dtk23.04初始化

start_fast.md

+# Fast Setup instructions
+
+This quick instructions document contains 3 steps:
+
+1. installing software
+2. preparing data
+3. running the script
+
+This is useful if you need to ask someone to reproduce problems with `Megatron-Deepspeed`
+
+## 1. Software
+
+Please follow this exact order.
+
+
+0. Create a new conda env if need be or activate an existing environment.
+
+1. Install `pytorch`. Choose the desired version install instructions [here](https://pytorch.org/get-started/locally/), but for conda it'd be:
+
+```
+conda install pytorch torchvision torchaudio cudatoolkit=11.3 -c pytorch
+```
+
+2. Install system-wide `cuda` if you don't have it already. [NVIDIA instructions](https://docs.nvidia.com/cuda/cuda-installation-guide-linux/index.html). Of course ideally use [the premade packages for your distro](https://docs.nvidia.com/cuda/cuda-installation-guide-linux/index.html#package-manager-installation).
+Use the same major version as pytorch's cuda build. To check use:
+
+```
+python -c 'import torch; print(f"pt={torch.__version__}, cuda={torch.version.cuda}")'
+```
+
+The minor versions don't actually have to match, but then you will need to hack `apex` installer to ignore minor version changes, see below.
+
+3. Install `apex`
+
+```
+git clone https://github.com/NVIDIA/apex
+cd apex
+pip install --global-option="--cpp_ext" --global-option="--cuda_ext" --no-cache -v --disable-pip-version-check .  2>&1 | tee build.log
+cd -
+```
+
+If the pytorch and system-wide cuda minor versions mismatch, it's not a problem, you just need to hack `apex`'s build to bypass the check by applying this patch first and then build it.
+```
+diff --git a/setup.py b/setup.py
+index d76e998..f224dae 100644
+--- a/setup.py
++++ b/setup.py
+@@ -31,6 +31,8 @@ def check_cuda_torch_binary_vs_bare_metal(cuda_dir):
+     print(raw_output + "from " + cuda_dir + "/bin\n")
+
+     if (bare_metal_major != torch_binary_major) or (bare_metal_minor != torch_binary_minor):
++        # allow minor diffs
++        if bare_metal_minor != torch_binary_minor: return
+         raise RuntimeError(
+             "Cuda extensions are being compiled with a version of Cuda that does "
+             "not match the version used to compile Pytorch binaries.  "
+```
+
+
+4. Checkout and prepare `Megatron-DeepSpeed` and install its requirements
+
+```
+git clone https://github.com/bigscience-workshop/Megatron-DeepSpeed
+cd Megatron-DeepSpeed
+pip install -r requirements.txt
+```
+
+
+
+
+## 2. Data
+
+Will work under the `Megatron-DeepSpeed` clone
+
+```
+cd Megatron-DeepSpeed
+```
+
+
+
+Prepare data for preprocessing
+```
+mkdir -p data
+wget https://s3.amazonaws.com/models.huggingface.co/bert/gpt2-vocab.json -O data/gpt2-vocab.json
+wget https://s3.amazonaws.com/models.huggingface.co/bert/gpt2-merges.txt -O data/gpt2-merges.txt
+python -c 'from datasets import load_dataset; ds = load_dataset("stas/oscar-en-10k", split="train", keep_in_memory=False); ds.to_json(f"data/oscar-en-10k.jsonl", orient="records", lines=True, force_ascii=False)'
+```
+
+Pre-process a small dataset to be used for training
+
+```
+python tools/preprocess_data.py \
+    --input data/oscar-en-10k.jsonl \
+    --output-prefix data/meg-gpt2-oscar-en-10k \
+    --dataset-impl mmap \
+    --tokenizer-type GPT2BPETokenizer \
+    --merge-file data/gpt2-merges.txt \
+    --vocab data/gpt2-vocab.json \
+    --append-eod \
+    --workers 4
+```
+
+now you have data/meg-gpt2-oscar-en-10k, vocab and merges files to pass as arguments to training, the next section shows how to use them.
+
+Note that Megatron wants `data/meg-gpt2-oscar-en-10k_text_document` prefix later in `--data-path`
+
+## 3. Train
+
+Here is a tiny model training setup configured over 2 gpus to train on the data we prepared in step 2.
+
+Put it in a script or run it directly.
+
+If you have only 1 gpu, change these 2 lines below to:
+
+```
+N_GPUS=1
+TP_SIZE=1
+```
+
+The script:
+
+```
+CHECKPOINT_PATH=checkpoints/gpt2
+
+VOCAB_FILE=data/gpt2-vocab.json
+MERGE_FILE=data/gpt2-merges.txt
+DATA_PATH=data/meg-gpt2-oscar-en-10k_text_document
+TENSORBOARD_PATH=output_dir/tensorboard
+
+N_GPUS=2
+MICRO_BATCH_SIZE=1
+GLOBAL_BATCH_SIZE=16
+TP_SIZE=2
+PP_SIZE=1
+
+NLAYERS=2
+NHIDDEN=8
+NHEADS=2
+SEQ_LEN=512
+VOCAB_SIZE=50257
+
+SAVE_INTERVAL=50
+
+TRAIN_SAMPLES=10_000
+
+GPT_ARGS=" \
+    --num-layers $NLAYERS \
+    --hidden-size $NHIDDEN \
+    --num-attention-heads $NHEADS \
+    --seq-length $SEQ_LEN \
+    --max-position-embeddings $SEQ_LEN \
+    --micro-batch-size $MICRO_BATCH_SIZE \
+    --rampup-batch-size 2 2 1_000 \
+    --global-batch-size $GLOBAL_BATCH_SIZE \
+    --train-samples $TRAIN_SAMPLES \
+    --optimizer adam \
+    --adam-beta1 0.9 \
+    --adam-beta2 0.95 \
+    --adam-eps 1e-8 \
+    --lr 1e-4 \
+    --lr-warmup-samples 5 \
+    --min-lr 1e-6 \
+    --lr-decay-style cosine \
+    --lr-decay-samples 12 \
+    --clip-grad 1.0 \
+    --weight-decay 1e-1 \
+    --embed-layernorm \
+    --fp16 \
+    --partition-activations \
+    --seed 42 \
+    --vocab-file $VOCAB_FILE \
+    --merge-file $MERGE_FILE \
+    "
+
+OUTPUT_ARGS=" \
+    --exit-interval 100 \
+    --log-interval 10 \
+    --save-interval $SAVE_INTERVAL \
+    --eval-interval 100 \
+    --eval-iters 10 \
+    --checkpoint-activations \
+    "
+
+DATA_ARGS=" \
+    --save $CHECKPOINT_PATH \
+    --load $CHECKPOINT_PATH \
+    --data-path $DATA_PATH \
+    --tensorboard-dir $TENSORBOARD_PATH \
+    --tensorboard-queue-size 5 \
+    --log-timers-to-tensorboard \
+    --log-batch-size-to-tensorboard \
+    --log-validation-ppl-to-tensorboard \
+    --kill-switch-path /tmp/kill-switch \
+    "
+
+ZERO_STAGE=1
+
+config_json="./ds_config.json"
+
+# Deepspeed figures out GAS dynamically from dynamic GBS via set_train_batch_size()
+cat <<EOT > $config_json
+{
+  "train_micro_batch_size_per_gpu": $MICRO_BATCH_SIZE,
+  "train_batch_size": $GLOBAL_BATCH_SIZE,
+  "gradient_clipping": 1.0,
+  "zero_optimization": {
+    "stage": $ZERO_STAGE
+  },
+  "fp16": {
+    "enabled": true,
+    "loss_scale": 0,
+    "loss_scale_window": 500,
+    "hysteresis": 2,
+    "min_loss_scale": 1,
+    "initial_scale_power": 12
+  },
+  "steps_per_print": 2000,
+  "wall_clock_breakdown": false
+}
+EOT
+
+DEEPSPEED_ARGS=" \
+    --deepspeed \
+    --deepspeed_config ${config_json} \
+    --zero-stage ${ZERO_STAGE} \
+    --deepspeed-activation-checkpointing \
+    "
+
+ALL_ARGS="$GPT_ARGS $OUTPUT_ARGS $DATA_ARGS $DEEPSPEED_ARGS"
+
+MASTER_ADDR=localhost
+MASTER_PORT=6777
+
+export LAUNCHER="python -u -m torch.distributed.run \
+    --nproc_per_node $N_GPUS \
+    --nnodes 1 \
+    --rdzv_endpoint $MASTER_ADDR:$MASTER_PORT \
+    --rdzv_backend c10d \
+    --max_restarts 0 \
+    --tee 3 \
+    "
+export CMD=" \
+    $LAUNCHER pretrain_gpt.py \
+    --tensor-model-parallel-size $TP_SIZE \
+    --pipeline-model-parallel-size $PP_SIZE \
+    --distributed-backend nccl \
+    $ALL_ARGS \
+    "
+
+echo $CMD
+
+$CMD
+
+```
+
+You can, of course, run this as a slurm script, but here is [a full slurm script example](https://github.com/bigscience-workshop/bigscience/blob/d57b76bb592832bb4d2054cd5cbf132796be2d83/train/tr11-176B-ml/setup-test-n2.slurm), which has some tweaks to get `MASTER_ADDR` and a few other bits right under the SLURM environment on JeanZay, which may or may not be needed if you run it elsewhere.
+
+Remember to wipe out `$CHECKPOINT_PATH`, if you change the model shape and there is a checkpoint with the old shapes saved already.

tasks/data_utils.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.
+
+""" Tasks data utility."""
+
+import re
+import numpy as np
+
+
+def clean_text(text):
+    """Remove new lines and multiple spaces and adjust end of sentence dot."""
+
+    text = text.replace("\n", " ")
+    text = re.sub(r'\s+', ' ', text)
+    for _ in range(3):
+        text = text.replace(' . ', '. ')
+
+    return text
+
+
+def build_sample(ids, types, paddings, label, unique_id):
+    """Convert to numpy and return a sample consumed by the batch producer."""
+
+    ids_np = np.array(ids, dtype=np.int64)
+    types_np = np.array(types, dtype=np.int64)
+    paddings_np = np.array(paddings, dtype=np.int64)
+    sample = ({'text': ids_np,
+               'types': types_np,
+               'padding_mask': paddings_np,
+               'label': int(label),
+               'uid': int(unique_id)})
+
+    return sample
+
+
+def build_tokens_types_paddings_from_text(text_a, text_b,
+                                          tokenizer, max_seq_length):
+    """Build token types and paddings, trim if needed, and pad if needed."""
+
+    text_a_ids = tokenizer.tokenize(text_a)
+    text_b_ids = None
+    if text_b is not None:
+        text_b_ids = tokenizer.tokenize(text_b)
+
+    return build_tokens_types_paddings_from_ids(text_a_ids, text_b_ids,
+                                                max_seq_length, tokenizer.cls,
+                                                tokenizer.sep, tokenizer.pad)
+
+
+def build_tokens_types_paddings_from_ids(text_a_ids, text_b_ids, max_seq_length,
+                                         cls_id, sep_id, pad_id):
+    """Build token types and paddings, trim if needed, and pad if needed."""
+
+    ids = []
+    types = []
+    paddings = []
+
+    # [CLS].
+    ids.append(cls_id)
+    types.append(0)
+    paddings.append(1)
+
+    # A.
+    len_text_a = len(text_a_ids)
+    ids.extend(text_a_ids)
+    types.extend([0] * len_text_a)
+    paddings.extend([1] * len_text_a)
+
+    # [SEP].
+    ids.append(sep_id)
+    types.append(0)
+    paddings.append(1)
+
+    # B.
+    if text_b_ids is not None:
+        len_text_b = len(text_b_ids)
+        ids.extend(text_b_ids)
+        types.extend([1] * len_text_b)
+        paddings.extend([1] * len_text_b)
+
+    # Cap the size.
+    trimmed = False
+    if len(ids) >= max_seq_length:
+        max_seq_length_m1 = max_seq_length - 1
+        ids = ids[0:max_seq_length_m1]
+        types = types[0:max_seq_length_m1]
+        paddings = paddings[0:max_seq_length_m1]
+        trimmed = True
+
+    # [SEP].
+    if (text_b_ids is not None) or trimmed:
+        ids.append(sep_id)
+        if text_b_ids is None:
+            types.append(0)
+        else:
+            types.append(1)
+        paddings.append(1)
+
+    # Padding.
+    padding_length = max_seq_length - len(ids)
+    if padding_length > 0:
+        ids.extend([pad_id] * padding_length)
+        types.extend([pad_id] * padding_length)
+        paddings.extend([0] * padding_length)
+
+    return ids, types, paddings

