Projects
Open-source work on efficient LLM training and systems.
Research projects
SlideFormer
Full-parameter LLM fine-tuning on a single GPU through layer-sliding execution and CPU/GPU/NVMe pipelining.
Result. Fine-tunes 123B+ models on a single RTX 4090; the paper reports 1.40–6.27× higher throughput than evaluated baselines.
Runtime, reproducibility scripts, and documentation available.
SlideDP
Scaling host-resident LLM fine-tuning across multiple GPUs with synchronous data parallelism.
Result. Fine-tunes Qwen2.5-72B on four RTX 4090 GPUs; matched-batch experiments report 1.46–2.64× geometric-mean throughput ratios over evaluated baselines.
Reference multi-GPU extension available; the optimized runtime is forthcoming.
GitHub profile
GitHub Repositories
View all repositories on GitHub.
Code availability
-
SlideFormer: Runtime, reproducibility scripts, and documentation available. Paper · DAC 2026.
-
SlideDP: Reference multi-GPU extension available; the optimized runtime is forthcoming. Paper · Preprint, 2026.