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.
- Getting Started
- Runtime Performance Results
- Comparison To Alternatives
- Development
- Features and Diagnostics
We recommend starting with a single host first and then moving to multihost.
You need to run these steps once per project prior to any local development or cluster experiments.
- Create two gcs buckets in your project, one for to downloading and retrieving the dataset and the other for storing the logs.
- Download the dataset in your gcs bucket
bash download_dataset.sh {GCS_PROJECT} {GCS_BUCKET_NAME}
- Set config values for
base_output_directory
anddataset_path
inconfigs/base.yml
.tokenizer_path
is full path 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 inconfigs/base.yml
, for instance you may change thesteps
orlogging_period
by either modifyingconfigs/base.yml
or by passing insteps
andlogging_period
as additional args to thetrain.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.
Local development is a convenient way to run MaxText on a single host. It doesn't scale to multiple hosts.
- Create and SSH to the single-host TPU of your choice. We recommend a
v4-8
. - Clone MaxText onto that TPUVM.
- Within the root directory of that
git
repo, install dependencies by running:
bash setup.sh
- After installation completes, run training with the command:
python3 MaxText/train.py MaxText/configs/base.yml run_name=$YOUR_JOB_NAME
- If you want to decode, you can decode as follows.
python3 MaxText/decode.py MaxText/configs/base.yml run_name=$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.
- Use
bash docker_build_dependency_image.sh DEVICE=gpu
can be used to build a container with the required dependencies. - After installation is completed, run training with the command:
python3 MaxText/train.py MaxText/configs/base.yml run_name=$YOUR_JOB_NAME
- If you want to decode, you can decode as follows.
python3 MaxText/decode.py MaxText/configs/base.yml run_name=$YOUR_JOB_NAME
- If you see the following error when running inside a container, set a larger
--shm-size
(e.g.--shm-size=1g
)
Failed to execute XLA Runtime executable: run time error: custom call 'xla.gpu.all_reduce' failed: external/xla/xla/service/gpu/nccl_utils.cc:297: NCCL operation ncclCommInitRank(&comm, nranks, id, rank) failed: unhandled cuda error (run with NCCL_DEBUG=INFO for details); current tracing scope: all-reduce-start.2; current profiling annotation: XlaModule:#hlo_module=jit__unnamed_wrapped_function_,program_id=7#.
There are three patterns for running MaxText with more than one host.
- [GKE, recommended] Running Maxtext with xpk - Quick Experimentation and Production support
- [GCE] Running Maxtext with Multihost Jobs - Long Running Production Jobs with Queued Resources
- [GCE] Running Maxtext with Multihost Runner - Fast experiments via multiple ssh connections.
Once you've gotten workloads running, there are important optimizations you might want to put on your cluster. Please check the doc PREFLIGHT.md
For a 22B model. See full run configs in MaxText/configs/v4/
as 22b.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/v4/
as 52b.sh
.
Hardware | TFLOP/sec/chip | MFU |
---|---|---|
1x v4-384 | 154 | 56.0% |
2x v4-384 | 162 | 58.9% |
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% |
More details on reproducing these results can be found in MaxText/configs/README.md
.
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.
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.
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
:
- 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, setcollect_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. Settingstack_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.
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 dependencies, so we recommend running setup.sh
to install these if you have not already done so.
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.
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:https://my-output-bucket dataset_path=gs:https://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.
MaxText supports training and inference of various open models. Follow user guides under getting started section to know more.
- 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 instructions mentioned here.