Spaces:
Running
on
Zero
Running
on
Zero
Update optimized.py
Browse files- optimized.py +20 -20
optimized.py
CHANGED
|
@@ -8,14 +8,7 @@ from accelerate import init_empty_weights
|
|
| 8 |
|
| 9 |
huggingface_token = os.getenv("HUGGINFACE_TOKEN")
|
| 10 |
|
| 11 |
-
|
| 12 |
-
import xformers
|
| 13 |
-
pipe.enable_xformers_memory_efficient_attention()
|
| 14 |
-
except ImportError:
|
| 15 |
-
print("XFormers missing! Using PyTorch attention instead")
|
| 16 |
-
# Fallback to PyTorch 2.0+ memory efficient attention
|
| 17 |
-
pipe.enable_sdp_attention()
|
| 18 |
-
torch.backends.cuda.enable_flash_sdp(True)
|
| 19 |
|
| 20 |
good_vae = AutoencoderKL.from_pretrained("black-forest-labs/FLUX.1-dev", subfolder="vae",
|
| 21 |
torch_dtype=torch.bfloat16,
|
|
@@ -29,24 +22,31 @@ controlnet = FluxControlNetModel.from_pretrained(
|
|
| 29 |
"jasperai/Flux.1-dev-Controlnet-Upscaler",
|
| 30 |
torch_dtype=torch.bfloat16
|
| 31 |
)
|
| 32 |
-
with init_empty_weights():
|
| 33 |
-
|
| 34 |
-
|
| 35 |
-
|
| 36 |
-
|
| 37 |
-
|
| 38 |
-
|
| 39 |
-
|
| 40 |
-
|
| 41 |
-
|
| 42 |
pipe.enable_model_cpu_offload(device="cuda")
|
| 43 |
# Add to your pipeline initialization:
|
| 44 |
# pipe.enable_xformers_memory_efficient_attention()
|
| 45 |
# pipe.enable_vae_slicing() # Batch processing of VAE
|
| 46 |
# pipe.enable_model_cpu_offload() # Use with accelerate
|
| 47 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 48 |
# Convert all models to memory-efficient format
|
| 49 |
-
pipe.to(memory_format=torch.channels_last)
|
| 50 |
pipe.to("cuda")
|
| 51 |
|
| 52 |
@spaces.GPU
|
|
|
|
| 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,
|
|
|
|
| 22 |
"jasperai/Flux.1-dev-Controlnet-Upscaler",
|
| 23 |
torch_dtype=torch.bfloat16
|
| 24 |
)
|
| 25 |
+
#with init_empty_weights():
|
| 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 |
+
use_safetensors=True,
|
| 33 |
+
token=huggingface_token
|
| 34 |
+
)
|
| 35 |
pipe.enable_model_cpu_offload(device="cuda")
|
| 36 |
# Add to your pipeline initialization:
|
| 37 |
# pipe.enable_xformers_memory_efficient_attention()
|
| 38 |
# pipe.enable_vae_slicing() # Batch processing of VAE
|
| 39 |
# pipe.enable_model_cpu_offload() # Use with accelerate
|
| 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
|