Spaces:
Running
on
Zero
Running
on
Zero
Joseph Pollack
commited on
skip examples caching
Browse files
app.py
CHANGED
|
@@ -21,6 +21,7 @@ import os
|
|
| 21 |
HF_TOKEN = os.getenv("HF_TOKEN")
|
| 22 |
if not HF_TOKEN:
|
| 23 |
logger.warning("HF_TOKEN not found in environment variables. Model access may be restricted.")
|
|
|
|
| 24 |
|
| 25 |
class LOperatorDemo:
|
| 26 |
def __init__(self):
|
|
@@ -29,15 +30,18 @@ class LOperatorDemo:
|
|
| 29 |
self.is_loaded = False
|
| 30 |
|
| 31 |
def load_model(self):
|
| 32 |
-
"""Load the L-Operator model and processor"""
|
| 33 |
try:
|
|
|
|
|
|
|
| 34 |
logger.info(f"Loading model {MODEL_ID} on device {DEVICE}")
|
| 35 |
-
|
| 36 |
# Check if token is available
|
| 37 |
if not HF_TOKEN:
|
| 38 |
return "β HF_TOKEN not found. Please set HF_TOKEN in Spaces secrets."
|
| 39 |
-
|
| 40 |
-
# Load model
|
|
|
|
| 41 |
self.model = AutoModelForImageTextToText.from_pretrained(
|
| 42 |
MODEL_ID,
|
| 43 |
device_map="auto",
|
|
@@ -46,17 +50,20 @@ class LOperatorDemo:
|
|
| 46 |
)
|
| 47 |
|
| 48 |
# Load processor
|
|
|
|
| 49 |
self.processor = AutoProcessor.from_pretrained(
|
| 50 |
MODEL_ID,
|
| 51 |
trust_remote_code=True
|
| 52 |
-
)
|
|
|
|
| 53 |
if DEVICE == "cpu":
|
| 54 |
self.model = self.model.to(DEVICE)
|
| 55 |
-
|
| 56 |
self.is_loaded = True
|
| 57 |
-
|
| 58 |
-
|
| 59 |
-
|
|
|
|
| 60 |
except Exception as e:
|
| 61 |
logger.error(f"Error loading model: {str(e)}")
|
| 62 |
return f"β Error loading model: {str(e)} - This may be a custom model requiring special handling"
|
|
@@ -163,74 +170,72 @@ class LOperatorDemo:
|
|
| 163 |
# Initialize demo
|
| 164 |
demo_instance = LOperatorDemo()
|
| 165 |
|
| 166 |
-
#
|
| 167 |
-
|
| 168 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 169 |
try:
|
| 170 |
-
logger.info("
|
| 171 |
result = demo_instance.load_model()
|
| 172 |
-
logger.info(f"
|
| 173 |
return result
|
|
|
|
|
|
|
|
|
|
| 174 |
except Exception as e:
|
| 175 |
-
logger.error(f"Error
|
| 176 |
-
return f"β Error
|
| 177 |
-
|
| 178 |
-
#
|
| 179 |
-
|
| 180 |
-
|
| 181 |
-
print("β
Model loading completed!")
|
| 182 |
|
| 183 |
-
# Load example episodes
|
| 184 |
def load_example_episodes():
|
| 185 |
-
"""Load example episodes from the extracted data
|
| 186 |
examples = []
|
| 187 |
|
| 188 |
try:
|
| 189 |
-
# Load episode
|
| 190 |
-
|
| 191 |
-
|
| 192 |
-
|
| 193 |
-
# Load episode 53
|
| 194 |
-
with open("extracted_episodes_duckdb/episode_53/metadata.json", "r") as f:
|
| 195 |
-
episode_53 = json.load(f)
|
| 196 |
|
| 197 |
-
|
| 198 |
-
with open("extracted_episodes_duckdb/episode_73/metadata.json", "r") as f:
|
| 199 |
-
episode_73 = json.load(f)
|
| 200 |
-
|
| 201 |
-
# Create examples with simple identifiers
|
| 202 |
-
examples = [
|
| 203 |
-
(
|
| 204 |
-
"extracted_episodes_duckdb/episode_13/screenshots/screenshot_1.png",
|
| 205 |
-
f"Episode 13: {episode_13.get('goal', 'Navigate app interface')[:50]}..."
