3 projects
dynquant
Training-dynamics-driven mixed-precision LLM quantization, with prebuilt CUDA kernels.
dynquant-kernels
Compiled CUDA kernels for DynQuant: n-bit dequant, decode GEMV, tensor-core GEMM, MoE grouped GEMM.
dynquant-core
Training-dynamics-driven mixed-precision LLM quantization: signals, scoring, allocation, packing. Pure Python.
