Bitdecoding fp4 boosts large language model memory efficiency and speed
BitDecoding-FP4: Native FP4 KV Cache Decoding via Algorithm-System Co-Design
Artificial IntelligenceHardware ArchitectureComputation and LanguagePerformance
Summary
Large language models use memory to store information, but this can be slow and costly. The authors found a way to shrink and speed up this memory storage by improving how data is encoded and decoded in a special 4-bit format. Their method keeps accuracy high while making the models run 2 to 3 times faster during use. This helps in managing the huge amount of data long conversations or texts require.
What this means in practice
- •For llm infrastructure teams: Increase serving throughput of long-context large language models by improving FP4 key-value cache decoding with minimal accuracy loss.
- •For cloud ai service operators: Deploy optimized FP4 decoding to deliver faster and more cost-efficient LLM inference on NVIDIA hardware platforms.$Commercial implications: Enables commercial AI services to offer cheaper, high-performance LLM hosting with improved memory efficiency and speed.
Authors
Dayou Du, Woon Yee Ng, Guangming Sheng, Nick Lau, Jianyi Cheng, Luo Mai
Abstract
The key-value (KV) cache dominates memory capacity and bandwidth in long-context LLM serving. While 4-bit microscaling formats (NVFP4, MXFP4) shrink the cache and run attention on FP4 Tensor Cores, a fundamental gap remains between algorithmic quantization and system-aware acceleration: existing FP4 KV systems use vanilla quantization that causes severe accuracy drops, whereas accuracy-oriented 4-bit algorithms rely on vector codebooks whose lookup overhead degrades decoding throughput. We present BitDecoding-FP4, an algorithm--system co-design for accurate and efficient native FP4 KV decoding. We mitigate three error sources: Dual-Residual Query Quantization splits queries into two FP4 tensors, cutting query logit error from 60% to 1%; Underflow-Aware Probability Scaling prevents E4M3 scales from rounding long-context probabilities to zero; and Outlier-Aware Key Quantization combines static smoothing, Hadamard rotation, and scale search to suppress key outliers under token-wise scaling. At the system level, we remove the softmax bottleneck via transposed softmax with fused probability quantization and stacked query MMA on paged KV caches. Across dense and MoE models with GQA and MLA, BitDecoding-FP4 stays within 0.5 points of BF16 KV accuracy with NVFP4 (1 point with MXFP4), and tracks BF16 KV in RL training rollouts. In end-to-end serving on NVIDIA B300 with Qwen3-4B and Qwen3-30B-A3B, it delivers 2.4-3.0$\times$ higher peak serving throughput than BF16 KV. The implementation is available at https://github.com/OpenBitSys/BitDecoding-FP4.