Feature Request: LambdaLabs Automatic Runner Allocation

[ei-grad]

Lambda Labs is gaining popularity in the machine learning community due to its competitive pricing, availability of the latest GPU models, and overall ease of use. Integrating Lambda Labs cloud API support into CML (Continuous Machine Learning) would provide users with an additional cost-effective cloud option for running ML workloads in CI/CD pipelines.

Motivation

• Cost Efficiency: Lambda Labs offers lower pricing compared to major cloud providers, making it an attractive option for ML engineers and researchers.
• GPU Availability: Easy access to NVIDIA GH200, H100, A100, etc., which may require additional effort, such as quota requests or meeting specific terms, on other cloud providers.
• Simplicity: Pre-configured ML environment simplify the workflow without the need for complex setup and configuration.

Challenges

Lambda Labs provides a minimalistic and straightforward API compared to major cloud providers. While this makes integration relatively simple, there are some limitations to consider:

• No Custom AMIs: Unlike AWS or GCP, Lambda Labs does not allow using custom machine images. It only provides an Ubuntu 22.04 with their so-called "Lambda Stack" preconfigured.
• No Cloud-Init Support: There is no option to pass cloud-init metadata for instance customization at launch.

Since Lambda Labs lacks support for custom AMIs and cloud-init, the typical approach to setting up an instance involves manual instance configuration via SSH after the instance creation.

Additional Context

• Lambda Labs official website: https://lambdalabs.com
• Lambda Cloud API documentation: https://docs.lambdalabs.com/public-cloud/cloud-api/
• FAQ: https://lambdalabs.com/service/gpu-cloud/faqs
• Lambda Stack: https://lambdalabs.com/lambda-stack-deep-learning-software