Saved in:
| Main Authors: | , , , , , , |
|---|---|
| Format: | Preprint |
| Published: |
2025
|
| Subjects: | |
| Online Access: | https://arxiv.org/abs/2512.05409 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866909945840009216 |
|---|---|
| author | Huang, Ruixuan Zeng, Hao Huang, Hantao Shi, Jinyuan Yu, Minghui Yen, Ian En-Hsu Wang, Shuai |
| author_facet | Huang, Ruixuan Zeng, Hao Huang, Hantao Shi, Jinyuan Yu, Minghui Yen, Ian En-Hsu Wang, Shuai |
| contents | Post-training quantization (PTQ) plays a crucial role in the democratization of large language models (LLMs). However, existing low-bit quantization and sparsification techniques are difficult to balance accuracy and efficiency due to the limited hardware support. For example, W4A8 can only achieve the same peak TOPS as W8A8 whereas the GPU-supported sparse data format (2:4 semi-structure sparse) is seldomly adopted due to the loss of accuracy. To bridge this gap, in this paper, we propose the Sparse-Quantized Format (SQ-format), which is a unified data format for quantization and sparsification potentially easily supported by new hardware and existing GPUs. SQ-format makes use of the fact that sparse matrix can be accelerated in high-precision, and low-precision matrix multiplication can also be accelerated accordingly. As such, SQ-format is proposed to achieve Pareto improvement between performance and throughput. This format is particularly suitable for activations with outlier inequality status and makes their static compression possible. We show the state-of-the-art PTQ performance with SQ-format, propose the hardware required to support it, and further offer the design exploration and insights for the next-generation AI accelerators. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2512_05409 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | SQ-format: A Unified Sparse-Quantized Hardware-friendly Data Format for LLMs Huang, Ruixuan Zeng, Hao Huang, Hantao Shi, Jinyuan Yu, Minghui Yen, Ian En-Hsu Wang, Shuai Computation and Language Post-training quantization (PTQ) plays a crucial role in the democratization of large language models (LLMs). However, existing low-bit quantization and sparsification techniques are difficult to balance accuracy and efficiency due to the limited hardware support. For example, W4A8 can only achieve the same peak TOPS as W8A8 whereas the GPU-supported sparse data format (2:4 semi-structure sparse) is seldomly adopted due to the loss of accuracy. To bridge this gap, in this paper, we propose the Sparse-Quantized Format (SQ-format), which is a unified data format for quantization and sparsification potentially easily supported by new hardware and existing GPUs. SQ-format makes use of the fact that sparse matrix can be accelerated in high-precision, and low-precision matrix multiplication can also be accelerated accordingly. As such, SQ-format is proposed to achieve Pareto improvement between performance and throughput. This format is particularly suitable for activations with outlier inequality status and makes their static compression possible. We show the state-of-the-art PTQ performance with SQ-format, propose the hardware required to support it, and further offer the design exploration and insights for the next-generation AI accelerators. |
| title | SQ-format: A Unified Sparse-Quantized Hardware-friendly Data Format for LLMs |
| topic | Computation and Language |
| url | https://arxiv.org/abs/2512.05409 |