Update RockPaperScissor/services/LLM_service.py

RockPaperScissor/services/LLM_service.py

@@ -1,21 +1,12 @@
+import spaces  # Import unconditionally at the top
 from transformers import AutoModelForCausalLM, AutoTokenizer
 import torch
 from typing import List, Dict, Optional
 import re
 import os
 
-# Import spaces GPU decorator - this should work on HF Spaces
-try:
-    import spaces
-    from spaces import GPU
-    print("[LLMService] spaces.GPU imported successfully")
-    SPACES_AVAILABLE = True
-except ImportError:
-    print("[LLMService] spaces.GPU not available, using fallback")
-    SPACES_AVAILABLE = False
-    def GPU(f): 
-        return f
-    spaces = None
+print("[LLMService] Starting LLM service initialization...")
+print("[LLMService] Spaces imported successfully")
 
 # Global model variables for GPU functions
 _model = None
@@ -91,6 +82,9 @@ def generate_text_with_gpu(prompt_text: str, max_tokens: int = 150):
     print("[LLMService] GPU generation completed.")
     return response
 
+# Print confirmation that GPU function is registered
+print(f"[LLMService] GPU function registered: {generate_text_with_gpu.__name__}")
+
 class LLMService:
     def __init__(self):
         # Use a reliable model that works well with Zero GPU
@@ -107,7 +101,7 @@ class LLMService:
         # Model will be loaded on-demand in GPU context
         self.device = "cuda" if torch.cuda.is_available() else "cpu"
         print(f"[LLMService] Target device: {self.device}")
-        print(f"[LLMService] Spaces GPU available: {SPACES_AVAILABLE}")
+        print("[LLMService] Service initialization completed")
 
     def generate_with_model(self, prompt_text: str, max_tokens: int = 150):
         """Use the top-level GPU function"""
@@ -258,4 +252,5 @@ Your advice:
             print(f"[LLMService] Error during cleanup: {e}")
 
 # Create a singleton instance
-llm_service_instance = LLMService()
+llm_service_instance = LLMService()
+print("[LLMService] Service instance created successfully")