Flax Masked Language Modeling training example (#8728)

* Remove "Model" suffix from Flax models to look more 🤗

Signed-off-by: Morgan Funtowicz <morgan@huggingface.co>

* Initial working (forward + backward) for Flax MLM training example.

Signed-off-by: Morgan Funtowicz <morgan@huggingface.co>

* Simply code

Signed-off-by: Morgan Funtowicz <morgan@huggingface.co>

* Addressing comments, using module and moving to LM task.

Signed-off-by: Morgan Funtowicz <morgan@huggingface.co>

* Restore parameter name "module" wrongly renamed model.

Signed-off-by: Morgan Funtowicz <morgan@huggingface.co>

* Restore correct output ordering...

Signed-off-by: Morgan Funtowicz <morgan@huggingface.co>

* Actually commit the example 😅

Signed-off-by: Morgan Funtowicz <morgan@huggingface.co>

* Add FlaxBertModelForMaskedLM after rebasing.

Signed-off-by: Morgan Funtowicz <funtowiczmo@gmail.com>

* Make it possible to initialize the training from scratch

Signed-off-by: Morgan Funtowicz <funtowiczmo@gmail.com>

* Reuse flax linen example of cross entropy loss

Signed-off-by: Morgan Funtowicz <funtowiczmo@gmail.com>

* Added specific data collator for flax

Signed-off-by: Morgan Funtowicz <funtowiczmo@gmail.com>

* Remove todo for data collator

Signed-off-by: Morgan Funtowicz <funtowiczmo@gmail.com>

* Added evaluation step

Signed-off-by: Morgan Funtowicz <funtowiczmo@gmail.com>

* Added ability to provide dtype to support bfloat16 on TPU

Signed-off-by: Morgan Funtowicz <funtowiczmo@gmail.com>

* Enable flax tensorboard output

Signed-off-by: Morgan Funtowicz <funtowiczmo@gmail.com>

* Enable jax.pmap support.

Signed-off-by: Morgan Funtowicz <funtowiczmo@gmail.com>

* Ensure batches are correctly sized to be dispatched with jax.pmap

Signed-off-by: Morgan Funtowicz <funtowiczmo@gmail.com>

* Enable bfloat16 with --fp16 cmdline args

Signed-off-by: Morgan Funtowicz <funtowiczmo@gmail.com>

* Correctly export metrics to tensorboard

Signed-off-by: Morgan Funtowicz <funtowiczmo@gmail.com>

* Added dropout and ability to use it.

Signed-off-by: Morgan Funtowicz <funtowiczmo@gmail.com>

* Effectively enable & disable during training and evaluation steps.

Signed-off-by: Morgan Funtowicz <funtowiczmo@gmail.com>

* Oops.

Signed-off-by: Morgan Funtowicz <funtowiczmo@gmail.com>

* Enable specifying kernel initializer scale

Signed-off-by: Morgan Funtowicz <funtowiczmo@gmail.com>

* Style.

Signed-off-by: Morgan Funtowicz <funtowiczmo@gmail.com>

* Added warmup step to the learning rate scheduler.

Signed-off-by: Morgan Funtowicz <funtowiczmo@gmail.com>

* Fix typo.

Signed-off-by: Morgan Funtowicz <funtowiczmo@gmail.com>

* Print training loss

Signed-off-by: Morgan Funtowicz <funtowiczmo@gmail.com>

* Make style

Signed-off-by: Morgan Funtowicz <funtowiczmo@gmail.com>

* fix linter issue (flake8)

Signed-off-by: Morgan Funtowicz <funtowiczmo@gmail.com>

* Fix model matching

Signed-off-by: Morgan Funtowicz <funtowiczmo@gmail.com>

* Fix dummies

Signed-off-by: Morgan Funtowicz <funtowiczmo@gmail.com>

* Fix non default dtype on Flax models

Signed-off-by: Morgan Funtowicz <funtowiczmo@gmail.com>

* Use the same create_position_ids_from_input_ids for FlaxRoberta

Signed-off-by: Morgan Funtowicz <funtowiczmo@gmail.com>

* Make Roberta attention as Bert

Signed-off-by: Morgan Funtowicz <funtowiczmo@gmail.com>

* fix copy

Signed-off-by: Morgan Funtowicz <funtowiczmo@gmail.com>

* Wording.

