Branch summary

This branch contains all of the patches from the following branches:

• soft (can load soft prompts (but not train them); see that branch's README.md for more details)
• modelcompat (can load GPT-Neo, fairseq and GPT-NeoX models; see that branch's README.md for more details)
• lowmem-fastinstall (installs faster in Google Colab)
• main

Patches in this branch:

Inherited from soft:

• (/mesh_transformer/layers.py and /mesh_transformer/transformer_shard.py) Modified EmbeddingShard to support soft prompting (arXiv:2104.08691).

Inherited from modelcompat:

• (/mesh_transformer/layers.py) Added support for more types of positional embedding:
  • "pe": "sinusoidal": The original sinusoidal positional embedding described in the original paper on transformer models, "Attention Is All You Need" (https://arxiv.org/pdf/1706.03762.pdf), in section 3.5 Positional Encoding.
  • "pe": "fairseq_sinusoidal": A variant of sinusoidal positional embedding currently used by fairseq. Most fairseq causal language models use this type of positional embedding.
  • "pe": "neox_rotary": A variant of rotary positional embedding used by GPT-NeoX. When using this, you can still specify pe_rotary_dims like with the rotary PE type.
  • "pe": "alibi": Attention with Linear Biases (ALiBi), described in https://arxiv.org/pdf/2108.12409.pdf. The implementation of ALiBi in this repository differs from the paper in that, in figure 3 of the paper, instead of using [0], [-1, 0], [-2, -1, 0], etc. as the coefficients of the bias matrix, we use [0], [0, 1], [0, 1, 2], etc. which allows for more efficient implementation. This implementation is compatible with the original one (models trained with either implementation produce identical results when run with either implementation).
• (/mesh_transformer/layers.py and /mesh_transformer/transformer_shard.py) Changed the implementation of the loss function in ProjectionShard so that loss is only computed for tokens with token ID less than n_vocab.
• (/convert_neo_pytorch_model_to_jax.ipynb) Created this notebook, which converts GPT-Neo models from pytorch_model.bin format to a format usable by this branch.
• (/mesh_transformer/layers.py and /mesh_transformer/transformer_shard.py) Added some optional GPT-Neo/fairseq/GPT-NeoX/OPT compatibility config options to the v1 transformer:
  • compat: A string that can be set to "j", "neo", "fairseq", "neox", "opt" or "bloom". Setting this to something other than "j" changes the architecture of the transformer network slightly to better conform to that of the corresponding model. Defaults to "j".
  • attention_layers: A list with layers strings inside of it, each of which is either "global" or "local", specifying whether each layer should use global or local attention. If compat is set to "neo", this defaults to a list with alternating "global" and "local", otherwise defaults to all "global".
  • local_attention_window: A positive integer that specifies the window size for layers with local attention. Has no effect if all of your layers are global attention layers. Defaults to 256.
  • n_vocab_padding: A nonnegative integer, the amount of padding your input and output embeddings have. Defaults to 0.
  • pe_shift: A nonnegative integer -- what the first position ID should be when doing positional embedding. If compat is set to "fairseq", this defaults to 2, otherwise 0.
  • activation: A string describing which activation function to use. Possible options are listed in https://github.com/huggingface/transformers/blob/main/src/transformers/activations.py. The default activation function depends on what compat is set to:
    • "fairseq_lm": "gelu"
    • "neox" or "bloom": "gelu_fast"
    • "opt": "relu"
    • All other values: "gelu_new"
  • combined_qkv: A Boolean value. If set to True, uses GPT-NeoX-style strided QKV linear transformation with weight shape (d_model, d_model * 3) and bias shape (d_model * 3,) instead of three separate linear transformations each with weight shape (d_model, d_model) and bias shape (d_model,). Defaults to True if compat is "neox" or "bloom", otherwise False.
    • The first group of d_model // n_heads columns of the query weight and the first group of d_model // n_heads elements of the query bias correspond to the first group of d_model // n_heads columns of the strided QKV weight and the first group of d_model // n_heads elements of the strided QKV bias.
    • The first group of d_model // n_heads columns of the key weight and the first group of d_model // n_heads elements of the key bias correspond to the second group of d_model // n_heads columns of the strided QKV weight and the second group of d_model // n_heads elements of the strided QKV bias.
    • The first group of d_model // n_heads columns of the value weight and the first group of d_model // n_heads elements of the value bias correspond to the third group of d_model // n_heads columns of the strided QKV weight and the third group of d_model // n_heads elements of the strided QKV bias.
    • The second group of d_model // n_heads columns of the query weight and the second group of d_model // n_heads elements of the query bias correspond to the fourth group of d_model // n_heads columns of the strided QKV weight and the fourth group of d_model // n_heads elements of the strided QKV bias.
    • etc.
  • neox_gpt_j_residual: If compat is "neox", this corresponds to the gpt_j_residual config option of GPT-NeoX models and defaults to True. If compat is set to anything else, this has no effect.
  • pe_rotary_pct: A floating point value. If set to a value greater than or equal to 0 and less than or equal to 1, we will ignore the value of pe_rotary_dims if it is set and instead use math.floor(pe_rotary_pct * (d_model // n_heads)) as the value for pe_rotary_dims. Defaults to being unset.
  • do_layer_norm_before: A Boolean value. If True, causes the first layer norm to be done prior to self attention, and the second layer norm (if any) to be done before the MLP. Setting this to False causes the self attention to be done prior to the first layer norm and the MLP to be done prior to the second layer norm. Defaults to True. You should set this to False if you're using specifically the 350 million parameter OPT model, and leave this at True if you're using any OPT model with a different number of parameters or any non-OPT model.
• (/mesh_transformer/layers.py) The implementation of standard positional embedding has been changed because the original implementation would've thrown an error. This does not affect GPT-J since it uses rotary positional embedding.

