api.py

# coding=utf-8
# Copyright (c) 2020, NVIDIA CORPORATION.  All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

"""Inference API."""


import torch

from megatron import mpu
from .communication import broadcast_float_list
from .generation import generate_tokens_probs_and_return_on_first_stage
from .tokenization import (
    tokenize_prompts,
    detokenize_generations)



def generate_and_post_process(model,
                              prompts=None,
                              tokens_to_generate=0,
                              return_output_log_probs=False,
                              return_all_log_probs=False,
                              greedy_sampling=False,
                              top_k_sampling=0,
                              top_p_sampling=0.0,
                              temperature=1.0,
                              add_BOS=False,
                              use_eod_token_for_early_termination=True,
                              just_score=False):
    """Run inference and post-process outputs, i.e., detokenize,
    move to cpu and convert to list."""

    # Main inference.
    tokens, lengths, output_log_probs, all_log_probs = generate(
        model,
        prompts=prompts,
        tokens_to_generate=tokens_to_generate,
        return_output_log_probs=return_output_log_probs,
        return_all_log_probs=return_all_log_probs,
        greedy_sampling=greedy_sampling,
        top_k_sampling=top_k_sampling,
        top_p_sampling=top_p_sampling,
        temperature=temperature,
        add_BOS=add_BOS,
        use_eod_token_for_early_termination=use_eod_token_for_early_termination,
        just_score=just_score)

    # Only post-process on first stage.
    if mpu.is_pipeline_first_stage():

        tokens, prompts_plus_generations, prompts_plus_generations_segments = \
            detokenize_generations(tokens, lengths, True)

        if return_output_log_probs:
            output_log_probs = output_log_probs.cpu().numpy().tolist()
        if return_all_log_probs:
            all_log_probs = all_log_probs.cpu().numpy().tolist()

        return prompts_plus_generations, prompts_plus_generations_segments, \
            output_log_probs, all_log_probs, tokens

    return None



def generate(model,
             prompts=None,
             tokens_to_generate=0,
             return_output_log_probs=False,
             return_all_log_probs=False,
             greedy_sampling=False,
             top_k_sampling=0,
             top_p_sampling=0.0,
             temperature=1.0,
             add_BOS=False,
             use_eod_token_for_early_termination=True,
             just_score=False):
    """Given prompts and input parameters, run inference and return:
       tokens: prompts plus the generated tokens.
       lengths: length of the prompt + generations. Note that we can
           discard tokens in the tokens tensor that are after the
           corresponding length.
       output_log_probs: log probs of the tokens.
       all_log_probs: full log probs for all of tokens.
    """

    # Make sure input params are avaialble to all ranks.
    values = [tokens_to_generate,
              return_output_log_probs, return_all_log_probs,
              greedy_sampling, top_k_sampling, top_p_sampling,
              temperature, add_BOS, use_eod_token_for_early_termination, just_score]
    values_float_tensor = broadcast_float_list(10, float_list=values)
    tokens_to_generate = int(values_float_tensor[0].item())
    return_output_log_probs = bool(values_float_tensor[1].item())
    return_all_log_probs = bool(values_float_tensor[2].item())
    greedy_sampling = bool(values_float_tensor[3].item())
    top_k_sampling = int(values_float_tensor[4].item())
    top_p_sampling = values_float_tensor[5].item()
    temperature = values_float_tensor[6].item()
    add_BOS = bool(values_float_tensor[7].item())
    use_eod_token_for_early_termination = bool(values_float_tensor[8].item())
    just_score = bool(values_float_tensor[9].item())

    # Tokenize prompts and get the batch.
    # Note that these tensors are broadcaseted to all ranks.
    if torch.distributed.get_rank() == 0:
        assert prompts is not None
        #assert tokens_to_generate > 0
    context_tokens_tensor, context_length_tensor = tokenize_prompts(
        prompts=prompts, tokens_to_generate=tokens_to_generate, add_BOS=add_BOS)

    # Main inference function.
    # Note that the outputs are available on the first stage.
    return generate_tokens_probs_and_return_on_first_stage(
        model, context_tokens_tensor, context_length_tensor,
        return_output_log_probs=return_output_log_probs,
        return_all_log_probs=return_all_log_probs,
        greedy=greedy_sampling, top_k=top_k_sampling, top_p=top_p_sampling,
        temperature=temperature,
        use_eod_token_for_early_termination=use_eod_token_for_early_termination,
        just_score=just_score)