|
| 206 |
-
),
|
| 207 |
-
(
|
| 208 |
-
"extracted_episodes_duckdb/episode_53/screenshots/screenshot_1.png",
|
| 209 |
-
f"Episode 53: {episode_53.get('goal', 'App interaction example')[:50]}..."
|
| 210 |
-
),
|
| 211 |
-
(
|
| 212 |
-
"extracted_episodes_duckdb/episode_73/screenshots/screenshot_1.png",
|
| 213 |
-
f"Episode 73: {episode_73.get('goal', 'Device control task')[:50]}..."
|
| 214 |
-
)
|
| 215 |
-
]
|
| 216 |
-
|
| 217 |
-
# Validate each example by checking if image file exists and is readable
|
| 218 |
-
for image_path, description in examples:
|
| 219 |
try:
|
| 220 |
-
|
| 221 |
-
|
| 222 |
-
|
| 223 |
-
|
| 224 |
-
|
| 225 |
-
|
| 226 |
-
logger.warning(f"Skipping invalid image {image_path}: {str(img_error)}")
|
| 227 |
continue
|
| 228 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 229 |
except Exception as e:
|
| 230 |
logger.error(f"Error loading examples: {str(e)}")
|
| 231 |
examples = []
|
| 232 |
|
| 233 |
-
logger.info(f"Loaded {len(examples)}
|
| 234 |
return examples
|
| 235 |
|
| 236 |
# Create Gradio interface
|
|
@@ -322,7 +327,8 @@ def create_demo():
|
|
| 322 |
title="L-Operator Chat",
|
| 323 |
description="Chat with L-Operator using screenshots and text instructions",
|
| 324 |
examples=load_example_episodes(),
|
| 325 |
-
type="messages"
|
|
|
|
| 326 |
)
|
| 327 |
|
| 328 |
gr.Markdown("### π― Action Output")
|
|
@@ -349,20 +355,26 @@ def create_demo():
|
|
| 349 |
except:
|
| 350 |
return {"raw_response": response}
|
| 351 |
|
| 352 |
-
# Update model status on page load
|
| 353 |
def update_model_status():
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 354 |
if demo_instance.is_loaded:
|
| 355 |
return "β
L-Operator model loaded and ready!"
|
| 356 |
else:
|
| 357 |
return "β Model failed to load. Please check logs."
|
| 358 |
-
|
| 359 |
generate_btn.click(
|
| 360 |
fn=on_generate_action,
|
| 361 |
inputs=[image_input, goal_input, step_input],
|
| 362 |
outputs=action_output
|
| 363 |
)
|
| 364 |
-
|
| 365 |
-
#
|
| 366 |
demo.load(
|
| 367 |
fn=update_model_status,
|
| 368 |
outputs=model_status
|
|
@@ -404,14 +416,23 @@ def create_demo():
|
|
| 404 |
|
| 405 |
return demo
|
| 406 |
|
| 407 |
-
# Create and launch the demo
|
| 408 |
if __name__ == "__main__":
|
| 409 |
-
|
| 410 |
-
|
| 411 |
-
|
| 412 |
-
|
| 413 |
-
|
| 414 |
-
|
| 415 |
-
|
| 416 |
-
|
| 417 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 21 |
HF_TOKEN = os.getenv("HF_TOKEN")
|
| 22 |
if not HF_TOKEN:
|
| 23 |
logger.warning("HF_TOKEN not found in environment variables. Model access may be restricted.")
|
| 24 |
+
logger.warning("Please set HF_TOKEN in your environment variables or Spaces secrets.")
|
| 25 |
|
| 26 |
class LOperatorDemo:
|
| 27 |
def __init__(self):
|
|
|
|
| 30 |
self.is_loaded = False
|
| 31 |
|
| 32 |
def load_model(self):
|
| 33 |
+
"""Load the L-Operator model and processor with timeout handling"""
|
| 34 |
try:
|
| 35 |
+
import time
|
| 36 |
+
start_time = time.time()
|
| 37 |
logger.info(f"Loading model {MODEL_ID} on device {DEVICE}")
|
| 38 |
+
|
| 39 |
# Check if token is available
|
| 40 |
if not HF_TOKEN:
|
| 41 |
return "β HF_TOKEN not found. Please set HF_TOKEN in Spaces secrets."
