I work on practical GPU enablement for 3D AI, differentiable rendering, Gaussian splatting, neural graphics, and reconstruction pipelines.

My main AMD target stack:

• AMD Radeon AI PRO R9700 / RDNA4 / gfx1201
• ROCm 7.2
• Python 3.12
• PyTorch ROCm builds
• Linux-based local AI / 3D workflows

I also maintain several CUDA wheel helper repositories for newer NVIDIA / Blackwell-oriented environments where upstream wheels are missing or delayed.

• amd-nvdiffrast-rocm72-gfx1201
  AMD ROCm 7.2 / RDNA4 / gfx1201 patch and validation bundle for NVLabs nvdiffrast.

• diff-gaussian-rasterization-rocm72-gfx1201
  ROCm 7.2 / gfx1201 porting work for diff-gaussian-rasterization.

• amd-gsplat-rocm72-gfx1201
  AMD ROCm / gfx1201 porting work for gsplat-style Gaussian splatting workflows.

• amd-gaussian-splatting-rocm72-gfx1201
  AMD ROCm / RDNA4 porting work for Gaussian Splatting pipelines.

• simple-knn-rocm72-gfx1201
  ROCm / gfx1201 porting work for simple-knn.

• fused-ssim-rocm72-gfx1201
  ROCm / gfx1201 porting work for fused-ssim.

• amd-sugar-rocm72-gfx1201
  AMD ROCm / gfx1201 related porting work for SuGaR-style reconstruction workflows.

• amd-nerfstudio-rocm72-gfx1201 Qualified AMD ROCm 7.2 / RDNA4 / gfx1201 integration for the Nerfacto training chain from Nerfstudio, using tiny-rdna4-nn and a ROCm-compatible nerfacc runtime.

• tiny-rdna4-nn
  RDNA4-focused tiny neural network / fused MLP kernel work for AMD gfx1201.

These repositories are primarily wheel/build helper repositories. The wheel projects are not the same as the ROCm/gfx1201 ports above.

These repositories target CUDA / Blackwell-era environments, including cases where prebuilt upstream wheels are unavailable or not yet aligned with newer CUDA/PyTorch stacks:

• pytorch3d-Wheel
• nvdiffrast_wheel
• torch_scatter_wheel
• xformers-wheel
• xformers-wheel_cu130
• kaolin_cu130_wheel
• gaussian-wheels

• amd-gpu-torch-runtime-patch
  Runtime patching / helper work for AMD GPU detection and PyTorch workflows.

• torch-fade-logger
  Utility project around PyTorch logging / diagnostics.

• PIXIE-to-SMPL-X-Converter
  Conversion helper work around PIXIE / SMPL-X workflows.

The goal of these repositories is practical GPU enablement:

CUDA-centered or vendor-specific 3D/AI project
        ↓
porting / build repair / validation
        ↓
AMD ROCm/RDNA4 or CUDA/Blackwell target stack
        ↓
reproducible install, wheel, or patch bundle
        ↓
usable local 3D AI workflow

Most projects are community ports, validation bundles, wheel helpers, or experimental compatibility layers.