Inherited from lowmem-fastinstall:

• (/mesh_transformer/util.py) ClipByGlobalNormState is now a subclass of optax.EmptyState instead of optax.OptState in order to maintain compatibility with optax 0.1.0 and higher.
• (/requirements.txt) Removed requirements that aren't needed to use inference and relaxed most of the still-existing requirements. Also now requires progressbar2. Installation takes less time now.

Inherited from lowmem:

• (/mesh_transformer/checkpoint.py) Copied the implementation of read_ckpt from https://github.com/VE-FORBRYDERNE/mtj-softtuner/commit/0b8a98f2a4cf84ff9dedd71808d7aa23e0551bc4

Inherited from main:

• (/mesh_transformer/transformer_shard.py, /mesh_transformer/util.py) Optax is no longer required for model inference (you still need it to train models).
• (/.gitattributes, /setup.cfg, /setup.py, /versioneer.py, /mesh_transformer/__init.py__, /mesh_transformer/_version.py) python-versioneer is now used to automatically generate the version string.
• (/requirements.txt) A new requirement for chex >= 0.0.7, < 0.1.3 has been added.
• (/mesh_transformer/checkpoint.py) If smart_open is not found, we will use Python's builtin open instead.
• (/mesh_transformer/checkpoint.py) Ray is now imported when you call read_sharded_v2() (the only function in the file that uses Ray) so that you don't need to have Ray installed if you don't use the function.
• (/mesh_transformer/transformer_shard.py) CausalTransformer now accepts dematerialized keyword argument, which defaults to False and can be set to True to attempt to predict the structure of the model parameter dictionary without using JAX.
• (/mesh_transformer/layers.py) All einops calls have been replaced with equivalent JAX calls, removing the need for the Python package einops.
• (/requirements.txt) No longer requires einops.
• (/requirements.txt) Changed to specifically require JAX 0.2.12
• (/mesh_transformer/sampling.py) nucleaus_filter now subtracts infinity from logits to be removed instead of 1e10.

Table of contents

1. Mesh Transformer JAX
   1. Updates
2. Pretrained Models
   1. GPT-J-6B
      1. Links
      2. Acknowledgments
      3. License
      4. Model Details
      5. Zero-Shot Evaluations
3. Architecture and Usage
   1. Fine-tuning
   2. JAX Dependency
4. TODO

Mesh Transformer JAX

A haiku library using the xmap/pjit operators in JAX for model parallelism of transformers.