Co-authored-by: Marc van Zee <marcvanzee@gmail.com>

Co-authored-by: Marc van Zee <marcvanzee@gmail.com>

examples/language-modeling/run_mlm_flax.py

@@ -0,0 +1,636 @@
+# coding=utf-8
+# Copyright 2020 The HuggingFace Team 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.
+"""
+Fine-tuning the library models for masked language modeling (BERT, ALBERT, RoBERTa...) with whole word masking on a
+text file or a dataset.
+
+Here is the full list of checkpoints on the hub that can be fine-tuned by this script:
+https://huggingface.co/models?filter=masked-lm
+"""
+import logging
+import os
+import sys
+from dataclasses import dataclass, field
+
+# You can also adapt this script on your own masked language modeling task. Pointers for this are left as comments.
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import numpy as np
+from datasets import load_dataset
+from tqdm import tqdm
+
+import jax
+import jax.numpy as jnp
+from flax import jax_utils
+from flax.optim import Adam
+from flax.training import common_utils
+from flax.training.common_utils import get_metrics
+from jax.nn import log_softmax
+from transformers import (
+    CONFIG_MAPPING,
+    MODEL_FOR_MASKED_LM_MAPPING,
+    AutoConfig,
+    AutoTokenizer,
+    FlaxBertForMaskedLM,
+    HfArgumentParser,
+    PreTrainedTokenizerBase,
+    TensorType,
+    TrainingArguments,
+    is_tensorboard_available,
+    set_seed,
+)
+
+
+# Cache the result
+has_tensorboard = is_tensorboard_available()
+if has_tensorboard:
+    try:
+        from flax.metrics.tensorboard import SummaryWriter
+    except ImportError as ie:
+        has_tensorboard = False
+        print(f"Unable to display metrics through TensorBoard because some package are not installed: {ie}")
+
+else:
+    print(
+        "Unable to display metrics through TensorBoard because the package is not installed: "
+        "Please run pip install tensorboard to enable."
+    )
+
+
+MODEL_CONFIG_CLASSES = list(MODEL_FOR_MASKED_LM_MAPPING.keys())
+MODEL_TYPES = tuple(conf.model_type for conf in MODEL_CONFIG_CLASSES)
+
+
+@dataclass
+class ModelArguments:
+    """
+    Arguments pertaining to which model/config/tokenizer we are going to fine-tune, or train from scratch.
+    """
+
+    model_name_or_path: Optional[str] = field(
+        default=None,
+        metadata={
+            "help": "The model checkpoint for weights initialization."
+            "Don't set if you want to train a model from scratch."
+        },
+    )
+    model_type: Optional[str] = field(
+        default=None,
+        metadata={"help": "If training from scratch, pass a model type from the list: " + ", ".join(MODEL_TYPES)},
+    )
+    config_name: Optional[str] = field(
+        default=None, metadata={"help": "Pretrained config name or path if not the same as model_name"}
+    )
+    tokenizer_name: Optional[str] = field(
+        default=None, metadata={"help": "Pretrained tokenizer name or path if not the same as model_name"}
+    )
+    cache_dir: Optional[str] = field(
+        default=None, metadata={"help": "Where do you want to store the pretrained models downloaded from s3"}
+    )
+    use_fast_tokenizer: bool = field(
+        default=True,
+        metadata={"help": "Whether to use one of the fast tokenizer (backed by the tokenizers library) or not."},
+    )
+
+
+@dataclass
+class DataTrainingArguments:
+    """
+    Arguments pertaining to what data we are going to input our model for training and eval.
+    """
+
+    dataset_name: Optional[str] = field(
+        default=None, metadata={"help": "The name of the dataset to use (via the datasets library)."}
+    )
+    dataset_config_name: Optional[str] = field(
+        default=None, metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
+    )
+    train_file: Optional[str] = field(default=None, metadata={"help": "The input training data file (a text file)."})
+    validation_file: Optional[str] = field(
+        default=None,
+        metadata={"help": "An optional input evaluation data file to evaluate the perplexity on (a text file)."},
+    )
+    train_ref_file: Optional[str] = field(
+        default=None,
+        metadata={"help": "An optional input train ref data file for whole word masking in Chinese."},
+    )
+    validation_ref_file: Optional[str] = field(
+        default=None,
+        metadata={"help": "An optional input validation ref data file for whole word masking in Chinese."},
+    )
+    overwrite_cache: bool = field(
+        default=False, metadata={"help": "Overwrite the cached training and evaluation sets"}
+    )
+    max_seq_length: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": "The maximum total input sequence length after tokenization. Sequences longer "
+            "than this will be truncated. Default to the max input length of the model."
+        },
+    )
+    preprocessing_num_workers: Optional[int] = field(
+        default=None,
+        metadata={"help": "The number of processes to use for the preprocessing."},
+    )
+    mlm_probability: float = field(
+        default=0.15, metadata={"help": "Ratio of tokens to mask for masked language modeling loss"}
+    )
+    pad_to_max_length: bool = field(
+        default=False,
+        metadata={
+            "help": "Whether to pad all samples to `max_seq_length`. "
+            "If False, will pad the samples dynamically when batching to the maximum length in the batch."
+        },
+    )
+
+    def __post_init__(self):
+        if self.dataset_name is None and self.train_file is None and self.validation_file is None:
+            raise ValueError("Need either a dataset name or a training/validation file.")
+        else:
+            if self.train_file is not None:
+                extension = self.train_file.split(".")[-1]
+                assert extension in ["csv", "json", "txt"], "`train_file` should be a csv, a json or a txt file."
+            if self.validation_file is not None:
+                extension = self.validation_file.split(".")[-1]
+                assert extension in ["csv", "json", "txt"], "`validation_file` should be a csv, a json or a txt file."
+
+
+# Adapted from transformers/data/data_collator.py
+# Letting here for now, let's discuss where it should live
+@dataclass
+class FlaxDataCollatorForLanguageModeling:
+    """
+    Data collator used for language modeling. Inputs are dynamically padded to the maximum length of a batch if they
+    are not all of the same length.
+
+    Args:
+        tokenizer (:class:`~transformers.PreTrainedTokenizer` or :class:`~transformers.PreTrainedTokenizerFast`):
+            The tokenizer used for encoding the data.
+        mlm (:obj:`bool`, `optional`, defaults to :obj:`True`):
+            Whether or not to use masked language modeling. If set to :obj:`False`, the labels are the same as the
+            inputs with the padding tokens ignored (by setting them to -100). Otherwise, the labels are -100 for
+            non-masked tokens and the value to predict for the masked token.
+        mlm_probability (:obj:`float`, `optional`, defaults to 0.15):
+            The probability with which to (randomly) mask tokens in the input, when :obj:`mlm` is set to :obj:`True`.
+
+    .. note::
+
+        For best performance, this data collator should be used with a dataset having items that are dictionaries or
+        BatchEncoding, with the :obj:`"special_tokens_mask"` key, as returned by a
+        :class:`~transformers.PreTrainedTokenizer` or a :class:`~transformers.PreTrainedTokenizerFast` with the
+        argument :obj:`return_special_tokens_mask=True`.
+    """
+
+    tokenizer: PreTrainedTokenizerBase
+    mlm: bool = True
+    mlm_probability: float = 0.15
+
+    def __post_init__(self):
+        if self.mlm and self.tokenizer.mask_token is None:
+            raise ValueError(
+                "This tokenizer does not have a mask token which is necessary for masked language modeling. "
+                "You should pass `mlm=False` to train on causal language modeling instead."
+            )
+
+    def __call__(self, examples: List[Dict[str, np.ndarray]], pad_to_multiple_of: int) -> Dict[str, np.ndarray]:
+        # Handle dict or lists with proper padding and conversion to tensor.
+        batch = self.tokenizer.pad(examples, pad_to_multiple_of=pad_to_multiple_of, return_tensors=TensorType.NUMPY)
+
+        # If special token mask has been preprocessed, pop it from the dict.
+        special_tokens_mask = batch.pop("special_tokens_mask", None)
+        if self.mlm:
+            batch["input_ids"], batch["labels"] = self.mask_tokens(
+                batch["input_ids"], special_tokens_mask=special_tokens_mask
+            )
+        else:
+            labels = batch["input_ids"].copy()
+            if self.tokenizer.pad_token_id is not None:
+                labels[labels == self.tokenizer.pad_token_id] = -100
+            batch["labels"] = labels
+        return batch
+
+    def mask_tokens(
+        self, inputs: np.ndarray, special_tokens_mask: Optional[np.ndarray]
+    ) -> Tuple[jnp.ndarray, jnp.ndarray]:
+        """
+        Prepare masked tokens inputs/labels for masked language modeling: 80% MASK, 10% random, 10% original.
+        """
+        labels = inputs.copy()
+        # We sample a few tokens in each sequence for MLM training (with probability `self.mlm_probability`)
+        probability_matrix = np.full(labels.shape, self.mlm_probability)
+        special_tokens_mask = special_tokens_mask.astype("bool")
+
+        probability_matrix[special_tokens_mask] = 0.0
+        masked_indices = np.random.binomial(1, probability_matrix).astype("bool")
+        labels[~masked_indices] = -100  # We only compute loss on masked tokens
+
+        # 80% of the time, we replace masked input tokens with tokenizer.mask_token ([MASK])
+        indices_replaced = np.random.binomial(1, np.full(labels.shape, 0.8)).astype("bool") & masked_indices
+        inputs[indices_replaced] = self.tokenizer.convert_tokens_to_ids(self.tokenizer.mask_token)
+
+        # 10% of the time, we replace masked input tokens with random word
+        indices_random = np.random.binomial(1, np.full(labels.shape, 0.5)).astype("bool")
+        indices_random &= masked_indices & ~indices_replaced
+
+        random_words = np.random.randint(self.tokenizer.vocab_size, size=labels.shape, dtype="i4")
+        inputs[indices_random] = random_words[indices_random]
+
+        # The rest of the time (10% of the time) we keep the masked input tokens unchanged
+        return inputs, labels
+
+
+def create_learning_rate_scheduler(
+    factors="constant * linear_warmup * rsqrt_decay",
+    base_learning_rate=0.5,
+    warmup_steps=1000,
+    decay_factor=0.5,
+    steps_per_decay=20000,
+    steps_per_cycle=100000,
+):
+    """Creates learning rate schedule.
+    Interprets factors in the factors string which can consist of:
+    * constant: interpreted as the constant value,
+    * linear_warmup: interpreted as linear warmup until warmup_steps,
+    * rsqrt_decay: divide by square root of max(step, warmup_steps)
+    * rsqrt_normalized_decay: divide by square root of max(step/warmup_steps, 1)
+    * decay_every: Every k steps decay the learning rate by decay_factor.
+    * cosine_decay: Cyclic cosine decay, uses steps_per_cycle parameter.
+    Args:
+      factors: string, factors separated by "*" that defines the schedule.
+      base_learning_rate: float, the starting constant for the lr schedule.
+      warmup_steps: int, how many steps to warm up for in the warmup schedule.
+      decay_factor: float, the amount to decay the learning rate by.
+      steps_per_decay: int, how often to decay the learning rate.
+      steps_per_cycle: int, steps per cycle when using cosine decay.
+    Returns:
+      a function learning_rate(step): float -> {"learning_rate": float}, the
+      step-dependent lr.
+    """
+    factors = [n.strip() for n in factors.split("*")]
+
+    def step_fn(step):
+        """Step to learning rate function."""
+        ret = 1.0
+        for name in factors:
+            if name == "constant":
+                ret *= base_learning_rate
+            elif name == "linear_warmup":
+                ret *= jnp.minimum(1.0, step / warmup_steps)
+            elif name == "rsqrt_decay":
+                ret /= jnp.sqrt(jnp.maximum(step, warmup_steps))
+            elif name == "rsqrt_normalized_decay":
+                ret *= jnp.sqrt(warmup_steps)
+                ret /= jnp.sqrt(jnp.maximum(step, warmup_steps))
+            elif name == "decay_every":
+                ret *= decay_factor ** (step // steps_per_decay)
+            elif name == "cosine_decay":
+                progress = jnp.maximum(0.0, (step - warmup_steps) / float(steps_per_cycle))
+                ret *= jnp.maximum(0.0, 0.5 * (1.0 + jnp.cos(jnp.pi * (progress % 1.0))))
+            else:
+                raise ValueError("Unknown factor %s." % name)
+        return jnp.asarray(ret, dtype=jnp.float32)
+
+    return step_fn
+
+
+def compute_metrics(logits, labels, weights, label_smoothing=0.0):
+    """Compute summary metrics."""
+    loss, normalizer = cross_entropy(logits, labels, weights, label_smoothing)
+    acc, _ = accuracy(logits, labels, weights)
+    metrics = {"loss": loss, "accuracy": acc, "normalizer": normalizer}
+    metrics = jax.lax.psum(metrics, axis_name="batch")
+    return metrics
+
+
