Deferred / rejected pod decisions #

Parked or rejected ideas around the pod setup. Nothing here is scheduled; items under Rejected are decided unless new information arrives.

Parked #

Re-pin the pod image off the rc tag #

The pod image is stock runpod/base:1.3.0-rc.164-ubuntu2404, pinned by digest (immutable), so this is not urgent; when a stable 1.3.0+ tag exists, re-pin the template / pod create image reference and re-check the uv version pinned in scripts/setup-pod.sh (constraint >=0.12.7,<0.13.0) at the same time.

Exercise the Jupyter option #

Documented in README_RUNPOD.md §3 as untested (JUPYTER_PASSWORD + port 8888/http). Try it once on a live pod, then either bless it in the runbook or drop the section.

Rejected #

Add torchvision (flipped the preprocessing backend; resolved 2026-09-14) #

Parked here as "ship with the next uv.lock change, and don't mix runs from before/after" — then adding dreamsim for dltb-distance pulled torchvision into uv.lock as a side effect, closing the item. The feared reproducibility break did not materialize: NOTES.md ("Runtime log noise" item 3) verified that the torchvision-backed transformers image processors are reachable only from the safety checker and image-encoder code paths, both dead in dltb's pipelines, so pixel output is unchanged. The caveat stands: re-check before ever enabling a safety checker or an image encoder, and compare deliberately against pre-change runs instead of mixing them if that happens.

Build a custom pod image (retired 2026-09-13) #

The first commissioning baked the locked venv into a ~12 GB custom image (published as refinementsystems/imgiter on Docker Hub; the old tags remain there, unchanged) to skip the multi-GB uv sync on pod boot. Retired because Runpod starts billing when the image pull starts: the ~12 GB pull (from Docker Hub, often throttled) cost more billed GPU time than the ~6 GB PyPI sync it skipped — and every uv.lock change forced an emulated linux/amd64 rebuild, a push, and a template digest re-pin. The stock base image + scripts/setup-pod.sh (README_RUNPOD.md §4) does the same job with no build train. For scale: the ~87.5 GB of HF model downloads every fresh pod pays dwarfs both sides of the trade anyway.