Saxml is an experimental system that serves Paxml, JAX, and PyTorch models for inference.

A Sax cell (aka Sax cluster) consists of an admin server and a group of model servers. The admin server keeps track of model servers, assigns published models to model servers to serve, and helps clients locate model servers serving specific published models.

The example below walks through setting up a Sax cell and starting a TPU or GPU model server in the cell.

Install the gcloud CLI and set the default account and project:

gcloud config set account <your-email-account> gcloud config set project <your-project> 

Create a Cloud Storage bucket:

GSBUCKET=sax-data gcloud storage buckets create gs://${GSBUCKET} 

Create a Compute Engine VM instance:

gcloud compute instances create sax-admin \ --zone=us-central1-b \ --machine-type=e2-standard-8 \ --boot-disk-size=200GB \ --scopes=https://www.googleapis.com/auth/cloud-platform 

Use this guide to enable the Cloud TPU API in a Google Cloud project.

Create a Cloud TPU VM instance:

 gcloud compute tpus tpu-vm create sax-tpu \ --zone=us-central2-b \ --accelerator-type=v4-8 \ --version=tpu-vm-v4-base \ --scopes=https://www.googleapis.com/auth/cloud-platform 

Alternatively or in addition to the Cloud TPU VM instance, create a Compute Engine VM instance with GPUs:

gcloud compute instances create sax-gpu \ --zone=us-central1-b \ --machine-type=n1-standard-32 \ --accelerator=count=4,type=nvidia-tesla-v100 \ --maintenance-policy=TERMINATE \ --boot-disk-size=200GB \ --scopes=https://www.googleapis.com/auth/cloud-platform 

Consider creating a VM instance using the "GPU-optimized Debian 10 with CUDA 11.0" image instead, so the Nvidia CUDA stack doesn't need to be manually installed as described below.

SSH to the Compute Engine VM instance:

gcloud compute ssh --zone=us-central1-b sax-admin 

Inside the VM instance, clone the Sax repo and initialize the environment:

git clone https://github.com/google/saxml.git cd saxml saxml/tools/init_cloud_vm.sh 

Configure the Sax admin server. This only needs to be done once:

bazel run saxml/bin:admin_config -- \ --sax_cell=/sax/test \ --sax_root=gs://${GSBUCKET}/sax-root \ --fs_root=gs://${GSBUCKET}/sax-fs-root \ --alsologtostderr 

Start the Sax admin server:

bazel run saxml/bin:admin_server -- \ --sax_cell=/sax/test \ --sax_root=gs://${GSBUCKET}/sax-root \ --port=10000 \ --alsologtostderr 

SSH to the Cloud TPU VM instance:

gcloud compute tpus tpu-vm ssh --zone=us-central2-b sax-tpu 

Inside the VM instance, clone the Sax repo and initialize the environment:

git clone https://github.com/google/saxml.git cd saxml saxml/tools/init_cloud_vm.sh 

Start the Sax model server:

SAX_ROOT=gs://${GSBUCKET}/sax-root \ bazel run saxml/server:server -- \ --sax_cell=/sax/test \ --port=10001 \ --platform_chip=tpuv4 \ --platform_topology=2x2x1 \ --alsologtostderr 

You should see a log message "Joined [admin server IP:port]" from the model server to indicate it has successfully joined the admin server.

SSH to the Compute Engine VM instance:

gcloud compute ssh --zone=us-central1-b sax-gpu 

Install the Nvidia GPU driver, CUDA, and cuDNN. Note that Sax by default requires CUDA 11. To switch to CUDA 12, edit requirements-cuda.txt and replace jaxlib==0.4.7+cuda11.cudnn86 with jaxlib==0.4.7+cuda12.cudnn88.

Inside the VM instance, clone the Sax repo and initialize the environment:

git clone https://github.com/google/saxml.git cd saxml saxml/tools/init_cloud_vm.sh 

Enable the GPU-specific requirements.txt file:

cp requirements-cuda.txt requirements.txt 

Start the Sax model server:

SAX_ROOT=gs://${GSBUCKET}/sax-root \ bazel run saxml/server:server -- \ --sax_cell=/sax/test \ --port=10001 \ --platform_chip=v100 \ --platform_topology=4 \ --jax_platforms=cuda \ --alsologtostderr 

You should see a log message "Joined [admin server IP:port]" from the model server to indicate it has successfully joined the admin server.