+def accuracy(logits, targets, weights=None):
+    """Compute weighted accuracy for log probs and targets.
+    Args:
+     logits: [batch, length, num_classes] float array.
+     targets: categorical targets [batch, length] int array.
+     weights: None or array of shape [batch, length]
+    Returns:
+      Tuple of scalar loss and batch normalizing factor.
+    """
+    if logits.ndim != targets.ndim + 1:
+        raise ValueError(
+            "Incorrect shapes. Got shape %s logits and %s targets" % (str(logits.shape), str(targets.shape))
+        )
+
+    loss = jnp.equal(jnp.argmax(logits, axis=-1), targets)
+    loss *= weights
+
+    return loss.sum(), weights.sum()
+
+
+def cross_entropy(logits, targets, weights=None, label_smoothing=0.0):
+    """Compute cross entropy and entropy for log probs and targets.
+    Args:
+     logits: [batch, length, num_classes] float array.
+     targets: categorical targets [batch, length] int array.
+     weights: None or array of shape [batch, length]
+     label_smoothing: label smoothing constant, used to determine the on and off values.
+    Returns:
+      Tuple of scalar loss and batch normalizing factor.
+    """
+    if logits.ndim != targets.ndim + 1:
+        raise ValueError(
+            "Incorrect shapes. Got shape %s logits and %s targets" % (str(logits.shape), str(targets.shape))
+        )
+
+    vocab_size = logits.shape[-1]
+    confidence = 1.0 - label_smoothing
+    low_confidence = (1.0 - confidence) / (vocab_size - 1)
+    normalizing_constant = -(
+        confidence * jnp.log(confidence) + (vocab_size - 1) * low_confidence * jnp.log(low_confidence + 1e-20)
+    )
+    soft_targets = common_utils.onehot(targets, vocab_size, on_value=confidence, off_value=low_confidence)
+
+    loss = -jnp.sum(soft_targets * log_softmax(logits), axis=-1)
+    loss = loss - normalizing_constant
+
+    if weights is not None:
+        loss = loss * weights
+        normalizing_factor = weights.sum()
+    else:
+        normalizing_factor = np.prod(targets.shape)
+
+    return loss.sum(), normalizing_factor
+
+
+def training_step(optimizer, batch, dropout_rng):
+    dropout_rng, new_dropout_rng = jax.random.split(dropout_rng)
+
+    def loss_fn(params):
+        targets = batch.pop("labels")
+
+        # Hide away tokens which doesn't participate in the optimization
+        token_mask = jnp.where(targets > 0, 1.0, 0.0)
+
+        pooled, logits = model(**batch, params=params, dropout_rng=dropout_rng, train=True)
+        loss, weight_sum = cross_entropy(logits, targets, token_mask)
+        return loss / weight_sum
+
+    step = optimizer.state.step
+    lr = lr_scheduler_fn(step)
+    grad_fn = jax.value_and_grad(loss_fn)
+    loss, grad = grad_fn(optimizer.target)
+    grad = jax.lax.pmean(grad, "batch")
+    optimizer = optimizer.apply_gradient(grad, learning_rate=lr)
+
+    return loss, optimizer, new_dropout_rng
+
+
+def eval_step(params, batch):
+    """
+    Calculate evaluation metrics on a batch.
+    """
+    targets = batch.pop("labels")
+
+    # Hide away tokens which doesn't participate in the optimization
+    token_mask = jnp.where(targets > 0, 1.0, 0.0)
+    _, logits = model(**batch, params=params, train=False)
+
+    return compute_metrics(logits, targets, token_mask)
+
+
+def generate_batch_splits(samples_idx: jnp.ndarray, batch_size: int) -> jnp.ndarray:
+    nb_samples = len(samples_idx)
+    samples_to_remove = nb_samples % batch_size
+
+    if samples_to_remove != 0:
+        samples_idx = samples_idx[:-samples_to_remove]
+    sections_split = nb_samples // batch_size
+    batch_idx = jnp.split(samples_idx, sections_split)
+    return batch_idx
+
+
+if __name__ == "__main__":
+    # See all possible arguments in src/transformers/training_args.py
+    # or by passing the --help flag to this script.
+    # We now keep distinct sets of args, for a cleaner separation of concerns.
+
+    parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments))
+    if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
+        # If we pass only one argument to the script and it's the path to a json file,
+        # let's parse it to get our arguments.
+        model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
+    else:
+        model_args, data_args, training_args = parser.parse_args_into_dataclasses()
+
+    if (
+        os.path.exists(training_args.output_dir)
+        and os.listdir(training_args.output_dir)
+        and training_args.do_train
+        and not training_args.overwrite_output_dir
+    ):
+        raise ValueError(
+            f"Output directory ({training_args.output_dir}) already exists and is not empty."
+            "Use --overwrite_output_dir to overcome."
+        )
+
+    # Setup logging
+    logging.basicConfig(
+        format="%(asctime)s - %(levelname)s - %(name)s -   %(message)s",
+        level="NOTSET",
+        datefmt="[%X]",
+    )
+
+    # Log on each process the small summary:
+    logger = logging.getLogger(__name__)
+    logger.warning(
+        f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
+        + f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
+    )
+
+    # Set the verbosity to info of the Transformers logger (on main process only):
+    logger.info("Training/evaluation parameters %s", training_args)
+
+    # Set seed before initializing model.
+    set_seed(training_args.seed)
+
+    # Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below)
+    # or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
+    # (the dataset will be downloaded automatically from the datasets Hub).
+    #
+    # For CSV/JSON files, this script will use the column called 'text' or the first column if no column called
+    # 'text' is found. You can easily tweak this behavior (see below).
+    #
+    # In distributed training, the load_dataset function guarantees that only one local process can concurrently
+    # download the dataset.
+    if data_args.dataset_name is not None:
+        # Downloading and loading a dataset from the hub.
+        datasets = load_dataset(data_args.dataset_name, data_args.dataset_config_name)
+    else:
+        data_files = {}
+        if data_args.train_file is not None:
+            data_files["train"] = data_args.train_file
+        if data_args.validation_file is not None:
+            data_files["validation"] = data_args.validation_file
+        extension = data_args.train_file.split(".")[-1]
+        if extension == "txt":
+            extension = "text"
+        datasets = load_dataset(extension, data_files=data_files)
+    # See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
+    # https://huggingface.co/docs/datasets/loading_datasets.html.
+
+    # Load pretrained model and tokenizer
+
+    # Distributed training:
+    # The .from_pretrained methods guarantee that only one local process can concurrently
+    # download model & vocab.
+    if model_args.config_name:
+        config = AutoConfig.from_pretrained(model_args.config_name, cache_dir=model_args.cache_dir)
+    elif model_args.model_name_or_path:
+        config = AutoConfig.from_pretrained(model_args.model_name_or_path, cache_dir=model_args.cache_dir)
+    else:
+        config = CONFIG_MAPPING[model_args.model_type]()
+        logger.warning("You are instantiating a new config instance from scratch.")
+
+    if model_args.tokenizer_name:
+        tokenizer = AutoTokenizer.from_pretrained(
+            model_args.tokenizer_name, cache_dir=model_args.cache_dir, use_fast=model_args.use_fast_tokenizer
+        )
+    elif model_args.model_name_or_path:
+        tokenizer = AutoTokenizer.from_pretrained(
+            model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_fast=model_args.use_fast_tokenizer
+        )
+    else:
+        raise ValueError(
+            "You are instantiating a new tokenizer from scratch. This is not supported by this script."
+            "You can do it from another script, save it, and load it from here, using --tokenizer_name."
+        )
+
+    # Preprocessing the datasets.
+    # First we tokenize all the texts.
+    if training_args.do_train:
+        column_names = datasets["train"].column_names
+    else:
+        column_names = datasets["validation"].column_names
+    text_column_name = "text" if "text" in column_names else column_names[0]
+
+    padding = "max_length" if data_args.pad_to_max_length else False
+
+    def tokenize_function(examples):
+        # Remove empty lines
+        examples["text"] = [line for line in examples["text"] if len(line) > 0 and not line.isspace()]
+        return tokenizer(
+            examples["text"],
+            return_special_tokens_mask=True,
+            padding=padding,
+            truncation=True,
+            max_length=data_args.max_seq_length,
+        )
+
+    tokenized_datasets = datasets.map(
+        tokenize_function,
+        batched=True,
+        num_proc=data_args.preprocessing_num_workers,
+        remove_columns=[text_column_name],
+        load_from_cache_file=not data_args.overwrite_cache,
+    )
+
+    # Enable tensorboard only on the master node
+    if has_tensorboard and jax.host_id() == 0:
+        summary_writer = SummaryWriter(log_dir=Path(training_args.output_dir).joinpath("logs").as_posix())
+
+    # Data collator
+    # This one will take care of randomly masking the tokens.
+    data_collator = FlaxDataCollatorForLanguageModeling(tokenizer=tokenizer, mlm_probability=data_args.mlm_probability)
+
+    # Initialize our training
+    rng = jax.random.PRNGKey(training_args.seed)
+    dropout_rngs = jax.random.split(rng, jax.local_device_count())
+
+    model = FlaxBertForMaskedLM.from_pretrained("bert-base-cased", dtype=jnp.float32, dropout_rate=0.1)
+    model.init(jax.random.PRNGKey(training_args.seed), (training_args.train_batch_size, model.config.max_length))
+
+    # Setup optimizer
+    optimizer = Adam(
+        learning_rate=training_args.learning_rate,
+        weight_decay=training_args.weight_decay,
+        beta1=training_args.adam_beta1,
+        beta2=training_args.adam_beta2,
+    ).create(model.params)
+
+    # Create learning rate scheduler
+    lr_scheduler_fn = create_learning_rate_scheduler(
+        base_learning_rate=training_args.learning_rate, warmup_steps=training_args.warmup_steps
+    )
+
+    # Create parallel version of the training and evaluation steps
+    p_training_step = jax.pmap(training_step, "batch", donate_argnums=(0,))
+    p_eval_step = jax.pmap(eval_step, "batch", donate_argnums=(0,))
+
+    # Replicate the optimizer on each device
+    optimizer = jax_utils.replicate(optimizer)
+
+    # Store some constant
+    nb_epochs = int(training_args.num_train_epochs)
+    batch_size = int(training_args.train_batch_size)
+    eval_batch_size = int(training_args.eval_batch_size)
+
+    epochs = tqdm(range(nb_epochs), desc=f"Epoch ... (1/{nb_epochs})", position=0)
+    for epoch in epochs:
+
+        # ======================== Training ================================
+        # Create sampling rng
+        rng, training_rng, eval_rng = jax.random.split(rng, 3)
+
+        # Generate an epoch by shuffling sampling indices from the train dataset
+        nb_training_samples = len(tokenized_datasets["train"])
+        training_samples_idx = jax.random.permutation(training_rng, jnp.arange(nb_training_samples))
+        training_batch_idx = generate_batch_splits(training_samples_idx, batch_size)
+
+        # Gather the indexes for creating the batch and do a training step
+        for batch_idx in tqdm(training_batch_idx, desc="Training...", position=1):
+            samples = [tokenized_datasets["train"][int(idx)] for idx in batch_idx]
+            model_inputs = data_collator(samples, pad_to_multiple_of=16)
+
+            # Model forward
+            model_inputs = common_utils.shard(model_inputs.data)
+            loss, optimizer, dropout_rngs = p_training_step(optimizer, model_inputs, dropout_rngs)
+
+        epochs.write(f"Loss: {loss}")
+
+        # ======================== Evaluating ==============================
+        nb_eval_samples = len(tokenized_datasets["test"])
+        eval_samples_idx = jnp.arange(nb_eval_samples)
+        eval_batch_idx = generate_batch_splits(eval_samples_idx, eval_batch_size)
+
+        eval_metrics = []
+        for i, batch_idx in enumerate(tqdm(eval_batch_idx, desc="Evaluating ...", position=2)):
+            samples = [tokenized_datasets["test"][int(idx)] for idx in batch_idx]
+            model_inputs = data_collator(samples, pad_to_multiple_of=16)
+
+            # Model forward
+            model_inputs = common_utils.shard(model_inputs.data)
+            metrics = p_eval_step(optimizer.target, model_inputs)
+            eval_metrics.append(metrics)
+
+        eval_metrics_np = get_metrics(eval_metrics)
+        eval_metrics_np = jax.tree_map(jnp.sum, eval_metrics_np)
+        eval_normalizer = eval_metrics_np.pop("normalizer")
+        eval_summary = jax.tree_map(lambda x: x / eval_normalizer, eval_metrics_np)
+
+        # Update progress bar
+        epochs.desc = (
+            f"Epoch... ({epoch + 1}/{nb_epochs} | Loss: {eval_summary['loss']}, Acc: {eval_summary['accuracy']})"
+        )
+
+        # Save metrics
+        if has_tensorboard and jax.host_id() == 0:
+            for name, value in eval_summary.items():
+                summary_writer.scalar(name, value, epoch)

