Pin ComfyUI to 0.29.0, build with CUDA 12.8, expand allowed CUDA versions to 12.8–13.3

[lukepiette]

What

• Pin ComfyUI to 0.29.0 (latest stable release, 2026-07-29) instead of latest, so builds are reproducible and don't silently change between deploys. comfy-cli is also pinned (1.13.0) for the same reason.
• Build with CUDA 12.8: default base image moves from nvidia/cuda:12.6.3-cudnn-runtime-ubuntu24.04 to nvidia/cuda:12.8.1-cudnn-runtime-ubuntu24.04 (Dockerfile defaults + docker-bake.hcl globals).
• Pin the RUNTIME PyTorch to 2.11.0+cu128: the venv that actually runs ComfyUI (/opt/venv) previously got its torch from default PyPI via the requirements mirror step — and PyPI torch ≥ 2.11 ships CUDA 13 builds (nvidia-*-cu13 deps) that require driver ≥ 580. That's incompatible with hosts advertising CUDA 12.8/12.9 (driver 570/575) and is likely the source of the current torch-related worker failures. Installing torch==2.11.0+cu128 before the mirror step satisfies ComfyUI's bare torch requirement, so the runtime torch is now a known cu128 build that runs on driver ≥ 570.
• .runpod/hub.json: allowedCudaVersions updated from ["12.7", "12.6"] to ["12.8", "12.9", "13.0"]. The disabled .runpod/tests_.json is synced to the same list.
• Changeset added (minor) so the change ships with the next release cut.

Validated on real hosts (2026-07-30)

This exact build (base target, PR Dockerfile defaults) was pushed to a test registry and run as a serverless endpoint pinned to each CUDA version, one job per host type:

Host CUDA	GPU	Result
12.8	RTX 4090	✅ COMPLETED — worker healthy (GPU pre-flight passed), workflow executed, image returned (exec 3.0s)
12.9	RTX 5090 / H200	⚠️ Could not schedule — no 12.9 host became available during the test window (capacity, not compatibility; cu128 requires driver ≥ 570, which 12.9 hosts satisfy by definition). Kept in the allowlist — harmless, and correct when 12.9 hosts appear.
13.0	RTX 4090	✅ COMPLETED — worker healthy, workflow executed, image returned (exec 3.1s)

Additional validation notes:

• The allowlist ceiling of 13.0 matches Runpod's API schema exactly: the platform's CUDA version enum currently ends at "13.0" — values above it don't exist yet, so this list covers every currently-valid version ≥ 12.8. When the platform adds 13.1+, extend the list (or switch to minCudaVersion if hub.json gains support for it — floor-only semantics avoid this ceiling maintenance entirely).
• The old 12.6/12.7 list is nearly dead capacity-wise (zero community-cloud 4090 stock, Low secure on 12.7 only); current ADA_24 stock is concentrated on 12.8 and 13.0 hosts, so this change is a strict capacity improvement and unlocks community cloud.
• CUDA 12.8 + cu128 torch also unlocks Blackwell (sm_120 / RTX 5090) support.

Known limitations / follow-ups

• comfy --install's workspace venv (/comfyui/.venv) still receives its own unused torch (~5–7 GB installed). Removing it is a follow-up image-size win, left out here to keep this change low-risk.
• The base-cuda12-8-1 bake target now produces nearly the same image as base (both end with a cu128 runtime torch). Left untouched since release.yml/manual-build-all.yml reference it by name; consolidating is a follow-up.
• Existing Hub endpoints keep the old release and its exhausted 12.6/12.7 host pool — release notes should tell users to update their endpoint.

Verification (build-time)

• torch==2.11.0+cu128 / torchvision==0.26.0+cu128 / torchaudio==2.11.0+cu128 cp312 manylinux wheels confirmed present on the cu128 index (it caps at 2.11.0 — newer torch is cu129/cu13x only).
• ComfyUI 0.29.0's requirements verified compatible with the existing transformers>=4.50.3,<5 / huggingface-hub<1.0 pins (DR-1170); --quick-test-for-ci --cpu flags still exist in 0.29.0, so the build-time smoke test keeps catching startup-breaking deps.
• Pinning --version 0.29.0 resolves the tag by string construction in comfy-cli — no GitHub API call, removing the rate-limit risk that latest resolution carries.

🤖 Generated with Claude Code

[Chmokachka]

LGTM