pytrio

What problem does it solve?

Claude Code and other assistants generate invalid PyTRIO usage because they do not know the remote training and inference flow, credential requirements, or naming differences from HuggingFace APIs, so the same mistakes keep recurring.

Core Features & Use Cases

• Credential and concept checklist: Covers trio login, ServiceClient creation, PyTorch-to-PyTRIO concept mapping, and how loss functions, ModelInput wrapping, and AdamParams differ from local training.
• API quick references: Details the training, sampling, and rest clients plus Datum/SamplingParams structures so you can construct correct pipelines, save checkpoints, and resume or sample from LoRA weights.
• Trap list and best practices: Explains issues such as max_tokens vs max_new_tokens, weights masking, prompt autoshift rules, and the requirement to use Qwen/Qwen3-4B-Instruct-2507 for training, with runnable examples stored under examples/ and best-practices/.

Quick Start

Ask Claude Code to plan how to log in via trio login, wrap training samples with ModelInput, call forward_backward and optim_step, and save sampler weights for inference.