|
| 42 |
+
|
| 43 |
+
# Load model with progress logging
|
| 44 |
+
logger.info("Downloading and loading model weights...")
|
| 45 |
self.model = AutoModelForImageTextToText.from_pretrained(
|
| 46 |
MODEL_ID,
|
| 47 |
device_map="auto",
|
|
|
|
| 50 |
)
|
| 51 |
|
| 52 |
# Load processor
|
| 53 |
+
logger.info("Loading processor...")
|
| 54 |
self.processor = AutoProcessor.from_pretrained(
|
| 55 |
MODEL_ID,
|
| 56 |
trust_remote_code=True
|
| 57 |
+
)
|
| 58 |
+
|
| 59 |
if DEVICE == "cpu":
|
| 60 |
self.model = self.model.to(DEVICE)
|
| 61 |
+
|
| 62 |
self.is_loaded = True
|
| 63 |
+
load_time = time.time() - start_time
|
| 64 |
+
logger.info(".1f")
|
| 65 |
+
return ".1f"
|
| 66 |
+
|
| 67 |
except Exception as e:
|
| 68 |
logger.error(f"Error loading model: {str(e)}")
|
| 69 |
return f"β Error loading model: {str(e)} - This may be a custom model requiring special handling"
|
|
|
|
| 170 |
# Initialize demo
|
| 171 |
demo_instance = LOperatorDemo()
|
| 172 |
|
| 173 |
+
def load_model_with_timeout(timeout_seconds=600): # 10 minutes timeout
|
| 174 |
+
"""Load model with timeout protection"""
|
| 175 |
+
import signal
|
| 176 |
+
import time
|
| 177 |
+
|
| 178 |
+
def timeout_handler(signum, frame):
|
| 179 |
+
raise TimeoutError("Model loading timed out")
|
| 180 |
+
|
| 181 |
+
# Set up the signal handler for timeout
|
| 182 |
+
old_handler = signal.signal(signal.SIGALRM, timeout_handler)
|
| 183 |
+
signal.alarm(timeout_seconds)
|
| 184 |
+
|
| 185 |
try:
|
| 186 |
+
logger.info("Loading L-Operator model with timeout protection...")
|
| 187 |
result = demo_instance.load_model()
|
| 188 |
+
logger.info(f"Model loading result: {result}")
|
| 189 |
return result
|
| 190 |
+
except TimeoutError:
|
| 191 |
+
logger.error("Model loading timed out - this may be due to network issues or large model size")
|
| 192 |
+
return "β Model loading timed out. Please try again or check your internet connection."
|
| 193 |
except Exception as e:
|
| 194 |
+
logger.error(f"Error loading model: {str(e)}")
|
| 195 |
+
return f"β Error loading model: {str(e)}"
|
| 196 |
+
finally:
|
| 197 |
+
# Restore the original signal handler
|
| 198 |
+
signal.alarm(0)
|
| 199 |
+
signal.signal(signal.SIGALRM, old_handler)
|
|
|
|
| 200 |
|
| 201 |
+
# Load example episodes (lazy loading to avoid startup timeout)
|
| 202 |
def load_example_episodes():
|
| 203 |
+
"""Load example episodes from the extracted data - simplified for fast startup"""
|
| 204 |
examples = []
|
| 205 |
|
| 206 |
try:
|
| 207 |
+
# Load episode metadata quickly without PIL validation
|
| 208 |
+
episodes_data = []
|
| 209 |
+
episode_dirs = ["episode_13", "episode_53", "episode_73"]
|
|
|
|
|
|
|
|
|
|
|
|
|
| 210 |
|
| 211 |
+
for episode_dir in episode_dirs:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 212 |
try:
|
| 213 |
+
metadata_path = f"extracted_episodes_duckdb/{episode_dir}/metadata.json"
|
| 214 |
+
with open(metadata_path, "r") as f:
|
| 215 |
+
metadata = json.load(f)
|
| 216 |
+
episodes_data.append(metadata)
|
| 217 |
+
except Exception as e:
|
| 218 |
+
logger.warning(f"Could not load metadata for {episode_dir}: {str(e)}")
|
|
|
|
| 219 |
continue
|
| 220 |
|
| 221 |
+
# Create examples with simple path checks (no PIL validation)
|
| 222 |
+
for i, metadata in enumerate(episodes_data):
|
| 223 |
+
episode_num = ["13", "53", "73"][i]
|
| 224 |
+
image_path = f"extracted_episodes_duckdb/episode_{episode_num}/screenshots/screenshot_1.png"
|
| 225 |
+
|
| 226 |
+
# Simple file existence check instead of PIL validation
|
| 227 |
+
if os.path.exists(image_path):
|
| 228 |
+
goal_text = metadata.get('goal', f'Episode {episode_num} example')
|
| 229 |
+
examples.append([
|
| 230 |
+
image_path,
|
| 231 |
+
f"Episode {episode_num}: {goal_text[:50]}..."
