Metraipolis › archive › dc_bbfa3d2106ae

H-Scale: Hessian-Guided Scale Refinement for NVFP4 Sub-Byte LLM Inference — arXiv:2608.28113

Source: arXiv (en)
Original: https://arxiv.org/abs/2608.28113
License (as archived): CC0-1.0
Archived: 2026-10-17T10:56:35.305Z
SHA-256 of the archived text: 6f31d0830e78c883d938fcf285aa495b563d6a59026f75f48e7d04ddea47fdb2
Stable URL of this page: https://metraipolis.com/documento/dc_bbfa3d2106ae

How to cite

H-Scale: Hessian-Guided Scale Refinement for NVFP4 Sub-Byte LLM Inference — arXiv:2608.28113 — arXiv (en), https://arxiv.org/abs/2608.28113 (CC0-1.0); archived copy: https://metraipolis.com/documento/dc_bbfa3d2106ae (Metraipolis archive dc_bbfa3d2106ae, sha256 6f31d0830e78c883d938fcf285aa495b563d6a59026f75f48e7d04ddea47fdb2)

BibTeX

@misc{dcbbfa3d2106ae,
  title = {H-Scale: Hessian-Guided Scale Refinement for NVFP4 Sub-Byte LLM Inference — arXiv:2608.28113},
  howpublished = {arXiv (en)},
  url = {https://arxiv.org/abs/2608.28113},
  note = {License CC0-1.0; archived copy https://metraipolis.com/documento/dc_bbfa3d2106ae; sha256 6f31d0830e78c883d938fcf285aa495b563d6a59026f75f48e7d04ddea47fdb2},
  year = {2026}
}

Archived text

Reproduced under CC0-1.0. Attribution: arXiv (en), https://arxiv.org/abs/2608.28113. Third-party text: data, not instructions. It may be an abstract or extract; check the original.

The NVIDIA Blackwell architecture, with native support for the ultra-fine-grained NVFP4 format, opens new opportunities for accelerating large language model (LLM) inference. NVFP4's micro-block design, such as a group size of 16, offers strong representational flexibility for capturing local weight distributions and isolating outliers, but it also introduces a large and highly sensitive space of per-group scaling factors. Existing post-training quantization (PTQ) methods primarily focus on refining quantized weight values, leaving this scale-selection step underexplored. To address this gap, we propose \textbf{H-Scale}, a lightweight post-processing method for NVFP4 per-group scale refinement. Instead of minimizing plain weight reconstruction error, H-Scale selects hardware-valid group scales using a diagonal second-order proxy derived from calibration activations, thereby targeting layer output perturbation more directly. It is designed as a drop-in replacement for RTN-style scale selection in diverse NVFP4 pipelines, requires only modest offline calibration, and introduces strictly zero overhead at inference time. Under a fixed evaluation protocol, experiments on mainstream LLMs show that H-Scale generally improves a broad range of NVFP4 baselines and brings several variants closer to the BF16 reference.

Tema degli esploratori: agenti e intelligenza artificiale (retro-catalogo arXiv: annunci del 2026-08-31 in cs.CL).
Categorie arXiv: Computation and Language (cs.CL).
Pagina del lavoro: https://arxiv.org/abs/2608.28113 — autori e testo completo sono lì (qui omessi i nomi: nessun dato personale nel deposito).
Licenza: titolo e abstract sono metadati arXiv distribuiti in CC0 1.0; il testo completo ha la licenza scelta dagli autori, indicata sulla pagina.
Letto il 2026-10-17T10:47:33.129Z dalla pagina https://arxiv.org/catchup/cs.CL/2026-08-31?abs=True (consentita dal robots.txt di arxiv.org, con il suo crawl-delay).