src/transformers/__init__.py

@@ -936,7 +936,7 @@ else:
 
 if is_flax_available():
     from .models.auto import FLAX_MODEL_MAPPING, FlaxAutoModel
-    from .models.bert import FlaxBertModel
+    from .models.bert import FlaxBertForMaskedLM, FlaxBertModel
     from .models.roberta import FlaxRobertaModel
 else:
     # Import the same objects as dummies to get them in the namespace.

src/transformers/modeling_flax_utils.py

@@ -65,13 +65,12 @@ class FlaxPreTrainedModel(ABC):
     base_model_prefix = ""
     model_class = None
 
-    def __init__(self, config: PretrainedConfig, module: nn.Module, params: Dict, seed: int = 0):
+    def __init__(
+        self, config: PretrainedConfig, module: nn.Module, params: Dict, seed: int = 0, dtype: jnp.dtype = jnp.float32
+    ):
         if config is None:
             raise ValueError("config cannot be None")
 
-        if module is None:
-            raise ValueError("module cannot be None")
-
         if params is None:
             raise ValueError("state cannot be None")
 
@@ -82,19 +81,23 @@ class FlaxPreTrainedModel(ABC):
         # Those are public as their type is generic to every derived classes.
         self.key = PRNGKey(seed)
         self.params = params
-        self.model = module
+        self.dtype = dtype
 
     @property
     def config(self) -> PretrainedConfig:
         return self._config
 
+    @property
+    def module(self) -> nn.Module:
+        return self._module
+
     @staticmethod
     @abstractmethod
     def convert_from_pytorch(pt_state: Dict, config: PretrainedConfig) -> Dict:
         raise NotImplementedError()
 
     @classmethod
-    def from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs):
+    def from_pretrained(cls, pretrained_model_name_or_path, dtype: jnp.dtype = jnp.float32, *model_args, **kwargs):
         r"""
         Instantiate a pretrained Flax model from a pre-trained model configuration.
         """
@@ -127,6 +130,9 @@ class FlaxPreTrainedModel(ABC):
         else:
             model_kwargs = kwargs
 
+        # Add the dtype to model_kwargs
+        model_kwargs["dtype"] = dtype
+
         # Load model
         if pretrained_model_name_or_path is not None:
             if os.path.isfile(pretrained_model_name_or_path) or is_remote_url(pretrained_model_name_or_path):

src/transformers/models/bert/__init__.py

@@ -59,4 +59,4 @@ if is_tf_available():
     )
 
 if is_flax_available():
-    from .modeling_flax_bert import FlaxBertModel
+    from .modeling_flax_bert import FlaxBertForMaskedLM, FlaxBertModel

src/transformers/models/bert/modeling_flax_bert.py

@@ -13,13 +13,14 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from typing import Callable, Dict
+from typing import Callable, Dict, Tuple
 
 import numpy as np
 
 import flax.linen as nn
 import jax
 import jax.numpy as jnp
+from jax.random import PRNGKey
 
 from ...file_utils import add_start_docstrings, add_start_docstrings_to_model_forward
 from ...modeling_flax_utils import FlaxPreTrainedModel, gelu
@@ -101,8 +102,8 @@ class FlaxBertLayerNorm(nn.Module):
     scale: bool = True  # If True, multiply by scale (gamma). When the next layer is linear
     # (also e.g. nn.relu), this can be disabled since the scaling will be
     # done by the next layer.
-    bias_init: jnp.ndarray = nn.initializers.zeros
+    scale_init: Callable[..., np.ndarray] = jax.nn.initializers.ones
-    scale_init: jnp.ndarray = nn.initializers.ones
+    bias_init: Callable[..., np.ndarray] = jax.nn.initializers.zeros
 
     @nn.compact
     def __call__(self, x):
@@ -122,11 +123,13 @@ class FlaxBertLayerNorm(nn.Module):
         mean2 = jnp.mean(jax.lax.square(x), axis=-1, keepdims=True)
         var = mean2 - jax.lax.square(mean)
         mul = jax.lax.rsqrt(var + self.epsilon)
+
         if self.scale:
-            mul = mul * jnp.asarray(self.param("gamma", self.scale_init, (features,)), self.dtype)
+            mul = mul * jnp.asarray(self.param("gamma", self.scale_init, (features,)))
         y = (x - mean) * mul
+
         if self.bias:
-            y = y + jnp.asarray(self.param("beta", self.bias_init, (features,)), self.dtype)
+            y = y + jnp.asarray(self.param("beta", self.bias_init, (features,)))
         return y
 
 
@@ -138,7 +141,9 @@ class FlaxBertEmbedding(nn.Module):
 
     vocab_size: int
     hidden_size: int
-    emb_init: Callable[..., np.ndarray] = nn.initializers.normal(stddev=0.1)
+    kernel_init_scale: float = 0.2
+    emb_init: Callable[..., np.ndarray] = jax.nn.initializers.normal(stddev=kernel_init_scale)
+    dtype: jnp.dtype = jnp.float32  # the dtype of the computation
 
     @nn.compact
     def __call__(self, inputs):
@@ -153,63 +158,105 @@ class FlaxBertEmbeddings(nn.Module):
     hidden_size: int
     type_vocab_size: int
     max_length: int
+    kernel_init_scale: float = 0.2
+    dropout_rate: float = 0.0
+    dtype: jnp.dtype = jnp.float32  # the dtype of the computation
 
     @nn.compact
-    def __call__(self, input_ids, token_type_ids, position_ids, attention_mask):
+    def __call__(self, input_ids, token_type_ids, position_ids, attention_mask, deterministic: bool = True):
 
         # Embed
-        w_emb = FlaxBertEmbedding(self.vocab_size, self.hidden_size, name="word_embeddings")(
+        w_emb = FlaxBertEmbedding(
-            jnp.atleast_2d(input_ids.astype("i4"))
+            self.vocab_size,
-        )
+            self.hidden_size,
-        p_emb = FlaxBertEmbedding(self.max_length, self.hidden_size, name="position_embeddings")(
+            kernel_init_scale=self.kernel_init_scale,
-            jnp.atleast_2d(position_ids.astype("i4"))
+            name="word_embeddings",
-        )
+            dtype=self.dtype,
-        t_emb = FlaxBertEmbedding(self.type_vocab_size, self.hidden_size, name="token_type_embeddings")(
+        )(jnp.atleast_2d(input_ids.astype("i4")))
-            jnp.atleast_2d(token_type_ids.astype("i4"))
+        p_emb = FlaxBertEmbedding(
-        )
+            self.max_length,
+            self.hidden_size,
+            kernel_init_scale=self.kernel_init_scale,
+            name="position_embeddings",
+            dtype=self.dtype,
+        )(jnp.atleast_2d(position_ids.astype("i4")))
+        t_emb = FlaxBertEmbedding(
+            self.type_vocab_size,
+            self.hidden_size,
+            kernel_init_scale=self.kernel_init_scale,
+            name="token_type_embeddings",
+            dtype=self.dtype,
+        )(jnp.atleast_2d(token_type_ids.astype("i4")))
 
         # Sum all embeddings
         summed_emb = w_emb + jnp.broadcast_to(p_emb, w_emb.shape) + t_emb
 
         # Layer Norm
-        layer_norm = FlaxBertLayerNorm(name="layer_norm")(summed_emb)
+        layer_norm = FlaxBertLayerNorm(name="layer_norm", dtype=self.dtype)(summed_emb)
-
+        embeddings = nn.Dropout(rate=self.dropout_rate)(layer_norm, deterministic=deterministic)
-        return layer_norm
+        return embeddings
 
 
 class FlaxBertAttention(nn.Module):
     num_heads: int
     head_size: int
+    dropout_rate: float = 0.0
+    kernel_init_scale: float = 0.2
+    dtype: jnp.dtype = jnp.float32  # the dtype of the computation
 
     @nn.compact
-    def __call__(self, hidden_state, attention_mask):
+    def __call__(self, hidden_state, attention_mask, deterministic: bool = True):
         # Attention mask comes in as attention_mask.shape == (*batch_sizes, kv_length)
         # FLAX expects: attention_mask.shape == (*batch_sizes, 1, 1, kv_length) such that it is broadcastable
         # with attn_weights.shape == (*batch_sizes, num_heads, q_length, kv_length)
         attention_mask = jnp.expand_dims(attention_mask, axis=(-3, -2))
-        self_att = nn.attention.SelfAttention(num_heads=self.num_heads, qkv_features=self.head_size, name="self")(
+        self_att = nn.attention.SelfAttention(
-            hidden_state, attention_mask
+            num_heads=self.num_heads,
-        )
+            qkv_features=self.head_size,
+            dropout_rate=self.dropout_rate,
+            deterministic=deterministic,
+            kernel_init=jax.nn.initializers.normal(self.kernel_init_scale, self.dtype),
+            bias_init=jax.nn.initializers.zeros,
+            name="self",
+            dtype=self.dtype,
+        )(hidden_state, attention_mask)
 
-        layer_norm = FlaxBertLayerNorm(name="layer_norm")(self_att + hidden_state)
+        layer_norm = FlaxBertLayerNorm(name="layer_norm", dtype=self.dtype)(self_att + hidden_state)
         return layer_norm
 
 
 class FlaxBertIntermediate(nn.Module):
     output_size: int
+    kernel_init_scale: float = 0.2
+    dtype: jnp.dtype = jnp.float32  # the dtype of the computation
 
     @nn.compact
     def __call__(self, hidden_state):
         # TODO: Add ACT2FN reference to change activation function
-        dense = nn.Dense(features=self.output_size, name="dense")(hidden_state)
+        dense = nn.Dense(
+            features=self.output_size,
+            kernel_init=jax.nn.initializers.normal(self.kernel_init_scale, self.dtype),
+            name="dense",
+            dtype=self.dtype,
+        )(hidden_state)
         return gelu(dense)
 
 
 class FlaxBertOutput(nn.Module):
+    dropout_rate: float = 0.0
+    kernel_init_scale: float = 0.2
+    dtype: jnp.dtype = jnp.float32  # the dtype of the computation
+
     @nn.compact
-    def __call__(self, intermediate_output, attention_output):
+    def __call__(self, intermediate_output, attention_output, deterministic: bool = True):
-        hidden_state = nn.Dense(attention_output.shape[-1], name="dense")(intermediate_output)
+        hidden_state = nn.Dense(
-        hidden_state = FlaxBertLayerNorm(name="layer_norm")(hidden_state + attention_output)
+            attention_output.shape[-1],
+            kernel_init=jax.nn.initializers.normal(self.kernel_init_scale, self.dtype),
+            name="dense",
+            dtype=self.dtype,
+        )(intermediate_output)
+        hidden_state = nn.Dropout(rate=self.dropout_rate)(hidden_state, deterministic=deterministic)
+        hidden_state = FlaxBertLayerNorm(name="layer_norm", dtype=self.dtype)(hidden_state + attention_output)
         return hidden_state
 
 
@@ -217,12 +264,26 @@ class FlaxBertLayer(nn.Module):
     num_heads: int
     head_size: int
     intermediate_size: int
+    dropout_rate: float = 0.0
+    kernel_init_scale: float = 0.2
+    dtype: jnp.dtype = jnp.float32  # the dtype of the computation
 
     @nn.compact
-    def __call__(self, hidden_state, attention_mask):
+    def __call__(self, hidden_state, attention_mask, deterministic: bool = True):
-        attention = FlaxBertAttention(self.num_heads, self.head_size, name="attention")(hidden_state, attention_mask)
+        attention = FlaxBertAttention(
-        intermediate = FlaxBertIntermediate(self.intermediate_size, name="intermediate")(attention)
+            self.num_heads,
-        output = FlaxBertOutput(name="output")(intermediate, attention)
+            self.head_size,
+            kernel_init_scale=self.kernel_init_scale,
+            dropout_rate=self.dropout_rate,
+            name="attention",
+            dtype=self.dtype,
+        )(hidden_state, attention_mask, deterministic=deterministic)
+        intermediate = FlaxBertIntermediate(
+            self.intermediate_size, kernel_init_scale=self.kernel_init_scale, name="intermediate", dtype=self.dtype
+        )(attention)
+        output = FlaxBertOutput(
+            kernel_init_scale=self.kernel_init_scale, dropout_rate=self.dropout_rate, name="output", dtype=self.dtype
+        )(intermediate, attention, deterministic=deterministic)
 
         return output
 
@@ -236,9 +297,12 @@ class FlaxBertLayerCollection(nn.Module):
     num_heads: int
     head_size: int
     intermediate_size: int
+    dropout_rate: float = 0.0
+    kernel_init_scale: float = 0.2
+    dtype: jnp.dtype = jnp.float32  # the dtype of the computation
 
     @nn.compact
-    def __call__(self, inputs, attention_mask):
+    def __call__(self, inputs, attention_mask, deterministic: bool = True):
         assert self.num_layers > 0, f"num_layers should be >= 1, got ({self.num_layers})"
 
         # Initialize input / output
@@ -246,8 +310,16 @@ class FlaxBertLayerCollection(nn.Module):
 
         # Forward over all encoders
         for i in range(self.num_layers):
-            layer = FlaxBertLayer(self.num_heads, self.head_size, self.intermediate_size, name=f"{i}")
+            layer = FlaxBertLayer(
-            input_i = layer(input_i, attention_mask)
+                self.num_heads,
+                self.head_size,
+                self.intermediate_size,
+                kernel_init_scale=self.kernel_init_scale,
+                dropout_rate=self.dropout_rate,
+                name=f"{i}",
+                dtype=self.dtype,
+            )
+            input_i = layer(input_i, attention_mask, deterministic=deterministic)
         return input_i
 
 
@@ -256,21 +328,39 @@ class FlaxBertEncoder(nn.Module):
     num_heads: int
     head_size: int
     intermediate_size: int
+    dropout_rate: float = 0.0
+    kernel_init_scale: float = 0.2
+    dtype: jnp.dtype = jnp.float32  # the dtype of the computation
 
     @nn.compact
-    def __call__(self, hidden_state, attention_mask):
+    def __call__(self, hidden_state, attention_mask, deterministic: bool = True):
         layer = FlaxBertLayerCollection(
-            self.num_layers, self.num_heads, self.head_size, self.intermediate_size, name="layer"
+            self.num_layers,
-        )(hidden_state, attention_mask)
+            self.num_heads,
+            self.head_size,
+            self.intermediate_size,
+            kernel_init_scale=self.kernel_init_scale,
+            dropout_rate=self.dropout_rate,
+            name="layer",
+            dtype=self.dtype,
+        )(hidden_state, attention_mask, deterministic=deterministic)
         return layer
 
 
 class FlaxBertPooler(nn.Module):
+    kernel_init_scale: float = 0.2
+    dtype: jnp.dtype = jnp.float32  # the dtype of the computation
+
     @nn.compact
     def __call__(self, hidden_state):
         cls_token = hidden_state[:, 0]
-        out = nn.Dense(hidden_state.shape[-1], name="dense")(cls_token)
+        out = nn.Dense(
-        return jax.lax.tanh(out)
+            hidden_state.shape[-1],
+            kernel_init=jax.nn.initializers.normal(self.kernel_init_scale, self.dtype),
+            name="dense",
+            dtype=self.dtype,
+        )(cls_token)
+        return nn.tanh(out)
 
 
 class FlaxBertModule(nn.Module):
@@ -282,24 +372,104 @@ class FlaxBertModule(nn.Module):
     num_heads: int
     head_size: int
     intermediate_size: int
+    dropout_rate: float = 0.0
+    kernel_init_scale: float = 0.2
+    dtype: jnp.dtype = jnp.float32  # the dtype of the computation
 
     @nn.compact
-    def __call__(self, input_ids, attention_mask, token_type_ids, position_ids):
+    def __call__(self, input_ids, attention_mask, token_type_ids, position_ids, deterministic: bool = True):
 
         # Embedding
         embeddings = FlaxBertEmbeddings(
-            self.vocab_size, self.hidden_size, self.type_vocab_size, self.max_length, name="embeddings"
+            self.vocab_size,
-        )(input_ids, token_type_ids, position_ids, attention_mask)
+            self.hidden_size,
+            self.type_vocab_size,
+            self.max_length,
+            kernel_init_scale=self.kernel_init_scale,
+            dropout_rate=self.dropout_rate,
+            name="embeddings",
+            dtype=self.dtype,
+        )(input_ids, token_type_ids, position_ids, attention_mask, deterministic=deterministic)
 
         # N stacked encoding layers
         encoder = FlaxBertEncoder(
-            self.num_encoder_layers, self.num_heads, self.head_size, self.intermediate_size, name="encoder"
+            self.num_encoder_layers,
-        )(embeddings, attention_mask)
+            self.num_heads,
+            self.head_size,
+            self.intermediate_size,
+            kernel_init_scale=self.kernel_init_scale,
+            dropout_rate=self.dropout_rate,
+            name="encoder",
+            dtype=self.dtype,
+        )(embeddings, attention_mask, deterministic=deterministic)
 
