Qihang Wu

AI systems / model efficiency / technical notes

Qihang Wu

Training and inference framework engineer focused on LLM and VLM systems.

Writing about attention, memory, multimodal pipelines, distributed training, and practical deployment constraints.

Focus LLM / VLM systems
Research NeurIPS 2025 · IEEE TIP 2025
Archive 13 technical notes
01

Selected work

Ongoing Model efficiency studies Implementation notes on attention, multimodal internals, JAX / TPU practice, and long-context systems. NeurIPS 2025 CroPe Cross-modal semantic compensation for adverse-scene understanding. IEEE TIP 2025 UCDS Cross-domain collaboration for semantic segmentation.
02

Recent notes

Pallas TPU Kernel 写法与优化文档 Markdown note JAX 到 VLIW，以及 Pallas / Splash Attention 复习文档 Markdown note Pallas TPU Distributed Collectives 复习文档 Markdown note MoE 优化的探索 Markdown note DeepSeek V4笔记 Markdown note ZERO训练技术 Markdown note
View complete archive