๐ 1T or bust my dudes ๐
An implementation of model & data parallel GPT3-like models using the mesh-tensorflow library.
If you're just here to play with our pre-trained models, we strongly recommend you try out the HuggingFace Transformer integration.
Training and inference is officially supported on TPU and should work on GPU as well. This repository will be (mostly) archived as we move focus to our GPU-specific repo, GPT-NeoX.
In addition to the functionality offered by GPT-3, we also offer the following:
NB, while neo can technically run a training step at 200B+ parameters, it is very inefficient at those scales. This, as well as the fact that many GPUs became available to us, among other things, prompted us to move development over to GPT-NeoX.
Update 21/03/2021:
We're proud to release two pretrained GPT-Neo models trained on The Pile, the weights and configs can be freely downloaded from the-eye.eu.
1.3B: https://mystic.the-eye.eu/public/AI/gptneo-release/GPT3_XL/
2.7B: https://mystic.the-eye.eu/public/AI/gptneo-release/GPT3_2-7B/
For more information on how to get these set up, see the colab notebook, or read through the rest of the readme.
Model and Size | Pile BPB | Pile PPL | Wikitext PPL | Lambada PPL | Lambada Acc | Winogrande | Hellaswag |
---|---|---|---|---|---|---|---|
GPT-Neo 125M | ----- | ----- | 32.285 | 30.266 | 37.36% | 50.43% | 28.67% |
GPT-3 125M | ----- | ----- | ----- | 18.6 | 42.7% | 52.0% | 33.7% |
GPT-Neo 350M | ----- | ----- | 22.5657 | 13.876 | 47.27% | 51.14% | 32.16% |
GPT-3 350M | ----- | ----- | ----- | 9.09 | 54.3% | 52.1% | 43.6% |
GPT-3 Ada | 0.9631 | ----- | ----- | 9.954 | 51.60% | 52.90% | 35.93% |
GPT-Neo 1.3B | 0.7527 | 6.159 | 13.10 | 7.498 | 57.23% | 55.01% | 38.66% |
GPT-3 1.3B | ----- | ----- | ----- | 5.44 | 63.6% | 58.7% | 54.7% |
GPT-2 1.5B | 1.0468 | ----- | 17.48 | 10.634 | 51.21% | 59.40% | 40.03% |
GPT-Neo 2.7B | 0.7165 | 5.646 | 11.39 | 5.626 | 62.22% | 56.50% | 42.73% |
GPT-3 2.7B | ----- | ----- | ----- | 4.60 | 67.1% | 62.3% | 62.8% |
Model and Size | MathQA | PubMedQA | Piqa |
---|---|---|---|
GPT-Neo 125M | 22.78% | 55.10% | 63.06% |
GPT-3 125M | ----- | ----- | 64.6% |
GPT-Neo 350M | 23.45% | 53.80% | 65.07% |
GPT-3 350M | ----- | ----- | 70.2% |
GPT-3 Ada | 24.29% | 52.80% | 68.88% |
GPT-Neo 1.3B | 24.05% | 54.40% | 71.11% |
GPT-3 1.3B | ----- | ----- | 75.1% |
GPT-2 1.5B | 23.64% | 58.33% | 70.78% |
GPT-Neo 2.7B | 24.72% | 57.54% | 72.14% |
GPT-3 2.7B | ----- | ----- | 75.6% |
Note: All evaluations were done using our evaluation harness. Some results for GPT-2 and GPT-3 are inconsistent with the values reported in the respective papers. We are currently looking into why, and would greatly appreciate feedback and further testing of our eval harness.
git clone https://github.com/EleutherAI/GPTNeo
cd GPTNeo
pip3 install -r requirements.txt
Sign up for Google Cloud Platform, and create a storage bucket.
Create your VM through a google shell (https://ssh.cloud.google.com/
) with ctpu up --vm-only
so that it can connect to your Google bucket and TPUs and install the requirements with pip (see above).
Google colab provides tpu-v8s for free, which should be enough to finetune our models up to GPT3XL (1.5B parameter) sizes.
Click to run through our example colab notebook.
For more detailed instructions, run through our Training Guide below.
You can also choose to train GPTNeo locally on your GPUs. To do so, you can omit the Google cloud setup steps above, and git clone the repo locally. Run through the Training Guide below, then when running main.py, you simply have to omit the tpu
flag, and pass in GPU ids instead.
Note: Some users have reported having difficulty getting MTF to recognize their GPUs. See here for details and instructions on how to fix it.
