Refactor app.py: lazy imports and fallback to avoid missing dependencies

app.py

@@ -1,13 +1,15 @@
 import gradio as gr
-from transformers import AutoModelForCausalLM, AutoTokenizer
-import torch
 
 MODEL_NAME = "likhonhfai/mysterious-coding-model"
 
-
 def load_model():
-    """Attempt to load the CodeAI model. Returns (model, tokenizer) or (None, None) on failure."""
+    """
+    Attempt to lazily import transformers and torch and load the CodeAI model.
+    Returns (model, tokenizer) if loaded successfully, otherwise (None, None).
+    """
     try:
+        from transformers import AutoModelForCausalLM, AutoTokenizer
+        import torch
         model = AutoModelForCausalLM.from_pretrained(
             MODEL_NAME,
             torch_dtype=torch.float16,
@@ -17,24 +19,24 @@ def load_model():
         tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
         return model, tokenizer
     except Exception:
-        # Fallback when the large model cannot be loaded due to resource constraints.
         return None, None
 
 
-# Load the model at module import time
+# Load the model once at startup
 model, tokenizer = load_model()
 
 
 def respond(message, history):
-    """Respond to user messages using the loaded model or a fallback."""
-    # If the model is loaded successfully, generate a response from it.
+    """
+    Generate a response using the loaded model or provide a placeholder message.
+    """
+    # If the model is available, generate a response using it
     if model is not None and tokenizer is not None:
-        # Build conversation prompt from history and the current message
+        import torch  # Safe to import since it was available during model loading
         prompt = ""
         for user_msg, bot_msg in history:
             prompt += f"User: {user_msg}\nAssistant: {bot_msg}\n"
         prompt += f"User: {message}\nAssistant:"
-        # Encode and generate output
         inputs = tokenizer.encode(prompt, return_tensors="pt").to(model.device)
         with torch.no_grad():
             output_ids = model.generate(
@@ -45,41 +47,33 @@ def respond(message, history):
                 pad_token_id=tokenizer.eos_token_id,
             )
         output_text = tokenizer.decode(output_ids[0], skip_special_tokens=True)
-        # Extract the assistant's response after the last 'Assistant:' marker
         if "Assistant:" in output_text:
-            response = output_text.split("Assistant:")[-1].strip()
+            return output_text.split("Assistant:")[-1].strip()
         else:
-            response = output_text.strip()
-        return response
+            return output_text.strip()
 
     # Fallback responses when the model is unavailable
     lower = message.lower()
     if "hello" in lower:
         return (
-            "Hello! I'm a placeholder chatbot while the full CodeAI model loads. "
-            "Ask me about our capabilities like long-context processing, multimodal understanding, "
-            "and advanced code generation."
+            "Hello! I'm a placeholder chatbot while the full CodeAI model loads. Ask me about long-context processing, "
+            "multimodal understanding, or code generation."
         )
     if "code" in lower:
         return (
-            "Our model specializes in coding tasks such as code generation, completion, bug fixing, "
-            "refactoring, and documentation. For example, try asking: 'write a python function to add two numbers'."
+            "Our model excels at code generation, completion, bug fixing, refactoring and documentation. "
+            "Try asking: 'write a python function to add two numbers'."
         )
     if "image" in lower:
-        return (
-            "The CodeAI model supports image understanding tasks like visual question answering and image captioning."
-        )
+        return "The CodeAI model supports image understanding tasks like visual question answering and image captioning."
     if "audio" in lower or "speech" in lower:
-        return (
-            "Our model can process audio for speech recognition and audio understanding tasks."
-        )
+        return "Our model can process audio for speech recognition and audio understanding."
     if "thanks" in lower or "thank you" in lower:
-        return "You're welcome! Let me know if you have any more questions."
-    # Default fallback summary
+        return "You're welcome! Let me know if you have more questions."
     return (
         "This is a demo placeholder response. The CodeAI model uses safetensors storage, supports 8-bit and mxfp4 "
-        "mixed-precision variants, is compatible with the vLLM inference engine, and is trained using Hugging Face AutoTrain. "
-        "It can handle long contexts (up to 200,000 tokens) and perform text, image, audio, and multimodal reasoning tasks."
+        "mixed-precision variants, is compatible with the vLLM engine, and is trained using Hugging Face AutoTrain. "
+        "It handles long contexts (up to 200,000 tokens) and performs text, image, audio, and multimodal reasoning tasks."
     )