training_hub: Skill for Claude Code

Training Hub Guide

Training Hub is an abstraction layer for LLM post-training algorithms. It packages SFT, OSFT, and LoRA behind a unified interface so users do not need to learn multiple backend APIs. Backends are wired together internally; users interact with a single API surface.

For API reference and conceptual overviews, consult the live documentation at https://ai-innovation.team/training_hub/#/ and the docs/ directory in the repo root. This skill covers practical knowledge, decision frameworks, and troubleshooting that supplements the official docs.

Installation

Install targets

# Minimal (no backends, no GPU training)
uv pip install training_hub

# SFT + OSFT (high-scale distributed fine-tuning via CUDA backends)
# IMPORTANT: base install MUST come first, then [cuda] with --no-build-isolation
uv pip install training_hub && uv pip install training_hub[cuda] --no-build-isolation

# LoRA (budget-friendly, single/few-GPU via Unsloth — does NOT require [cuda])
uv pip install training_hub[lora]

The [cuda] extra is only needed for SFT and OSFT algorithms. LoRA uses the Unsloth backend which handles its own CUDA dependencies through [lora].

The two-step install for [cuda] is required because flash-attn and other CUDA packages need torch and packaging to already be present at build time.

Third-party loggers

Loggers are not bundled. Install separately as needed:

uv pip install wandb       # Weights & Biases
uv pip install mlflow      # MLflow
uv pip install tensorboard # TensorBoard

Fixing CUDA/kernel import errors

When users encounter errors like cannot import from flash_attn: unknown symbol or similar issues with optimized kernels (flash attention, liger, causal-conv1d, mamba-ssm), the root cause is usually stale cached builds. Fix with:

1. uv cache clean
2. Remove GPU-related caches from ~/.cache/ (torch, triton, flash_attn, vllm, and similar)
3. Remove ~/.triton/ if it exists (triton kernel cache)
4. Delete the current venv and recreate it fresh
5. Reinstall with the two-step process above