Once you have a trained model, or you've downloaded one of our pre-trained models, generating text is as simple as running the main.py script with the --predict
flag on. You can pass a path to your prompt txt file with the --prompt
flag, like so:
python3 main.py --predict --prompt <example_prompt.txt> --tpu <tpu_name> --model <config_name>
or, if using GPUs:
python3 main.py --predict --prompt <example_prompt.txt> --gpu_ids <device:GPU:0 device:GPU:1> --model <config_name>
We recommend you use Huggingface's pretrained GPT2 tokenizer with our repo (instructions provided below), but if you want to train a model with a different vocabulary size, we provide facilities to train your own tokenizer like so:
python data/train_tokenizer.py \
--base_dir ./path/to/your/txt/files \
--output_dir ./output/path \
--file_type txt \
--vocab_size 50257
# if it succeeded, you should see the message
# 'tokenizer saved at ./output/path/byte-level-bpe.tokenizer.json'
If you just want to test training, you can skip this step and download some dummy data like so:
wget https://storage.googleapis.com/connors-datasets/bundestag/bundestag_0.tfrecords
Then copy the data to your bucket, or if using GPUs, a local directory:
gsutil cp bundestag_0.tfrecords gs://<your bucket>/
If using your own data to train, you can use the data/create_tfrecords.py
script to encode your text data into tfrecords.
Your data must either be in the form of lots of normal .txt files (one document per file), or in any format supported by lm_dataformat.
You can run the script without parameters to see help for all options.
In document mode Each example in the tfrecords is one (variably sized) document. This is to be used with the documents_fixed
and documents_random
sampling modes (For more details see the parameters reference section).
Document mode is the default mode.
The below command will tokenize all files in acceptable formats in base_dir using gpt2 tokenizer and save them to output_dir
python3 create_tfrecords.py --mode documents --input_dir <base> --name <name> --output_dir <output> --use_gpt2_tokenizer --minimum_size <min>
input_dir
: Defines the folder where your data is located. The script will encode all files present in this folder.name
: Name of output files will bename_i.tfrecords
where i is the number of the file.output_dir
: Where to save the tfrecords touse_gpt2_tokenizer
: Whether to use the pretrained HuggingFace GPT2 tokenizer, in which case the separator will be set to [50256].encoder_path
: if not using the pretrained gpt2 tokenizer, use this flag to provide a path to your generated tokenizer json.separator
: Written in list format, the separator token(s) to insert between documents (e.g. "[0]"). Will depend on your encoder.minimum_size
: The minimum size (in tokens) a document must have, otherwise it is discarded. This is what will later determine yourstitch
parameter:stitch * minimum_size
must always be greater or equaln_ctx
(For more details see the parameters reference section).
To use a dataset in a model, you must first register that dataset under ./configs/dataset_configs
folder. First choose a filename with a .json
extension. That filename will serve as the dataset identification. The config should be filled out the following manner.
If you have a dataset encoded using the pretrained gpt2 tokenizer, you can specify that like so:
{
"n_vocab": 50257,
"path": "gs://neo-datasets/openwebtext-documents/openwebtext_*.tfrecords",
"eval_path": "gs://neo-datasets/openwebtext-documents/openwebtext_*.tfrecords",
"tokenizer_is_pretrained": true,
"tokenizer_path": "gpt2"
}
or if you've trained a custom tokenizer, like so:
{
"n_vocab": 32768,
"path": "./path/to/your/*.tfrecords",
"eval_path": "./path/to/your/eval/*.tfrecords",
"tokenizer_path": "./path/to/your/byte-level-bpe.tokenizer.json"
}
Finally, in your model config, add the filename that you created above to the datasets
array.
The <dataset id>
will be the filename, excluding the .json
, that you created above
"datasets": [[<dataset id>, <stitch>, <datatype>, <weight>]] # datasets key defines at run time how each dataset is processed for training
Once you have your datasets set up, find a suitable config in /configs
.
Here we use a GPT3-XL sized model as an example, but there are many more in ./configs
, all of which have short summaries in the Available Configs section.
All you need to do is edit the dataset id as described above, and edit model_path
(where logs and checkpoints will be saved) to point to a cloud bucket you have write access to (or local path, if using GPUs).
