FastAFD

Hao AI Lab · LLM Inference

Ongoing contribution

Currently contributing to a new LLM inference engine that targets high-throughput workloads through Attention–FFN disaggregation.

AutoPallas & TPU Inference

ByteDance Seed Infra · Student Researcher Intern

Internal project

As code owner of AutoPallas, I developed an internal TPU kernel optimization agent with interactive task specification, configurable single-/multi-agent loops, parallel exploration, a strict correctness gate, and reusable optimization trajectories.

I also contributed to Seed’s internal vllm tpu-inference framework. My kernel work included optimizing Pallas paged attention through DMA pipelining and compute scheduling, reducing latency from 184.81μs to 117.16μs at 128K context. I used HLO/LLO analysis and XProf profiling, with XLA as a performance baseline.

Hierarchical Sparse Attention Done Right

Tencent Hunyuan Team · LLM Serving & Sparse Attention (Core Contributor)

arXiv Preprint, Jul 2026

Built the official SGLang serving backend and reference inference system for learned chunk-wise sparse attention. Enables 512K-token serving and 64× context-length extrapolation, with up to 13.5× prefill and 15.7× decode speedup over dense attention at 512K context.

Flash Multi-Head Feed-Forward Network

Minshen Zhang*, Xiang Hu*, Jianguo Li, Wei Wu, Kewei Tu

arXiv Preprint, 2025

Software–hardware co-design for an efficient Transformer FFN architecture. Wrote Flash-style kernels in ThunderKittens/CUDA and Triton that keep intermediates in SRAM. FlashMHF reduces peak memory by 3–5× and achieves up to 1.08× inference speedup over the SwiGLU baseline while improving model quality.

FastVideo

Hao AI Lab · Training & Inference Infrastructure · One of the Code Owners

Open-Source Project, Oct 2025 - Present

Training and inference infrastructure for video generation and world models. Contributed custom GPU kernels, inference optimization, and quantization-aware distillation, and helped develop the Blackwell-optimized DreamVerse runtime.

DreamVerse: Real-Time Video Generation

Hao AI Lab (Core Contributor)

Open-Source Release, May 2026

A real-time video generation workspace for "vibe directing" — steering generation through natural-language iteration instead of one-shot prompting. Built on the open-weights LTX-2 model with a FastVideo backend runtime and a Blackwell-optimized pipeline (NVFP4 inference, FA4, torch.compile), streaming 30s 1080p clips with under 5s wait on a single NVIDIA B200. As a core contributor, I built session-based audio–visual continuity for seamless multi-segment generation and accelerated backend inference through kernel benchmarking and fusion.

Enhancing 3D Character Generation with ControlNet and LoRA

Congrong Xu, Zhanhe Shi, Minshen Zhang, Qingcheng Zhao

EECS 182/282A | Deep Neural Networks, UC Berkeley, 2023

A project exploring enhanced 3D character generation techniques using ControlNet and LoRA for improved control and quality in generative models.

CUDA/C++ Parallel Image Rendering

Minshen Zhang

Personal Project, 2023

Built a C++ path tracer supporting Lambertian, metal, dielectric, and emissive materials. Implemented motion blur, depth of field, and volumetric effects. Accelerated rendering via CUDA parallelization and importance sampling, achieving ~200× speedup vs. single-threaded CPU baseline.