[ROCm] End-to-end inference verified on AMD MI300X

[ZJLi2013]

GeometryForcing runs successfully on AMD MI300X GPUs with ROCm 6.4 — no CUDA-specific kernel changes needed.

Environment

Component	Version
GPU	8x AMD Instinct MI300X (gfx942)
Docker	rocm/pytorch:rocm6.4.3_ubuntu22.04_py3.10_pytorch_release_2.6.0
PyTorch	2.6.0 (ROCm)
Flash Attention	AOTriton (via F.scaled_dot_product_attention)

What Works

• All module imports (DFoTVideo, DFoTVideoPose, VGGT, U-ViT3D)
• Checkpoint loading (geometry_forcing_state_dict.ckpt, 458.8M params)
• Multi-GPU DDP inference (8 GPUs)
• End-to-end video generation with realestate10k_mini (16 frames, 50 denoising steps)

Fixes

A few PyTorch 2.6 / torchmetrics v1.x compatibility fixes were needed (not ROCm-specific). Submitted as #17.

Results

Metric	Value
Inference time	~69s per batch (16 frames, 50 steps)
Per-step speed	~1.0–1.4 it/s
End-to-end (incl. model loading)	~3.5 min