Mathurin Videau*, Badr Youbi Idrissi*, Daniel Haziza, Luca Wehrstedt, Jade Copet, Olivier Teytaud, David Lopez-Paz. *Equal and main contribution
Meta Lingua is a minimal and fast LLM training and inference library designed for research. Meta Lingua uses easy-to-modify PyTorch components in order to try new architectures, losses, data, etc. We aim for this code to enable end to end training, inference and evaluation as well as provide tools to better understand speed and stability. While Meta Lingua is currently under development, we provide you with multiple apps
to showcase how to use this codebase.
The following commands launch a SLURM job that creates an environment for Meta Lingua. The env creation should take around 5 minutes without counting downloads.
git clone https://github.com/facebookresearch/lingua
cd lingua
bash setup/create_env.sh
# or if you have access to a SLURM cluster
sbatch setup/create_env.sh
Once that is done your can activate the environment
conda activate lingua_<date>
use the provided script to download and prepare data from huggingface (among fineweb_edu
, fineweb_edu_10bt
, or dclm_baseline_1.0
).
This command will download the fineweb_edu
and prepare it for training in the ./data
directory, specifying the amount of memory terashuf
(the tool used to shuffle samples) will be allocated. By default, the number of chunks (nchunks
) is 32. If you are running on fewer than 32 GPUs, it is recommended to set nchunks
to 1 or to match nchunks
with the number of GPUs (nchunks
= NGPUs). See here for more details.
python setup/download_prepare_hf_data.py fineweb_edu <MEMORY> --data_dir ./data --seed 42 --nchunks <NCHUNKS>
to download tokenizer (here llama3), use the folowing script:
python setup/download_tokenizer.py llama3 <SAVE_PATH> --api_key <HUGGINGFACE_TOKEN>
Now launch a debug job to check if everything works. The provided configurations are templates, you need to adapt them for them to work (change dump_dir
, data.root_dir
, data.tokenizer.path
, etc ...)
# stool stands for SLURM tool !
python -m lingua.stool script=apps.main.train config=apps/main/configs/debug.yaml nodes=1 partition=<partition>
# if you want to launch locally you can use torchrun
torchrun --nproc-per-node 8 -m apps.main.train config=apps/main/configs/debug.yaml
# or you can also launch on 1 GPU
python -m apps.main.train config=apps/main/configs/debug.yaml
When using stool
, if a job crashes, it can be relaunched using sbatch:
sbatch path/to/dump_dir/submit.slurm
We get very strong performance on many downstream tasks and match the performance of DCLM baseline 1.0.
name | arc_challenge | arc_easy | boolq | copa | hellaswag | obqa | piqa | siqa | winogrande | nq | tqa |
---|---|---|---|---|---|---|---|---|---|---|---|
Transformer 1B | 36.48 | 62.83 | 62.57 | 79.00 | 63.62 | 37.40 | 75.14 | 45.19 | 61.64 | 8.75 | 26.31 |
minGRU 1B | 30.82 | 57.89 | 62.05 | 74.00 | 50.27 | 37.00 | 72.31 | 43.76 | 52.49 | 3.24 | 9.03 |
minLSTM 1B | 31.76 | 60.04 | 62.02 | 73.00 | 53.39 | 36.40 | 72.36 | 45.09 | 52.80 | 4.52 | 12.73 |
Hawk 1B | 34.94 | 63.68 | 62.42 | 76.00 | 63.10 | 38.20 | 73.23 | 46.01 | 55.33 | 8.42 | 23.58 |
Mamba 1B | 35.54 | 63.42 | 62.63 | 74.00 | 64.16 | 38.80 | 75.24 | 45.14 | 60.14 | 8.84 | 26.64 |
name | arc_challenge | arc_easy | boolq | copa | hellaswag | obqa | piqa | siqa | winogrande | mmlu | nq | tqa | bbh |
---|---|---|---|---|---|---|---|---|---|---|---|---|---|
Mamba 7B 200B tokens | 47.21 | 76.03 | 65.63 | 84.00 | 77.80 | 44.00 | 80.25 | 49.69 | 70.24 | 32.81 | 20.53 | 51.93 | 20.35 |
Llama 7B 200B tokens | 46.95 | 75.73 | 64.80 | 84.00 | 77.45 | 45.00 | 80.20 | 48.26 | 70.32 | 48.64 | 20.66 | 51.01 | 31.47 |
Llama 7B squared relu 1T tokens | 49.61 | 76.74 | 72.45 | 89.00 | 81.19 | 44.80 | 82.05 | 49.95 | 72.14 | 60.56 | 25.68 | 59.52 | 42.11 |
Meta Lingua is structured as follows:
📦meta-lingua
┣ 📂lingua # Core library
┃ ┣ 📜args.py
┃ ┣ 📜checkpoint.py
┃ ┣ 📜data.py
┃ ┣ 📜distributed.py
┃ ┣ 📜float8.py
┃ ┣ 📜logger.py
┃ ┣ 📜metrics.py
┃ ┣ 📜optim.py
┃ ┣ 📜probe.py
┃ ┣ 📜profiling.py
┃ ┣ 📜stool.py
┃ ┣ 📜tokenizer.py
┃ ┗ 📜transformer.py
┣ 📂setup
┃ ┣ 📜create_env.sh
┃ ┗ 📜download_prepare_hf_data.py
┗ 📂apps # Apps that put components together
┣ 📂main # Main language modeling app with llama
┃ ┣ 📂configs
┃ ┣ 📜eval.py
┃ ┣ 📜generate.py
┃ ┣ 📜train.py
┃ ┗ 📜transformer.py
┣ 📂fastRNN
┃ ┣ 📂component
┃ ┣ 📂hawk
┃ ┣ 📂minGRU
┃ ┣ 📂minLSTM
┣ 📂mamba
┣ 📂mtp # Multi token prediction
┗ 📂plots
The lingua
folder contains some essential and reusable components, while the apps
folder contains scripts that put those components together. For instance the main training loop is in apps/main
. We highly encourage you to use that as a template and modify it however you please to suit your experiments.
Nothing is sacred in Meta Lingua. We've specifically tried to make it as easily modifiable as possible! So feel free to branch out and modify anything.
Here's a quick description of the most important files and features:
transformer.py
: Defines model architecture. This is pure PyTorchnn.Module
! Nothing fancy here.distributed.py
: Handles distributing the model on multiple GPUs. This is done throughparallelize_module
function which wraps your vanillann.Module
and applies nearly any combination of Data Parallel, Fully Sharded Data Parallel, Model Parallelism,torch.compile
, activation checkpointing andfloat8
.data.py
: Dataloader for LLM pretraining.
profiling.py
: Small wrapper around xformers' profiler which provides automatic MFU and HFU calculation and dumps profile traces in profiling folder in your dump directory. It also has memory profiling trace.checkpoint.py
: Manages model checkpoints. It saves model in checkpoints folder in your dump dir in .distcp format which is the new PyTorch distributed saving method. This format allows to reload the model with a different number of GPUs and with a different sharding. You can also convert those into normal PyTorch checkpoints withtorch.distributed.checkpoint.format_utils.dcp_to_torch_save
and the other way aroundtorch_save_to_dcp
.args.py
: Utilities to work with configs.
Most components need configuration and we chose to use data classes to represent these configuration objects. args.py
helps with converting between config.yaml
and config dictionaries into the respective data classes.
So for examples the TrainArgs
in apps/main/train.py
has a LMTransformerArgs
, OptimArgs
, etc ... as children.
