vLLM inference on v6e TPUs

This tutorial shows you how to run vLLM inference on v6e TPUs. It also shows you how to run the benchmark script for the Meta Llama-3.1 8B model.

To get started with vLLM on v6e TPUs, see the vLLM quickstart.

If you are using GKE, also see the GKE tutorial.

Before you begin

You must sign the consent agreement to use Llama3 family of models in the HuggingFace repo. Go to https://huggingface.co/meta-llama/Llama-3.1-8B, fill out the consent agreement, and wait until you are approved.

Prepare to provision a TPU v6e with 4 chips:

1. Follow Set up the Cloud TPU environment guide to ensure you have appropriate access to use Cloud TPUs.

2. Create a service identity for the TPU VM.

   gcloud alpha compute tpus tpu-vm service-identity create --zone=zone

3. Create a TPU service account and grant access to Google Cloud services.

   Service accounts allow the Google Cloud TPU service to access other Google Cloud services. A user-managed service account is recommended. You can create a service account from the Google Cloud console or through the gcloud command.

   Create a service account using the gcloud command-line tool:

   gcloud iam service-accounts create your-service-account-name \
   --description="your-sa-description" \
   --display-name="your-sa-display-name"
   export SERVICE_ACCOUNT_NAME=your-service-account-name

   Create a service account from the Google Cloud console:

   1. Go to the Service Accounts page in the Google Cloud console.
   2. Click Create service account.
   3. Enter the service account name.
   4. (Optional) Enter a description for the service account.
   5. Click Create and continue.
   6. Choose the roles you want to grant to the service account.
   7. Click Continue.
   8. (Optional) Specify users or groups that can manage the service account.
   9. Click Done to finish creating the service account.

   After creating your service account, follow these steps to grant service account roles.

   The following roles are necessary:

   • TPU Admin: Needed to create a TPU
   • Storage Admin: Needed for accessing Cloud Storage
   • Logs Writer
   • Monitoring Metric Writer: Needed for writing metrics to Cloud Monitoring

   Your administrator must grant you the roles/resourcemanager.projectIamAdmin in order for you to assign IAM roles to users. A user with the Project IAM Admin roles/resourcemanager.projectIamAdmin role can also grant this role.

   Use the following gcloud commands to add service account roles:

   gcloud projects add-iam-policy-binding ${PROJECT_ID} \
      --member serviceAccount:${SERVICE_ACCOUNT_NAME}@${PROJECT_ID}.iam.gserviceaccount.com \
      --role roles/tpu.admin
   gcloud projects add-iam-policy-binding ${PROJECT_ID} \
      --member serviceAccount:${SERVICE_ACCOUNT_NAME}@${PROJECT_ID}.iam.gserviceaccount.com \
      --role roles/storage.admin
   gcloud projects add-iam-policy-binding ${PROJECT_ID} \
      --member serviceAccount:${SERVICE_ACCOUNT_NAME}@${PROJECT_ID}.iam.gserviceaccount.com \
      --role roles/logging.logWriter
   gcloud projects add-iam-policy-binding ${PROJECT_ID} \
      --member serviceAccount:${SERVICE_ACCOUNT_NAME}@${PROJECT_ID}.iam.gserviceaccount.com \
      --role roles/monitoring.metricWriter

   You can also assign roles using the Google Cloud console.

   From the Google Cloud console, select the following roles:

   1. Select your service account and click Add Principal.
   2. In the New Principals field, enter the email address of your service account.
   3. In the Select a role drop-down, search for role (for example, Storage Admin) and select it.
   4. Click Save.

4. Authenticate with Google Cloud and configure the default project and zone for Google Cloud CLI.

   gcloud auth login
   gcloud config set project PROJECT_ID
   gcloud config set compute/zone ZONE

Secure capacity

When you are ready to secure TPU capacity, review the quotas page to learn about the Cloud Quotas system. If you have additional questions about securing capacity, contact your Cloud TPU sales or account team.

Provision the Cloud TPU environment

You can provision TPU VMs with GKE, with GKE and XPK, or as queued resources.