tasks/ensemble_classifier.py

+import os
+import argparse
+import collections
+
+import numpy as np
+import torch
+
+
+def process_files(args):
+    all_predictions = collections.OrderedDict()
+    all_labels = collections.OrderedDict()
+    all_uid = collections.OrderedDict()
+    for path in args.paths:
+        path = os.path.join(path, args.prediction_name)
+        try:
+            data = torch.load(path)
+            for dataset in data:
+                name, d = dataset
+                predictions, labels, uid = d
+                if name not in all_predictions:
+                    all_predictions[name] = np.array(predictions)
+                    if args.labels is None:
+                        args.labels = [i for i in range(all_predictions[name].shape[1])]
+                    if args.eval:
+                        all_labels[name] = np.array(labels)
+                    all_uid[name] = np.array(uid)
+                else:
+                    all_predictions[name] += np.array(predictions)
+                    assert np.allclose(all_uid[name], np.array(uid))
+        except Exception as e:
+            print(e)
+            continue
+    return all_predictions, all_labels, all_uid
+
+
+def get_threshold(all_predictions, all_labels, one_threshold=False):
+    if one_threshold:
+        all_predictons = {'combined': np.concatenate(list(all_predictions.values()))}
+        all_labels = {'combined': np.concatenate(list(all_predictions.labels()))}
+    out_thresh = []
+    for dataset in all_predictions:
+        preds = all_predictions[dataset]
+        labels = all_labels[dataset]
+        out_thresh.append(calc_threshold(preds, labels))
+    return out_thresh
+
+
+def calc_threshold(p, l):
+    trials = [(i) * (1. / 100.) for i in range(100)]
+    best_acc = float('-inf')
+    best_thresh = 0
+    for t in trials:
+        acc = ((apply_threshold(p, t).argmax(-1) == l).astype(float)).mean()
+        if acc > best_acc:
+            best_acc = acc
+            best_thresh = t
+    return best_thresh
+
+
+def apply_threshold(preds, t):
+    assert (np.allclose(preds.sum(-1), np.ones(preds.shape[0])))
+    prob = preds[:, -1]
+    thresholded = (prob >= t).astype(int)
+    preds = np.zeros_like(preds)
+    preds[np.arange(len(thresholded)), thresholded.reshape(-1)] = 1
+    return preds
+
+
+def threshold_predictions(all_predictions, threshold):
+    if len(threshold) != len(all_predictions):
+        threshold = [threshold[-1]] * (len(all_predictions) - len(threshold))
+    for i, dataset in enumerate(all_predictions):
+        thresh = threshold[i]
+        preds = all_predictions[dataset]
+        all_predictions[dataset] = apply_threshold(preds, thresh)
+    return all_predictions
+
+
+def postprocess_predictions(all_predictions, all_labels, args):
+    for d in all_predictions:
+        all_predictions[d] = all_predictions[d] / len(args.paths)
+
+    if args.calc_threshold:
+        args.threshold = get_threshold(all_predictions, all_labels, args.one_threshold)
+        print('threshold', args.threshold)
+
+    if args.threshold is not None:
+        all_predictions = threshold_predictions(all_predictions, args.threshold)
+
+    return all_predictions, all_labels
+
+
+def write_predictions(all_predictions, all_labels, all_uid, args):
+    all_correct = 0
+    count = 0
+    for dataset in all_predictions:
+        preds = all_predictions[dataset]
+        preds = np.argmax(preds, -1)
+        if args.eval:
+            correct = (preds == all_labels[dataset]).sum()
+            num = len(all_labels[dataset])
+            accuracy = correct / num
+            count += num
+            all_correct += correct
+            accuracy = (preds == all_labels[dataset]).mean()
+            print(accuracy)
+        if not os.path.exists(os.path.join(args.outdir, dataset)):
+            os.makedirs(os.path.join(args.outdir, dataset))
+        outpath = os.path.join(
+            args.outdir, dataset, os.path.splitext(
+                args.prediction_name)[0] + '.tsv')
+        with open(outpath, 'w') as f:
+            f.write('id\tlabel\n')
+            f.write('\n'.join(str(uid) + '\t' + str(args.labels[p])
+                              for uid, p in zip(all_uid[dataset], preds.tolist())))
+    if args.eval:
+        print(all_correct / count)
+
+
+def ensemble_predictions(args):
+    all_predictions, all_labels, all_uid = process_files(args)
+    all_predictions, all_labels = postprocess_predictions(all_predictions, all_labels, args)
+    write_predictions(all_predictions, all_labels, all_uid, args)
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--paths', required=True, nargs='+',
+                        help='paths to checkpoint directories used in ensemble')
+    parser.add_argument('--eval', action='store_true',
+                        help='compute accuracy metrics against labels (dev set)')
+    parser.add_argument('--outdir',
+                        help='directory to place ensembled predictions in')
+    parser.add_argument('--prediction-name', default='test_predictions.pt',
+                        help='name of predictions in checkpoint directories')
+    parser.add_argument('--calc-threshold', action='store_true',
+                        help='calculate threshold classification')
+    parser.add_argument('--one-threshold', action='store_true',
+                        help='use on threshold for all subdatasets')
+    parser.add_argument('--threshold', nargs='+', default=None, type=float,
+                        help='user supplied threshold for classification')
+    parser.add_argument('--labels', nargs='+', default=None,
+                        help='whitespace separated list of label names')
+    args = parser.parse_args()
+    ensemble_predictions(args)
+
+
+if __name__ == '__main__':
+    main()

tasks/eval_harness/download.py

+# Downloads the specified taks in the evaluation harness
+# This is particularly useful when running in environments where the GPU nodes 
+# do not have internet access. This way we can pre-download them and use the cached data-set during evaluation.
+
+from lm_eval import tasks
+from lm_eval.tasks import ALL_TASKS
+import argparse
+import os
+
+
+parser = argparse.ArgumentParser(description='Download evaluation harness', allow_abbrev=False)
+parser.add_argument('--task_list', type=str, default = "all", help='Either "all" or comma separated list of tasks to download.')
+args = parser.parse_args()
+
+def main():
+    task_list = ALL_TASKS if args.task_list == 'all' else args.task_list.split(',')
+    tasks.get_task_dict(task_list)
+
+if __name__ == '__main__':
+    main()
+
+
+    
\ No newline at end of file

