Cramming the training of a (BERT-type) language model into limited compute.

Cramming Language Model (Pretraining)

This repository contains code to replicate our research described in "Cramming: Training a Language Model on a Single GPU in One Day". We experiment with language model pretraining a BERT-type model with limited compute, wondering "how bad can it really be"?

You can find our paper here: https://arxiv.org/abs/2212.14034, and the abstract below:

Recent trends in language modeling have focused on increasing performance through scaling, and have resulted in an environment where training language models is out of reach for most researchers and practitioners. While most in the community are asking how to push the limits of extreme computation, we ask the opposite question:
How far can we get with a single GPU in just one day?

We investigate the downstream performance achievable with a transformer-based language model trained completely from scratch with masked language modeling for a single day on a single consumer GPU. Aside from re-analyzing nearly all components of the pretraining pipeline for this scenario and providing a modified pipeline with performance close to BERT, we investigate why scaling down is hard, and which modifications actually improve performance in this scenario. We provide evidence that even in this constrained setting, performance closely follows scaling laws observed in large-compute settings. Through the lens of scaling laws, we categorize a range of recent improvements to training and architecture and discuss their merit and practical applicability (or lack thereof) for the limited compute setting.

The Rules for Cramming

Setting:

• A transformer-based language model of arbitrary size is trained with masked-language modeling, completely from scratch.
• Existing pretrained models cannot be included in any part of the pipeline.
• Any raw text (excluding downstream data) can be included for training. This means that one can achieve speedups by making judicious choices about how and when to sample data, provided the sampling mechanism does not require a pre-trained model.
• The downloading and pre-processing of raw data is exempted from the total compute budget. Pre-processing may include CPU-based tokenizer construction, tokenization, and filtering, but cannot include representation learning (e.g. pre-training a word embedding is not allowed, unless it is counted towards the final runtime).
• Training proceeds on a single GPU for 24 hours.
• Downstream performance is evaluated on GLUE (https://gluebenchmark.com/). Downstream finetuning on GLUE is limited to brief training with only the training data of the downstream task (we consider 5 epochs or less) and needs to work with hyperparameters set globally for all GLUE tasks. Downstream finetuning is excluded from the total compute budget.

How to run the code

Requirements:

• PyTorch: torch
• huggingface: transformers, tokenizers, datasets
• hydra-core
• [OPTIONAL]deepspeed
• [OPTIONAL] flash-attention
• psutil
• einops
• [OPTIONAL] For The-Pile data, install zstandard

Installation

• Just clone for now, and install packages as described
• [Optional] Follow the instructions at https://pre-commit.com/ to install the pre-commit hooks.
• [Optional] For deduplication, first install rust curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh , then git clone https://github.com/google-research/deduplicate-text-datasets/tree/dev-v1 and then run cargo install --target-dir ../cramming/dedup
• [Optional] For FlashAttention, install package as instructed at https://github.com/HazyResearch/flash-attention

General Usage

Use the pretrain.py script to pretrain with limited compute. This repository uses hydra (https://hydra.cc/docs/intro/), so all fields in cramming/config can be modified on the command line. For example, the budget can be modified by providing budget=48 as additional argument, or the learning rate can be modified via train.optim.lr=1e-4. Check out the configuration folder to see all arguments.

Your first step should be to verify the installed packages. To do so, you can run python pretrain.py dryrun=True, which will run the default sanity check for a single iteration. From there, you can enable additional functionality. For example, modify the architecture, e.g. arch=bert-original and training setup train=bert-original. To really train a language model, you need to switch away from the sanity check dataset to at least data=bookcorpus-wikipedia.

Data Handling

The data sources from data.sources will be read, normalized and pretokenized before training starts and cached into a database. Subsequent calls with the same configuration will reused this database of tokenized sequences. By default, a new tokenizer will also be constructed and saved during this process. Important data options are data.max_entries_in_raw_dataset, which defines how much raw data will be loaded. For example, for a large data source such as C4, only a subset of raw data will be downloaded. Then, max_seq_in_tokenized_dataset bottlenecks how many processed sequences will be stored in the database. This number should be larger than the number of sequences expected to be read within the budget.

Additional Notes:

• A simple trick to run dataset preprocessing only is to run python pretrain.py data=... dryrun=True, which dry-runs the training, but runs the full data preprocessing. Later runs can then re-use the cached data.
• Dataset preprocessing is heavily parallelized. This might be a problem for your RAM. If this happens, reduce impl.threads. Especially the deduplication code does require substantial amounts of RAM.
• I would run first experiments with bookcorpus-wikipedia only, which preprocesses comparatively quickly and only then look into the full processed and filtered C4.

Evaluation

To evaluate pretrained models on GLUE (or some GLUE tasks), use eval.py. This script searches for saved models in the base directory. Given the name of a previous run, this script will, by default, retrieve the latest checkpoint saved with this name, and then run evaluations.

WandB

You can log runs to your weights&biases account. To do so, simply modify wandb.entity and wandb.project on the command line or at cramming/config/wandb/default.yaml.

Replicate the final recipe

To replicate the final recipe discussed in the paper, run

python pretrain.py name=amp_b4096_c5_o3_final arch=bert-c5 train=bert-o3 train.batch_size=4096 data=c4-subset-processed

to pretrain and

python eval.py eval=GLUE_sane name=amp_b4096_c5_o3_final eval.checkpoint=latest impl.microbatch_size=16 impl.shuffle_in_dataloader=True

to evaluate the model. The recipe called "crammed BERT" in the paper corresponds to the architecture called bert-c5 trained with training setup bert-o3 on data c4-subset-processed.

Additional Recipes

Pretraining: Single GPU:

python pretrain.py name=bert data=bookcorpus-wikipedia arch=bert-original train=bert-original

Multi-GPU:

torchrun --nproc_per_node=4 --standalone pretrain.py name=bert4gpu  data=bookcorpus-wikipedia arch=bert-original train=bert-original

Eval a huggingface checkpoint:

python eval.py dryrun=True eval=rte name=bert-finetuning eval.checkpoint=hf://bert-base-uncased

Sanity check for distributed code on CPU:

torchrun --nproc_per_node=4 --standalone  pretrain.py name=speedtest1 dryrun=True data=sanity-check-2  impl.backend=gloo

Additional examples for recipes can be found in the /scripts folder.

Todos:

The following options are currently broken/limited/work-in-progress. Use these at your own discretion, or open a pull-request with a fix.

• The-Pile needs to be downloaded in its entirety to be used, but the code could be updated to stream, just like C4.
• Data Preprocessing is wasteful in terms of RAM.
• Token Dropping is simplistic, a more involved version could be better.
• Code currently uses the "old" jit.script fusion, should move toward new torch.compile implementation at some point. The current inductor hook is also non-functional.
• Shampoo (see discussion at https://twitter.com/_arohan_/status/1608577721818546176?s=20)
• Causal Attention [I broke this shortly before release, if you want to re-test CA, you'd have to fix it first]
• LAWA

Contact

Please, feel free to contact us with any questions, or open an issue on Github.