Upload app.py with huggingface_hub

app.py

@@ -1,57 +1,63 @@
 import gradio as gr
-from huggingface_hub import InferenceClient
+from transformers import AutoModelForCausalLM, AutoTokenizer
+import torch
 import os
 import re
+import spaces
 
 # The model to call
 model_id = "Atlas-labs/mini-fable-5-qwen-merged"
 token = os.getenv("HF_TOKEN")
 
-# Initialize the high-speed client
-client = InferenceClient(model=model_id, token=token)
+print(f"Loading tokenizer and model...")
+tokenizer = AutoTokenizer.from_pretrained(model_id, token=token)
+model = AutoModelForCausalLM.from_pretrained(
+    model_id, 
+    torch_dtype=torch.float16, 
+    device_map="auto",
+    token=token
+)
 
-def parse_fable_response(raw_text):
-    # Extract thought
+def parse_fable_response(raw_text, user_input):
     thought = ""
     thought_match = re.search(r'<thought>(.*?)</thought>', raw_text, re.DOTALL | re.IGNORECASE)
     if thought_match:
         thought = thought_match.group(1).strip()
     
-    # Extract answer
     answer = raw_text
     if "</thought>" in raw_text.lower():
         answer = re.split(r'</thought>', raw_text, flags=re.IGNORECASE)[-1]
     
-    # Clean up tags
-    answer = re.sub(r'</?thought>', '', answer, flags=re.IGNORECASE).strip()
-    answer = re.sub(r'</?response>', '', answer, flags=re.IGNORECASE).strip()
+    answer = re.sub(r'</?thought>', '', answer, flags=re.IGNORECASE)
+    answer = re.sub(r'</?response>', '', answer, flags=re.IGNORECASE)
+    answer = answer.replace(user_input, "").strip()
 
     if thought:
         return f"💡 **Reasoning:**\n> *{thought}*\n\n{answer}"
     return answer
 
+@spaces.GPU
 def chat(message, history):
-    # We send the prompt and get a fast response from the HF GPU cluster
     prompt = f"<thought>\nAnalyzing request: {message}\n"
+    inputs = tokenizer(prompt + message, return_tensors="pt").to(model.device)
     
-    try:
-        # Using text_generation for maximum speed and control over the CoT tags
-        response = client.text_generation(
-            prompt + message,
-            max_new_tokens=512,
-            temperature=0.7,
+    with torch.no_grad():
+        outputs = model.generate(
+            **inputs, 
+            max_new_tokens=512, 
+            temperature=0.7, 
             do_sample=True,
-            return_full_text=False
+            pad_token_id=tokenizer.eos_token_id
         )
-        return parse_fable_response(response)
-    except Exception as e:
-        return f"⚠️ **Inference Error:** The model is currently loading on the server. Please try again in 30 seconds. ({str(e)})"
+    
+    raw_response = tokenizer.decode(outputs[0], skip_special_tokens=True)
+    return parse_fable_response(raw_response, message)
 
 demo = gr.ChatInterface(
     fn=chat, 
-    title="Mini Fable 5 (Turbo)", 
-    description="High-speed reasoning engine by Atlas Labs. Powered by Hugging Face Inference API.",
-    examples=["Explain the theory of relativity.", "How do I build a search engine?", "Why is the sky blue?"]
+    title="Mini Fable 5 (ZeroGPU)", 
+    description="High-speed reasoning engine by Atlas Labs. Powered by Hugging Face ZeroGPU.",
+    examples=["Explain the theory of relativity.", "Write a Python script for a binary search.", "Why is the sky blue?"]
 )
 
 if __name__ == "__main__":