v0.3.3

chore: correct parameters in runpod

v0.3.2 — RunPod serverless complete

Changes

• Dockerfile: add runpod SDK to pip install (was missing — caused import error at container startup)
• handler.py: read TRAINING_STEPS / MATCHING_TYPE env vars set via RunPod Hub config; forward --matching to pipeline; stream pipeline stdout+stderr live via Popen (no more silent 10–60 min jobs)
• scripts/test_runpod.sh: new script to test a deployed RunPod endpoint via API — submits job, polls status, validates ply_base64, saves decoded PLY locally
• .runpod/tests.json: expand allowedCudaVersions to 12.4–12.9 (was 12.9 only); bump test timeout to 900 s to absorb cold-start Vulkan init

v0.4.0 — Three-repo architecture + frontmarket site

Architecture refactor.

Dockerfile now pulls compiled COLMAP and Brush binaries from upstream images:

• ghcr.io/alx/colmap-serverless:latest — COLMAP 4.0.3, built from source
• ghcr.io/alx/brush-serverless:latest — Brush (main), built from source

allowedCudaVersions broadened to 12.4–12.9 (was 12.9 only), matching the upstream images built on CUDA 12.4.1 (minimum for ubuntu24.04).

New:

• scripts/test_local_colmap.sh and scripts/test_local_brush.sh
• docs/frontmarket/ Hugo marketing site → https://alx.github.io/runsplat/
• .github/workflows/frontmarket.yml — deploys frontmarket to gh-pages on push

Companion repos:

• https://github.com/alx/colmap-serverless
• https://github.com/alx/brush-serverless

v0.3.1 — Fix allowedCudaVersions format

Use "12.9" instead of "12.9.3" in .runpod/hub.json and .runpod/tests.json.

v0.3.0 — Fix CUDA version requirement

fix: allowed cuda 12.9.3

v0.2.9 — Fix brush SIGSEGV on headless server

Fix

brush (built with eframe/wgpu) crashed with SIGSEGV ~28s after launch on every run. Root cause: the binary requires a display connection on Linux even in headless training mode — without DISPLAY set, Vulkan/X11 surface initialization fails fatally.

Changes:

• scripts/pipeline.py: wrap brush invocation with xvfb-run -a to provide a virtual X11 framebuffer. wgpu still uses the NVIDIA GPU for compute via Vulkan; only the window surface goes through Xvfb.
• Dockerfile: add xvfb package; create /usr/share/vulkan/icd.d and /etc/vulkan/icd.d directories so the NVIDIA Container Toolkit can inject the Vulkan ICD at runtime.

v0.2.9 — Fix CUDA version requirement

fix: force CUDA 12.9 to allowedCudaVersions

v0.2.8 — Fix CUDA version requirement

Add 12.9 to allowedCudaVersions in both .runpod/hub.json and .runpod/tests.json. The colmap/colmap:latest image requires cuda>=12.9; without this, RunPod was scheduling pods on hosts with older drivers, causing the OCI runtime error.

v0.2.7 — Single-stage Dockerfile

Remove separate brush-builder stage; brush binary is now downloaded directly in the runtime stage alongside apt packages.

v0.2.6 — Use pre-built COLMAP image

Changes

Both build stages are now pre-built binaries — no C++ or Rust compilation happens during the Docker build:

• brush: downloaded from ArthurBrussee/brush v0.3.0 (SHA256 verified)
• COLMAP: runtime image rebased on colmap/colmap:latest (v4.0.3, CUDA 12.9.1), which ships the binary and all ~200 runtime dependencies

Build time drops from ~15+ minutes to under 2 minutes. The OOM-kill issue is fully resolved.