tasks/eval_harness/evaluate.py

+from functools import reduce
+from logging import logMultiprocessing
+import os
+import sys
+import datetime
+sys.path.append(os.path.abspath(os.path.join(os.path.dirname(__file__),
+                                             os.path.pardir,os.path.pardir)))
+
+from lm_eval.models.gpt2 import GPT2LM
+from lm_eval import evaluator, tasks, utils
+from lm_eval.base import CacheHook
+from tqdm import tqdm
+import torch.nn.functional as F
+
+from lm_eval.tasks import ALL_TASKS
+from pretrain_gpt import model_provider
+import numpy as np
+
+import torch
+from megatron import get_args
+from megatron import print_rank_0
+from megatron import get_tokenizer
+from megatron import mpu
+from megatron.training import setup_model_and_optimizer, get_model
+from megatron.mpu.mappings import gather_from_tensor_model_parallel_region
+
+from megatron.utils import get_ltor_masks_and_position_ids, unwrap_model
+from megatron.p2p_communication import recv_forward, send_forward
+import pickle
+import json
+
+from torch.nn.parallel.distributed import DistributedDataParallel as torchDDP
+from megatron.model.distributed import DistributedDataParallel as LocalDDP
+from megatron.model.module import Float16Module
+from deepspeed.runtime.pipe import schedule
+
+class EvalHarnessAdaptor(GPT2LM):
+    def __init__(self, model, tokenizer):
+        args = get_args()
+        self.args = args
+        self.model = model
+        self.tokenizer = tokenizer
+        self.VOCAB_SIZE = tokenizer.vocab_size
+        self.EOT_TOKEN_ID = tokenizer.eod
+
+        self._max_length = args.seq_length
+
+        # For ds we split into mini batches and then micro batches to keep pipelining api happy.
+        # With Megatron we just go to micro_batches directly
+        self._batch_size = args.micro_batch_size * args.micro_bs_multiplier
+
+        self.cache_hook = CacheHook(None)
+        self.is_main = args.rank == 0
+        self.is_local_main = args.local_rank == 0
+        self._device = torch.cuda.current_device()
+        self.is_model_parallel = mpu.get_tensor_model_parallel_world_size() > 1
+        self.is_pipe_parallel = mpu.get_pipeline_model_parallel_world_size() > 1
+        self.is_data_parallel = mpu.get_data_parallel_world_size() > 1
+        self.adaptive_seq_len = args.adaptive_seq_len
+        if self.is_data_parallel:
+            raise NotImplementedError("Data parallelism is currently not supported for evaluation")
+
+        self.is_last_stage = True if not self.is_pipe_parallel else mpu.is_pipeline_last_stage()  # only the last stage of the pipeline model will receive the logits
+
+    @property
+    def max_length(self):
+        return self._max_length
+
+    @property
+    def batch_size(self):
+        return self._batch_size
+
+    @property
+    def device(self):
+        return self._device
+
+
+    def loglikelihood(self, requests):
+        new_reqs = []
+        for context, continuation in requests:
+            if context == "":
+                # end of text as context
+                context_enc = [self.EOT_TOKEN_ID]
+            else:
+                context_enc = self.tokenizer_encode(context)
+
+            continuation_enc = self.tokenizer_encode(continuation)
+
+            new_reqs.append(((context, continuation), context_enc, continuation_enc))
+
+        return self._loglikelihood_tokens(new_reqs)
+
+    def loglikelihood_rolling(self, requests):
+        # TODO: Implement caching once we've confirmed the perplexity implementation
+        # TODO: automatic batch size detection for vectorization
+
+        loglikelihoods = []
+        with torch.no_grad():
+            for string, in tqdm(requests):
+                rolling_token_windows = list(map(utils.make_disjoint_window, utils.get_rolling_token_windows(
+                    token_list=self.tokenizer_encode(string),
+                    prefix_token=self.EOT_TOKEN_ID,
+                    max_seq_len=self.max_length,
+                    context_len=1,
+                )))
+
+                rolling_token_windows = [(None,) + x for x in rolling_token_windows]
+
+                # TODO: extract out this call so it only gets called once and also somehow figure out partial caching for that
+                string_nll = self._loglikelihood_tokens(rolling_token_windows, disable_tqdm=True)
+
+                # discard is_greedy
+                string_nll = [x[0] for x in string_nll]
+
+                string_nll = sum(string_nll)
+                loglikelihoods.append(string_nll)
+
+        return loglikelihoods
+
+    def _loglikelihood_tokens(self, requests, disable_tqdm=False):
+        disable_tqdm = disable_tqdm if self.is_main else True
+        res = []
+        res_len = 0  # storing the result length for later
+        self.model.eval()
+        with torch.no_grad():
+            def _collate(x):
+                toks = x[1] + x[2]
+                return (-len(toks), tuple(toks))
+
+            reord = utils.Reorderer(requests, _collate)
+            for chunk in utils.chunks(tqdm(reord.get_reordered(), disable=disable_tqdm), self.batch_size):
+                inps, contlens, inplens, padding_length = [], [], [], None
+                for _, context_enc, continuation_enc in chunk:
+                    # when too long to fit in context, truncate from the left
+                    inp = torch.tensor(
+                        (context_enc + continuation_enc)[-(self.max_length + 1):][:-1]
+                        , dtype=torch.long).to(self.device)
+                    inplen, = inp.shape
+
+                    cont = continuation_enc
+
+                    # since in _collate we make sure length is descending, the longest is always the first one.
+                    padding_length = padding_length if padding_length is not None else inplen
+                    if not self.adaptive_seq_len:
+                        padding_length = self.max_length
+                    # pad to length
+                    inp = torch.cat([
+                        inp,  # [seq]
+                        torch.zeros(padding_length - inplen, dtype=torch.long).to(inp.device)  # [padding_length - seq]
+                    ], dim=0)
+
+                    inps.append(inp.unsqueeze(0))
+
+                    contlens.append(cont)
+                    inplens.append(inplen)
+
+                logits = self._model_call(torch.cat(inps, dim=0))
+                res_len += len(chunk)
+                if logits is not None:
+                    multi_logits = F.log_softmax(logits, dim=-1).cpu()  # [batch, seq, vocab]
+
+                    for (cache_key, _, _), logits, inp, inplen, cont_toks in zip(chunk, multi_logits, inps, inplens, contlens):
+                        contlen = len(cont_toks)
+                        logits = logits[inplen - contlen:inplen].unsqueeze(0)  # [1, seq, vocab]
+                        greedy_tokens = logits.argmax(dim=-1)
+                        # cont_toks :: [1, seq]
+                        cont_toks = torch.tensor(cont_toks, dtype=torch.long).unsqueeze(0)
+                        max_equal = (greedy_tokens == cont_toks).all()
+                        # last_token_slice = logits[:, -1, :].squeeze(0).tolist()
+
+                        logits = torch.gather(logits, 2, cont_toks.unsqueeze(-1)).squeeze(-1)  # [1, seq]
+                        answer = (float(logits.sum()), bool(max_equal))
+                        # partial caching
+                        if cache_key is not None:
+                            self.cache_hook.add_partial("loglikelihood", cache_key, answer)
+                        res.append(answer)
+
+        if not mpu.is_pipeline_last_stage():
+            # @HACK: To make the eval harness happy on threads that don't have access to the results.
+            #        We just randomly generate some data.
+            res = [(np.random.rand(), np.random.rand()>0.5) for _ in requests]
+
+        return reord.get_original(res)
+
+    def create_model_inputs(self, tokens):
+        args = get_args()
+
+        attention_mask, loss_mask, position_ids = get_ltor_masks_and_position_ids(
+            tokens,
+            self.EOT_TOKEN_ID,
+            args.reset_position_ids,
+            args.reset_attention_mask,
+            args.eod_mask_loss,
+            prefix_indices=None,
+            loss_on_targets_only=False)
+
+        return (tokens, position_ids, attention_mask), (tokens, loss_mask)
+
+    def _model_call(self, inps):
+        args = get_args()
+
+        if args.deepspeed:
+            self.model.set_batch_fn(self.create_model_inputs)
+            # round up to multiple of micro_batch_size
+            new_size = ((len(inps) + args.micro_batch_size-1)  // args.micro_batch_size) * args.micro_batch_size
+            padded = F.pad(inps, (0, 0, 0, new_size-len(inps)), value = 0)
+            # dummy data iterator for pipelining.
+            data_iterator = list((torch.stack(inp) for inp in utils.chunks(padded, args.micro_batch_size)))
+            self.model.micro_batches = len(data_iterator)
+            
+            if self.adaptive_seq_len:
+                # Allow different shapes than the default seq_len to be communicated across pipes
+                # Without this Deepspeed will hang when trying to receive activations
+                self.model.reset_activation_shape()
+            
+            output = self.model.eval_batch(iter(data_iterator), compute_loss = False, reduce_output = None)
+
+
+            if output is not None:
+                output = torch.cat(output, 0)[:len(inps)]
+            else:
+                output = None
+
+            # hack #2 for adaptive_seq_len to work as total_loss gets appended to and shapes aren't the same
+            if args.adaptive_seq_len:
+                self.model.total_loss = None
+        else:
+            # Since the shape of the micro-batch will change
+            # We need set the correct shapes here
+            # So that latter pipeline stages knows which shapes to expect.
+            # Otherwise we will deadlock.
+
+            args.micro_batch_size = len(inps)
+            args.seq_length = len(inps[0])
+            args.max_position_embeddings = args.seq_length
+
+            input_tensor = recv_forward()
+
+            # Forward pass through the model.
+            unwrapped_model = unwrap_model(self.model, (torchDDP, LocalDDP, Float16Module))
+            unwrapped_model.set_input_tensor(input_tensor)
+            output = self.model(*self.create_model_inputs(inps)[0])
+            send_forward(output)
+
+        if mpu.is_pipeline_last_stage():
+            return gather_from_tensor_model_parallel_region(output)[..., :self.tokenizer.vocab_size]
+        else:
+            return None
+
+    def tokenizer_encode(self, text):
+        """Tokenize text *without* adding special tokens."""
+        # Splitting this into its own method in case we need to handle special cases for different tokenizers
+        from megatron.tokenizer.gpt2_tokenization import GPT2Tokenizer
+        if isinstance(self.tokenizer.tokenizer, GPT2Tokenizer):
+            return self.tokenizer.tokenizer.encode(text)
+        else:
+            return self.tokenizer.tokenizer.encode(text, add_special_tokens=False)
+
+
+from megatron.initialize import initialize_megatron
+import megatron
+
+from tools.convert_checkpoint.deepspeed_checkpoint import DeepSpeedCheckpoint
+from tools.convert_checkpoint.deepspeed_to_megatron import _create_rank_checkpoint
+
+def override_args(args, override_args, skip_keys, skip_if_specified_keys):
+    for k, v in vars(override_args).items():
+        if k in skip_keys:
+            continue
+        if k in skip_if_specified_keys and getattr(args, k) is not None:
+            continue
+        setattr(args, k, v)
+
+
+# Note(Hesslow):
+# The model loading is a bit convoluted.
+# We want to parse out the model arguments from the checkpoint and use those to initialize megatron-ds.
+#
+# However megatron-ds expects its arguments on the command line.
+# And at that point we don't know them.
+#
+# Instead we use Jasons way: we load the arguments form the checkpoint and then override _parse_args to return whatever args we want.
+#
+# If the checkpoint is old, some new arguments may have been introduced and the code will expect these arguments to exist.
+# In order to support this we _first_ parse the arguments normally, and then override them with the arguments from the checkpoint.
+# Keeping the default-value of newer arguments.
+#
+# We then use the megatron deepspeed converter to load the deepspeed checkpoints as if they we're megatron checkpoints.
+def load_ds_checkpoint_and_setup_megatron(args):
+    _print_args = megatron.arguments._print_args
+    megatron.arguments._print_args = lambda *_args, **kwarg: None
+
+    if not os.path.exists(args.load):
+        raise ValueError(f"checkpoint path {args.load} doesn't exit")
+
+    ds_checkpoint = DeepSpeedCheckpoint(args.load,
+                                        tp_degree=args.tensor_model_parallel_size,
+                                        pp_degree=args.pipeline_model_parallel_size)
+
+
+    cp_args = ds_checkpoint.get_args()
+    # Merge the current args with the checkpoint args.
+    skip_keys = [
+        'abort_on_unmet_fused_kernel_constraints',
+        'batch_size',
+        'data_parallel_size',
+        'deepspeed',
+        'deepspeed_config',
+        'device_count',
+        'global_batch_size',
+        'inference',
+        'iteration',
+        'load',
+        'local_rank',
+        'micro_batch_size',
+        'pipeline_model_parallel_size',
+        'rampup_batch_size',
+        'rank',
+        'tensor_model_parallel_size',
+        'tensorboard_dir',
+        'world_size',
+    ]
+
+    skip_if_specified = ['merge_file', 'vocab_file']
+
+    if args.eval_fp32:
+        cp_args.fp16 = False
+        cp_args.bf16 = False
+        cp_args.params_dtype = torch.float32
+
+    override_args(args, cp_args, skip_keys, skip_if_specified)
+
+    # stop megatron from reparsing the arguments.
+    megatron.global_vars._parse_args = lambda *_args, **kwarg: args
+    megatron.global_vars._GLOBAL_ARGS = args
+
+    initialize_megatron()
+    torch.distributed.barrier()
+
+    # Initializing megatron will update eg. tokenizer size. Override again.
+    override_args(args, cp_args, skip_keys, skip_if_specified)
+
+    # print final arguments.
+    _print_args(args)
+    if args.deepspeed:
+
+        # Hack #3:
+        # Loading pipelined models in deepspeed with different TP than it was originally trained on fails
+        # due to a sanity check, that makes sure that all state_dicts that we merge contains attention layers.
+        # This, however, is not true for pipelining when we will merge the state_dict for the embeddings which
+        # which does not contain these attention-specific keys.
+        #
+        # Deepspeed does however manage to load the model if we just turn off this sanity check.
+        import deepspeed
+        deepspeed.runtime.state_dict_factory.MegatronSDLoader.sanity_check = lambda self, ckpt_file_name: None
+
+
+        cp_path = args.load
+        args.load = None
+        model, _, _ = setup_model_and_optimizer(model_provider)
+        model = model[0]
+        zero_enabled = model._config.zero_enabled
+        model._config.zero_enabled = False
+        _, _ = model.load_checkpoint(cp_path, tag = '.', load_optimizer_states=False, load_lr_scheduler_states=False, load_module_only=True)
+        model._config.zero_enabled = zero_enabled
+    else:
+        model = get_model(model_provider)[0]
+        # Initialize megatron model using the parsed state dict.
+        sd = _create_rank_checkpoint(ds_checkpoint, None, mpu.get_tensor_model_parallel_rank(), mpu.get_pipeline_model_parallel_rank(), True)
+
+        model.load_state_dict(sd['model'], strict=True)
+
+    if args.eval_fp32:
+        model = model.float()
+
+    torch.distributed.barrier()
+    return model
+
+def tasks_args(parser):
+
+    results_path_default = f"results-{datetime.datetime.now().strftime('%Y-%m-%d-%H-%M-%S')}.json"
+
+    """Provide extra arguments required for tasks."""
+    group = parser.add_argument_group(title='Evaluation options')
+    group.add_argument('--task_list', type=str, default = "all", help='Either "all" or comma separated list of tasks.')
+    group.add_argument('--results_path', type=str, default = results_path_default, help='Path to where the results will be stored.')
+    group.add_argument('--adaptive_seq_len',  default = False, action='store_true',
+                       help='Should the sequence length be adapted to the batch during evaluation, if in fp16 the results will be slightly different due to numerical errors but greatly speed up evaluation.')
+    group.add_argument('--eval_fp32',  default = False, action='store_true', help='Should the evaluation run in fp32')
+    group.add_argument('--intermed_results',  default = False, action='store_true', help='Whether to print & write intermediate results for each task')
+    group.add_argument('--bootstrap_iters', type=int, default=100000, help='How many iterations to use for stderr estimation')
+    group.add_argument('--micro_bs_multiplier', type=int, default=1, help='Increase the global batch size to remove bubble when pipeline parallel')
+    return parser
+
+from megatron.global_vars import _parse_args
+
+def main():
+    # parse the megatron args. But wait with initalizing megatron.
+    # avoid printing the arguments, since they will later be overridden.
+    args = _parse_args(tasks_args)
+    load_path = args.load
+    model = load_ds_checkpoint_and_setup_megatron(args)
+
+    args = get_args()
+    if args.deepspeed and args.adaptive_seq_len:
+        # adaptive_seq_len hack #1:
+        # CL automatically enables reset_activation_shape() which allows us to change input shapes
+        # and it also reshapes the attenion scores in attention_mask_func
+        args.curriculum_learning = 1
+
+    task_list = ALL_TASKS if args.task_list == 'all' else args.task_list.split(',')
+    task_dict = tasks.get_task_dict(task_list)
+
+    model.module.activation_checkpoint_interval = 0
+    model._compute_loss = False
+    model.fwd_outputs = []
+
+    tokenizer = get_tokenizer()
+    adaptor = EvalHarnessAdaptor(model, tokenizer)
+    
+    if args.intermed_results:
+        global_results = {"results": {}, "versions": {}}
+        timestamp = datetime.datetime.now().strftime('%Y-%m-%d-%H-%M-%S')
+        iteration_id = load_path.split("/")[-1].replace("/", "")
+        results_path = args.results_path.replace(".json", f"_lm-eval_{iteration_id}_{timestamp}.json")
+        # Backup file in case of interruption during writing
+        results_path_backup = args.results_path.replace(".json", f"_lm-eval_{iteration_id}_{timestamp}_backup.json")
+        for task_name, task in task_dict.items():
+            results = evaluator.evaluate(adaptor, {task_name: task}, False, 0, None, bootstrap_iters=args.bootstrap_iters)
+            global_results["results"] = {**global_results["results"], **results["results"]}
+            global_results["versions"] = {**global_results["versions"], **results["versions"]}
+            if mpu.is_pipeline_last_stage() and mpu.get_tensor_model_parallel_rank() == 0:
+                print(json.dumps(results, indent=2))
+                with open(results_path, 'w') as outfile:
+                    json.dump(global_results, outfile, indent=4)
+                with open(results_path_backup, 'w') as outfile:
+                    json.dump(global_results, outfile, indent=4)
+    else:
+        global_results = evaluator.evaluate(adaptor, task_dict, False, 0, None, bootstrap_iters=args.bootstrap_iters)
+        if mpu.is_pipeline_last_stage() and mpu.get_tensor_model_parallel_rank() == 0:
+            print(json.dumps(global_results, indent=2))
+            with open(args.results_path, 'w') as outfile:
+                json.dump(global_results, outfile, indent=4)
+
+if __name__ == '__main__':
+    main()

