Second-year CS undergraduate, Turing Class, Zhejiang University.
I work on model inference and post-training.
|
FlashML-org/FreeToken — serving engine for frontier MoE models on desktop hardware Contributed to the block-FP8 CPU MoE expert GEMV path: found and fixed a silent-correctness bug in the quantization scale-bank stride handling, added an AVX-512F kernel tier for CPUs without the BF16 extension, and built a float64 parity harness for the fp8 decode. |
Most other repositories here are tools I wrote for my own use.
