Overview

MaxText is a high performance, arbitrarily scalable, open-source, simple, easily forkable, well-tested, batteries included LLM written in pure Python/Jax and targeting Google Cloud TPUs. MaxText typically achieves 55% to 60% model-flop utilization and scales from single host to very large clusters while staying simple and "optimization-free" thanks to the power of Jax and the XLA compiler.

MaxText aims to be a launching off point for ambitious LLM projects both in research and production. We encourage users to start by experimenting with MaxText out of the box and then fork and modify MaxText to meet their needs.

Table of Contents

• Getting Started
• Runtime Performance Results
• Comparison To Alternatives
• Development
• Features and Diagnostics

Getting Started

There are three recommended patterns for running MaxText. You can run locally, run on a cluster experimentally or spawn a production-style that is managed by Google Compute Engine. We recommend starting with Local Development, moving to Cluster Experimentation for some ad hoc development and ultimately running your long running jobs with Queued Resources.

Getting Started: Download Dataset and Configure

You need to run these steps once per project prior to any local development or cluster experiments.

1. Create two gcs buckets in your project, one for to downloading and retrieving the dataset and the other for storing the logs.
2. Download the dataset in your gcs bucket

bash download_dataset.sh {GCS_PROJECT} {GCS_BUCKET_NAME}

3. Set config values for base_output_directory and dataset_path in configs/base.yml. vocab_relative_path is relative to base_output_directory for loading the tokenizer. MaxText assumes these GCS buckets are created in the same project and that it has permissions to read and write from them. We also recommend reviewing the configurable options in configs/base.yml, for instance you may change the steps or logging_period by either modifying configs/base.yml or by passing in steps and logging_period as additional args to the train.py call.

To run maxtext the TPUVMs must have permission to read the gcs bucket. These permissions are granted by service account roles, such as the STORAGE ADMIN role.

Getting Started: Local Development

Local development is the faster and most convenient way to run MaxText. However, it doesn't scale to multiple hosts.

1. Create and SSH to the single-host TPU of your choice. We recommend a v4-8.
2. Clone MaxText onto that TPUVM.
3. Within the root directory of that git repo, install dependencies by running:

bash setup.sh

4. After installation completes, run training with the command:

python3 MaxText/train.py MaxText/configs/base.yml run_name=$YOUR_JOB_NAME

5. If you want to decode, you can decode as follows.

python3 MaxText/decode.py MaxText/configs/base.yml run_name=$YOUR_JOB_NAME

Be aware, these decodings will be random. To get high quality decodings you need pass in a checkpoint, typically via the load_parameters_path argument.

Getting Started: Quick Experiments on Multiple Hosts (or Multiple Slices)

This workflow using multihost_runner.py is optimized for quick experiments, repeatedly re-using the same TPUs. Because the multihost_runner.py script depends on long-lived ssh connections, we do not recommend it for any long-running jobs.

We call the runner machine the one that multihost_runner.py is called from. This script will ssh into TPUVM worker machines that are found from the --TPU_PREFIX flag, and must be different than the runner machine. If the runner machine is a cloud VM, it must be in the same project as the workers.

The multihost_runner.py script:

• Distributes your code by recursively copying the current state of the chosen directory to multiple worker TPUVM.
• Runs the code on the workers
• Logs and monitors the processes' error statuses and brings the logs back to the runner machine.

Although there are several steps below, most are for the initial setup. Once setup you can continually make changes to your code and re-run your code with only step 5.

1. Choose a directory on your runner machine to develop and clone MaxText into. The runner machine can either be a TPUVM or not, but it cannot be one of the workers. If your runner machine is a TPUVM, it needs service account roles that grant it permission to create queued resources and ssh into them, such as the TPU ADMIN role. Clone MaxText, and cd into the root of the repo.

2. Set your project, zone, and ssh keys.

   Set your gcloud config, see https://cloud.google.com/sdk/gcloud/reference/config for more.

   PROJECT=<project>
   

   ZONE=<zone>
   

   gcloud config set project $PROJECT
   gcloud config set compute/zone $ZONE
   

   Create ssh keys for gcloud, we recommend leaving a blank password (hit enter twice after running the below command). If you are prompted that the the file already exists you can choose not to overwrite by selecting "n".

   ssh-keygen -f ~/.ssh/google_compute_engine 
   

3. Create your instances via Queued Resource (QR). Choose names for your TPUs and QR:

   TPU_PREFIX=$YOUR_TPU_NAME # Use new names when you create new TPUs
   QR_ID=$TPU_PREFIX # Convenient to re-use the node names, but can be different
   

   Choose the number of nodes (we use 2 below, but you may customize this and other feature of your TPU(s))

   NODE_COUNT=2
   

   Create a multislice environment of nodes using create queued resources

   gcloud alpha compute tpus queued-resources create $QR_ID --accelerator-type=v4-8 --runtime-version=tpu-ubuntu2204-base --node-count=$NODE_COUNT --node-prefix=$TPU_PREFIX  --reserved
   

   We target the reserved pool above, but you may instead target the on-demand pool by omitting this flag, or target pre-emptible capacity with the --best-effort flag, which may be necessary if your reservation is full.

   You have to wait for the QR to become ACTIVE (as opposed to ACCEPTED or PROVISIONING) which corresponds to the worker nodes becoming READY (as opposed to CREATING). This may take a minute or two and can be checked via

   gcloud alpha compute tpus queued-resources list --filter=$QR_ID 
   

4. Install dependencies. Install the dependencies of train.py on each worker using multihost_runner.py:

   python3 multihost_runner.py --TPU_PREFIX=$TPU_PREFIX --COMMAND="bash setup.sh"
   

   If you are running the multihost_runner.py script from a TPUVM, you will need to set --INTERNAL_IP=true.

5. Run your training job.

   Set a RUN_NAME for your job:

   RUN_NAME=$YOUR_JOB_NAME # You may set this to any unique name for a fresh run.
   

   Set config values for base_output_directory and dataset_path in configs/base.yml if not set already.

   python3 multihost_runner.py --TPU_PREFIX=$TPU_PREFIX --COMMAND="python3 MaxText/train.py MaxText/configs/base.yml run_name=$RUN_NAME"
   

   If you are running the multihost_runner.py script from a TPUVM, you will need to set --INTERNAL_IP=true.

6. Clean up TPUs and QR when finished.

   gcloud alpha compute tpus queued-resources delete $QR_ID --force --async
   

   The --force flag deletes both the queued resources and the TPU VMs, without it only a SUSPENDED queued resource whose TPUs have already been deleted can itself be deleted. We highly recommend the --async flag since deleting the TPUs and QR will take a minute or two.

Getting Started: Production Jobs On Multiple Slices

The workflow using multihost_job.py is optimized for long running experiments, providing resiliency against hardware failure and avoiding long running ssh connections. Its latency is much higher than multihost_runner.py because it needs to provision new capacity each time. The multihost_job.py script ends once the request to create the TPUs is issued. Logs are written both to gcloud in real time and also sent to GCS at the end of the job.

The multihost_job.py script:

• Copies your code to your GCS bucket
• Spins up specified TPU VM(s) via CQR
• Directs the TPU's to download then run that code. Because this logic is within the CQR's startup script, if there hardware is interrupted, the job will be rescheduled and resumed.
• Logs to gcloud, and additionally sends the logs to GCS at the job end
• Delete the TPUs and QR at the end of the job.

1. Choose a directory on your runner machine to develop and clone MaxText into. The runner machine can either be a TPUVM or not. If your runner machine is a TPUVM, it needs service account roles that grant it permission to create queued resources and has write access to GCS, such as the TPU ADMIN and STORAGE ADMIN roles. Clone MaxText, and cd into the root of the repo.

2. Set your project, zone. Set your gcloud config, see https://cloud.google.com/sdk/gcloud/reference/config for more.

   PROJECT=<project>
   

   ZONE=<zone>
   

   gcloud config set project $PROJECT
   gcloud config set compute/zone $ZONE
   

3. Link to a GCS bucket. Create a bucket if you don't already have one, see: https://cloud.google.com/storage/docs/creating-buckets for instructions to create one. Once you've identified your bucket:

   BUCKET_NAME=<your-bucket>
   

4. Run your training job.

   *** IMPORTANT *** multihost_job creates a request for new capacity for each run! You cannot use this tool on existing capacity, instead we recommend multihost_runner for this purpose.

