This repository implements a hybrid sample generation pipeline for imaging data to extend existing training datasets.

For me, the goal of this project was to handle transformers fine tuning by implementing and applying optimization training techniques using limited computational resources.

4-GPU H800 low-resource post-training optimization project for Qwen3.6-35B-A3B MoE.

GPU memory-efficient training for PyTorch - 90%+ memory savings through gradient compression

Drop-in, idempotent speed patches for PGSR surface-reconstruction training — removes per-iteration PCIe transfers, redundant ones-kernel convolutions, and GPU-CPU logging syncs (plus a real EMA-logging bug) with CPU-proven numerical parity: same geometry, lower GPU bill.