Exploring HiFloat8: A Tapered Format Complementing the FP8 Ecosystem for Robust Model T...
Talkat KubeCon + CloudNativeCon + OpenInfra Summit + PyTorch Conference China 2026 · 2026-09-07..09-09·
summary
Introduces HiFloat8 (HiF8), a tapered 8-bit floating format designed to handle neural network weight/gradient distributions and high-magnitude outliers with less aggressive scaling than standard FP8. Covers integration and evaluation of HiF8 in vllm and DeepSpeed on GPUs, reports training-stability and final-loss comparisons showing similar accuracy to FP16 with 1.5–1.7× GEMM performance, and discusses ongoing work toward dedicated hardware support.
AI-written summary · about
related talks
25mFp8 Training From Hopper To Blackwell
26mLow-Precision Inference without Quality Loss...
17mMiniFloat-NN: A RISC-V ISA Extension for Low-Precision NN Training
25mmxfp8, mxfp4, nvfp4 formats and applications in PyTorch
48mDemystifying Floating Point
14mOcto: INT8 Training with Loss-aware Compensation and Backward Quantization for Tiny
09mOpen models and open compression
25mThe Golden Triangle of Inference Optimization: Balancing Latency, Throughput, and Quality