   Choose the number of nodes (we use 2 below, but you may customize this and other feature of your TPU(s))

   NODE_COUNT=2
   

   RUN_NAME=$YOUR_JOB_NAME # You may set this to any unique name for a fresh run.
   python3 multihost_job.py --NUM_SLICES=$NODE_COUNT --RUN_NAME=$RUN_NAME --BUCKET_NAME=$BUCKET_NAME --CQR_EXTRA_ARGS="--reserved" --COMMAND="bash setup.sh && python3 MaxText/train.py MaxText/configs/base.yml run_name=$RUN_NAME"
   

   We tell multihost_job to target the reserved pool by by including --reserved as extra arguments to the CQR request, but you may instead target the on-demand pool by removing the --CQR_EXTRA_ARGS flag (on-demand is default), or the pre-emptible pool with --CQR_EXTRA_ARGS="--best-effort", which may be necessary if your reservation is full.

5. View the job's logs in cloud logging.

   The link to your job's cloud logging is printed at the end of multihost_job output. Additionally logs are saved to GCS when your job finishes, and this bucket's URL is also printed by multihost_job.

Runtime Performance Results

TPU v4

For a 22B model. See full run configs in MaxText/configs/ as 1xv4-128.sh, 2xv4-128.sh, 4xv4-128.sh, and 8xv4-128.sh.

Hardware	TFLOP/sec/chip	MFU
1x v4-128	156	56.7%
2x v4-128	152	55.2%
4x v4-128	149	54.3%
8x v4-128	146	53.2%

For a 52B model. See full run configs in MaxText/configs/ as 1xv4-384.sh and 2xv4-384.sh.

Hardware	TFLOP/sec/chip	MFU
1x v4-384	154	56.0%
2x v4-384	162	58.9%

TPU v5e

For 16B, 32B, 64B, and 128B models. See full run configs in MaxText/configs/largest_job/ as 16b.sh, 32b.sh, 64b.sh, 128b.sh.

Hardware	16B TFLOP/sec/chip	16B MFU	32B TFLOP/sec/chip	32B MFU	64B TFLOP/sec/chip	64B MFU	128B TFLOP/sec/chip	128B MFU
1x v5e-256	120	61.10%	132	66.86%	118	59.90%	110	56.06%
2x v5e-256	117	59.37%	128	64.81%	112	56.66%	110	55.82%
4x v5e-256	117	59.14%	126	64.10%	110	55.85%	108	54.93%
8x v5e-256	115	58.27%	125	63.67%	108	54.96%	104	52.93%
16x v5e-256	111	56.56%	123	62.26%	105	53.29%	100	50.86%
32x v5e-256	108	54.65%	119	60.40%	99	50.18%	91	46.25%

More details on reproducing these results on v5e can be found in v5e_high_performance.md.

Comparison to Alternatives

MaxText is heavily inspired by MinGPT/NanoGPT, elegant standalone GPT implementations written in PyTorch and targeting Nvidia GPUs. MaxText is more complex but has an MFU more than three times the 17% reported most recently with that codebase, is massively scalable and implements a key-value cache for efficient auto-regressive decoding.

MaxText is more similar to Nvidia/Megatron-LM, a very well tuned LLM implementation targeting Nvidia GPUs. The two implementations achieve comparable MFUs. The difference in the codebases highlights the different programming strategies. MaxText is pure Python, relying heavily on the XLA compiler to achieve high performance. By contrast, Megatron-LM is a mix of Python and CUDA, relying on well-optimized CUDA kernels to achieve high performance.

MaxText is also comparable to Pax. Like Pax, MaxText provides high-performance and scalable implementations of LLMs in Jax. Pax focuses on enabling powerful configuration parameters, enabling developers to change the model by editing config parameters. By contrast, MaxText is a simple, concrete implementation of an LLM that encourages users to extend by forking and directly editing the source code. The right choice depends on your project's priorities.

Development

Whether you are forking MaxText for your own needs or intending to contribute back to the community, we wanted to offer simple testing recipes.

To run unit tests and lint, simply run:

bash unit_test_and_lint.sh

The full suite of end-to-end tests is in end_to_end/. We run them with a nightly cadence.

Features and Diagnostics

Collect Stack Traces

When running a Single Program, Multiple Data (SPMD) job on TPU VMs, the overall process can hang if there is any error or any VM hangs/crashes for some reason. In this scenario, capturing stack traces will help to identify and troubleshoot the issues for the jobs running on TPU VMs.

