Liquid AI
May 2026 — now
ML Engineer Intern, GPU Inference
Model bring-up and performance work across SGLang, vLLM, TensorRT-LLM, llama.cpp, and Transformers.
Vincent Zhong
I work on GPU inference.
I'm studying Computer Engineering at the University of Waterloo and currently work at Liquid AI. Most days involve model bring-up, profiling, kernels, and serving systems that are slightly stranger than a standard Transformer.
I also spend a fair amount of time in SGLang. I care about low-precision numerics, decoding algorithms, cache design, and the point where a model architecture meets an actual GPU.
Work experience
Shopify
Sep — Dec 2025
ML Engineer Intern, Search
Typo correction and query rewriting for Shop.app's retrieval engine, including SFT/RL experiments and large-scale relevance evaluation.
University of Waterloo · Data Systems Group
Nov 2024 — May 2025
Systems Researcher
Dense and sparse retrieval in Anserini; co-authored a resource paper published at SIGIR 2025.
Yupp AI
Feb — May 2025
Software Engineer Intern
Built LLM tool-use infrastructure and pairwise preference-data systems.
Some things
A disorganized list of some things that were useful to me at some point.
SGLang
GLM-5.x and DeepSeek-V3.2 fused indexer kernels and MXFP4 index-K caching; Kimi K2.x linear-attention, MLA, and radix-cache optimization; Qwen3-Next and Qwen3.5 Spec V2; and ReplaySSM for GDN speculative verification with Yuan Luo.
DeepSeek-V4 dataflow
mHC hyper-connections, HCA, compressed sparse attention with the C4 indexer, and a 384-expert MoE.
decontaminate
Python bindings and packaging for AI2's Rust train–test contamination detector.
GDN-3 ablations
Controlled preconditioner ablations on Gated DeltaNet-2 from 200M to 1.3B parameters, with multiple seeds and positive controls.
DeepSeek-V3.2 dataflow
Absorbed-MLA decode, the lightning indexer, grouped MoE routing, and multi-token prediction, following the reference model.py.
Anserini retrieval
Dense and sparse retrieval support in Anserini, ending in a resource paper at SIGIR 2025.
GLM-5.2 × SGLang
An NVFP4 model and its serving path: TP4, modelopt_fp4, EAGLE 5-1-6, chunked prefill, and full/shared DSA indexer state.
cs336-a1-sol
A speed-of-light study of the Stanford CS336 A1 OpenWebText leaderboard task.
Small kernel comparisons
Fused SGLang elementwise kernels across Triton, Inductor, and CuTe DSL, plus gated short-convolution implementations.
Model integrations
Text and multimodal serving work for Molmo 2, LFM2.5-VL, and NVIDIA Nemotron Parse.