File size: 8,648 Bytes
23bd7af |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 |
# 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,
score_and_return_on_first_stage,
beam_search_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,
top_k_sampling=0,
top_p_sampling=0.0,
top_p_decay=0.0,
top_p_bound=0.0,
temperature=1.0,
add_BOS=False,
use_eod_token_for_early_termination=True,
stop_on_double_eol=False,
stop_on_eol=False,
random_seed=-1):
"""Run inference and post-process outputs, i.e., detokenize,
move to cpu and convert to list."""
# Main inference.
tokens, lengths, output_log_probs = generate(
model,
prompts=prompts,
tokens_to_generate=tokens_to_generate,
return_output_log_probs=return_output_log_probs,
top_k_sampling=top_k_sampling,
top_p_sampling=top_p_sampling,
top_p_decay=top_p_decay,
top_p_bound=top_p_bound,
temperature=temperature,
add_BOS=add_BOS,
use_eod_token_for_early_termination=use_eod_token_for_early_termination,
stop_on_double_eol=stop_on_double_eol,
stop_on_eol=stop_on_eol,
random_seed=random_seed)
# 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()
for i, (prob, seg) in enumerate(zip(output_log_probs, prompts_plus_generations_segments)):
output_log_probs[i] = prob[:len(seg)-1]
return prompts_plus_generations, prompts_plus_generations_segments, \
output_log_probs, tokens
return None
def generate(model,
prompts=None,
tokens_to_generate=0,
return_output_log_probs=False,
top_k_sampling=0,
top_p_sampling=0.0,
top_p_decay=0.0,
top_p_bound=0.0,
temperature=1.0,
add_BOS=False,
use_eod_token_for_early_termination=True,
stop_on_double_eol=False,
stop_on_eol=False,
random_seed=-1):
"""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.
"""
# Make sure input params are avaialble to all ranks.
values = [tokens_to_generate,
return_output_log_probs,
top_k_sampling, top_p_sampling, top_p_decay, top_p_bound,
temperature, add_BOS, use_eod_token_for_early_termination,
stop_on_double_eol,
stop_on_eol,
random_seed]
values_float_tensor = broadcast_float_list(12, float_list=values)
tokens_to_generate = int(values_float_tensor[0].item())
return_output_log_probs = bool(values_float_tensor[1].item())
top_k_sampling = int(values_float_tensor[2].item())
top_p_sampling = values_float_tensor[3].item()
top_p_decay = values_float_tensor[4].item()
top_p_bound = 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())
stop_on_double_eol = bool(values_float_tensor[9].item())
stop_on_eol = bool(values_float_tensor[10].item())
random_seed = int(values_float_tensor[11].item())
if random_seed != -1:
torch.random.manual_seed(random_seed)
# 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
context_tokens_tensor, context_length_tensor = tokenize_prompts(
prompts=prompts, tokens_to_generate=tokens_to_generate, add_BOS=add_BOS)
if tokens_to_generate == 0:
return score_and_return_on_first_stage(
model, context_tokens_tensor, context_length_tensor)
# 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,
top_k=top_k_sampling,
top_p=top_p_sampling,
top_p_decay=top_p_decay,
top_p_bound=top_p_bound,
temperature=temperature,
use_eod_token_for_early_termination=use_eod_token_for_early_termination,
stop_on_double_eol=stop_on_double_eol,
stop_on_eol=stop_on_eol)
def beam_search_and_post_process(model,
prompts=None,
tokens_to_generate=0,
beam_size=0,
add_BOS=False,
stop_token=50256,
num_return_gen=1,
length_penalty=1):
"""Run beam search and post-process outputs, i.e., detokenize,
move to cpu and convert to list."""
# Main inference.
tokens, scores = beam_search(model,
prompts=prompts,
tokens_to_generate=tokens_to_generate,
beam_size=beam_size,
add_BOS=add_BOS,
stop_token=stop_token,
num_return_gen=num_return_gen,
length_penalty=length_penalty)
# Only post-process on first stage.
if mpu.is_pipeline_first_stage():
lengths = tokens.size(1)*torch.ones(beam_size, dtype=torch.int64, device=torch.cuda.current_device())
tokens, prompts_plus_generations, prompts_plus_generations_segments = detokenize_generations(tokens, lengths, True)
scores = scores.cpu().numpy().tolist()
return prompts_plus_generations, prompts_plus_generations_segments, scores
return None
def beam_search(model, prompts=None, tokens_to_generate=0, beam_size=0, add_BOS=False, stop_token=50256, num_return_gen=1, length_penalty=1):
# Make sure input params are avaialble to all ranks.
values = [tokens_to_generate,
beam_size,
add_BOS,
stop_token,
num_return_gen,
length_penalty]
values_float_tensor = broadcast_float_list(6, float_list=values)
tokens_to_generate = int(values_float_tensor[0].item())
beam_size = int(values_float_tensor[1].item())
add_BOS = bool(values_float_tensor[2].item())
stop_token = int(values_float_tensor[3].item())
num_return_gen = int(values_float_tensor[4].item())
length_penalty = values_float_tensor[5].item()
context_tokens_tensor, context_length_tensor = tokenize_prompts(
prompts=prompts, tokens_to_generate=tokens_to_generate, add_BOS=add_BOS)
return beam_search_and_return_on_first_stage(model, context_tokens_tensor, context_length_tensor,
beam_size, stop_token=stop_token, num_return_gen=num_return_gen, length_penalty=length_penalty)
|