The following configurations will help to debug a fault or when a program is stuck or hung somewhere by collecting stack traces. Change the parameter values accordingly in MaxText/configs/base.yml:

1. Set collect_stack_trace: True to enable collection of stack traces on faults or when the program is hung. This setting will periodically dump the traces for the program to help in debugging. To disable this, set collect_stack_trace: False.
2. Set stack_trace_to_cloud: False to display stack traces on console. stack_trace_to_cloud: True will create a temporary file in /tmp/debugging in the TPUs to store the stack traces. There is an agent running on TPU VMs that will periodically upload the traces from the temporary directory to cloud logging in the gcp project. You can view the traces in Logs Explorer on Cloud Logging using the following query:

logName="projects/<project_name>/logs/tpu.googleapis.com%2Fruntime_monitor"
jsonPayload.verb="stacktraceanalyzer"

3. stack_trace_interval_seconds signifies the duration in seconds between each stack trace collection event. Setting stack_trace_interval_seconds: 600 will collect the stack traces every 600 seconds (10 minutes).

Here is the related PyPI package: https://pypi.org/project/cloud-tpu-diagnostics.

Ahead of Time Compilation (AOT)

To compile your training run ahead of time, we provide a tool train_compile.py. This tool allows you to compile the main train_step in train.py for target hardware (e.g. a large number of v5e devices) without using the target hardware, and instead you may use only a CPU or a single VM from a different family. This compilation helps with two main goals:

• It will flag any out of memory (OOM) information, such as when the per_device_batch_size is set too high, with an identical OOM stack trace as if it was compiled on the target hardware.

• The ahead of time compilation can be saved and then loaded for fast startup and restart times on the target hardware.

The tool train_compile.py is tightly linked to train.py and uses the same configuration file configs/base.yml. Although you don't need to run on a TPU, you do need to install jax[tpu] in addition to other dependenices, so we recommend running setup.sh to install these if you have not already done so.

Example AOT 1: Compile ahead of time basics

After installing the dependencies listed above, you are ready to compile ahead of time:

# Run the below on a single machine, e.g. a CPU
python3 MaxText/train_compile.py MaxText/configs/base.yml compile_topology=v5e-256 compile_topology_num_slices=2 \
global_parameter_scale=16 per_device_batch_size=4

This will compile a 16B parameter MaxText model on 2 v5e pods.

Example AOT 2: Save compiled function, then load and run it

Here is an example that saves then loads the compiled train_step, starting with the save:

Step 1: Run AOT and save compiled function

# Run the below on a single machine, e.g. a CPU
export LIBTPU_INIT_ARGS="--xla_enable_async_all_gather=true" 
python3 MaxText/train_compile.py MaxText/configs/base.yml compile_topology=v5e-256 \
compile_topology_num_slices=2 \
compiled_trainstep_file=my_compiled_train.pickle global_parameter_scale=16 \
per_device_batch_size=4 steps=10000 learning_rate=1e-3

Step 2: Run train.py and load the compiled function

To load the compiled train_step, you just need to pass compiled_trainstep_file=my_compiled_train.pickle into train.py:

# Run the below on each host of the target hardware, e.g. each host on 2 slices of v5e-256
export LIBTPU_INIT_ARGS="--xla_enable_async_all_gather=true" 
python3 MaxText/train.py MaxText/configs/base.yml run_name=example_load_compile \
compiled_trainstep_file=my_compiled_train.pickle \
global_parameter_scale=16  per_device_batch_size=4 steps=10000 learning_rate=1e-3 \
base_output_directory=gs://my-output-bucket dataset_path=gs://my-dataset-bucket

In the save step of example 2 above we included exporting the compiler flag LIBTPU_INIT_ARGS and learning_rate because those affect the compiled object my_compiled_train.pickle. The sizes of the model (e.g. global_parameter_scale, max_sequence_length and per_device_batch) are fixed when you initally compile via compile_train.py, you will see a size error if you try to run the saved compiled object with different sizes than you compiled with. However a subtle note is that the learning rate schedule is also fixed when you run compile_train - which is determined by both steps and learning_rate. The optimizer parameters such as adam_b1 are passed only as shaped objects to the compiler - thus their real values are determined when you run train.py, not during the compilation. If you do pass in different shapes (e.g. per_device_batch), you will get a clear error message reporting that the compiled signature has different expected shapes than what was input. If you attempt to run on different hardware than the compilation targets requested via compile_topology, you will get an error saying there is a failure to map the devices from the compiled to your real devices. Using different XLA flags or a LIBTPU than what was compiled will probably run silently with the environment you compiled in without error. However there is no guaranteed behavior in this case; you should run in the same environment you compiled in.