Function chat

SwissArmyTransformer/examples/llama/chat_sat.py:8–30

(model, tokenizer, 
        max_length: int = 256, num_beams=1, top_p=0.7, top_k=0, temperature=0.95)

from sat.generation.sampling_strategies import BaseStrategy, BeamSearchStrategy

def chat(model, tokenizer, 
        max_length: int = 256, num_beams=1, top_p=0.7, top_k=0, temperature=0.95):
    prompt = "The capital of China is"
    inputs = tokenizer([prompt], return_tensors="pt", add_special_tokens=False).to(model.parameters().__next__().device)['input_ids'][0]
    seq = torch.cat(
        [inputs, torch.tensor([-1]*(max_length-len(inputs)), device=inputs.device)], dim=0
    )
    strategy = BaseStrategy(temperature=temperature, top_p=top_p, top_k=0, end_tokens=[tokenizer.eos_token_id])
    strategy = BeamSearchStrategy(temperature=temperature, top_p=top_p, top_k=0, end_tokens=[tokenizer.eos_token_id], num_beams=num_beams, consider_end=True)
    
    output = filling_sequence(
        model, seq,
        batch_size=1,
        strategy=strategy
    )[0] # drop memory
    
    # ---------------
    # port from inference_glm.py, more general than chat mode
    # clip -1s and fill back generated things into seq
    output_list = list(output)

    response = tokenizer.decode(output_list[0])
    print(response)

if __name__ == "__main__":
    import argparse