tasks/eval_harness/report-to-csv.py

+#!/usr/bin/env python
+
+# this script converts results.json:
+#
+#   "results": {
+#     "arc_challenge": {
+#       "acc": 0.24232081911262798,
+#       "acc_stderr": 0.01252159329580012,
+#       "acc_norm": 0.2764505119453925,
+#       "acc_norm_stderr": 0.013069662474252425
+#     },
+#
+# into a format expected by a spreadsheet, which is:
+#
+#   task          metric   value    err
+#   arc_challenge acc      xxx      yyy
+#   arc_challenge acc_norm xxx      yyy
+#   arc_challenge f1       xxx      yyy
+#
+# usage:
+# report-to-csv.py results.json
+
+
+import sys
+import json
+import io
+import csv
+
+results_file = sys.argv[1]
+
+csv_file = results_file.replace("json", "csv")
+
+print(f"Converting {results_file} to {csv_file}")
+
+with io.open(results_file, 'r', encoding='utf-8') as f:
+    results = json.load(f)
+
+with io.open(csv_file, 'w', encoding='utf-8') as f:
+
+    writer = csv.writer(f)
+    writer.writerow(["task", "metric", "value", "err", "version"])
+
+    versions = results["versions"]
+
+    for k,v in sorted(results["results"].items()):
+        if k not in versions:
+            versions[k] = -1
+
+        if "acc" in v:
+            writer.writerow([k, "acc", v["acc"], v["acc_stderr"], versions[k]])
+        if "acc_norm" in v:
+            writer.writerow([k, "acc_norm", v["acc_norm"], v["acc_norm_stderr"], versions[k]])
+        if "f1" in v:
+            writer.writerow([k, "f1", v["f1"], v["f1_stderr"] if "f1_stderr" in v else "", versions[k]])
+        # if "ppl" in v:
+        #     writer.writerow([k, "ppl", v["ppl"], v["ppl_stderr"], versions[k]])
+        # if "em" in v:
+        #     writer.writerow([k, "em", v["em"], v["em_stderr"] if "em_stderr" in v else "", versions[k]])

tasks/eval_utils.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.
+
+"""Evaluation utilities."""
+
+import os
+import time
+from functools import partial
+
+import torch
+
+from megatron import get_args
+from megatron import print_rank_last, is_last_rank
+from megatron import mpu
+from megatron.schedules import get_forward_backward_func
+from tasks.finetune_utils import build_data_loader
+from tasks.finetune_utils import process_batch
+
+
+def accuracy_func_provider(single_dataset_provider):
+    """Provide function that calculates accuracies."""
+    args = get_args()
+
+    # Build dataloaders.
+    datapaths = args.valid_data
+    dataloaders = []
+    for datapath in datapaths:
+        dataset = single_dataset_provider(datapath)
+        dataloader = build_data_loader(
+            dataset, args.orig_micro_batch_size, num_workers=args.num_workers,
+            drop_last=(mpu.get_data_parallel_world_size() > 1))
+        dataloaders.append((dataset.dataset_name, dataloader))
+
+    def metrics_func(model, epoch, output_predictions=False):
+        print_rank_last('calculating metrics ...')
+        correct = 0
+        total = 0
+        if output_predictions:
+            assert mpu.get_data_parallel_world_size() == 1
+            named_predictions = []
+            names = 'predictions'
+        for name, dataloader in dataloaders:
+            output = calculate_correct_answers(name, model, dataloader,
+                                               epoch, output_predictions)
+            if not output_predictions:
+                correct_ans, total_count = output
+            else:
+                correct_ans, total_count, predictions = output
+                named_predictions.append((name, predictions))
+                names += '_' + name
+            correct += correct_ans
+            total += total_count
+        if is_last_rank():
+            percent = float(correct) * 100.0 / float(total)
+            print(' >> |epoch: {}| overall: correct / total = {} / {} = '
+                  '{:.4f} %'.format(epoch, correct, total, percent))
+
+        if output_predictions and is_last_rank():
+            assert args.load is not None
+            filename = os.path.join(args.load, names + '.pt')
+            torch.save(named_predictions, filename)
+
+    return metrics_func
+
+
+def calculate_correct_answers(name, model, dataloader,
+                              epoch, output_predictions):
+    """Calculate correct over total answers and return prediction if the
+    `output_predictions` is true."""
+    args = get_args()
+    forward_backward_func = get_forward_backward_func()
+    start_time = time.time()
+    for m in model:
+        m.eval()
+    saved_micro_batch_size = args.micro_batch_size
+    saved_global_batch_size = args.global_batch_size
+
+    ds = dataloader.dataset
+    if hasattr(ds, 'sample_multiplier'):
+        # If our dataset as a sample_multiplier attribute that means
+        # each "sample" from the dataset actually has multiple samples
+        # that will collapse into the batch dimension (for example in
+        # the RACE dataset that has several options), we need to
+        # account for that when setting the micro batch size.
+        sample_multiplier = ds.sample_multiplier
+    else:
+        sample_multiplier = 1
+    micro_batch_size_times_data_parallel = args.orig_micro_batch_size * args.data_parallel_size
+    num_micro_batches = args.orig_global_batch_size // micro_batch_size_times_data_parallel
+
+    def loss_func(output_predictions, labels, output_tensor):
+        logits = output_tensor
+
+        loss_dict = {}
+        # Add output predictions.
+        if output_predictions:
+            assert False
+            loss_dict['softmaxes'] = torch.nn.Softmax(dim=-1)(
+                logits.float()).data.cpu().numpy().tolist()
+            loss_dict['labels'] = labels.data.cpu().numpy().tolist()
+            loss_dict['ids'] = batch['uid'].cpu().numpy().tolist()
+        # Compute the correct answers.
+        predicted = torch.argmax(logits, dim=-1)
+        corrects = (predicted == labels)
+        # Add to the counters.
+        loss_dict['total'] = labels.size(0)
+        loss_dict['correct'] = corrects.sum().item()
+
+        return 0, loss_dict
+
+    # defined inside to capture output_predictions
+    def correct_answers_forward_step(batch, model):
+        try:
+            batch_ = next(batch)
+        except BaseException:
+            batch_ = batch
+        tokens, types, labels, attention_mask = process_batch(batch_)
+
+        # Forward model.
+        args = get_args()
+        output_tensor = model(tokens, attention_mask, tokentype_ids=types)
+
+        return output_tensor, partial(loss_func, output_predictions, labels)
+
+    with torch.no_grad():
+        # For all the batches in the dataset.
+        total = 0
+        correct = 0
+        if output_predictions:
+            # This option is only possible when data parallel size is 1.
+            assert mpu.get_data_parallel_world_size() == 1
+            softmaxes = []
+            labels = []
+            ids = []
+        for _, batch in enumerate(dataloader):
+            # For evaluation only mode we use drop_last = False to get all the
+            # samples, which means we might not have a full batch, so we
+            # adjust batch_size here to actual batch size of data
+            actual_batch_size = len(batch['label'])
+            # ... applying sample_multiplier if necessary
+            args.micro_batch_size = actual_batch_size * sample_multiplier
+            args.global_batch_size = actual_batch_size * sample_multiplier * num_micro_batches
+
+            loss_dicts = forward_backward_func(correct_answers_forward_step, batch, model,
+                                               optimizer=None, timers=None, forward_only=True)
+
+            for loss_dict in loss_dicts:
+                if output_predictions:
+                    softmaxes.extend(loss_dict['softmaxes'])
+                    labels.extend(loss_dict['labels'])
+                    ids.extend(loss_dict['ids'])
+                total += loss_dict['total']
+                correct += loss_dict['correct']
+
+
+    for m in model:
+        m.train()
+    args.micro_batch_size = saved_micro_batch_size
+    args.global_batch_size = saved_global_batch_size
+
+    # Reduce.
+    if mpu.is_pipeline_last_stage():
+        unreduced = torch.cuda.LongTensor([correct, total])
+        torch.distributed.all_reduce(unreduced,
+                                     group=mpu.get_data_parallel_group())
+
+        # Print on screen.
+
+        correct_ans = unreduced[0].item()
+        total_count = unreduced[1].item()
+        percent = float(correct_ans) * 100.0 / float(total_count)
+        elapsed_time = time.time() - start_time
+        print_rank_last(' > |epoch: {}| metrics for {}: correct / total '
+                        '= {} / {} = {:.4f} %, elapsed time (sec): {:.3f}'.format(
+                            epoch, name, correct_ans, total_count,
+                            percent, elapsed_time))
+
+        if output_predictions:
+            return correct_ans, total_count, (softmaxes, labels, ids)
+        return correct_ans, total_count
+    if output_predictions:
+        return 0, 0, ()
+    return 0, 0