Sax comes with a command-line tool called saxutil for easy usage:

# From the `saxml` repo root directory: alias saxutil='bazel run saxml/bin:saxutil -- --sax_root=gs://${GSBUCKET}/sax-root' 

saxutil supports the following commands:

• saxutil help: Show general help or help about a particular command.
• saxutil ls: List all cells, all models in a cell, or a particular model.
• saxutil publish: Publish a model.
• saxutil unpublish: Unpublish a model.
• saxutil update: Update a model.
• saxutil lm.generate: Use a language model generate suffixes from a prefix.
• saxutil lm.score: Use a language model to score a prefix and suffix.
• saxutil lm.embed: Use a language model to embed text into a vector.
• saxutil vm.generate: Use a vision model to generate images from text.
• saxutil vm.classify: Use a vision model to classify an image.
• saxutil vm.embed: Use a vision model to embed an image into a vector.

As an example, Sax comes with a Pax language model servable on a Cloud TPU VM v4-8 instance. You can use it to verify Sax is correctly set up by publishing and using the model with a dummy checkpoint.

saxutil publish \ /sax/test/lm2b \ saxml.server.pax.lm.params.lm_cloud.LmCloudSpmd2BTest \ None \ 1 

Check if the model is loaded by looking at the "selected replica address" column of this command's output:

saxutil ls /sax/test/lm2b 

When the model is loaded, issue a query:

saxutil lm.generate /sax/test/lm2b "Q: Who is Harry Porter's mother? A: " 

The result will be printed in the terminal.

To use a real checkpoint with the model, follow the Paxml tutorial to generate a checkpoint. The model can then be published in Sax like this:

saxutil publish \ /sax/test/lm2b \ saxml.server.pax.lm.params.lm_cloud.LmCloudSpmd2B \ gs://${GSBUCKET}/checkpoints/checkpoint_00000000 \ 1 

Use the same saxutil lm.generate command as above to query the model.

First get LLaMA pytorch_vars from Meta, then run the following script to convert the LLaMA PyTorch checkpoint to SAX format

python3 -m saxml/tools/convert_llama_ckpt --base llama_7b --pax pax_7b 

For the 7B model, this script roughly needs 50-60GB memory. For larger models, for example, the 70B model, this script would need 500-600GB memory to run.

The script load and save weights in a single pass. To fit less memory, modify convert() function to load/save weights in multiple passes. In each pass, load and save partial weights (subset of all weight variables).

After converting the checkpoint, the checkpoint folder should have the following structure

checkpoint_00000000 metadata/	metadata	state/	mdl_vars.params.lm*/	...	...	step/ 

Please create empty files “commit_success.txt” and put one in each folder. This will let SAX know this checkpoint is ready to use when loading the model. So the fully ready checkpoint should be as following:

checkpoint_00000000	commit_success.txt metadata/	commit_success.txt	metadata	state/	commit_success.txt	mdl_vars.params.lm*/	...	...	step/ 

Now the checkpoint is fully ready.

Then start the SAX server

GPU server:

SAX_ROOT=gs://${GSBUCKET}/sax-root \ bazel run saxml/server:server -- \ --sax_cell=/sax/test \ --port=10001 \ --platform_chip=a100 \ --platform_topology=1 \ --jax_platforms=cuda \ --alsologtostderr 

TPU server:

SAX_ROOT=gs://${GSBUCKET}/sax-root \ bazel run saxml/server:server -- \ --sax_cell=/sax/test \ --port=10001 \ --platform_chip=tpuv4 \ --platform_topology=2x2x1 \ --alsologtostderr 

Finally move the converted ckpt to your google cloud data bucket and publish the model

7B model

saxutil publish \ /sax/test/llama-7b \ saxml.server.pax.lm.params.lm_cloud.LLaMA7BFP16 \ gs://sax-data/pax-llama/7B \ 1 

70B model

saxutil publish \ /sax/test/llama-7b \ saxml.server.pax.lm.params.lm_cloud.LLaMA70BFP16TPUv5e \ gs://sax-data/pax-llama/70B \ 1