interpreta

Interpreta is a GKE-powered managed notebook environment for mechanistic interpretability research. Born from the resource limitations of online notebook environments, it provides researchers complete control over dependencies and the compute resources they want for running large-scale experiments.

Features

• Python and jupyter notebook environment are already setup
• Mech interp packages are setup
• Whatever machine you want within the Google ecosystem is setup. Currently it is an n1-standard-1.

Packages

Python environment: python 3.11.6

    "torch==2.5.1" \
    "einops==0.8.0" \
    "jaxtyping==0.2.36" \
    "tqdm==4.67.1" \
    "numpy==1.26.4" \
    "transformer-lens==2.9.1" \
    "sae-lens==5.2.0" \
    "plotly==5.24.0" \
    "transformers==4.47.0" \
    "huggingface_hub==0.27.0"

Stack

terraform, helm, docker

Setup instructions

FYI, I aliased kubectl with k.

1. Assuming you've set up your terraform backend within GCP, run a terraform apply within infrastructure/environments/dev. This will spin up your cluster.

2. Run k get nodes. You should see two nodes.

   NAME                                                  STATUS   ROLES    AGE     VERSION
   gke-interpreta-gke-interpreta-node-po-756e7c1c-l47x   Ready    <none>   20h     v1.29.7-gke.1104000
   gke-interpreta-gke-interpreta-node-po-756e7c1c-v159   Ready    <none>   9m20s   v1.29.7-gke.1104000
   

3. To get the jupyter environment running: cd helm and then make install

4. After that completes, you should be able to see the services: k get svc --namespace jhub

   You should hopefully see a few different services that are not exposed to the internet:

   NAME           TYPE        CLUSTER-IP       EXTERNAL-IP   PORT(S)    AGE
   hub            ClusterIP   34.118.226.209   <none>        8081/TCP   99s
   proxy-api      ClusterIP   34.118.229.142   <none>        8001/TCP   99s
   proxy-public   ClusterIP   34.118.235.32    <none>        80/TCP     99s
   

5. To access the jupyter login page, you need to port-forward the traffic to the proxy-public service.

   make proxy

   Then you should be able to access the jupyter login page on http://localhost:8080