vggt

tmp_amur/vggt workspace

Prepared on 216 (user-4U-GPU-Server) for VGGT-style geometry guidance smoke tests.

Contents

• repos/vggt: official VGGT code, editable-installed in .venv.
• repos/GeometryForcing: reference code for VGGT feature-alignment style video geometry guidance.
• repos/fantasy-world: reference code for VGGT-style geometry branch + video backbone design.
• weights/facebook--VGGT-1B: downloaded facebook/VGGT-1B pretrained weights.
• .venv: uv virtual environment with PyTorch 2.3.1+cu121 and VGGT installed.
• uv_python/cpython-3.10.20-linux-x86_64-gnu: uv-managed CPython runtime copy, included because .venv/bin/python is an absolute symlink on the source machine.
• data_views/smoke_uav: relative symlinks to the read-only source dataset mirror location.
• scripts/smoke_vggt.py: smoke script for local VGGT inference.
• outputs/smoke_vggt: smoke outputs: camera extrinsics/intrinsics, depth, confidence, and summary.

Smoke command

export HF_HOME=weights/hf_home
.venv/bin/python scripts/smoke_vggt.py \
  --image-dir data_views/smoke_uav \
  --out-dir outputs/smoke_vggt \
  --max-images 4 \
  --device cuda:0 \
  --model-path weights/facebook--VGGT-1B

Verified smoke summary:

• torch: 2.3.1+cu121
• CUDA visible devices: 8
• input images: 4 UAV frames from 2025-03-27_16-37-55_log
• extrinsic shape: [1, 4, 3, 4]
• intrinsic shape: [1, 4, 3, 3]
• depth shape: [1, 4, 518, 518, 1]

Upload retry

export HF_TOKEN=<token>
export HTTPS_PROXY=http://10.130.159.169:7890
export HTTP_PROXY=http://10.130.159.169:7890
export ALL_PROXY=http://10.130.159.169:7890
scripts/retry_upload_to_hf.sh

The source dataset was only read. The smoke image folder contains relative symlinks only.