Here is an example configuration file that will be converted to TrainArgs
:
# This is where Meta Lingua will store anything related to the experiment.
dump_dir: /path/to/dumpdir
name: "debug"
steps: 1000
seed: 12
optim:
lr: 3e-4
warmup: 2000
lr_min_ratio: 0.000001
clip: 10.0
distributed:
fsdp_type: full_shard
compile: true
selective_activation_checkpointing: false
model:
dim: 1024
n_layers: 8
n_heads: 8
data:
root_dir: data/shuffled
sources:
wikipedia: 80.0
arxiv: 20.0
batch_size: 32
seq_len: 1024
load_async: true
tokenizer:
name: sp
path: tokenizers/llama2.model
The command line interface in all scripts (train.py
, eval.py
, stool.py
) uses OmegaConf
This accepts arguments as a dot list
So if the dataclass looks like
@dataclass
class DummyArgs:
name: str = "blipbloup"
mode: LMTransformerArgs = LMTransformerArgs()
@dataclass
class LMTransformerArgs:
dim: int = 512
n_layers: int = 12
Then you can pass model.dim = 32
to change values in LMTransformerArgs
or just name = tictac
for top level attributes.
train.py
simply takes as argument the path to a config file and will load that config. The behavior here is as follows:
- We instantiate
TrainArgs
with its default values - We override those default values with the ones in the provided config file
- We override the result with the additional arguments provided through command line
If we take the DummyArgs
example above, calling train.py
with train.py config=debug.yaml model.dim=64 name=tictac
where debug.yaml
contains
model:
n_layers: 24
will launch training with the config
DummyArgs(name="tictac", LMTransformerArgs(dim=64, n_layers=24))
Since we want to do distributed training, we need train.py
to run N times (with N being the number of GPUs)
The easiest way to do this is through SLURM. And in order to make that simpler, we provide lingua/stool.py
which is a simple python script that
- Saves the provided config to
dump_dir
- Copies your current code to
dump_dir
in order to back it up - Creates an sbatch file
submit.slurm
which is then used to launch the job with the provided config.
It can either be used through command line
python -m lingua.stool config=apps/main/configs/debug.yaml nodes=1 account=fair_amaia_cw_codegen qos=lowest
Or the launch_job
function directly. This allows you for example to create many arbitrary configs (to sweep parameters, do ablations) in a jupyter notebook and launch jobs directly from there.
Since the configuration file is copied to dump_dir
, an easy way to iterate is to simply change the config file and launch the same command above.
In order to iterate quickly, it is preferable not to have to wait for a SLURM allocation every time. You can instead ask SLURM to allocate resources for you, then once they're allocated you can run multiple commands on that same allocation.
For example you can do :
salloc --nodes 2 --cpus-per-gpu 16 --mem 1760GB --gres=gpu:8 --exclusive --time=72:00:00
Which will give you access to 2 nodes in your current terminal. Once the allocation is done, you will see some SLURM environement variables that were automatically added such as $SLURM_JOB_ID
and others... This allows you for example to do in the same terminal
srun -n 16 python -m apps.main.train config=apps/main/configs/debug.yaml
Which will run the python -m apps.main.train config=apps/main/configs/debug.yaml
command on each of the 16 GPUs. If this crashes or ends you can just relaunch srun
again because the nodes are already allocated to you and you don't have to wait for SLURM to give you the resources again.
This will also show you the outputs of all those commands in the same terminal which might become cumbersome.
Instead you can use stool
directly to configure logs to be separated into different files per GPU.
python -m lingua.stool config=apps/main/configs/debug.yaml nodes=2 launcher=bash dirs_exists_ok=true
Notice that we added launcher=bash
which basically means that the generated submit.slurm
will simply be executed instead of submitting it through sbatch
. The submit.slurm
has an srun
command also so this is very similar to the above srun
command. We also add dirs_exists_ok=true
to tell stool
that it is okay to override things in an existing folder (code, config, etc)
If you want to use pdb
to step through your code, you should use -n 1
to run only on 1 GPU.
