cuDSS OOM in global_positioning on RTX 3090 (24GB): `CUDSS_STATUS_ALLOC_FAILED`

[DuVogel87]

Reproducible cuDSS allocation failure during global_positioning on 4262-image fisheye dataset

I am seeing a reproducible cuDSS allocation failure during global_positioning on a 4262-image fisheye dataset.

This looks similar to #215 / #191, but I’m posting full reproducible details from my setup.

---

Environment

• OS: Ubuntu 24.04.3 LTS
• CPU: AMD Ryzen 9 5900XT (32 logical CPUs)
• RAM: 64 GB
• GPU: NVIDIA GeForce RTX 3090 (24 GB VRAM)
• NVIDIA driver: 570.211.01

Build / Versions

• GLOMAP binary: /home/ubu/dev/src/glomap_cuda120/install/bin/glomap
• glomap --help shows: This version was compiled with CUDA!
• Local glomap repo revision used for this build:
  14e9104c2e60bc3f68fbc8b645b0d79a62f97f7a
• COLMAP in the same pipeline:
  COLMAP 3.13.0 (Commit d69fddb2 on 2026-02-15 with CUDA)

Dataset

• 4262 images (2-camera rig), total image size approximately 2.4–2.6 GB
• Image dimensions: 3840x3840

Feature extraction settings

• SiftExtraction.max_image_size=3200
• SiftExtraction.max_num_features=16384
• SiftExtraction.peak_threshold=0.0033
• GPU feature extraction enabled

Extracted features (from logs)

• min: 10426
• max: 21180
• mean: 15638.68 per image

---

Command used

/home/ubu/dev/src/glomap_cuda120/run_glomap_120.sh mapper \
  --database_path /.../database/database.db \
  --image_path /.../distorted/images \
  --output_path /.../distorted/sparse \
  --GlobalPositioning.use_gpu 1 \
  --GlobalPositioning.gpu_index -1 \
  --BundleAdjustment.use_gpu 1 \
  --BundleAdjustment.gpu_index -1 \
  --BundleAdjustment.optimize_intrinsics 1 \
  --BundleAdjustment.optimize_principal_point 1 \
  --BundleAdjustment.optimize_rig_poses 1 \
  --BundleAdjustment.max_num_iterations 300 \
  --skip_pruning 0

I also reproduced the same failure with --skip_pruning 1.

---

Error output

During Running global positioning ...:

Linear solver fatal error: cudssExecute with CUDSS_PHASE_FACTORIZATION failed Got error: CUDSS_STATUS_ALLOC_FAILED
Terminating: Linear solver failed due to unrecoverable non-numeric causes.
Ceres Solver Report ... Termination: FAILURE

---

Important behavior

Even after this solver failure, mapper continues and exports a COLMAP model.

In that case, the model quality can be invalid or corrupted (for example, extremely large reprojection error in analyzer output).

If I force CPU solvers (GlobalPositioning.use_gpu=0, BundleAdjustment.use_gpu=0), the reconstruction completes with normal stats.

---

Questions

1. Is this expected behavior when cuDSS runs out of memory (i.e. continue execution and still export a model), or should mapper fail hard in this case?
2. Are there any recommended flags or settings to reduce cuDSS memory pressure in global_positioning for datasets of this scale?

[ahojnnes]

Note that glomap development moved to the colmap repository. Please file your issues there. Can you see if the following fixes your issues: colmap/colmap#4217