tmux-trainsh 1.2026.120

GPU training workflow automation with tmux

tmux-trainsh

tmux-trainsh is a terminal-first workflow runner for GPU and remote automation work.

README is the quick overview and landing page. The canonical command reference stays in the CLI:

train help
train --help

Those two commands are the canonical help entry points.

Install

Install from PyPI with uv:

uv tool install -U tmux-trainsh
train help

Or use the install script:

curl -LsSf https://raw.githubusercontent.com/binbinsh/tmux-trainsh/main/install.sh | bash

Main Command Groups

• train recipe for recipe files, execution, status, logs, jobs, and schedules
• train host for named SSH or Colab hosts
• train vllm for managed remote vLLM servers and local batch clients
• train storage for named storage backends
• train transfer for local, host, and storage copies
• train secrets for credentials
• train config for config and tmux settings
• train vast for Vast.ai instances
• train runpod for RunPod Pods
• train colab for one-off Colab tunnels
• train pricing for exchange rates and cost estimates

Quick Start

train secrets set GITHUB_TOKEN
train secrets set VAST_API_KEY
train secrets set RUNPOD_API_KEY
train secrets set POE_API_KEY
train host add
train vllm serve gpu-box Qwen/Qwen2.5-32B-Instruct --gpus 0
train storage add

train recipe show nanochat
train recipe new demo --template remote-train
train exec nanochat
train host clone gpu-box https://github.com/org/private-repo.git /srv/private-repo
train recipe status --last

Bundled Examples

Current bundled examples: aptup, brewup, hello, nanochat

Recipe Authoring

Public imports:

from trainsh import Recipe, Host, RunpodHost, VastHost, HostPath, Storage, StoragePath, load_python_recipe, local
from trainsh import flash_attn_install_script

Main authoring model:

• Recipe
• Host / VastHost / RunpodHost / HostPath
• Storage / StoragePath
• host.tmux(...)
• local.tmux(...)
• tmux.install_flash_attn(...)
• tmux.script(...)
• recipe.storage_wait_count(...)

Runtime Guarantees

• .pyrecipe recipes run as: load -> dependency graph from depends_on -> executor run
• Airflow-like retry / timeout / callback / trigger-rule semantics remain supported
• Supported executor aliases include sequential, thread_pool, process_pool, local, airflow, celery, dask, and debug
• Kubernetes executor remains unsupported

Maintenance

To refresh this file after editing the canonical help catalog:

python3 scripts/sync_cli_docs.py

Regression commands:

python3 tests/test_commands.py
python3 -m unittest tests.test_runtime_persist tests.test_pyrecipe_runtime tests.test_runtime_semantics tests.test_provider_dispatch tests.test_ti_dependencies
uv run --with coverage python scripts/check_runtime_coverage.py