krauncher-jupyter

%%krauncher — a Jupyter cell magic that runs a marked notebook cell as an ephemeral task on the cheapest suitable remote GPU, via the Krauncher broker. The notebook kernel stays local; only cells you mark go remote. Price is quoted before the cell runs.

%pip install krauncher-jupyter
%load_ext krauncher_magic

%%krauncher --pip torch
import torch
import torch.nn as nn

model = build_model().to("cuda")
losses = train(model, epochs, batch_size)   # epochs, batch_size taken from the notebook
accuracy = evaluate(model)

krauncher: auto --in epochs,batch_size --out model,losses,accuracy
krauncher: estimate — cv_training, ≥6 GB VRAM, 214.0 CU
Executing on worker-a1514da4: RTX 5060 Ti, 15GB, storage 2884 MB/s, net 800 Mbps
Epoch 1/3  avg_loss=2.3409          ← streamed live, with a GPU metric line
krauncher: done on RTX 5060 Ti in 131.2s — 3.42 KU → losses, accuracy

losses and accuracy are regular variables in the notebook afterwards, produced on a GPU picked for the lowest price for this specific workload. No flags are required: the transfer set is detected from the cell's code.

Why per-cell

Each marked cell is one ephemeral task: no GPU session, no idle time billed, no instance to choose. The broker classifies the cell's code (VRAM, workload class, compute units), quotes it, and dispatches it to the cheapest GPU that fits — across providers. Unmarked cells cost nothing.

No GPU state persists between cells (load-model-in-one-cell, use-in-the-next does not work — group related work into a single cell). Consecutive cells with the same requirements do reuse a warm worker when one is still around (session affinity), and an opt-in idle hold in your account settings keeps it around longer.

What the magic does for you

• Auto namespace — free variables of the cell that exist in the notebook are sent as inputs; assigned names come back as outputs. Values are JSON-safe data; models/tensors stay on the worker and are reported, not fatal. Credential-shaped values and values over 1 MB are held back unless passed explicitly with --in.
• Quote before run — the analysis phase prices the cell (VRAM, compute units, IO) before anything is submitted; --estimate stops there.
• Live feedback — remote stdout/stderr stream into the cell output as they happen, plus a GPU utilisation/VRAM/elapsed progress line.
• Hugging Face / S3 pre-fetch — literal load_dataset("org/name"), from_pretrained("org/name") and "s3://bucket/key" references are downloaded by the data bridge before the container starts and served to your unmodified code. Credentials come from your account defaults (Storage → Credentials).
• Non-blocking cells — --async submits and returns a handle for later cells (await kr_task / kr_task.result()).
• Restart the kernel = cancel — in-flight tasks are cancelled and their holds released when the kernel exits.

Options

Install & credentials

pip install krauncher-jupyter          # pulls the krauncher SDK

CAS_API_KEY (and optionally CAS_BROKER_URL) as environment variables or in a .env next to the notebook — same as the SDK. Get a key at krauncher.com → Account → API Keys.

Limitations

• Transferred values are JSON-safe data only (numbers, strings, lists, dicts) within a 16 MB inline budget; larger or binary data goes through data sources / volumes / the HF-S3 pre-fetch above.
• No GPU state across cells (see "Why per-cell").
• Dynamic Hub/S3 references (f-strings, variables) cannot be pre-fetched — they download inside execution instead of the pre-fetch phase (a warning says so).
• Top-level global in a cell is not supported (the body runs as a function).

Development

PYTHONPATH=.:../cas-client python -m pytest tests/

License

MIT © 2026 Ilya Sergeev