第一次测试提交

Megatron-LM-core_r0.7.0.beta/megatron/inference/text_generation/api.py

+# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
+
+"""Inference API."""
+
+
+import torch
+
+from megatron.core import mpu
+from .communication import broadcast_float_list
+from .generation import (
+        generate_tokens_probs_and_return_on_first_stage,
+        score_and_return_on_first_stage,
+        beam_search_and_return_on_first_stage)
+from .tokenization import (
+    tokenize_prompts,
+    detokenize_generations)
+
+def generate_and_post_process(model,
+                              prompts=None,
+                              tokens_to_generate=0,
+                              return_output_log_probs=False,
+                              top_k_sampling=0,
+                              top_p_sampling=0.0,
+                              top_p_decay=0.0,
+                              top_p_bound=0.0,
+                              temperature=1.0,
+                              add_BOS=False,
+                              use_eod_token_for_early_termination=True,
+                              stop_on_double_eol=False,
+                              stop_on_eol=False,
+                              prevent_newline_after_colon=False,
+                              random_seed=-1,
+                              return_logits=False):
+    """Run inference and post-process outputs, i.e., detokenize,
+    move to cpu and convert to list."""
+
+    # Main inference.
+    tokens, lengths, output_log_probs, logits = generate(
+        model,
+        prompts=prompts,
+        tokens_to_generate=tokens_to_generate,
+        return_output_log_probs=return_output_log_probs,
+        top_k_sampling=top_k_sampling,
+        top_p_sampling=top_p_sampling,
+        top_p_decay=top_p_decay,
+        top_p_bound=top_p_bound,
+        temperature=temperature,
+        add_BOS=add_BOS,
+        use_eod_token_for_early_termination=use_eod_token_for_early_termination,
+        stop_on_double_eol=stop_on_double_eol,
+        stop_on_eol=stop_on_eol,
+        prevent_newline_after_colon=prevent_newline_after_colon,
+        random_seed=random_seed)
+
+    # Only post-process on first stage.
+    if mpu.is_pipeline_first_stage():
+        tokens, prompts_plus_generations, prompts_plus_generations_segments = \
+            detokenize_generations(tokens, lengths, True)
+
+        if return_output_log_probs:
+            output_log_probs = output_log_probs.cpu().numpy().tolist()
+            for i, (prob, seg) in enumerate(zip(output_log_probs, prompts_plus_generations_segments)):
+                output_log_probs[i] = prob[:len(seg)-1]
+
+        if return_logits:
+            assert(tokens_to_generate == 0)
+            assert(mpu.get_pipeline_model_parallel_world_size() == 1)
+            return prompts_plus_generations, prompts_plus_generations_segments, \
+            output_log_probs, tokens, logits
+        else:
+            return prompts_plus_generations, prompts_plus_generations_segments, \
+            output_log_probs, tokens
+
+    return None
+
+def generate(model,
+             prompts=None,
+             tokens_to_generate=0,
+             return_output_log_probs=False,
+             top_k_sampling=0,
+             top_p_sampling=0.0,
+             top_p_decay=0.0,
+             top_p_bound=0.0,
+             temperature=1.0,
+             add_BOS=False,
+             use_eod_token_for_early_termination=True,
+             stop_on_double_eol=False,
+             stop_on_eol=False,
+             prevent_newline_after_colon=False,
+             random_seed=-1):
+    """Given prompts and input parameters, run inference and return:
+       tokens: prompts plus the generated tokens.
+       lengths: length of the prompt + generations. Note that we can
+           discard tokens in the tokens tensor that are after the
+           corresponding length.
+       output_log_probs: log probs of the tokens.
+    """
+
+    # Make sure input params are avaialble to all ranks.
+    values = [tokens_to_generate,
+              return_output_log_probs,
+              top_k_sampling, top_p_sampling, top_p_decay, top_p_bound,
+              temperature, add_BOS, use_eod_token_for_early_termination,
+              stop_on_double_eol,
+              stop_on_eol,
+              prevent_newline_after_colon,
+              random_seed]
+    values_float_tensor = broadcast_float_list(len(values), float_list=values)
+    tokens_to_generate = int(values_float_tensor[0].item())
+    return_output_log_probs = bool(values_float_tensor[1].item())
+    top_k_sampling = int(values_float_tensor[2].item())
+    top_p_sampling = values_float_tensor[3].item()
+    top_p_decay = values_float_tensor[4].item()
+    top_p_bound = values_float_tensor[5].item()
+    temperature = values_float_tensor[6].item()
+    add_BOS = bool(values_float_tensor[7].item())
+    use_eod_token_for_early_termination = bool(values_float_tensor[8].item())
+    stop_on_double_eol = bool(values_float_tensor[9].item())
+    stop_on_eol = bool(values_float_tensor[10].item())
+    prevent_newline_after_colon = bool(values_float_tensor[11].item())
+    random_seed = int(values_float_tensor[12].item())
+
+    if random_seed != -1:
+        torch.random.manual_seed(random_seed)
+
+    # Tokenize prompts and get the batch.
+    # Note that these tensors are broadcaseted to all ranks.
+    if torch.distributed.get_rank() == 0:
+        assert prompts is not None
+    
+    context_tokens_tensor, context_length_tensor = tokenize_prompts(
+        prompts=prompts, tokens_to_generate=tokens_to_generate, add_BOS=add_BOS)
+
+    if tokens_to_generate == 0:
+        return score_and_return_on_first_stage(
+            model, context_tokens_tensor, context_length_tensor)
+    
+    # Main inference function.
+    # Note that the outputs are available on the first stage.
+    return generate_tokens_probs_and_return_on_first_stage(
+        model, context_tokens_tensor, context_length_tensor,
+        return_output_log_probs=return_output_log_probs,
+        top_k=top_k_sampling,
+        top_p=top_p_sampling,
+        top_p_decay=top_p_decay,
+        top_p_bound=top_p_bound,
+        temperature=temperature,
+        use_eod_token_for_early_termination=use_eod_token_for_early_termination,
+        stop_on_double_eol=stop_on_double_eol,
+        stop_on_eol=stop_on_eol,
+        prevent_newline_after_colon=prevent_newline_after_colon)
+
+def beam_search_and_post_process(model,
+                                 prompts=None,
+                                 tokens_to_generate=0,
+                                 beam_size=0,
+                                 add_BOS=False,
+                                 stop_token=50256,
+                                 num_return_gen=1,
+                                 length_penalty=1,
+                                 prevent_newline_after_colon=False):
+    """Run beam search and post-process outputs, i.e., detokenize,
+    move to cpu and convert to list."""
+
+    # Main inference.
+    tokens, scores = beam_search(model,
+                                 prompts=prompts,
+                                 tokens_to_generate=tokens_to_generate,
+                                 beam_size=beam_size,
+                                 add_BOS=add_BOS,
+                                 stop_token=stop_token,
+                                 num_return_gen=num_return_gen,
+                                 length_penalty=length_penalty,
+                                 prevent_newline_after_colon=prevent_newline_after_colon)
+    # Only post-process on first stage.
+    if mpu.is_pipeline_first_stage():
+        lengths = tokens.size(1)*torch.ones(beam_size, dtype=torch.int64, device=torch.cuda.current_device()) 
+        tokens, prompts_plus_generations, prompts_plus_generations_segments = detokenize_generations(tokens, lengths, True)
+        scores = scores.cpu().numpy().tolist()
+        return prompts_plus_generations, prompts_plus_generations_segments, scores
+
+    return None
+
+def beam_search(model, prompts=None, tokens_to_generate=0, beam_size=0, add_BOS=False, stop_token=50256, num_return_gen=1, length_penalty=1, prevent_newline_after_colon=False):
+    # Make sure input params are avaialble to all ranks.
+    values = [tokens_to_generate,
+              beam_size,
+              add_BOS,
+              stop_token,
+              num_return_gen,
+              length_penalty,
+              prevent_newline_after_colon]
+    values_float_tensor = broadcast_float_list(len(values), float_list=values)
+    tokens_to_generate = int(values_float_tensor[0].item())
+    beam_size = int(values_float_tensor[1].item())
+    add_BOS = bool(values_float_tensor[2].item())
+    stop_token = int(values_float_tensor[3].item())
+    num_return_gen = int(values_float_tensor[4].item())
+    length_penalty = values_float_tensor[5].item()
+    prevent_newline_after_colon = values_float_tensor[6].item()
+
+    context_tokens_tensor, context_length_tensor = tokenize_prompts(
+        prompts=prompts, tokens_to_generate=tokens_to_generate, add_BOS=add_BOS)
+    
+    return beam_search_and_return_on_first_stage(model, context_tokens_tensor, context_length_tensor, 
+            beam_size, stop_token=stop_token, num_return_gen=num_return_gen, length_penalty=length_penalty,
+            prevent_newline_after_colon=prevent_newline_after_colon)

Megatron-LM-core_r0.7.0.beta/megatron/inference/text_generation/beam_utils.py

+# coding=utf-8
+# Copyright 2018 The Google AI Language Team Authors, Facebook AI Research authors and The HuggingFace Inc. team.
+# Copyright (c) 2018, 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.
+
+
+## from huggingface beam search
+class BeamHypotheses(object):
+    def __init__(self, num_beams, length_penalty=1.0, early_stopping=False):
+        """
+        Initialize n-best list of hypotheses.
+        """
+        self.length_penalty = length_penalty
+        self.early_stopping = early_stopping
+        self.num_beams = num_beams
+        self.beams = []
+        self.worst_score = 1e9
+
+    def __len__(self):
+        """
+        Number of hypotheses in the list.
+        """
+        return len(self.beams)
+
+    def add(self, hyp, sum_logprobs, length):
+        """
+        Add a new hypothesis to the list.
+        """
+        score = sum_logprobs / length ** self.length_penalty
+        if len(self) < self.num_beams or score > self.worst_score:
+            self.beams.append((score, hyp))
+            if len(self) > self.num_beams:
+                sorted_scores = sorted([(s, idx) for idx, (s, _) in enumerate(self.beams)])
+                del self.beams[sorted_scores[0][1]]
+                self.worst_score = sorted_scores[1][0]
+            else:
+                self.worst_score = min(score, self.worst_score)
+
+    def is_done(self, best_sum_logprobs, cur_len):
+        """
+        If there are enough hypotheses and that none of the hypotheses being generated
+        can become better than the worst one in the heap, then we are done with this sentence.
+        """
+
+        if len(self) < self.num_beams:
+            return False
+        elif self.early_stopping:
+            return True
+        else:
+            cur_score = best_sum_logprobs / cur_len ** self.length_penalty
+            ret = self.worst_score >= cur_score
+            return ret
+

Megatron-LM-core_r0.7.0.beta/megatron/inference/text_generation/communication.py

+# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
+
+"""Communications utilities."""
+
+
+import torch
+
+from megatron.core import mpu
+
+
+
+# TODO: use functions from megatron/p2p
+def recv_from_prev_pipeline_rank_(recv_buffer=None):
+    """Receive from previous pipeline stage and update the
+    input buffer inplace."""
+    if not mpu.is_pipeline_first_stage():
+        assert recv_buffer is not None
+        recv_prev_op = torch.distributed.P2POp(
+            torch.distributed.irecv, recv_buffer,
+            mpu.get_pipeline_model_parallel_prev_rank())
+        reqs = torch.distributed.batch_isend_irecv([recv_prev_op])
+        for req in reqs:
+            req.wait()
+        # To protect against race condition when using batch_isend_irecv().
+        torch.cuda.synchronize()
+
+
+
+# TODO: use functions from megatron/p2p
+def send_to_next_pipeline_rank(tensor=None):
+    """Send output to the next pipeline stage."""
+    if not mpu.is_pipeline_last_stage():
+        assert tensor is not None
+        send_next_op = torch.distributed.P2POp(
+            torch.distributed.isend, tensor,
+            mpu.get_pipeline_model_parallel_next_rank())
+        reqs = torch.distributed.batch_isend_irecv([send_next_op])
+        for req in reqs:
+            req.wait()
+        # To protect against race condition when using batch_isend_irecv().
+        torch.cuda.synchronize()
+
+
+
+def _is_cuda(tensor):
+    """Check if a tensor is not none and is cuda."""
+    assert tensor is not None
+    assert tensor.is_cuda
+
+
+
+def _is_cuda_contiguous(tensor):
+    """Check if a tensor is not none, is cuda, and is contiguous."""
+    _is_cuda(tensor)
+    assert tensor.is_contiguous()
+
+
+
+def broadcast_from_last_pipeline_stage(size, dtype, tensor=None):
+    """Broadcast a tensor from last pipeline stage to all ranks."""
+
+    is_last_stage = mpu.is_pipeline_last_stage()
+    # If first stage and last state are the same, then there is no
+    # pipeline parallelism and no need to communicate.
+    if mpu.is_pipeline_first_stage() and is_last_stage:
+        return tensor
+
+    if is_last_stage:
+        _is_cuda_contiguous(tensor)
+    else:
+        tensor = torch.empty(size,
+                             dtype=dtype,
+                             device=torch.cuda.current_device())
+    # Get the group and corresponding source rank.
+    src = mpu.get_pipeline_model_parallel_last_rank()
+    group = mpu.get_pipeline_model_parallel_group()
+    torch.distributed.broadcast(tensor, src, group)
+
+    return tensor
+
+
+
+def broadcast_from_last_to_first_pipeline_stage(size, dtype, tensor=None):
+    """Broadcast tensor values from last stage into the first stage."""
+
+    is_last_stage = mpu.is_pipeline_last_stage()
+    is_first_stage = mpu.is_pipeline_first_stage()
+    # If first stage and last state are the same, then there is no
+    # pipeline parallelism and no need to communicate.
+    if is_first_stage and is_last_stage:
+        return tensor
+    # Only first and last stage pipeline stages need to be involved.
+    if is_last_stage or is_first_stage:
+        if is_last_stage:
+            _is_cuda_contiguous(tensor)
+        else:
+            tensor = torch.empty(size,
+                                 dtype=dtype,
+                                 device=torch.cuda.current_device())
+        src = mpu.get_pipeline_model_parallel_last_rank()
+        group = mpu.get_embedding_group()
+        # Broadcast from last stage into the first stage.
+        torch.distributed.broadcast(tensor, src, group)
+    else:
+        tensor = None
+
+    return tensor
+
+
+
+def copy_from_last_to_first_pipeline_stage(size, dtype, tensor=None):
+    """Copy tensor values from last stage into the first stage.
+    Note that the input tensor is updated in place."""
+
+    is_last_stage = mpu.is_pipeline_last_stage()
+    is_first_stage = mpu.is_pipeline_first_stage()
+    # If first stage and last state are the same, then there is no
+    # pipeline parallelism and no need to communicate.
+    if is_first_stage and is_last_stage:
+        return
+    # Only first and last stage pipeline stages need to be involved.
+    if is_last_stage or is_first_stage:
+        _is_cuda(tensor)
+        is_contiguous = tensor.is_contiguous()
+        src = mpu.get_pipeline_model_parallel_last_rank()
+        group = mpu.get_embedding_group()
+        if is_contiguous:
+            tensor_ = tensor
+        else:
+            if is_last_stage:
+                tensor_ = tensor.contiguous()
+            else:
+                tensor_ = torch.empty(size,
+                                      dtype=dtype,
+                                      device=torch.cuda.current_device())
+        # Broadcast from last stage into the first stage.
+        torch.distributed.broadcast(tensor_, src, group)
+        # Update the first stage tensor
+        if is_first_stage and not is_contiguous:
+            tensor[...] = tensor_
+
+
+
+def broadcast_tensor(size, dtype, tensor=None, rank=0):
+    """ Given size and type of a tensor on all ranks and the tensor value
+        only on a specific rank, broadcast from that rank to all other ranks.
+    """
+
+    if torch.distributed.get_rank() == rank:
+        _is_cuda_contiguous(tensor)
+    else:
+        tensor = torch.empty(size,
+                             dtype=dtype,
+                             device=torch.cuda.current_device())
+
+    torch.distributed.broadcast(tensor, rank)
+
+    return tensor
+
+
+
+def broadcast_list(size, dtype, list_values=None, rank=0):
+    """Broadcast a list of values with a given type."""
+
+    tensor = None
+    if torch.distributed.get_rank() == rank:
+        tensor = torch.tensor(list_values, dtype=dtype,
+                              device=torch.cuda.current_device())
+
+    return broadcast_tensor(size, dtype, tensor=tensor, rank=rank)
+
+
+
+def broadcast_int_list(size, int_list=None, rank=0):
+    """Broadcast a list of interger values."""
+
+    return broadcast_list(size, torch.int64, list_values=int_list, rank=rank)
+
+
+
+def broadcast_float_list(size, float_list=None, rank=0):
+    """Broadcast a list of float values."""
+
+    return broadcast_list(size, torch.float32, list_values=float_list,
+                          rank=rank)

