all

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, forward_step, 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.
+        forward_step (ForwardStep): Class for running the model forward step.
+        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 = forward_step(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, forward_step, 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 = forward_step(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/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/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',
+                                           'HuggingFaceTokenizer',
+                                           'Llama2Tokenizer',
+                                           'MistralTokenizer']:
+                    word = tokenizer.decoder[token]
+                elif args.tokenizer_type == 'Llama3Tokenizer':
+                    word = tokenizer.decode([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/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/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/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/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 = []