Spaces:
Running
on
Zero
Running
on
Zero
Update optimized.py
Browse files- optimized.py +53 -40
optimized.py
CHANGED
@@ -8,70 +8,83 @@ from accelerate import init_empty_weights
|
|
8 |
|
9 |
huggingface_token = os.getenv("HUGGINFACE_TOKEN")
|
10 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
11 |
|
12 |
-
|
13 |
-
good_vae = AutoencoderKL.from_pretrained("black-forest-labs/FLUX.1-dev", subfolder="vae",
|
14 |
-
torch_dtype=torch.bfloat16,
|
15 |
-
# variant="4bit",
|
16 |
-
device_map="balanced",
|
17 |
-
use_safetensors=True,
|
18 |
-
token=huggingface_token).to("cuda")
|
19 |
-
|
20 |
-
# Load pipeline
|
21 |
controlnet = FluxControlNetModel.from_pretrained(
|
22 |
"jasperai/Flux.1-dev-Controlnet-Upscaler",
|
23 |
torch_dtype=torch.bfloat16
|
24 |
)
|
25 |
-
|
|
|
26 |
pipe = FluxControlNetPipeline.from_pretrained(
|
27 |
"LPX55/FLUX.1-merged_uncensored",
|
28 |
controlnet=controlnet,
|
29 |
-
torch_dtype=torch.bfloat16,
|
30 |
-
device_map="balanced",
|
31 |
vae=good_vae,
|
32 |
-
|
|
|
|
|
33 |
token=huggingface_token
|
34 |
)
|
35 |
-
|
36 |
-
#
|
37 |
-
|
38 |
-
|
39 |
-
|
|
|
|
|
40 |
try:
|
41 |
import xformers
|
42 |
pipe.enable_xformers_memory_efficient_attention()
|
43 |
except ImportError:
|
44 |
print("XFormers missing! Using PyTorch attention instead")
|
45 |
-
# Fallback to PyTorch 2.0+ memory efficient attention
|
46 |
pipe.enable_sdp_attention()
|
47 |
torch.backends.cuda.enable_flash_sdp(True)
|
48 |
-
# Convert all models to memory-efficient format
|
49 |
-
#pipe.to(memory_format=torch.channels_last)
|
50 |
-
pipe.to("cuda")
|
51 |
|
|
|
|
|
|
|
52 |
@spaces.GPU
|
53 |
def generate_image(prompt, scale, steps, control_image, controlnet_conditioning_scale, guidance_scale):
|
54 |
-
#
|
55 |
-
control_image = control_image.resize((int(w * scale), int(h * scale)), PIL.Image.BICUBIC)
|
56 |
-
# control_image = load_image(control_image)
|
57 |
w, h = control_image.size
|
58 |
-
#
|
59 |
-
|
60 |
-
|
61 |
-
|
62 |
-
|
63 |
-
|
64 |
-
|
65 |
-
|
66 |
-
|
67 |
-
|
68 |
-
|
69 |
-
|
70 |
-
|
71 |
-
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
72 |
torch.cuda.empty_cache()
|
|
|
|
|
73 |
return image
|
74 |
-
|
75 |
# Create Gradio interface
|
76 |
iface = gr.Interface(
|
77 |
fn=generate_image,
|
|
|
8 |
|
9 |
huggingface_token = os.getenv("HUGGINFACE_TOKEN")
|
10 |
|
11 |
+
good_vae = AutoencoderKL.from_pretrained(
|
12 |
+
"black-forest-labs/FLUX.1-dev",
|
13 |
+
subfolder="vae",
|
14 |
+
torch_dtype=torch.bfloat16,
|
15 |
+
use_safetensors=True,
|
16 |
+
device_map=None, # Disable automatic mapping
|
17 |
+
token=huggingface_token
|
18 |
+
)
|
19 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
20 |
controlnet = FluxControlNetModel.from_pretrained(
|
21 |
"jasperai/Flux.1-dev-Controlnet-Upscaler",
|
22 |