-        pooled = FlaxBertPooler(name="pooler")(encoder)
+        pooled = FlaxBertPooler(kernel_init_scale=self.kernel_init_scale, name="pooler", dtype=self.dtype)(encoder)
         return encoder, pooled
 
 
+class FlaxBertPredictionHeadTransform(nn.Module):
+    dtype: jnp.dtype = jnp.float32
+
+    @nn.compact
+    def __call__(self, hidden_states):
+        hidden_states = nn.Dense(hidden_states.shape[-1], name="dense", dtype=self.dtype)(hidden_states)
+        hidden_states = nn.elu(hidden_states)  # TODO: ACT2FN[config.hidden_act]
+        return FlaxBertLayerNorm(name="LayerNorm", dtype=self.dtype)(hidden_states)
+
+
+class FlaxBertLMPredictionHead(nn.Module):
+    vocab_size: int
+    dtype: jnp.dtype = jnp.float32
+
+    @nn.compact
+    def __call__(self, hidden_states):
+        # TODO: The output weights are the same as the input embeddings, but there is
+        #   an output-only bias for each token.
+        #   Need a link between the two variables so that the bias is correctly
+        #   resized with `resize_token_embeddings`
+
+        hidden_states = FlaxBertPredictionHeadTransform(name="transform", dtype=self.dtype)(hidden_states)
+        hidden_states = nn.Dense(self.vocab_size, name="decoder", dtype=self.dtype)(hidden_states)
+        return hidden_states
+
+
+class FlaxBertOnlyMLMHead(nn.Module):
+    vocab_size: int
+    hidden_size: int
+    intermediate_size: int
+    head_size: int
+    num_heads: int
+    num_encoder_layers: int
+    type_vocab_size: int
+    max_length: int
+    dropout_rate: float = 0.0
+    dtype: jnp.dtype = jnp.float32
+
+    @nn.compact
+    def __call__(
+        self, input_ids, attention_mask=None, token_type_ids=None, position_ids=None, deterministic: bool = True
+    ):
+        # Model
+        encoder, pooled = FlaxBertModule(
+            vocab_size=self.vocab_size,
+            type_vocab_size=self.type_vocab_size,
+            hidden_size=self.hidden_size,
+            intermediate_size=self.intermediate_size,
+            head_size=self.hidden_size,
+            num_heads=self.num_heads,
+            num_encoder_layers=self.num_encoder_layers,
+            max_length=self.max_length,
+            dropout_rate=self.dropout_rate,
+            dtype=self.dtype,
+        )(input_ids, attention_mask, token_type_ids, position_ids, deterministic=deterministic)
+
+        # Compute the prediction scores
+        encoder = nn.Dropout(rate=self.dropout_rate)(encoder, deterministic=deterministic)
+        logits = FlaxBertLMPredictionHead(vocab_size=self.vocab_size, name="predictions", dtype=self.dtype)(encoder)
+
+        return logits, pooled
+
+
 @add_start_docstrings(
     "The bare Bert Model transformer outputting raw hidden-states without any specific head on top.",
     BERT_START_DOCSTRING,
@@ -385,8 +555,8 @@ class FlaxBertModel(FlaxPreTrainedModel):
 
         return jax_state
 
-    def __init__(self, config: BertConfig, state: dict, seed: int = 0, **kwargs):
+    def __init__(self, config: BertConfig, state: dict, seed: int = 0, dtype: jnp.dtype = jnp.float32):
-        model = FlaxBertModule(
+        module = FlaxBertModule(
             vocab_size=config.vocab_size,
             hidden_size=config.hidden_size,
             type_vocab_size=config.type_vocab_size,
@@ -395,16 +565,43 @@ class FlaxBertModel(FlaxPreTrainedModel):
             num_heads=config.num_attention_heads,
             head_size=config.hidden_size,
             intermediate_size=config.intermediate_size,
+            dropout_rate=config.hidden_dropout_prob,
+            dtype=dtype,
         )
 
-        super().__init__(config, model, state, seed)
+        super().__init__(config, module, state, seed)
-
-    @property
-    def module(self) -> nn.Module:
-        return self._module
 
     @add_start_docstrings_to_model_forward(BERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
-    def __call__(self, input_ids, attention_mask=None, token_type_ids=None, position_ids=None):
+    def __call__(
+        self,
+        input_ids,
+        attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        params: dict = None,
+        dropout_rng: PRNGKey = None,
+        train: bool = False,
+    ):
+        input_ids, attention_mask, token_type_ids, position_ids = self._check_inputs(
+            input_ids, attention_mask, token_type_ids, position_ids
+        )
+
+        # Handle any PRNG if needed
+        rngs = {}
+        if dropout_rng is not None:
+            rngs["dropout"] = dropout_rng
+
+        return self.module.apply(
+            {"params": params or self.params},
+            jnp.array(input_ids, dtype="i4"),
+            jnp.array(attention_mask, dtype="i4"),
+            jnp.array(token_type_ids, dtype="i4"),
+            jnp.array(position_ids, dtype="i4"),
+            not train,
+            rngs=rngs,
+        )
+
+    def _check_inputs(self, input_ids, attention_mask, token_type_ids, position_ids):
         if token_type_ids is None:
             token_type_ids = jnp.ones_like(input_ids)
 
@@ -414,10 +611,62 @@ class FlaxBertModel(FlaxPreTrainedModel):
         if attention_mask is None:
             attention_mask = jnp.ones_like(input_ids)
 
-        return self.model.apply(
+        return input_ids, attention_mask, token_type_ids, position_ids
-            {"params": self.params},
+
+    def init(self, rng: jax.random.PRNGKey, input_shape: Tuple):
+        input_ids, attention_mask, token_type_ids, position_ids = self._check_inputs(
+            jnp.zeros(input_shape, dtype="i4"), None, None, None
+        )
+
+        params_rng, dropout_rng = jax.random.split(rng)
+        rngs = {"params": params_rng, "dropout": dropout_rng}
+
+        self.params = self.module.init(rngs, input_ids, attention_mask, token_type_ids, position_ids)["params"]
+
+
+class FlaxBertForMaskedLM(FlaxBertModel):
+    def __init__(self, config: BertConfig, state: dict, seed: int = 0, dtype: jnp.dtype = jnp.float32, **kwargs):
+        super().__init__(config, state, seed, dtype)
+
+        self._module = FlaxBertOnlyMLMHead(
+            vocab_size=config.vocab_size,
+            type_vocab_size=config.type_vocab_size,
+            hidden_size=config.hidden_size,
+            intermediate_size=config.intermediate_size,
+            head_size=config.hidden_size,
+            num_heads=config.num_attention_heads,
+            num_encoder_layers=config.num_hidden_layers,
+            max_length=config.max_length,
+            **kwargs,
+        )
+
+    def __call__(
+        self,
+        input_ids,
+        attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        params: dict = None,
+        dropout_rng: PRNGKey = None,
+        train: bool = False,
+    ):
+        input_ids, attention_mask, token_type_ids, position_ids = self._check_inputs(
+            input_ids, attention_mask, token_type_ids, position_ids
+        )
+
+        # Handle any PRNG if needed
+        rngs = {}
+        if dropout_rng is not None:
+            rngs["dropout"] = dropout_rng
+
+        pooled, logits = self.module.apply(
+            {"params": params or self.params},
             jnp.array(input_ids, dtype="i4"),
             jnp.array(attention_mask, dtype="i4"),
             jnp.array(token_type_ids, dtype="i4"),
             jnp.array(position_ids, dtype="i4"),
+            not train,
+            rngs=rngs,
         )
+
+        return logits, pooled

src/transformers/models/roberta/modeling_flax_roberta.py

@@ -12,13 +12,14 @@
 # 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 typing import Callable, Dict
+from typing import Callable, Dict, Tuple
 
 import numpy as np
 
 import flax.linen as nn
 import jax
 import jax.numpy as jnp
+from jax.random import PRNGKey
 
 from ...file_utils import add_start_docstrings, add_start_docstrings_to_model_forward
 from ...modeling_flax_utils import FlaxPreTrainedModel, gelu
@@ -101,8 +102,8 @@ class FlaxRobertaLayerNorm(nn.Module):
     scale: bool = True  # If True, multiply by scale (gamma). When the next layer is linear
     # (also e.g. nn.relu), this can be disabled since the scaling will be
     # done by the next layer.
-    bias_init: jnp.ndarray = nn.initializers.zeros
+    scale_init: Callable[..., np.ndarray] = jax.nn.initializers.ones
-    scale_init: jnp.ndarray = nn.initializers.ones
+    bias_init: Callable[..., np.ndarray] = jax.nn.initializers.zeros
 
     @nn.compact
     def __call__(self, x):
@@ -122,11 +123,13 @@ class FlaxRobertaLayerNorm(nn.Module):
         mean2 = jnp.mean(jax.lax.square(x), axis=-1, keepdims=True)
         var = mean2 - jax.lax.square(mean)
         mul = jax.lax.rsqrt(var + self.epsilon)
+
         if self.scale:
-            mul = mul * jnp.asarray(self.param("gamma", self.scale_init, (features,)), self.dtype)
+            mul = mul * jnp.asarray(self.param("gamma", self.scale_init, (features,)))
         y = (x - mean) * mul
+
         if self.bias:
-            y = y + jnp.asarray(self.param("beta", self.bias_init, (features,)), self.dtype)
+            y = y + jnp.asarray(self.param("beta", self.bias_init, (features,)))
         return y
 
