Manage ML workflows via HTTP APIs, CLI and SDKs.
npx skills add https://github.com/thinking-machines-lab/tinker-cookbook --skill tinker-sdk
Please help me install this Agent Skill. Skill: tinker-sdk Source: https://github.com/thinking-machines-lab/tinker-cookbook/tree/main/.claude/skills/tinker-sdk Command: npx skills add https://github.com/thinking-machines-lab/tinker-cookbook --skill tinker-sdk
The Tinker SDK provides a unified, Python-based interface to manage training, sampling, and REST interactions for fine-tuning LLMs, reducing boilerplate and integration effort.
Instantiate a ServiceClient and create a TrainingClient to start a fine-tuning workflow on your base model.
High-intent search queries and answers about installing and using this skill.
To set up an async LLM fine-tuning pipeline, instantiate a ServiceClient, then create a TrainingClient to manage workflows. The SDK uses futures and awaitable async patterns to overlap computation and IO for scalable experimentation.
To generate samples from a fine-tuned model, create a SamplingClient via the Tinker SDK. This client manages sampling workflows asynchronously, allowing you to retrieve generated text outputs efficiently after training.
Yes, you can resume training runs from checkpoints. The TrainingClient within the Tinker SDK supports resuming workflows, allowing you to continue fine-tuning your base model from a previously saved state without restarting.
Yes, the Tinker SDK supports REST interactions for managing deployments through the RestClient. It provides a unified Python interface to handle REST API calls, reducing boilerplate when integrating deployment management into your workflows.
Async patterns for LLM fine-tuning and sampling are used to overlap computation and IO operations. This async-friendly approach enables scalable experimentation by preventing resources from blocking during long-running training or sampling tasks.
An end-to-end LLM fine-tuning workflow requires ServiceClient, TrainingClient, SamplingClient, and RestClient. These unified API clients cover starting training runs, resuming checkpoints, generating samples, and managing deployments.
0xFANGOSet up videocut dependencies and download models for offline operation.
NVIDIA-NeMoOptimize deployed NeMo agents with routing splits, hyperparameter tuning, and model swaps.
Sanali209Implement asynchronous Python patterns with asyncio, gather, and queues.
SwanHubXDocument PyTRIO SDK usage for remote training and inference with Qwen models.
