Update app.py

app.py

@@ -1,64 +1,137 @@
 import gradio as gr
-from huggingface_hub import InferenceClient
+import torch
+from transformers import AutoModelForCausalLM, AutoTokenizer
+import gc # Garbage collector
 
-"""
-For more information on `huggingface_hub` Inference API support, please check the docs: https://huggingface.co/docs/huggingface_hub/v0.22.2/en/guides/inference
-"""
-client = InferenceClient("HuggingFaceH4/zephyr-7b-beta")
+# --- Configuration ---
+MODEL_ID = "naver-hyperclovax/HyperCLOVAX-SEED-Text-Instruct-0.5B"
+MAX_NEW_TOKENS = 512 # Limit output length for faster response on CPU
+SYSTEM_PROMPT = "- AI 언어모델의 이름은 \"CLOVA X\" 이며 네이버에서 만들었다.\n- 사용자의 질문에 대해 친절하고 자세하게 답변해야 한다." # Simplified system prompt
 
+# --- Model Loading ---
+# Load model and tokenizer explicitly on CPU
+# Using device_map='cpu' and torch_dtype=torch.float32 ensures it targets the CPU
+print(f"Loading model: {MODEL_ID}...")
+try:
+    model = AutoModelForCausalLM.from_pretrained(
+        MODEL_ID,
+        torch_dtype=torch.float32, # Use float32 for CPU compatibility
+        device_map="cpu" # Explicitly load on CPU
+    )
+    tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
+    print("Model and tokenizer loaded successfully on CPU.")
+    # Ensure model is in evaluation mode
+    model.eval()
+except Exception as e:
+    print(f"Error loading model: {e}")
+    # If loading fails, exit or handle gracefully
+    raise gr.Error(f"Failed to load the model {MODEL_ID}. Check logs. Error: {e}")
 
-def respond(
-    message,
-    history: list[tuple[str, str]],
-    system_message,
-    max_tokens,
-    temperature,
-    top_p,
-):
-    messages = [{"role": "system", "content": system_message}]
+# Define stop tokens based on the example, get their IDs
+stop_token_strings = ["<|endofturn|>", "<|stop|>"]
+stop_token_ids = [tokenizer.convert_tokens_to_ids(token) for token in stop_token_strings]
+# Also include the standard EOS token if it's different
+if tokenizer.eos_token_id not in stop_token_ids:
+    stop_token_ids.append(tokenizer.eos_token_id)
 
-    for val in history:
-        if val[0]:
-            messages.append({"role": "user", "content": val[0]})
-        if val[1]:
-            messages.append({"role": "assistant", "content": val[1]})
+# --- Inference Function ---
+def predict(message, history):
+    """
+    Generates a response using the HyperCLOVAX model based on user message and chat history.
+    """
+    # Reconstruct chat history in the required format
+    chat_history_formatted = [
+        {"role": "tool_list", "content": ""}, # As per model card example structure
+        {"role": "system", "content": SYSTEM_PROMPT}
+    ]
+    for user_msg, ai_msg in history:
+        chat_history_formatted.append({"role": "user", "content": user_msg})
+        # Add the endofturn token manually if needed, or rely on template
+        chat_history_formatted.append({"role": "assistant", "content": ai_msg }) # Ensure assistant response is captured correctly
 
-    messages.append({"role": "user", "content": message})
+    # Add the current user message
+    chat_history_formatted.append({"role": "user", "content": message})
 
-    response = ""
+    # Apply the chat template
+    try:
+        inputs = tokenizer.apply_chat_template(
+            chat_history_formatted,
+            add_generation_prompt=True, # Crucial for instruct models
+            return_dict=True,
+            return_tensors="pt"
+        ).to(model.device) # Ensure inputs are on the CPU device
+    except Exception as e:
+        print(f"Error applying chat template: {e}")
+        return f"Error formatting input: {e}"
 
