Spaces:
Running
on
Zero
Running
on
Zero
import gradio as gr | |
import spaces | |
from transformers import AutoModelForCausalLM, AutoTokenizer | |
import torch | |
model_name = "Qwen/Qwen2.5-7B-Instruct" | |
model = AutoModelForCausalLM.from_pretrained( | |
model_name, | |
torch_dtype=torch.bfloat16, | |
device_map="auto" | |
) | |
tokenizer = AutoTokenizer.from_pretrained(model_name) | |
def generate(prompt, history): | |
messages = [ | |
{"role": "system", "content": "Je bent een vriendelijke, behulpzame assistent."}, | |
{"role": "user", "content": prompt} | |
] | |
text = tokenizer.apply_chat_template( | |
messages, | |
tokenize=False, | |
add_generation_prompt=True | |
) | |
model_inputs = tokenizer([text], return_tensors="pt").to(model.device) | |
generated_ids = model.generate( | |
**model_inputs, | |
max_new_tokens=512 | |
) | |
generated_ids = [ | |
output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids) | |
] | |
response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0] | |
return response | |
chat_interface = gr.ChatInterface( | |
fn=generate, | |
) | |
chat_interface.launch(share=True) | |