Megatron-LM-core_r0.7.0.beta/megatron/inference/text_generation/forward_step.py

+# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
+
+"""Forward step utilities."""
+
+from collections.abc import Iterable
+
+import torch
+
+from megatron.training import get_args
+from megatron.core import mpu, InferenceParams
+from .communication import (
+    send_to_next_pipeline_rank,
+    recv_from_prev_pipeline_rank_)
+
+
+class ForwardStep:
+    """Forward step function with all the communications.
+    We use a class here to hide the inference parameters
+    from the outside caller."""
+
+    def __init__(self, model, max_batch_size, max_sequence_length):
+        """Set values so we don't need to do it multiple times."""
+        # Make sure model is in eval mode.
+        assert not isinstance(model, Iterable), \
+            'interleaving schedule is not supported for inference'
+        model.eval()
+        self.model = model
+        # Initialize inference parameters.
+        self.inference_params = InferenceParams(max_batch_size,
+                                                max_sequence_length)
+        # Pipelining arguments.
+        args = get_args()
+        self.pipeline_size_larger_than_one = (
+            args.pipeline_model_parallel_size > 1)
+        # Threshold of pipelining.
+        self.pipelining_batch_x_seqlen = \
+            args.inference_batch_times_seqlen_threshold
+
+
+    def __call__(self, tokens, position_ids, attention_mask):
+        """Invocation of the forward methods. Note that self.inference_params
+        is being modified by the forward step."""
+        # Pipelining case.
+        if self.pipeline_size_larger_than_one:
+            current_batch_x_seqlen = tokens.size(0) * tokens.size(1)
+            if current_batch_x_seqlen >= self.pipelining_batch_x_seqlen:
+                micro_batch_size = \
+                    max(1, self.pipelining_batch_x_seqlen // tokens.size(1))
+                return _with_pipelining_forward_step(self.model,
+                                                     tokens,
+                                                     position_ids,
+                                                     attention_mask,
+                                                     self.inference_params,
+                                                     micro_batch_size)
+
+        return _no_pipelining_forward_step(self.model,
+                                           tokens,
+                                           position_ids,
+                                           attention_mask,
+                                           self.inference_params)
+
+
+
+def _get_recv_buffer_dtype(args):
+    """Receive happens between the layers."""
+    if args.fp32_residual_connection:
+        return torch.float
+    return args.params_dtype
+
+
+
+def _allocate_recv_buffer(batch_size, sequence_length):
+    """Receive happens between the layers with size [s, b, h]."""
+    if mpu.is_pipeline_first_stage():
+        return None
+    args = get_args()
+    recv_size = (sequence_length, batch_size, args.hidden_size)
+    return torch.empty(recv_size,
+                       dtype=_get_recv_buffer_dtype(args),
+                       device=torch.cuda.current_device())
+
+
+
+def _forward_step_helper(model, tokens, position_ids, attention_mask,
+                         inference_params, recv_buffer=None):
+    """Single forward step. Update the allocate memory flag so
+    only the first time the memory is allocated."""
+    batch_size = tokens.size(0)
+    sequence_length = tokens.size(1)
+    if recv_buffer is None:
+        recv_buffer = _allocate_recv_buffer(batch_size, sequence_length)
+
+    # Receive from previous stage.
+    recv_from_prev_pipeline_rank_(recv_buffer)
+
+    # Forward pass through the model.
+    model.set_input_tensor(recv_buffer)
+    output_tensor = model(tokens, position_ids, attention_mask,
+                          inference_params=inference_params)
+
+    # Send output to the next stage.
+    send_to_next_pipeline_rank(output_tensor)
+
+    return output_tensor
+
+
+
+def _no_pipelining_forward_step(model, tokens, position_ids, attention_mask,
+                                inference_params, recv_buffer=None):
+    """If recv_buffer is none, we will allocate one on the fly."""
+    # Run a simple forward pass.
+    output_tensor = _forward_step_helper(model, tokens, position_ids,
+                                         attention_mask, inference_params,
+                                         recv_buffer=recv_buffer)
+    # Update the sequence length offset.
+    inference_params.sequence_len_offset += tokens.size(1)
+
+    logits = None
+    if mpu.is_pipeline_last_stage():
+        logits = output_tensor
+
+    return logits
+
+
+
+def _with_pipelining_forward_step(model, tokens, position_ids, attention_mask,
+                                  inference_params, micro_batch_size):
+    """No interleaving is supported."""
+    sequence_length = tokens.size(1)
+    batch_size = tokens.size(0)
+
+    # Divide the batch dimension into micro batches.
+    num_micro_batches, last_chunk = divmod(batch_size,
+                                           micro_batch_size)
+    if last_chunk > 0:
+        num_micro_batches += 1
+
+    # Preallocate memory for output logits.
+    logits = None
+    if mpu.is_pipeline_last_stage():
+        args = get_args()
+        logits = torch.empty(
+            (batch_size, sequence_length, args.padded_vocab_size),
+            dtype=torch.float32, device=torch.cuda.current_device())
+
+    # Preallocate recv buffer.
+    recv_buffer = _allocate_recv_buffer(micro_batch_size, sequence_length)
+
+    for micro_batch_index in range(num_micro_batches):
+        # Slice among the batch dimenion.
+        start = micro_batch_index * micro_batch_size
+        end = min(start + micro_batch_size, batch_size)
+        this_micro_batch_size = end - start
+        tokens2use = tokens[start:end, ...]
+        position_ids2use = position_ids[start:end, ...]
+
+        # Run a simple forward pass.
+        if this_micro_batch_size != micro_batch_size:
+            recv_buffer = None
+        output = _forward_step_helper(model, tokens2use, position_ids2use,
+                                      attention_mask, inference_params,
+                                      recv_buffer=recv_buffer)
+
+        # Adjust the batch size offset to account for the micro-batch.
+        inference_params.batch_size_offset += this_micro_batch_size
+
+        # Copy logits.
+        if mpu.is_pipeline_last_stage():
+            logits[start:end, ...] = output
+
+    # Once we are done with all the micro-batches, we can
+    # adjust the sequence length offset.
+    inference_params.sequence_len_offset += sequence_length
+    # and reset the batch size offset
+    inference_params.batch_size_offset = 0
+
+    return logits

