Report: PartCrafter runs on AMD ROCm (RDNA4 gfx1201) - no code changes needed

[bioritmovideo]

Sharing a compatibility datapoint, since we found no prior reports of PartCrafter on AMD.

PartCrafter works out of the box on ROCm — no patches to the repo code.

Setup

• GPU: AMD Radeon AI PRO R9700 32GB (gfx1201, RDNA4), ROCm 7.2.1, Ubuntu 24.04
• torch 2.12.0+rocm7.2 (official wheel), Python 3.12
• Installed requirements with three adjustments:
  • skipped deepspeed / wandb (not needed for inference);
  • skipped torch-cluster (CUDA-only wheel) — its fps is never called in the
    inference path anyway (we stubbed it to be safe);
  • default SDPA attention (no flash-attn on ROCm) — works fine.

Results

• Sample robot image, --num_parts 3, seed 42: 110.8 s full pipeline,
  peak 8.7 GB VRAM, 3 clean semantically-sensible parts.
• Custom character/prop images, --num_parts 4–8, --rmbg: 118–161 s, ~9.5 GB peak.
• Denoising ~3 it/s (50 steps, fp16).

One observation for other users: on stylized orthographic top-down illustrations the
part decomposition degrades noticeably (most geometry collapses into one part) — clearly
the Objaverse render domain gap the README mentions; a 3/4 perspective view of the same
subject decomposes well. --style_transfer presumably helps but we haven't tested it
(needs a Gemini key).

Happy to provide more details or run tests on this hardware if useful.