 
@@ -139,7 +142,9 @@ class FlaxRobertaEmbedding(nn.Module):
 
     vocab_size: int
     hidden_size: int
-    emb_init: Callable[..., np.ndarray] = nn.initializers.normal(stddev=0.1)
+    kernel_init_scale: float = 0.2
+    emb_init: Callable[..., np.ndarray] = jax.nn.initializers.normal(stddev=kernel_init_scale)
+    dtype: jnp.dtype = jnp.float32  # the dtype of the computation
 
     @nn.compact
     def __call__(self, inputs):
@@ -155,66 +160,108 @@ class FlaxRobertaEmbeddings(nn.Module):
     hidden_size: int
     type_vocab_size: int
     max_length: int
+    kernel_init_scale: float = 0.2
+    dropout_rate: float = 0.0
+    dtype: jnp.dtype = jnp.float32  # the dtype of the computation
 
     @nn.compact
-    def __call__(self, input_ids, token_type_ids, position_ids, attention_mask):
+    def __call__(self, input_ids, token_type_ids, position_ids, attention_mask, deterministic: bool = True):
 
         # Embed
-        w_emb = FlaxRobertaEmbedding(self.vocab_size, self.hidden_size, name="word_embeddings")(
+        w_emb = FlaxRobertaEmbedding(
-            jnp.atleast_2d(input_ids.astype("i4"))
+            self.vocab_size,
-        )
+            self.hidden_size,
-        p_emb = FlaxRobertaEmbedding(self.max_length, self.hidden_size, name="position_embeddings")(
+            kernel_init_scale=self.kernel_init_scale,
-            jnp.atleast_2d(position_ids.astype("i4"))
+            name="word_embeddings",
-        )
+            dtype=self.dtype,
-        t_emb = FlaxRobertaEmbedding(self.type_vocab_size, self.hidden_size, name="token_type_embeddings")(
+        )(jnp.atleast_2d(input_ids.astype("i4")))
-            jnp.atleast_2d(token_type_ids.astype("i4"))
+        p_emb = FlaxRobertaEmbedding(
-        )
+            self.max_length,
+            self.hidden_size,
+            kernel_init_scale=self.kernel_init_scale,
+            name="position_embeddings",
+            dtype=self.dtype,
+        )(jnp.atleast_2d(position_ids.astype("i4")))
+        t_emb = FlaxRobertaEmbedding(
+            self.type_vocab_size,
+            self.hidden_size,
+            kernel_init_scale=self.kernel_init_scale,
+            name="token_type_embeddings",
+            dtype=self.dtype,
+        )(jnp.atleast_2d(token_type_ids.astype("i4")))
 
         # Sum all embeddings
         summed_emb = w_emb + jnp.broadcast_to(p_emb, w_emb.shape) + t_emb
 
         # Layer Norm
-        layer_norm = FlaxRobertaLayerNorm(name="layer_norm")(summed_emb)
+        layer_norm = FlaxRobertaLayerNorm(name="layer_norm", dtype=self.dtype)(summed_emb)
-
+        embeddings = nn.Dropout(rate=self.dropout_rate)(layer_norm, deterministic=deterministic)
-        return layer_norm
+        return embeddings
 
 
 # Copied from transformers.models.bert.modeling_flax_bert.FlaxBertAttention with Bert->Roberta
 class FlaxRobertaAttention(nn.Module):
     num_heads: int
     head_size: int
+    dropout_rate: float = 0.0
+    kernel_init_scale: float = 0.2
+    dtype: jnp.dtype = jnp.float32  # the dtype of the computation
 
     @nn.compact
-    def __call__(self, hidden_state, attention_mask):
+    def __call__(self, hidden_state, attention_mask, deterministic: bool = True):
         # Attention mask comes in as attention_mask.shape == (*batch_sizes, kv_length)
         # FLAX expects: attention_mask.shape == (*batch_sizes, 1, 1, kv_length) such that it is broadcastable
         # with attn_weights.shape == (*batch_sizes, num_heads, q_length, kv_length)
         attention_mask = jnp.expand_dims(attention_mask, axis=(-3, -2))
-        self_att = nn.attention.SelfAttention(num_heads=self.num_heads, qkv_features=self.head_size, name="self")(
+        self_att = nn.attention.SelfAttention(
-            hidden_state, attention_mask
+            num_heads=self.num_heads,
-        )
+            qkv_features=self.head_size,
+            dropout_rate=self.dropout_rate,
+            deterministic=deterministic,
+            kernel_init=jax.nn.initializers.normal(self.kernel_init_scale, self.dtype),
+            bias_init=jax.nn.initializers.zeros,
+            name="self",
+            dtype=self.dtype,
+        )(hidden_state, attention_mask)
 
-        layer_norm = FlaxRobertaLayerNorm(name="layer_norm")(self_att + hidden_state)
+        layer_norm = FlaxRobertaLayerNorm(name="layer_norm", dtype=self.dtype)(self_att + hidden_state)
         return layer_norm
 
 
 # Copied from transformers.models.bert.modeling_flax_bert.FlaxBertIntermediate with Bert->Roberta
 class FlaxRobertaIntermediate(nn.Module):
     output_size: int
+    kernel_init_scale: float = 0.2
+    dtype: jnp.dtype = jnp.float32  # the dtype of the computation
 
     @nn.compact
     def __call__(self, hidden_state):
         # TODO: Add ACT2FN reference to change activation function
-        dense = nn.Dense(features=self.output_size, name="dense")(hidden_state)
+        dense = nn.Dense(
+            features=self.output_size,
+            kernel_init=jax.nn.initializers.normal(self.kernel_init_scale, self.dtype),
+            name="dense",
+            dtype=self.dtype,
+        )(hidden_state)
         return gelu(dense)
 
 
 # Copied from transformers.models.bert.modeling_flax_bert.FlaxBertOutput with Bert->Roberta
 class FlaxRobertaOutput(nn.Module):
+    dropout_rate: float = 0.0
+    kernel_init_scale: float = 0.2
+    dtype: jnp.dtype = jnp.float32  # the dtype of the computation
+
     @nn.compact
-    def __call__(self, intermediate_output, attention_output):
+    def __call__(self, intermediate_output, attention_output, deterministic: bool = True):
-        hidden_state = nn.Dense(attention_output.shape[-1], name="dense")(intermediate_output)
+        hidden_state = nn.Dense(
-        hidden_state = FlaxRobertaLayerNorm(name="layer_norm")(hidden_state + attention_output)
+            attention_output.shape[-1],
+            kernel_init=jax.nn.initializers.normal(self.kernel_init_scale, self.dtype),
+            name="dense",
+            dtype=self.dtype,
+        )(intermediate_output)
+        hidden_state = nn.Dropout(rate=self.dropout_rate)(hidden_state, deterministic=deterministic)
+        hidden_state = FlaxRobertaLayerNorm(name="layer_norm", dtype=self.dtype)(hidden_state + attention_output)
         return hidden_state
 
 
@@ -222,14 +269,29 @@ class FlaxRobertaLayer(nn.Module):
     num_heads: int
     head_size: int
     intermediate_size: int
+    dropout_rate: float = 0.0
+    kernel_init_scale: float = 0.2
+    dtype: jnp.dtype = jnp.float32  # the dtype of the computation
 
     @nn.compact
-    def __call__(self, hidden_state, attention_mask):
+    def __call__(self, hidden_state, attention_mask, deterministic: bool = True):
-        attention = FlaxRobertaAttention(self.num_heads, self.head_size, name="attention")(
+        attention = FlaxRobertaAttention(
-            hidden_state, attention_mask
+            self.num_heads,
-        )
+            self.head_size,
-        intermediate = FlaxRobertaIntermediate(self.intermediate_size, name="intermediate")(attention)
+            kernel_init_scale=self.kernel_init_scale,
-        output = FlaxRobertaOutput(name="output")(intermediate, attention)
+            dropout_rate=self.dropout_rate,
+            name="attention",
+            dtype=self.dtype,
+        )(hidden_state, attention_mask, deterministic=deterministic)
+        intermediate = FlaxRobertaIntermediate(
+            self.intermediate_size,
+            kernel_init_scale=self.kernel_init_scale,
+            name="intermediate",
+            dtype=self.dtype,
+        )(attention)
+        output = FlaxRobertaOutput(
+            kernel_init_scale=self.kernel_init_scale, dropout_rate=self.dropout_rate, name="output", dtype=self.dtype
+        )(intermediate, attention, deterministic=deterministic)
 
         return output
 
@@ -244,9 +306,12 @@ class FlaxRobertaLayerCollection(nn.Module):
     num_heads: int
     head_size: int
     intermediate_size: int
+    dropout_rate: float = 0.0
+    kernel_init_scale: float = 0.2
+    dtype: jnp.dtype = jnp.float32  # the dtype of the computation
 
     @nn.compact
-    def __call__(self, inputs, attention_mask):
+    def __call__(self, inputs, attention_mask, deterministic: bool = True):
         assert self.num_layers > 0, f"num_layers should be >= 1, got ({self.num_layers})"
 
         # Initialize input / output
@@ -254,8 +319,16 @@ class FlaxRobertaLayerCollection(nn.Module):
 
         # Forward over all encoders
         for i in range(self.num_layers):
-            layer = FlaxRobertaLayer(self.num_heads, self.head_size, self.intermediate_size, name=f"{i}")
+            layer = FlaxRobertaLayer(
-            input_i = layer(input_i, attention_mask)
+                self.num_heads,
+                self.head_size,
+                self.intermediate_size,
+                kernel_init_scale=self.kernel_init_scale,
+                dropout_rate=self.dropout_rate,
+                name=f"{i}",
+                dtype=self.dtype,
+            )
+            input_i = layer(input_i, attention_mask, deterministic=deterministic)
         return input_i
 
