[justinwlin]

Installs the pinned torchvision in Dockerfile-lb, matching what #101 did for Dockerfile.

Why

flash build never bundles torchvision, because a bundled copy would shadow one matched to the image's torch. The queue-based GPU image has installed torchvision==0.24.1 since #101. The load-balanced GPU image never got it, so LB endpoints that import torchvision fail with ModuleNotFoundError.

Change

• ARG TORCHVISION_VERSION=0.24.1, the release paired with the base torch 2.9.1.
• Install from the cu128 index, the same install block as Dockerfile.
• Add import torchvision to the build-time verify step.

Verification

• The install block is identical to Dockerfile (diff of the two blocks is empty).
• The published images confirm the gap: runpod/flash:py3.12-latest has the torchvision step in its build history, and runpod/flash-lb:py3.12-latest doesn't.
• torch==2.9.1+cu128 + torchvision==0.24.1 resolves cleanly against the cu128 index. PyPI's latest torchvision 0.29.1 needs torch>=2.14, which is why the pin matters.
• I didn't run a local image build (amd64 GPU image). CI builds Dockerfile-lb for Python 3.10 through 3.13, and the non-3.12 path reinstalls torch before this step.

Related Flash PRs

runpod-flash stops getting fixes on 2026-10-29, so the runpod/flash changes need to merge and release before then.

PR	What it does	Status (2026-10-26)	Requires before merge
#113	Installs torchvision in the GPU load-balancer image	Ready, no review	Nothing. Must be merged and released before runpod/flash#390.
runpod/flash#390	Keeps torch's dependency tree out of flash build archives	Approved	#113 merged and released.
runpod/docs#923	Adds the flash deploy step and fixes stale Python and build guidance	Draft	Nothing. Documents runpod-flash 1.21.0 as released.
runpod/flash#369	Reads runpodctl's top-level apikey as a credentials fallback	Draft	Nothing.