# VaultLayer

> VaultLayer is a BYOC (bring-your-own-cloud) AI training control plane. It keeps
> PyTorch, JAX, and Hugging Face training and fine-tuning jobs running reliably on
> your own cloud accounts, GPU contracts, reserved instances, or credits — adding
> orchestration, checkpointing, monitoring, and automatic failure recovery on top,
> with external GPU capacity available as overflow when your own fleet is constrained.

## What VaultLayer does
- Runs any training script unchanged: `vl run train.py` wraps your existing
  PyTorch / JAX / Hugging Face command — no SDK, decorators, or framework migration.
- BYOC-first: route jobs to your own connected cloud account, reserved instances,
  or GPU contracts. BYOC jobs never touch a managed pool and carry no per-run
  VaultLayer charge.
- Automatic recovery: when a GPU is reclaimed, crashes, or goes unhealthy,
  VaultLayer resumes from the last checkpoint on available compute.
- Elastic GPU capacity: add external GPUs (H100, A100, L40S, A10G, RTX 4090, etc.)
  on demand for overflow or urgent jobs.
- Per-job isolation: each job runs on its own provisioned GPU with job-scoped
  credentials — no shared containers.

## Who it is for
Funded AI startups, research labs, and individual ML engineers running real GPU
training and fine-tuning workloads who already have cloud credits, reserved
instances, or GPU contracts and need jobs to finish reliably.

## Key facts
- Install: `pip install vaultlayer && vaultlayer init`
- Frameworks: PyTorch, JAX/Flax, Hugging Face Transformers/TRL/Accelerate/PEFT,
  PyTorch Lightning, DeepSpeed, Axolotl.
- Focus: training and fine-tuning, not real-time inference serving.
- Access: open signup at https://dashboard.vaultlayer.cloud/signup — install the
  CLI and submit a first job in minutes.
- Contact: rahuljain@vaultlayer.cloud (founder-direct) — for BYOC pilots, GPU
  contracts, or team evaluations.

## Guides
- BYOC GPU training: https://vaultlayer.cloud/byoc-gpu-training
- Fault-tolerant training (auto-resume after interruptions): https://vaultlayer.cloud/fault-tolerant-training
- Fine-tune LLMs on your own cloud: https://vaultlayer.cloud/fine-tune-llms
- VaultLayer vs SkyPilot: https://vaultlayer.cloud/vs/skypilot
- VaultLayer vs renting GPUs directly (RunPod, Lambda, Vast.ai): https://vaultlayer.cloud/vs/diy-runpod-lambda
- Managed vs self-hosted GPU training: https://vaultlayer.cloud/managed-vs-self-hosted-gpu-training
- What is a GPU training control plane: https://vaultlayer.cloud/learn/gpu-training-control-plane
- What is BYOC for AI training: https://vaultlayer.cloud/learn/byoc-ai-training
- How checkpoint & resume works on interruptible GPUs: https://vaultlayer.cloud/learn/checkpoint-resume-spot-gpus
- Distributed multi-node GPU training: https://vaultlayer.cloud/distributed-multi-node-training
- VaultLayer vs AWS SageMaker: https://vaultlayer.cloud/vs/aws-sagemaker
- VaultLayer vs Modal: https://vaultlayer.cloud/vs/modal
- GPU types for training (H100 vs A100 vs L40S): https://vaultlayer.cloud/learn/gpu-types-for-training
- How to resume PyTorch training after a crash: https://vaultlayer.cloud/guides/resume-pytorch-training-after-a-crash
- Run PyTorch training on cloud GPUs: https://vaultlayer.cloud/pytorch-cloud-training
- Train Hugging Face models on cloud GPUs: https://vaultlayer.cloud/huggingface-trainer-cloud
- VaultLayer vs CoreWeave: https://vaultlayer.cloud/vs/coreweave
- VaultLayer vs RunPod: https://vaultlayer.cloud/vs/runpod
- VaultLayer vs Slurm: https://vaultlayer.cloud/vs/slurm
- QLoRA vs LoRA vs full fine-tuning: https://vaultlayer.cloud/learn/qlora-vs-lora-vs-full-finetune
- How much GPU memory to fine-tune an LLM (7B/13B/70B): https://vaultlayer.cloud/learn/gpu-memory-to-fine-tune-llm
- Run AI training on your own AWS GPUs (BYOC): https://vaultlayer.cloud/aws-gpu-training
- Run AI training on your own GCP GPUs (BYOC): https://vaultlayer.cloud/gcp-gpu-training
- Use your cloud credits & reserved GPUs for training: https://vaultlayer.cloud/use-cloud-credits-for-training
- Spot GPU training that actually finishes: https://vaultlayer.cloud/spot-gpu-training
- FSDP vs DDP (PyTorch parallelism): https://vaultlayer.cloud/learn/fsdp-vs-ddp
- VaultLayer vs Google Vertex AI: https://vaultlayer.cloud/vs/vertex-ai
- Fix CUDA out of memory during training: https://vaultlayer.cloud/guides/cuda-out-of-memory-fix
- Training job keeps getting preempted: https://vaultlayer.cloud/guides/training-job-keeps-getting-preempted
- Checkpoint training to S3 / object storage: https://vaultlayer.cloud/guides/checkpoint-to-s3

## Links
- Home: https://vaultlayer.cloud/
- VaultLayer for startups: https://vaultlayer.cloud/startups
- Security & credential handling: https://vaultlayer.cloud/security
- Docs: https://vaultlayer.cloud/docs
- Contact: rahuljain@vaultlayer.cloud