Update langgraph_nodes.py

langgraph_nodes.py

@@ -1,6 +1,6 @@
 """
-LangGraph Nodes - FIXED VERSION
-Better error handling and debugging for API calls
+LangGraph Nodes - LAZY LOADING VERSION
+Initializes HF client when needed, not at module import
 """
 
 import os
@@ -18,17 +18,29 @@ warnings.filterwarnings('ignore')
 from langgraph_state import ReviewState, BatchState
 from database_enhanced import EnhancedDatabase
 
-# FIXED: Initialize HF client with better error handling
-HF_TOKEN = os.getenv("HUGGINGFACE_API_KEY")
+# FIXED: Don't initialize client at module import
+# Initialize LAZILY when first needed
+_hf_client = None
+
+def get_hf_client():
+    """Get or initialize HuggingFace client (lazy loading)"""
+    global _hf_client
+    
+    if _hf_client is not None:
+        return _hf_client
+    
+    # Try to get token from environment
+    HF_TOKEN = os.getenv("HUGGINGFACE_API_KEY")
+    
+    if not HF_TOKEN or HF_TOKEN.strip() == "":
+        # No token available
+        return None
+    
+    # Initialize client with token
+    print(f"✅ Initializing HF client with token: {HF_TOKEN[:10]}...")
+    _hf_client = InferenceClient(token=HF_TOKEN)
+    return _hf_client
 
-# Check if token exists
-if not HF_TOKEN or HF_TOKEN.strip() == "":
-    print("❌ WARNING: HUGGINGFACE_API_KEY not set!")
-    print("   API calls will fail. Please set your API key.")
-    hf_client = None
-else:
-    print(f"✅ HF Token found: {HF_TOKEN[:8]}...")
-    hf_client = InferenceClient(token=HF_TOKEN)
 
 # Initialize sentiment models (singleton) - load once
 _sentiment_models_loaded = False
@@ -51,9 +63,12 @@ def load_sentiment_models():
     _best_model = AutoModelForSequenceClassification.from_pretrained("cardiffnlp/twitter-roberta-base-sentiment-latest")
     _best_model.eval()
     
-    # Alternate Model
+    # Alternate Model - FIXED: Proper loading
     _alt_tokenizer = AutoTokenizer.from_pretrained("finiteautomata/bertweet-base-sentiment-analysis")
-    _alt_model = AutoModelForSequenceClassification.from_pretrained("finiteautomata/bertweet-base-sentiment-analysis")
+    _alt_model = AutoModelForSequenceClassification.from_pretrained(
+        "finiteautomata/bertweet-base-sentiment-analysis",
+        torch_dtype=torch.float32  # FIXED: Explicit dtype to avoid meta tensors
+    )
     _alt_model.eval()
     
     _sentiment_models_loaded = True
@@ -67,10 +82,10 @@ def load_sentiment_models():
 def llm1_classify(review: Dict[str, Any]) -> Dict[str, Any]:
     """LLM1: Type, Department, Priority classification"""
     
-    # FIXED: Check if client exists
+    # FIXED: Get client lazily
+    hf_client = get_hf_client()
+    
     if hf_client is None:
-        print("❌ ERROR: HuggingFace client not initialized!")
-        print("   Make sure HUGGINGFACE_API_KEY environment variable is set")
         return {
             'type': 'unknown',
             'department': 'unknown',
@@ -124,7 +139,6 @@ Respond ONLY in valid JSON format:
 }}"""
 
     try:
-        # FIXED: Better error logging
         print(f"   🔍 Calling Qwen API...")
         
         response = hf_client.text_generation(
@@ -151,10 +165,7 @@ Respond ONLY in valid JSON format:
         return result
         
     except Exception as e:
-        # FIXED: Show the actual error
         print(f"❌ LLM1 ERROR: {type(e).__name__}: {str(e)}")
-        import traceback
-        traceback.print_exc()
         
         return {
             'type': 'unknown',
@@ -169,9 +180,10 @@ Respond ONLY in valid JSON format:
 def llm2_analyze(review: Dict[str, Any]) -> Dict[str, Any]:
     """LLM2: User type, Emotion, Context analysis"""
     
