Initial commit

megatron/legacy/data/multimodal_dataset.py

+# Copyright (c) 2023, NVIDIA CORPORATION. All rights reserved.
+
+from PIL import Image, UnidentifiedImageError
+import numpy as np
+import io
+import torch
+
+try:
+    from torchvision.transforms import InterpolationMode
+    BICUBIC = InterpolationMode.BICUBIC
+except ImportError:
+    BICUBIC = Image.BICUBIC
+
+from torchvision.transforms import Compose, ToTensor, Normalize, ToPILImage, RandomResizedCrop, Resize
+
+def _convert_image_to_rgb(image):
+    return image.convert("RGB")
+
+def _transform(img_h, img_w):
+    return Compose([
+        ToPILImage(),
+        RandomResizedCrop((img_h, img_w), scale=(0.5, 1.0), interpolation=BICUBIC),
+        _convert_image_to_rgb,
+        ToTensor(),
+        Normalize((0.48145466, 0.4578275, 0.40821073), (0.26862954, 0.26130258, 0.27577711)),
+    ])
+
+class MultiModalDataset(torch.utils.data.Dataset):
+
+    def __init__(self, name, data_prefix, indexed_dataset,
+                 num_samples, seq_length, seed, img_h, img_w):
+
+        self.name = name
+        self.indexed_dataset = indexed_dataset
+        self.doc_idx = indexed_dataset.get_document_indices()
+        self.visual_transform = _transform(img_h, img_w)
+
+    def __len__(self):
+        return self.indexed_dataset.sequence_lengths.shape[0]
+
+    def __getitem__(self, idx):
+        text_sample, mode = self.indexed_dataset.get(self.doc_idx[idx])
+        assert mode == 0
+        img_sample, mode = self.indexed_dataset.get(self.doc_idx[idx]+1)
+        assert mode == 1
+        img_pad = img_sample[0].item()
+        xs = img_sample[1:].tobytes(order='C')
+        xs = xs[:len(xs)-img_pad]
+
+        img_sample = np.array(Image.open(io.BytesIO(xs)))
+        img_sample = self.visual_transform(img_sample).reshape(-1)
+
+        return {'text': np.array(text_sample, dtype=np.int64),
+                'img': np.array(img_sample, dtype=np.float32)}

megatron/legacy/data/orqa_wiki_dataset.py

+# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
+
+"""Wikipedia dataset from DPR code for ORQA."""
+
+from abc import ABC
+import csv
+import numpy as np
+import random
+import torch
+from torch.utils.data import Dataset
+
+from megatron.training import print_rank_0, get_args, get_tokenizer
+from megatron.core import tensor_parallel
+from megatron.legacy.data.biencoder_dataset_utils import make_attention_mask
+
+def get_open_retrieval_wiki_dataset():
+    args = get_args()
+    tokenizer = get_tokenizer()
+
+    dataset = OpenRetrievalEvidenceDataset('2018 Wikipedia from DPR codebase',
+                                           'evidence',
+                                           args.evidence_data_path,
+                                           tokenizer,
+                                           args.retriever_seq_length)
+    return dataset
+
+
+def get_open_retrieval_batch(data_iterator):
+    # Items and their type.
+    keys = ['row_id', 'context', 'context_mask', 'context_types', 
+        'context_pad_mask']
+    datatype = torch.int64
+
+    # Broadcast data.
+    data = None if data_iterator is None else next(data_iterator)
+    data_b = tensor_parallel.broadcast_data(keys, data, datatype)
+
+    # Unpack.
+    row_id = data_b['row_id'].long()
+    context = data_b['context'].long()
+
+    # TODO: make the context mask a binary one
+    context_mask = (data_b['context_mask'] < 0.5)
+
+    context_types = data_b['context_types'].long()
+    context_pad_mask = data_b['context_pad_mask'].long()
+
+    return row_id, context, context_mask, context_types, context_pad_mask
+
+
+def build_tokens_types_paddings_from_text(row, tokenizer, max_seq_length):
+    """Build token types and paddings, trim if needed, and pad if needed."""
+
+    title_ids = tokenizer.tokenize(row['title'])
+    context_ids = tokenizer.tokenize(row['text'])
+
+    # Appending the title of the context at front
+    extended_context_ids = title_ids + [tokenizer.sep_id] + context_ids
+
+    context_ids, context_types, context_pad_mask = \
+        build_tokens_types_paddings_from_ids(extended_context_ids, 
+            max_seq_length, tokenizer.cls, tokenizer.sep, tokenizer.pad)
+
+    return context_ids, context_types, context_pad_mask
+
+
+# noinspection DuplicatedCode
+def build_tokens_types_paddings_from_ids(text_ids, max_seq_length,
+                                         cls_id, sep_id, pad_id):
+    """Build token types and paddings, trim if needed, and pad if needed."""
+    enc_ids = []
+    tokentypes_enc = []
+
+    # [CLS].
+    enc_ids.append(cls_id)
+    tokentypes_enc.append(0)
+
+    # A.
+    len_src = len(text_ids)
+    enc_ids.extend(text_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)
+
+    pad_mask = ([1] * num_tokens_enc) + ([0] * padding_length)
+    pad_mask = np.array(pad_mask, dtype=np.int64)
+
+    return enc_ids, tokentypes_enc, pad_mask
+
+
+def build_sample(row_id, context_ids, context_types, context_pad_mask):
+    """Convert to numpy and return a sample consumed by the batch producer."""
+
+    context_ids = np.array(context_ids, dtype=np.int64)
+    context_types = np.array(context_types, dtype=np.int64)
+    context_mask = make_attention_mask(context_ids, context_ids)
+
+    sample = ({
+        'row_id': row_id,
+        'context': context_ids,
+        'context_mask': context_mask,
+        'context_types': context_types,
+        'context_pad_mask': context_pad_mask
+    })
+    return sample
+
+
+class OpenRetrievalEvidenceDataset(ABC, Dataset):
+    """Open Retrieval Evidence dataset class."""
+
+    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))
+        # Process the files.
+        print_rank_0(datapath)
+        self.samples, self.id2text = self.process_samples_from_single_path(
+                                        datapath)
+
+        args = get_args()
+        if args.sample_rate < 1:  # subsample
+            k = int(len(self.samples) * args.sample_rate)
+            self.samples = random.sample(self.samples, k)
+
+        print_rank_0('  >> total number of samples: {}'.format(
+            len(self.samples)))
+
+    def __len__(self):
+        return len(self.samples)
+
+    def __getitem__(self, idx):
+        row = self.samples[idx]
+
+        context_ids, context_types, context_pad_mask = \
+            build_tokens_types_paddings_from_text(row, self.tokenizer, 
+                self.max_seq_length)
+
+        sample = build_sample(row['doc_id'],
+                              context_ids,
+                              context_types,
+                              context_pad_mask)
+        return sample
+
+    @staticmethod
+    def process_samples_from_single_path(filename):
+        print_rank_0(' > Processing {} ...'.format(filename))
+        total = 0
+
+        rows = []
+        id2text = {}
+
+        with open(filename) as tsvfile:
+            reader = csv.reader(tsvfile, delimiter='\t')
+            next(reader, None)  # skip the headers
+            for row in reader:
+                # file format: doc_id, doc_text, title
+                doc_id = int(row[0])
+                text = row[1]
+                title = row[2]
+
+                rows.append({'doc_id': doc_id,
+                             'text': text,
+                             'title': title})
+
+                assert doc_id not in id2text
+                id2text[doc_id] = (text, title)
+
+                total += 1
+                if total % 100000 == 0:
+                    print_rank_0('  > processed {} rows so far ...'.format(
+                        total))
+
+        print_rank_0(' >> processed {} samples.'.format(len(rows)))
+        return rows, id2text

