Using TPUs for fine-tuning and deploying LLMs

If you’re using or planning to use TPUs with Google Cloud, you can now do so via dstack. Just specify the TPU version and the number of cores (separated by a dash), in the gpu property under resources.

Read below to find out how to use TPUs with dstack for fine-tuning and deploying LLMs, leveraging open-source tools like Hugging Face’s Optimum TPU and vLLM .

Below is an example of a dev environment:

type: dev-environment
name: vscode-tpu    

python: 3.11
ide: vscode

resources:
  gpu: v2-8

If you've configured the gcp backend, dstack will automatically provision the dev environment with a TPU.

Currently, maximum 8 TPU cores can be specified, so the maximum supported values are v2-8, v3-8, v4-8, v5litepod-8, and v5e-8. Multi-host TPU support, allowing for larger numbers of cores, is coming soon.

Deployment

You can use any serving framework, such as vLLM, TGI. Here's an example of a service that deploys Llama 3.1 8B using Optimum TPU and vLLM .

Optimum TPUvLLM

type: service
name: llama31-service-optimum-tpu

image: dstackai/optimum-tpu:llama31
env:
  - HUGGING_FACE_HUB_TOKEN
  - MODEL_ID=meta-llama/Meta-Llama-3.1-8B-Instruct
  - MAX_TOTAL_TOKENS=4096
  - MAX_BATCH_PREFILL_TOKENS=4095
commands:
  - text-generation-launcher --port 8000
port: 8000

spot_policy: auto
resources:
  gpu: v5litepod-4 

model:
  format: tgi
  type: chat
  name: meta-llama/Meta-Llama-3.1-8B-Instruct

Once the pull request is merged, the official Docker image can be used instead of dstackai/optimum-tpu:llama31.

type: service
name: llama31-service-vllm-tpu

env:
  - MODEL_ID=meta-llama/Meta-Llama-3.1-8B-Instruct
  - HUGGING_FACE_HUB_TOKEN
  - DATE=20240828
  - TORCH_VERSION=2.5.0
  - VLLM_TARGET_DEVICE=tpu
  - MAX_MODEL_LEN=4096
commands:
  - pip install https://storage.googleapis.com/pytorch-xla-releases/wheels/tpuvm/torch-${TORCH_VERSION}.dev${DATE}-cp311-cp311-linux_x86_64.whl
  - pip3 install https://storage.googleapis.com/pytorch-xla-releases/wheels/tpuvm/torch_xla-${TORCH_VERSION}.dev${DATE}-cp311-cp311-linux_x86_64.whl
  - pip install torch_xla[tpu] -f https://storage.googleapis.com/libtpu-releases/index.html
  - pip install torch_xla[pallas] -f https://storage.googleapis.com/jax-releases/jax_nightly_releases.html -f https://storage.googleapis.com/jax-releases/jaxlib_nightly_releases.html
  - git clone https://github.com/vllm-project/vllm.git
  - cd vllm
  - pip install -r requirements-tpu.txt
  - apt-get install -y libopenblas-base libopenmpi-dev libomp-dev
  - python setup.py develop
  - vllm serve $MODEL_ID 
      --tensor-parallel-size 4 
      --max-model-len $MAX_MODEL_LEN
      --port 8000
port:
  - 8000

spot_policy: auto
resources:
  gpu: v5litepod-4

model:
  format: openai
  type: chat
  name: meta-llama/Meta-Llama-3.1-8B-Instruct

Control plane

If you specify model when running a service, dstack will automatically register the model on the gateway's global endpoint and allow you to use it for chat via the control plane UI.

Memory requirements

Below are the approximate memory requirements for serving LLMs with their corresponding TPUs.

Model size	bfloat16	TPU	int8	TPU
8B	16GB	v5litepod-4	8GB	v5litepod-4
70B	140GB	v5litepod-16	70GB	v5litepod-16
405B	810GB	v5litepod-64	405GB	v5litepod-64

Note, v5litepod is optimized for serving transformer-based models. Each core is equipped with 16GB of memory.

Supported frameworks

Framework	Quantization	Note
TGI	bfloat16	To deploy with TGI, Optimum TPU must be used.
vLLM	int8, bfloat16	int8 quantization still requires the same memory because the weights are first moved to the TPU in bfloat16, and then converted to int8. See the pull request for more details.

Running a configuration

Once the configuration is ready, run dstack apply -f <configuration file>, and dstack will automatically provision the cloud resources and run the configuration.

Fine-tuning

Below is an example of fine-tuning Llama 3.1 8B using Optimum TPU and the Abirate/english_quotes dataset.

type: task
name: optimum-tpu-llama-train

python: "3.11"

env:
  - HUGGING_FACE_HUB_TOKEN
commands:
  - git clone -b add_llama_31_support https://github.com/dstackai/optimum-tpu.git
  - mkdir -p optimum-tpu/examples/custom/
  - cp examples/fine-tuning/optimum-tpu/llama31/train.py optimum-tpu/examples/custom/train.py
  - cp examples/fine-tuning/optimum-tpu/llama31/config.yaml optimum-tpu/examples/custom/config.yaml
  - cd optimum-tpu
  - pip install -e . -f https://storage.googleapis.com/libtpu-releases/index.html
  - pip install datasets evaluate
  - pip install accelerate -U
  - pip install peft
  - python examples/custom/train.py examples/custom/config.yaml


resources:
  gpu: v5litepod-8

Memory requirements

Below are the approximate memory requirements for fine-tuning LLMs with their corresponding TPUs.

Model size	LoRA	TPU
8B	16GB	v5litepod-8
70B	160GB	v5litepod-16
405B	950GB	v5litepod-64

Note, v5litepod is optimized for fine-tuning transformer-based models. Each core is equipped with 16GB of memory.

Supported frameworks

Framework	Quantization	Note
TRL	bfloat16	To fine-tune using TRL, Optimum TPU is recommended. TRL doesn't support Llama 3.1 out of the box.
Pytorch XLA	bfloat16

What's next?

1. Browse Optimum TPU , Optimum TPU TGI and vLLM .
2. Check dev environments, tasks, services, and fleets.

Multi-host TPUs

If you’d like to use dstack with more than eight TPU cores, upvote the corresponding issue .