init

app.py

@@ -0,0 +1,50 @@
+import os
+import torch
+from transformers import AutoTokenizer, AutoModelForCausalLM
+import gradio as gr
+import spaces
+
+huggingface_token = os.getenv("HUGGINGFACE_TOKEN")
+if not huggingface_token:
+    raise ValueError("HUGGINGFACE_TOKEN environment variable is not set")
+
+model_id = "meta-llama/Llama-Guard-3-8B"
+device = "cuda"
+dtype = torch.bfloat16
+
+tokenizer = AutoTokenizer.from_pretrained(model_id, token=huggingface_token)
+
+
+@spaces.GPU
+def moderate(chat):
+    model = AutoModelForCausalLM.from_pretrained(
+        model_id, torch_dtype=dtype, device_map=device, token=huggingface_token
+    )
+    input_ids = tokenizer.apply_chat_template(chat, return_tensors="pt").to(device)
+    output = model.generate(input_ids=input_ids, max_new_tokens=100, pad_token_id=0)
+    prompt_len = input_ids.shape[-1]
+    return tokenizer.decode(output[0][prompt_len:], skip_special_tokens=True)
+
+
+def gradio_moderate(user_input, assistant_response):
+    chat = [
+        {"role": "user", "content": user_input},
+    ]
+    if assistant_response:
+        chat.append({"role": "assistant", "content": assistant_response})
+    return moderate(chat)
+
+
+iface = gr.Interface(
+    fn=gradio_moderate,
+    inputs=[
+        gr.Textbox(lines=3, label="User Input"),
+        gr.Textbox(lines=3, label="Assistant Response (optional)"),
+    ],
+    outputs=gr.Textbox(label="Moderation Result"),
+    title="Llama Guard Moderation",
+    description="Enter a user input and an optional assistant response to check for content moderation.",
+)
+
+if __name__ == "__main__":
+    iface.launch()

requirements.txt

@@ -0,0 +1,6 @@
+numpy
+torch
+spaces
+bitsandbytes
+accelerate
+transformers