megatron/legacy/data/realm_dataset_utils.py

+import os
+import time
+
+import numpy as np
+import torch
+
+from megatron.training import print_rank_0
+from megatron.core import mpu, tensor_parallel
+from megatron.legacy.data.dataset_utils import create_masked_lm_predictions, pad_and_convert_to_numpy
+from megatron.training import get_args, get_tokenizer, print_rank_0
+
+
+def get_one_epoch_dataloader(dataset, micro_batch_size=None):
+    """Specifically one epoch to be used in an indexing job."""
+    args = get_args()
+
+    world_size = mpu.get_data_parallel_world_size()
+    rank = mpu.get_data_parallel_rank()
+    if micro_batch_size is None:
+        micro_batch_size = args.micro_batch_size
+    global_batch_size = micro_batch_size * world_size
+    num_workers = args.num_workers
+
+    sampler = torch.utils.data.SequentialSampler(dataset)
+    # importantly, drop_last must be False to get all the data.
+    assert False, 'DistributedBatchSampler deprecated, change the implementation'
+    from megatron.legacy.data.samplers import DistributedBatchSampler
+    batch_sampler = DistributedBatchSampler(sampler,
+                                            batch_size=global_batch_size,
+                                            drop_last=False,
+                                            rank=rank,
+                                            world_size=world_size)
+
+    return torch.utils.data.DataLoader(dataset,
+                                       batch_sampler=batch_sampler,
+                                       num_workers=num_workers,
+                                       pin_memory=True)
+
+
+def get_ict_batch(data_iterator):
+    # Items and their type.
+    keys = ['query_tokens', 'query_pad_mask',
+            'block_tokens', 'block_pad_mask', 'block_data']
+    datatype = torch.int64
+
+    # Broadcast data.
+    if data_iterator is None:
+        data = None
+    else:
+        data = next(data_iterator)
+    data_b = tensor_parallel.broadcast_data(keys, data, datatype)
+
+    # Unpack.
+    query_tokens = data_b['query_tokens'].long()
+    query_pad_mask = data_b['query_pad_mask'].long()
+    block_tokens = data_b['block_tokens'].long()
+    block_pad_mask = data_b['block_pad_mask'].long()
+    block_indices = data_b['block_data'].long()
+
+    return query_tokens, query_pad_mask,\
+           block_tokens, block_pad_mask, block_indices
+
+
+def join_str_list(str_list):
+    """Join a list of strings, handling spaces appropriately"""
+    result = ""
+    for s in str_list:
+        if s.startswith("##"):
+            result += s[2:]
+        else:
+            result += " " + s
+    return result
+
+
+class BlockSampleData(object):
+    """A struct for fully describing a fixed-size block of data as used in REALM
+
+    :param start_idx: for first sentence of the block
+    :param end_idx: for last sentence of the block (may be partially truncated in sample construction)
+    :param doc_idx: the index of the document from which the block comes in the original indexed dataset
+    :param block_idx: a unique integer identifier given to every block.
+    """
+    def __init__(self, start_idx, end_idx, doc_idx, block_idx):
+        self.start_idx = start_idx
+        self.end_idx = end_idx
+        self.doc_idx = doc_idx
+        self.block_idx = block_idx
+
+    def as_array(self):
+        return np.array([self.start_idx, self.end_idx, self.doc_idx, self.block_idx]).astype(np.int64)
+
+    def as_tuple(self):
+        return self.start_idx, self.end_idx, self.doc_idx, self.block_idx
+
+
+class BlockSamplesMapping(object):
+    def __init__(self, mapping_array):
+        # make sure that the array is compatible with BlockSampleData
+        assert mapping_array.shape[1] == 4
+        self.mapping_array = mapping_array
+
+    def __len__(self):
+        return self.mapping_array.shape[0]
+
+    def __getitem__(self, idx):
+        """Get the data associated with an indexed sample."""
+        sample_data = BlockSampleData(*self.mapping_array[idx])
+        return sample_data
+
+
+def get_block_samples_mapping(block_dataset, title_dataset, data_prefix, num_epochs,
+                              max_num_samples, max_seq_length, seed, name, use_one_sent_docs=False):
+    """Get samples mapping for a dataset over fixed size blocks. This function also requires
+    a dataset of the titles for the source documents since their lengths must be taken into account.
+
+    :return: samples_mapping (BlockSamplesMapping)
+    """
+
+    if not num_epochs:
+        if not max_num_samples:
+            raise ValueError("Need to specify either max_num_samples "
+                             "or num_epochs")
+        num_epochs = np.iinfo(np.int32).max - 1
+    if not max_num_samples:
+        max_num_samples = np.iinfo(np.int64).max - 1
+
+    # Filename of the index mapping
+    indexmap_filename = data_prefix
+    indexmap_filename += '_{}_indexmap'.format(name)
+    if num_epochs != (np.iinfo(np.int32).max - 1):
+        indexmap_filename += '_{}ep'.format(num_epochs)
+    if max_num_samples != (np.iinfo(np.int64).max - 1):
+        indexmap_filename += '_{}mns'.format(max_num_samples)
+    indexmap_filename += '_{}msl'.format(max_seq_length)
+    indexmap_filename += '_{}s'.format(seed)
+    if use_one_sent_docs:
+        indexmap_filename += '_1sentok'
+    indexmap_filename += '.npy'
+
+    # Build the indexed mapping if not exist.
+    if mpu.get_data_parallel_rank() == 0 and \
+            not os.path.isfile(indexmap_filename):
+        print(' > WARNING: could not find index map file {}, building '
+              'the indices on rank 0 ...'.format(indexmap_filename))
+
+        # Make sure the types match the helpers input types.
+        assert block_dataset.document_indices.dtype == np.int64
+        assert block_dataset.sequence_lengths.dtype == np.int32
+
+        # Build samples mapping
+        verbose = torch.distributed.get_rank() == 0
+        start_time = time.time()
+        print_rank_0(' > building samples index mapping for {} ...'.format(
+            name))
+
+        from megatron.core.datasets import helpers
+        mapping_array = helpers.build_blocks_mapping(
+            block_dataset.document_indices,
+            block_dataset.sequence_lengths,
+            title_dataset.sequence_lengths,
+            num_epochs,
+            max_num_samples,
+            max_seq_length - 3,  # account for added tokens
+            seed,
+            verbose,
+            use_one_sent_docs)
+
+
+        print_rank_0(' > done building samples index mapping')
+        np.save(indexmap_filename, mapping_array, allow_pickle=True)
+        print_rank_0(' > saved the index mapping in {}'.format(
+            indexmap_filename))
+        # Make sure all the ranks have built the mapping
+        print_rank_0(' > elapsed time to build and save samples mapping '
+                     '(seconds): {:4f}'.format(
+            time.time() - start_time))
+
+    # This should be a barrier but nccl barrier assumes
+    # device_index=rank which is not the case for model
+    # parallel case
+    counts = torch.tensor([1], dtype=torch.long, device='cuda')
+    torch.distributed.all_reduce(counts, group=mpu.get_data_parallel_group())
+    assert counts[0].item() == torch.distributed.get_world_size(
+        group=mpu.get_data_parallel_group())
+
+    # Load indexed dataset.
+    print_rank_0(' > loading indexed mapping from {}'.format(
+        indexmap_filename))
+    start_time = time.time()
+
+    mapping_array = np.load(indexmap_filename, allow_pickle=True, mmap_mode='r')
+    samples_mapping = BlockSamplesMapping(mapping_array)
+
+    print_rank_0('    loaded indexed file in {:3.3f} seconds'.format(
+        time.time() - start_time))
+    print_rank_0('    total number of samples: {}'.format(
+        mapping_array.shape[0]))
+
+    return samples_mapping