Megatron-LM-core_r0.7.0.beta/megatron/inference/text_generation/generation.py

+# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
+
+"""Generation utilities."""
+
+import torch
+import torch.nn.functional as F
+
+from megatron.training import get_args, get_tokenizer
+from megatron.core import mpu
+from megatron.training.utils import get_ltor_masks_and_position_ids
+from .communication import (
+    copy_from_last_to_first_pipeline_stage,
+    broadcast_from_last_pipeline_stage,
+    broadcast_from_last_to_first_pipeline_stage)
+from .forward_step import ForwardStep
+from .sampling import sample
+from .beam_utils import BeamHypotheses
+
+def score_and_return_on_first_stage(model, tokens, lengths):
+    """Function for just scoring.
+
+    Args:
+        model: no interleaving is supported.
+        tokens: prompt tokens extended to be of size [b, max_prompt_length]
+        lengths: original prompt length, size: [b]
+    Note: Outside of model, other parameters only need to be available on
+          rank 0.
+
+    Returns:
+        output_log_probs: log probability of the selected tokens. size: [b, s]
+    """
+
+    args = get_args()
+
+    batch_size = tokens.size(0)
+    max_prompt_length = lengths.max().item()
+    assert max_prompt_length == tokens.size(1)
+    
+    if max_prompt_length > args.max_position_embeddings:
+        raise ValueError("Length of prompt + tokens_to_generate longer than allowed")
+    
+    if max_prompt_length * batch_size > args.max_tokens_to_oom:
+        raise ValueError("Too many tokens.  " + str(max_prompt_length*batch_size)+ " is greater than "+str(args.max_tokens_to_oom))
+
+    # forward step.
+    forward_step = ForwardStep(model, batch_size, max_prompt_length)
+
+    # ===================
+    # Pre-allocate memory
+    # ===================
+
+    # Log probability of the sequence (prompt + generated tokens).
+    output_log_probs = None
+    output_log_probs_size = (batch_size, max_prompt_length - 1)
+    
+    if mpu.is_pipeline_last_stage():
+        output_log_probs = torch.empty(output_log_probs_size,
+                                       dtype=torch.float32,
+                                       device=torch.cuda.current_device())
+    
+    # =============
+    # Run infernece
+    # =============
+    with torch.no_grad():
+        attention_mask, position_ids = _build_attention_mask_and_position_ids(tokens)
+        
+        # logits will be meanigful only in the last pipeline stage.
+        logits = forward_step(tokens, position_ids, attention_mask)
+
+        if mpu.is_pipeline_last_stage():
+            # Always the last stage should have an output.
+            assert logits is not None
+            log_probs = F.log_softmax(logits, dim=2)
+            
+            # Pick the tokens that we need to get the log
+            # probabilities for. Note that next input token is
+            # the token which we selected in the current logits,
+            # so shift by 1.
+            indices = torch.unsqueeze(tokens[:, 1:], 2)
+            output_log_probs = torch.gather(log_probs, 2, indices).squeeze(2)
+    
+    # ======================================
+    # Broadcast to the first pipeline stage.
+    # ======================================
+    output_log_probs = broadcast_from_last_to_first_pipeline_stage(
+        output_log_probs_size, torch.float32, output_log_probs)
+    
+    return tokens, lengths, output_log_probs, logits
+
+def generate_tokens_probs_and_return_on_first_stage(
+        model, tokens, lengths,
+        return_output_log_probs=False,
+        top_k=0, top_p=0.0, top_p_decay=0.0, top_p_bound=0.0,
+        temperature=1.0,
+        use_eod_token_for_early_termination=True,
+        stop_on_double_eol=False,
+        stop_on_eol=False,
+        prevent_newline_after_colon=True
+        ):
+    """Main token generation function.
+
+    Args:
+        model: no interleaving is supported.
+        tokens: prompt tokens extended to be of size [b, max-sequence-length]
+        lengths: original prompt length, size: [b]
+        return_output_log_probs: flag to calculate the log probability of
+            the generated tokens. Note that the log probability is the one
+            from the original logit.
+        top_k, top_p: top-k and top-p sampling parameters.
+            Note that top-k = 1 is gready. Also, these paramters are
+            exclusive meaning that:
+                if top-k > 0 then we expect top-p=0.
+                if top-p > 0 then we check for top-k=0.
+        temperature: sampling temperature.
+        use_eod_token_for_early_termination: if True, do early termination if
+            all the sequences have reached this token.
+        prevent_newline_after_colon: if True, it will disable generating new line \n after :
+    Note: Outside of model, other parameters only need to be available on
+          rank 0.
+
+    Returns: Note that is size is adjusted to a lower value than
+             max-sequence-length if generation is terminated early.
+        tokens: prompt and generated tokens. size: [b, :]
+        generated_sequence_lengths: total length (including prompt) of
+            the generated sequence. size: [b]
+        output_log_probs: log probability of the selected tokens. size: [b, s]
+    """
+
+    args = get_args()
+    tokenizer = get_tokenizer()
+
+    batch_size = tokens.size(0)
+    min_prompt_length = lengths.min().item()
+    max_sequence_length = tokens.size(1)
+
+    if max_sequence_length > args.max_position_embeddings:
+        raise ValueError("Length of prompt + tokens_to_generate longer than allowed")
+    
+    if max_sequence_length * batch_size > args.max_tokens_to_oom:
+        raise ValueError("Too many tokens.  " + str(max_sequence_length*batch_size)+ " is greater than "+str(args.max_tokens_to_oom))
+
+    # forward step.
+    forward_step = ForwardStep(model, batch_size, max_sequence_length)
+
+    # Added termination_id to support the case that we want to terminate the
+    # generation once that id is generated.
+    if hasattr(args, 'eos_id'):
+        termination_id = args.eos_id
+    else:
+        termination_id = tokenizer.eod
+
+    # ===================
+    # Pre-allocate memory
+    # ===================
+
+    # Log probability of the sequence (prompt + generated tokens).
+    output_log_probs = None
+    output_log_probs_size = (batch_size, max_sequence_length - 1)
+    # Lengths of generated seuquence including including prompts.
+    generated_sequence_lengths = None
+    if mpu.is_pipeline_last_stage():
+        if return_output_log_probs:
+            output_log_probs = torch.empty(output_log_probs_size,
+                                           dtype=torch.float32,
+                                           device=torch.cuda.current_device())
+        generated_sequence_lengths = torch.ones(
+                batch_size, dtype=torch.int64,
+                device=torch.cuda.current_device()) * max_sequence_length
+    
+    # Whether we have reached a termination id.
+    is_generation_done = torch.zeros(batch_size, dtype=torch.uint8,
+                                     device=torch.cuda.current_device())
+
+    # =============
+    # Run infernece
+    # =============
+
+    with torch.no_grad():
+        attention_mask, position_ids = _build_attention_mask_and_position_ids(
+            tokens)
+        prev_context_length = 0
+        for context_length in range(min_prompt_length, max_sequence_length):
+
+            # Pick the slice that we need to pass through the network.
+            tokens2use = tokens[:, prev_context_length:context_length]
+            positions2use = position_ids[:, prev_context_length:context_length]
+            attention_mask2use = attention_mask[
+                ..., prev_context_length:context_length, :context_length]
+
+            # logits will be meanigful only in the last pipeline stage.
+            logits = forward_step(tokens2use, positions2use, attention_mask2use)
+
+            if mpu.is_pipeline_last_stage():
+                if prevent_newline_after_colon:
+                    logits[tokens2use[:, -1] == tokenizer.tokenize(':')[0], -1, tokenizer.tokenize('\n')[0]] = -1e10 # disable "\n" after ":"
+                # Always the last stage should have an output.
+                assert logits is not None
+
+                # Sample.
+                last_token_logits = logits[:, -1, :]
+                new_sample = sample(last_token_logits,
+                                    top_k=top_k,
+                                    top_p=top_p,
+                                    temperature=temperature,
+                                    vocab_size=tokenizer.vocab_size)
+                if top_p > 0.0 and top_p_decay > 0.0:
+                    top_p = top_p * top_p_decay
+                    if top_p_bound > 0.0:
+                        top_p = max(top_p, top_p_bound)
+
+                # If a prompt length is smaller or equal th current context
+                # length, it means we have started generating tokens
+                started = lengths <= context_length
+                # Update the tokens.
+                tokens[started, context_length] = new_sample[started]
+
+                # Calculate the log probabilities.
+                if return_output_log_probs:
+                    log_probs = F.log_softmax(logits, dim=2)
+                    if return_output_log_probs:
+                        # Pick the tokens that we need to get the log
+                        # probabilities for. Note that next input token is
+                        # the token which we selected in the current logits,
+                        # so shift by 1.
+                        indices = torch.unsqueeze(
+                            tokens[
+                                :,
+                                (prev_context_length + 1):(context_length + 1)],
+                            2)
+                        output_log_probs[:,
+                                         prev_context_length:context_length] = \
+                            torch.gather(log_probs, 2, indices).squeeze(2)
+
+            # Update the tokens on the first stage so the next input to
+            # the network is correct.
+            copy_from_last_to_first_pipeline_stage(batch_size, torch.int64,
+                                                   tokens[:, context_length])
+
+            # Update the context length for the next token generation.
+            prev_context_length = context_length
+
+            # Check if all the sequences have hit the termination_id.
+            done = None
+            if mpu.is_pipeline_last_stage():
+                # TODO(rprenger) These stopping methods are tokenizer dependent
+                # instead tokenization should be in the inference loop so stop sequences can be used
+                if stop_on_double_eol:
+                    hit_double_eol = (new_sample == 628).byte() & started.byte()
+                    hit_two_eols = (new_sample == 198).byte() & (tokens[:, context_length-1] == 198).byte() & started.byte()
+                    done_token = hit_double_eol | hit_two_eols
+                elif stop_on_eol:
+                    hit_double_eol = (new_sample == 628).byte() & started.byte()
+                    hit_eol = (new_sample == 198).byte() & started.byte()
+                    done_token = hit_double_eol | hit_eol
+                else: 
+                    done_token = (new_sample == termination_id).byte() & \
+                        started.byte()
+                
+                just_finished = (done_token & ~is_generation_done).bool()
+                generated_sequence_lengths[just_finished.view(-1)] = \
+                    context_length + 1
+                is_generation_done = is_generation_done | done_token
+                done = torch.all(is_generation_done)
+            done = broadcast_from_last_pipeline_stage(1, torch.uint8,
+                                                      tensor=done)
+            if use_eod_token_for_early_termination and done:
+                break
+            
+    # ===================================================
+    # Update the length of based on max generated length.
+    # ===================================================
+
+    tokens = tokens[:, :(context_length + 1)]
+    if mpu.is_pipeline_last_stage():
+        if return_output_log_probs:
+            output_log_probs = output_log_probs[:, :context_length]
+
+    # ======================================
+    # Broadcast to the first pipeline stage.
+    # ======================================
+
+    generated_sequence_lengths = broadcast_from_last_to_first_pipeline_stage(
+        batch_size, torch.int64, generated_sequence_lengths)
+    if return_output_log_probs:
+        output_log_probs_size = (batch_size, context_length)
+        output_log_probs = broadcast_from_last_to_first_pipeline_stage(
+            output_log_probs_size, torch.float32, output_log_probs)
+
+    return tokens, generated_sequence_lengths, output_log_probs, None
+
+def beam_search_and_return_on_first_stage(model, tokens, lengths, beam_size, stop_token, num_return_gen, length_penalty, prevent_newline_after_colon=True):
+    args = get_args()
+    tokenizer = get_tokenizer()
+
+    batch_size = tokens.size(0)
+    assert(batch_size == 1)
+    prompt_length = lengths.item()
+    final_sequence_length = tokens.size(1)
+    final_sequence_length = min(final_sequence_length, args.max_position_embeddings)
+    
+    # If the context is too big, this happens
+    if prompt_length >= final_sequence_length:
+        raise ValueError("context length + tokens_to_generate too large")
+
+    # forward step.
+    forward_step = ForwardStep(model, beam_size, final_sequence_length)
+
+    beam_hyp = BeamHypotheses(beam_size, length_penalty)
+    best_batches = None
+    done = torch.zeros(1, dtype=torch.uint8, device=torch.cuda.current_device())
+    scores = torch.zeros(beam_size,
+                         dtype=torch.float32,
+                         device=torch.cuda.current_device()).unsqueeze(1)
+    scores_size_tensor, tokens_size_tensor = None, None
+    # =============
+    # Run infernece
+    # =============
+    with torch.no_grad():
+        tokens = tokens.repeat(beam_size, 1)
+        attention_mask, position_ids = _build_attention_mask_and_position_ids(tokens)
+        prev_context_length = 0
+        for context_length in range(prompt_length, final_sequence_length):
+
+            # Pick the slice that we need to pass through the network.
+            tokens2use = tokens[:, prev_context_length:context_length]
+            positions2use = position_ids[:, prev_context_length:context_length]
+            attention_mask2use = attention_mask[
+                ..., prev_context_length:context_length, :context_length]
+
+            # logits will be meanigful only in the last pipeline stage.
+            logits = forward_step(tokens2use, positions2use, attention_mask2use)
+
+            if mpu.is_pipeline_last_stage():
+                if prevent_newline_after_colon:
+                    logits[tokens2use[:, -1] == tokenizer.tokenize(':')[0], -1, tokenizer.tokenize('\n')[0]] = -1e10 # disable "\n" after ":"
+                vocab_size = logits.size(2)
+                log_probs = F.log_softmax(logits, dim=2)
+                new_scores = log_probs[:, -1, :] + scores
+
+                if context_length == prompt_length:  # if this is the first one
+                    sorted_scores, indices = torch.sort(new_scores[0,:], descending=True)
+                else:
+                    sorted_scores, indices = torch.sort(new_scores.view(-1), descending=True)
+
+                best_beam_ids = torch.div(indices[: 2 * beam_size], vocab_size).trunc().long()
+                best_words = indices[:2 * beam_size] % vocab_size
+                best_scores = sorted_scores[: 2 * beam_size]
+
+                next_beams = []
+                for beam_token_rank, (token_id, beam_score, beam_id) in enumerate(
+                    zip(best_words, best_scores, best_beam_ids)
+                ):
+                    if token_id.item() == stop_token:
+                        # if beam_token does not belong to top num_beams tokens, it should not be added
+                        is_beam_token_worse_than_top_num_beams = beam_token_rank >= beam_size
+                        if is_beam_token_worse_than_top_num_beams:
+                            continue
+                        beam_hyp.add(
+                            tokens[beam_id].clone(),
+                            beam_score,
+                            context_length + 1 - prompt_length
+                        )
+                    else:
+                        # add next predicted token since it is not eos_token
+                        next_beams.append((token_id, beam_score, beam_id))
+
+                    if len(next_beams) == beam_size:
+                        break
+
+                if beam_hyp.is_done(best_scores.max().item(), context_length + 1 - prompt_length):
+                    done = torch.ones(1, dtype=torch.uint8, device=torch.cuda.current_device())
+            
+                best_batches = tokens.new([item[2] for item in next_beams])
+                tokens = tokens[best_batches,:]
+                tokens[:, context_length] = tokens.new([item[0] for item in next_beams])
+                scores = scores.new([item[1] for item in next_beams]).unsqueeze(1)
+          
+            # torch.distributed.barrier()
+            done = broadcast_from_last_pipeline_stage(1, torch.uint8, done)
+            if done:
+                break
+
+            # Update the tokens on the first stage so the next input to
+            # the network is correct.
+            copy_from_last_to_first_pipeline_stage(tokens.size(), torch.int64,
+                                                   tokens)
+
+            # set inference key values to make it consistent with best beam index
+            best_batches = broadcast_from_last_pipeline_stage(beam_size, torch.int64, best_batches)
+            forward_step.inference_params.swap_key_value_dict(best_batches)
+
+            # Update the context length for the next token generation.
+            prev_context_length = context_length
+
+        if mpu.is_pipeline_last_stage():
+            # if cannot find stop token, add open beams to hyps
+            if not done:
+                for beam_id in range(beam_size):
+                    beam_hyp.add(tokens[beam_id].clone(), scores[beam_id].squeeze(), context_length + 1 - prompt_length)
+
+            # rank based on scores
+            sorted_hyps = sorted(beam_hyp.beams, key=lambda x: x[0], reverse=True)
+            num_return_gen = min(num_return_gen, len(sorted_hyps))
+            scores = [sorted_hyps[i][0] for i in range(num_return_gen)]
+            tokens = [sorted_hyps[i][1] for i in range(num_return_gen)]
+            scores = torch.stack(scores, dim=0)
+            tokens = torch.stack(tokens, dim=0)
+            scores_size_tensor = torch.tensor(scores.shape, dtype=torch.int64, device=torch.cuda.current_device())
+            tokens_size_tensor = torch.tensor(tokens.shape, dtype=torch.int64, device=torch.cuda.current_device())
+
+        scores_size_tensor = broadcast_from_last_pipeline_stage(1, torch.int64, scores_size_tensor)
+        tokens_size_tensor = broadcast_from_last_pipeline_stage(2, torch.int64, tokens_size_tensor)
+
+        scores = broadcast_from_last_to_first_pipeline_stage(tuple(scores_size_tensor), torch.float32, scores)
+        tokens = broadcast_from_last_to_first_pipeline_stage(tuple(tokens_size_tensor), torch.int64, tokens)
+
+    return tokens, scores
+
+
+def _build_attention_mask_and_position_ids(tokens):
+    """Build the attention mask and postition ids for the input tokens."""
+
+    # Since we are not interested in loss-mask and reset attention/position
+    # is also False, eod_token is not used so it is safe to set it to None.
+    attention_mask, _, position_ids = get_ltor_masks_and_position_ids(
+        data=tokens,
+        eod_token=None,
+        reset_position_ids=False,
+        reset_attention_mask=False,
+        eod_mask_loss=False)
+
+    return attention_mask, position_ids

Megatron-LM-core_r0.7.0.beta/megatron/inference/text_generation/sampling.py

+# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
+
+"""Sampling utilities.
+Part of this code is inspired by:
+ - https://github.com/ari-holtzman/degen/blob/master/gen.py
+ - https://huggingface.co/transformers/_modules/transformers/generation_logits_process.html
+"""
+
+
+import torch
+
+
+
+def modify_logits_for_top_k_filtering(logits, top_k):
+    """Set the logits for none top-k values to -inf."""
+
+    filter_ = logits < torch.topk(logits, top_k)[0][..., -1, None]
+    logits.masked_fill_(filter_, float('-Inf'))
+
+
+
+def modify_logits_for_top_p_filtering(logits, top_p):
+    """Set the logits for none top-p values to -inf."""
+
+    # First sort and calculate cumulative sum of probabilities.
+    sorted_logits, sorted_indices = torch.sort(logits, descending=True)
+    cumulative_probs = sorted_logits.softmax(dim=-1).cumsum(dim=-1)
+
+    # Filteration based on the cumulative sum.
+    filter_ = cumulative_probs > top_p
+    # This shift by 1 is weird and I cannot justify it. This existed
+    # in the original implementation:
+    #   https://github.com/ari-holtzman/degen/blob/master/gen.py
+    # and I guess it is needed so keeping it for now.
+    filter_[:, 1:] = filter_[:, :-1].clone()
+    # Make sure we at least have one token to select from.
+    filter_[..., 0] = 0
+
+    # Fill in the filtered part
+    filter_ = filter_.scatter(1, sorted_indices, filter_)
+    logits.masked_fill_(filter_, float('-Inf'))
+
+
+
+def sample(logits, top_k=0, top_p=0.0, temperature=1.0, vocab_size=None):
+    """ Sample and generate a token.
+    Note: logits has the dimension [b, v] where b is the batch size
+          and v is the vocabulary size.
+    If vocab_size is provided, we will make sure the sample that is
+    generated is in [0, vocab-size). This will avoid out of vocabulary
+    generations due to padding.
+    """
+
+    # Check logits for consistency.
+    assert logits.ndim == 2, 'expected the logits to be of [b, v] shape.'
+    assert logits.type() == 'torch.cuda.FloatTensor', \
+        'input logits should be floats.'
+
+
+    # Greedy is just simple argmax.
+    if top_k == 1:
+        assert top_p == 0.0, 'cannot set both greedy and top-p samplings.'
+        samples = torch.argmax(logits, dim=-1)
+
+    # Top-k or top-p sampling.
+    else:
+        # Clone so we do not modify the inputs,
+        logits = logits.clone()
+        # Apply temperature in place.
+        if temperature != 1.0:
+            logits.div_(temperature)
+
+        if top_k > 1:
+            assert top_p == 0.0, 'cannot set both top-k and top-p samplings.'
+            assert top_k <= logits.size(1), 'top-k is larger than logit size.'
+            if vocab_size:
+                assert top_k < vocab_size, 'top-k is larger than vocab size.'
+            modify_logits_for_top_k_filtering(logits, top_k)
+
+        elif top_p > 0.0:
+            assert top_p <= 1.0, 'top-p should be in (0, 1].'
+            modify_logits_for_top_p_filtering(logits, top_p)
+
+        # After filtering, we need to recalculate the distribution.
+        probs = logits.softmax(dim=-1)
+        samples = torch.multinomial(probs, num_samples=1).view(-1)
+
+    # If vocab size is provided, make sure the samples are in
+    # in the range [0, vocab-size).
+    if vocab_size:
+        samples = torch.clamp(samples, min=0, max=(vocab_size - 1))
+
+    return samples

