Spaces:
Running
on
Zero
Running
on
Zero
Update app.py
Browse files
app.py
CHANGED
@@ -44,7 +44,6 @@ except Exception as e:
|
|
44 |
|
45 |
print("Initialization complete. Gradio is starting...")
|
46 |
|
47 |
-
|
48 |
@spaces.GPU()
|
49 |
def generate(prompt, negative_prompt, width=1024, height=1024, num_inference_steps=30, lora_id=None, progress=gr.Progress(track_tqdm=True)):
|
50 |
|
@@ -56,7 +55,6 @@ def generate(prompt, negative_prompt, width=1024, height=1024, num_inference_ste
|
|
56 |
if causvid_path:
|
57 |
try:
|
58 |
print(f"Loading base LoRA '{BASE_LORA_NAME}'...")
|
59 |
-
# THE CORRECT FIX: Use device_map to load the LoRA directly to the GPU.
|
60 |
pipe.load_lora_weights(causvid_path, adapter_name=BASE_LORA_NAME, device_map={"":device})
|
61 |
active_adapters.append(BASE_LORA_NAME)
|
62 |
adapter_weights.append(1.0)
|
@@ -69,7 +67,6 @@ def generate(prompt, negative_prompt, width=1024, height=1024, num_inference_ste
|
|
69 |
if clean_lora_id:
|
70 |
try:
|
71 |
print(f"Loading custom LoRA '{CUSTOM_LORA_NAME}' from '{clean_lora_id}'...")
|
72 |
-
# THE CORRECT FIX: Also use device_map for the custom LoRA.
|
73 |
pipe.load_lora_weights(clean_lora_id, adapter_name=CUSTOM_LORA_NAME, device_map={"":device})
|
74 |
active_adapters.append(CUSTOM_LORA_NAME)
|
75 |
adapter_weights.append(1.0)
|
@@ -83,6 +80,7 @@ def generate(prompt, negative_prompt, width=1024, height=1024, num_inference_ste
|
|
83 |
if active_adapters:
|
84 |
print(f"Activating adapters: {active_adapters} with weights: {adapter_weights}")
|
85 |
pipe.set_adapters(active_adapters, adapter_weights)
|
|
|
86 |
else:
|
87 |
# Ensure LoRA is disabled if no adapters were loaded
|
88 |
pipe.disable_lora()
|
@@ -104,24 +102,21 @@ def generate(prompt, negative_prompt, width=1024, height=1024, num_inference_ste
|
|
104 |
return Image.fromarray(image)
|
105 |
finally:
|
106 |
# --- PROPER CLEANUP ---
|
107 |
-
# The most reliable way to clean up in this complex environment is to unload ALL LoRAs.
|
108 |
-
# This avoids leaving dangling configs.
|
109 |
print("Unloading all LoRAs to ensure a clean state...")
|
110 |
pipe.unload_lora_weights()
|
111 |
-
gc.collect()
|
112 |
-
torch.cuda.empty_cache()
|
113 |
print("✅ LoRAs unloaded and memory cleaned.")
|
114 |
|
115 |
-
|
116 |
iface = gr.Interface(
|
117 |
fn=generate,
|
118 |
inputs=[
|
119 |
gr.Textbox(label="Input prompt"),
|
120 |
-
gr.Textbox(label="Negative prompt", value
|
121 |
gr.Slider(label="Width", minimum=480, maximum=1280, step=16, value=1024),
|
122 |
gr.Slider(label="Height", minimum=480, maximum=1280, step=16, value=1024),
|
123 |
gr.Slider(minimum=1, maximum=80, step=1, label="Inference Steps", value=10),
|
124 |
-
gr.Textbox(label="LoRA ID (Optional)"),
|
125 |
],
|
126 |
outputs=gr.Image(label="output"),
|
127 |
)
|
|
|
44 |
|
45 |
print("Initialization complete. Gradio is starting...")
|
46 |
|
|
|
47 |
@spaces.GPU()
|
48 |
def generate(prompt, negative_prompt, width=1024, height=1024, num_inference_steps=30, lora_id=None, progress=gr.Progress(track_tqdm=True)):
|
49 |
|
|
|
55 |
if causvid_path:
|
56 |
try:
|
57 |
print(f"Loading base LoRA '{BASE_LORA_NAME}'...")
|
|
|
58 |
pipe.load_lora_weights(causvid_path, adapter_name=BASE_LORA_NAME, device_map={"":device})
|
59 |
active_adapters.append(BASE_LORA_NAME)
|
60 |
adapter_weights.append(1.0)
|
|
|
67 |
if clean_lora_id:
|
68 |
try:
|
69 |
print(f"Loading custom LoRA '{CUSTOM_LORA_NAME}' from '{clean_lora_id}'...")
|
|
|
70 |
pipe.load_lora_weights(clean_lora_id, adapter_name=CUSTOM_LORA_NAME, device_map={"":device})
|
71 |
active_adapters.append(CUSTOM_LORA_NAME)
|
72 |
adapter_weights.append(1.0)
|
|
|
80 |
if active_adapters:
|
81 |
print(f"Activating adapters: {active_adapters} with weights: {adapter_weights}")
|
82 |
pipe.set_adapters(active_adapters, adapter_weights)
|
83 |
+
pipe.transformer.to(device) # Explicitly move transformer to GPU after setting adapters
|
84 |
else:
|
85 |
# Ensure LoRA is disabled if no adapters were loaded
|
86 |
pipe.disable_lora()
|
|
|
102 |
return Image.fromarray(image)
|
103 |
finally:
|
104 |
# --- PROPER CLEANUP ---
|
|
|
|
|
105 |
print("Unloading all LoRAs to ensure a clean state...")
|
106 |
pipe.unload_lora_weights()
|
107 |
+
gc.collect() # Force garbage collection
|
108 |
+
torch.cuda.empty_cache() # Clear CUDA cache
|
109 |
print("✅ LoRAs unloaded and memory cleaned.")
|
110 |
|
|
|
111 |
iface = gr.Interface(
|
112 |
fn=generate,
|
113 |
inputs=[
|
114 |
gr.Textbox(label="Input prompt"),
|
115 |
+
gr.Textbox(label="Negative prompt", value="Bright tones, overexposed, static, blurred details, subtitles, style, works, paintings, images, static, overall gray, worst quality, low quality, JPEG compression residue, ugly, incomplete, extra fingers, poorly drawn hands, poorly drawn faces, deformed, disfigured, misshapen limbs, fused fingers, still picture, messy background, three legs, many people in the background, walking backwards"),
|
116 |
gr.Slider(label="Width", minimum=480, maximum=1280, step=16, value=1024),
|
117 |
gr.Slider(label="Height", minimum=480, maximum=1280, step=16, value=1024),
|
118 |
gr.Slider(minimum=1, maximum=80, step=1, label="Inference Steps", value=10),
|
119 |
+
gr.Textbox(label="LoRA ID (Optional, loads dynamically)"),
|
120 |
],
|
121 |
outputs=gr.Image(label="output"),
|
122 |
)
|