Ecosystem
Tuneline Studio builds on open-source libraries and connects to services you may already use; this page lists them
and what Tuneline uses each one for.
Version ranges are the ones Tuneline 0.1.0 requires (from pyproject.toml, requirements.txt and
web/package.json). Where a tested version is given, that is the version the runner's tests ran with.
Training
| Project |
What Tuneline uses it for |
Version |
| PyTorch |
The deep-learning framework under everything. The installers add the CUDA 12.8 build on NVIDIA machines and the CPU build otherwise. |
Any recent release; cloud job images use 2.8 |
| Transformers |
Loads base models and tokenizers, applies chat templates and generates answers during evaluation. |
5.x (>=5,<6), tested with 5.17 |
| TRL |
The trainers behind the training blocks: SFT, DPO / ORPO / SimPO / KTO, GRPO and vision SFT. |
1.x (>=1,<2), tested with 1.14 |
| PEFT |
LoRA and QLoRA adapters, starting from an existing adapter, and merging an adapter into its base model. |
>=0.18, tested with 0.21 |
| Datasets |
Loads Hugging Face Hub datasets (with streaming) and holds data between steps. |
>=4.7, tested with 5.0.1 |
| Accelerate |
Device placement, including spreading a model's layers over several GPUs. |
>=1.4, tested with 1.15 |
| bitsandbytes |
4-bit quantization for QLoRA and 8-bit optimizers. Not installed on macOS; on Windows, most reliable under WSL2. |
>=0.46 |
| Unsloth (optional) |
Faster LoRA and QLoRA training when a training step turns it on. Current releases need an older TRL, so Tuneline falls back to TRL for now. |
Installed on first use |
Evaluation
| Project |
What Tuneline uses it for |
Version |
| lm-evaluation-harness |
The Benchmarks (lm-eval) block, for standard benchmarks such as IFEval. Installed on first use. |
>=0.4.8 |
Golden-set evaluation (Evaluate on test set), the metrics (exact match, ROUGE-L, label accuracy and recall,
valid JSON and others) and the Quality gate are part of Tuneline itself.
Export and serving
| Project |
What Tuneline uses it for |
Version |
| llama.cpp |
The Export GGUF block: converts models to GGUF and quantizes them (for example Q4_K_M, Q8_0). The runner downloads and builds it on first use. The gguf Python package is also used. |
gguf>=0.10 |
| Ollama |
The Ollama package block writes an Ollama Modelfile for your GGUF export. On your own computer with Ollama installed, it also runs ollama create so the model is ready to chat. |
Your installed version |
| vLLM |
Serves OpenAI-compatible endpoints on RunPod Serverless, Hugging Face Inference Endpoints and Modal. Also an optional faster backend for evaluation, which currently falls back to Transformers in jobs. |
Set by each endpoint image |
Compute
| Service |
What Tuneline uses it for |
| Kaggle |
Free notebooks with two T4 GPUs: Tuneline starts a private GPU notebook with your Kaggle token. |
| Google Colab, Lightning AI and any Jupyter server |
The Notebook target: download a notebook for a run and run it there. |
| Modal |
Serverless GPU jobs billed per second, and Modal endpoints. |
| RunPod |
GPU pods billed per second (removed when the run ends), and RunPod Serverless endpoints. |
| Hugging Face Jobs |
Runs on Hugging Face hardware, billed to your Hugging Face account. |
| Your own GPU |
The My computer target: a worker on your machine picks up runs. See Your computer. |
| Vast.ai |
Live GPU prices on the Hardware page (no runs). |
Hugging Face Hub
| Service |
What Tuneline uses it for |
| Hugging Face Hub |
Base models and datasets, the model catalog's live search, and your private repositories for adapters, merged models, GGUF files and cached steps. Uses huggingface_hub 1.x (>=1.5,<2). |
| Inference Endpoints |
One of the three places to deploy a model, with scale to zero after an idle timeout. |
| Inference Providers |
Lets the Playground chat with Hub models without deploying them, and powers the LLM judge and synthetic data blocks when you choose Hugging Face. |
Tracking
| Service |
What Tuneline uses it for |
| Weights & Biases |
Optional. Mirrors training charts to your W&B account when a training step's Also log to setting is Weights & Biases and you saved a W&B key. Tuneline's own Metrics tab works without it. |
These run the Tuneline Studio app itself. You only deal with them if you host your own copy.
| Project |
What Tuneline uses it for |
Version |
| Supabase |
Sign-in, the Postgres database with row level security, Vault for encrypted keys, Storage for uploaded datasets, Realtime for live run updates, and pg_cron for background checks. |
@supabase/supabase-js 2.x |
| Vercel |
Hosts the web app and the Python API (a FastAPI function). |
– |
| FastAPI |
The API that launches runs, tests keys, estimates hardware, searches the Hub and manages deployments. |
0.141 |
| React |
The web app. Built with Vite, TypeScript and Tailwind CSS. |
19 |
| React Flow |
The pipeline canvas (@xyflow/react). |
12 |
| TanStack Query and Recharts |
Data loading in the web app, and the charts on the run and hardware pages. |
5 and 3 |