Fine-tune at any scale.
From one GPU to a thousand.
Fine-tune open-source models on single or multi-node GPU clusters in minutes. Bring-your-own framework — PyTorch, JAX, DeepSpeed and HF Accelerate all supported.
1from asc import Cluster, gpu 2 3cluster = Cluster( 4 name="llama3-finetune", 5 nodes=4, 6 gpu_per_node=gpu.H100(count=8), 7 image="asc/torch:2.4-cu124", 8) 9 10@cluster.run11def train():12 import torch13 # ... your distributed training loop ...14 15train()Designed for production.
Multi-node out of the box
NCCL + InfiniBand pre-wired. Spin up an 8×H100 cluster with one flag.
Checkpoint anywhere
Stream checkpoints directly to our object store. Resume on a different cluster size.
Datasets at rest
Mount petabyte datasets via DFS with near-local IOPS. No data movement tax.
Per-second billing
Crashes don't cost a full hour. You only pay while GPUs are actively training.
Spot + preemption safe
Automatic checkpoint-and-resume on spot reclamation — no babysitting.
Sweeps + W&B
Native sweeps, hyperparameter search and W&B / MLflow integration.
Metered. No markup.
Pay per active second / per GiB. Free tier covers small projects; $200/mo cap until you opt in. See the full calculator.
| Line item | Unit | Rate (USD) |
|---|---|---|
| GPU A10G pod | per hour | $0.35 |
| GPU L4 (functions) | per GPU-second | $0.000095 |
| Storage | per GiB-month | $0.026 |
| Egress | per GiB | $0.098 |
Ship your first deploy in minutes.
Free $30/month of compute. No card required.