Megatron-LM-core_r0.7.0.beta/megatron/inference/text_generation/tokenization.py

+# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
+
+"""Tokenization utilities."""
+
+
+import torch
+
+
+from megatron.training import get_tokenizer, get_args
+from .communication import broadcast_int_list, broadcast_tensor
+
+
+def detokenize_generations(tokens_gpu_tensor,
+                           lengths_gpu_tensor,
+                           return_segments):
+    """Detokenize the generated tokens."""
+
+    tokenizer = get_tokenizer()
+    args = get_args()
+    prompts_plus_generations = []
+    if return_segments:
+        prompts_plus_generations_segments = []
+
+    tokens = tokens_gpu_tensor.cpu().numpy().tolist()
+    lengths = lengths_gpu_tensor.cpu().numpy().tolist()
+    for sequence_tokens, length in zip(tokens, lengths):
+        sequence_tokens = sequence_tokens[:length]
+        prompts_plus_generations.append(
+            tokenizer.detokenize(sequence_tokens))
+        if return_segments:
+            words = []
+            for token in sequence_tokens:
+                if args.tokenizer_type in ['SentencePieceTokenizer', 
+                                           'GPTSentencePieceTokenizer',
+                                           'Llama2Tokenizer']:
+                    word = tokenizer.decoder[token]
+                elif args.tokenizer_type == 'NullTokenizer':
+                    word = str(token)
+                else:
+                    word = tokenizer.tokenizer.decoder[token]
+                    word = bytearray(
+                        [tokenizer.tokenizer.byte_decoder[c] for c in word]).decode(
+                            'utf-8', errors='replace')
+                words.append(word)
+            prompts_plus_generations_segments.append(words)
+
+    if return_segments:
+        return tokens, prompts_plus_generations, \
+            prompts_plus_generations_segments
+
+    return tokens, prompts_plus_generations
+
+
+def tokenize_prompts(prompts=None, tokens_to_generate=None,
+                     add_BOS=None, rank=0):
+    """Tokenize prompts and make them avaiable on all ranks."""
+
+    # On all ranks set to None so we can pass them to functions
+    sizes_list = None
+    prompts_tokens_cuda_long_tensor = None
+    prompts_length_cuda_long_tensor = None
+
+    # On the specified rank, build the above.
+    if torch.distributed.get_rank() == rank:
+        assert prompts is not None
+        assert tokens_to_generate is not None
+        # Tensor of tokens padded and their unpadded length.
+        prompts_tokens_cuda_long_tensor, prompts_length_cuda_long_tensor = \
+            _tokenize_prompts_and_batch(prompts, tokens_to_generate, add_BOS)
+        # We need the sizes of these tensors for the boradcast
+        sizes_list = [prompts_tokens_cuda_long_tensor.size(0), # Batch size
+                      prompts_tokens_cuda_long_tensor.size(1)] # Sequence lenght
+
+    # First, broadcast the sizes.
+    sizes_tensor = broadcast_int_list(2, int_list=sizes_list, rank=rank)
+
+    # Now that we have the sizes, we can boradcast the tokens
+    # and length tensors.
+    sizes = sizes_tensor.tolist()
+    prompts_tokens_cuda_long_tensor = broadcast_tensor(
+        sizes, torch.int64, tensor=prompts_tokens_cuda_long_tensor, rank=rank)
+    prompts_length_cuda_long_tensor = broadcast_tensor(
+        sizes[0], torch.int64, tensor=prompts_length_cuda_long_tensor,
+        rank=rank)
+
+    return prompts_tokens_cuda_long_tensor, prompts_length_cuda_long_tensor
+
+
+def _tokenize_prompts_and_batch(prompts, tokens_to_generate, add_BOS):
+    """Given a set of prompts and number of tokens to generate:
+        - tokenize prompts
+        - set the sequence length to be the max of length of prompts
+          plus the number of tokens we would like to generate
+        - pad all the sequences to this length so we can convert them
+          into a 2D tensor.
+    """
+
+    # Tokenize all the prompts.
+    tokenizer = get_tokenizer()
+    if add_BOS:
+        prompts_tokens = [[tokenizer.eod] + tokenizer.tokenize(prompt)
+                          for prompt in prompts]
+    else:
+        prompts_tokens = [tokenizer.tokenize(prompt) for prompt in prompts]
+
+    # Now we have a list of list of tokens which each list has a different
+    # size. We want to extend this list to:
+    #   - incorporate the tokens that need to be generated
+    #   - make all the sequences equal length.
+    # Get the prompts length.
+    prompts_length = [len(prompt_tokens) for prompt_tokens in prompts_tokens]
+    # Get the max prompts length.
+    max_prompt_len = max(prompts_length)
+    # Number of tokens in the each sample of the batch.
+    samples_length = max_prompt_len + tokens_to_generate
+    # Now update the list of list to be of the same size: samples_length.
+    for prompt_tokens, prompt_length in zip(prompts_tokens, prompts_length):
+        padding_size = samples_length - prompt_length
+        prompt_tokens.extend([tokenizer.eod] * padding_size)
+
+    # Now we are in a structured format, we can convert to tensors.
+    prompts_tokens_tensor = torch.tensor(prompts_tokens, dtype=torch.long, device='cuda')
+    prompts_length_tensor = torch.tensor(prompts_length, dtype=torch.long, device='cuda')
+
+    return prompts_tokens_tensor, prompts_length_tensor

Megatron-LM-core_r0.7.0.beta/megatron/inference/text_generation_server.py

+# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
+import datetime
+import torch
+import json
+import threading
+from flask import Flask, request, jsonify, current_app
+from flask_restful import Resource, Api
+from megatron.training import get_args
+from megatron.inference.text_generation import generate_and_post_process
+from megatron.inference.text_generation import beam_search_and_post_process
+
+
+GENERATE_NUM = 0
+BEAM_NUM = 1
+lock = threading.Lock()
+
+class MegatronGenerate(Resource):
+    def __init__(self, model):
+        self.model = model
+
+    @staticmethod
+    def send_do_generate():
+        choice = torch.tensor([GENERATE_NUM], dtype=torch.long, device='cuda')
+        torch.distributed.broadcast(choice, 0)
+     
+    @staticmethod
+    def send_do_beam_search():
+        choice = torch.tensor([BEAM_NUM], dtype=torch.long, device='cuda')
+        torch.distributed.broadcast(choice, 0)
+    
+    def put(self):
+        args = get_args()
+       
+        if not "prompts" in request.get_json():
+            return "prompts argument required", 400
+        
+        if "max_len" in request.get_json():
+            return "max_len is no longer used.  Replace with tokens_to_generate", 400
+        
+        if "sentences" in request.get_json():
+            return "sentences is no longer used.  Replace with prompts", 400
+
+        prompts = request.get_json()["prompts"]
+        if not isinstance(prompts, list):
+            return "prompts is not a list of strings", 400
+
+        if len(prompts) == 0:
+            return "prompts is empty", 400
+        
+        if len(prompts) > 128:
+            return "Maximum number of prompts is 128", 400
+        
+        tokens_to_generate = 64  # Choosing hopefully sane default.  Full sequence is slow
+        if "tokens_to_generate" in request.get_json():
+            tokens_to_generate = request.get_json()["tokens_to_generate"]
+            if not isinstance(tokens_to_generate, int):
+                return "tokens_to_generate must be an integer greater than 0"
+            if tokens_to_generate < 0:
+                return "tokens_to_generate must be an integer greater than or equal to 0"
+
+        logprobs = False
+        if "logprobs" in request.get_json():
+            logprobs = request.get_json()["logprobs"]
+            if not isinstance(logprobs, bool):
+                return "logprobs must be a boolean value"
+        
+        if tokens_to_generate == 0 and not logprobs:
+            return "tokens_to_generate=0 implies logprobs should be True"
+        
+        temperature = 1.0
+        if "temperature" in request.get_json():
+            temperature = request.get_json()["temperature"]
+            if not (type(temperature) == int or type(temperature) == float):
+                return "temperature must be a positive number less than or equal to 100.0"
+            if not (0.0 < temperature <= 100.0):
+                return "temperature must be a positive number less than or equal to 100.0"
+        
+        top_k = 0.0
+        if "top_k" in request.get_json():
+            top_k = request.get_json()["top_k"]
+            if not (type(top_k) == int):
+                return "top_k must be an integer equal to or greater than 0 and less than or equal to 1000"
+            if not (0 <= top_k <= 1000):
+                return "top_k must be equal to or greater than 0 and less than or equal to 1000"
+        
+        top_p = 0.0
+        if "top_p" in request.get_json():
+            top_p = request.get_json()["top_p"]
+            if not (type(top_p) == float):
+                return "top_p must be a positive float less than or equal to 1.0"
+            if top_p > 0.0 and top_k > 0.0:
+                return "cannot set both top-k and top-p samplings."
+            if not (0 <= top_p <= 1.0):
+                return "top_p must be less than or equal to 1.0"
+        
+        top_p_decay = 0.0
+        if "top_p_decay" in request.get_json():
+            top_p_decay = request.get_json()["top_p_decay"]
+            if not (type(top_p_decay) == float):
+                return "top_p_decay must be a positive float less than or equal to 1.0"
+            if top_p == 0.0:
+                return "top_p_decay cannot be set without top_p"
+            if not (0 <= top_p_decay <= 1.0):
+                return "top_p_decay must be less than or equal to 1.0"
+        
+        top_p_bound = 0.0
+        if "top_p_bound" in request.get_json():
+            top_p_bound = request.get_json()["top_p_bound"]
+            if not (type(top_p_bound) == float):
+                return "top_p_bound must be a positive float less than or equal to top_p"
+            if top_p == 0.0:
+                return "top_p_bound cannot be set without top_p"
+            if not (0.0 < top_p_bound <= top_p):
+                return "top_p_bound must be greater than 0 and less than top_p"
+        
+        add_BOS = False
+        if "add_BOS" in request.get_json():
+            add_BOS = request.get_json()["add_BOS"]
+            if not isinstance(add_BOS, bool):
+                return "add_BOS must be a boolean value"
+        
+        if any([len(prompt) == 0 for prompt in prompts]) and not add_BOS:
+            return "Empty prompts require add_BOS=true"
+
+        stop_on_double_eol = False
+        if "stop_on_double_eol" in request.get_json():
+            stop_on_double_eol = request.get_json()["stop_on_double_eol"]
+            if not isinstance(stop_on_double_eol, bool):
+                return "stop_on_double_eol must be a boolean value"
+        
+        stop_on_eol = False
+        if "stop_on_eol" in request.get_json():
+            stop_on_eol = request.get_json()["stop_on_eol"]
+            if not isinstance(stop_on_eol, bool):
+                return "stop_on_eol must be a boolean value"
+
+        prevent_newline_after_colon = False
+        if "prevent_newline_after_colon" in request.get_json():
+            prevent_newline_after_colon = request.get_json()["prevent_newline_after_colon"]
+            if not isinstance(prevent_newline_after_colon, bool):
+                return "prevent_newline_after_colon must be a boolean value"
+
+        random_seed = -1
+        if "random_seed" in request.get_json():
+            random_seed = request.get_json()["random_seed"]
+            if not isinstance(random_seed, int):
+                return "random_seed must be integer"
+            if random_seed < 0: 
+                return "random_seed must be a positive integer"
+
+        no_log = False
+        if "no_log" in request.get_json():
+            no_log = request.get_json()["no_log"]
+            if not isinstance(no_log, bool):
+                return "no_log must be a boolean value"
+        
+        beam_width = None
+        if "beam_width" in request.get_json():
+            beam_width = request.get_json()["beam_width"]
+            if not isinstance(beam_width, int):
+                return "beam_width must be integer"
+            if beam_width < 1:
+                return "beam_width must be an integer > 1"
+            if len(prompts) > 1:
+                return "When doing beam_search, batch size must be 1"
+
+        stop_token=50256
+        if "stop_token" in request.get_json():
+            stop_token = request.get_json()["stop_token"]
+            if not isinstance(stop_token, int):
+                return "stop_token must be an integer"
+        
+        length_penalty = 1 
+        if "length_penalty" in request.get_json():
+            length_penalty = request.get_json()["length_penalty"]
+            if not isinstance(length_penalty, float):
+                return "length_penalty must be a float"
+        
+        with lock:  # Need to get lock to keep multiple threads from hitting code
+            
+            if not no_log:
+                print("request IP: " + str(request.remote_addr))
+                print(json.dumps(request.get_json()),flush=True)
+                print("start time: ", datetime.datetime.now())
+            
+            try:
+                if beam_width is not None:
+                    MegatronGenerate.send_do_beam_search()  # Tell other ranks we're doing beam_search
+                    response, response_seg, response_scores = \
+                        beam_search_and_post_process(
+                        self.model,
+                        prompts=prompts,
+                        tokens_to_generate=tokens_to_generate,
+                        beam_size = beam_width,
+                        add_BOS=add_BOS,
+                        stop_token=stop_token,
+                        num_return_gen=beam_width,  # Returning whole beam
+                        length_penalty=length_penalty,
+                        prevent_newline_after_colon=prevent_newline_after_colon
+                        )
+                    
+                    return jsonify({"text": response,
+                        "segments": response_seg,
+                        "scores": response_scores})
+                else:
+                    MegatronGenerate.send_do_generate()  # Tell other ranks we're doing generate
+                    response, response_seg, response_logprobs, _ = \
+                        generate_and_post_process(
+                        self.model,
+                        prompts=prompts,
+                        tokens_to_generate=tokens_to_generate,
+                        return_output_log_probs=logprobs,
+                        top_k_sampling=top_k,
+                        top_p_sampling=top_p,
+                        top_p_decay=top_p_decay,
+                        top_p_bound=top_p_bound,
+                        temperature=temperature,
+                        add_BOS=add_BOS,
+                        use_eod_token_for_early_termination=True,
+                        stop_on_double_eol=stop_on_double_eol,
+                        stop_on_eol=stop_on_eol,
+                        prevent_newline_after_colon=prevent_newline_after_colon,
+                        random_seed=random_seed)
+
+                    return jsonify({"text": response,
+                        "segments": response_seg,
+                        "logprobs": response_logprobs})
+
+            except ValueError as ve:
+                return ve.args[0]
+            print("end time: ", datetime.datetime.now())
+        
+
+class MegatronServer(object):
+    def __init__(self, model):
+        self.app = Flask(__name__, static_url_path='')
+        api = Api(self.app)
+        api.add_resource(MegatronGenerate, '/api', resource_class_args=[model])
+        
+    def run(self, url, port): 
+        self.app.run(url, threaded=True, debug=False, port=port)

