BitDecoding-FP4: Native FP4 KV Cache Decoding via Algorithm-System Co-Design

BitDecoding-FP4, an algorithm-system co-design for native FP4 KV cache decoding released on GitHub by OpenBitSys, stays within 0.5 points of BF16 KV accuracy with NVFP4 (1 point with MXFP4) across dense and MoE models with GQA and MLA, and delivers 2.4-3.0× higher peak serving throughput than BF16 KV on NVIDIA B300 with Qwen3-4B and Qwen3-30B-A3B. The method cuts query logit error from 60% to 1% via Dual-Residual Query Quantization, prevents E4M3 scales from rounding long-context probabilities to zero with Underflow-Aware Probability Scaling, and suppresses key outliers with Outlier-Aware Key Quantization combining static smoothing, Hadamard rotation and scale search. At the system level it removes the softmax bottleneck using transposed softmax with fused probability quantization and stacked query MMA on paged KV caches, and tracks BF16 KV in RL training rollouts.

arXiv:2611.09825v1 Announce Type: cross Abstract: 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 accuracy-oriented 4-bit algorithms rely on vector codebooks whose lookup overhead degrades decoding throughput. We present BitDecoding-FP4, an algorithm--system co-design for accurate and efficient native FP4 KV decoding. We mitigate three error sources: Dual-Residual Query Quantization splits queries into two FP4 tensors, cutting query logit error from 60% to 1%; Underflow-Aware Probability Scaling prevents E4M3 scales from rounding long-context probabilities to zero; and Outlier-Aware Key Quantization combines static smoothing, Hadamard rotation, and scale search to suppress key outliers under token-wise scaling. At the system level, we remove the softmax bottleneck via transposed softmax with fused probability quantization and stacked query MMA on paged KV caches. Across dense and MoE models with GQA and MLA, BitDecoding-FP4 stays within 0.5 points of BF16 KV accuracy with NVFP4 1 point with MXFP4 , and tracks BF16 KV in RL training rollouts. In end-to-end serving on NVIDIA B300 with Qwen3-4B and Qwen3-30B-A3B, it delivers 2.4-3.0$\times$ higher peak serving throughput than BF16 KV. The implementation is available at https://github.com/OpenBitSys/BitDecoding-FP4.