tasks/finetune_utils.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.
+
+"""Finetune utilities."""
+
+from functools import partial
+
+import torch
+
+from megatron import get_args
+from megatron import print_rank_0
+from megatron import get_timers
+from megatron import mpu
+from megatron.checkpointing import load_checkpoint
+from megatron.checkpointing import save_checkpoint
+from megatron.training import evaluate_and_print_results
+from megatron.training import setup_model_and_optimizer
+from megatron.training import train_step
+from megatron.training import training_log
+from megatron.utils import average_losses_across_data_parallel_group
+from megatron.utils import calc_params_l2_norm
+from megatron.utils import check_adlr_autoresume_termination
+
+
+def process_batch(batch):
+    """Process batch and produce inputs for the model."""
+    args = get_args()
+
+    tokens = batch['text'].long().cuda().contiguous()
+    types = batch['types'].long().cuda().contiguous()
+    labels = batch['label'].long().cuda().contiguous()
+    attention_mask = batch['padding_mask'].float().cuda().contiguous()
+    if args.fp16:
+        attention_mask = attention_mask.half()
+
+    return tokens, types, labels, attention_mask
+
+
+def cross_entropy_loss_func(labels, output_tensor):
+    logits = output_tensor
+
+    # Cross-entropy loss.
+    loss_func = torch.nn.CrossEntropyLoss()
+    loss = loss_func(logits.contiguous().float(), labels)
+
+    # Reduce loss for logging.
+    averaged_loss = average_losses_across_data_parallel_group([loss])
+
+    return loss, {'lm loss': averaged_loss[0]}
+
+
+def _cross_entropy_forward_step(batch, model):
+    """Simple forward step with cross-entropy loss."""
+    timers = get_timers()
+
+    # Get the batch.
+    timers('batch-generator').start()
+    try:
+        batch_ = next(batch)
+    except BaseException:
+        batch_ = batch
+    tokens, types, labels, attention_mask = process_batch(batch_)
+    timers('batch-generator').stop()
+
+    # Forward model.
+    output_tensor = model(tokens, attention_mask, tokentype_ids=types)
+
+    return output_tensor, partial(cross_entropy_loss_func, labels)
+
+
+def build_data_loader(dataset, micro_batch_size, num_workers, drop_last):
+    """Data loader. Note that batch-size is the local (per GPU) batch-size."""
+
+    # Sampler.
+    world_size = mpu.get_data_parallel_world_size()
+    rank = mpu.get_data_parallel_rank()
+    sampler = torch.utils.data.distributed.DistributedSampler(
+        dataset, num_replicas=world_size, rank=rank)
+
+    # Data loader. Note that batch size is the per GPU batch size.
+    data_loader = torch.utils.data.DataLoader(dataset,
+                                              batch_size=micro_batch_size,
+                                              sampler=sampler,
+                                              shuffle=False,
+                                              num_workers=num_workers,
+                                              drop_last=drop_last,
+                                              pin_memory=True)
+
+    return data_loader
+
+
+def _build_infinite_size_dataloader(dataloader):
+    """Build a looped dataloader with infinite size."""
+
+    iterator = dataloader.__iter__()
+    while True:
+        try:
+            yield iterator.__next__()
+        except StopIteration:
+            iterator = dataloader.__iter__()
+
+
+def _build_train_valid_dataloaders(train_dataset, valid_dataset):
+    """Traing and validation dataloaders."""
+    args = get_args()
+
+    print_rank_0('building train and validation dataloaders ...')
+    # Training dataset.
+    train_dataloader = build_data_loader(train_dataset, args.micro_batch_size,
+                                         args.num_workers, not args.keep_last)
+    # Set the training iterations.
+    args.train_iters_per_epoch = len(train_dataloader)
+    args.train_iters = args.epochs * args.train_iters_per_epoch
+    # Validation dataset. For this dataset, we do not need to set up
+    # shuffling so we can just use a simple infinite loop.
+    valid_dataloader_ = build_data_loader(valid_dataset, args.micro_batch_size,
+                                          args.num_workers, not args.keep_last)
+    valid_dataloader = _build_infinite_size_dataloader(valid_dataloader_)
+
+    # Now that we've built the data loaders, set batch_size arguments
+    # to the actual batch size the model will see for this dataset.
+    # This is necessary so pipeline transfers know what size they are
+    # and the LR schedule, which is based on samples seen, gets set
+    # correctly.
+    args.orig_micro_batch_size = args.micro_batch_size
+    args.orig_global_batch_size = args.global_batch_size
+    if hasattr(train_dataset, 'sample_multiplier'):
+        # If our dataset as a sample_multiplier attribute that means
+        # each "sample" from the dataset actually has multiple samples
+        # that will collapse into the batch dimension (for example in
+        # the RACE dataset that has several options), we need to
+        # account for that when setting the micro batch size.
+        args.micro_batch_size *= train_dataset.sample_multiplier
+        args.global_batch_size *= train_dataset.sample_multiplier
+
+    return train_dataloader, valid_dataloader
+
+
+def _train(model, optimizer, lr_scheduler, forward_step,
+           train_dataloader, valid_dataloader, end_of_epoch_callback):
+    """Train the model."""
+    args = get_args()
+    timers = get_timers()
+
+    # Turn on training mode which enables dropout.
+    for m in model:
+        m.train()
+
+    # Tracking loss.
+    losses_dict_sum = {}
+
+    # Starting epoch and iteration
+    start_epoch = args.iteration // args.train_iters_per_epoch
+    start_iteration = args.iteration % args.train_iters_per_epoch
+    iteration = args.iteration
+
+    # Memory reporting flag.
+    report_memory_flag = True
+
+    # For each remaining epoch
+    timers('interval-time').start()
+    for epoch in range(start_epoch, args.epochs):
+        print_rank_0('working on epoch {} ...'.format(epoch + 1))
+
+        # Set the data loader epoch to shuffle the index iterator.
+        train_dataloader.sampler.set_epoch(args.seed + epoch)
+
+        # For all the batches in the dataset.
+        for iteration_, batch in enumerate(train_dataloader):
+
+            # Ignore the iterations before starting value
+            if iteration_ < start_iteration:
+                continue
+            # Set to zero so the next epoch does not skip any batches.
+            start_iteration = 0
+
+            # Train for one step.
+            out = train_step(forward_step, batch, model, optimizer, lr_scheduler)
+            losses_dict, skipped_iter, grad_norm, num_zeros_in_grad = out
+            iteration += 1
+
+            # Logging.
+            params_norm = None
+            if args.log_params_norm:
+                params_norm = calc_params_l2_norm(model)
+            report_memory_flag = training_log(losses_dict, losses_dict_sum,
+                                              optimizer.param_groups[0]['lr'],
+                                              iteration,
+                                              optimizer.get_loss_scale().item(),
+                                              report_memory_flag, skipped_iter,
+                                              grad_norm, params_norm, num_zeros_in_grad)
+
+            # Autoresume
+            if args.adlr_autoresume and \
+               (iteration % args.adlr_autoresume_interval == 0):
+                check_adlr_autoresume_termination(iteration, model,
+                                                  optimizer, lr_scheduler)
+
+            # Checkpointing
+            if args.save and args.save_interval and \
+               iteration % args.save_interval == 0:
+                save_checkpoint(iteration, model, optimizer, lr_scheduler)
+
+            # Evaluation
+            if args.eval_interval and iteration % args.eval_interval == 0:
+                prefix = 'iteration {}'.format(iteration)
+                evaluate_and_print_results(prefix, forward_step,
+                                           valid_dataloader, model,
+                                           iteration, False)
+
+        # Checkpointing at the end of each epoch.
+        if args.save:
+            save_checkpoint(iteration, model, optimizer, lr_scheduler)
+
+        # Callback at the end of each epoch.
+        if end_of_epoch_callback is not None:
+            end_of_epoch_callback(model, epoch)
+
+
+def finetune(train_valid_datasets_provider, model_provider,
+             forward_step=_cross_entropy_forward_step,
+             end_of_epoch_callback_provider=None):
+    """Main finetune function used across all tasks."""
+    args = get_args()
+    timers = get_timers()
+
+    assert args.rampup_batch_size is None, \
+        'batch size scaling is not supported for finetuning'
+
+    # Train and validation data loaders.
+    timers('train/valid/test dataset/dataloder').start()
+    if args.epochs > 0:
+        train_dataset, valid_dataset = train_valid_datasets_provider()
+        train_dataloader, valid_dataloader = _build_train_valid_dataloaders(
+            train_dataset, valid_dataset)
+    else:
+        args.train_iters = 0
+    timers('train/valid/test dataset/dataloder').stop()
+
+    # Build calback function.
+    timers('callback function').start()
+    end_of_epoch_callback = None
+    if end_of_epoch_callback_provider is not None:
+        end_of_epoch_callback = end_of_epoch_callback_provider()
+    timers('callback function').stop()
+
+    # Build model, optimizer and learning rate scheduler.
+    timers('model and optimizer').start()
+    model, optimizer, lr_scheduler = setup_model_and_optimizer(model_provider)
+    timers('model and optimizer').stop()
+
+    # If pretrained checkpoint is provided and we have not trained for
+    # any iteration (i.e., iteration is zero), then load the pretrained
+    # checkpoint.
+    timers('pretrained checkpoint').start()
+    if args.iteration == 0 and args.pretrained_checkpoint is not None:
+        original_load = args.load
+        args.load = args.pretrained_checkpoint
+        _ = load_checkpoint(model, None, None)
+        args.load = original_load
+        # This is critical when only model is loaded. We should make sure
+        # main parameters are also updated.
+        optimizer.reload_model_params()
+    timers('pretrained checkpoint').stop()
+
+    # Print setup timing.
+    print_rank_0('done with setups ...')
+    timers.log(['train/valid/test dataset/dataloder', 'callback function',
+                'model and optimizer', 'pretrained checkpoint'])
+    print_rank_0('training ...')
+
+    # Finetune the model.
+    if args.epochs > 0:
+        _train(model, optimizer, lr_scheduler, forward_step,
+               train_dataloader, valid_dataloader, end_of_epoch_callback)
+    # Or just evaluate.
+    else:
+        if end_of_epoch_callback is not None:
+            print_rank_0('evaluation only mode, setting epoch to -1')
+            end_of_epoch_callback(model, epoch=-1, output_predictions=True)
+    print_rank_0('done :-)')

tasks/glue/data.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.
+
+"""GLUE dataset."""
+
+from abc import ABC
+from abc import abstractmethod
+
+from torch.utils.data import Dataset
+
+from megatron import print_rank_0
+from tasks.data_utils import build_sample
+from tasks.data_utils import build_tokens_types_paddings_from_text
+
+
+class GLUEAbstractDataset(ABC, Dataset):
+    """GLUE base dataset class."""
+
+    def __init__(self, task_name, dataset_name, datapaths,
+                 tokenizer, max_seq_length):
+        # Store inputs.
+        self.task_name = task_name
+        self.dataset_name = dataset_name
+        self.tokenizer = tokenizer
+        self.max_seq_length = max_seq_length
+        print_rank_0(' > building {} dataset for {}:'.format(self.task_name,
+                                                             self.dataset_name))
+        # Process the files.
+        string = '  > paths:'
+        for path in datapaths:
+            string += ' ' + path
+        print_rank_0(string)
+        self.samples = []
+        for datapath in datapaths:
+            self.samples.extend(self.process_samples_from_single_path(datapath))
+        print_rank_0('  >> total number of samples: {}'.format(
+            len(self.samples)))
+
+    def __len__(self):
+        return len(self.samples)
+
+    def __getitem__(self, idx):
+        raw_sample = self.samples[idx]
+        ids, types, paddings = build_tokens_types_paddings_from_text(
+            raw_sample['text_a'], raw_sample['text_b'],
+            self.tokenizer, self.max_seq_length)
+        sample = build_sample(ids, types, paddings,
+                              raw_sample['label'], raw_sample['uid'])
+        return sample
+
+    @abstractmethod
+    def process_samples_from_single_path(self, datapath):
+        """Abstract method that takes a single path / filename and
+        returns a list of dataset samples, each sample being a dict of
+            {'text_a': string, 'text_b': string, 'label': int, 'uid': int}
+        """
+        pass

tasks/glue/finetune.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.
+
+"""GLUE finetuning/evaluation."""
+
+from megatron import get_args
+from megatron import print_rank_0
+from megatron import get_tokenizer
+from megatron import mpu
+from megatron.model.classification import Classification
+from tasks.eval_utils import accuracy_func_provider
+from tasks.finetune_utils import finetune
+
+
+def glue_classification(num_classes, Dataset,
+                        name_from_datapath_func):
+
+    def train_valid_datasets_provider():
+        """Build train and validation dataset."""
+        args = get_args()
+        tokenizer = get_tokenizer()
+
+        train_dataset = Dataset('training', args.train_data,
+                                tokenizer, args.seq_length)
+        valid_dataset = Dataset('validation', args.valid_data,
+                                tokenizer, args.seq_length)
+
+        return train_dataset, valid_dataset
+
+    def model_provider(pre_process=True, post_process=True):
+        """Build the model."""
+        args = get_args()
+
+        print_rank_0('building classification model for {} ...'.format(
+            args.task))
+        model = Classification(num_classes=num_classes, num_tokentypes=2,
+                               pre_process=pre_process, post_process=post_process)
+
+        return model
+
+    def metrics_func_provider():
+        """Privde metrics callback function."""
+        def single_dataset_provider(datapath):
+            args = get_args()
+            tokenizer = get_tokenizer()
+
+            name = name_from_datapath_func(datapath)
+            return Dataset(name, [datapath], tokenizer, args.seq_length)
+        return accuracy_func_provider(single_dataset_provider)
+
+    """Finetune/evaluate."""
+    finetune(train_valid_datasets_provider, model_provider,
+             end_of_epoch_callback_provider=metrics_func_provider)
+
+
+def main():
+    args = get_args()
+
+    if args.task == 'MNLI':
+
+        num_classes = 3
+        from tasks.glue.mnli import MNLIDataset as Dataset
+
+        def name_from_datapath(datapath):
+            return datapath.split('MNLI')[-1].strip(
+                '.tsv').strip('/').replace('_', '-')
+
+    elif args.task == 'QQP':
+
+        num_classes = 2
+        from tasks.glue.qqp import QQPDataset as Dataset
+
+        def name_from_datapath(datapath):
+            return datapath.split('QQP')[-1].strip(
+                '.tsv').strip('/').replace('_', '-')
+
+    else:
+        raise NotImplementedError('GLUE task {} is not implemented.'.format(
+            args.task))
+
+    glue_classification(num_classes, Dataset, name_from_datapath)

tasks/glue/mnli.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.
+
+"""MNLI dataset."""
+
+from megatron import print_rank_0
+from tasks.data_utils import clean_text
+from .data import GLUEAbstractDataset
+
+
+LABELS = {'contradiction': 0, 'entailment': 1, 'neutral': 2}
+
+
+class MNLIDataset(GLUEAbstractDataset):
+
+    def __init__(self, name, datapaths, tokenizer, max_seq_length,
+                 test_label='contradiction'):
+        self.test_label = test_label
+        super().__init__('MNLI', name, datapaths,
+                         tokenizer, max_seq_length)
+
+    def process_samples_from_single_path(self, filename):
+        """"Implement abstract method."""
+        print_rank_0(' > Processing {} ...'.format(filename))
+
+        samples = []
+        total = 0
+        first = True
+        is_test = False
+        with open(filename, 'r') as f:
+            for line in f:
+                row = line.strip().split('\t')
+                if first:
+                    first = False
+                    if len(row) == 10:
+                        is_test = True
+                        print_rank_0(
+                            '   reading {}, {} and {} columns and setting '
+                            'labels to {}'.format(
+                                row[0].strip(), row[8].strip(),
+                                row[9].strip(), self.test_label))
+                    else:
+                        print_rank_0('    reading {} , {}, {}, and {} columns '
+                                     '...'.format(
+                                         row[0].strip(), row[8].strip(),
+                                         row[9].strip(), row[-1].strip()))
+                    continue
+
+                text_a = clean_text(row[8].strip())
+                text_b = clean_text(row[9].strip())
+                unique_id = int(row[0].strip())
+                label = row[-1].strip()
+                if is_test:
+                    label = self.test_label
+
+                assert len(text_a) > 0
+                assert len(text_b) > 0
+                assert label in LABELS
+                assert unique_id >= 0
+
+                sample = {'text_a': text_a,
+                          'text_b': text_b,
+                          'label': LABELS[label],
+                          'uid': unique_id}
+                total += 1
+                samples.append(sample)
+
+                if total % 50000 == 0:
+                    print_rank_0('  > processed {} so far ...'.format(total))
+
+        print_rank_0(' >> processed {} samples.'.format(len(samples)))
+        return samples