Prerequisites

• This tutorial has been tested with Python 3.10 or later.
• Verify that your project has enough TPUS_PER_TPU_FAMILY quota, which specifies the maximum number of chips you can access within your Google Cloud project.
• Verify that your project has enough TPU quota for:
  • TPU VM quota
  • IP Address quota
  • Hyperdisk balanced quota
• User project permissions
  • If you are using GKE with XPK, see Cloud Console Permissions on the user or service account for the permissions needed to run XPK.

Provision a TPU v6e

   gcloud alpha compute tpus queued-resources create QUEUED_RESOURCE_ID \
      --node-id TPU_NAME \
       --project PROJECT_ID \
       --zone ZONE \
       --accelerator-type v6e-4 \
       --runtime-version v2-alpha-tpuv6e \
       --service-account SERVICE_ACCOUNT

Use the list or describe commands to query the status of your queued resource.

   gcloud alpha compute tpus queued-resources describe ${QUEUED_RESOURCE_ID}  \
      --project ${PROJECT_ID} --zone ${ZONE}

For a complete list of queued resource request statuses, see the Queued Resources documentation.

Connect to the TPU using SSH

  gcloud compute tpus tpu-vm ssh TPU_NAME

Install dependencies

1. Create a directory for Miniconda:

   mkdir -p ~/miniconda3

2. Download the Miniconda installer script:

   wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh -O ~/miniconda3/miniconda.sh

3. Install Miniconda:

   bash ~/miniconda3/miniconda.sh -b -u -p ~/miniconda3

4. Remove the Miniconda installer script:

   rm -rf ~/miniconda3/miniconda.sh

5. Add Miniconda to your PATH variable:

   export PATH="$HOME/miniconda3/bin:$PATH"

6. Reload ~/.bashrc to apply the changes to the PATH variable:

   source ~/.bashrc

7. Create a Conda environment:

   conda create -n vllm python=3.11 -y
   conda activate vllm

8. Clone the vLLM repository and navigate to the vLLM directory:

   git clone https://github.com/vllm-project/vllm.git && cd vllm
   

9. Clean up the existing torch and torch-xla packages:

   pip uninstall torch torch-xla -y
   

10. Install other build dependencies:

    pip install -r requirements-tpu.txt
    VLLM_TARGET_DEVICE="tpu" python setup.py develop
    sudo apt-get install libopenblas-base libopenmpi-dev libomp-dev
    

Get access to the model

Generate a new Hugging Face token if you don't already have one:

1. Click Your Profile > Settings > Access Tokens.
2. Select New Token.
3. Specify a Name of your choice and a Role with at least Read permissions.
4. Select Generate a token.

5. Copy the generated token to your clipboard, set it as an environment variable and authenticate with the huggingface-cli:

   export TOKEN=YOUR_TOKEN
   git config --global credential.helper store
   huggingface-cli login --token $TOKEN

Download benchmarking data

1. Create a /data directory and download the ShareGPT dataset from Hugging Face.

   mkdir ~/data && cd ~/data
   wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json
   

Launch the vLLM server

The following command downloads the model weights from Hugging Face Model Hub to the TPU VM's /tmp directory, pre-compiles a range of input shapes, and writes the model compilation to ~/.cache/vllm/xla_cache.

For more details, refer to the vLLM docs.

   cd ~/vllm
   vllm serve "meta-llama/Meta-Llama-3.1-8B" --download_dir /tmp --num-scheduler-steps 4 --swap-space 16 --disable-log-requests --tensor_parallel_size=4 --max-model-len=2048 &> serve.log &

Run vLLM benchmarks

Run the vLLM benchmarking script:

   python benchmarks/benchmark_serving.py \
       --backend vllm \
       --model "meta-llama/Meta-Llama-3.1-8B"  \
       --dataset-name sharegpt \
       --dataset-path ~/data/ShareGPT_V3_unfiltered_cleaned_split.json  \
       --num-prompts 1000

Clean up

Delete the TPU:

gcloud compute tpus queued-resources delete QUEUED_RESOURCE_ID \
    --project PROJECT_ID \
    --zone ZONE \
    --force \
    --async