megatron/legacy/data/realm_index.py

+import itertools
+import os
+import pickle
+import shutil
+
+import numpy as np
+import torch
+
+from megatron.training import get_args
+from megatron.core import mpu
+
+
+def detach(tensor):
+    return tensor.detach().cpu().numpy()
+
+
+class OpenRetreivalDataStore(object):
+    """
+    Serializable data structure for holding data for blocks --
+    embeddings and necessary metadata for Retriever
+    """
+    def __init__(self, embedding_path=None, load_from_path=True, rank=None):
+        self.embed_data = dict()
+        if embedding_path is None:
+            args = get_args()
+            embedding_path = args.embedding_path
+            rank = args.rank
+        self.embedding_path = embedding_path
+        self.rank = rank
+
+        if load_from_path:
+            self.load_from_file()
+
+        block_data_name = os.path.splitext(self.embedding_path)[0]
+        self.temp_dir_name = block_data_name + '_tmp'
+
+    def state(self):
+        return {
+            'embed_data': self.embed_data,
+        }
+
+    def clear(self):
+        """
+        Clear the embedding data structures to save memory.
+        The metadata ends up getting used, and is also much smaller in
+        dimensionality so it isn't really worth clearing.
+        """
+        self.embed_data = dict()
+
+    def load_from_file(self):
+        """Populate members from instance saved to file"""
+
+        if not mpu.model_parallel_is_initialized() or mpu.get_data_parallel_rank() == 0:
+            print("\n> Unpickling BlockData", flush=True)
+        state_dict = pickle.load(open(self.embedding_path, 'rb'))
+        if not mpu.model_parallel_is_initialized() or mpu.get_data_parallel_rank() == 0:
+            print(">> Finished unpickling BlockData\n", flush=True)
+
+        self.embed_data = state_dict['embed_data']
+
+    def add_block_data(self, row_id, block_embeds, allow_overwrite=False):
+        """
+        Add data for set of blocks
+        :param row_id: 1D array of unique int ids for the blocks
+        :param block_embeds: 2D array of embeddings of the blocks
+            In the case of retriever this will be [start_idx, end_idx, doc_idx]
+        """
+        for idx, embed in zip(row_id, block_embeds):
+            if not allow_overwrite and idx in self.embed_data:
+                raise ValueError("Unexpectedly tried to overwrite block data")
+
+            self.embed_data[idx] = np.float16(embed)
+
+    def save_shard(self):
+        """
+        Save the block data that was created this in this process
+        """
+        if not os.path.isdir(self.temp_dir_name):
+            os.makedirs(self.temp_dir_name, exist_ok=True)
+
+        # save the data for each shard
+        with open('{}/{}.pkl'.format(self.temp_dir_name, self.rank), 'wb') \
+            as writer:
+            pickle.dump(self.state(), writer)
+
+    def merge_shards_and_save(self):
+        #Combine all the shards made using save_shard
+        shard_names = os.listdir(self.temp_dir_name)
+        seen_own_shard = False
+
+        for fname in os.listdir(self.temp_dir_name):
+            shard_rank = int(os.path.splitext(fname)[0])
+            if shard_rank == self.rank:
+                seen_own_shard = True
+                continue
+
+            with open('{}/{}'.format(self.temp_dir_name, fname), 'rb') as f:
+                data = pickle.load(f)
+                old_size = len(self.embed_data)
+                shard_size = len(data['embed_data'])
+
+                # add the shard's data and check to make sure there
+                # is no overlap
+                self.embed_data.update(data['embed_data'])
+                assert len(self.embed_data) == old_size + shard_size
+
+        assert seen_own_shard
+
+        # save the consolidated shards and remove temporary directory
+        with open(self.embedding_path, 'wb') as final_file:
+            pickle.dump(self.state(), final_file)
+        shutil.rmtree(self.temp_dir_name, ignore_errors=True)
+
+        print("Finished merging {} shards for a total of {} embeds".format(
+            len(shard_names), len(self.embed_data)), flush=True)
+
+
+class FaissMIPSIndex(object):
+    """
+    Wrapper object for a BlockData which similarity search via FAISS under the hood
+    """
+    def __init__(self, embed_size, embed_data=None, use_gpu=False):
+        self.embed_size = embed_size
+        self.embed_data = embed_data
+        self.use_gpu = use_gpu
+
+        self.mips_index = None
+        self._set_mips_index()
+
+    def _set_mips_index(self):
+        """
+        Create a Faiss Flat index with inner product as the metric
+        to search against
+        """
+        try:
+            import faiss
+        except ImportError:
+            raise Exception("Error: Please install faiss to use FaissMIPSIndex")
+
+        if not mpu.model_parallel_is_initialized() or mpu.get_data_parallel_rank() == 0:
+            print("\n> Building index", flush=True)
+
+        cpu_index = faiss.IndexFlatIP(self.embed_size)
+
+        if self.use_gpu:
+            # create resources and config for GpuIndex
+            config = faiss.GpuMultipleClonerOptions()
+            config.shard = True
+            config.useFloat16 = True
+            gpu_index = faiss.index_cpu_to_all_gpus(cpu_index, co=config)
+            self.mips_index = faiss.IndexIDMap(gpu_index)
+            if not mpu.model_parallel_is_initialized() or mpu.get_data_parallel_rank() == 0:
+                print(">> Initialized index on GPU", flush=True)
+        else:
+            # CPU index supports IDs so wrap with IDMap
+            self.mips_index = faiss.IndexIDMap(cpu_index)
+            if not mpu.model_parallel_is_initialized() or mpu.get_data_parallel_rank() == 0:
+                print(">> Initialized index on CPU", flush=True)
+
+        # if we were constructed with a BlockData, then automatically load it
+        # when the FAISS structure is built
+        if self.embed_data is not None:
+            self.add_embed_data(self.embed_data)
+
+    def reset_index(self):
+        """Delete existing index and create a new"""
+        del self.mips_index
+
+        # reset the block data so that _set_block_index will reload it as well
+        if self.embed_data is not None:
+            embed_data_path = self.embed_data.embedding_path
+            del self.embed_data
+            self.embed_data = OpenRetreivalDataStore(embed_data_path)
+
+        self._set_mips_index()
+
+    def update_index(self):
+        """Delete existing index and create a new"""
+        del self.mips_index
+
+        # reset the block data so that _set_mips_index will reload it as well
+        if self.embed_data is not None:
+            self.embed_data.load_from_file()
+        self._set_mips_index()
+
+    def add_embed_data(self, all_embed_data):
+        """Add the embedding of each block to the underlying FAISS index"""
+
+        # this assumes the embed_data is a dict : {int: np.array<float>}
+        block_indices, block_embeds = zip(*all_embed_data.embed_data.items())
+
+        # the embeddings have to be entered in as float32 even though the math
+        # internally is done with float16.
+        embeds_arr = np.float32(np.array(block_embeds))
+        indices_arr = np.array(block_indices)
+
+        # we no longer need the embedding data since it's in the index now
+        all_embed_data.clear()
+
+        self.mips_index.add_with_ids(embeds_arr, indices_arr)
+
+        if not mpu.model_parallel_is_initialized() or mpu.get_data_parallel_rank() == 0:
+            print(">>> Finished adding block data to index", flush=True)
+
+    def search_mips_index(self, query_embeds, top_k, reconstruct=True):
+        """
+        Get the top-k blocks by the index distance metric.
+
+        :param reconstruct: if True: return a [num_queries x k x embed_dim]
+                                array of blocks
+                            if False: return [num_queries x k] array of
+                                distances, and another for indices
+        """
+        query_embeds = np.float32(detach(query_embeds))
+
+        if reconstruct:
+            # get the vectors themselves
+            top_k_block_embeds = self.mips_index.search_and_reconstruct(\
+                query_embeds, top_k)
+            return top_k_block_embeds
+        else:
+            # get distances and indices of closest vectors
+            distances, block_indices = self.mips_index.search(query_embeds, top_k)
+            return distances, block_indices