-    for message in client.chat_completion(
-        messages,
-        max_tokens=max_tokens,
-        stream=True,
-        temperature=temperature,
-        top_p=top_p,
-    ):
-        token = message.choices[0].delta.content
+    print(f"Input tokens: {inputs['input_ids'].shape[1]}") # Log input length
 
-        response += token
-        yield response
+    # Generate response - Use torch.no_grad() to save memory
+    try:
+        with torch.no_grad():
+            output_ids = model.generate(
+                **inputs,
+                max_new_tokens=MAX_NEW_TOKENS,
+                eos_token_id=stop_token_ids, # Use the identified stop token IDs
+                pad_token_id=tokenizer.eos_token_id, # Often set pad = eos
+                do_sample=True, # Sample for more diverse outputs
+                temperature=0.7,
+                top_p=0.9,
+                # Note: The original example used 'stop_strings' and passed the tokenizer,
+                # which isn't standard in `generate`. Using eos_token_id is preferred.
+                # If specific stopping behavior is needed beyond EOS, StoppingCriteria can be used.
+            )
+    except Exception as e:
+        print(f"Error during model generation: {e}")
+        # Clean up GPU memory if an error occurs (though we are on CPU, good practice)
+        gc.collect()
+        # Consider torch.cuda.empty_cache() if GPU was accidentally involved
+        return f"Error during generation: {e}"
 
+    # Decode only the newly generated tokens
+    input_length = inputs['input_ids'].shape[1]
+    new_tokens = output_ids[0, input_length:]
+    response = tokenizer.decode(new_tokens, skip_special_tokens=True)
+
+    print(f"Output tokens: {len(new_tokens)}") # Log output length
+    print(f"Raw response: {response}")
+
+    # Clean up memory (especially important in constrained environments)
+    del inputs
+    del output_ids
+    gc.collect()
+    # Consider torch.cuda.empty_cache() if GPU was involved
+
+    return response
+
+# --- Gradio Interface ---
+# Use ChatInterface for a conversational experience
+chatbot = gr.Chatbot(label="HyperCLOVA X SEED (0.5B)", height=600)
 
-"""
-For information on how to customize the ChatInterface, peruse the gradio docs: https://www.gradio.app/docs/chatinterface
-"""
 demo = gr.ChatInterface(
-    respond,
-    additional_inputs=[
-        gr.Textbox(value="You are a friendly Chatbot.", label="System message"),
-        gr.Slider(minimum=1, maximum=2048, value=512, step=1, label="Max new tokens"),
-        gr.Slider(minimum=0.1, maximum=4.0, value=0.7, step=0.1, label="Temperature"),
-        gr.Slider(
-            minimum=0.1,
-            maximum=1.0,
-            value=0.95,
-            step=0.05,
-            label="Top-p (nucleus sampling)",
-        ),
+    fn=predict,
+    chatbot=chatbot,
+    title="🇰🇷 HyperCLOVA X SEED (0.5B) - CPU Demo",
+    description=(
+        f"Chat with the `naver-hyperclovax/HyperCLOVAX-SEED-Text-Instruct-0.5B` model. "
+        f"This model excels in Korean language understanding. \n"
+        f"**Note:** Running on a free CPU Hugging Face Space. Responses will be **slow**. "
+        f"Max new tokens set to {MAX_NEW_TOKENS}."
+    ),
+    examples=[
+        ["슈뢰딩거 방정식과 양자역학의 관계를 최대한 자세히 알려줘."],
+        ["네이버 하이퍼클로바X에 대해 설명해줘."],
+        ["오늘 날씨 어때?"], # Example showing it might not know real-time data
+        ["간단한 파이썬 코드 예제를 보여줘."],
     ],
+    cache_examples=False, # Caching might consume too much memory/disk on free tier
+    theme="soft",
+    retry_btn=None,
+    undo_btn="Delete Previous Turn",
+    clear_btn="Clear Conversation",
 )
 
-
+# --- Launch the App ---
 if __name__ == "__main__":
-    demo.launch()
+    # queue() is important for handling multiple users, especially with slow inference
+    demo.queue().launch()