tasks/glue/qqp.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.
+
+"""QQP dataset."""
+
+from megatron import print_rank_0
+from tasks.data_utils import clean_text
+from .data import GLUEAbstractDataset
+
+
+LABELS = [0, 1]
+
+
+class QQPDataset(GLUEAbstractDataset):
+
+    def __init__(self, name, datapaths, tokenizer, max_seq_length,
+                 test_label=0):
+        self.test_label = test_label
+        super().__init__('QQP', name, datapaths,
+                         tokenizer, max_seq_length)
+
+    def process_samples_from_single_path(self, filename):
+        """"Implement abstract method."""
+        print_rank_0(' > Processing {} ...'.format(filename))
+
+        samples = []
+        total = 0
+        first = True
+        is_test = False
+        with open(filename, 'r') as f:
+            for line in f:
+                row = line.strip().split('\t')
+                if first:
+                    first = False
+                    if len(row) == 3:
+                        is_test = True
+                        print_rank_0('   reading {}, {}, and {} columns and '
+                                     'setting labels to {}'.format(
+                                         row[0].strip(), row[1].strip(),
+                                         row[2].strip(), self.test_label))
+                    else:
+                        assert len(row) == 6
+                        print_rank_0('    reading {}, {}, {}, and {} columns'
+                                     ' ...'.format(
+                                         row[0].strip(), row[3].strip(),
+                                         row[4].strip(), row[5].strip()))
+                    continue
+
+                if is_test:
+                    assert len(row) == 3, 'expected length 3: {}'.format(row)
+                    uid = int(row[0].strip())
+                    text_a = clean_text(row[1].strip())
+                    text_b = clean_text(row[2].strip())
+                    label = self.test_label
+                    assert len(text_a) > 0
+                    assert len(text_b) > 0
+                else:
+                    if len(row) == 6:
+                        uid = int(row[0].strip())
+                        text_a = clean_text(row[3].strip())
+                        text_b = clean_text(row[4].strip())
+                        label = int(row[5].strip())
+                    else:
+                        print_rank_0('***WARNING*** index error, '
+                                     'skipping: {}'.format(row))
+                        continue
+                    if len(text_a) == 0:
+                        print_rank_0('***WARNING*** zero length a, '
+                                     'skipping: {}'.format(row))
+                        continue
+                    if len(text_b) == 0:
+                        print_rank_0('***WARNING*** zero length b, '
+                                     'skipping: {}'.format(row))
+                        continue
+                assert label in LABELS
+                assert uid >= 0
+
+                sample = {'uid': uid,
+                          'text_a': text_a,
+                          'text_b': text_b,
+                          'label': label}
+                total += 1
+                samples.append(sample)
+
+                if total % 50000 == 0:
+                    print_rank_0('  > processed {} so far ...'.format(total))
+
+        print_rank_0(' >> processed {} samples.'.format(len(samples)))
+        return samples

tasks/main.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.
+
+"""Main tasks functionality."""
+
+import os
+import sys
+sys.path.append(os.path.abspath(os.path.join(os.path.dirname(__file__),
+                                             os.path.pardir)))
+
+from megatron import get_args
+from megatron.initialize import initialize_megatron
+
+
+def get_tasks_args(parser):
+    """Provide extra arguments required for tasks."""
+    group = parser.add_argument_group(title='tasks')
+
+    group.add_argument('--task', type=str, required=True,
+                       help='Task name.')
+    group.add_argument('--epochs', type=int, default=None,
+                       help='Number of finetunning epochs. Zero results in '
+                       'evaluation only.')
+    group.add_argument('--pretrained-checkpoint', type=str, default=None,
+                       help='Pretrained checkpoint used for finetunning.')
+    group.add_argument('--keep-last', action='store_true',
+                       help='Keep the last batch (maybe incomplete) in'
+                       'the data loader')
+    group.add_argument('--train-data', nargs='+', default=None,
+                       help='Whitespace separated paths or corpora names '
+                       'for training.')
+    group.add_argument('--valid-data', nargs='*', default=None,
+                       help='path(s) to the validation data.')
+    group.add_argument('--overlapping-eval', type=int, default=32,
+                       help='Sliding window for overlapping evaluation.')
+    group.add_argument('--strict-lambada', action='store_true',
+                       help='Use more difficult formulation of lambada.')
+    # Retriever args
+    group.add_argument('--qa-data-dev', type=str, default=None,
+                       help='Path to the QA dataset dev file.')
+    group.add_argument('--qa-data-test', type=str, default=None,
+                       help='Path to the QA dataset test file.')
+
+    # Faiss arguments for retriever
+    group.add_argument('--faiss-use-gpu', action='store_true',
+                       help='Whether create the FaissMIPSIndex on GPU')
+    group.add_argument('--faiss-match', type=str, default='string', \
+                        choices=['regex', 'string'], help="Answer matching '\
+                        'logic type")
+    group.add_argument('--faiss-topk-retrievals', type=int, default=100,
+                       help='Number of blocks to use as top-k during retrieval')
+
+    return parser
+
+
+if __name__ == '__main__':
+
+    initialize_megatron(extra_args_provider=get_tasks_args)
+
+    args = get_args()
+
+    if args.num_layers_per_virtual_pipeline_stage is not None:
+        print("Interleaved pipeline schedule is not yet supported for downstream tasks.")
+        exit()
+
+    if args.task == 'RACE':
+        from race.finetune import main
+    elif args.task in ['MNLI', 'QQP']:
+        from glue.finetune import main
+    elif args.task in ['LAMBADA', 'WIKITEXT103']:
+        from zeroshot_gpt.evaluate import main
+    elif args.task in ['ICT-ZEROSHOT-NQ']:
+        from orqa.evaluate_orqa import main
+    else:
+        raise NotImplementedError('Task {} is not implemented.'.format(
+            args.task))
+
+    main()

tasks/orqa/evaluate_orqa.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.
+
+"""Main tasks functionality."""
+
+import os
+import sys
+
+from megatron import get_args
+from tasks.orqa.evaluate_utils import ORQAEvaluator
+
+def main():
+    """
+    Main program
+    """
+
+    args = get_args()
+
+    # Set up the model and evaluator
+    evaluator = ORQAEvaluator()
+
+    # Run evaluation
+    if args.qa_data_dev is not None:
+        evaluator.evaluate(args.qa_data_dev, "DEV")
+
+    if args.qa_data_test is not None:
+        evaluator.evaluate(args.qa_data_test, "TEST")
+

tasks/orqa/evaluate_utils.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.
+
+import torch
+
+from megatron import get_args, print_rank_0
+from megatron.checkpointing import load_biencoder_checkpoint
+from megatron.data.orqa_wiki_dataset import get_open_retrieval_wiki_dataset
+from tasks.orqa.natural_questions.nq import get_nq_dataset
+from tasks.orqa.natural_questions.nq import get_one_epoch_nq_dataloader
+from tasks.orqa.natural_questions.nq import process_nq_batch
+from tasks.orqa.natural_questions.qa_utils import calculate_matches
+from megatron.data.realm_index import OpenRetreivalDataStore, FaissMIPSIndex
+from megatron.model.biencoder_model import biencoder_model_provider
+from megatron.training import get_model
+
+class ORQAEvaluator(object):
+    def __init__(self):
+        args = get_args()
+        self.embedding_size = args.hidden_size
+        self.faiss_use_gpu = args.faiss_use_gpu
+        self.evidence_embedder_obj = None
+        self.evidence_dataset = None
+        self.mips_index = None
+        self.eval_dataset = None
+
+        # Get Evidence (Wikipedia) dataset
+        self.get_evidence_dataset()
+
+        # Load query encoder checkpoint
+        only_query_model = True
+        if args.biencoder_shared_query_context_model:
+            only_query_model = False
+
+        model = get_model(lambda: biencoder_model_provider(only_query_model=\
+            only_query_model, biencoder_shared_query_context_model=\
+            args.biencoder_shared_query_context_model))
+
+        self.model = load_biencoder_checkpoint(model,
+                only_query_model=only_query_model)
+
+        assert len(self.model) == 1
+        self.model[0].eval()
+
+        # Load faiss indexer
+        self.faiss_wrapper()
+
+    def get_evidence_embedding(self):
+        # This will load the embedding from the embedding path
+        self.evidence_embedder_obj = OpenRetreivalDataStore(load_from_path=True)
+
+    def get_evidence_dataset(self):
+        self.evidence_dataset = get_open_retrieval_wiki_dataset()
+
+    def faiss_wrapper(self):
+        # Initialize FAISS wrapper on local rank = 0 as the evidence embeddings
+        # is distributed over all the GPUs in a node and FAISS is not 
+        # thread-safe
+        args = get_args()
+        if args.local_rank == 0:
+            # Get evidence embeddings computed using context encoder
+            self.get_evidence_embedding()
+
+            assert self.evidence_embedder_obj is not None
+            self.mips_index = FaissMIPSIndex(embed_size=self.embedding_size,
+                                        embed_data=self.evidence_embedder_obj,
+                                        use_gpu=self.faiss_use_gpu)
+
+        # Wait for the FAISS index to be initialized in all the nodes
+        torch.distributed.barrier()
+
+    def generate_query_vectors(self, qa_data, split):
+
+        self.eval_dataset = get_nq_dataset(qa_data, split)
+        dataloader = get_one_epoch_nq_dataloader(self.eval_dataset)
+
+        query_vectors = []
+        reference_list = []
+
+        for batch in dataloader:
+            # batch also has query_tokens and query_pad_data
+            query_tokens, query_mask, query_types, \
+                query_len, reference = process_nq_batch(batch)
+
+            assert len(self.model) == 1
+            unwrapped_model = self.model[0]
+            while not hasattr(unwrapped_model, 'embed_text'):
+                unwrapped_model = unwrapped_model.module
+
+            with torch.no_grad():
+                query_logits = unwrapped_model.embed_text(
+                    unwrapped_model.query_model, query_tokens, 
+                    query_mask, query_types)
+
+            reference_list.extend(reference)
+            query_vectors.extend(query_logits.split(1, dim=0))
+            if len(query_vectors) % 100 == 0:
+                print_rank_0('Encoded queries {}'.format(len(query_vectors)))
+
+        query_tensor = torch.cat(query_vectors, dim=0)
+        print_rank_0('Total encoded queries tensor {}'.format(query_tensor.size()))
+
+        assert query_tensor.size(0) == len(self.eval_dataset)
+        return query_tensor, reference_list
+
+    def evaluate(self, qa_data, split):
+        args = get_args()
+        query_tensor, reference_list = self.generate_query_vectors(qa_data, \
+                                                                    split)
+        local_rank = args.local_rank
+        rank = torch.distributed.get_rank()
+        device_count = torch.cuda.device_count()
+        num_nodes = torch.distributed.get_world_size() // device_count
+        node_id = rank // device_count
+
+        for node in range(num_nodes):
+            start_rank = node * device_count
+            end_rank = (node + 1) * device_count
+            ranks_list = list(range(start_rank, end_rank))
+            node_group = torch.distributed.new_group(ranks=ranks_list)
+
+            if node_id == node:
+                device_start_rank = start_rank
+                group = node_group
+        
+        input_ = torch.empty_like(query_tensor).copy_(query_tensor).detach_()
+        tensor_list = [torch.empty_like(input_) for _ in range(device_count)]
+        torch.distributed.all_gather(tensor_list, query_tensor, group=group)
+
+        if local_rank == 0 and self.mips_index is not None:
+            all_query_tensor = torch.cat(tensor_list, dim=0).contiguous()
+
+            distance, topkindex = self.mips_index.search_mips_index(
+                all_query_tensor, top_k=args.faiss_topk_retrievals, 
+                reconstruct=False)
+            distance = torch.from_numpy(distance).cuda()
+            topkindex = torch.LongTensor(topkindex).cuda()
+
+        if local_rank != 0:
+            distance = torch.empty(device_count * len(query_tensor), \
+                args.faiss_topk_retrievals, dtype=torch.float32).cuda()
+            topkindex = torch.empty(device_count * len(query_tensor), \
+                args.faiss_topk_retrievals, dtype=torch.int64).cuda()
+
+        torch.distributed.broadcast(distance, src=device_start_rank, \
+            group=group)
+        torch.distributed.broadcast(topkindex, src=device_start_rank, \
+            group=group)
+
+        distance = torch.split(distance, len(query_tensor), dim=0)\
+            [local_rank]
+        topkindex = torch.split(topkindex, len(query_tensor), dim=0)\
+            [local_rank]
+
+        top_ids_and_scores = []
+        for darray, topkarray in zip(distance, topkindex):
+            top_ids_and_scores.append((topkarray.tolist(), darray.tolist()))
+
+        passages = self.evidence_dataset.id2text
+        match_stats = calculate_matches(passages,
+                                        reference_list,
+                                        top_ids_and_scores,
+                                        workers_num=args.num_workers,
+                                        match_type=args.faiss_match)
+        top_k_hits = match_stats.top_k_hits
+
+        print_rank_0("{} SET RESULTS".format(split))
+        print_rank_0("topk-{} documents hits {}".format(
+            args.faiss_topk_retrievals, top_k_hits))
+        top_k_hits = [v / len(top_ids_and_scores) for v in top_k_hits]
+        print_rank_0("top-k documents hits accuracy {}".format(top_k_hits))
+
+        for i in args.retriever_report_topk_accuracies:
+            print_rank_0("top-{}: {:.2f}".format(i, top_k_hits[i-1] * 100))
+
+        return