Evaluations can run either during training periodically or you directly launch evals on a given checkpoint as follows:
srun -n 8 python -u -m apps.main.eval config=apps/main/configs/eval.yaml
You need to specify the checkpoint and dump dir of the evaluation in that config
Or through stool
with
python -m lingua.stool script=apps.main.eval config=apps/main/configs/eval.yaml nodes=1 account=fair_amaia_cw_codegen qos=lowest
📂example_dump_dir
┣ 📂checkpoints
┃ ┣ 📂0000001000
┃ ┣ 📂0000002000
┃ ┣ 📂0000003000
┃ ┣ 📂0000004000
┃ ┣ 📂0000005000
┃ ┣ 📂0000006000
┃ ┣ 📂0000007000 # Checkpoint and train state saved every 1000 steps here
┃ ┃ ┣ 📜.metadata
┃ ┃ ┣ 📜__0_0.distcp
┃ ┃ ┣ 📜__1_0.distcp
┃ ┃ ┣ 📜params.json
┃ ┃ ┣ 📜train_state_00000.json
┃ ┃ ┗ 📜train_state_00001.json
┣ 📂code # Backup of the code at the moment the job was launched
┣ 📂logs
┃ ┗ 📂166172 # Logs for each GPU in this SLURM job.
┃ ┃ ┣ 📜166172.stderr
┃ ┃ ┣ 📜166172.stdout
┃ ┃ ┣ 📜166172_0.err
┃ ┃ ┣ 📜166172_0.out
┃ ┃ ┣ 📜166172_1.err
┃ ┃ ┗ 📜166172_1.out
┣ 📂profiling
┃ ┣ 📂memory_trace_plot # Trace of memory usage through time for all GPUs
┃ ┃ ┣ 📜000102_h100-192-145_451082.html
┃ ┃ ┣ 📜000102_h100-192-145_451083.html
┃ ┗ 📂profile_CPU_CUDA_000104 # Profiling traces for all GPUs
┃ ┃ ┣ 📜h100-192-145_451082.1720183858874741723.pt.trace.json.gz
┃ ┃ ┗ 📜h100-192-145_451083.1720183858865656716.pt.trace.json.gz
┣ 📜base_config.yaml
┣ 📜config.yaml
┣ 📜metrics.jsonl
┗ 📜submit.slurm
Here we highlight some related work that is complementary to this one. Most important being torchtitan and torchtune.
Lingua is designed for researchers who want to experiment with new ideas for LLM pretraining and get quick feedback on both training/inference speed and downstream benchmarks. Our goal is to lower the barrier to entry for LLM research by providing a lightweight and focused codebase.
We see torchtitan, torchtune, and lingua as complementary tools. Torchtitan is excellent for large-scale work because it features 3D parallelism and is likely to integrate the latest PyTorch distributed training features more quickly, thanks to its close ties to the PyTorch team. On the other hand, Torchtune excels at fine-tuning, especially when GPU resources are limited, by offering various fine-tuning strategies like LoRA, QLoRA, DPO, and PPO.
A typical workflow could look like this: you might first test a new idea in Lingua, then scale it up further with Torchtitan, and finally use Torchtune for instruction or preference fine-tuning.
Although there's definitely some overlap among these codebases, we think it's valuable to have focused tools for different aspects of LLM work. For example, Torchtitan aims to showcase the latest distributed training features of PyTorch in a clean, minimal codebase, but for most research, you really don't need every feature PyTorch has to offer or the capability to scale to 100B parameters on 4096 GPUs. For instance, we think that FSDP + torch compile will cover 90% of all needs of a researcher. With lingua, we tried to ask "What's the minimal set of features needed to draw solid conclusions on the scalability of idea X?"
We believe this targeted approach helps researchers make progress faster without the mental overhead of using many techniques that might not be needed.
@misc{meta_lingua,
author = {Mathurin Videau, Badr Youbi Idrissi, Daniel Haziza, Luca Wehrstedt, Jade Copet, Olivier Teytaud, David Lopez-Paz},
title = {{Meta Lingua}: A minimal {PyTorch LLM} training library},
url = {https://github.com/facebookresearch/lingua},
year = {2024}
}
Meta Lingua is licensed under BSD-3-Clause license. Refer to the LICENSE file in the top level directory.