Skip to content

akash-trainer

Directory actions

More options

Directory actions

More options

Latest commit

 

History

History

akash-trainer

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
.Rhistory
.Rhistory
 
 
README.md
README.md
 
 
config.json
config.json
 
 
deploy.yaml
deploy.yaml
 
 
logo.png
logo.png
 
 

README.md

🤖 Akash Trainer

An openclaw-powered agent that let you train any ML/AI model on Akash GPUs with just a message. The agent would automate the whole training process. No setup. Just deploy.

Point it at a GitHub repo → it clones, installs, trains on a GPU, and pushes results back to your repo + close the deployment automatically.

Quick Start

  1. Go to Akash Console → New Deployment → paste the deploy.yaml
  2. Set three env vars:
Variable Example
REPO_URL https://github.com/you/your-project.git
TRAIN_CMD python3 train.py
GITHUB_TOKEN Your token with repo scope (create one here)
  1. Deploy → accept a bid → done

Your trained model and outputs appear on GitHub as a new branch: trained-output/YYYYMMDD-HHMMSS

How It Works

You deploy on Akash
        ↓
  Clone your repo
        ↓
  Install all dependencies (auto-detected)
        ↓
  Verify GPU + CUDA works
        ↓
  Run your training command
        ↓
  Collect all new files (models, logs, metrics)
        ↓
  Push to GitHub as a new branch
        ↓
  Done. Close deployment.

Supports: PyTorch, TensorFlow, HuggingFace, JAX, and any Python ML framework. Private repos via GITHUB_TOKEN. Custom setup commands for anything else.

Environment Variables

Variable Required Description
REPO_URL ✅ GitHub repo URL
TRAIN_CMD ✅ Training command (e.g. python3 train.py --epochs 50)
GITHUB_TOKEN Recommended For auto-pushing results to GitHub
REPO_BRANCH No Branch to clone (default: main, falls back to master)
SETUP_CMD No Run before training (e.g. pip3 install -e . && python3 download_data.py)
REQUIREMENTS No Custom requirements file (default: requirements.txt)

Resource Guide

Pick the right size for your project:

Project CPU RAM GPU Storage Est. Cost
MNIST / CIFAR demo 4 8Gi 1x any 20Gi ~$0.20-0.50
ResNet / YOLO / standard CV 8 16Gi 1x RTX4090+ 50Gi ~$1-5
LLM fine-tuning (LoRA) 16 32Gi 1x A100/H100 100Gi ~$5-20

Edit the profiles.compute section in deploy.yaml to adjust.

Optional: akash.yaml

Drop an akash.yaml in your repo root to customize training without changing the deployment:

# All fields optional — only set what you need
train_cmd: python3 scripts/train.py --epochs 50 --lr 0.001
setup_cmd: pip3 install -e . && python3 scripts/download_data.py
requirements: requirements-gpu.txt
exclude:
  - data/raw/
  - "*.gz"
env:
  WANDB_API_KEY: "your-key"
  HF_TOKEN: "hf_xxx"

Without akash.yaml, the trainer auto-detects everything from your repo.

What Gets Pushed to GitHub

A new branch trained-output/YYYYMMDD-HHMMSS containing:

  • TRAINING_REPORT.md — duration, GPU used, files collected
  • models/ — saved model files
  • output/ — training outputs
  • All new files created during training (except downloaded datasets and caches)

If training fails → TRAINING_FAILED.md with the last 20 lines of error log is pushed instead, so you can debug without opening the Akash dashboard.

Files over 100MB are auto-excluded (GitHub limit).

Pre-installed

PyTorch 2.5.1 (CUDA 12.4) · TensorFlow · NumPy · Pandas · scikit-learn · matplotlib · OpenCV · W&B · TensorBoard · tqdm · rich

Your requirements.txt can add or override anything.

Links