Megatron-LM-core_r0.7.0.beta/megatron/legacy/data/autoaugment.py

+"""AutoAugment data augmentation policy for ImageNet.
+
+-- Begin license text.
+
+MIT License
+
+Copyright (c) 2018 Philip Popien
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
+
+-- End license text.
+
+Code adapted from https://github.com/DeepVoltaire/AutoAugment.
+
+This module implements the fixed AutoAugment data augmentation policy for ImageNet provided in
+Appendix A, Table 9 of reference [1]. It does not include any of the search code for augmentation
+policies.
+
+Reference:
+[1] https://arxiv.org/abs/1805.09501
+"""
+
+import random
+
+import numpy as np
+from PIL import Image
+from PIL import ImageEnhance
+from PIL import ImageOps
+
+_MAX_LEVEL = 10  # Maximum integer strength of an augmentation, if applicable.
+
+
+class ImageNetPolicy:
+    """Definition of an ImageNetPolicy.
+
+    Implements a fixed AutoAugment data augmentation policy targeted at
+    ImageNet training by randomly applying at runtime one of the 25 pre-defined
+    data augmentation sub-policies provided in Reference [1].
+
+    Usage example as a Pytorch Transform:
+    >>> transform=transforms.Compose([transforms.Resize(256),
+    >>>                               ImageNetPolicy(),
+    >>>                               transforms.ToTensor()])
+    """
+
+    def __init__(self, fillcolor=(128, 128, 128)):
+        """Initialize an ImageNetPolicy.
+
+        Args:
+            fillcolor (tuple): RGB color components of the color to be used for
+            filling when needed (default: (128, 128, 128), which
+            corresponds to gray).
+        """
+        # Instantiate a list of sub-policies.
+        # Each entry of the list is a SubPolicy which consists of
+        # two augmentation operations,
+        # each of those parametrized as operation, probability, magnitude.
+        # Those two operations are applied sequentially on the image upon call.
+        self.policies = [
+            SubPolicy("posterize", 0.4, 8, "rotate", 0.6, 9, fillcolor),
+            SubPolicy("solarize", 0.6, 5, "autocontrast", 0.6, 5, fillcolor),
+            SubPolicy("equalize", 0.8, 8, "equalize", 0.6, 3, fillcolor),
+            SubPolicy("posterize", 0.6, 7, "posterize", 0.6, 6, fillcolor),
+            SubPolicy("equalize", 0.4, 7, "solarize", 0.2, 4, fillcolor),
+            SubPolicy("equalize", 0.4, 4, "rotate", 0.8, 8, fillcolor),
+            SubPolicy("solarize", 0.6, 3, "equalize", 0.6, 7, fillcolor),
+            SubPolicy("posterize", 0.8, 5, "equalize", 1.0, 2, fillcolor),
+            SubPolicy("rotate", 0.2, 3, "solarize", 0.6, 8, fillcolor),
+            SubPolicy("equalize", 0.6, 8, "posterize", 0.4, 6, fillcolor),
+            SubPolicy("rotate", 0.8, 8, "color", 0.4, 0, fillcolor),
+            SubPolicy("rotate", 0.4, 9, "equalize", 0.6, 2, fillcolor),
+            SubPolicy("equalize", 0.0, 7, "equalize", 0.8, 8, fillcolor),
+            SubPolicy("invert", 0.6, 4, "equalize", 1.0, 8, fillcolor),
+            SubPolicy("color", 0.6, 4, "contrast", 1.0, 8, fillcolor),
+            SubPolicy("rotate", 0.8, 8, "color", 1.0, 2, fillcolor),
+            SubPolicy("color", 0.8, 8, "solarize", 0.8, 7, fillcolor),
+            SubPolicy("sharpness", 0.4, 7, "invert", 0.6, 8, fillcolor),
+            SubPolicy("shearX", 0.6, 5, "equalize", 1.0, 9, fillcolor),
+            SubPolicy("color", 0.4, 0, "equalize", 0.6, 3, fillcolor),
+            SubPolicy("equalize", 0.4, 7, "solarize", 0.2, 4, fillcolor),
+            SubPolicy("solarize", 0.6, 5, "autocontrast", 0.6, 5, fillcolor),
+            SubPolicy("invert", 0.6, 4, "equalize", 1.0, 8, fillcolor),
+            SubPolicy("color", 0.6, 4, "contrast", 1.0, 8, fillcolor),
+            SubPolicy("equalize", 0.8, 8, "equalize", 0.6, 3, fillcolor),
+        ]
+
+    def __call__(self, img):
+        """Define call method for ImageNetPolicy class."""
+        policy_idx = random.randint(0, len(self.policies) - 1)
+        return self.policies[policy_idx](img)
+
+    def __repr__(self):
+        """Define repr method for ImageNetPolicy class."""
+        return "ImageNetPolicy"
+
+
+class SubPolicy:
+    """Definition of a SubPolicy.
+
+    A SubPolicy consists of two augmentation operations,
+    each of those parametrized as operation, probability, magnitude.
+    The two operations are applied sequentially on the image upon call.
+    """
+
+    def __init__(
+        self,
+        operation1,
+        probability1,
+        magnitude_idx1,
+        operation2,
+        probability2,
+        magnitude_idx2,
+        fillcolor,
+    ):
+        """Initialize a SubPolicy.
+
+        Args:
+            operation1 (str): Key specifying the first augmentation operation.
+            There are fourteen key values altogether (see supported_ops below
+            listing supported operations). probability1 (float): Probability
+            within [0., 1.] of applying the first augmentation operation.
+            magnitude_idx1 (int): Integer specifiying the strength of the first
+            operation as an index further used to derive the magnitude from a
+            range of possible values.
+            operation2 (str): Key specifying the second augmentation operation.
+            probability2 (float): Probability within [0., 1.] of applying the
+            second augmentation operation.
+            magnitude_idx2 (int): Integer specifiying the strength of the
+            second operation as an index further used to derive the magnitude
+            from a range of possible values.
+            fillcolor (tuple): RGB color components of the color to be used for
+            filling.
+        Returns:
+        """
+        # List of supported operations for operation1 and operation2.
+        supported_ops = [
+            "shearX",
+            "shearY",
+            "translateX",
+            "translateY",
+            "rotate",
+            "color",
+            "posterize",
+            "solarize",
+            "contrast",
+            "sharpness",
+            "brightness",
+            "autocontrast",
+            "equalize",
+            "invert",
+        ]
+        assert (operation1 in supported_ops) and (
+            operation2 in supported_ops
+        ), "SubPolicy:one of oper1 or oper2 refers to an unsupported operation."
+
+        assert (
+            0.0 <= probability1 <= 1.0 and 0.0 <= probability2 <= 1.0
+        ), "SubPolicy: prob1 and prob2 should be within [0., 1.]."
+
+        assert (
+            isinstance(magnitude_idx1, int) and 0 <= magnitude_idx1 <= 10
+        ), "SubPolicy: idx1 should be specified as an integer within [0, 10]."
+
+        assert (
+            isinstance(magnitude_idx2, int) and 0 <= magnitude_idx2 <= 10
+        ), "SubPolicy: idx2 should be specified as an integer within [0, 10]."
+
+        # Define a dictionary where each key refers to a specific type of
+        # augmentation and the corresponding value is a range of ten possible
+        # magnitude values for that augmentation.
+        num_levels = _MAX_LEVEL + 1
+        ranges = {
+            "shearX": np.linspace(0, 0.3, num_levels),
+            "shearY": np.linspace(0, 0.3, num_levels),
+            "translateX": np.linspace(0, 150 / 331, num_levels),
+            "translateY": np.linspace(0, 150 / 331, num_levels),
+            "rotate": np.linspace(0, 30, num_levels),
+            "color": np.linspace(0.0, 0.9, num_levels),
+            "posterize": np.round(np.linspace(8, 4, num_levels), 0).astype(
+                np.int32
+            ),
+            "solarize": np.linspace(256, 0, num_levels),  # range [0, 256]
+            "contrast": np.linspace(0.0, 0.9, num_levels),
+            "sharpness": np.linspace(0.0, 0.9, num_levels),
+            "brightness": np.linspace(0.0, 0.9, num_levels),
+            "autocontrast": [0]
+            * num_levels,  # This augmentation doesn't use magnitude parameter.
+            "equalize": [0]
+            * num_levels,  # This augmentation doesn't use magnitude parameter.
+            "invert": [0]
+            * num_levels,  # This augmentation doesn't use magnitude parameter.
+        }
+
+        def rotate_with_fill(img, magnitude):
+            """Define rotation transformation with fill.
+
+            The input image is first rotated, then it is blended together with
+            a gray mask of the same size. Note that fillcolor as defined
+            elsewhere in this module doesn't apply here.
+
+            Args:
+                magnitude (float): rotation angle in degrees.
+            Returns:
+                rotated_filled (PIL Image): rotated image with gray filling for
+                disoccluded areas unveiled by the rotation.
+            """
+            rotated = img.convert("RGBA").rotate(magnitude)
+            rotated_filled = Image.composite(
+                rotated, Image.new("RGBA", rotated.size, (128,) * 4), rotated
+            )
+            return rotated_filled.convert(img.mode)
+
+        # Define a dictionary of augmentation functions where each key refers
+        # to a specific type of augmentation and the corresponding value defines
+        # the augmentation itself using a lambda function.
+        # pylint: disable=unnecessary-lambda
+        func_dict = {
+            "shearX": lambda img, magnitude: img.transform(
+                img.size,
+                Image.AFFINE,
+                (1, magnitude * random.choice([-1, 1]), 0, 0, 1, 0),
+                Image.BICUBIC,
+                fillcolor=fillcolor,
+            ),
+            "shearY": lambda img, magnitude: img.transform(
+                img.size,
+                Image.AFFINE,
+                (1, 0, 0, magnitude * random.choice([-1, 1]), 1, 0),
+                Image.BICUBIC,
+                fillcolor=fillcolor,
+            ),
+            "translateX": lambda img, magnitude: img.transform(
+                img.size,
+                Image.AFFINE,
+                (
+                    1,
+                    0,
+                    magnitude * img.size[0] * random.choice([-1, 1]),
+                    0,
+                    1,
+                    0,
+                ),
+                fillcolor=fillcolor,
+            ),
+            "translateY": lambda img, magnitude: img.transform(
+                img.size,
+                Image.AFFINE,
+                (
+                    1,
+                    0,
+                    0,
+                    0,
+                    1,
+                    magnitude * img.size[1] * random.choice([-1, 1]),
+                ),
+                fillcolor=fillcolor,
+            ),
+            "rotate": lambda img, magnitude: rotate_with_fill(img, magnitude),
+            "color": lambda img, magnitude: ImageEnhance.Color(img).enhance(
+                1 + magnitude * random.choice([-1, 1])
+            ),
+            "posterize": lambda img, magnitude: ImageOps.posterize(
+                img, magnitude
+            ),
+            "solarize": lambda img, magnitude: ImageOps.solarize(
+                img, magnitude
+            ),
+            "contrast": lambda img, magnitude: ImageEnhance.Contrast(
+                img
+            ).enhance(1 + magnitude * random.choice([-1, 1])),
+            "sharpness": lambda img, magnitude: ImageEnhance.Sharpness(
+                img
+            ).enhance(1 + magnitude * random.choice([-1, 1])),
+            "brightness": lambda img, magnitude: ImageEnhance.Brightness(
+                img
+            ).enhance(1 + magnitude * random.choice([-1, 1])),
+            "autocontrast": lambda img, magnitude: ImageOps.autocontrast(img),
+            "equalize": lambda img, magnitude: ImageOps.equalize(img),
+            "invert": lambda img, magnitude: ImageOps.invert(img),
+        }
+
+        # Store probability, function and magnitude of the first augmentation
+        # for the sub-policy.
+        self.probability1 = probability1
+        self.operation1 = func_dict[operation1]
+        self.magnitude1 = ranges[operation1][magnitude_idx1]
+
+        # Store probability, function and magnitude of the second augmentation
+        # for the sub-policy.
+        self.probability2 = probability2
+        self.operation2 = func_dict[operation2]
+        self.magnitude2 = ranges[operation2][magnitude_idx2]
+
+    def __call__(self, img):
+        """Define call method for SubPolicy class."""
+        # Randomly apply operation 1.
+        if random.random() < self.probability1:
+            img = self.operation1(img, self.magnitude1)
+
+        # Randomly apply operation 2.
+        if random.random() < self.probability2:
+            img = self.operation2(img, self.magnitude2)
+
+        return img

Megatron-LM-core_r0.7.0.beta/megatron/legacy/data/biencoder_dataset_utils.py

+import os
+import time
+
+import numpy as np
+import torch
+
+from megatron.training import get_args, get_tokenizer, 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.legacy.data.data_samplers import MegatronPretrainingSampler
+
+def make_attention_mask(source_block, target_block):
+    """
+    Returns a 2-dimensional (2-D) attention mask
+    :param source_block: 1-D array
+    :param target_block: 1-D array
+    """
+    mask = (target_block[None, :] >= 1) * (source_block[:, None] >= 1)
+    mask = mask.astype(np.int64)
+    # (source_length, target_length)
+    return mask
+
+def get_one_epoch_dataloader(dataset, micro_batch_size=None):
+    """Specifically one epoch to be used in an indexing job."""
+    args = get_args()
+
+    if micro_batch_size is None:
+        micro_batch_size = args.micro_batch_size
+    num_workers = args.num_workers
+
+    # Use megatron's sampler with consumed samples set to 0 as
+    # this is only for evaluation and don't intend to resume half way.
+    # Also, set the drop last to false as don't intend to remove
+    # the last batch
+    batch_sampler = MegatronPretrainingSampler(
+        total_samples=len(dataset),
+        consumed_samples=0,
+        micro_batch_size=args.micro_batch_size,
+        data_parallel_rank=mpu.get_data_parallel_rank(),
+        data_parallel_size=mpu.get_data_parallel_world_size(),
+        drop_last=False)
+
+    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_mask',
+            'context_tokens', 'context_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_mask = data_b['query_mask'] < 0.5
+    context_tokens = data_b['context_tokens'].long()
+    context_mask = data_b['context_mask'] < 0.5
+    block_indices = data_b['block_data'].long()
+
+    return query_tokens, query_mask,\
+           context_tokens, context_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-LM-core_r0.7.0.beta/megatron/legacy/data/data_samplers.py

+# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
+
+"""Dataloaders."""
+
+
+import random
+import torch
+import numpy as np
+from torch.utils.data import Dataset
+from megatron.training import get_args
+from megatron.core import mpu
+
+
+def build_pretraining_data_loader(dataset, consumed_samples):
+    """Build dataloader given an input dataset."""
+
+    if dataset is None:
+        return None
+    args = get_args()
+
+    # Megatron sampler
+    if args.dataloader_type == 'single':
+        batch_sampler = MegatronPretrainingSampler(
+            total_samples=len(dataset),
+            consumed_samples=consumed_samples,
+            micro_batch_size=args.micro_batch_size,
+            data_parallel_rank=mpu.get_data_parallel_rank(),
+            data_parallel_size=mpu.get_data_parallel_world_size())
+    elif args.dataloader_type == 'cyclic':
+        batch_sampler = MegatronPretrainingRandomSampler(
+            dataset,
+            total_samples=len(dataset),
+            consumed_samples=consumed_samples,
+            micro_batch_size=args.micro_batch_size,
+            data_parallel_rank=mpu.get_data_parallel_rank(),
+            data_parallel_size=mpu.get_data_parallel_world_size(),
+            data_sharding=args.data_sharding)
+    elif args.dataloader_type == "external":
+        # External dataloaders are passed through. User is expected to provide a
+        # torch-compatible dataloader and define samplers, if needed.
+        return dataset
+    else:
+        raise Exception('{} dataloader type is not supported.'.format(
+                args.dataloader_type))
+
+    # Torch dataloader.
+    return torch.utils.data.DataLoader(dataset,
+                                       batch_sampler=batch_sampler,
+                                       num_workers=args.num_workers,
+                                       pin_memory=True,
+                                       persistent_workers=True if args.num_workers > 0 else False,
+                                       )
+
+class MegatronPretrainingSampler:
+
+    def __init__(self, total_samples, consumed_samples, micro_batch_size,
+                 data_parallel_rank, data_parallel_size, drop_last=True):
+        # Keep a copy of input params for later use.
+        self.total_samples = total_samples
+        self.consumed_samples = consumed_samples
+        self.micro_batch_size = micro_batch_size
+        self.data_parallel_rank = data_parallel_rank
+        self.micro_batch_times_data_parallel_size = \
+            self.micro_batch_size * data_parallel_size
+        self.drop_last = drop_last
+
+        # Sanity checks.
+        assert self.total_samples > 0, \
+            'no sample to consume: {}'.format(self.total_samples)
+        assert self.consumed_samples < self.total_samples, \
+            'no samples left to consume: {}, {}'.format(self.consumed_samples,
+                                                        self.total_samples)
+        assert self.micro_batch_size > 0
+        assert data_parallel_size > 0
+        assert self.data_parallel_rank < data_parallel_size, \
+            'data_parallel_rank should be smaller than data size: {}, ' \
+            '{}'.format(self.data_parallel_rank, data_parallel_size)
+
+    def __len__(self):
+        return self.total_samples
+
+    def get_start_end_idx(self):
+        start_idx = self.data_parallel_rank * self.micro_batch_size
+        end_idx = start_idx + self.micro_batch_size
+        return start_idx, end_idx
+
+    def __iter__(self):
+        batch = []
+        # Last batch will be dropped if drop_last is not set False
+        for idx in range(self.consumed_samples, self.total_samples):
+            batch.append(idx)
+            if len(batch) == self.micro_batch_times_data_parallel_size:
+                start_idx, end_idx = self.get_start_end_idx()
+                yield batch[start_idx:end_idx]
+                batch = []
+
+        # Check the last partial batch and see drop_last is set
+        if len(batch) > 0 and not self.drop_last:
+            start_idx, end_idx = self.get_start_end_idx()
+            yield batch[start_idx:end_idx]
+
+
+class RandomSeedDataset(Dataset):
+
+    def __init__(self, dataset):
+        args = get_args()
+        self.base_seed = args.seed
+        self.curr_seed = args.seed
+        self.dataset = dataset
+
+    def __len__(self):
+        return len(self.dataset)
+
+    def set_epoch(self, epoch):
+        self.curr_seed = self.base_seed + epoch
+
+    def __getitem__(self, idx):
+        seed = idx + self.curr_seed
+        torch.manual_seed(seed)
+        random.seed(seed)
+        np.random.seed(seed)
+        return self.dataset[idx]
+
+
+class MegatronPretrainingRandomSampler:
+
+    def __init__(self, dataset, total_samples, consumed_samples, micro_batch_size,
+                 data_parallel_rank, data_parallel_size, data_sharding):
+        # Keep a copy of input params for later use.
+        self.dataset = dataset
+        self.total_samples = total_samples
+        self.consumed_samples = consumed_samples
+        self.micro_batch_size = micro_batch_size
+        self.data_parallel_rank = data_parallel_rank
+        self.data_parallel_size = data_parallel_size
+        self.data_sharding = data_sharding
+        self.micro_batch_times_data_parallel_size = \
+            self.micro_batch_size * data_parallel_size
+        self.last_batch_size = \
+            self.total_samples % self.micro_batch_times_data_parallel_size
+
+        # Sanity checks.
+        assert self.total_samples > 0, \
+            'no sample to consume: {}'.format(self.total_samples)
+        assert self.micro_batch_size > 0
+        assert data_parallel_size > 0
+        assert self.data_parallel_rank < data_parallel_size, \
+            'data_parallel_rank should be smaller than data size: {}, ' \
+            '{}'.format(self.data_parallel_rank, data_parallel_size)
+
+    def __len__(self):
+        return self.total_samples
+
+    def __iter__(self):
+        active_total_samples = self.total_samples - self.last_batch_size
+        self.epoch = self.consumed_samples // active_total_samples
+        current_epoch_samples = self.consumed_samples % active_total_samples
+        assert current_epoch_samples % self.micro_batch_times_data_parallel_size == 0
+
+        if isinstance(self.dataset, RandomSeedDataset):
+            self.dataset.set_epoch(self.epoch)
+
+        # data sharding and random sampling
+        if self.data_sharding:
+            bucket_size = (self.total_samples // self.micro_batch_times_data_parallel_size) \
+                           * self.micro_batch_size
+            bucket_offset = current_epoch_samples // self.data_parallel_size
+            start_idx = self.data_parallel_rank * bucket_size
+
+            g = torch.Generator()
+            g.manual_seed(self.epoch)
+            random_idx = torch.randperm(bucket_size, generator=g).tolist()
+            idx_range = [start_idx + x for x in random_idx[bucket_offset:]]
+        else:
+            full_bucket_size = (self.total_samples // self.micro_batch_size) \
+                                * self.micro_batch_size
+            full_bucket_offset = current_epoch_samples
+            g = torch.Generator()
+            g.manual_seed(self.epoch)
+            idx_range_total = \
+                torch.randperm(full_bucket_size, generator=g).tolist()
+            idx_range_active = idx_range_total[full_bucket_offset:]
+            idx_range = idx_range_active[self.data_parallel_rank::self.data_parallel_size]
+
+        batch = []
+        # Last batch if not complete will be dropped.
+        for idx in idx_range:
+            batch.append(idx)
+            if len(batch) == self.micro_batch_size:
+                self.consumed_samples += self.micro_batch_times_data_parallel_size
+                yield batch
+                batch = []

