mikon 0.0.5

Self-hosted AI development manager for GPU job execution on your own servers

mikon

日本語版: README-ja.md

A self-hosted tool for managing AI development (training and evaluation jobs) on GPU servers from a browser.

Write Python the way you normally would, and just add a decorator to your functions. mikon automatically discovers your jobs, generates config forms, allocates GPUs, launches runs, and displays metrics, logs, and artifacts in real time.

---

Features

• Decorator-only discovery — Just add @mikon.job. No registration step, no code changes required.
• Auto-generated config UI — Pydantic fields on class Config(mikon.Config) become form widgets automatically (sliders, selects, checkboxes, etc.)
• NVIDIA and AMD support — Specify GPUs in unified nvidia:0 / amd:0 format. CUDA_VISIBLE_DEVICES / ROCR_VISIBLE_DEVICES are set automatically.
• Live monitoring — Metric charts and log streams update every few seconds via SSE.
• Artifact management — Call ctx.log_artifact() to make files downloadable from the browser.
• Lineage tracking — ctx.use_dataset() / ctx.use_artifact() automatically build an upstream/downstream graph.
• Module system — Register swappable components with @mikon.module; the UI generates a module-selection form automatically.
• Dataset management — Register existing paths or create datasets with a builder function.
• Document viewer — Markdown, Typst, and TypMark files placed in docs/ are viewable in the dashboard.
• File-based persistence — No SQL database required. Everything is stored as text files under .mikon/.
• Independent processes — Jobs keep running even if the dashboard is restarted.

---

Requirements

• Python 3.11+
• GPU server with NVIDIA drivers or AMD ROCm installed
• typst CLI (optional, for Typst document rendering)
• typmark-cli CLI (optional, for TypMark document rendering)

---

Quick Start

# Install
uv tool install mikon   # or: pip install mikon

# Initialize project
mikon init

# src/train.py
import mikon
from mikon import Config, RunContext
from pydantic import Field
from typing import Literal

class TrainConfig(Config):
    lr: float = Field(1e-3, gt=0, le=1)
    epochs: int = Field(10, ge=1, le=1000)
    optimizer: Literal["adam", "sgd"] = "adam"

@mikon.job
def train(config: TrainConfig, ctx: RunContext) -> None:
    for epoch in range(config.epochs):
        loss = 1.0 / (epoch + 1)
        ctx.log_metric("loss", loss, step=epoch)

# Start the server on the GPU machine
mikon serve

# Launch a job from the CLI
mikon run train --gpu nvidia:0

# Access the dashboard from your local machine via SSH port forwarding
ssh -L 8000:localhost:8000 you@gpu-server
# → Open http://localhost:8000 in your browser

---

Installation

uv tool install mikon   # recommended
pip install mikon

Dependencies (pynvml, psutil, watchfiles, fastapi, etc.) are installed automatically.

---

Documentation

For the full usage guide, SDK reference, CLI options, and API error model, see USAGE.md.

---

License

MIT