{
"n_head": 32,
"n_vocab": 50257,
"embed_dropout": 0.1,
"lr": 0.0002,
"lr_decay": "cosine",
"warmup_steps": 3000,
"beta1": 0.9,
"beta2": 0.95,
"epsilon": 1e-8,
"opt_name": "adam",
"weight_decay": 0.1,
"train_batch_size": 512,
"attn_dropout": 0.1,
"train_steps": 286150,
"eval_steps": 0,
"predict_steps": 1,
"res_dropout": 0.1,
"eval_batch_size": 128,
"predict_batch_size": 1,
"iterations": 2500,
"n_embd": 2048,
"datasets": [["your_dataset_name", 25, "documents_random", 1.0]],
"model_path": "gs://neo-models/GPT3_XL",
"n_ctx": 2048,
"n_layer": 24,
"scale_by_depth": true,
"scale_by_in": false,
"attention_types" : [[["global"],24]],
"mesh_shape": "x:128,y:2",
"layout": "batch:x,memory_length:y,embd:y",
"activation_function": "gelu",
"recompute_grad": true,
"gradient_clipping": 1.0,
"tokens_per_mb_per_replica": 2048
}
python3 main.py --model <your_config_name> --steps_per_checkpoint <n> --tpu <tpu-name>
tpu
: Name of the TPU to use.steps_per_checkpoint
: The frequency in steps at which to save checkpoints.--auto_layout
and--auto_layout_and_mesh_shape
(Optional): Disable training and instead auto generate a memory efficientlayout
(andmesh_shape
)gpu_ids
: if training using GPUs, omit thetpu
flag and pass in the ids of your gpus. In the example below, we train on 3 GPUs, specifying their device ids delimited by spaces:
python3 main.py --model <your_config_name> --steps_per_checkpoint <n> --gpu_ids <device:GPU:0 device:GPU:1>
We have several model sizes available, but some of our configs require large TPUs and will need tweaking to run on smaller machines, or GPUs. Below is a short guide to each model in the configs directory:
TODO
Sacred helps track experiments and is much nicer to work with than tensorboard.
To setup:
-
Install Docker and Docker-compose
-
Run
docker-compose up
To use:
-
Ensure model_dir doesn't have any metric logs in it (it trips up the metric stuff for tensorboard, which assumes that it's a continuation of the existing run). You can use
gsutil rm -r ...
to delete model dir -
Run
python3 run_experiment.py --tpu sometpuhere --model someconfig.json
Options are the same asmain.py
. -
You can go to http://server_ip_goes_here:8081/ to see the Omniboard overview. If you prefer to see a tensorboard, the script also spins one up and automatically assigns it a port. The script should print out the tensorboard port near the top of the log.
If you are ever confused by the dataset of a particular config file, you can easily check the minimum and maximum token ids with a single command. This is useful for making sure that the vocabulary size of the model is at least as large as the maximum token id. Tensorflow will not error if you try to gather on a matrix with out of bounds indices, so you need to make sure your vocabulary size is sufficiently large.
python main --model {config_name} --check_dataset
In addition to being able to train large GPT's, this repository also allows you to easily do masked language modeling (BERT, RoBERTa). In order to do so, you must follow two additional steps.
-
When tokenizing your dataset, you must reserve a special id for the
[mask]
token. -
In the configs, you will have to define two additional fields
"mlm_training": true, # must be set to true
"mlm_mask_id": <mask id> # the mask id that you reserved from above
That's all you need to train a model with the MLM objective, good for any type of data that you have encoded properly. If you would like to tweak the other related hyperparameters, please continue reading.
"mlm_cls_token_id": <cls token id>, # auto append specified CLS token id on the left
"mlm_mask_prob": 0.15, # the probability of masking a token, defaults to 15%
"mlm_same_token_prob": 0.10, # probability of keeping the token the same, defaults to 10%
"mlm_random_token_prob": 0.10, # probability of tokens that are replaced with random tokens, 10% was recommended by the BERT paper
"mlm_mask_ignore_ids": [<cls token>, <sep token>] # ignore masking other special tokens, if any
Pick a valid config from /configs
and tweak the parameters as needed:
n_heads
: The number of attention heads.n_embd
: Size of the hidden layers, must be divisible byn_heads
.n_vocab
: Vocabulary size.embed_dropout
,res_dropout
,attn_dropout
: Dropout probability for word embedding/residuals/attentionlr
: Learning ratewarmup_steps
: Number of steps before full learning rate is reached (linear ramp from0
tolr
).lr_decay
:cosine
orlinear
.opt_name
:adam
oradafactor
.beta1
,beta2
andepsilon
:adam
optimizer params.beta1
,ada_epsilon1
andada_epsilon2
:adafactor
optimizer params.weight_decay
: Weight decay parameter, if not present no weight decay is used (the weight decay fix for Adam is used) (default: 0.01) (optional).train_batch_size
: Batch size during training.train_steps
: Number of training steps (batches), set to roughly ~1 epoch for now (total number of tokens in your dataset / number of tokens per batch (=train_batch_size
/n_ctx
)).eval_steps
: Number of steps to run for each evaluation. Set to0
for no eval. i.e After every checkpoint, the model is tested foreval_steps
iterations
: Number of steps queued to the TPU, must be smaller thansteps_per_checkpoint
