BitDecoding-FP4 Cuts KV Cache to FP4 With 2.4-3.0x Throughput Gains
BitDecoding-FP4: Native FP4 KV Cache Decoding via Algorithm-System Co-Design
The co-design removes the trade-off that previously forced a choice between quantization accuracy and decode speed: system-level changes like transposed softmax eliminate lookup overheads that plagued earlier FP4 KV methods. Teams running long-context inference on NVIDIA Blackwell hardware can apply it directly, since the implementation is open-sourced and works across dense and MoE models with GQA and MLA, and it even holds up in RL training rollouts.
BitDecoding-FP4 is an algorithm-system co-design for native FP4 KV cache decoding in long-context LLM serving, released on arXiv (2611.09825) with open code. It addresses three error sources: Dual-Residual Query Quantization cuts query logit error from 60% to 1%, Underflow-Aware Probability Scaling keeps E4M3 scales from zeroing probabilities, and Outlier-Aware Key Quantization suppresses key outliers. On NVIDIA B300 serving Qwen3-4B and Qwen3-30B-A3B, it delivers 2.4-3.0x higher peak serving throughput than BF16 KV while staying within 0.5 points of BF16 accuracy under NVFP4 (1 point under MXFP4).
BitDecoding-FP4: Native FP4 KV Cache Decoding via Algorithm-System Co-Design
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.