megatron/legacy/data/vit_dataset.py

+# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
+import os
+import random
+import numpy as np
+import torch
+import torchvision.transforms as T
+from torchvision import datasets
+from megatron.training import get_args
+from megatron.legacy.data.image_folder import ImageFolder
+from megatron.legacy.data.autoaugment import ImageNetPolicy
+from megatron.legacy.data.data_samplers import RandomSeedDataset
+from PIL import Image, ImageFilter, ImageOps
+
+
+class GaussianBlur(object):
+    """
+    Apply Gaussian Blur to the PIL image.
+    """
+    def __init__(self, p=0.5, radius_min=0.1, radius_max=2.):
+        self.prob = p
+        self.radius_min = radius_min
+        self.radius_max = radius_max
+
+    def __call__(self, img):
+        do_it = random.random() <= self.prob
+        if not do_it:
+            return img
+
+        return img.filter(
+            ImageFilter.GaussianBlur(
+                radius=random.uniform(self.radius_min, self.radius_max)
+            )
+        )
+
+
+class Solarization(object):
+    """
+    Apply Solarization to the PIL image.
+    """
+    def __init__(self, p):
+        self.p = p
+
+    def __call__(self, img):
+        if random.random() < self.p:
+            return ImageOps.solarize(img)
+        else:
+            return img
+
+
+class ClassificationTransform():
+    def __init__(self, image_size, train=True):
+        args = get_args()
+        assert args.fp16 or args.bf16
+        self.data_type = torch.half if args.fp16 else torch.bfloat16
+        if train:
+            self.transform = T.Compose([
+                T.RandomResizedCrop(image_size),
+                T.RandomHorizontalFlip(),
+                T.ColorJitter(0.4, 0.4, 0.4, 0.1),
+                ImageNetPolicy(),
+                T.ToTensor(),
+                T.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225)),
+                T.ConvertImageDtype(self.data_type)
+            ])
+        else:
+            self.transform = T.Compose([
+                T.Resize(image_size),
+                T.CenterCrop(image_size),
+                T.ToTensor(),
+                T.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225)),
+                T.ConvertImageDtype(self.data_type)
+            ])
+
+    def __call__(self, input):
+        output = self.transform(input)
+        return output
+
+
+class InpaintingTransform():
+    def __init__(self, image_size, train=True):
+
+        args = get_args()
+        self.mask_factor = args.mask_factor
+        self.mask_type = args.mask_type
+        self.image_size = image_size
+        self.patch_size = args.patch_dim
+        self.mask_size = int(self.mask_factor*(image_size[0]/self.patch_size)*(image_size[1]/self.patch_size))
+        self.train = train
+        assert args.fp16 or args.bf16
+        self.data_type = torch.half if args.fp16 else torch.bfloat16
+     
+        if self.train:
+            self.transform = T.Compose([
+                T.RandomResizedCrop(self.image_size),
+                T.RandomHorizontalFlip(),
+                T.ColorJitter(0.4, 0.4, 0.4, 0.1),
+                ImageNetPolicy(),
+                T.ToTensor(),
+                T.ConvertImageDtype(self.data_type)
+            ])
+        else:
+            self.transform = T.Compose([
+                T.Resize(self.image_size, interpolation=2),
+                T.CenterCrop(self.image_size),
+                T.ToTensor(),
+                T.ConvertImageDtype(self.data_type)
+            ])
+
+    def gen_mask(self, image_size, mask_size, mask_type, patch_size):
+        # output: mask as a list with indices for missing patches
+        action_list = [[0, 1], [0, -1], [1, 0], [-1, 0]]
+        assert image_size[0] == image_size[1]
+        img_size_patch = image_size[0] // patch_size
+
+        # drop masked patches
+        mask = torch.zeros((image_size[0], image_size[1]), dtype=torch.float)
+
+        if mask_type == 'random':
+            x = torch.randint(0, img_size_patch, ())
+            y = torch.randint(0, img_size_patch, ())
+            for i in range(mask_size):
+                r = torch.randint(0, len(action_list), ())
+                x = torch.clamp(x + action_list[r][0], min=0, max=img_size_patch - 1)
+                y = torch.clamp(y + action_list[r][1], min=0, max=img_size_patch - 1)
+                x_offset = x * patch_size
+                y_offset = y * patch_size
+                mask[x_offset:x_offset+patch_size, y_offset:y_offset+patch_size] = 1
+        else:
+            assert mask_type == 'row'
+            count = 0
+            for x in reversed(range(img_size_patch)):
+                for y in reversed(range(img_size_patch)):
+                    if (count < mask_size):
+                        count += 1
+                        x_offset = x * patch_size
+                        y_offset = y * patch_size
+                        mask[x_offset:x_offset+patch_size, y_offset:y_offset+patch_size] = 1
+        return mask
+
+    def __call__(self, input):
+        trans_input = self.transform(input)
+        mask = self.gen_mask(self.image_size, self.mask_size, 
+			     self.mask_type, self.patch_size)
+        mask = mask.unsqueeze(dim=0)
+        return trans_input, mask
+
+
+class DinoTransform(object):
+    def __init__(self, image_size, train=True):
+        args = get_args()
+        self.data_type = torch.half if args.fp16 else torch.bfloat16
+
+        flip_and_color_jitter = T.Compose([
+            T.RandomHorizontalFlip(p=0.5),
+            T.RandomApply(
+                [T.ColorJitter(brightness=0.4, contrast=0.4,
+			       saturation=0.2, hue=0.1)],
+                p=0.8
+            ),
+            T.RandomGrayscale(p=0.2),
+        ])
+
+        if args.fp16 or args.bf16:
+            normalize = T.Compose([
+                T.ToTensor(),
+                T.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225)),
+                T.ConvertImageDtype(self.data_type)
+            ])
+        else:
+            normalize = T.Compose([
+                T.ToTensor(),
+                T.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225)),
+            ])
+
+        # first global crop
+        scale_const = 0.4
+        self.global_transform1 = T.Compose([
+            T.RandomResizedCrop(image_size,
+                                scale=(scale_const, 1),
+                                interpolation=Image.BICUBIC),
+            flip_and_color_jitter,
+            GaussianBlur(1.0),
+            normalize
+        ])
+        # second global crop
+        self.global_transform2 = T.Compose([
+            T.RandomResizedCrop(image_size,
+                                scale=(scale_const, 1),
+                                interpolation=Image.BICUBIC),
+            flip_and_color_jitter,
+            GaussianBlur(0.1),
+            Solarization(0.2),
+            normalize
+        ])
+        # transformation for the local small crops
+        self.local_crops_number = args.dino_local_crops_number
+        self.local_transform = T.Compose([
+            T.RandomResizedCrop(args.dino_local_img_size,
+                                scale=(0.05, scale_const),
+                                interpolation=Image.BICUBIC),
+            flip_and_color_jitter,
+            GaussianBlur(p=0.5),
+            normalize
+        ])
+
+    def __call__(self, image):
+        crops = []
+        crops.append(self.global_transform1(image))
+        crops.append(self.global_transform2(image))
+        for _ in range(self.local_crops_number):
+            crops.append(self.local_transform(image))
+        return crops
+
+
+def build_train_valid_datasets(data_path, image_size=224):
+    args = get_args()
+
+    if args.vision_pretraining_type == 'classify':
+        train_transform = ClassificationTransform(image_size)
+        val_transform = ClassificationTransform(image_size, train=False)
+    elif args.vision_pretraining_type == 'inpaint':
+        train_transform = InpaintingTransform(image_size, train=False)
+        val_transform = InpaintingTransform(image_size, train=False)
+    elif args.vision_pretraining_type == 'dino':
+        train_transform = DinoTransform(image_size, train=True)
+        val_transform = ClassificationTransform(image_size, train=False)
+    else:
+        raise Exception('{} vit pretraining type is not supported.'.format(
+                args.vit_pretraining_type))
+
+    # training dataset
+    train_data_path = data_path[0] if len(data_path) <= 2 else data_path[2]
+    train_data = ImageFolder(
+        root=train_data_path,
+        transform=train_transform,
+        classes_fraction=args.classes_fraction,
+        data_per_class_fraction=args.data_per_class_fraction
+    )
+    train_data = RandomSeedDataset(train_data)
+
+    # validation dataset
+    val_data_path = data_path[1]
+    val_data = ImageFolder(
+        root=val_data_path,
+        transform=val_transform
+    )
+    val_data = RandomSeedDataset(val_data)
+
+    return train_data, val_data

