Overview

MaxText is a high performance, highly scalable, open-source LLM written in pure Python/Jax and targeting Google Cloud TPUs and GPUs for training and inference. MaxText achieves high MFUs 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.

We have used MaxText to demonstrate high-performance, well-converging training in int8 and scale training to ~51K chips.

Key supported features:

TPUs and GPUs (in preview)
Training and Inference (in preview)
Models: Llama2, Mistral and Gemma

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

Getting Started

For your first time running MaxText, we provide specific instructions.

MaxText supports training and inference of various open models. Follow user guides in the getting started folder to know more.

Some extra helpful guides:

Gemma: a family of open-weights Large Language Model (LLM) by Google DeepMind, based on Gemini research and technology. You can run decode and finetuning using these instructions.
Llama2: a family of open-weights Large Language Model (LLM) by Meta. You can run decode and finetuning using these instructions.
Mixtral: a family of open-weights sparse mixture-of-experts (MoE) model by Mistral AI. You can run decode and finetuning using these instructions

In addition to the getting started guides, there are always other MaxText capabilities that are being constantly being added! The full suite of end-to-end tests is in end_to_end. We run them with a nightly cadence. They can be a good source for understanding MaxText Alternatively you can see the continuous unit tests which are run almost continuously.

Runtime Performance Results

More details on reproducing these results can be found in MaxText/configs/README.md.

TPU v5p

No. of params	Accelerator Type	TFLOP/chip/sec	Model flops utilization (MFU)
32B	v5p-128	3.28e+02	71.47%
64B	v5p-128	3.23e+02	70.31%
128B	v5p-256	3.15e+02	68.68%
128B	v5p-512	3.15e+02	68.53%
256B	v5p-1024	3.16e+02	68.82%
512B	v5p-1024	2.94e+02	63.99%
1024B	v5p-2048	2.49e+02	64.05%
1024B	v5p-4096	2.97e+02	64.80%
1160B	v5p-7680	2.95e+02	64.27%
1160B	v5p-12288	3.04e+02	66.23%

TPU v5e

For 16B, 32B, 64B, and 128B models. See full run configs in MaxText/configs/v5e/ 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%

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, supporting more industry standard models and scaling to tens of thousands of chips. Ultimately MaxText 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 various LLMs that encourages users to extend by forking and directly editing the source code.

Features and Diagnostics

Collect Stack Traces

When running a Single Program, Multiple Data (SPMD) job on accelerators, 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:

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.
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"
```
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 full cluster.

TPU Support

You may use only a CPU or a single VM from a different family to pre-compile for a TPU cluster. 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 dependencies, 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 initially 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.

GPU Support

Ahead-of-time compilation is also supported for GPUs with some differences from TPUs:

GPU does not support compilation across hardware: A GPU host is still required to run AoT compilation, but a single GPU host can compile a program for a larger cluster of the same hardware.
For A3 Cloud GPUs, the maximum "slice" size is a single host, and the compile_topology_num_slices parameter represents the number of A3 machines to precompile for.

Example

This example illustrates the flags to use for a multihost GPU compilation targeting a cluster of 4 A3 hosts: