Spaces:
Running
on
A10G
Running
on
A10G
fix diffusers
Browse files
app.py
CHANGED
@@ -1,11 +1,7 @@
|
|
1 |
import numpy as np
|
2 |
import PIL.Image
|
3 |
import torch
|
4 |
-
from typing import List
|
5 |
-
from diffusers.utils import numpy_to_pil
|
6 |
from diffusers import StableCascadeDecoderPipeline, StableCascadePriorPipeline
|
7 |
-
from diffusers.pipelines.wuerstchen import DEFAULT_STAGE_C_TIMESTEPS
|
8 |
-
from fastapi import FastAPI
|
9 |
import uvicorn
|
10 |
from fastapi.middleware.cors import CORSMiddleware
|
11 |
from fastapi.responses import RedirectResponse, StreamingResponse
|
@@ -16,7 +12,6 @@ from db import Database
|
|
16 |
import uuid
|
17 |
import logging
|
18 |
from fastapi import FastAPI, Request, HTTPException
|
19 |
-
from fastapi.middleware.cors import CORSMiddleware
|
20 |
from asyncio import Lock
|
21 |
|
22 |
|
@@ -40,13 +35,11 @@ dtype = torch.bfloat16
|
|
40 |
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
|
41 |
if torch.cuda.is_available():
|
42 |
prior_pipeline = StableCascadePriorPipeline.from_pretrained(
|
43 |
-
"stabilityai/stable-cascade-prior", torch_dtype=
|
44 |
-
)
|
45 |
decoder_pipeline = StableCascadeDecoderPipeline.from_pretrained(
|
46 |
-
"stabilityai/stable-cascade", torch_dtype=
|
47 |
-
)
|
48 |
-
prior_pipeline.to(device)
|
49 |
-
decoder_pipeline.to(device)
|
50 |
|
51 |
if USE_TORCH_COMPILE:
|
52 |
prior_pipeline.prior = torch.compile(
|
@@ -67,16 +60,14 @@ def generate(
|
|
67 |
prior_guidance_scale: float = 4.0,
|
68 |
decoder_num_inference_steps: int = 10,
|
69 |
decoder_guidance_scale: float = 0.0,
|
70 |
-
num_images_per_prompt: int =
|
71 |
) -> PIL.Image.Image:
|
72 |
-
|
73 |
generator = torch.Generator().manual_seed(seed)
|
74 |
prior_output = prior_pipeline(
|
75 |
prompt=prompt,
|
76 |
height=height,
|
77 |
width=width,
|
78 |
num_inference_steps=prior_num_inference_steps,
|
79 |
-
timesteps=DEFAULT_STAGE_C_TIMESTEPS,
|
80 |
negative_prompt=negative_prompt,
|
81 |
guidance_scale=prior_guidance_scale,
|
82 |
num_images_per_prompt=num_images_per_prompt,
|
@@ -133,7 +124,6 @@ async def generate_image(
|
|
133 |
|
134 |
logging.info(f"Image not found in cache, generating new image")
|
135 |
async with generate_lock:
|
136 |
-
|
137 |
pil_image = generate(prompt, negative_prompt, seed)
|
138 |
img_id = str(uuid.uuid4())
|
139 |
img_path = IMGS_PATH / f"{img_id}.jpg"
|
|
|
1 |
import numpy as np
|
2 |
import PIL.Image
|
3 |
import torch
|
|
|
|
|
4 |
from diffusers import StableCascadeDecoderPipeline, StableCascadePriorPipeline
|
|
|
|
|
5 |
import uvicorn
|
6 |
from fastapi.middleware.cors import CORSMiddleware
|
7 |
from fastapi.responses import RedirectResponse, StreamingResponse
|
|
|
12 |
import uuid
|
13 |
import logging
|
14 |
from fastapi import FastAPI, Request, HTTPException
|
|
|
15 |
from asyncio import Lock
|
16 |
|
17 |
|
|
|
35 |
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
|
36 |
if torch.cuda.is_available():
|
37 |
prior_pipeline = StableCascadePriorPipeline.from_pretrained(
|
38 |
+
"stabilityai/stable-cascade-prior", variant="bf16", torch_dtype=torch.bfloat16
|
39 |
+
).to(device)
|
40 |
decoder_pipeline = StableCascadeDecoderPipeline.from_pretrained(
|
41 |
+
"stabilityai/stable-cascade", variant="bf16", torch_dtype=torch.bfloat16
|
42 |
+
).to(device)
|
|
|
|
|
43 |
|
44 |
if USE_TORCH_COMPILE:
|
45 |
prior_pipeline.prior = torch.compile(
|
|
|
60 |
prior_guidance_scale: float = 4.0,
|
61 |
decoder_num_inference_steps: int = 10,
|
62 |
decoder_guidance_scale: float = 0.0,
|
63 |
+
num_images_per_prompt: int = 1,
|
64 |
) -> PIL.Image.Image:
|
|
|
65 |
generator = torch.Generator().manual_seed(seed)
|
66 |
prior_output = prior_pipeline(
|
67 |
prompt=prompt,
|
68 |
height=height,
|
69 |
width=width,
|
70 |
num_inference_steps=prior_num_inference_steps,
|
|
|
71 |
negative_prompt=negative_prompt,
|
72 |
guidance_scale=prior_guidance_scale,
|
73 |
num_images_per_prompt=num_images_per_prompt,
|
|
|
124 |
|
125 |
logging.info(f"Image not found in cache, generating new image")
|
126 |
async with generate_lock:
|
|
|
127 |
pil_image = generate(prompt, negative_prompt, seed)
|
128 |
img_id = str(uuid.uuid4())
|
129 |
img_path = IMGS_PATH / f"{img_id}.jpg"
|