During October 2025, Ppadjin enhanced the tenstorrent/tt-mlir repository by implementing GlobalAvgPool2d support in TTIR and TTNN, addressing out-of-memory issues in fusion patterns, and expanding validation tests to improve model coverage. They stabilized the Conv2d path by removing a slicing workaround in tt-metal after resolving the underlying bug. In tenstorrent/tt-forge-models, Ppadjin integrated Stable Diffusion XL UNet loading and variant support into the testing framework using PyTorch and Hugging Face models. Additionally, they extended tt-xla to support conditional generation testing for UNet on a single device, strengthening end-to-end validation across MLIR-based engines.