Zengxiao He

Tessera — From-Scratch LLM Distillation

A from-scratch LLM systems stack built around teacher–student distillation and efficient serving — paged-KV continuous batching, speculative decoding, int8/AWQ/FP8 quantization, custom Triton/CUDA kernels, and a Rust (tokio/axum) gateway. 500+ stars on GitHub.

Triton LLM Inference Kernel Lab

A compact GPU kernel lab for LLM inference primitives in Triton and CUDA: row-wise softmax, tiled FP16 GEMM with FP32 accumulation, and a FlashAttention-style fused attention kernel — each validated against a PyTorch reference. Measured on H100: fused softmax at 2.8 TB/s, ~83% of HBM3 peak bandwidth.

How a one-program-per-row Triton softmax gets within 17% of the memory roofline on H100 — the bandwidth math, the measured numbers, and the shapes where it honestly loses to torch.softmax.

I rebuilt vLLM's two core ideas — a block-paged KV cache and iteration-level scheduling — in a from-scratch engine. Here's the design logic, the ownership contract that prevents KV-block leaks, and the preemption policy that fell out of it.