lambda-labs-gpu-cloud

Launch a Lambda Labs GPU instance via the cloud console, the lambda-cloud-client Python API, or curl against the REST API. You must add an SSH key before launching, select a GPU type and region, and wait 3-15 minutes for the instance to become active.

For fine-tuning a 7B model, a single A100 40GB at $1.29/hr is sufficient. For 70B models, use an 8x H100 SXM instance. A10 or A6000 GPUs work well for inference and development workloads.

Lambda Labs provides dedicated instances with pre-installed Lambda Stack, persistent filesystems, and 1-Click Slurm clusters. RunPod offers cheaper spot instances and serverless endpoints, while Vast.ai is a marketplace with the lowest prices but less managed infrastructure.

Yes, Lambda Labs supports multi-node training via torchrun with NCCL across instances in the same region, or through 1-Click Clusters providing 16-512 H100 or B200 GPUs with Slurm and 400 Gb/s InfiniBand interconnect.

SSH connection refused usually means the instance is still initializing, which takes 3-5 minutes for single-GPU and 10-15 minutes for multi-GPU instances. Verify the instance status is active and that your SSH key was added before launch.

The root volume is ephemeral, so all data in /home/ubuntu is lost on termination. Attach a persistent filesystem at launch time and store datasets, checkpoints, and outputs under /lambda/nfs/<filesystem_name> to preserve them.