|
| 232 |
+
])
|
| 233 |
+
|
| 234 |
except Exception as e:
|
| 235 |
logger.error(f"Error loading examples: {str(e)}")
|
| 236 |
examples = []
|
| 237 |
|
| 238 |
+
logger.info(f"Loaded {len(examples)} examples (without validation for faster startup)")
|
| 239 |
return examples
|
| 240 |
|
| 241 |
# Create Gradio interface
|
|
|
|
| 327 |
title="L-Operator Chat",
|
| 328 |
description="Chat with L-Operator using screenshots and text instructions",
|
| 329 |
examples=load_example_episodes(),
|
| 330 |
+
type="messages",
|
| 331 |
+
cache_examples=False
|
| 332 |
)
|
| 333 |
|
| 334 |
gr.Markdown("### π― Action Output")
|
|
|
|
| 355 |
except:
|
| 356 |
return {"raw_response": response}
|
| 357 |
|
| 358 |
+
# Update model status on page load (with timeout-protected model loading)
|
| 359 |
def update_model_status():
|
| 360 |
+
if not demo_instance.is_loaded:
|
| 361 |
+
logger.info("Loading model on Gradio startup with timeout protection...")
|
| 362 |
+
result = load_model_with_timeout(timeout_seconds=900) # 15 minutes for Spaces
|
| 363 |
+
logger.info(f"Model loading result: {result}")
|
| 364 |
+
return result
|
| 365 |
+
|
| 366 |
if demo_instance.is_loaded:
|
| 367 |
return "β
L-Operator model loaded and ready!"
|
| 368 |
else:
|
| 369 |
return "β Model failed to load. Please check logs."
|
| 370 |
+
|
| 371 |
generate_btn.click(
|
| 372 |
fn=on_generate_action,
|
| 373 |
inputs=[image_input, goal_input, step_input],
|
| 374 |
outputs=action_output
|
| 375 |
)
|
| 376 |
+
|
| 377 |
+
# Load model and update status on page load
|
| 378 |
demo.load(
|
| 379 |
fn=update_model_status,
|
| 380 |
outputs=model_status
|
|
|
|
| 416 |
|
| 417 |
return demo
|
| 418 |
|
| 419 |
+
# Create and launch the demo with optimized settings
|
| 420 |
if __name__ == "__main__":
|
| 421 |
+
try:
|
| 422 |
+
logger.info("Creating Gradio demo interface...")
|
| 423 |
+
demo = create_demo()
|
| 424 |
+
|
| 425 |
+
logger.info("Launching Gradio server...")
|
| 426 |
+
demo.launch(
|
| 427 |
+
server_name="0.0.0.0",
|
| 428 |
+
server_port=7860,
|
| 429 |
+
share=False,
|
| 430 |
+
debug=False, # Disable debug to reduce startup time
|
| 431 |
+
show_error=True,
|
| 432 |
+
ssr_mode=False,
|
| 433 |
+
max_threads=2, # Limit threads to prevent resource exhaustion
|
| 434 |
+
quiet=True # Reduce startup logging noise
|
| 435 |
+
)
|
| 436 |
+
except Exception as e:
|
| 437 |
+
logger.error(f"Failed to launch Gradio app: {str(e)}")
|
| 438 |
+
raise
|