Native-Windows operator scripts for a managed Unsloth Studio installation rooted at D:\AI\Unsloth, with models stored under D:\AI\Models.
The repository intentionally keeps the Windows environment isolated: scripts/env.ps1 builds a process-local environment for the managed runtime and does not persistently modify the user or machine PATH.
start.ps1— starts the managed Unsloth Studio onhttp://localhost:8888.doctor.ps1— reports package/GPU/runtime/database/credential/isolation health.update.ps1— stops on an active Studio instance, snapshots important state, runs the official Studio updater, then validates CUDA and database state.model-sync.ps1— safely synchronizes Hugging Face model content into the managed model tree. It is dry-run by default;-Applyis required to download.scripts/env.ps1— canonical process-local paths, cache locations, Python/uv settings, and CUDA backend safety rails.
The current scripts target the managed layout below:
D:\AI\Unsloth
D:\AI\Models
They expect a previously provisioned Unsloth Studio runtime. This repository does not currently contain a generic installer and CI does not pretend to provide a CUDA/GPU acceptance environment.
Start Studio:
.\start.ps1Inspect the installation:
.\doctor.ps1Update a stopped Studio installation:
.\update.ps1Preview a filtered Hugging Face model sync:
.\model-sync.ps1 owner/model -Include '*.safetensors','*.json'Apply the same sync only after reviewing the dry run:
.\model-sync.ps1 owner/model -Include '*.safetensors','*.json' -Applymodel-sync.ps1obtains the saved Hugging Face token through Studio's credential storage rather than exporting the token into the process environment.- Temporary sync parameters are removed in a
finallyblock. update.ps1snapshots Studio/auth databases before updating and validates CUDA/database state afterwards.- Explicit CUDA/backend settings are treated as safety rails; do not silently replace them with source-build fallbacks.
- Runtime files, models, caches, credentials, and databases do not belong in this repository.
Public pull-request CI performs source-level validation only: PowerShell parsing, static analysis, and GitHub Actions security analysis. GPU/CUDA/runtime acceptance belongs in a trusted Windows environment with the managed installation present; public PR code must not be executed on a home/self-hosted GPU runner.
See LICENSE.