. (default: 500)datasets
: List of tfrecords datasets to use. Each dataset is a list with the following parameters:[train glob , eval glob, stitch, sampling_mode, weight]
. So for example for a single dataset (note the double list):[["bundestag_*.tfrecords", "", 10, "random_sample", 1.0]]
dataset_id
: The name of a dataset configuration file in./configs/dataset_configs
stitch
: Ifsampling_mode
random_sample
is used, the input pipeline samples this amount of texts into one to sample from. You must select stitch so thatstitch * minimum_document_length >= n_ctx
sampling_mode
:chunks
(tfrecords are preprocessed into the correct length and are read sequentially) ordocuments_random
(stitch
amount of documents are concatenated and then an_ctx
chunk is randomly subsampled)weights
: How much relative weight this dataset should have compared to others
model
: Which model to train. Currently onlyGPT
is supported, and it defaults to this if not present.model_path
: Google storage bucket location (or local path, if using GPUs) to save model checkpoints and logs.n_ctx
: Size of context window. Default is 2048n_layer
: Number of layers (blocks) in the model.scale_by_depth
: If true, the weight initialization of layers are scaled by their depth as in the GPT2 paper.scale_by_in
: If true, the weight initialization of layers are scaled by their number of inputs as in the GPT2 paper.mesh_shape
: A Mesh is an n-dimensional array of processors with named dimensions used for parallelism in the mesh-tensorflow library. Each Tensor is split evenly across mesh dimensions according to the layout (see below). The 'mesh_shape' is the shape of this array, and must be equal to the number of processors. e.g., for a v3-128 TPU "mesh_shape": โx:16,y:8โ.layout
: A Tensor is laid out on its mesh with one slice on each processor. A Tensor "layout", is an injective partial map specifying which dimensions of the tensor are (evenly) split across which dimensions of the mesh. No dimension of a tensor may be split across two dimensions of its mesh and no two dimensions of a tensor may be split across the same dimension of its mesh. The user defines a global set of layout rules in the form of (tensor-dimension-name, mesh-dimension-name) pairs. A dimension of a tensor is split across a dimension of its mesh if there is a matching rule, e.g. (for the above example mesh_shape: "layout":"batch:x,heads:y"activation_function
:selu
(self normalizing) orgelu
(used by OA), activation function used in feed-forward passes. (default: gelu)attention_types
: the type of attention for each layer in a list of the following format [[["attention_type"], n_layers]]. e.g. for a 12 layer net [[["global"], 12]] or [[["local"], 10], [["global"], 2]].- Choose from:
linear
,global
,local
ornone
. We have found a 50/50 mix ofglobal
andlinear
to work well.none
allows you to create feed-forward only layers for more efficient PAR Transformer models.
- Choose from:
precision
:float32
orbfloat16
.tokens_per_mb_per_replica
: If not None, will split the batch up into smaller microbatches containingtokens_per_mb_per_replica
tokens to avoid OOMs. Gradients are accumulated locally and reduced once. IMPORTANT: mb refers to minibatch not megabyte here.
Mixture of Experts
moe_layers
: A list of layer numbers to append a mixture of experts layer onto. E.G:[2,4,6,8,10,12]
. We have experimentally found a moe layer for every two self-attention layers to work well.moe_params
: a dictionary of additional kwargs to pass in to the moe layer. E.G{"moe_dropout_rate": 0.0 }
Experimental features
axial_pos_emb_
: If true, uses [axial positional embedding](https://arxiv.org/abs/1912.12180.mlp_glu
: If true, uses a gated linear unit variant of feed forward layers.scalenorm
: If true, uses scalenorm instead of layernorm.rezero
: If true, uses rezero instead of layernorm.num_mem_kv
: adds memory / key values from the all-attention paper. Param is an int with the number of desired mem/key values.macaron
: if true - uses a macaron transformer for each layer block.
- finalize documentation
- update configs
If you have found GPT-Neo helpful in your work, you can cite this repository as
@software{gpt-neo,
author = {Black, Sid and
Gao, Leo and
Wang, Phil and
Leahy, Connor and
Biderman, Stella},
title = {{GPT-Neo: Large Scale Autoregressive Language
Modeling with Mesh-Tensorflow}},
month = mar,
year = 2021,
note = {{If you use this software, please cite it using
these metadata.}},
publisher = {Zenodo},
version = {1.0},
doi = {10.5281/zenodo.5297715},
url = {https://doi.org/10.5281/zenodo.5297715}
}
The version number should be replaced with the version number you are using, and the year corresponds to the project's open-source release.
If you are specifically interested in citing the GPT-Neo models trained on the Pile, we would appreciate also citing
@article{gao2020pile,
title={The Pile: An 800GB Dataset of Diverse Text for Language Modeling},
author={Gao, Leo and Biderman, Stella and Black, Sid and Golding, Laurence and Hoppe, Travis and Foster, Charles and Phang, Jason and He, Horace and Thite, Anish and Nabeshima, Noa and others},
journal={arXiv preprint arXiv:2101.00027},
year={2020}
}