Megatron-LM-core_r0.7.0.beta/megatron/legacy/data/dataset_utils.py

+# coding=utf-8
+# Copyright 2018 The Google AI Language Team Authors, and NVIDIA.
+#
+# 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.
+
+
+# Most of the code here has been copied from:
+#   https://github.com/google-research/albert/blob/master/create_pretraining_data.py
+# with some modifications.
+
+import math
+import os
+import time
+import collections
+
+import numpy as np
+import torch
+
+from megatron.training import (
+    get_args,
+    print_rank_0
+)
+from megatron.core import mpu
+from megatron.core.datasets.indexed_dataset import IndexedDataset
+
+
+DSET_TYPE_BERT = 'standard_bert'
+DSET_TYPE_ICT = 'ict'
+DSET_TYPE_T5  = 't5'
+DSET_TYPE_MULTIMODAL = 'multimodal'
+
+DSET_TYPES = [DSET_TYPE_BERT, DSET_TYPE_ICT, DSET_TYPE_T5, DSET_TYPE_MULTIMODAL]
+
+
+def get_datasets_weights_and_num_samples(data_prefix,
+                                         train_valid_test_num_samples):
+
+    # The data prefix should be in the format of:
+    #   weight-1, data-prefix-1, weight-2, data-prefix-2, ..
+    assert len(data_prefix) % 2 == 0
+    num_datasets = len(data_prefix) // 2
+    weights = [0]*num_datasets
+    prefixes = [0]*num_datasets
+    for i in range(num_datasets):
+        weights[i] = float(data_prefix[2*i])
+        prefixes[i] = (data_prefix[2*i+1]).strip()
+    # Normalize weights
+    weight_sum = 0.0
+    for weight in weights:
+        weight_sum += weight
+    assert weight_sum > 0.0
+    weights = [weight / weight_sum for weight in weights]
+
+    # Add 0.5% (the 1.005 factor) so in case the bleding dataset does
+    # not uniformly distribute the number of samples, we still have
+    # samples left to feed to the network.
+    if isinstance(train_valid_test_num_samples, list):
+        datasets_train_valid_test_num_samples = []
+        for weight in weights:
+            datasets_train_valid_test_num_samples.append(
+                [int(math.ceil(val * weight * 1.005))
+                for val in train_valid_test_num_samples])
+    else:
+        # Used when separate dataset files are provided for train,
+        # valid and test
+        datasets_train_valid_test_num_samples = [
+            int(math.ceil(train_valid_test_num_samples * weight * 1.005))
+            for weight in weights]
+
+    return prefixes, weights, datasets_train_valid_test_num_samples
+
+
+def get_a_and_b_segments(sample, np_rng):
+    """Divide sample into a and b segments."""
+
+    # Number of sentences in the sample.
+    n_sentences = len(sample)
+    # Make sure we always have two sentences.
+    assert n_sentences > 1, 'make sure each sample has at least two sentences.'
+
+    # First part:
+    # `a_end` is how many sentences go into the `A`.
+    a_end = 1
+    if n_sentences >= 3:
+        # Note that randin in numpy is exclusive.
+        a_end = np_rng.randint(1, n_sentences)
+    tokens_a = []
+    for j in range(a_end):
+        tokens_a.extend(sample[j])
+
+    # Second part:
+    tokens_b = []
+    for j in range(a_end, n_sentences):
+        tokens_b.extend(sample[j])
+
+    # Random next:
+    is_next_random = False
+    if np_rng.random() < 0.5:
+        is_next_random = True
+        tokens_a, tokens_b = tokens_b, tokens_a
+
+    return tokens_a, tokens_b, is_next_random
+
+
+def truncate_segments(tokens_a, tokens_b, len_a, len_b, max_num_tokens, np_rng):
+    """Truncates a pair of sequences to a maximum sequence length."""
+    #print(len_a, len_b, max_num_tokens)
+    assert len_a > 0
+    if len_a + len_b <= max_num_tokens:
+        return False
+    while len_a + len_b > max_num_tokens:
+        if len_a > len_b:
+            len_a -= 1
+            tokens = tokens_a
+        else:
+            len_b -= 1
+            tokens = tokens_b
+        if np_rng.random() < 0.5:
+            del tokens[0]
+        else:
+            tokens.pop()
+    return True
+
+
+def create_tokens_and_tokentypes(tokens_a, tokens_b, cls_id, sep_id):
+    """Merge segments A and B, add [CLS] and [SEP] and build tokentypes."""
+
+    tokens = []
+    tokentypes = []
+    # [CLS].
+    tokens.append(cls_id)
+    tokentypes.append(0)
+    # Segment A.
+    for token in tokens_a:
+        tokens.append(token)
+        tokentypes.append(0)
+    # [SEP].
+    tokens.append(sep_id)
+    tokentypes.append(0)
+    # Segment B.
+    for token in tokens_b:
+        tokens.append(token)
+        tokentypes.append(1)
+    if tokens_b:
+        # [SEP].
+        tokens.append(sep_id)
+        tokentypes.append(1)
+
+    return tokens, tokentypes
+
+
+MaskedLmInstance = collections.namedtuple("MaskedLmInstance",
+                                          ["index", "label"])
+
+
+def is_start_piece(piece):
+    """Check if the current word piece is the starting piece (BERT)."""
+    # When a word has been split into
+    # WordPieces, the first token does not have any marker and any subsequence
+    # tokens are prefixed with ##. So whenever we see the ## token, we
+    # append it to the previous set of word indexes.
+    return not piece.startswith("##")
+
+
+def create_masked_lm_predictions(tokens,
+                                 vocab_id_list, vocab_id_to_token_dict,
+                                 masked_lm_prob,
+                                 cls_id, sep_id, mask_id,
+                                 max_predictions_per_seq,
+                                 np_rng,
+                                 max_ngrams=3,
+                                 do_whole_word_mask=True,
+                                 favor_longer_ngram=False,
+                                 do_permutation=False,
+                                 geometric_dist=False,
+                                 masking_style="bert"):
+    """Creates the predictions for the masked LM objective.
+    Note: Tokens here are vocab ids and not text tokens."""
+
+    cand_indexes = []
+    # Note(mingdachen): We create a list for recording if the piece is
+    # the starting piece of current token, where 1 means true, so that
+    # on-the-fly whole word masking is possible.
+    token_boundary = [0] * len(tokens)
+
+    for (i, token) in enumerate(tokens):
+        if token == cls_id or token == sep_id:
+            token_boundary[i] = 1
+            continue
+        # Whole Word Masking means that if we mask all of the wordpieces
+        # corresponding to an original word.
+        #
+        # Note that Whole Word Masking does *not* change the training code
+        # at all -- we still predict each WordPiece independently, softmaxed
+        # over the entire vocabulary.
+        if (do_whole_word_mask and len(cand_indexes) >= 1 and
+                not is_start_piece(vocab_id_to_token_dict[token])):
+            cand_indexes[-1].append(i)
+        else:
+            cand_indexes.append([i])
+            if is_start_piece(vocab_id_to_token_dict[token]):
+                token_boundary[i] = 1
+
+    output_tokens = list(tokens)
+
+    masked_lm_positions = []
+    masked_lm_labels = []
+
+    if masked_lm_prob == 0:
+        return (output_tokens, masked_lm_positions,
+                masked_lm_labels, token_boundary)
+
+    num_to_predict = min(max_predictions_per_seq,
+                         max(1, int(round(len(tokens) * masked_lm_prob))))
+
+    ngrams = np.arange(1, max_ngrams + 1, dtype=np.int64)
+    if not geometric_dist:
+        # Note(mingdachen):
+        # By default, we set the probilities to favor shorter ngram sequences.
+        pvals = 1. / np.arange(1, max_ngrams + 1)
+        pvals /= pvals.sum(keepdims=True)
+        if favor_longer_ngram:
+            pvals = pvals[::-1]
+
+    ngram_indexes = []
+    for idx in range(len(cand_indexes)):
+        ngram_index = []
+        for n in ngrams:
+            ngram_index.append(cand_indexes[idx:idx + n])
+        ngram_indexes.append(ngram_index)
+
+    np_rng.shuffle(ngram_indexes)
+
+    (masked_lms, masked_spans) = ([], [])
+    covered_indexes = set()
+    for cand_index_set in ngram_indexes:
+        if len(masked_lms) >= num_to_predict:
+            break
+        if not cand_index_set:
+            continue
+        # Note(mingdachen):
+        # Skip current piece if they are covered in lm masking or previous ngrams.
+        for index_set in cand_index_set[0]:
+            for index in index_set:
+                if index in covered_indexes:
+                    continue
+
+        if not geometric_dist:
+            n = np_rng.choice(ngrams[:len(cand_index_set)],
+                              p=pvals[:len(cand_index_set)] /
+                              pvals[:len(cand_index_set)].sum(keepdims=True))
+        else:
+            # Sampling "n" from the geometric distribution and clipping it to
+            # the max_ngrams. Using p=0.2 default from the SpanBERT paper
+            # https://arxiv.org/pdf/1907.10529.pdf (Sec 3.1)
+            n = min(np_rng.geometric(0.2), max_ngrams)
+
+        index_set = sum(cand_index_set[n - 1], [])
+        n -= 1
+        # Note(mingdachen):
+        # Repeatedly looking for a candidate that does not exceed the
+        # maximum number of predictions by trying shorter ngrams.
+        while len(masked_lms) + len(index_set) > num_to_predict:
+            if n == 0:
+                break
+            index_set = sum(cand_index_set[n - 1], [])
+            n -= 1
+        # If adding a whole-word mask would exceed the maximum number of
+        # predictions, then just skip this candidate.
+        if len(masked_lms) + len(index_set) > num_to_predict:
+            continue
+        is_any_index_covered = False
+        for index in index_set:
+            if index in covered_indexes:
+                is_any_index_covered = True
+                break
+        if is_any_index_covered:
+            continue
+        for index in index_set:
+            covered_indexes.add(index)
+            masked_token = None
+            if masking_style == "bert":
+                # 80% of the time, replace with [MASK]
+                if np_rng.random() < 0.8:
+                    masked_token = mask_id
+                else:
+                    # 10% of the time, keep original
+                    if np_rng.random() < 0.5:
+                        masked_token = tokens[index]
+                    # 10% of the time, replace with random word
+                    else:
+                        masked_token = vocab_id_list[np_rng.randint(0, len(vocab_id_list))]
+            elif masking_style == "t5":
+                masked_token = mask_id
+            else:
+                raise ValueError("invalid value of masking style")
+
+            output_tokens[index] = masked_token
+            masked_lms.append(MaskedLmInstance(index=index, label=tokens[index]))
+
+        masked_spans.append(MaskedLmInstance(
+            index=index_set,
+            label=[tokens[index] for index in index_set]))
+
+    assert len(masked_lms) <= num_to_predict
+    np_rng.shuffle(ngram_indexes)
+
+    select_indexes = set()
+    if do_permutation:
+        for cand_index_set in ngram_indexes:
+            if len(select_indexes) >= num_to_predict:
+                break
+            if not cand_index_set:
+                continue
+            # Note(mingdachen):
+            # Skip current piece if they are covered in lm masking or previous ngrams.
+            for index_set in cand_index_set[0]:
+                for index in index_set:
+                    if index in covered_indexes or index in select_indexes:
+                        continue
+
+            n = np.random.choice(ngrams[:len(cand_index_set)],
+                                 p=pvals[:len(cand_index_set)] /
+                                 pvals[:len(cand_index_set)].sum(keepdims=True))
+            index_set = sum(cand_index_set[n - 1], [])
+            n -= 1
+
+            while len(select_indexes) + len(index_set) > num_to_predict:
+                if n == 0:
+                    break
+                index_set = sum(cand_index_set[n - 1], [])
+                n -= 1
+            # If adding a whole-word mask would exceed the maximum number of
+            # predictions, then just skip this candidate.
+            if len(select_indexes) + len(index_set) > num_to_predict:
+                continue
+            is_any_index_covered = False
+            for index in index_set:
+                if index in covered_indexes or index in select_indexes:
+                    is_any_index_covered = True
+                    break
+            if is_any_index_covered:
+                continue
+            for index in index_set:
+                select_indexes.add(index)
+        assert len(select_indexes) <= num_to_predict
+
+        select_indexes = sorted(select_indexes)
+        permute_indexes = list(select_indexes)
+        np_rng.shuffle(permute_indexes)
+        orig_token = list(output_tokens)
+
+        for src_i, tgt_i in zip(select_indexes, permute_indexes):
+            output_tokens[src_i] = orig_token[tgt_i]
+            masked_lms.append(MaskedLmInstance(index=src_i, label=orig_token[src_i]))
+
+    masked_lms = sorted(masked_lms, key=lambda x: x.index)
+    # Sort the spans by the index of the first span
+    masked_spans = sorted(masked_spans, key=lambda x: x.index[0])
+
+    for p in masked_lms:
+        masked_lm_positions.append(p.index)
+        masked_lm_labels.append(p.label)
+    return (output_tokens, masked_lm_positions, masked_lm_labels, token_boundary, masked_spans)
