OOM during global positioning on ~10k fisheye images (3.9M tracks) even with cuDSS

[DuVogel87]

• Context: COLMAP DB ~34 GB; ~10,130 fisheye images (~3 GB total). Running glomap mapper (cuDSS build, CUDA 12.8) on Ubuntu, 64 GB RAM, GPU: NVIDIA GeForce RTX 3090.
• Observed: After track establishment (10.3M → 3.885M filtered), global positioning begins, then the kernel OOM-kills glomap at ~42 GB RSS. Happens consistently. cuDSS is enabled; --GlobalPositioning.use_gpu/--BundleAdjustment.use_gpu are at defaults (1).
• Logs/syslog snippet: Out of memory: Killed process XXXX (glomap) total-vm:~59G, anon-rss:~42G … right after “Solving the global positioner problem.”
• Questions/requests:
  1. Recommended flags/workflow for large scenes to reduce peak RAM during global positioning/BA?
  2. Is there a built-in way to chunk/stream the solver (e.g., batch tracks/poses) or an out-of-core mode?
  3. Any known memory regressions/tunables for cuDSS builds in recent commits?
  4. Guidance on safe limits (tracks, matches, overlaps) for ~10k-image rigs?
• Environment: glomap commit as of 2025-11-26 (COLMAP 3.13.0, cuDSS 0.7.1, CUDA 12.8), Ubuntu, 64 GB RAM, NVIDIA GeForce RTX 3090.
• Repro command (example):

  glomap mapper \
    --database_path /path/to/database.db \
    --image_path /path/to/images \
    --output_path /path/to/glomap_out \
    --BundleAdjustment.optimize_intrinsics 1 \
    --BundleAdjustment.optimize_principal_point 0 \
    --BundleAdjustment.optimize_rig_poses 1 \
    --BundleAdjustment.max_num_iterations 300
  

  (Global positioning starts, then OOM.)