Gated DeltaNet 可承受 4-bit 量化：混合 27B 模型的 NVFP4 W4A4 方案
原标题：Paper page - Why Gated DeltaNet Survives 4-Bit Quantization: NVFP4 W4A4 for the Recurrent Half of a Hybrid 27B LLM
AI 摘要
Minima AI 团队发布论文，提出对混合大模型 Qwen3.8-27B 的全部 496 个线性层（包括 Gated DeltaNet 循环层）进行 NVFP4 W4A4 4-bit 量化，仅用 PTQ 校准，无需 QAT 或蒸馏。实验表明，该量化模型在 MMLU-Pro、GSM8K、AIME'25 等基准上与 BF16 基线性能相当，模型大小降至 17.5 GiB，prefill 速度提升 14-19%。研究揭示门控投影对量化不敏感、delta-rule 循环可自然遗忘噪声等机制，并修复了内核融合导致的缩放不匹配问题。
正文节选
Why Gated DeltaNet Survives 4-Bit Quantization: NVFP4 W4A4 for the Recurrent Half of a Hybrid 27B LLM Abstract Fully quantizing hybrid LLMs—including recurrent Gated DeltaNet layers—to 4-bit NVFP4 preserves accuracy across long-context and reasoning benchmarks by localizing outliers and exploiting robust delta-rule dynamics. Hybrid LLMs pair softmax attention with linear-attention layers such as Gated DeltaNet (GDN), whose recurrent state summarizes the context in fixed size. Early community 4-b