-    # FIXED: Check if client exists
+    # FIXED: Get client lazily
+    hf_client = get_hf_client()
+    
     if hf_client is None:
-        print("❌ ERROR: HuggingFace client not initialized!")
         return {
             'user_type': 'unknown',
             'emotion': 'unknown',
@@ -222,7 +234,6 @@ Respond ONLY in valid JSON format:
 }}"""
 
     try:
-        # FIXED: Better error logging
         print(f"   🔍 Calling Mistral API...")
         
         response = hf_client.text_generation(
@@ -249,10 +260,7 @@ Respond ONLY in valid JSON format:
         return result
         
     except Exception as e:
-        # FIXED: Show the actual error
         print(f"❌ LLM2 ERROR: {type(e).__name__}: {str(e)}")
-        import traceback
-        traceback.print_exc()
         
         return {
             'user_type': 'unknown',
@@ -267,9 +275,10 @@ Respond ONLY in valid JSON format:
 def manager_synthesize(llm1_result: Dict, llm2_result: Dict, review: Dict) -> Dict[str, Any]:
     """Manager: Synthesize LLM1 and LLM2 results"""
     
-    # FIXED: Check if client exists
+    # FIXED: Get client lazily
+    hf_client = get_hf_client()
+    
     if hf_client is None:
-        print("❌ ERROR: HuggingFace client not initialized!")
         return {
             'final_type': llm1_result.get('type', 'unknown'),
             'final_department': llm1_result.get('department', 'unknown'),
@@ -308,7 +317,6 @@ Respond ONLY in valid JSON format:
 }}"""
 
     try:
-        # FIXED: Better error logging
         print(f"   🔍 Calling Llama Manager API...")
         
         response = hf_client.text_generation(
@@ -334,7 +342,6 @@ Respond ONLY in valid JSON format:
         return result
         
     except Exception as e:
-        # FIXED: Show the actual error
         print(f"❌ MANAGER ERROR: {type(e).__name__}: {str(e)}")
         
         return {
@@ -432,7 +439,7 @@ def analyze_best_sentiment(text: str) -> Dict[str, Any]:
 
 
 def analyze_alt_sentiment(text: str) -> Dict[str, Any]:
-    """Alternate Model: BERTweet"""
+    """Alternate Model: BERTweet - FIXED version"""
     load_sentiment_models()
     
     try:
@@ -440,7 +447,16 @@ def analyze_alt_sentiment(text: str) -> Dict[str, Any]:
         
         with torch.no_grad():
             outputs = _alt_model(**inputs)
-            probs = torch.nn.functional.softmax(outputs.logits, dim=-1)
+            logits = outputs.logits
+            
+            # FIXED: Check if logits are on meta device
+            if logits.device.type == 'meta':
+                print("⚠️ Warning: Model on meta device, moving to CPU")
+                _alt_model.to('cpu')
+                outputs = _alt_model(**inputs)
+                logits = outputs.logits
+            
+            probs = torch.nn.functional.softmax(logits, dim=-1)
             prediction = torch.argmax(probs, dim=-1).item()
             confidence = probs[0][prediction].item()
         
@@ -553,11 +569,10 @@ def stage3_finalization_node(state: ReviewState) -> Dict[str, Any]:
     
     start_time = time.time()
     
-    # FIXED: Check if client exists
+    # FIXED: Get client lazily
+    hf_client = get_hf_client()
+    
     if hf_client is None:
-        print("❌ ERROR: HuggingFace client not initialized!")
-        print("   Skipping Stage 3 (requires API key)")
-        
         result = {
             'final_sentiment': state.get('sentiment', 'NEUTRAL'),
             'confidence': state.get('sentiment_confidence', 0.0),
@@ -630,7 +645,6 @@ Respond ONLY in valid JSON format:
 }}"""
 
     try:
-        # FIXED: Better error logging
         print(f"   🔍 Calling Llama 70B API...")
         
         response = hf_client.text_generation(
@@ -653,10 +667,7 @@ Respond ONLY in valid JSON format:
         result['model'] = 'meta-llama/Llama-3.1-70B-Instruct'
         
     except Exception as e:
-        # FIXED: Show the actual error
         print(f"❌ STAGE 3 ERROR: {type(e).__name__}: {str(e)}")
-        import traceback
-        traceback.print_exc()
         
         result = {
             'final_sentiment': state.get('sentiment', 'NEUTRAL'),