docs/transformers/examples/flax/language-modeling/README.md

<!---
Copyright 2021 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.
-->

# Language model training and inference examples

The following example showcases how to train a language model from scratch
using the JAX/Flax backend.

JAX/Flax allows you to trace pure functions and compile them into efficient, fused accelerator code on both GPU and TPU.
Models written in JAX/Flax are **immutable** and updated in a purely functional
way which enables simple and efficient model parallelism.

## Masked language modeling

In the following, we demonstrate how to train a bi-directional transformer model
using masked language modeling objective as introduced in [BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding](https://arxiv.org/abs/1810.04805).
More specifically, we demonstrate how JAX/Flax can be leveraged
to pre-train [**`FacebookAI/roberta-base`**](https://huggingface.co/FacebookAI/roberta-base)
in Norwegian on a single TPUv3-8 pod.

The example script uses the 🤗 Datasets library. You can easily customize them to your needs if you need extra processing on your datasets.

To setup all relevant files for training, let's create a directory.

```bash
mkdir ./norwegian-roberta-base
```

### Train tokenizer

In the first step, we train a tokenizer to efficiently process the text input for the model. Similar to how it is shown in [How to train a new language model from scratch using Transformers and Tokenizers](https://huggingface.co/blog/how-to-train), we use a **`ByteLevelBPETokenizer`**.
The tokenizer is trained on the complete Norwegian dataset of OSCAR
and consequently saved in the cloned model directory.
This can take up to 10 minutes depending on your hardware ☕.

```python
from datasets import load_dataset
from tokenizers import trainers, Tokenizer, normalizers, ByteLevelBPETokenizer

# load dataset
dataset = load_dataset("oscar", "unshuffled_deduplicated_no", split="train")

# Instantiate tokenizer
tokenizer = ByteLevelBPETokenizer()

def batch_iterator(batch_size=1000):
    for i in range(0, len(dataset), batch_size):
        yield dataset[i: i + batch_size]["text"]

# Customized training
tokenizer.train_from_iterator(batch_iterator(), vocab_size=50265, min_frequency=2, special_tokens=[
    "<s>",
    "<pad>",
    "</s>",
    "<unk>",
    "<mask>",
])

# Save files to disk
tokenizer.save("./norwegian-roberta-base/tokenizer.json")
```

### Create configuration

Next, we create the model's configuration file. This is as simple
as loading and storing [`**FacebookAI/roberta-base**`](https://huggingface.co/FacebookAI/roberta-base)
in the local model folder:

```python
from transformers import RobertaConfig

config = RobertaConfig.from_pretrained("FacebookAI/roberta-base", vocab_size=50265)
config.save_pretrained("./norwegian-roberta-base")
```

Great, we have set up our model repository. During training, we will automatically
push the training logs and model weights to the repo.

### Train model

Next we can run the example script to pretrain the model:

```bash
python run_mlm_flax.py \
    --output_dir="./norwegian-roberta-base" \
    --model_type="roberta" \
    --config_name="./norwegian-roberta-base" \
    --tokenizer_name="./norwegian-roberta-base" \
    --dataset_name="oscar" \
    --dataset_config_name="unshuffled_deduplicated_no" \
    --max_seq_length="128" \
    --weight_decay="0.01" \
    --per_device_train_batch_size="128" \
    --per_device_eval_batch_size="128" \
    --learning_rate="3e-4" \
    --warmup_steps="1000" \
    --overwrite_output_dir \
    --num_train_epochs="18" \
    --adam_beta1="0.9" \
    --adam_beta2="0.98" \
    --logging_steps="500" \
    --save_steps="2500" \
    --eval_steps="2500" \
    --push_to_hub
```

Training should converge at a loss and accuracy
of 1.78 and 0.64 respectively after 18 epochs on a single TPUv3-8.
This should take less than 18 hours.
Training statistics can be accessed on [tfhub.dev](https://tensorboard.dev/experiment/GdYmdak2TWeVz0DDRYOrrg).

For a step-by-step walkthrough of how to do masked language modeling in Flax, please have a
look at [this](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/masked_language_modeling_flax.ipynb) google colab.

