Update app.py

app.py

@@ -0,0 +1,45 @@
+import torch, gradio as gr
+from transformers import AutoModelForCausalLM, AutoTokenizer
+from peft import PeftModel
+import os
+
+model_id = "Qwen/Qwen2.5-0.5B-Instruct"
+# Path to adapter - local if exists, else load from Hub
+local_adapter = "outputs/qwen-fine-tuned"
+# Environment variables from HF Space secret settings
+hf_username = os.getenv("HF_USERNAME")
+hf_model_name = os.getenv("HF_MODEL_NAME")
+hub_adapter = f"{hf_username}/{hf_model_name}" if hf_username and hf_model_name else None
+
+# Prioritize local folder but fallback to hub repo
+adapter_path = local_adapter if os.path.exists(local_adapter) else hub_adapter
+
+# Handle device detection for varied environments
+device = "cuda" if torch.cuda.is_available() else "cpu"
+if not torch.cuda.is_available() and torch.backends.mps.is_available():
+    device = "mps"
+
+print(f"Loading model on {device}...")
+print(f"Using adapter path: {adapter_path}")
+
+tokenizer = AutoTokenizer.from_pretrained(model_id)
+# Load in float16 for memory efficiency
+model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.float16, device_map="auto")
+
+if adapter_path:
+    try:
+        model = PeftModel.from_pretrained(model, adapter_path)
+        print("Adapter loaded successfully!")
+    except Exception as e:
+        print(f"Warning: Could not load adapter from {adapter_path}: {e}")
+else:
+    print("Warning: No adapter found. Using base model.")
+
+def chat(message, history):
+    msgs = [{"role": "user", "content": message}]
+    text = tokenizer.apply_chat_template(msgs, tokenize=False, add_generation_prompt=True)
+    model_inputs = tokenizer([text], return_tensors="pt").to(device)
+    ids = model.generate(**model_inputs, max_new_tokens=512, pad_token_id=tokenizer.eos_token_id)
+    return tokenizer.decode(ids[0][len(model_inputs.input_ids[0]):], skip_special_tokens=True)
+
+gr.ChatInterface(chat).launch()