Function generate_samples_eval

codegeex/megatron/code_generation_utils.py:225–249

(model, context, max_gen_length, eos_token_id)

# codebase. The lm-evaluation-harness code can now call this function
# similar to their current generate function call used for gpt style models.
def generate_samples_eval(model, context, max_gen_length, eos_token_id):
    # Generate samples for lm evaluation
    # NEED TO THINK ABOUT eos token

    args = get_args()
    tokenizer = get_tokenizer()

    raw_text_len = len(context)
    model.eval()

    context_tokens = tokenizer.tokenize(context)
    args.out_seq_length = max_gen_length + len(context_tokens)
    args.eos_id = eos_token_id

    with torch.no_grad():
        token_stream = get_token_stream(model, [context_tokens])
        for counter, decode_tokens in enumerate(token_stream):
            if counter == args.out_seq_length:
                break

    decode_tokens, _ = decode_tokens
    decode_tokens = decode_tokens[0].cpu().numpy().tolist()
    trim_decode_tokens = tokenizer.detokenize(decode_tokens)[raw_text_len:]

    return trim_decode_tokens


def generate_samples_interactive_code_contest(model, print_frequency=10):