The parallelism scheme is similar to the original Megatron-LM, which is efficient on TPUs due to the high speed 2d mesh network. There is also an experimental model version which implements ZeRo style sharding.

This library is designed for scalability up to approximately 40B parameters on TPUv3s, beyond which different parallelism strategies should be used. See other implementations such as GPT-NeoX or DeepSpeed for that.

One future direction for research is integrating this codebase with swarm-jax, to achieve further scalability with pipeline parallelism.

Updates

12-07-21: Added guide to fine tuning

Pretrained Models

GPT-J-6B

A 6 billion parameter, autoregressive text generation model trained on The Pile.

Links

Slim weights (bf16 weights only, for inference, 9GB)

Full weights (including optimizer params, 61GB)

Colab demo

Web demo

Aran's blog post

Acknowledgments

This project would not have been possible without compute generously provided by the TPU Research Cloud with assistance from EleutherAI.

Thanks to the Cloud TPU team at Google for providing early access to the Cloud TPU VM alpha (now publicly available!)

Thanks to everyone who have helped out one way or another (listed alphabetically):

• Aran Komatsuzaki for advice with experiment design and writing the blog posts.
• James Bradbury for valuable assistance with debugging JAX issues.
• Janko Prester for creating the web demo frontend.
• Laurence Golding for adding some features to the web demo.
• Leo Gao for running zero shot evaluations for the baseline models for the table.

License

The weights of GPT-J-6B are licensed under version 2.0 of the Apache License.

Model Details

Hyperparameter	Value
n_parameters	6,053,381,344
n_layers	28*
d_model	4,096
d_ff	16,384
n_heads	16
d_head	256
n_ctx	2,048
n_vocab	50,257 (same tokenizer as GPT-2/3)
position encoding	Rotary position encodings (RoPE)
RoPE dimensions	64

* each layer consists of one feedforward block and one self attention block

The model consists of 28 layers with a model dimension of 4096, and a feedforward dimension of 16384. The model dimension is split into 16 heads, each with a dimension of 256. Rotary position encodings (RoPE) was applied to 64 dimensions of each head. The model is trained with a tokenization vocabulary of 50257, using the same set of BPEs as GPT-2/GPT-3.

Zero-Shot Evaluations

Models roughly sorted by performance, or by FLOPs if not available.

Model	Weights	Training FLOPs	LAMBADA PPL ↓	LAMBADA Acc ↑	Winogrande ↑	Hellaswag ↑	PIQA ↑	Dataset Size (GB)
Chance	✔	0	~a lot	~0%	50%	25%	25%	0
GPT-3-Ada‡	✘	-----	9.95	51.6%	52.9%	43.4%	70.5%	-----
GPT-2-1.5B	✔	-----	10.63	51.21%	59.4%	50.9%	70.8%	40
GPTNeo-1.3B‡	✔	3.0e21	7.50	57.2%	55.0%	48.9%	71.1%	825
Megatron-2.5B*	✘	2.4e21	-----	61.7%	-----	-----	-----	174
GPTNeo-2.7B‡	✔	6.8e21	5.63	62.2%	56.5%	55.8%	73.0%	825
GPT-3-1.3B*‡	✘	2.4e21	5.44	63.6%	58.7%	54.7%	75.1%	~800
GPT-3-Babbage‡	✘	-----	5.58	62.4%	59.0%	54.5%	75.5%	-----
Megatron-8.3B*	✘	7.8e21	-----	66.5%	-----	-----	-----	174
GPT-3-2.7B*‡	✘	4.8e21	4.60	67.1%	62.3%	62.8%	75.6%	~800
Megatron-11B†	✔	1.0e22	-----	-----	-----	-----	-----	161
GPT-J-6B‡	✔	1.5e22	3.99	69.7%	65.3%	66.1%	76.5%	825
GPT-3-6.7B*‡	✘	1.2e22	4.00	70.3%	64.5%	67.4%	78.0%	~800
GPT-3-Curie‡	✘	-----	4.00	69.3%	65.6%	68.5%	77.9%	-----
GPT-3-13B*‡	✘	2.3e22	3.56	72.5%	67.9%	70.9%	78.5%	~800
GPT-3-175B*‡	✘	3.1e23	3.00	76.2%	70.2%	78.9%	81.0%	~800
GPT-3-Davinci‡	✘	-----	3.0	75%	72%	78%	80%	-----
Gopher 230B*	✘	6.31E+23	-----	74.50%	70.10%	79.20%	81.80%	1344
MT-NLG 530B*‡	✘	-----	-----	76.6%	73.0%	80.2%	82.0%	-----

