LanesLanes
Use cases/Lanes Compute/Fine-tune LLMs

Fine-tune LLMs

Fine-tune open models on the right GPU with your own stack, from a quick LoRA run to a full multi-node job.

Fine-tuning ranges from a single-GPU LoRA run to a multi-node full fine-tune. Lanes Compute gives you the card that fits the job and lets you bring the tools you already use.

Lanes Compute is a GPU service in limited early access. You describe the job, get matched to hardware, and pay per second while the environment is live. Your pipeline does not have to change to fit it: the same containers and the same training code run on rented cards, whether that is one L40S or several H100 nodes.

How Lanes Compute does it

  • Right-sized GPUs. A100, H100, H200, or L40S depending on the model and method. A LoRA pass on a small model fits a single card, while a full fine-tune of a large one wants H100 or H200 SXM and sometimes several nodes. You choose per job instead of committing to one card size for everything.
  • Bring your own stack. PyTorch, JAX, vLLM, Axolotl, DeepSpeed, whatever your pipeline uses. Anything that runs in a container works out of the box, so the recipe you tested on a small card is the recipe that runs on the big one.
  • Multi-node when needed. InfiniBand and NVLink for larger fine-tunes. Nodes are wired together over fast InfiniBand, with NVLink inside each node, and jobs scale across boxes without extra setup.
  • Per-second billing. Short runs cost like short runs. The meter runs while the environment is live and stops at teardown, so an hour of LoRA is billed as an hour, not as a day of reserved hardware.

Why it matters

You are not boxed into one framework or one GPU size. You pick per job, and only pay for the time the job takes.

Fine-tuning also tends to involve proprietary data, which makes two details load-bearing. Datasets and checkpoints sit on persistent network volumes you control, in a region you choose, and Lanes does not train on your data or share it. The tuned weights end up as checkpoints on your volumes, ready for the next run or for serving.

Set it up

  1. Request access. Lanes Compute is in limited early access; request it through the form on the Compute page and say what you are tuning.
  2. Describe the job. Base model, method, and data size. A quick LoRA pass and a full fine-tune get matched to very different hardware, and that is the point.
  3. Get onboarded. Usually within a day or two.
  4. Run and tear down. Stand up the environment, mount the dataset volume, tune, write the checkpoint, tear down. Billing ends there, and the checkpoint waits for the next run.
Fine-tuning is where one-size hardware hurts twice. A standing big card makes every small LoRA run expensive, and a small card cannot hold a full fine-tune at all. Per-job sizing resolves both, and per-second billing is what makes per-job sizing affordable.

Further reading

See Lanes Compute.