Lambda

Lambda rents on-demand GPU instances (full VMs) in several US regions, billed per minute. Instances boot the Lambda Stack image (Ubuntu with NVIDIA drivers, CUDA and common ML frameworks).

export LAMBDA_API_KEY=...          # or [lambda] api_key in config.toml
gpuprov provision --cloud lambda --name dev --gpus A6000:1
gpuprov provision --cloud lambda --name big --gpus H100:8 --region us-east-1

Create an API key at https://cloud.lambda.ai/api-keys.

Acceptance testing pending

This adapter is covered by unit and mock-API tests. It has not yet passed the end-to-end acceptance run against a live account (allocate → SSH → k3s/GPU ready → terminate → nothing billable left), so treat it as preview.

Planning: live capacity

One call, GET /api/v1/instance-types, returns every instance type with its price and the regions that can launch it right now. Each type × region with capacity becomes one offer:

• instance_type = Lambda's type, e.g. gpu_1x_a6000 or gpu_8x_h100_sxm5
• region = Lambda region (us-east-1, us-west-1, …); there are no zones
• ordered by hourly price, cheapest first

Without --region, only US regions (us-*) are considered. Pass --region explicitly for any other region. --zone also pins a region, because Lambda has no zones.

--gpus matches the GPU description by whole tokens: A10 does not match A100, A6000 matches the RTX A6000, A100-80GB matches only the 80 GB A100, and H100:1 picks the single-GPU PCIe type while H100:8 picks the 8× SXM5 box. --instance-type takes an exact type name, including CPU types.

Capacity is re-checked right before launching, so a stale offer fails before anything is registered or created.

What gets created

Resource	Notes
instance	named and hostnamed gpuprov-<name>, tagged gpuprov-vm-name=<name> and gpuprov-managed=true
firewall ruleset	only when extra ports are needed (--port, or the k3s API); named gpuprov-<name>, allows SSH plus those ports, deleted on teardown
SSH key	your gpuprov public key, registered once and reused; never removed

The login user is ubuntu, created by gpuprov's cloud-init user data. --image takes a Lambda image family (e.g. lambda-stack-24-04) or an image id.

Kubernetes

Launches accept cloud-init user data, so the k3s installer runs while the VM boots and gpuprov verifies it over SSH. Firewall rulesets attach per instance, so clusters use direct API access: ports 6443, 80 and 443 are opened in the VM's ruleset.

Limits

Feature	Why
--spot	Lambda offers on-demand instances only

Lambda's legacy account-wide firewall rules do not apply in us-south-1, and gpuprov uses per-instance rulesets instead. If a region rejects ruleset creation, that offer is skipped and the next one is tried.

Teardown

gpuprov down terminates the instance, waits until Lambda reports it gone, then deletes the VM's firewall ruleset in that region. While the instance is still terminating the ruleset is "in use", so the delete is retried until it frees up.