 
@@ -265,22 +338,40 @@ class FlaxRobertaEncoder(nn.Module):
     num_heads: int
     head_size: int
     intermediate_size: int
+    dropout_rate: float = 0.0
+    kernel_init_scale: float = 0.2
+    dtype: jnp.dtype = jnp.float32  # the dtype of the computation
 
     @nn.compact
-    def __call__(self, hidden_state, attention_mask):
+    def __call__(self, hidden_state, attention_mask, deterministic: bool = True):
         layer = FlaxRobertaLayerCollection(
-            self.num_layers, self.num_heads, self.head_size, self.intermediate_size, name="layer"
+            self.num_layers,
-        )(hidden_state, attention_mask)
+            self.num_heads,
+            self.head_size,
+            self.intermediate_size,
+            kernel_init_scale=self.kernel_init_scale,
+            dropout_rate=self.dropout_rate,
+            name="layer",
+            dtype=self.dtype,
+        )(hidden_state, attention_mask, deterministic=deterministic)
         return layer
 
 
 # Copied from transformers.models.bert.modeling_flax_bert.FlaxBertPooler with Bert->Roberta
 class FlaxRobertaPooler(nn.Module):
+    kernel_init_scale: float = 0.2
+    dtype: jnp.dtype = jnp.float32  # the dtype of the computation
+
     @nn.compact
     def __call__(self, hidden_state):
         cls_token = hidden_state[:, 0]
-        out = nn.Dense(hidden_state.shape[-1], name="dense")(cls_token)
+        out = nn.Dense(
-        return jax.lax.tanh(out)
+            hidden_state.shape[-1],
+            kernel_init=jax.nn.initializers.normal(self.kernel_init_scale, self.dtype),
+            name="dense",
+            dtype=self.dtype,
+        )(cls_token)
+        return nn.tanh(out)
 
 
 # Copied from transformers.models.bert.modeling_flax_bert.FlaxBertModule with Bert->Roberta
@@ -293,21 +384,38 @@ class FlaxRobertaModule(nn.Module):
     num_heads: int
     head_size: int
     intermediate_size: int
+    dropout_rate: float = 0.0
+    kernel_init_scale: float = 0.2
+    dtype: jnp.dtype = jnp.float32  # the dtype of the computation
 
     @nn.compact
-    def __call__(self, input_ids, attention_mask, token_type_ids, position_ids):
+    def __call__(self, input_ids, attention_mask, token_type_ids, position_ids, deterministic: bool = True):
 
         # Embedding
         embeddings = FlaxRobertaEmbeddings(
-            self.vocab_size, self.hidden_size, self.type_vocab_size, self.max_length, name="embeddings"
+            self.vocab_size,
-        )(input_ids, token_type_ids, position_ids, attention_mask)
+            self.hidden_size,
+            self.type_vocab_size,
+            self.max_length,
+            kernel_init_scale=self.kernel_init_scale,
+            dropout_rate=self.dropout_rate,
+            name="embeddings",
+            dtype=self.dtype,
+        )(input_ids, token_type_ids, position_ids, attention_mask, deterministic=deterministic)
 
         # N stacked encoding layers
         encoder = FlaxRobertaEncoder(
-            self.num_encoder_layers, self.num_heads, self.head_size, self.intermediate_size, name="encoder"
+            self.num_encoder_layers,
-        )(embeddings, attention_mask)
+            self.num_heads,
+            self.head_size,
+            self.intermediate_size,
+            kernel_init_scale=self.kernel_init_scale,
+            dropout_rate=self.dropout_rate,
+            name="encoder",
+            dtype=self.dtype,
+        )(embeddings, attention_mask, deterministic=deterministic)
 
-        pooled = FlaxRobertaPooler(name="pooler")(encoder)
+        pooled = FlaxRobertaPooler(kernel_init_scale=self.kernel_init_scale, name="pooler", dtype=self.dtype)(encoder)
         return encoder, pooled
 
 
@@ -396,8 +504,8 @@ class FlaxRobertaModel(FlaxPreTrainedModel):
 
         return jax_state
 
-    def __init__(self, config: RobertaConfig, state: dict, seed: int = 0, **kwargs):
+    def __init__(self, config: RobertaConfig, state: dict, seed: int = 0, dtype: jnp.dtype = jnp.float32):
-        model = FlaxRobertaModule(
+        module = FlaxRobertaModule(
             vocab_size=config.vocab_size,
             hidden_size=config.hidden_size,
             type_vocab_size=config.type_vocab_size,
@@ -406,31 +514,78 @@ class FlaxRobertaModel(FlaxPreTrainedModel):
             num_heads=config.num_attention_heads,
             head_size=config.hidden_size,
             intermediate_size=config.intermediate_size,
+            dropout_rate=config.hidden_dropout_prob,
+            dtype=dtype,
         )
 
-        super().__init__(config, model, state, seed)
+        super().__init__(config, module, state, seed)
-
-    @property
-    def module(self) -> nn.Module:
-        return self._module
 
     @add_start_docstrings_to_model_forward(ROBERTA_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
-    def __call__(self, input_ids, token_type_ids=None, attention_mask=None, position_ids=None):
+    def __call__(
-        if token_type_ids is None:
+        self,
-            token_type_ids = jnp.ones_like(input_ids)
+        input_ids,
-
+        token_type_ids=None,
-        if position_ids is None:
+        attention_mask=None,
-            position_ids = jnp.arange(
+        position_ids=None,
-                self.config.pad_token_id + 1, jnp.atleast_2d(input_ids).shape[-1] + self.config.pad_token_id + 1
+        params: dict = None,
+        dropout_rng: PRNGKey = None,
+        train: bool = False,
+    ):
+        input_ids, attention_mask, token_type_ids, position_ids = self._check_inputs(
+            input_ids, attention_mask, token_type_ids, position_ids
         )
 
-        if attention_mask is None:
+        # Handle any PRNG if needed
-            attention_mask = jnp.ones_like(input_ids)
+        rngs = {}
+        if dropout_rng is not None:
+            rngs["dropout"] = dropout_rng
 
-        return self.model.apply(
+        return self.module.apply(
-            {"params": self.params},
+            {"params": params or self.params},
             jnp.array(input_ids, dtype="i4"),
             jnp.array(attention_mask, dtype="i4"),
             jnp.array(token_type_ids, dtype="i4"),
             jnp.array(position_ids, dtype="i4"),
+            not train,
+            rngs=rngs,
         )
+
+    def init(self, rng: jax.random.PRNGKey, input_shape: Tuple):
+        input_ids, attention_mask, token_type_ids, position_ids = self._check_inputs(
+            jnp.zeros(input_shape, dtype="i4"), None, None, None
+        )
+
+        params_rng, dropout_rng = jax.random.split(rng)
+        rngs = {"params": params_rng, "dropout": dropout_rng}
+
+        self.params = self.module.init(rngs, input_ids, attention_mask, token_type_ids, position_ids)["params"]
+
+    def _check_inputs(self, input_ids, attention_mask, token_type_ids, position_ids):
+
+        if token_type_ids is None:
+            token_type_ids = jnp.ones_like(input_ids)
+
+        if position_ids is None:
+            position_ids = create_position_ids_from_input_ids(input_ids, self.config.pad_token_id)
+
+        if attention_mask is None:
+            attention_mask = jnp.ones_like(input_ids)
+
+        return input_ids, attention_mask, token_type_ids, position_ids
+
+
+def create_position_ids_from_input_ids(input_ids, padding_idx):
+    """
+    Replace non-padding symbols with their position numbers. Position numbers begin at padding_idx+1. Padding symbols
+    are ignored. This is modified from fairseq's `utils.make_positions`.
+
+    Args:
+        input_ids: jnp.ndarray
+        padding_idx: int
+
+    Returns: jnp.ndarray
+    """
+    # The series of casts and type-conversions here are carefully balanced to both work with ONNX export and XLA.
+    mask = (input_ids != padding_idx).astype("i4")
+    incremental_indices = jnp.cumsum(mask, axis=1).astype("i4") * mask
+    return incremental_indices.astype("i4") + padding_idx

src/transformers/utils/dummy_flax_objects.py

@@ -14,6 +14,15 @@ class FlaxAutoModel:
         requires_flax(self)
 
 
+class FlaxBertForMaskedLM:
+    def __init__(self, *args, **kwargs):
+        requires_flax(self)
+
+    @classmethod
+    def from_pretrained(self, *args, **kwargs):
+        requires_flax(self)
+
+
 class FlaxBertModel:
     def __init__(self, *args, **kwargs):
         requires_flax(self)

tests/test_modeling_flax_roberta.py

@@ -57,7 +57,7 @@ class FlaxRobertaModelTest(unittest.TestCase):
                 self.assertEqual(len(fx_outputs), len(pt_outputs), "Output lengths differ between Flax and PyTorch")
 
                 for fx_output, pt_output in zip(fx_outputs, pt_outputs.to_tuple()):
-                    self.assert_almost_equals(fx_output, pt_output.numpy(), 6e-4)
+                    self.assert_almost_equals(fx_output, pt_output.numpy(), 5e-3)
 
     def test_multiple_sequences(self):
         tokenizer = RobertaTokenizerFast.from_pretrained("roberta-base")