tasks/orqa/natural_questions/nq.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.
+
+"""
+ Data Loader for Google NQ dataset
+"""
+
+from abc import ABC
+import csv
+from collections import OrderedDict
+import numpy as np
+
+import torch
+from torch.utils.data import DataLoader
+from torch.utils.data import Dataset, BatchSampler
+
+from megatron import print_rank_0, get_args, get_tokenizer, mpu
+from megatron.data.biencoder_dataset_utils import make_attention_mask
+
+def get_nq_dataset(qa_data, split):
+    args = get_args()
+    tokenizer = get_tokenizer()
+
+    dataset = NQDataset('Google NQ {} Split'.format(split),
+                        'Google Natural Questions',
+                        qa_data,
+                        tokenizer,
+                        args.retriever_seq_length)
+    return dataset
+
+
+def process_nq_batch(batch):
+    query_tokens = batch['token_ids'].long().cuda()
+    query_mask = (batch['token_mask'] < 0.5).cuda()
+    query_types = batch['token_types'].long().cuda()
+    query_len = batch['seq_len'].long().cuda()
+    reference = batch['reference']
+
+    return query_tokens, query_mask, query_types, query_len, reference
+
+
+class CustomDataLoader(DataLoader):
+    def __init__(self, dataset, eval=False, **kwargs):
+        if kwargs.get('collate_fn', None) is None:
+            kwargs['collate_fn'] = self._collate_fn
+        self.eval = eval
+        super().__init__(dataset, **kwargs)
+
+    def _collate_fn(self, batch_data):
+        # generate batch
+        batch_size = len(batch_data)
+        tensorized = OrderedDict()
+        for d in batch_data:
+            for k, v in d.items():
+                tensorized.setdefault(k, []).append(v)
+        assert len(tensorized) == 5
+
+        tensorized['token_ids'] = torch.LongTensor(tensorized['token_ids'])
+        tensorized['token_mask'] = torch.LongTensor(tensorized['token_mask'])
+        tensorized['token_types'] = torch.LongTensor(tensorized['token_types'])
+        tensorized['seq_len'] = torch.LongTensor(tensorized['seq_len'])
+        return tensorized
+
+
+def get_one_epoch_nq_dataloader(dataset, micro_batch_size=None):
+    """Data loader. Note that batch-size is the local (per GPU) batch-size.
+       NOTE: This dataloader is not distributed !!!
+    """
+
+    args = get_args()
+    if micro_batch_size is None:
+        micro_batch_size = args.micro_batch_size
+    num_workers = args.num_workers
+
+    sampler = torch.utils.data.SequentialSampler(dataset)
+    # importantly, drop_last must be False to get all the data.
+    batch_sampler = BatchSampler(sampler,
+                                 batch_size=micro_batch_size,
+                                 drop_last=False)
+
+    # Data loader. Note that batch size is the per GPU batch size.
+    data_loader = CustomDataLoader(dataset,
+                                   batch_sampler=batch_sampler,
+                                   num_workers=num_workers,
+                                   pin_memory=True)
+    return data_loader
+
+
+def build_tokens_types_paddings_from_text(src_text, tokenizer, max_seq_length):
+    """Build token types and paddings, trim if needed, and pad if needed."""
+
+    src_text_ids = tokenizer.tokenize(src_text)
+
+    return build_tokens_types_paddings_from_ids(src_text_ids,
+                                                max_seq_length,
+                                                tokenizer.cls,
+                                                tokenizer.sep,
+                                                tokenizer.pad)
+
+
+def build_tokens_types_paddings_from_ids(src_ids, max_seq_length, cls_id, \
+    sep_id, pad_id):
+    """
+    Build token types and paddings, trim if needed, and pad if needed.
+
+    TODO: Design modular interface to reuse this function. This is getting
+    repeated multiple times in different tasks
+    """
+
+    enc_ids = []
+    tokentypes_enc = []
+
+    # [CLS].
+    enc_ids.append(cls_id)
+    tokentypes_enc.append(0)
+
+    # A.
+    len_src = len(src_ids)
+    enc_ids.extend(src_ids)
+    tokentypes_enc.extend([0] * len_src)
+
+    # Cap the size.
+    if len(enc_ids) > max_seq_length - 1:
+        enc_ids = enc_ids[0: max_seq_length - 1]
+        tokentypes_enc = tokentypes_enc[0: max_seq_length - 1]
+
+    # [SEP].
+    enc_ids.append(sep_id)
+    tokentypes_enc.append(0)
+
+    num_tokens_enc = len(enc_ids)
+    # Padding.
+    padding_length = max_seq_length - len(enc_ids)
+    if padding_length > 0:
+        enc_ids.extend([pad_id] * padding_length)
+        tokentypes_enc.extend([pad_id] * padding_length)
+
+    return enc_ids, tokentypes_enc, num_tokens_enc
+
+
+def build_sample(token_ids, token_types, num_tokens, reference):
+    """
+    Convert to numpy and return a sample consumed by the
+    batch producer.
+    """
+
+    token_ids = np.array(token_ids, dtype=np.int64)
+    token_types = np.array(token_types, dtype=np.int64)
+    token_mask = make_attention_mask(token_ids, token_ids)
+
+    sample = ({
+        'token_ids': token_ids,
+        'token_mask': token_mask,
+        'token_types': token_types,
+        'seq_len': num_tokens,
+        'reference': reference
+    })
+    return sample
+
+
+class NQDataset(ABC, Dataset):
+    """
+    Open Retrieval Question Answering evaluation using Google NQ dataset.
+    """
+
+    def __init__(self, task_name, dataset_name, datapath,
+                 tokenizer, max_seq_length):
+        # Store inputs.
+        self.task_name = task_name
+        self.dataset_name = dataset_name
+        self.tokenizer = tokenizer
+        self.max_seq_length = max_seq_length
+        print_rank_0(' > building {} dataset for {}:'.format(self.task_name,
+                                                             self.dataset_name))
+        print_rank_0(datapath)
+        self.samples = self.process_samples_from_single_path(datapath)
+        print_rank_0('  >> total number of samples: {}'.format(\
+                                                        len(self.samples)))
+
+    def __len__(self):
+        return len(self.samples)
+
+    def __getitem__(self, idx):
+        raw_sample = self.samples[idx]
+
+        ques_tokens, tokentypes_enc, num_tokens_ques = \
+            build_tokens_types_paddings_from_text(raw_sample['question'],
+                self.tokenizer, self.max_seq_length)
+
+        sample = build_sample(ques_tokens,
+                              tokentypes_enc,
+                              num_tokens_ques,
+                              raw_sample['answers'])
+        return sample
+
+    @staticmethod
+    def process_samples_from_single_path(filename):
+        print_rank_0(' > Processing {} ...'.format(filename))
+        samples = []
+        total = 0
+
+        with open(filename, 'r') as ifile:
+            reader = csv.reader(ifile, delimiter='\t')
+            for row in reader:
+                question = row[0]
+                answers = eval(row[1])
+
+                sample = {'question': question, 'answers': answers}
+                total += 1
+                samples.append(sample)
+
+                if total % 1000 == 0:
+                    print_rank_0('  > processed {} so far ...'.format(total))
+
+        print_rank_0(' >> processed {} samples.'.format(len(samples)))
+        return samples

tasks/orqa/natural_questions/qa_utils.py

+#!/usr/bin/env python3
+# Copyright (c) Facebook, Inc. and its affiliates.
+# All rights reserved.
+#
+
+# The following code has been taken from
+# https://github.com/facebookresearch/DPR, which is CC-BY-NC 4.0
+# licensed as of now. More details on the license can be found
+# at https://github.com/facebookresearch/DPR/blob/master/LICENSE
+
+"""
+ Set of utilities for Q&A results validation tasks - Retriver passage
+ validation and Reader predicted answer validation
+"""
+
+import collections
+import string
+import unicodedata
+from functools import partial
+from multiprocessing import Pool as ProcessPool
+from typing import Tuple, List, Dict
+
+import regex as re
+
+from megatron import logging
+from tasks.orqa.natural_questions.tokenizers import SimpleTokenizer
+
+logger = logging.get_logger(__name__)
+
+QAMatchStats = collections.namedtuple('QAMatchStats', ['top_k_hits',\
+                                        'questions_doc_hits'])
+
+def calculate_matches(all_docs: Dict[object, Tuple[str, str]], 
+    answers: List[List[str]], closest_docs: List[Tuple[List[object], 
+    List[float]]], workers_num: int, match_type: str) -> QAMatchStats:
+    """
+    Evaluates answers presence in the set of documents. This function is 
+    supposed to be used with a large collection of documents and results. 
+    It internally forks multiple sub-processes for evaluation and then 
+    merges results
+    :param all_docs: dictionary of the entire documents database. 
+        doc_id -> (doc_text, title)
+    :param answers: list of answers's list. One list per question
+    :param closest_docs: document ids of the top results along with their
+        scores
+    :param workers_num: amount of parallel threads to process data
+    :param match_type: type of answer matching. Refer to has_answer code for
+        available options
+    :return: matching information tuple.
+    top_k_hits - a list where the index is the amount of top documents retrieved
+        and the value is the total amount of valid matches across an entire
+        dataset.
+    questions_doc_hits - more detailed info with answer matches for every
+        question and every retrieved document
+    """
+    global dpr_all_documents
+    dpr_all_documents = all_docs
+
+    tok_opts = {}
+    tokenizer = SimpleTokenizer(**tok_opts)
+
+    processes = ProcessPool(
+        processes=workers_num,
+    )
+
+    logger.info('Matching answers in top docs...')
+
+    get_score_partial = partial(check_answer, match_type=match_type,
+                                    tokenizer=tokenizer)
+
+    questions_answers_docs = zip(answers, closest_docs)
+
+    scores = processes.map(get_score_partial, questions_answers_docs)
+
+    logger.info('Per question validation results len=%d', len(scores))
+
+    n_docs = len(closest_docs[0][0])
+    top_k_hits = [0] * n_docs
+    for question_hits in scores:
+        best_hit = next((i for i, x in enumerate(question_hits) if x), None)
+        if best_hit is not None:
+            top_k_hits[best_hit:] = [v + 1 for v in top_k_hits[best_hit:]]
+
+    return QAMatchStats(top_k_hits, scores)
+
+
+def check_answer(questions_answers_docs, tokenizer, match_type) -> List[bool]:
+    """
+    Search through all the top docs to see if they have any of the answers.
+    """
+    answers, (doc_ids, doc_scores) = questions_answers_docs
+
+    global dpr_all_documents
+    hits = []
+
+    for i, doc_id in enumerate(doc_ids):
+        doc = dpr_all_documents[doc_id]
+        text = doc[0]
+
+        answer_found = False
+        if text is None:  # cannot find the document for some reason
+            logger.warning("no doc in db")
+            hits.append(False)
+            continue
+
+        if has_answer(answers, text, tokenizer, match_type):
+            answer_found = True
+        hits.append(answer_found)
+    return hits
+
+
+def has_answer(answers, text, tokenizer, match_type) -> bool:
+    """
+    Check if a document contains an answer string.
+    If `match_type` is string, token matching is done between the text 
+        and answer.
+    If `match_type` is regex, we search the whole text with the regex.
+    """
+    text = _normalize(text)
+
+    if match_type == 'string':
+        # Answer is a list of possible strings
+        text = tokenizer.tokenize(text).words(uncased=True)
+
+        for single_answer in answers:
+            single_answer = _normalize(single_answer)
+            single_answer = tokenizer.tokenize(single_answer)
+            single_answer = single_answer.words(uncased=True)
+
+            for i in range(0, len(text) - len(single_answer) + 1):
+                if single_answer == text[i: i + len(single_answer)]:
+                    return True
+
+    elif match_type == 'regex':
+        # Answer is a regex
+        for single_answer in answers:
+            single_answer = _normalize(single_answer)
+            if regex_match(text, single_answer):
+                return True
+    return False
+
+
+def regex_match(text, pattern):
+    """Test if a regex pattern is contained within a text."""
+    try:
+        pattern = re.compile(
+            pattern,
+            flags=re.IGNORECASE + re.UNICODE + re.MULTILINE,
+        )
+    except BaseException:
+        return False
+    return pattern.search(text) is not None
+
+
+# function for the reader model answer validation
+def exact_match_score(prediction, ground_truth):
+    return _normalize_answer(prediction) == _normalize_answer(ground_truth)
+
+
+def _normalize_answer(s):
+    def remove_articles(text):
+        return re.sub(r'\b(a|an|the)\b', ' ', text)
+
+    def white_space_fix(text):
+        return ' '.join(text.split())
+
+    def remove_punc(text):
+        exclude = set(string.punctuation)
+        return ''.join(ch for ch in text if ch not in exclude)
+
+    def lower(text):
+        return text.lower()
+
+    return white_space_fix(remove_articles(remove_punc(lower(s))))
+
+
+def _normalize(text):
+    return unicodedata.normalize('NFD', text)