megatron/legacy/fp16_deprecated/loss_scaler.py

+# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
+
+"""For backward compatibility, we need the class definitions to deserialize."""
+
+class LossScaler:
+    def __init__(self, scale=1):
+        self.cur_scale = scale
+
+class DynamicLossScaler:
+    def __init__(self,
+                 init_scale=2**32,
+                 scale_factor=2.,
+                 scale_window=1000,
+                 min_scale=1,
+                 delayed_shift=1,
+                 consecutive_hysteresis=False):
+        self.cur_scale = init_scale
+        self.cur_iter = 0
+        self.last_overflow_iter = -1
+        self.scale_factor = scale_factor
+        self.scale_window = scale_window
+        self.min_scale = min_scale
+        self.delayed_shift = delayed_shift
+        self.cur_hysteresis = delayed_shift
+        self.consecutive_hysteresis = consecutive_hysteresis
+

megatron/legacy/fused_kernels/__init__.py

+# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
+
+import os
+import pathlib
+import subprocess
+
+from torch.utils import cpp_extension
+
+# Setting this param to a list has a problem of generating different
+# compilation commands (with diferent order of architectures) and
+# leading to recompilation of fused kernels. Set it to empty string
+# to avoid recompilation and assign arch flags explicity in
+# extra_cuda_cflags below
+os.environ["TORCH_CUDA_ARCH_LIST"] = ""
+
+
+def load(args):
+
+    # Check if cuda 11 is installed for compute capability 8.0
+    cc_flag = []
+    _, bare_metal_major, bare_metal_minor = _get_cuda_bare_metal_version(
+        cpp_extension.CUDA_HOME
+    )
+    if int(bare_metal_major) >= 11:
+        cc_flag.append('-gencode')
+        cc_flag.append('arch=compute_80,code=sm_80')
+        if int(bare_metal_minor) >= 8:
+            cc_flag.append('-gencode')
+            cc_flag.append('arch=compute_90,code=sm_90')
+
+    # Build path
+    srcpath = pathlib.Path(__file__).parent.absolute()
+    buildpath = srcpath / "build"
+    _create_build_dir(buildpath)
+
+    # Helper function to build the kernels.
+    def _cpp_extention_load_helper(name, sources, extra_cuda_flags):
+        return cpp_extension.load(
+            name=name,
+            sources=sources,
+            build_directory=buildpath,
+            extra_cflags=[
+                "-O3",
+            ],
+            extra_cuda_cflags=[
+                "-O3",
+                "-gencode",
+                "arch=compute_70,code=sm_70",
+                "--use_fast_math",
+            ]
+            + extra_cuda_flags
+            + cc_flag,
+            verbose=(args.rank == 0),
+        )
+
+
+def _get_cuda_bare_metal_version(cuda_dir):
+    raw_output = subprocess.check_output(
+        [cuda_dir + "/bin/nvcc", "-V"], universal_newlines=True
+    )
+    output = raw_output.split()
+    release_idx = output.index("release") + 1
+    release = output[release_idx].split(".")
+    bare_metal_major = release[0]
+    bare_metal_minor = release[1][0]
+
+    return raw_output, bare_metal_major, bare_metal_minor
+
+
+def _create_build_dir(buildpath):
+    try:
+        os.mkdir(buildpath)
+    except OSError:
+        if not os.path.isdir(buildpath):
+            print(f"Creation of the build directory {buildpath} failed")

megatron/legacy/fused_kernels/compat.h

+/* Copyright (c) 2022, NVIDIA CORPORATION.  All rights reserved. */
+
+/*This code is copied fron NVIDIA apex:
+ *     https://github.com/NVIDIA/apex
+ *     with minor changes. */
+
+
+
+#ifndef TORCH_CHECK
+#define TORCH_CHECK AT_CHECK
+#endif
+
+#ifdef VERSION_GE_1_3
+#define DATA_PTR data_ptr
+#else
+#define DATA_PTR data
+#endif