## Causal language modeling

In the following, we demonstrate how to train an auto-regressive causal transformer model
in JAX/Flax.
More specifically, we pretrain a randomly initialized [**`openai-community/gpt2`**](https://huggingface.co/openai-community/gpt2) model in Norwegian on a single TPUv3-8.
to pre-train 124M [**`openai-community/gpt2`**](https://huggingface.co/openai-community/gpt2)
in Norwegian on a single TPUv3-8 pod.

The example script uses the 🤗 Datasets library. You can easily customize them to your needs if you need extra processing on your datasets.


To setup all relevant files for training, let's create a directory.

```bash
mkdir ./norwegian-gpt2
```

### Train tokenizer

In the first step, we train a tokenizer to efficiently process the text input for the model. Similar to how it is shown in [How to train a new language model from scratch using Transformers and Tokenizers](https://huggingface.co/blog/how-to-train), we use a **`ByteLevelBPETokenizer`**.
The tokenizer is trained on the complete Norwegian dataset of OSCAR
and consequently saved in the cloned model directory.
This can take up to 10 minutes depending on your hardware ☕.

```python
from datasets import load_dataset
from tokenizers import trainers, Tokenizer, normalizers, ByteLevelBPETokenizer

# load dataset
dataset = load_dataset("oscar", "unshuffled_deduplicated_no", split="train")

# Instantiate tokenizer
tokenizer = ByteLevelBPETokenizer()

def batch_iterator(batch_size=1000):
    for i in range(0, len(dataset), batch_size):
        yield dataset[i: i + batch_size]["text"]

# Customized training
tokenizer.train_from_iterator(batch_iterator(), vocab_size=50257, min_frequency=2, special_tokens=[
    "<s>",
    "<pad>",
    "</s>",
    "<unk>",
    "<mask>",
])

# Save files to disk
tokenizer.save("./norwegian-gpt2/tokenizer.json")
```

### Create configuration

Next, we create the model's configuration file. This is as simple
as loading and storing [`**openai-community/gpt2**`](https://huggingface.co/openai-community/gpt2)
in the local model folder:

```python
from transformers import GPT2Config

config = GPT2Config.from_pretrained("openai-community/gpt2", resid_pdrop=0.0, embd_pdrop=0.0, attn_pdrop=0.0, vocab_size=50257)
config.save_pretrained("./norwegian-gpt2")
```

Great, we have set up our model repository. During training, we will now automatically
push the training logs and model weights to the repo.

### Train model

Finally, we can run the example script to pretrain the model:

```bash
python run_clm_flax.py \
    --output_dir="./norwegian-gpt2" \
    --model_type="gpt2" \
    --config_name="./norwegian-gpt2" \
    --tokenizer_name="./norwegian-gpt2" \
    --dataset_name="oscar" \
    --dataset_config_name="unshuffled_deduplicated_no" \
    --do_train --do_eval \
    --block_size="512" \
    --per_device_train_batch_size="64" \
    --per_device_eval_batch_size="64" \
    --learning_rate="5e-3" --warmup_steps="1000" \
    --adam_beta1="0.9" --adam_beta2="0.98" --weight_decay="0.01" \
    --overwrite_output_dir \
    --num_train_epochs="20" \
    --logging_steps="500" \
    --save_steps="2500" \
    --eval_steps="2500" \
    --push_to_hub
```

Training should converge at a loss and perplexity
of 3.24 and 25.72 respectively after 20 epochs on a single TPUv3-8.
This should take less than ~21 hours.
Training statistics can be accessed on [tfhub.dev](https://tensorboard.dev/experiment/2zEhLwJ0Qp2FAkI3WVH9qA).

For a step-by-step walkthrough of how to do causal language modeling in Flax, please have a
look at [this](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/causal_language_modeling_flax.ipynb) google colab.

## T5-like span-masked language modeling

In the following, we demonstrate how to train a T5 model using the span-masked language model
objective as proposed in the [Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer](https://arxiv.org/abs/1910.10683).
More specifically, we demonstrate how JAX/Flax can be leveraged
to pre-train [**`google/t5-v1_1-base`**](https://huggingface.co/google/t5-v1_1-base)
in Norwegian on a single TPUv3-8 pod.

The example script uses the 🤗 Datasets library. You can easily customize them to your needs if you need extra processing on your datasets.

Let's start by creating a model repository to save the trained model and logs.
Here we call the model `"norwegian-t5-base"`, but you can change the model name as you like.

