Training crash: "n_instances exceeds int range" (MCMC, 10M splats)

[msultan32]

Description

Headless training crashes when MCMC densification reaches --max-cap with a large dataset.

Environment

• LichtFeld Studio: v0.5.2-128-gc4eb3dea (tag c4eb3dea)
• Platform: GCP Cloud Run, NVIDIA GPU, CUDA driver 13.0
• Mode: --headless --train
• Dataset loader: COLMAP (binary), 6 cameras, 816 images, 121,219 initial points

Command

LichtFeld-Studio \
  -d /path/to/lichtfeld_input \
  -o /path/to/training_output \
  -i 30000 \
  --headless --train \
  --mask-mode ignore \
  --enable-mip \
  --max-cap 10000000 \
  --strategy mcmc

Error

These are some of the logs from my pipeline:

Training [...] 16400/30000 | Loss: 0.0828 | Splats: 10000000 (+)
[PipelinedImageLoader] Shutting down...
[PipelinedImageLoader] Done: 16508 loaded, 15692 hits, 816 misses
[error] application.cpp:190 Training error: n_instances exceeds int range

Additional context

Same dataset trains successfully with lower --max-cap (e.g. 1M–2M).