unsloth llm3-1B

app.py

@@ -29,7 +29,7 @@ class ModelManager:
         if HAS_UNSLOTH and self.device != "cpu":
             # GPU via Unsloth + LoRA
             backbone, tokenizer = FastLanguageModel.from_pretrained(
-                "Qwen/Qwen2.5-Coder-1.5B-Instruct",
+                "unsloth/Llama-3.2-1B-Instruct-bnb-4bit",
                 load_in_4bit=True,
                 dtype=torch.float16,
                 device_map="auto",
@@ -50,7 +50,7 @@ class ModelManager:
 
         # --- Fallback: CPU-only via HF Transformers + PEFT ---
         print("Falling back to CPU-only Transformers + PEFT")
-        base_name = "Qwen/Qwen2.5-Coder-1.5B-Instruct"  # non-4bit to run on CPU
+        base_name = "unsloth/Llama-3.2-1B-Instruct"  # non-4bit to run on CPU
         tokenizer = AutoTokenizer.from_pretrained(base_name, use_fast=True)
         base = AutoModelForCausalLM.from_pretrained(
             base_name,
@@ -109,4 +109,4 @@ demo = gr.Interface(
 )
 
 if __name__ == "__main__":
-    demo.launch()
+    demo.launch(share=True)