megatron/legacy/fused_kernels/tests/test_fused_kernels.py

+import math
+
+import torch
+from torch.nn import LayerNorm
+
+from megatron.legacy.model.enums import AttnMaskType
+from megatron.legacy.model.fused_layer_norm import MixedFusedLayerNorm
+from megatron.legacy.model.fused_softmax import FusedScaleMaskSoftmax
+from megatron.legacy.model.utils import attention_mask_func
+from megatron.legacy.fused_kernels import load
+
+def test_load_fused_kernels():
+    try:
+        import fused_layer_norm_cuda
+        import scaled_masked_softmax_cuda
+        import scaled_upper_triang_masked_softmax_cuda
+        import torch
+
+        print("[Success] load_fused_kernels")
+    except ImportError as e:
+        print("[Fail] load_fused_kernels")
+        raise e
+
+def test_fused_softmax():
+    bert = BertModel.from_pretrained("bert-base-cased").cuda().half()
+    tokenizer = BertTokenizer.from_pretrained("bert-base-cased")
+    test_text = (
+        "Hello. How are you? I am fine thank you and you? yes Good. "
+        "hi hi hi hi hi hi hi hi hi hi hi hi hi"  # 32
+    )
+
+    tokens = tokenizer(
+        [test_text] * 4,
+        return_tensors="pt",
+    )
+
+    embedding_output = bert.embeddings(
+        input_ids=tokens["input_ids"].cuda(),
+        position_ids=None,
+        token_type_ids=tokens["token_type_ids"].cuda(),
+        inputs_embeds=None,
+        past_key_values_length=0,
+    )
+
+    # (bsz, 1, 1, seq_len)
+    mask = bert.get_extended_attention_mask(
+        attention_mask=tokens["attention_mask"].cuda(),
+        input_shape=tokens["input_ids"].shape,
+        device=bert.device,
+    )
+    # (bsz, 1, seq_len, seq_len)
+    mask = mask.repeat(1, 1, mask.size()[-1], 1)
+
+    attention = bert.encoder.layer[0].attention.self
+    key_layer = attention.transpose_for_scores(attention.key(embedding_output))
+    query_layer = attention.transpose_for_scores(attention.query(embedding_output))
+
+    attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2))
+    attention_scores /= math.sqrt(key_layer.size()[-1])
+
+    fused_softmax = (
+        FusedScaleMaskSoftmax(
+            input_in_fp16=True,
+            input_in_bf16=False,
+            mask_func=attention_mask_func,
+            scale=None,
+            softmax_in_fp32=False,
+            attn_mask_type=AttnMaskType.padding,
+            scaled_masked_softmax_fusion=True,
+        )
+        .cuda()
+        .half()
+    )
+
+    fused_softmax_output = fused_softmax(
+        attention_scores,
+        (mask != 0),
+    )
+
+    torch_softmax = (
+        FusedScaleMaskSoftmax(
+            input_in_fp16=True,
+            input_in_bf16=False,
+            mask_func=attention_mask_func,
+            scale=None,
+            softmax_in_fp32=False,
+            attn_mask_type=AttnMaskType.padding,
+            scaled_masked_softmax_fusion=False,
+        )
+        .cuda()
+        .half()
+    )
+
+    torch_softmax_output = torch_softmax(
+        attention_scores,
+        (mask != 0),
+    )
+
+    test_result = (fused_softmax_output - torch_softmax_output).abs()
+
+    while test_result.dim() != 1:
+        test_result = test_result.mean(dim=-1)
+
+    diff = test_result.mean(dim=-1)
+
+    if diff <= 1e-3:
+        print(
+            f"\n[Success] test_fused_softmax"
+            f"\n > mean_difference={diff}"
+            f"\n > fused_values={fused_softmax_output[-1][-1][-1][:5].tolist()}"
+            f"\n > torch_values={torch_softmax_output[-1][-1][-1][:5].tolist()}"
+        )
+    else:
+        print(
+            f"\n[Fail] test_fused_softmax"
+            f"\n > mean_difference={diff}, "
+            f"\n > fused_values={fused_softmax_output[-1][-1][-1][:5].tolist()}, "
+            f"\n > torch_values={torch_softmax_output[-1][-1][-1][:5].tolist()}"
+        )
+
+
+def test_fused_upper_triangle_mask_softmax():
+    gpt = GPT2Model.from_pretrained("gpt2").cuda().half()
+    tokenizer = GPT2Tokenizer.from_pretrained("gpt2")
+    test_text = (
+        "Hello. How are you? I am fine thank you and you? yes Good. "
+        "hi hi hi hi hi hi hi"  # 24
+    )
+
+    tokens = tokenizer(
+        [test_text] * 4,
+        return_tensors="pt",
+    )
+
+    attention_mask = tokens["attention_mask"].cuda()
+    attention_mask = attention_mask.view(attention_mask.size(0), -1)
+    attention_mask = attention_mask[:, None, None, :]
+    attention_mask = (1.0 - attention_mask) * -10000.0
+    attention_mask = attention_mask.repeat(1, 1, attention_mask.size()[-1], 1)
+    attn = gpt.h[0]
+
+    hidden_states = gpt.wte(tokens["input_ids"].cuda())
+    q, k, v = attn.attn.c_attn(hidden_states).split(768, dim=-1)
+    q = attn.attn._split_heads(q, attn.attn.num_heads, attn.attn.head_dim)
+    k = attn.attn._split_heads(k, attn.attn.num_heads, attn.attn.head_dim)
+    attn_weights = torch.matmul(q, k.transpose(-1, -2))
+
+    sq, sk = q.size(-2), k.size(-2)
+    causal_mask = attn.attn.bias[:, :, sk - sq : sk, :sk].bool()
+    total_mask = ~(causal_mask & (attention_mask == 0))
+    """
+    tensor([[[[False,  True,  True,  ...,  True,  True,  True],
+              [False, False,  True,  ...,  True,  True,  True],
+              [False, False, False,  ...,  True,  True,  True],
+              ...,
+              [False, False, False,  ..., False,  True,  True],
+              [False, False, False,  ..., False, False,  True],
+              [False, False, False,  ..., False, False, False]]]
+    """
+
+    fused_softmax = (
+        FusedScaleMaskSoftmax(
+            input_in_fp16=True,
+            input_in_bf16=False,
+            mask_func=attention_mask_func,
+            scale=None,
+            softmax_in_fp32=False,
+            attn_mask_type=AttnMaskType.causal,
+            scaled_masked_softmax_fusion=True,
+        )
+        .cuda()
+        .half()
+    )
+
+    fused_softmax_output = fused_softmax(
+        attn_weights,
+        total_mask,
+    )
+
+    torch_softmax = (
+        FusedScaleMaskSoftmax(
+            input_in_fp16=True,
+            input_in_bf16=False,
+            mask_func=attention_mask_func,
+            scale=None,
+            softmax_in_fp32=False,
+            attn_mask_type=AttnMaskType.causal,
+            scaled_masked_softmax_fusion=False,
+        )
+        .cuda()
+        .half()
+    )
+
+    torch_softmax_output = torch_softmax(
+        attn_weights,
+        total_mask,
+    )
+
+    test_result = (fused_softmax_output - torch_softmax_output).abs()
+
+    while test_result.dim() != 1:
+        test_result = test_result.mean(dim=-1)
+
+    diff = test_result.mean(dim=-1)
+
+    if diff <= 1e-3:
+        print(
+            f"\n[Success] test_fused_upper_triangle_mask_softmax"
+            f"\n > mean_difference={diff}"
+            f"\n > fused_values={fused_softmax_output[-1][-1][-1][:5].tolist()}"
