Deploy a cluster of heterogenous Infernet nodes on Amazon Web Services (AWS) and / or Google Cloud Platform (GCP), using Terraform for infrastructure procurement and Docker compose for deployment.

1. Install Terraform
2. Configure nodes: A node configuration file for each node being deployed.
   • See example configuration.
   • They must have unique names
     • A straightforward approach would be 0.json, 1.json, etc...
   • They must be placed under the top-level configs/ directory.
   • Number and name of .json files must match the number and name of keys in the nodes variable in terraform.tfvars.
     • See terraform.tfvars.example.
     • Each key should correspond to the name of a .json file, excluding the .json postfix.
   • Each node strictly requires its own configuration .json file, even if those are identical.
   • For instructions on configuring individual nodes, refer to the Infernet Node.

The Infernet Router REST server is configured automatically by Terraform. However, if you plan to use it, you need to understand its implications:

IMPORTANT: When configuring a heterogeneous node cluster (i.e. 0.json, 1.json, etc. are not identical), container IDs should be reserved for a unique container setup at the cluster level, i.e. across nodes (and thus .json files). That is becuase the router uses container IDs to make routing decisions between services running across the cluster.

Example: Consider nodes A and B, each running a single LLM inference container; node A runs image1, and node B runs image2. If we set id: "llm-inference" in both containers (containers[0].id attribute in 0.json, 1.json), the router will be unable to disambiguate between the two services, and will consider them interchangeable, which they are not. Any requests for "llm-inference" will be routed to either container, which is an error.

Therefore, re-using a IDs across configuration files must imply an identical container configuration, including image, environment variables, command, etc. This will explicitly tell the router which containers are interchangeable, and allow it to distribute requests for those containers across all nodes running that container.

1. Create an AWS service account for deployment:

   cd procure/aws
   chmod 700 create_service_account.sh
   ./create_service_account.sh

   This will require local authentication with the AWS CLI. Add access_key_id and secret_access_key to your Terraform variables (see step 3).

2. Make a copy of the example configuration file terraform.tfvars.example:

   cd procure/aws
   cp terraform.tfvars.example terraform.tfvars

3. Configure your terraform.tfvars file. See variables.tf for config descriptions.

4. Run Terraform:

   # Initialize
   cd procure
   make init provider=aws
   
   # Print deployment plan
   make plan provider=aws
   
   # Deploy
   make apply provider=aws
   
   # WARNING: Destructive
   # Destroy deployment
   make destroy provider=aws

1. Create a GCP service account for deployment:

   cd procure/gcp
   chmod 700 create_service_account.sh
   ./create_service_account.sh

   This will require local authentication with the GCP CLI, and create a local credentials file. Add the path to the credentials file (gcp_credentials_file_path) to your Terraform variables (see step 3).

2. Make a copy of the example configuration file terraform.tfvars.example:

   cd procure/gcp
   cp terraform.tfvars.example terraform.tfvars

3. Configure your terraform.tfvars file. See variables.tf for config descriptions.

4. Run Terraform:

   # Initialize
   cd procure
   make init provider=gcp
   
   # Print deployment plan
   make plan provider=gcp
   
   # Deploy
   make apply provider=gcp
   
   # WARNING: Destructive
   # Destroy deployment
   make destroy provider=gcp

# Install tflint
brew install tflint

# Install plugins
tflint --init

# Run on all directories
tflint --recursive

# Format AWS files
cd procure/aws
terraform fmt

# Format GCP files
cd procure/gcp
terraform fmt

BSD 3-clause Clear