# AnswerDotAI/fsdp_qlora

Training LLMs with QLoRA + FSDP

Repository: https://github.com/AnswerDotAI/fsdp_qlora
Canonical: https://ross.abutalabs.com/products/fsdp_qlora
Language: Jupyter Notebook
License: Apache-2.0
License Family: permissive
Last push: 2024-11-09T00:13:14+00:00

## Health v2 (maintenance only)
Score: 26/100 (v2, computed 2026-09-03T02:20:16.233290+00:00)
- activity 0, release rhythm 35, longevity 68
- inputs: {"age_days": 963, "days_push": 663, "days_rel": null, "gap_med": null, "n_releases_24m": 0}
- flags: no_releases
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 1550, forks 201 (observed 2026-08-28T04:05:02.086094+00:00)

## What it is
A training script/library from Answer.AI that combines QLoRA (quantized LoRA) with PyTorch FSDP to fine-tune large language models like Llama-2 70B on limited GPU hardware. It supports bitsandbytes and HQQ quantization, gradient checkpointing, and CPU offloading.

## Use cases
- fine-tune llama 2 70b on two 24gb gpus
- train llms with qlora and fsdp
- quantized lora finetuning of large language models
- low-memory finetuning of llms with cpu offload
- finetune llama on alpaca dataset with limited vram

## When to choose
- you need to fine-tune very large LLMs on consumer or limited GPU hardware
- you want to combine QLoRA quantization with FSDP distributed training
- you're comfortable with alpha-stage tooling and debugging training runs

## When to avoid
- you need a production-stable, fully supported training framework
- you prefer higher-level fine-tuning tools like Axolotl or Hugging Face TRL
- you don't have CUDA GPUs or sufficient CPU RAM for offloading

## Facets
- artifact type: library
- maturity: experimental
- function: llm-training, machine-learning, gpu-computing
- domain: large-language-models, machine-learning, deep-learning
- platform: python
- tags: qlora, fsdp, fine-tuning, llama, quantization, bitsandbytes, hqq, distributed-training, gpu, linux

## Member repositories
- AnswerDotAI/fsdp_qlora (main) score 26

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:05:02.086094+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-30T04:30:18.504936+00:00, confidence not recorded.
  - readme: https://github.com/AnswerDotAI/fsdp_qlora (fetched 2026-08-28T04:05:02.086094+00:00, sha b8ca5e8ece45)
- Data as of 2026-08-30T08:39:29.467469+00:00.