Fix ZeroGPU: use direct decorator pattern per HF docs

README.md

@@ -8,7 +8,6 @@ sdk_version: 6.2.0
 app_file: app.py
 pinned: false
 license: apache-2.0
-hardware: zero-a10g
 short_description: Find anomalies in logs using AI. Paste logs, get insights.
 tags:
 - log-analysis

app.py

@@ -1,49 +1,17 @@
 """
 Cordon - Semantic Anomaly Detection for Log Files
-HuggingFace Space Demo
+HuggingFace Space Demo (ZeroGPU)
 """
 
 import tempfile
 from pathlib import Path
 
 import gradio as gr
+import spaces
 import torch
 
 from cordon import AnalysisConfig, SemanticLogAnalyzer
 
-# ZeroGPU support with fallback for local development
-try:
-    import spaces
-    ZEROGPU_AVAILABLE = True
-except ImportError:
-    ZEROGPU_AVAILABLE = False
-
-# Detect available device
-def get_device() -> str:
-    """Detect best available device."""
-    if torch.cuda.is_available():
-        return "cuda"
-    elif torch.backends.mps.is_available():
-        return "mps"
-    return "cpu"
-
-# Global model cache (keyed by model_name:device)
-_analyzer_cache: dict[str, SemanticLogAnalyzer] = {}
-
-def get_cached_analyzer(config: AnalysisConfig) -> SemanticLogAnalyzer:
-    """Get or create a cached analyzer instance.
-    
-    Caches the analyzer to avoid repeated model loading.
-    The config is updated on the cached instance for each call.
-    """
-    cache_key = f"{config.model_name}:{config.device}"
-    if cache_key not in _analyzer_cache:
-        _analyzer_cache[cache_key] = SemanticLogAnalyzer(config)
-    else:
-        # Update config params (window_size, k_neighbors, etc.)
-        _analyzer_cache[cache_key].config = config
-    return _analyzer_cache[cache_key]
-
 # Load sample log file
 SAMPLE_LOG_PATH = Path(__file__).parent / "sample_apache.log"
 FALLBACK_LOG_PATH = Path(__file__).parent.parent / "examples" / "apache_sample.log"
@@ -63,7 +31,8 @@ def estimate_tokens(text: str) -> int:
     return len(text) // 4 if text else 0
 
 
-def _analyze_logs_impl(
+@spaces.GPU(duration=120)
+def analyze_logs(
     log_content: str,
     window_size: int,
     k_neighbors: int,
@@ -73,7 +42,11 @@ def _analyze_logs_impl(
     range_max: float,
     progress: gr.Progress = gr.Progress(),
 ) -> tuple[str, str]:
-    """Analyze log content and return results."""
+    """Analyze log content and return results.
+    
+    This function is decorated with @spaces.GPU to run on ZeroGPU.
+    The GPU is allocated when this function is called.
+    """
     
     if not log_content.strip():
         return "", "Anomalous Sections"
@@ -88,7 +61,8 @@ def _analyze_logs_impl(
     try:
         progress(0.1, desc="Configuring analyzer...")
         
-        device = get_device()
+        # Check if CUDA is available (should be True on ZeroGPU)
+        device = "cuda" if torch.cuda.is_available() else "cpu"
         # Use larger batch size on GPU for better throughput
         batch_size = 128 if device == "cuda" else 32
         
@@ -111,25 +85,12 @@ def _analyze_logs_impl(
             )
         
         progress(0.15, desc="Loading embedding model...")
-        analyzer = get_cached_analyzer(config)
-        
-        progress(0.25, desc="Segmenting log into windows...")
-        # Small delay to show progress
-        import time
-        time.sleep(0.1)
+        analyzer = SemanticLogAnalyzer(config)
         
-        progress(0.4, desc="Generating embeddings...")
-        time.sleep(0.1)
-        
-        progress(0.6, desc="Scoring anomalies...")
-        time.sleep(0.1)
-        
-        progress(0.75, desc="Analyzing patterns...")
+        progress(0.4, desc="Analyzing patterns...")
         result = analyzer.analyze_file_detailed(temp_path)
         
         progress(0.9, desc="Formatting output...")
-        time.sleep(0.05)
-        
         progress(1.0, desc="Complete!")
         
         # Calculate token counts
@@ -145,13 +106,6 @@ def _analyze_logs_impl(
         temp_path.unlink(missing_ok=True)
 
 
-# Apply ZeroGPU decorator if available, otherwise use the function directly
-if ZEROGPU_AVAILABLE:
-    analyze_logs = spaces.GPU(duration=120)(_analyze_logs_impl)
-else:
-    analyze_logs = _analyze_logs_impl
-
-
 CUSTOM_CSS = """
 .main, .wrap, .contain, .gradio-container, .app {
     max-width: 100% !important;