torch_dtype=torch.bfloat16
|
23 |
)
|
24 |
+
|
25 |
+
# Initialize pipeline without automatic device mapping
|
26 |
pipe = FluxControlNetPipeline.from_pretrained(
|
27 |
"LPX55/FLUX.1-merged_uncensored",
|
28 |
controlnet=controlnet,
|
|
|
|
|
29 |
vae=good_vae,
|
30 |
+
torch_dtype=torch.bfloat16,
|
31 |
+
use_safetensors=True,
|
32 |
+
device_map=None, # Disable automatic device mapping
|
33 |
token=huggingface_token
|
34 |
)
|
35 |
+
print(f"VRAM used: {torch.cuda.memory_allocated()/1e9:.2f}GB")
|
36 |
+
# Proper CPU offloading sequence
|
37 |
+
pipe.enable_model_cpu_offload(device="cuda") # First enable offloading
|
38 |
+
pipe.enable_vae_slicing() # Then enable memory optimizations
|
39 |
+
pipe.enable_attention_slicing(1)
|
40 |
+
|
41 |
+
# Handle xformers/SDP attention after offloading
|
42 |
try:
|
43 |
import xformers
|
44 |
pipe.enable_xformers_memory_efficient_attention()
|
45 |
except ImportError:
|
46 |
print("XFormers missing! Using PyTorch attention instead")
|
|
|
47 |
pipe.enable_sdp_attention()
|
48 |
torch.backends.cuda.enable_flash_sdp(True)
|
|
|
|
|
|
|
49 |
|
50 |
+
# Memory format optimization (only after other configs)
|
51 |
+
pipe.to(memory_format=torch.channels_last)
|
52 |
+
print(f"VRAM used: {torch.cuda.memory_allocated()/1e9:.2f}GB")
|
53 |
@spaces.GPU
|
54 |
def generate_image(prompt, scale, steps, control_image, controlnet_conditioning_scale, guidance_scale):
|
55 |
+
# Clean up input handling
|
|
|
|
|
56 |
w, h = control_image.size
|
57 |
+
scale = min(scale, 2.0) # Cap scale factor
|
58 |
+
|
59 |
+
# Size calculation with safety limits
|
60 |
+
max_dim = 1536 # Set based on your VRAM
|
61 |
+
target_w = min(int(w * scale), max_dim)
|
62 |
+
target_h = min(int(h * scale), max_dim)
|
63 |
+
|
64 |
+
control_image = control_image.resize(
|
65 |
+
(target_w, target_h),
|
66 |
+
PIL.Image.BICUBIC
|
67 |
+
)
|
68 |
+
|
69 |
+
# Generation with memory-friendly parameters
|
70 |
+
with torch.autocast("cuda"): # Mixed precision
|
71 |
+
image = pipe(
|
72 |
+
prompt=prompt,
|
73 |
+
control_image=control_image,
|
74 |
+
controlnet_conditioning_scale=controlnet_conditioning_scale,
|
75 |
+
num_inference_steps=steps,
|
76 |
+
guidance_scale=guidance_scale,
|
77 |
+
height=target_h,
|
78 |
+
width=target_w,
|
79 |
+
output_type="pil", # Avoid extra latent decoding steps
|
80 |
+
generator=torch.Generator(device="cuda").manual_seed(0)
|
81 |
+
).images[0]
|
82 |
+
print(f"VRAM used: {torch.cuda.memory_allocated()/1e9:.2f}GB")
|
83 |
+
# Aggressive memory cleanup
|
84 |
torch.cuda.empty_cache()
|
85 |
+
torch.cuda.ipc_collect()
|
86 |
+
print(f"VRAM used: {torch.cuda.memory_allocated()/1e9:.2f}GB")
|
87 |
return image
|
|
|
88 |
# Create Gradio interface
|
89 |
iface = gr.Interface(
|
90 |
fn=generate_image,
|