Merge branch 'main' into inference

import datetime

import torch

import json

import threading

from flask import Flask, request, jsonify, current_app

from flask_restful import Resource, Api

from megatron import get_args

from megatron.text_generation_utils import generate

GENERATE_NUM = 0

lock = threading.Lock()

class MegatronGenerate(Resource):

    def __init__(self, model):

        print("request IP: " + str(request.remote_addr))

        print(json.dumps(request.get_json()),flush=True)

        print("current time: ", datetime.datetime.now())

        sentences = request.get_json()["sentences"]

        if len(sentences) > 128:

            return "Maximum number of sentences is 128", 400

               0.0 < temperature <= 100.0:

                return "temperature must be a positive float less than or equal to 100.0"

        add_BOS = False

        if "add_BOS" in request.get_json():

            add_BOS = request.get_json()["add_BOS"]

            if not isinstance(add_BOS, bool):

                return "add_BOS must be a boolean value"

        with lock:  # Need to get lock to keep multiple threads from hitting code

            MegatronGenerate.send_do_generate()  # Tell other ranks we're doing generate

        resp_sentences, resp_sentences_seg, output_logits, full_logits, tokens = generate(self.model, sentences, tokens_to_generate, all_probs, temperature)

            resp_sentences, resp_sentences_seg, output_logits, full_logits, tokens = generate(self.model, sentences, tokens_to_generate, all_probs, temperature, add_BOS) 

        if all_probs:

            return jsonify({"sentences": resp_sentences,

                "segments": resp_sentences_seg,

class MegatronServer(object):

    def __init__(self, model):

        self.app = Flask(__name__, static_folder='static', static_url_path='')

        self.app.config['SEND_FILE_MAX_AGE_DEFAULT'] = 0

        self.app = Flask(__name__, static_url_path='')

        api = Api(self.app)

        api.add_resource(MegatronGenerate, '/generate', resource_class_args=[model])

        context_lengths.append(context_length)

    return batch, context_lengths

def tokenize_batch(sentences, max_len):

def tokenize_batch(sentences, max_len, add_BOS):

    args = get_args()

    tokenizer = get_tokenizer()

    if add_BOS:

        context_tokens = [[tokenizer.eod] + tokenizer.tokenize(s) for s in sentences]

    else:

        context_tokens = [tokenizer.tokenize(s) for s in sentences]

    context_tokens, context_lengths = pad_batch(context_tokens,

                                                tokenizer.eod, max_len)

    if tokens is not None:

        return tokens[:, :context_length], output_logits, full_logits 

def generate(model, sentences=None, tokens_to_generate=0, all_probs=False, temperature=1.0):

def generate(model, sentences=None, tokens_to_generate=0, all_probs=False, temperature=1.0, add_BOS=False):

    model.eval()

    if torch.distributed.get_rank() == 0:

        context_tokens_tensor, context_length_tensor = tokenize_batch(sentences, tokens_to_generate)

        context_tokens_tensor, context_length_tensor = tokenize_batch(sentences, tokens_to_generate, add_BOS)

        send_generate_info(context_tokens_tensor, context_length_tensor, tokens_to_generate, all_probs)

    else:

        context_length_tensor, context_tokens_tensor, tokens_to_generate, all_probs = receive_generate_info()

            args.task))

    # Set up model and load checkpoint.

    model = get_model(get_model_provider(eval_metric))

    model = get_model(get_model_provider(eval_metric), wrap_with_ddp=False)

    if args.load is not None:

        _ = load_checkpoint(model, None, None)