Hugging Face Papers ↗ · 2026-11-06
BitDecoding-FP4: Native FP4 KV Cache Decoding via Algorithm-System Co-Design
Media reports Hugging Face Papers Source language: Chinese
Excerpt
论文摘要(作者报告,未经本站复测):The key-value (KV) cache dominates memory capacity and bandwidth in long-context LLM serving. While 4-bit microscaling formats (NVFP4, MXFP4) shrink the cache and run attention on FP4 Tensor Cores, a fundamental gap remains between algorithmic quantization and system-aware acceleration: existing FP4 KV systems use vanilla quantization that causes severe accuracy drops, whereas…
Short excerpt from the source — open the original for the full text.