PulseAugur
EN
LIVE 07:23:20

HiFA4 enables 4-bit FlashAttention on Ascend NPUs for LLM inference

Researchers have developed HiFA4, a novel post-training design for executing FlashAttention operations in 4-bit on Ascend HIF4 NPUs, aiming to improve LLM inference efficiency. This approach combines two key mechanisms: Smooth-QK for rescaling attention weights and P-Reordering for accumulating softmax normalizers. Evaluations across five LLMs, including Qwen3-8B and Gemma2-9B, demonstrate that HiFA4 significantly reduces quantization-induced accuracy regressions and decision drift, with notable improvements in MMLU scores. AI

IMPACT This research could lead to more efficient LLM inference on specialized hardware by enabling lower-bit quantization without significant accuracy loss.

RANK_REASON The cluster describes a novel technical approach presented in an academic paper on arXiv.

Read on arXiv cs.CL →

AI-generated summary · Google Gemini · from 2 sources. How we write summaries →

HiFA4 enables 4-bit FlashAttention on Ascend NPUs for LLM inference

COVERAGE [2]

  1. arXiv cs.AI TIER_1 English(EN) · Hui Dong, Yanzhao Li, Jie Gao, Chunlu Li, Zhiyuan Zhang, Yupeng Sun, Zhenyuan Chen, Zhiqiang Zou ·

    HiFA4: Training-Free 4-bit FlashAttention on Ascend HIF4 NPUs for LLM Inference

    arXiv:2607.04302v1 Announce Type: cross Abstract: We present HiFA4, a post-training operator-level design that executes both QK^T and PV in FlashAttention as 4-bit HIF4 Cube GEMMs for LLM inference on Ascend NPUs, while maintaining the online softmax state in FP16. To our knowled…

  2. arXiv cs.CL TIER_1 English(EN) · Zhiqiang Zou ·

    HiFA4: Training-Free 4-bit FlashAttention on Ascend HIF4 NPUs for LLM Inference

    We present HiFA4, a post-training operator-level design that executes both QK^T and PV in FlashAttention as 4-bit HIF4 Cube GEMMs for LLM inference on Ascend NPUs, while maintaining the online softmax state in FP16. To our knowledge, HiFA4 is the first Ascend-HIF4-targeted design…