fix(core): use --index-url for torch family + parameterise via TORCH_INDEX arg

[ductiletoaster]

Summary

ComfyUI's requirements.txt has bare torch / torchaudio / torchvision lines (no version pins, no wheel URLs). The core Dockerfile installed those with --extra-index-url https://download.pytorch.org/whl/cu128 — but that flag is additive. Pip picks the highest version found across PyPI + the extra index.

When PyPI ships a newer torch than the cu128 index (which happens regularly), pip splits the resolution:

Package	PyPI has	cu128 index has	Pip picks
torch	2.12.0 (cu130 ABI)	2.11.0+cu128	2.12.0 (PyPI) — newer wins
torchaudio	(none matching)	2.11.0+cu128	2.11.0+cu128
torchvision	0.27.0 (cu130 ABI)	older	0.27.0 (PyPI)

Result: torch / torchvision compiled against cu130 + torchaudio compiled against cu128 → runtime crash on first import torchaudio:

RuntimeError: Detected that PyTorch and TorchAudio were compiled with
different CUDA versions. PyTorch has CUDA version 13.0 whereas
TorchAudio has CUDA version 12.8.

The current published complete:cuda-latest image happens to work because it was built when PyPI didn't yet have torch 2.12.0 — pure timing luck. The next clean Weekly Fresh Rebuild would have failed regardless of any extras changes.

How I confirmed

Built complete-cuda locally from main (no PRs in flight). Resulting image errors on python -c "import torchaudio" with the exact crash above. Inspecting the build log shows the core layer downloading torch-2.12.0 from PyPI (no +cu128 suffix) and torchaudio-2.11.0+cu128 from the cu128 index.

Fix

Two-stage install in core:

1. pip install --index-url https://download.pytorch.org/whl/${TORCH_INDEX} torch torchaudio torchvision — --index-url replaces PyPI for this command, so all three are pulled from the same wheel index with a consistent CUDA ABI.
2. pip install -r /app/ComfyUI/requirements.txt runs normally afterwards — torch family already satisfied, pip skips re-resolution.

Wired through ARG TORCH_INDEX=cu128 and parameterised in docker-bake.hcl:

• core-cuda → TORCH_INDEX=cu128
• core-cpu → TORCH_INDEX=cpu

Adding a cu130 variant becomes a single bake-target addition (args = { TORCH_INDEX = "cu130" }), no Dockerfile changes.

Local verification

Building locally with the fix now to confirm consistent CUDA versions across the torch family. Will comment with the result before requesting merge.

Supersedes

#46 (closed). That one tried to fix it in complete with --extra-index-url cu130, which (a) targeted the wrong layer, (b) wouldn't fix the index-priority bug, and (c) would have unnecessarily moved the image to cu130.

Test plan

• Local docker buildx bake core-cuda complete-cuda produces consistent cu128 wheels for torch / torchaudio / torchvision (in-flight)
• docker run --rm <built-image> /app/.venv/bin/python -c "import torchaudio; print(torchaudio.__version__)" succeeds without CUDA-mismatch error
• CI rebuild passes
• In-cluster pod boots cleanly when harmony bumps to the new image SHA
• core-cpu build still works (TORCH_INDEX=cpu — separate validation since prod uses cuda)

🤖 Generated with Claude Code

[ductiletoaster]

✅ Local verification complete.

Force-rebuilt `core-cuda` and `complete-cuda` from scratch with this branch:

Build log — core layer's torch install:
```
Downloading torch-2.11.0+cu128-cp312-cp312-manylinux_2_28_x86_64.whl
Downloading torchaudio-2.11.0+cu128-cp312-cp312-manylinux_2_28_x86_64.whl
Downloading torchvision-0.26.0+cu128-cp312-cp312-manylinux_2_28_x86_64.whl
Successfully installed ... torch-2.11.0+cu128 torchaudio-2.11.0+cu128 torchvision-0.26.0+cu128 ...
```

All three pulled from the cu128 index with matching local-version IDs — no more PyPI generic torch sneaking in. The follow-up `pip install -r /app/ComfyUI/requirements.txt` ran cleanly without re-resolving torch family.

Runtime check in the built image (`docker run --entrypoint /app/.venv/bin/python ... -c "import torch, torchaudio, torchvision; ..."`):
```
torch = 2.11.0+cu128
torchaudio = 2.11.0+cu128
torchvision = 0.26.0+cu128
```

No `RuntimeError`, no CUDA mismatch. Ready for review.