+            f"\n > torch_values={torch_softmax_output[-1][-1][-1][:5].tolist()}"
+        )
+    else:
+        print(
+            f"\n[Fail] test_fused_upper_triangle_mask_softmax"
+            f"\n > mean_difference={diff}, "
+            f"\n > fused_values={fused_softmax_output[-1][-1][-1][:5].tolist()}, "
+            f"\n > torch_values={torch_softmax_output[-1][-1][-1][:5].tolist()}"
+        )
+
+
+def test_layer_norm():
+    bert = BertModel.from_pretrained("bert-base-cased").cuda().half()
+    tokenizer = BertTokenizer.from_pretrained("bert-base-cased")
+    test_text = (
+        "Hello. How are you? I am fine thank you and you? yes Good. "
+        "hi hi hi hi hi hi hi hi hi hi hi hi hi"  # 32
+    )
+
+    tokens = tokenizer(
+        [test_text] * 4,
+        return_tensors="pt",
+    )
+
+    # [bsz, seq_len, d_model]
+    embedding_output = (
+        bert.embeddings(
+            input_ids=tokens["input_ids"].cuda(),
+            position_ids=None,
+            token_type_ids=tokens["token_type_ids"].cuda(),
+            inputs_embeds=None,
+            past_key_values_length=0,
+        )
+        .cuda()
+        .half()
+    )
+
+    fused_layernorm_layer = (
+        MixedFusedLayerNorm(normalized_shape=embedding_output.size(-1)).cuda().half()
+    )
+
+    torch_layernorm_layer = (
+        LayerNorm(normalized_shape=embedding_output.size(-1)).cuda().half()
+    )
+
+    fused_output = fused_layernorm_layer(embedding_output)
+    torch_output = torch_layernorm_layer(embedding_output)
+    test_result = (fused_output - torch_output).abs()
+
+    while test_result.dim() != 1:
+        test_result = test_result.mean(dim=-1)
+
+    diff = test_result.mean(dim=-1)
+
+    if diff <= 1e-3:
+        print(
+            f"\n[Success] test_layer_norm"
+            f"\n > mean_difference={diff}"
+            f"\n > fused_values={fused_output[-1][-1][:5].tolist()}"
+            f"\n > torch_values={torch_output[-1][-1][:5].tolist()}"
+        )
+    else:
+        print(
+            f"\n[Fail] test_layer_norm"
+            f"\n > mean_difference={diff}, "
+            f"\n > fused_values={fused_output[-1][-1][:5].tolist()}, "
+            f"\n > torch_values={torch_output[-1][-1][:5].tolist()}"
+        )
+
+
+def attention_mask_func(attention_scores, attention_mask):
+    attention_scores.masked_fill_(attention_mask, -10000.0)
+    return attention_scores
+
+
+def forward_torch_softmax(input, mask, scale):
+    input = input * scale
+    mask_output = attention_mask_func(input, mask) if mask is not None else input
+    probs = torch.nn.Softmax(dim=-1)(mask_output)
+    return probs
+
+
+def test_masked_softmax_forward():
+    import scaled_masked_softmax_cuda
+
+    batch = 2
+    attn = 16
+    scale_t = torch.tensor([1.0])
+    for qlen in [128, 256, 1024, 2048, 4096]:
+        for klen in [128, 256, 1024, 2048]:
+            inputs = torch.normal(0, 2, (batch, attn, qlen, klen), dtype=torch.float16, device='cuda:0')
+            masks = torch.randint(0, 2, (batch, 1, qlen, klen), dtype=torch.bool, device='cuda:0')
+            softmax_results = scaled_masked_softmax_cuda.forward(inputs, masks, scale_t[0].item())
+            softmax_results_torch = forward_torch_softmax(inputs, masks, scale_t[0].item())
+            error = (softmax_results_torch - softmax_results).abs().max()
+            assert error < 1e-3
+
+def test_masked_softmax_backward():
+    import scaled_masked_softmax_cuda
+
+    batch = 2
+    attn = 16
+    scale_t = torch.tensor([1.0])
+    for qlen in [128, 256, 1024, 2048, 4096]:
+        for klen in [128, 256, 1024, 2048]:
+            inputs = torch.normal(0, 2, (batch, attn, qlen, klen), dtype=torch.float16, device='cuda:0')
+            backward = torch.rand_like(inputs, dtype=torch.float16, device='cuda:0')
+            masks = torch.randint(0, 2, (batch, 1, qlen, klen), dtype=torch.bool, device='cuda:0')
+            softmax_results = scaled_masked_softmax_cuda.forward(inputs, masks, scale_t[0].item())
+            back_grad = scaled_masked_softmax_cuda.backward(backward, softmax_results, scale_t[0].item())
+
+            inputs.requires_grad = True
+            softmax_results_torch = forward_torch_softmax(inputs, masks, scale_t[0].item())
+            softmax_results_torch.backward(backward)
+            error = (back_grad - inputs.grad).abs().max()
+            assert error < 1e-3
+
+
+def test_allmasked_softmax_forward():
+    import scaled_masked_softmax_cuda
+
+    batch = 2
+    attn = 16
+    scale_t = torch.tensor([1.0])
+    for qlen in [128, 256, 1024, 2048, 4096]:
+        for klen in [128, 256, 1024, 2048]:
+            inputs = torch.normal(0, 2, (batch, attn, qlen, klen), dtype=torch.float16, device='cuda:0')
+            masks = torch.ones((batch, 1, qlen, klen), dtype=torch.bool, device='cuda:0')
+            softmax_results = scaled_masked_softmax_cuda.forward(inputs, masks, scale_t[0].item())
+            softmax_results_torch = torch.zeros_like(inputs)
+            error = (softmax_results_torch - softmax_results).abs().max()
+            assert error == 0.0
+
+
+def test_allmasked_softmax_backward():
+    import scaled_masked_softmax_cuda
+
+    batch = 2
+    attn = 16
+    scale_t = torch.tensor([1.0])
+    for qlen in [128, 256, 1024, 2048, 4096]:
+        for klen in [128, 256, 1024, 2048]:
+            inputs = torch.normal(0, 2, (batch, attn, qlen, klen), dtype=torch.float16, device='cuda:0')
+            backward = torch.rand_like(inputs, dtype=torch.float16, device='cuda:0')
+            masks = torch.ones((batch, 1, qlen, klen), dtype=torch.bool, device='cuda:0')
+            softmax_results = scaled_masked_softmax_cuda.forward(inputs, masks, scale_t[0].item())
+            back_grad = scaled_masked_softmax_cuda.backward(backward, softmax_results, scale_t[0].item())
+            inputs.requires_grad = True
+            softmax_results_torch = forward_torch_softmax(inputs, masks, scale_t[0].item())
+            softmax_results_torch.backward(backward)
+            error = (back_grad - inputs.grad).abs().max()
+            assert error < 1e-3
+
+
+if __name__ == "__main__":
+    try:
+        from transformers import BertTokenizer, GPT2Tokenizer
+        from transformers.models.bert.modeling_bert import BertModel
+        from transformers.models.gpt2.modeling_gpt2 import GPT2Model
+        import transformers
+
+        transformers.logging.set_verbosity(
+            transformers.logging.FATAL,
+        )
+
+    except:
+        print("\n[Fail] Please install `transformers` package to test fused kernels\n")
+        exit(-1)
+
+    load()
+    test_masked_softmax_forward()
+    test_masked_softmax_backward()
+    test_allmasked_softmax_forward()
+    test_allmasked_softmax_backward()
+    test_load_fused_kernels()
+    test_fused_softmax()
+    test_fused_upper_triangle_mask_softmax()
+    test_layer_norm()

