Update app.py

app.py

@@ -4,22 +4,19 @@ import torch
 from transformers import AutoModelForCausalLM, AutoTokenizer, TextIteratorStreamer
 from threading import Thread
 
-@spaces.GPU
+# Remove GPU decorator since we are CPU-only
 def predict(message, history):
-    torch.set_default_device("cuda")
-    
-    # Load model and tokenizer
+    # Load model and tokenizer on CPU
     model_id = "kurakurai/Luth-LFM2-1.2B"
     tokenizer = AutoTokenizer.from_pretrained(model_id)
     model = AutoModelForCausalLM.from_pretrained(
         model_id,
-        device_map="auto",
-        torch_dtype=torch.bfloat16,
+        device_map="cpu",       # CPU only
+        torch_dtype=torch.float16,
         trust_remote_code=True,
-        load_in_4bit=True,  # Keeping 4-bit quantization for efficiency
-        # attn_implementation="flash_attention_2"  # Uncomment on compatible GPU
+        load_in_4bit=False      # 4-bit quantization not supported on CPU
     )
-    
+
     # Format conversation history for chat template
     messages = [{"role": "user" if i % 2 == 0 else "assistant", "content": msg} 
                 for conv in history for i, msg in enumerate(conv) if msg]
@@ -31,7 +28,7 @@ def predict(message, history):
         add_generation_prompt=True,
         return_tensors="pt",
         tokenize=True
-    ).to('cuda')
+    ).to('cpu')  # CPU device
     
     # Setup streamer for real-time output
     streamer = TextIteratorStreamer(tokenizer, timeout=10., skip_prompt=True, skip_special_tokens=True)
@@ -67,9 +64,9 @@ gr.ChatInterface(
     Chat with [Luth-LFM2-1.2B](https://huggingface.co/kurakurai/Luth-LFM2-1.2B), a French-tuned version of LFM2-1.2B.
     """,
     examples=[
-    "Peux-tu résoudre l'équation 3x - 7 = 11 pour x ?",
-    "Explique la photosynthèse en termes simples.",
-    "Écris un petit poème sur l'intelligence artificielle."
+        "Peux-tu résoudre l'équation 3x - 7 = 11 pour x ?",
+        "Explique la photosynthèse en termes simples.",
+        "Écris un petit poème sur l'intelligence artificielle."
     ],
     theme=gr.themes.Soft(primary_hue="blue"),
-).launch()
+).launch()