{
  "$type": "site.standard.document",
  "bskyPostRef": {
    "cid": "bafyreihyeons3vfkcqp2wyrfbovimxl2gcqlz26g5tuls4s5wkeu7lg5ny",
    "uri": "at://did:plc:pgryn3ephfd2xgft23qokfzt/app.bsky.feed.post/3mhimp7vgjct2"
  },
  "path": "/t/fastlora-v4-2-fine-tuning-library-that-never-crashes-pip-installable/174423#post_1",
  "publishedAt": "2026-03-20T08:41:11.000Z",
  "site": "https://discuss.huggingface.co",
  "textContent": "Hi everyone! I built **FastLoRA** , a drop-in alternative for fine-tuning LLMs.\n\nbash\n\n\n    pip install \"fastlora[full]\"\n\n\n**Why I built it:**\n\n  * Unsloth kept breaking on install and crashing mid-training\n\n  * Wanted something that just works, no matter what\n\n\n\n\n**What it does:**\n\n  * Never crashes — every error is caught, reported and recovered automatically\n\n  * Auto hardware detection — scans your GPU and applies best settings\n\n  * Unlimited model size — 1B to 1T+, automatic strategy selection\n\n  * Every feature is a `True`/`False` toggle + 0.0–1.0 power control\n\n  * Compiled kernel cache — 3min compile once, 5 seconds after\n\n\n\n\n**Quick start:**\n\npython\n\n\n    from fastlora import FastLoRA\n\n    fl = FastLoRA(\"meta-llama/Llama-3.2-3B\", lora=True, quantization=\"4bit\")\n    model, tokenizer = fl.load()\n\n\nBenchmarked on Tesla T4. Also benchmarked Unsloth on the same setup — Unsloth didn’t run.\n\nWould love feedback from the communit",
  "title": "FastLoRA v4.2 — Fine-tuning library that never crashes (pip installable)"
}