megatron/legacy/fused_kernels/type_shim.h

+/* Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. */
+
+
+#include <ATen/ATen.h>
+#include "compat.h"
+
+
+#define DISPATCH_HALF_AND_BFLOAT(TYPE, NAME, ...)			\
+  switch(TYPE)								\
+    {									\
+    case at::ScalarType::Half:						\
+      {									\
+	using scalar_t = at::Half;					\
+	__VA_ARGS__;							\
+	break;								\
+      }									\
+    case at::ScalarType::BFloat16:					\
+      {									\
+	using scalar_t = at::BFloat16;					\
+	__VA_ARGS__;							\
+	break;								\
+      }									\
+    default:								\
+      AT_ERROR(#NAME, " not implemented for '", toString(TYPE), "'");	\
+      }
+
+
+#define DISPATCH_HALF_BFLOAT_AND_FLOAT(TYPE, NAME, ...)			\
+  switch(TYPE)								\
+    {									\
+    case at::ScalarType::Half:						\
+      {									\
+	using scalar_t = at::Half;					\
+	__VA_ARGS__;							\
+	break;								\
+      }									\
+    case at::ScalarType::BFloat16:					\
+      {									\
+	using scalar_t = at::BFloat16;					\
+	__VA_ARGS__;							\
+	break;								\
+      }									\
+    case at::ScalarType::Float:						\
+      {									\
+	using scalar_t = float;					\
+	__VA_ARGS__;							\
+	break;								\
+      }									\
+    default:								\
+      AT_ERROR(#NAME, " not implemented for '", toString(TYPE), "'");	\
+      }
+
+
+
+#define DISPATCH_FLOAT_HALF_AND_BFLOAT_INOUT_TYPES(TYPEIN, TYPEOUT, NAME, ...) \
+  switch(TYPEIN)							\
+    {									\
+    case at::ScalarType::Float:						\
+      {									\
+	using scalar_t_in = float;					\
+	switch(TYPEOUT)							\
+	  {								\
+	  case at::ScalarType::Float:					\
+	    {								\
+	      using scalar_t_out = float;				\
+	      __VA_ARGS__;						\
+	      break;							\
+	    }								\
+	  case at::ScalarType::Half:					\
+	    {								\
+	      using scalar_t_out = at::Half;				\
+	      __VA_ARGS__;						\
+	      break;							\
+	    }								\
+	  case at::ScalarType::BFloat16:				\
+	    {								\
+	      using scalar_t_out = at::BFloat16;			\
+	      __VA_ARGS__;						\
+	      break;							\
+	    }								\
+	  default:							\
+	    AT_ERROR(#NAME, " not implemented for '", toString(TYPEOUT), "'"); \
+	  }								\
+	break;								\
+      }									\
+    case at::ScalarType::Half:						\
+      {									\
+	using scalar_t_in = at::Half;					\
+	using scalar_t_out = at::Half;					\
+	__VA_ARGS__;							\
+	break;								\
+      }									\
+    case at::ScalarType::BFloat16:					\
+      {									\
+	using scalar_t_in = at::BFloat16;				\
+	using scalar_t_out = at::BFloat16;				\
+	__VA_ARGS__;							\
+	break;								\
+      }									\
+    default:								\
+      AT_ERROR(#NAME, " not implemented for '", toString(TYPEIN), "'");	\
+    }
+

megatron/legacy/indexer.py

+import sys
+import time
+import torch
+import torch.distributed as dist
+
+from megatron.training import get_args, print_rank_0
+from megatron.core import mpu
+from megatron.training.checkpointing import load_biencoder_checkpoint
+from megatron.legacy.data.orqa_wiki_dataset import get_open_retrieval_wiki_dataset
+from megatron.legacy.data.orqa_wiki_dataset import get_open_retrieval_batch
+from megatron.legacy.data.biencoder_dataset_utils import get_one_epoch_dataloader
+from megatron.legacy.data.realm_index import detach, OpenRetreivalDataStore
+from megatron.legacy.model.biencoder_model import get_model_provider
+from megatron.training import get_model
+
+
+class IndexBuilder(object):
+    """
+    Object for taking one pass over a dataset and creating a BlockData of its
+    embeddings
+    """
+    def __init__(self):
+        args = get_args()
+        self.model = None
+        self.dataloader = None
+        self.evidence_embedder_obj = None
+        self.biencoder_shared_query_context_model = \
+            args.biencoder_shared_query_context_model
+
+        # need to know whether we're using a REALM checkpoint (args.load)
+        # or ICT checkpoint
+        assert not (args.load and args.ict_load)
+
+        self.log_interval = args.indexer_log_interval
+        self.batch_size = args.indexer_batch_size
+
+        self.load_attributes()
+        self.is_main_builder = mpu.get_data_parallel_rank() == 0
+        self.num_total_builders = mpu.get_data_parallel_world_size()
+        self.iteration = self.total_processed = 0
+
+    def load_attributes(self):
+        """
+        Load the necessary attributes: model, dataloader and empty BlockData
+        """
+        only_context_model = True
+        if self.biencoder_shared_query_context_model:
+            only_context_model = False
+
+        model = get_model(get_model_provider(only_context_model=\
+            only_context_model, biencoder_shared_query_context_model=\
+            self.biencoder_shared_query_context_model))
+
+        self.model = load_biencoder_checkpoint(model,
+                only_context_model=only_context_model)
+
+        assert len(self.model) == 1
+        self.model[0].eval()
+
+        self.dataset = get_open_retrieval_wiki_dataset()
+        self.dataloader = iter(get_one_epoch_dataloader(self.dataset, \
+            self.batch_size))
+
+        self.evidence_embedder_obj = OpenRetreivalDataStore( \
+            load_from_path=False)
+
+    def track_and_report_progress(self, batch_size):
+        """
+        Utility function for tracking progress
+        """
+        self.iteration += 1
+        self.total_processed += batch_size * self.num_total_builders
+        if self.is_main_builder and self.iteration % self.log_interval == 0:
+            print('Batch {:10d} | Total {:10d}'.format(self.iteration,
+                self.total_processed), flush=True)
+
+    def build_and_save_index(self):
+        """
+        Goes through one epoch of the dataloader and adds all data to this
+        instance's BlockData.
+
+        The copy of BlockData is saved as a shard, which when run in a
+        distributed setting will be consolidated by the rank 0 process
+        and saved as a final pickled BlockData.
+        """
+        assert len(self.model) == 1
+        unwrapped_model = self.model[0]
+
+        while not hasattr(unwrapped_model, 'embed_text'):
+            unwrapped_model = unwrapped_model.module
+
+        while True:
+            try:
+                # batch also has query_tokens and query_pad_data
+                row_id, context_tokens, context_mask, context_types, \
+                    context_pad_mask = get_open_retrieval_batch( \
+                    self.dataloader)
+            except (StopIteration, IndexError):
+                break
+
+            # TODO: can we add with torch.no_grad() to reduce memory usage
+            # detach, separate fields and add to BlockData
+            assert context_mask.dtype == torch.bool
+            context_logits = unwrapped_model.embed_text(
+                unwrapped_model.context_model, context_tokens, context_mask,
+                context_types)
+
+            context_logits = detach(context_logits)
+            row_id = detach(row_id)
+
+            self.evidence_embedder_obj.add_block_data(row_id, context_logits)
+            self.track_and_report_progress(batch_size=len(row_id))
+
+        # This process signals to finalize its shard and then synchronize with
+        # the other processes
+        self.evidence_embedder_obj.save_shard()
+        torch.distributed.barrier()
+        del self.model
+
+        # rank 0 process builds the final copy
+        if self.is_main_builder:
+            self.evidence_embedder_obj.merge_shards_and_save()
+            # make sure that every single piece of data was embedded
+            assert len(self.evidence_embedder_obj.embed_data) == \
+                len(self.dataset)
+        self.evidence_embedder_obj.clear()
+
+        # complete building the final copy
+        torch.distributed.barrier()

megatron/legacy/model/__init__.py

+# Copyright (c) 2023, NVIDIA CORPORATION. All rights reserved.
+
+from .fused_layer_norm import MixedFusedLayerNorm as LayerNorm
+from .rms_norm import RMSNorm
+#from lightop import RMSNorm
+from .bert_model import BertModel
+from .gpt_model import GPTModel
+from .t5_model import T5Model
+from .language_model import get_language_model
+from .module import Float16Module