+
+
+def pad_and_convert_to_numpy(tokens, tokentypes, masked_positions,
+                             masked_labels, pad_id, max_seq_length):
+    """Pad sequences and convert them to numpy."""
+
+    # Some checks.
+    num_tokens = len(tokens)
+    padding_length = max_seq_length - num_tokens
+    assert padding_length >= 0
+    assert len(tokentypes) == num_tokens
+    assert len(masked_positions) == len(masked_labels)
+
+    # Tokens and token types.
+    filler = [pad_id] * padding_length
+    tokens_np = np.array(tokens + filler, dtype=np.int64)
+    tokentypes_np = np.array(tokentypes + filler, dtype=np.int64)
+
+    # Padding mask.
+    padding_mask_np = np.array([1] * num_tokens + [0] * padding_length,
+                               dtype=np.int64)
+
+    # Lables and loss mask.
+    labels = [-1] * max_seq_length
+    loss_mask = [0] * max_seq_length
+    for i in range(len(masked_positions)):
+        assert masked_positions[i] < num_tokens
+        labels[masked_positions[i]] = masked_labels[i]
+        loss_mask[masked_positions[i]] = 1
+    labels_np = np.array(labels, dtype=np.int64)
+    loss_mask_np = np.array(loss_mask, dtype=np.int64)
+
+    return tokens_np, tokentypes_np, labels_np, padding_mask_np, loss_mask_np
+
+
+def build_train_valid_test_datasets_with_prefixes(train_valid_test_num_samples,
+                                                  max_seq_length,
+                                                  seed,
+                                                  train_data_prefix=None,
+                                                  valid_data_prefix=None,
+                                                  test_data_prefix=None,
+                                                  binary_head=False,
+                                                  max_seq_length_dec=None,
+                                                  dataset_type='standard_bert'):
+    print_rank_0("Separate data paths provided for train, valid & test.")
+
+    train_dataset, valid_dataset, test_dataset = None, None, None
+    # Single dataset.
+    if train_data_prefix is not None:
+        train_dataset = build_dataset("train", train_data_prefix,
+                                      train_valid_test_num_samples[0],
+                                      max_seq_length, seed,
+                                      binary_head, max_seq_length_dec,
+                                      dataset_type=dataset_type)
+
+    if valid_data_prefix is not None:
+        valid_dataset = build_dataset("valid", valid_data_prefix,
+                                      train_valid_test_num_samples[1],
+                                      max_seq_length, seed, False,
+                                      binary_head, max_seq_length_dec,
+                                      dataset_type=dataset_type)
+
+    if test_data_prefix is not None:
+        test_dataset = build_dataset("test", test_data_prefix,
+                                     train_valid_test_num_samples[2],
+                                     max_seq_length, seed, False,
+                                     binary_head, max_seq_length_dec,
+                                     dataset_type=dataset_type)
+
+    return (train_dataset, valid_dataset, test_dataset)
+
+
+def build_train_valid_test_datasets(data_prefix, splits_string,
+                                    train_valid_test_num_samples,
+                                    max_seq_length, seed,
+                                    binary_head=False,
+                                    max_seq_length_dec=None,
+                                    dataset_type='standard_bert'):
+
+    if len(data_prefix) == 1:
+        return _build_train_valid_test_datasets(data_prefix[0],
+                                                splits_string,
+                                                train_valid_test_num_samples,
+                                                max_seq_length, seed,
+                                                binary_head,
+                                                max_seq_length_dec,
+                                                dataset_type=dataset_type)
+
+    raise NotImplementedError("Blending currently unsupported for non-GPT dataset instances")
+
+
+def _build_train_valid_test_datasets(data_prefix, splits_string,
+                                     train_valid_test_num_samples,
+                                     max_seq_length, seed,
+                                     binary_head,
+                                     max_seq_length_dec,
+                                     dataset_type='standard_bert'):
+
+    # Indexed dataset.
+    indexed_dataset = get_indexed_dataset_(data_prefix,
+                                           dataset_type)
+
+    # Get start and end indices of train/valid/train into doc-idx
+    # Note that doc-idx is desinged to be num-docs + 1 so we can
+    # easily iterate over it.
+    total_num_of_documents = indexed_dataset.document_indices.shape[0] - 1
+    splits = get_train_valid_test_split_(splits_string, total_num_of_documents)
+
+    # Print stats about the splits.
+    print_rank_0(' > dataset split:')
+
+    def print_split_stats(name, index):
+        print_rank_0('    {}:'.format(name))
+        print_rank_0('     document indices in [{}, {}) total of {} '
+                     'documents'.format(splits[index], splits[index + 1],
+                                        splits[index + 1] - splits[index]))
+        start_index = indexed_dataset.document_indices[splits[index]]
+        end_index = indexed_dataset.document_indices[splits[index + 1]]
+        print_rank_0('     sentence indices in [{}, {}) total of {} '
+                     'sentences'.format(start_index, end_index,
+                                        end_index - start_index))
+    print_split_stats('train', 0)
+    print_split_stats('validation', 1)
+    print_split_stats('test', 2)
+
+    def build_split_dataset(index, name):
+        dataset = None
+        if splits[index + 1] > splits[index]:
+            # Get the pointer to the original doc-idx so we can set it later.
+            doc_idx_ptr = indexed_dataset.get_document_indices()
+            # Slice the doc-idx
+            start_index = splits[index]
+            # Add +1 so we can index into the dataset to get the upper bound.
+            end_index = splits[index + 1] + 1
+            # New doc_idx view.
+            indexed_dataset.set_document_indices(doc_idx_ptr[start_index:end_index])
+
+            dataset = build_dataset(
+                name, data_prefix,
+                train_valid_test_num_samples[index], max_seq_length,
+                seed, binary_head, max_seq_length_dec,
+                dataset_type, indexed_dataset)
+
+            # Set the original pointer so dataset remains the main dataset.
+            indexed_dataset.set_document_indices(doc_idx_ptr)
+            # Checks.
+            assert indexed_dataset.document_indices[0] == 0
+            assert indexed_dataset.document_indices.shape[0] == \
+                (total_num_of_documents + 1)
+        return dataset
+    
+    train_dataset = build_split_dataset(0, 'train')
+    valid_dataset = build_split_dataset(1, 'valid')
+    test_dataset = build_split_dataset(2, 'test')
+
+    return (train_dataset, valid_dataset, test_dataset)
+
+
+def build_dataset(name, data_prefix, max_num_samples,
+                  max_seq_length, seed, binary_head,
+                  max_seq_length_dec, dataset_type='standard_bert',
+                  indexed_dataset=None):
+
+    from megatron.legacy.data.ict_dataset import ICTDataset
+    from megatron.legacy.data.multimodal_dataset import MultiModalDataset
+
+    if dataset_type == DSET_TYPE_BERT or dataset_type == DSET_TYPE_T5:
+        raise ValueError("The Megatron-LM BERT and T5 datasets are deprecated.")
+
+    if dataset_type not in DSET_TYPES:
+        raise ValueError("Invalid dataset_type: ", dataset_type)
+
+    if indexed_dataset is None:
+        indexed_dataset = get_indexed_dataset_(data_prefix,
+                                               dataset_type)
+
+    kwargs = dict(
+        name=name,
+        data_prefix=data_prefix,
+        num_epochs=None,
+        max_num_samples=max_num_samples,
+        max_seq_length=max_seq_length,
+        seed=seed,
+    )
+
+    if dataset_type == DSET_TYPE_ICT:
+        args = get_args()
+
+        title_dataset = get_indexed_dataset_(
+            args.titles_data_path,
+            dataset_type)
+
+        dataset = ICTDataset(
+            block_dataset=indexed_dataset,
+            title_dataset=title_dataset,
+            query_in_block_prob=args.query_in_block_prob,
+            use_one_sent_docs=args.use_one_sent_docs,
+            binary_head=binary_head,
+            **kwargs
+        )
+    elif dataset_type == DSET_TYPE_MULTIMODAL:
+        args = get_args()
+        dataset = MultiModalDataset(
+            name=name,
+            data_prefix=data_prefix,
+            indexed_dataset=indexed_dataset,
+            num_samples=max_num_samples,
+            seq_length=max_seq_length,
+            seed=seed,
+            img_h=args.img_h,
+            img_w=args.img_w,
+        )
+    else:
+        raise NotImplementedError("Dataset type not fully implemented.")
+
+    return dataset
+
+
+def get_indexed_dataset_(data_prefix, dataset_type):
+
+    print_rank_0(' > building dataset index ...')
+
+    start_time = time.time()
+    multimodal = dataset_type == DSET_TYPE_MULTIMODAL
+    indexed_dataset = IndexedDataset(data_prefix, multimodal)
+    assert indexed_dataset.sequence_lengths.shape[0] == indexed_dataset.document_indices[-1]
+    print_rank_0(' > finished creating indexed dataset in {:4f} '
+                 'seconds'.format(time.time() - start_time))
+
+    print_rank_0(' > indexed dataset stats:')
+    print_rank_0('    number of documents: {}'.format(
+        indexed_dataset.document_indices.shape[0] - 1))
+    print_rank_0('    number of sentences: {}'.format(
+        indexed_dataset.sequence_lengths.shape[0]))
+
+    return indexed_dataset
+
+
+def get_train_valid_test_split_(splits_string, size):
+    """ Get dataset splits from comma or '/' separated string list."""
+
+    splits = []
+    if splits_string.find(',') != -1:
+        splits = [float(s) for s in splits_string.split(',')]
+    elif splits_string.find('/') != -1:
+        splits = [float(s) for s in splits_string.split('/')]
+    else:
+        splits = [float(splits_string)]
+    while len(splits) < 3:
+        splits.append(0.)
+    splits = splits[:3]
+    splits_sum = sum(splits)
+    assert splits_sum > 0.0
+    splits = [split / splits_sum for split in splits]
+    splits_index = [0]
+    for index, split in enumerate(splits):
+        splits_index.append(splits_index[index] +
+                            int(round(split * float(size))))
+    diff = splits_index[-1] - size
+    for index in range(1, len(splits_index)):
+        splits_index[index] -= diff
+    assert len(splits_index) == 4
+    assert splits_index[-1] == size
+    return splits_index
+
+def get_samples_mapping(indexed_dataset,
+                        data_prefix,
+                        num_epochs,
+                        max_num_samples,
+                        max_seq_length,
+                        short_seq_prob,
+                        seed,
+                        name,
+                        binary_head):
+    """Get a list that maps a sample index to a starting sentence index, end sentence index, and length"""
+
+    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 += '_{:0.2f}ssp'.format(short_seq_prob)
+    indexmap_filename += '_{}s'.format(seed)
+    indexmap_filename += '.npy'
+
+    # Build the indexed mapping if not exist.
+    if torch.distributed.get_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 indexed_dataset.document_indices.dtype == np.int64
+        assert indexed_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))
+        # First compile and then import.
+        from megatron.core.datasets import helpers
+        samples_mapping = helpers.build_mapping(
+            indexed_dataset.document_indices,
+            indexed_dataset.sequence_lengths,
+            num_epochs,
+            max_num_samples,
+            max_seq_length,
+            short_seq_prob,
+            seed,
+            verbose,
+            2 if binary_head else 1)
+        print_rank_0(' > done building samples index maping')
+        np.save(indexmap_filename, samples_mapping, 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(' > elasped 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())
+    torch.distributed.all_reduce(counts, group=mpu.get_pipeline_model_parallel_group())
+    assert counts[0].item() == (
+        torch.distributed.get_world_size() //
+        torch.distributed.get_world_size(group=mpu.get_tensor_model_parallel_group()))
+
+    # Load indexed dataset.
+    print_rank_0(' > loading indexed mapping from {}'.format(
+        indexmap_filename))
+    start_time = time.time()
+    samples_mapping = np.load(indexmap_filename, allow_pickle=True, mmap_mode='r')
+    print_rank_0('    loaded indexed file in {:3.3f} seconds'.format(
+        time.time() - start_time))
+    print_rank_0('    total number of samples: {}'.format(
+        samples_mapping.shape[0]))
+
+    return samples_mapping

