Update app.py

app.py

@@ -23,11 +23,25 @@ model = AutoModelForCausalLM.from_pretrained(config.base_model_name_or_path, qua
 model = PeftModel.from_pretrained(model, model_id)
 
 def greet(text):
-    with torch.no_grad():  # Disable gradient calculation for inference
-        batch = tokenizer(f'### Input:\n{text}\n\n### Answer:\n', return_tensors='pt')  # Move tensors to device
-        with torch.cuda.amp.autocast():  # Enable mixed-precision if available
-            output_tokens = model.generate(**batch, max_new_tokens=50)
-    return tokenizer.decode(output_tokens[0], skip_special_tokens=True)
+    with torch.no_grad():
+        # Include EOS token for better context
+        input_text = f'<s>### Input:\n{text}\n\n### Response:\n'
+        batch = tokenizer(input_text, return_tensors='pt', add_special_tokens=True).to(device)
+
+        with torch.cuda.amp.autocast():
+            output_tokens = model.generate(
+                **batch,
+                max_new_tokens=25,  # Limit response length
+                do_sample=True,  # Sample from the distribution
+                pad_token_id=tokenizer.eos_token_id,  # Stop at EOS
+            )
+
+        # Decode only the generated tokens
+        response = tokenizer.decode(output_tokens[0][len(batch['input_ids'][0]):], skip_special_tokens=True)
+
+        # Additional stopping condition at next "\n### Response:"
+        response_parts = response.split("\n### Response:")
+        return response_parts[0]  # Return only the first part
 
 iface = gr.Interface(fn=greet, inputs="text", outputs="text", title="PEFT Model for Big Brain")
 iface.launch()  # Share directly to Gradio Space