← All notes

Summary

GPTQ is a one-shot post-training weight quantization method that uses approximate second-order information. The authors report that it can quantize GPT models with 175 billion parameters in approximately four GPU hours, reducing bitwidth to 3 or 4 bits per weight with negligible accuracy degradation relative to the uncompressed baseline. They also report end-to-end inference speedups over FP16 of around 3.25x on NVIDIA A100 GPUs and 4.5x on NVIDIA A6000 GPUs. [S1]

Claims

• id: C1 claim: 'GPTQ is a one-shot weight quantization method based on approximate second-order information.' source: S1 evidence: "a new one-shot weight quantization method based on approximate second-order information"
• id: C2 claim: 'They report quantizing 175 billion parameter models in approximately four GPU hours to 3 or 4 bits per weight.' source: S1 evidence: "reducing the bitwidth down to 3 or 4 bits per weight"
• id: C3 claim: 'They report about 3.25x speedup on A100 and 4.5x on A6000 versus FP16.' source: S1 evidence: "around 3.25x when using high-end GPUs (NVIDIA A100)"

Verification

• Code available: Paper: https://github.com/IST-DASLab/gptq, which returned HTTP 200 on 2026-11-11. Not run.
• Reproduced: not reproduced (2026-11-11). Code not run.
• Related work: Live note 2305-14314-qlora. This is the paper-verification note for GPTQ itself.
• Limits: Accuracy and speedup figures are the authors' reported results on the models and hardware in the v2 paper (revised 22 Mar 2023). Extreme 2-bit or ternary settings are described as providing only reasonable accuracy, not as matching the baseline.
• License: CC BY 4.0 on the abs page (2026-11-11; icon by-4.0.png; http://creativecommons.org/licenses/by/4.0/). Submitted 31 Oct 2022 (v1); last revised 22 Mar 2023 (v2). Sources: [S1] https://arxiv.org/abs/2210.17323, v2 (22 Mar 2023). checked_on: 2026-11-11.

When to apply

Use when compressing a trained generative Transformer for inference, as distinct from quantization-aware training methods such as QLoRA.

Facts

• id: F1 claim: 'GPTQ is a one-shot post-training quantization method using approximate second-order information.' confidence: medium date: 2026-11-11 sources: [S1] license: CC BY 4.0
• id: F2 claim: 'The authors report quantizing 175 billion parameter models to 3 or 4 bits per weight in about four GPU hours.' confidence: medium date: 2026-11-11 sources: [S1] license: CC BY 4.0
• id: F3 claim: 'They report inference speedups over FP16 of about 3.25x on A100 and 4.5x on A6000 GPUs.' confidence: medium date: 2026-11-11 sources: [S1] license: CC BY 4.0

Sources

• [S1] arXiv 2210.17323, https://arxiv.org/abs/2210.17323. License: CC BY 4.0. v2 (22 Mar 2023). Accessed 2026-11-11.

License: this note is CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Summary by iinventedlight (AI-assisted). Paper is CC BY 4.0.

@misc{iil_note_nzmznexg,
  author = {{'iinventedlight (AI-assisted)'}},
  title = {{GPTQ: Accurate Post-Training Quantization for Generative Pre-trained Transformers}},
  howpublished = {iinventedlight Library},
  year = {2026},
  date = {2026-11-11},
  url = {https://iinventedlight.com/library/notes/2210-17323-gptq/},
  note = {Permanent ID iil:note:nzmznexg. License: CC-BY-4.0.},
}

Markdown

Spotted a mistake? Use the MCP feedback tool or email hello@iinventedlight.com.

For language models and agents: start at /llms.txt | /agents.md | /.well-known/agent-library.json. This page in Markdown: /library/notes/2210-17323-gptq.md.