fix(docker): SLS-377 install torchvision in GPU worker image

[deanq]

Problem

A deployed Flash app importing torchvision fails at worker runtime with ModuleNotFoundError (customer escalation SLS-377, Zendesk #41953).

Root cause is a three-part gap:

1. flash build/deploy unconditionally strips torchvision from the deployment tarball — SIZE_PROHIBITIVE_PACKAGES = {torch, torchvision, torchaudio, triton}, merged as a set union, so there is no CLI flag to force-include it. The justification is that these are "already provided by the GPU base images (runpod/pytorch:*)".
2. That assumption is false for torchvision. Probing the current base image (runpod/pytorch:1.0.3-cu1281-torch291-ubuntu2204): torch, torchaudio, triton are present — torchvision is not.
3. The worker Dockerfile re-adds numpy explicitly (it hit the same gap) but had no torchvision backfill.

Net: stripped from tarball -> absent from base image -> never reinstalled -> import fails at runtime.

Fix

Install torchvision in the GPU worker image, mirroring the existing numpy backfill:

• Pin TORCHVISION_VERSION=0.24.1, the release paired with the base torch 2.9.1 (bump both together on a torch upgrade).
• Install from the CUDA wheel index to stay CUDA-aligned.
• Add a build-time import torchvision assertion so a future torch bump that breaks the pairing fails the build, not the customer.

Not fixed by un-excluding torchvision from the tarball: its wheels are CUDA/platform-specific and depend on torch (which stays excluded), so bundling risks a CUDA mismatch and a second torch. The Dockerfile is the correct layer, consistent with the numpy precedent.

Test plan

• docker buildx build --platform linux/amd64 --build-arg PYTHON_VERSION=3.12 . — succeeds; torchvision resolves to 0.24.1+cu128 with torch==2.9.1 already satisfied (no torch re-pull).
• Build-time verify stage prints: Python 3.12 OK, torch 2.9.1+cu128, torchvision 0.24.1+cu128, numpy 2.5.1.
• make quality-check passes (14/14 handler tests + lint/format/typecheck/coverage).

Follow-ups (out of scope here)

• CPU images have the same gap — the tarball exclusion is unconditional, so a CPU Flash app importing torchvision would also fail. Left for a separate change/ticket since CPU torchvision is a different wheel and adds image weight.
• The SIZE_PROHIBITIVE_PACKAGES comment in flash build.py ("already provided by the GPU base images") is now inaccurate for torchvision and should be corrected in a flash-side PR.

[Copilot]

Pull request overview

Installs torchvision into the GPU worker Docker image to prevent runtime ModuleNotFoundError when Flash build/deploy strips torchvision from the deployment tarball and the runpod/pytorch:* base image does not provide it.

Changes:

• Add a pinned TORCHVISION_VERSION build arg with rationale and upgrade guidance alongside TORCH_VERSION.
• Install torchvision in the worker image from the CUDA wheel index (TORCH_INDEX_URL) to match the base image’s CUDA/torch build.
• Extend the build-time verification step to import and print the installed torchvision version.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.