tasks/orqa/natural_questions/tokenizers.py

+#!/usr/bin/env python3
+# Copyright (c) Facebook, Inc. and its affiliates.
+# All rights reserved.
+#
+
+# The following code has been taken from
+# https://github.com/facebookresearch/DPR, which is CC-BY-NC 4.0
+# licensed as of now. More details on the license can be found
+# at https://github.com/facebookresearch/DPR/blob/master/LICENSE
+
+"""
+Most of the tokenizers code here is copied from DrQA codebase to avoid adding extra dependency
+"""
+
+import copy
+
+import regex
+import spacy
+
+from megatron import logging
+
+logger = logging.get_logger(__name__)
+
+
+class Tokens(object):
+    """A class to represent a list of tokenized text."""
+    TEXT = 0
+    TEXT_WS = 1
+    SPAN = 2
+    POS = 3
+    LEMMA = 4
+    NER = 5
+
+    def __init__(self, data, annotators, opts=None):
+        self.data = data
+        self.annotators = annotators
+        self.opts = opts or {}
+
+    def __len__(self):
+        """The number of tokens."""
+        return len(self.data)
+
+    def slice(self, i=None, j=None):
+        """Return a view of the list of tokens from [i, j)."""
+        new_tokens = copy.copy(self)
+        new_tokens.data = self.data[i: j]
+        return new_tokens
+
+    def untokenize(self):
+        """Returns the original text (with whitespace reinserted)."""
+        return ''.join([t[self.TEXT_WS] for t in self.data]).strip()
+
+    def words(self, uncased=False):
+        """Returns a list of the text of each token
+
+        Args:
+            uncased: lower cases text
+        """
+        if uncased:
+            return [t[self.TEXT].lower() for t in self.data]
+        else:
+            return [t[self.TEXT] for t in self.data]
+
+    def offsets(self):
+        """Returns a list of [start, end) character offsets of each token."""
+        return [t[self.SPAN] for t in self.data]
+
+    def pos(self):
+        """Returns a list of part-of-speech tags of each token.
+        Returns None if this annotation was not included.
+        """
+        if 'pos' not in self.annotators:
+            return None
+        return [t[self.POS] for t in self.data]
+
+    def lemmas(self):
+        """Returns a list of the lemmatized text of each token.
+        Returns None if this annotation was not included.
+        """
+        if 'lemma' not in self.annotators:
+            return None
+        return [t[self.LEMMA] for t in self.data]
+
+    def entities(self):
+        """Returns a list of named-entity-recognition tags of each token.
+        Returns None if this annotation was not included.
+        """
+        if 'ner' not in self.annotators:
+            return None
+        return [t[self.NER] for t in self.data]
+
+    def ngrams(self, n=1, uncased=False, filter_fn=None, as_strings=True):
+        """Returns a list of all ngrams from length 1 to n.
+
+        Args:
+            n: upper limit of ngram length
+            uncased: lower cases text
+            filter_fn: user function that takes in an ngram list and returns
+              True or False to keep or not keep the ngram
+            as_string: return the ngram as a string vs list
+        """
+
+        def _skip(gram):
+            if not filter_fn:
+                return False
+            return filter_fn(gram)
+
+        words = self.words(uncased)
+        ngrams = [(s, e + 1)
+                  for s in range(len(words))
+                  for e in range(s, min(s + n, len(words)))
+                  if not _skip(words[s:e + 1])]
+
+        # Concatenate into strings
+        if as_strings:
+            ngrams = ['{}'.format(' '.join(words[s:e])) for (s, e) in ngrams]
+
+        return ngrams
+
+    def entity_groups(self):
+        """Group consecutive entity tokens with the same NER tag."""
+        entities = self.entities()
+        if not entities:
+            return None
+        non_ent = self.opts.get('non_ent', 'O')
+        groups = []
+        idx = 0
+        while idx < len(entities):
+            ner_tag = entities[idx]
+            # Check for entity tag
+            if ner_tag != non_ent:
+                # Chomp the sequence
+                start = idx
+                while (idx < len(entities) and entities[idx] == ner_tag):
+                    idx += 1
+                groups.append((self.slice(start, idx).untokenize(), ner_tag))
+            else:
+                idx += 1
+        return groups
+
+
+class Tokenizer(object):
+    """Base tokenizer class.
+    Tokenizers implement tokenize, which should return a Tokens class.
+    """
+
+    def tokenize(self, text):
+        raise NotImplementedError
+
+    def shutdown(self):
+        pass
+
+    def __del__(self):
+        self.shutdown()
+
+
+class SimpleTokenizer(Tokenizer):
+    ALPHA_NUM = r'[\p{L}\p{N}\p{M}]+'
+    NON_WS = r'[^\p{Z}\p{C}]'
+
+    def __init__(self, **kwargs):
+        """
+        Args:
+            annotators: None or empty set (only tokenizes).
+        """
+        self._regexp = regex.compile(
+            '(%s)|(%s)' % (self.ALPHA_NUM, self.NON_WS),
+            flags=regex.IGNORECASE + regex.UNICODE + regex.MULTILINE
+        )
+        if len(kwargs.get('annotators', {})) > 0:
+            logger.warning('%s only tokenizes! Skipping annotators: %s' %
+                           (type(self).__name__, kwargs.get('annotators')))
+        self.annotators = set()
+
+    def tokenize(self, text):
+        data = []
+        matches = [m for m in self._regexp.finditer(text)]
+        for i in range(len(matches)):
+            # Get text
+            token = matches[i].group()
+
+            # Get whitespace
+            span = matches[i].span()
+            start_ws = span[0]
+            if i + 1 < len(matches):
+                end_ws = matches[i + 1].span()[0]
+            else:
+                end_ws = span[1]
+
+            # Format data
+            data.append((
+                token,
+                text[start_ws: end_ws],
+                span,
+            ))
+        return Tokens(data, self.annotators)
+
+
+class SpacyTokenizer(Tokenizer):
+
+    def __init__(self, **kwargs):
+        """
+        Args:
+            annotators: set that can include pos, lemma, and ner.
+            model: spaCy model to use (either path, or keyword like 'en').
+        """
+        model = kwargs.get('model', 'en')
+        self.annotators = copy.deepcopy(kwargs.get('annotators', set()))
+        nlp_kwargs = {'parser': False}
+        if not any([p in self.annotators for p in ['lemma', 'pos', 'ner']]):
+            nlp_kwargs['tagger'] = False
+        if 'ner' not in self.annotators:
+            nlp_kwargs['entity'] = False
+        self.nlp = spacy.load(model, **nlp_kwargs)
+
+    def tokenize(self, text):
+        # We don't treat new lines as tokens.
+        clean_text = text.replace('\n', ' ')
+        tokens = self.nlp.tokenizer(clean_text)
+        if any([p in self.annotators for p in ['lemma', 'pos', 'ner']]):
+            self.nlp.tagger(tokens)
+        if 'ner' in self.annotators:
+            self.nlp.entity(tokens)
+
+        data = []
+        for i in range(len(tokens)):
+            # Get whitespace
+            start_ws = tokens[i].idx
+            if i + 1 < len(tokens):
+                end_ws = tokens[i + 1].idx
+            else:
+                end_ws = tokens[i].idx + len(tokens[i].text)
+
+            data.append((
+                tokens[i].text,
+                text[start_ws: end_ws],
+                (tokens[i].idx, tokens[i].idx + len(tokens[i].text)),
+                tokens[i].tag_,
+                tokens[i].lemma_,
+                tokens[i].ent_type_,
+            ))
+
+        # Set special option for non-entity tag: '' vs 'O' in spaCy
+        return Tokens(data, self.annotators, opts={'non_ent': ''})

tasks/race/data.py

+
+import glob
+import json
+import os
+import time
+
+from torch.utils.data import Dataset
+
+from megatron import print_rank_0
+from tasks.data_utils import build_sample
+from tasks.data_utils import build_tokens_types_paddings_from_ids
+from tasks.data_utils import clean_text
+
+
+NUM_CHOICES = 4
+MAX_QA_LENGTH = 128
+
+
+class RaceDataset(Dataset):
+
+    def __init__(self, dataset_name, datapaths, tokenizer, max_seq_length,
+                 max_qa_length=MAX_QA_LENGTH):
+
+        self.dataset_name = dataset_name
+        print_rank_0(' > building RACE dataset for {}:'.format(
+            self.dataset_name))
+
+        string = '  > paths:'
+        for path in datapaths:
+            string += ' ' + path
+        print_rank_0(string)
+
+        self.samples = []
+        for datapath in datapaths:
+            self.samples.extend(process_single_datapath(datapath, tokenizer,
+                                                        max_qa_length,
+                                                        max_seq_length))
+
+        print_rank_0('  >> total number of samples: {}'.format(
+            len(self.samples)))
+
+        # This indicates that each "sample" has multiple samples that
+        # will collapse into batch dimension
+        self.sample_multiplier = NUM_CHOICES
+
+    def __len__(self):
+        return len(self.samples)
+
+    def __getitem__(self, idx):
+        return self.samples[idx]
+
+
+def process_single_datapath(datapath, tokenizer, max_qa_length, max_seq_length):
+    """Read in RACE files, combine, clean-up, tokenize, and convert to
+    samples."""
+
+    print_rank_0('   > working on {}'.format(datapath))
+    start_time = time.time()
+
+    # Get list of files.
+    filenames = glob.glob(os.path.join(datapath, '*.txt'))
+
+    samples = []
+    num_docs = 0
+    num_questions = 0
+    num_samples = 0
+    # Load all the files
+    for filename in filenames:
+        with open(filename, 'r') as f:
+            for line in f:
+                data = json.loads(line)
+                num_docs += 1
+
+                context = data["article"]
+                questions = data["questions"]
+                choices = data["options"]
+                answers = data["answers"]
+                # Check the length.
+                assert len(questions) == len(answers)
+                assert len(questions) == len(choices)
+
+                # Context: clean up and convert to ids.
+                context = clean_text(context)
+                context_ids = tokenizer.tokenize(context)
+
+                # Loop over questions.
+                for qi, question in enumerate(questions):
+                    num_questions += 1
+                    # Label.
+                    label = ord(answers[qi]) - ord("A")
+                    assert label >= 0
+                    assert label < NUM_CHOICES
+                    assert len(choices[qi]) == NUM_CHOICES
+
+                    # For each question, build num-choices samples.
+                    ids_list = []
+                    types_list = []
+                    paddings_list = []
+                    for ci in range(NUM_CHOICES):
+                        choice = choices[qi][ci]
+                        # Merge with choice.
+                        if "_" in question:
+                            qa = question.replace("_", choice)
+                        else:
+                            qa = " ".join([question, choice])
+                        # Clean QA.
+                        qa = clean_text(qa)
+                        # Tokenize.
+                        qa_ids = tokenizer.tokenize(qa)
+                        # Trim if needed.
+                        if len(qa_ids) > max_qa_length:
+                            qa_ids = qa_ids[0:max_qa_length]
+
+                        # Build the sample.
+                        ids, types, paddings \
+                            = build_tokens_types_paddings_from_ids(
+                                qa_ids, context_ids, max_seq_length,
+                                tokenizer.cls, tokenizer.sep, tokenizer.pad)
+
+                        ids_list.append(ids)
+                        types_list.append(types)
+                        paddings_list.append(paddings)
+
+                    # Convert to numpy and add to samples
+                    samples.append(build_sample(ids_list, types_list,
+                                                paddings_list, label,
+                                                num_samples))
+                    num_samples += 1
+
+    elapsed_time = time.time() - start_time
+    print_rank_0('    > processed {} document, {} questions, and {} samples'
+                 ' in {:.2f} seconds'.format(num_docs, num_questions,
+                                             num_samples, elapsed_time))
+
+    return samples

tasks/race/finetune.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.
+
+"""Race."""
+
+from megatron import get_args
+from megatron import print_rank_0
+from megatron import get_tokenizer
+from megatron import mpu
+from megatron.model.multiple_choice import MultipleChoice
+from tasks.eval_utils import accuracy_func_provider
+from tasks.finetune_utils import finetune
+from tasks.race.data import RaceDataset
+
+
+def train_valid_datasets_provider():
+    """Provide train and validation datasets."""
+    args = get_args()
+    tokenizer = get_tokenizer()
+
+    train_dataset = RaceDataset('training', args.train_data,
+                                tokenizer, args.seq_length)
+    valid_dataset = RaceDataset('validation', args.valid_data,
+                                tokenizer, args.seq_length)
+
+    return train_dataset, valid_dataset
+
+
+def model_provider(pre_process=True, post_process=True):
+    """Build the model."""
+
+    print_rank_0('building multichoice model for RACE ...')
+    model = MultipleChoice(num_tokentypes=2,
+                           pre_process=pre_process,
+                           post_process=post_process)
+
+    return model
+
+
+def metrics_func_provider():
+    """Privde metrics callback function."""
+    args = get_args()
+    tokenizer = get_tokenizer()
+
+    def single_dataset_provider(datapath):
+        name = datapath.split('RACE')[-1].strip('/').replace('/', '-')
+        return RaceDataset(name, [datapath], tokenizer, args.seq_length)
+
+    return accuracy_func_provider(single_dataset_provider)
+
+
+def main():
+
+    finetune(train_valid_datasets_provider, model_provider,
+             end_of_epoch_callback_provider=metrics_func_provider)