To setup all relevant files for training, let's create a directory.

```bash
cd ./norwegian-t5-base
```

### Train tokenizer

In the first step, we train a tokenizer to efficiently process the text input for the model.
We make use of the [tokenizers](https://github.com/huggingface/tokenizers) library to train
a sentencepiece unigram tokenizer as shown in [t5_tokenizer_model.py](https://github.com/huggingface/transformers/tree/main/examples/flax/language-modeling/t5_tokenizer_model.py)
which is heavily inspired from [yandex-research/DeDLOC's tokenizer model](https://github.com/yandex-research/DeDLOC/blob/5c994bc64e573702a9a79add3ecd68b38f14b548/sahajbert/tokenizer/tokenizer_model.py) .

The tokenizer is trained on the complete Norwegian dataset of OSCAR
and consequently saved in the cloned model directory.
This can take up to 120 minutes depending on your hardware ☕☕☕ .

```python
import datasets

from t5_tokenizer_model import SentencePieceUnigramTokenizer


vocab_size = 32_000
input_sentence_size = None

# Initialize a dataset
dataset = datasets.load_dataset("oscar", name="unshuffled_deduplicated_no", split="train")

tokenizer = SentencePieceUnigramTokenizer(unk_token="<unk>", eos_token="</s>", pad_token="<pad>")


# Build an iterator over this dataset
def batch_iterator(input_sentence_size=None):
    if input_sentence_size is None:
        input_sentence_size = len(dataset)
    batch_length = 100
    for i in range(0, input_sentence_size, batch_length):
        yield dataset[i: i + batch_length]["text"]


# Train tokenizer
tokenizer.train_from_iterator(
    iterator=batch_iterator(input_sentence_size=input_sentence_size),
    vocab_size=vocab_size,
    show_progress=True,
)

# Save files to disk
tokenizer.save("./norwegian-t5-base/tokenizer.json")
```

### Create configuration

Next, we create the model's configuration file. This is as simple
as loading and storing [`**google/t5-v1_1-base**`](https://huggingface.co/google/t5-v1_1-base)
in the local model folder:

```python
from transformers import T5Config

config = T5Config.from_pretrained("google/t5-v1_1-base", vocab_size=tokenizer.get_vocab_size())
config.save_pretrained("./norwegian-t5-base")
```

Great, we have set up our model repository. During training, we will automatically
push the training logs and model weights to the repo.

### Train model

Next we can run the example script to pretrain the model:

```bash
python run_t5_mlm_flax.py \
	--output_dir="./norwegian-t5-base" \
	--model_type="t5" \
	--config_name="./norwegian-t5-base" \
	--tokenizer_name="./norwegian-t5-base" \
	--dataset_name="oscar" \
	--dataset_config_name="unshuffled_deduplicated_no" \
	--max_seq_length="512" \
	--per_device_train_batch_size="32" \
	--per_device_eval_batch_size="32" \
	--adafactor \
	--learning_rate="0.005" \
	--weight_decay="0.001" \
	--warmup_steps="2000" \
	--overwrite_output_dir \
	--logging_steps="500" \
	--save_steps="10000" \
	--eval_steps="2500" \
	--push_to_hub
```

Training should converge at a loss and accuracy
of 2.36 and 57.0 respectively after 3 epochs on a single TPUv3-8.
This should take around 4.5 hours.
Training statistics can be accessed on directly on the 🤗 [hub](https://huggingface.co/patrickvonplaten/t5-base-norwegian/tensorboard)

## BART: Denoising language modeling

In the following, we demonstrate how to train a BART model
using denoising language modeling objective as introduced in [BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension](https://arxiv.org/abs/1910.13461).
More specifically, we demonstrate how JAX/Flax can be leveraged
to pre-train [**`bart-base`**](https://huggingface.co/facebook/bart-base)
in Norwegian on a single TPUv3-8 pod.

The example script uses the 🤗 Datasets library. You can easily customize them to your needs if you need extra processing on your datasets.