* represents evaluation numbers reported by their respective authors, all other numbers are provided by running the lm-evaluation-harness either with the released weights or with API access. Due to subtle implementation differences as well as different zero shot task framing, these might not be directly comparable. See this blog post for more details.

† The Megatron-11B model provides no comparable metrics, and several implementations using the released weights do not reproduce the generation quality and evaluations. (see 1 2 3) Thus, evaluation was not attempted.

‡ These models have been trained with data which contains possible test set contamination. The OpenAI GPT-3 models failed to deduplicate training data for certain test sets, while the GPT-Neo models as well as this one is trained on The Pile, which has not been deduplicated against any test sets.

Architecture and Usage

Most scripts in this repository are designed to be run on TPUs, which under the TPU-VM architecture are virtual machines which can run arbitrary code. Most scripts are designed to spin up a TPU, SSH into it to set up the dependencies and copy code over from the local directory, and then start a Ray worker which can accept RPC calls.

The TPUVMs handles running model training steps and evaluation, checkpoint save and loading, while the driver python program handles data loading and general orchestration (such as when to save checkpoints etc).

This means that most scripts (train.py, eval_harness.py etc) expect to be running on a GCE virtual machine in the same region as the TPUs, to minimize RPC latency and data transfer cost. Other scripts (usually ones which don't take a --tpu argument, such as device_sample.py, device_serve.py or device_train.py) expect to be run directly on a TPUVM. The device_* scripts only work on a v3-8 and not on larger pods.

Furthermore, there is an example (resharding_example.py) of how to convert the provided checkpoints (which have 8 shards in the case of GPT-J-6B) down to a smaller number, such as for when running on GPU(s).

Fine-tuning

To fine-tune the model, run device_train.py on a TPU VM. Using a TPU v3-8, you can fine-tune at a rate of ~5000 tokens/second, which should be sufficient for small-to-medium-size datasets.

Please read the step by step guide for thorough fine-tuning instructions.

JAX Dependency

Note this library has some specific requirements for JAX version. Specifically, to use the v1 models (including GPT-J 6B), jax==0.2.12 is required. This in turn depends on jaxlib==0.1.68. If this is not done, you will get cryptic xmap errors

However, to use the v2 model code (no publicly released weights), the newest JAX version can be used.

Citation

To cite this repository:

@misc{mesh-transformer-jax,
  author = {Wang, Ben},
  title = {{Mesh-Transformer-JAX: Model-Parallel Implementation of Transformer Language Model with JAX}},
  howpublished = {\url{https://github.com/kingoflolz/mesh-transformer-jax}},
  year = 2021,
  month = May
}

To cite the weights of GPT-J-6B:

@misc{gpt-j,
  author = {Wang, Ben and Komatsuzaki, Aran},
  title = {{GPT-J-6B: A 6 Billion Parameter Autoregressive Language Model}},
  howpublished = {\url{https://github.com/kingoflolz/mesh-transformer-jax}},
  year = 2021,
  month = May
}

If you use this repository or any of the pretrained weights to do something cool, we would love to hear about it. Feel free to open a github issue or reach out over email (in profile).

TODO

• disentangle heads and shards
• test/benchmark on TPU
• implement gradient checkpointing
• fix initialization
• mixed precision
• deal with preemptible TPUs
• test and validate generation
• shard activations instead of replicating for memory efficiency (in v2)
• support ZeRO style sharding (in v2)