Tinker SDK compatibility

Compute hosts SkyRL’s Tinker-compatible server behind a session-specific *.api.compute.cx gateway. The official tinker SDK stays on your machine. Compute provides the base URL and a short-lived session capability.

This surface is a private alpha. Compatibility tests pass in automation, but unchanged cookbook SFT and one-step RL have not completed the public H100 and MI300X canaries. Public session quotes remain closed until those tests pass.

Create the client

The primary path is to run your program under the Compute CLI:

compute tinker run -- python train.py

Inside train.py:

import os

import tinker

service = tinker.ServiceClient(
    base_url=os.environ["TINKER_BASE_URL"],
    api_key=os.environ["TINKER_API_KEY"],
)

training = service.create_lora_training_client(
    base_model=os.environ["TINKER_MODEL"],
    rank=32,
)

The official SDK can also read TINKER_BASE_URL and TINKER_API_KEY when you call tinker.ServiceClient() with no arguments.

Supported first cut

Surface	Support
Model	`Qwen/Qwen3-4B-Instruct-2507`, pinned revision
Training	Single-tenant FSDP LoRA
Losses	`cross_entropy`, `importance_sampling`
GPUs	RunPod Secure H100-SXM or Hot Aisle MI300X
SDK protocol	Tinker API under `/api/v1`, including protobuf requests used by current SDK clients
Futures	Submit through the SDK and poll `retrieve_future` as usual
Checkpoints	Save during training; Compute harvests files when the session stops

The gateway supports the training calls used for session and model creation, forward and backward passes, optimizer steps, sampling, weight save or load, and future retrieval. It forwards supported content types without translating the training payload. ppo, cispo, dro, forward_backward_custom, full-parameter training, multi-GPU sessions, and arbitrary Hugging Face model IDs are outside the first cut.

Authentication boundary

Your Compute login authorizes the control-plane operation that creates the session. The Tinker SDK uses a different, session-scoped key at the training gateway. The CLI keeps that key in the child environment. The gateway accepts the SDK’s X-API-Key header and bearer authentication, verifies the Compute-issued capability, and removes it before forwarding the request to the on-box server. Do not use your normal Compute API token as TINKER_API_KEY. Do not copy an on-box placeholder such as tml-dummy; it is not a customer credential.

Billing and lifecycle

A Tinker session is billed while its GPU machine exists, in started minutes plus the platform fee shown by the quote. It is not billed at hosted Tinker token prices. CLI-owned run and shell sessions stop when the child exits. User-owned start sessions continue until you stop them or a safety control fires. One account can have one live Tinker session.

Checkpoints

Training code saves through the normal Tinker SDK. During stop, Compute harvests available checkpoint files before terminating the provider machine.

compute tinker stop tns_...
compute tinker checkpoints tns_...

The list includes each file’s kind, byte size, SHA-256 hash, and creation time. Download the file from the session page in the dashboard. Checkpoint download is supported. Starting a new session from an old checkpoint is not part of the first-cut customer workflow.

Tinker quickstart

Run a cookbook recipe with one Compute login.

Tinker CLI

Session ownership, safety flags, and teardown behavior.

Overview

Get started

Guides

CLI

Reference

Tinker training

Tinker SDK compatibility

Create the client

Supported first cut

Authentication boundary

Billing and lifecycle

Checkpoints

Tinker quickstart

Tinker CLI

​Create the client

​Supported first cut

​Authentication boundary

​Billing and lifecycle

​Checkpoints

Tinker quickstart

Tinker CLI

Create the client

Supported first cut

Authentication boundary

Billing and lifecycle

Checkpoints