Megatron-LM-core_r0.7.0.beta/megatron/legacy/data/ict_dataset.py

+import itertools
+import random
+
+import numpy as np
+from torch.utils.data import Dataset
+
+from megatron.training import get_tokenizer
+from megatron.training import get_args
+from megatron.legacy.data.dataset_utils import get_indexed_dataset_
+from megatron.legacy.data.realm_dataset_utils import get_block_samples_mapping
+
+def make_attention_mask(source_block, target_block):
+    """
+    Returns a 2-dimensional (2-D) attention mask
+    :param source_block: 1-D array
+    :param target_block: 1-D array
+    """
+    mask = (target_block[None, :] >= 1) * (source_block[:, None] >= 1)
+    mask = mask.astype(np.int64)
+    # (source_length, target_length)
+    return mask
+
+def get_ict_dataset(use_titles=True, query_in_block_prob=1):
+    """Get a dataset which uses block samples mappings to get ICT/block indexing data (via get_block())
+    rather than for training, since it is only built with a single epoch sample mapping.
+    """
+    args = get_args()
+    block_dataset = get_indexed_dataset_(args.data_path, 'mmap', True)
+    titles_dataset = get_indexed_dataset_(args.titles_data_path, 'mmap', True)
+
+    kwargs = dict(
+        name='full',
+        block_dataset=block_dataset,
+        title_dataset=titles_dataset,
+        data_prefix=args.data_path,
+        num_epochs=1,
+        max_num_samples=None,
+        max_seq_length=args.seq_length,
+        seed=1,
+        query_in_block_prob=query_in_block_prob,
+        use_titles=use_titles,
+        use_one_sent_docs=args.use_one_sent_docs
+    )
+    dataset = ICTDataset(**kwargs)
+    return dataset
+
+
+class ICTDataset(Dataset):
+    """Dataset containing sentences and their blocks for an inverse cloze task."""
+    def __init__(self, name, block_dataset, title_dataset, data_prefix,
+                 num_epochs, max_num_samples, max_seq_length, query_in_block_prob,
+                 seed, use_titles=True, use_one_sent_docs=False, binary_head=False):
+        self.name = name
+        self.seed = seed
+        self.max_seq_length = max_seq_length
+        self.query_in_block_prob = query_in_block_prob
+        self.block_dataset = block_dataset
+        self.title_dataset = title_dataset
+        self.rng = random.Random(self.seed)
+        self.use_titles = use_titles
+        self.use_one_sent_docs = use_one_sent_docs
+
+        self.samples_mapping = get_block_samples_mapping(
+            block_dataset, title_dataset, data_prefix, num_epochs,
+            max_num_samples, max_seq_length, seed, name, use_one_sent_docs)
+        self.tokenizer = get_tokenizer()
+        self.vocab_id_list = list(self.tokenizer.inv_vocab.keys())
+        self.vocab_id_to_token_list = self.tokenizer.inv_vocab
+        self.cls_id = self.tokenizer.cls
+        self.sep_id = self.tokenizer.sep
+        self.mask_id = self.tokenizer.mask
+        self.pad_id = self.tokenizer.pad
+
+    def __len__(self):
+        return len(self.samples_mapping)
+
+    def __getitem__(self, idx):
+        """Get an ICT example of a pseudo-query and the block of text from which it was extracted"""
+        sample_data = self.samples_mapping[idx]
+        start_idx, end_idx, doc_idx, block_idx = sample_data.as_tuple()
+
+        if self.use_titles:
+            title = self.title_dataset[int(doc_idx)]
+            title_pad_offset = 3 + len(title)
+        else:
+            title = None
+            title_pad_offset = 2
+        block = [self.block_dataset[i] for i in range(start_idx, end_idx)]
+        assert len(block) > 1 or self.use_one_sent_docs or self.query_in_block_prob == 1
+
+        # randint() is inclusive for Python rng
+        rand_sent_idx = self.rng.randint(0, len(block) - 1)
+
+        # keep the query in the context query_in_block_prob fraction of the time.
+        if self.rng.random() < self.query_in_block_prob:
+            query = block[rand_sent_idx].copy()
+        else:
+            query = block.pop(rand_sent_idx)
+
+        # still need to truncate because blocks are concluded when
+        # the sentence lengths have exceeded max_seq_length.
+        query = query[:self.max_seq_length - 2]
+        block = list(itertools.chain(*block))[:self.max_seq_length - title_pad_offset]
+
+        query_tokens, query_pad_mask = self.concat_and_pad_tokens(query)
+        context_tokens, context_pad_mask = self.concat_and_pad_tokens(block, title)
+
+        query_mask = make_attention_mask(query_tokens, query_tokens)
+        context_mask = make_attention_mask(context_tokens, context_tokens)
+
+        block_data = sample_data.as_array()
+
+        sample = {
+            'query_tokens': query_tokens,
+            'query_mask': query_mask,
+            'query_pad_mask': query_pad_mask,
+            'context_tokens': context_tokens,
+            'context_mask': context_mask,
+            'context_pad_mask': context_pad_mask,
+            'block_data': block_data,
+        }
+
+        return sample
+
+    def get_block(self, start_idx, end_idx, doc_idx):
+        """Get the IDs for an evidence block plus the title of the corresponding document"""
+        block = [self.block_dataset[i] for i in range(start_idx, end_idx)]
+        title = self.title_dataset[int(doc_idx)]
+
+        block = list(itertools.chain(*block))[:self.max_seq_length - (3 + len(title))]
+        block_tokens, block_pad_mask = self.concat_and_pad_tokens(block, title)
+
+        return block_tokens, block_pad_mask
+
+    def get_null_block(self):
+        """Get empty block and title - used in REALM pretraining"""
+        block, title = [], []
+        block_tokens, block_pad_mask = self.concat_and_pad_tokens(block, title)
+
+        return block_tokens, block_pad_mask
+
+    def concat_and_pad_tokens(self, tokens, title=None):
+        """Concat with special tokens and pad sequence to self.max_seq_length"""
+        tokens = list(tokens)
+        if title is None:
+            tokens = [self.cls_id] + tokens + [self.sep_id]
+        else:
+            title = list(title)
+            tokens = [self.cls_id] + title + [self.sep_id] + tokens + [self.sep_id]
+        assert len(tokens) <= self.max_seq_length
+
+        num_pad = self.max_seq_length - len(tokens)
+        pad_mask = [1] * len(tokens) + [0] * num_pad
+        tokens += [self.pad_id] * num_pad
+
+        return np.array(tokens), np.array(pad_mask)

Megatron-LM-core_r0.7.0.beta/megatron/legacy/data/image_folder.py

+# BSD 3-Clause License
+#
+# Copyright (c) Soumith Chintala 2016, 
+# All rights reserved.
+#
+# Redistribution and use in source and binary forms, with or without
+# modification, are permitted provided that the following conditions are met:
+#
+# * Redistributions of source code must retain the above copyright notice, this
+#   list of conditions and the following disclaimer.
+#
+# * Redistributions in binary form must reproduce the above copyright notice,
+#   this list of conditions and the following disclaimer in the documentation
+#   and/or other materials provided with the distribution.
+#
+# * Neither the name of the copyright holder nor the names of its
+#   contributors may be used to endorse or promote products derived from
+#   this software without specific prior written permission.
+
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+# AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+# DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
+# FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+# DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+# SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+# CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+# OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+# code taken from 
+# https://github.com/pytorch/vision/blob/main/torchvision/datasets/folder.py
+# added support for classes_fraction and data_per_class_fraction
+
+from torchvision.datasets import VisionDataset
+from PIL import Image
+
+import os
+import os.path
+from typing import Any, Callable, cast, Dict, List, Optional, Tuple
+import numpy as np
+
+def has_file_allowed_extension(filename: str, extensions: Tuple[str, ...]) -> bool:
+    """Checks if a file is an allowed extension.
+    Args:
+        filename (string): path to a file
+        extensions (tuple of strings): extensions to consider (lowercase)
+    Returns:
+        bool: True if the filename ends with one of given extensions
+    """
+    return filename.lower().endswith(extensions)
+
+
+def is_image_file(filename: str) -> bool:
+    """Checks if a file is an allowed image extension.
+    Args:
+        filename (string): path to a file
+    Returns:
+        bool: True if the filename ends with a known image extension
+    """
+    return has_file_allowed_extension(filename, IMG_EXTENSIONS)
+
+
+def make_dataset(
+    directory: str,
+    class_to_idx: Dict[str, int],
+    data_per_class_fraction: float,
+    extensions: Optional[Tuple[str, ...]] = None,
+    is_valid_file: Optional[Callable[[str], bool]] = None,
+) -> List[Tuple[str, int]]:
+    """Generates a list of samples of a form (path_to_sample, class).
+    Args:
+        directory (str): root dataset directory
+        class_to_idx (Dict[str, int]): dictionary mapping class name to class index
+        extensions (optional): A list of allowed extensions.
+            Either extensions or is_valid_file should be passed. Defaults to None.
+        is_valid_file (optional): A function that takes path of a file
+            and checks if the file is a valid file
+            (used to check of corrupt files) both extensions and
+            is_valid_file should not be passed. Defaults to None.
+    Raises:
+        ValueError: In case ``extensions`` and ``is_valid_file`` are None or both are not None.
+    Returns:
+        List[Tuple[str, int]]: samples of a form (path_to_sample, class)
+    """
+    instances = []
+    directory = os.path.expanduser(directory)
+    both_none = extensions is None and is_valid_file is None
+    both_something = extensions is not None and is_valid_file is not None
+    if both_none or both_something:
+        raise ValueError("Both extensions and is_valid_file cannot be None or not None at the same time")
+    if extensions is not None:
+        def is_valid_file(x: str) -> bool:
+            return has_file_allowed_extension(x, cast(Tuple[str, ...], extensions))
+    is_valid_file = cast(Callable[[str], bool], is_valid_file)
+    for target_class in sorted(class_to_idx.keys()):
+        class_index = class_to_idx[target_class]
+        target_dir = os.path.join(directory, target_class)
+        if not os.path.isdir(target_dir):
+            continue
+        local_instances = []
+        for root, _, fnames in sorted(os.walk(target_dir, followlinks=True)):
+            for fname in sorted(fnames):
+                path = os.path.join(root, fname)
+                if is_valid_file(path):
+                    item = path, class_index
+                    local_instances.append(item)
+
+        instances.extend(local_instances[0:int(len(local_instances) * data_per_class_fraction)])
+
+    return instances
+
+
+class DatasetFolder(VisionDataset):
+    """A generic data loader where the samples are arranged in this way: ::
+        root/class_x/xxx.ext
+        root/class_x/xxy.ext
+        root/class_x/[...]/xxz.ext
+        root/class_y/123.ext
+        root/class_y/nsdf3.ext
+        root/class_y/[...]/asd932_.ext
+    Args:
+        root (string): Root directory path.
+        loader (callable): A function to load a sample given its path.
+        extensions (tuple[string]): A list of allowed extensions.
+            both extensions and is_valid_file should not be passed.
+        transform (callable, optional): A function/transform that takes in
+            a sample and returns a transformed version.
+            E.g, ``transforms.RandomCrop`` for images.
+        target_transform (callable, optional): A function/transform that takes
+            in the target and transforms it.
+        is_valid_file (callable, optional): A function that takes path of a file
+            and check if the file is a valid file (used to check of corrupt files)
+            both extensions and is_valid_file should not be passed.
+     Attributes:
+        classes (list): List of the class names sorted alphabetically.
+        class_to_idx (dict): Dict with items (class_name, class_index).
+        samples (list): List of (sample path, class_index) tuples
+        targets (list): The class_index value for each image in the dataset
+    """
+
+    def __init__(
+            self,
+            root: str,
+            loader: Callable[[str], Any],
+            extensions: Optional[Tuple[str, ...]] = None,
+            transform: Optional[Callable] = None,
+            target_transform: Optional[Callable] = None,
+            classes_fraction=1.0,
+            data_per_class_fraction=1.0,
+            is_valid_file: Optional[Callable[[str], bool]] = None,
+    ) -> None:
+        super(DatasetFolder, self).__init__(root, transform=transform,
+                                            target_transform=target_transform)
+        self.classes_fraction = classes_fraction
+        self.data_per_class_fraction = data_per_class_fraction
+        classes, class_to_idx = self._find_classes(self.root)
+        samples = self.make_dataset(self.root,
+                                    class_to_idx,
+                                    self.data_per_class_fraction,
+                                    extensions,
+                                    is_valid_file)
+        if len(samples) == 0:
+            msg = "Found 0 files in subfolders of: {}\n".format(self.root)
+            if extensions is not None:
+                msg += "Supported extensions are: {}".format(",".join(extensions))
+            raise RuntimeError(msg)
+
+        self.loader = loader
+        self.extensions = extensions
+        self.total = len(samples)
+        self.classes = classes
+        self.class_to_idx = class_to_idx
+        self.samples = samples
+        self.targets = [s[1] for s in samples]
+
+    @staticmethod
+    def make_dataset(
+        directory: str,
+        class_to_idx: Dict[str, int],
+        data_per_class_fraction: float,
+        extensions: Optional[Tuple[str, ...]] = None,
+        is_valid_file: Optional[Callable[[str], bool]] = None,
+    ) -> List[Tuple[str, int]]:
+        return make_dataset(directory,
+                            class_to_idx,
+                            data_per_class_fraction,
+                            extensions=extensions,
+                            is_valid_file=is_valid_file)
+
+    def _find_classes(self, dir: str) -> Tuple[List[str], Dict[str, int]]:
+        """
+        Finds the class folders in a dataset.
+        Args:
+            dir (string): Root directory path.
+        Returns:
+            tuple: (classes, class_to_idx) where classes are relative to (dir), and class_to_idx is a dictionary.
+        Ensures:
+            No class is a subdirectory of another.
+        """
+        all_classes = [d.name for d in os.scandir(dir) if d.is_dir()]
+        classes = all_classes[0:int(len(all_classes) * self.classes_fraction)]
+        classes.sort()
+        class_to_idx = {cls_name: i for i, cls_name in enumerate(classes)}
+        return classes, class_to_idx
+
+    def __getitem__(self, index: int) -> Tuple[Any, Any]:
+        """
+        Args:
+            index (int): Index
+        Returns:
+            tuple: (sample, target) where target is class_index of the target class.
+        """
+        curr_index = index
+        for x in range(self.total):
+            try:
+                path, target = self.samples[curr_index]
+                sample = self.loader(path)
+                break
+            except Exception as e:
+                curr_index = np.random.randint(0, self.total)
+
+        if self.transform is not None:
+            sample = self.transform(sample)
+        if self.target_transform is not None:
+            target = self.target_transform(target)
+
+        return sample, target
+
+    def __len__(self) -> int:
+        return len(self.samples)
+
+
+IMG_EXTENSIONS = ('.jpg', '.jpeg', '.png', '.ppm', '.bmp', '.pgm', '.tif', '.tiff', '.webp')
+
+
+def pil_loader(path: str) -> Image.Image:
+    # open path as file to avoid ResourceWarning (https://github.com/python-pillow/Pillow/issues/835)
+    with open(path, 'rb') as f:
+        img = Image.open(f)
+        return img.convert('RGB')
+
+
+# TODO: specify the return type
+def accimage_loader(path: str) -> Any:
+    import accimage
+    try:
+        return accimage.Image(path)
+    except IOError:
+        # Potentially a decoding problem, fall back to PIL.Image
+        return pil_loader(path)
+
+
+def default_loader(path: str) -> Any:
+    from torchvision import get_image_backend
+    if get_image_backend() == 'accimage':
+        return accimage_loader(path)
+    else:
+        return pil_loader(path)
+
+
+class ImageFolder(DatasetFolder):
+    """A generic data loader where the images are arranged in this way: ::
+        root/dog/xxx.png
+        root/dog/xxy.png
+        root/dog/[...]/xxz.png
+        root/cat/123.png
+        root/cat/nsdf3.png
+        root/cat/[...]/asd932_.png
+    Args:
+        root (string): Root directory path.
+        transform (callable, optional): A function/transform that  takes in an PIL image
+            and returns a transformed version. E.g, ``transforms.RandomCrop``
+        target_transform (callable, optional): A function/transform that takes in the
+            target and transforms it.
+        loader (callable, optional): A function to load an image given its path.
+        is_valid_file (callable, optional): A function that takes path of an Image file
+            and check if the file is a valid file (used to check of corrupt files)
+     Attributes:
+        classes (list): List of the class names sorted alphabetically.
+        class_to_idx (dict): Dict with items (class_name, class_index).
+        imgs (list): List of (image path, class_index) tuples
+    """
+
+    def __init__(
+            self,
+            root: str,
+            transform: Optional[Callable] = None,
+            target_transform: Optional[Callable] = None,
+            classes_fraction=1.0,
+            data_per_class_fraction=1.0,
+            loader: Callable[[str], Any] = default_loader,
+            is_valid_file: Optional[Callable[[str], bool]] = None,
+    ):
+        super(ImageFolder, self).__init__(root, loader, IMG_EXTENSIONS if is_valid_file is None else None,
+                                          transform=transform,
+                                          target_transform=target_transform,
+                                          classes_fraction=classes_fraction,
+                                          data_per_class_fraction=data_per_class_fraction,
+                                          is_valid_file=is_valid_file)
+        self.imgs = self.samples
+

Megatron-LM-core_r0.7.0.beta/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-LM-core_r0.7.0.beta/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-LM-core_r0.7.0.beta/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-LM-core_r0.7.0.beta/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-LM-core_r0.7.0.beta/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