To setup all relevant files for training, let's create a directory.

```bash
mkdir ./norwegian-bart-base
```

### Train tokenizer
In the first step, we train a tokenizer to efficiently process the text input for the model. Similar to how it is shown in [How to train a new language model from scratch using Transformers and Tokenizers](https://huggingface.co/blog/how-to-train), we use a **`ByteLevelBPETokenizer`**.
The tokenizer is trained on the complete Norwegian dataset of OSCAR
and consequently saved in the cloned model directory.
This can take up to 10 minutes depending on your hardware ☕.

```python
from datasets import load_dataset
from tokenizers import trainers, Tokenizer, normalizers, ByteLevelBPETokenizer

# load dataset
dataset = load_dataset("oscar", "unshuffled_deduplicated_no", split="train")

# Instantiate tokenizer
tokenizer = ByteLevelBPETokenizer()

def batch_iterator(batch_size=1000):
    for i in range(0, len(dataset), batch_size):
        yield dataset[i: i + batch_size]["text"]

# Customized training
tokenizer.train_from_iterator(batch_iterator(), vocab_size=50265, min_frequency=2, special_tokens=[
    "<s>",
    "<pad>",
    "</s>",
    "<unk>",
    "<mask>",
])

# Save files to disk
tokenizer.save("./norwegian-bart-base/tokenizer.json")
```

### Create configuration

Next, we create the model's configuration file. This is as simple
as loading and storing [`**facebook/bart-base**`](https://huggingface.co/facebook/bart-base)
in the local model folder:

```python
from transformers import BartConfig
config = BartConfig.from_pretrained("facebook/bart-base", vocab_size=50265)
config.save_pretrained("./norwegian-bart-base")
```

Great, we have set up our model repository. During training, we will automatically
push the training logs and model weights to the repo.

### Train model

Next we can run the example script to pretrain the model:

```bash
python run_bart_dlm_flax.py \
    --output_dir="./norwegian-bart-base" \
    --config_name="./norwegian-bart-base" \
    --tokenizer_name="./norwegian-bart-base" \
    --dataset_name="oscar" \
    --dataset_config_name="unshuffled_deduplicated_no" \
    --max_seq_length="1024" \
    --per_device_train_batch_size="32" \
    --per_device_eval_batch_size="32" \
    --learning_rate="1e-4" \
    --warmup_steps="2000" \
    --overwrite_output_dir \
    --logging_steps="500" \
    --save_steps="2000" \
    --eval_steps="2000" \
    --push_to_hub
```

Training should converge at a loss and accuracy
of 1.36 and 0.77 respectively after 3 epochs on a single TPUv3-8.
This should take less than 6 hours.
Training statistics can be accessed on [tfhub.dev](https://tensorboard.dev/experiment/Maw62QlaSXWS0MOf2V2lbg/).

## Runtime evaluation

We also ran masked language modeling using PyTorch/XLA on a TPUv3-8, and PyTorch on 8 V100 GPUs. We report the
overall training time below.
For reproducibility, we state the training commands used for PyTorch/XLA and PyTorch further below.

| Task  | [TPU v3-8 (Flax)](https://tensorboard.dev/experiment/GdYmdak2TWeVz0DDRYOrrg/)  | [TPU v3-8 (Pytorch/XLA)](https://tensorboard.dev/experiment/7Jq1kcQQRAmy12KOdXek7A/)| [8 GPU (PyTorch)](https://tensorboard.dev/experiment/PJneV8FQRxa2unPw1QnVHA)  |
|-------|-----------|------------|------------|
| MLM   |  15h32m   |  23h46m    | 44h14m     |

*All experiments are ran on Google Cloud Platform.
GPU experiments are ran without further optimizations besides JAX
transformations. GPU experiments are ran with full precision (fp32). "TPU v3-8"
are 8 TPU cores on 4 chips (each chips has 2 cores), while "8 GPU" are 8 GPU chips.

