Humming 0.1.18 extends SGLang MoE quantization
Humming, a lightweight JIT-compiled kernel library for quantized inference, has extended its integration with SGLang to support additional quantization schemes and GPU architectures like Hopper and Blackwell. On DeepSeek-V4.1-Flash-1017, the integration delivers significant prefill speedups using mixed-precision and FP8 activations while preserving framework routing and distributed execution. Open article →