FIT — Flexible Interactive Training

A lightweight GPU experiment scheduler with a built-in web dashboard and real-time log viewer. Write a shell script, add GPU directives as comments, submit it — FIT queues, allocates GPUs, runs the job, and streams logs to your browser.

Quick Start

# Terminal 1: start the daemon
fit daemon

# Terminal 2: submit an experiment
fit test_samples/mlp_demo.exp

Open http://127.0.0.1:8080 for the dashboard.

Installation

Requires Go 1.25+ and Node.js (for building the frontend).

git clone https://github.com/rin/fit.git
cd fit

# Build frontend
cd web-ui && npm install && npm run build && cd ..

# Build binary
go build -o fit ./cmd/fit/

# (optional) install to PATH
cp fit ~/bin/

Usage

Daemon

fit daemon

Starts the background scheduler, REST API, and web dashboard on :8080. All state lives under ~/.fit/.

Submitting Jobs

fit <script.exp>

Parses the .exp file and submits it to the daemon via HTTP.

Experiment File Format

A .exp file is a shell script with embedded directives in comments:

#!/usr/bin/env fit
#F_NAME      MLP Function Fitting Demo
#F_GPUS      1
#F_TRACK_LOGS runs/**/metrics.csv

set -euo pipefail
SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
cd "$SCRIPT_DIR/dl_demo"

uv run python train.py --epochs 200 --log-interval 20

Directives

Directive	Example	Description
#F_NAME	#F_NAME My Experiment	Human-readable experiment name (defaults to filename)
#F_GPUS	#F_GPUS 2	Number of GPUs to request (default: 1)
#F_GPU_IDS	#F_GPU_IDS 0,1	Specific GPU indices to allocate (default: auto)
#F_TRACK_LOGS	#F_TRACK_LOGS runs/**/*.csv	Glob pattern for log files to watch

Lines starting with #F_ are parsed as metadata; everything else is the script body executed by bash -e.

Scheduling

• Jobs are queued in priority order (higher priority runs first)
• The scheduler polls every 3 seconds and launches one pending job per tick
• GPUs are allocated via CUDA_VISIBLE_DEVICES

Job Actions

From the web dashboard or API:

• Top — give a queued job highest priority
• Hold / Resume — pause or unpause a queued job
• Cancel — kill a running job
• Delete — remove a finished job and its logs

Web Dashboard

The dashboard at :8080 shows:

• GPU Status — per-GPU cards with VRAM usage, compute utilization, and power draw
• Job Queue — grouped by status (Running, Pending, Held, Finished) with priority controls
• Log Viewer — real-time terminal emulator via xterm.js with ANSI color support, syntax highlighting, line numbers, and horizontal scrolling

Log Viewer Features

• Full ANSI 256-color rendering (\x1b[...m escape codes)
• Carriage return (\r) support for progress bars (e.g., tqdm)
• Smart syntax highlighting for plain-text logs (log levels, timestamps, URLs, numbers, strings)
• Copy-safe line number gutter (physically isolated from terminal output)
• Horizontal scrolling for wide output
• Chunked rendering for large files via requestAnimationFrame

Environment Variables

Variable	Default	Description
FIT_FRONTEND_DIR	~/.fit/frontend/ (auto-extracted)	Override path to web UI build directory

Data Locations

Path	Purpose
~/.fit/	Base directory
~/.fit/fit.db	SQLite job database
~/.fit/logs/	Job log files
~/.fit/frontend/	Extracted web UI assets

API

Base: http://127.0.0.1:8080/api

Method	Endpoint	Description
GET	/system/status	GPU statuses
POST	/jobs	Submit a job
GET	/jobs	List all jobs
GET	/jobs/:id	Get job details
PATCH	/jobs/:id/priority	Update priority
POST	/jobs/:id/cancel	Cancel job
DELETE	/jobs/:id	Delete job
GET	/jobs/:id/logs	List log files
GET	/jobs/:id/logs/stream	SSE log stream
GET	/jobs/:id/logs/raw	Raw log download

Development

# Frontend dev server (proxies /api to :8080)
cd web-ui && npm run dev

# Run daemon with local frontend
FIT_FRONTEND_DIR=web-ui/dist fit daemon

# Lint & build
cd web-ui && npm run lint && npm run build
go build -o fit ./cmd/fit/

The frontend is embedded into the binary via //go:embed web-ui/dist (embed.go). Always rebuild the frontend before the Go binary.

Tech Stack

Layer	Technology
Runtime	Go 1.25
Web framework	Gin
Database	SQLite (pure Go via modernc.org/sqlite)
GPU monitoring	NVIDIA NVML (with graceful stub fallback)
Frontend	React 19, Vite 8, Tailwind CSS 4
Terminal	xterm.js 6
File watching	fsnotify