### Script to run MLM with PyTorch/XLA on TPUv3-8

For comparison one can run the same pre-training with PyTorch/XLA on TPU. To set up PyTorch/XLA on Cloud TPU VMs, please
refer to [this](https://cloud.google.com/tpu/docs/pytorch-xla-ug-tpu-vm) guide.
Having created the tokenizer and configuration in `norwegian-roberta-base`, we create the following symbolic links:

```bash
ln -s ~/transformers/examples/pytorch/language-modeling/run_mlm.py ./
ln -s ~/transformers/examples/pytorch/xla_spawn.py ./
```

, set the following environment variables:

```bash
export XRT_TPU_CONFIG="localservice;0;localhost:51011"
unset LD_PRELOAD

export NUM_TPUS=8
export TOKENIZERS_PARALLELISM=0
export MODEL_DIR="./norwegian-roberta-base"
mkdir -p ${MODEL_DIR}
```

, and start training as follows:

```bash
python3 xla_spawn.py --num_cores ${NUM_TPUS} run_mlm.py --output_dir="./runs" \
    --model_type="roberta" \
    --config_name="${MODEL_DIR}" \
    --tokenizer_name="${MODEL_DIR}" \
    --dataset_name="oscar" \
    --dataset_config_name="unshuffled_deduplicated_no" \
    --max_seq_length="128" \
    --weight_decay="0.01" \
    --per_device_train_batch_size="128" \
    --per_device_eval_batch_size="128" \
    --learning_rate="3e-4" \
    --warmup_steps="1000" \
    --overwrite_output_dir \
    --num_train_epochs="18" \
    --adam_beta1="0.9" \
    --adam_beta2="0.98" \
    --do_train \
    --do_eval \
    --logging_steps="500" \
    --eval_strategy="epoch" \
    --report_to="tensorboard" \
    --save_strategy="no"
```

### Script to compare pre-training with PyTorch on 8 GPU V100's

For comparison you can run the same pre-training with PyTorch on GPU. Note that we have to make use of `gradient_accumulation`
because the maximum batch size that fits on a single V100 GPU is 32 instead of 128.
Having created the tokenizer and configuration in `norwegian-roberta-base`, we create the following symbolic links:

```bash
ln -s ~/transformers/examples/pytorch/language-modeling/run_mlm.py ./
```

, set some environment variables:

```bash
export NUM_GPUS=8
export TOKENIZERS_PARALLELISM=0
export MODEL_DIR="./norwegian-roberta-base"
mkdir -p ${MODEL_DIR}
```

, and can start training as follows:

```bash
python3 -m torch.distributed.launch --nproc_per_node ${NUM_GPUS} run_mlm.py \
    --output_dir="${MODEL_DIR}" \
    --model_type="roberta" \
    --config_name="${MODEL_DIR}" \
    --tokenizer_name="${MODEL_DIR}" \
    --dataset_name="oscar" \
    --dataset_config_name="unshuffled_deduplicated_no" \
    --max_seq_length="128" \
    --weight_decay="0.01" \
    --per_device_train_batch_size="32" \
    --per_device_eval_batch_size="32" \
    --gradient_accumulation="4" \
    --learning_rate="3e-4" \
    --warmup_steps="1000" \
    --overwrite_output_dir \
    --num_train_epochs="18" \
    --adam_beta1="0.9" \
    --adam_beta2="0.98" \
    --do_train \
    --do_eval \
    --logging_steps="500" \
    --eval_strategy="steps" \
    --report_to="tensorboard" \
    --save_strategy="no"
```

## Language model inference with bfloat16

The following example demonstrates performing inference with a language model using the JAX/Flax backend.

The example script run_bert_flax.py uses bert-base-uncased, and the model is loaded into `FlaxBertModel`.
The input data are randomly generated tokens, and the model is also jitted with JAX.
By default, it uses float32 precision for inference. To enable bfloat16, add the flag shown in the command below.

```bash
python3 run_bert_flax.py --precision bfloat16
> NOTE: For JAX Versions after v0.4.33 or later, users will need to set the below environment variables as a \
> temporary workaround to use Bfloat16 datatype. \
> This restriction is expected to be removed in future version
```bash
export XLA_FLAGS=--xla_cpu_use_thunk_runtime=false
```
bfloat16 gives better performance on GPUs and also Intel CPUs (Sapphire Rapids or later) with Advanced Matrix Extension (Intel AMX).  
By changing the dtype for `FlaxBertModel `to `jax.numpy.bfloat16`, you get the performance benefits of the underlying hardware.
```python
import jax
model = FlaxBertModel.from_pretrained("bert-base-uncased", config=config, dtype=jax.numpy.bfloat16)
```
Switching from float32 to bfloat16 can increase the speed of an AWS c7i.4xlarge with Intel Sapphire Rapids by more than 2x.