Machine learning · Systems · Experiments
Suqin Zhang
I study efficient Transformer training, inference, and model evaluation. My work asks where computation is useful, where it is redundant, and how to tell the difference with an experiment.
Research archive
Retrospectives organized by project period2026
What to verify before pruning SwiGLU channels
A conservative protocol for SwiGLU pruning: inspect the right tensors, avoid underflow, and separate numerical agreement from task quality.
CUDA Graphs, batching, and end-to-end throughput
What a TensorRT experiment taught me about launch overhead, batch formation, and the boundary of a benchmark.
Correctness before scaling in JAX data parallel training
Data parallelism needs a correctness argument before a speedup claim. Start with the objective, the state, and the timing boundary.
Muon and AdamW under a fixed compute budget
Muon versus AdamW is also a systems question. The useful comparison is quality under a budget, not a utilization number in isolation.
2025
Exploring Power Creep in Magic: The Gathering
How I framed and led a team study combining numerical metrics, simulations, LLM annotations, and tournament data—and what each could establish.
2024
FusionNet: evaluating composed language models
A retrospective on nine Hugging Face releases: capability tradeoffs, benchmark provenance, and the controls a model-composition study needs.
Selected research
I contribute to KataGo, an open-source Go engine whose neural networks turn board positions into move probabilities and win-rate estimates. It is a useful setting for studying Transformer efficiency with a concrete downstream task.
My upstream work includes TensorRT performance and reliability changes, acknowledged in official releases. I also build and profile a JAX training path for TPU accelerators.
My earlier work includes FusionNet language-model releases and a team study of power creep in Magic: The Gathering. I led the latter project and am a co-first author, responsible for research framing, experimental design, task allocation, and evidence interpretation.
A little about me
I graduated early with a B.S. in Computer Science from the University of Washington’s Paul G. Allen School in March 2026.
My algorithmic background includes USACO Platinum, reached with a perfect score in the January 2024 Gold contest. Official result.
I’m interested in doctoral research on efficient learning and inference. These notes are a place to explain a result carefully, preserve a useful negative result, and identify the experiment that should come next.