start}")
|
122 |
+
|
123 |
+
dts = []
|
124 |
+
i = 0
|
125 |
+
for x in range(2):
|
126 |
+
|
127 |
+
start = time.time()
|
128 |
+
prompt = "Self-portrait oil painting, a beautiful cyborg with golden hair, 8k"
|
129 |
+
neg_prompt = ""
|
130 |
+
|
131 |
+
print(f"Prompt: {prompt}")
|
132 |
+
images = generate(prompt, neg_prompt)
|
133 |
+
t = time.time() - start
|
134 |
+
print(f"Inference in {t}")
|
135 |
+
|
136 |
+
dts.append(t)
|
137 |
+
for img in images:
|
138 |
+
img.save(f'{i:06d}.jpg')
|
139 |
+
i += 1
|
140 |
+
|
141 |
+
mean = np.mean(dts)
|
142 |
+
stdev = np.std(dts)
|
143 |
+
print(f"batches: {i}, Mean {mean:.2f} sec/batch± {stdev * 1.96 / np.sqrt(len(dts)):.2f} (95%)")
|
144 |
+
```
|
145 |
+
|