chore(deps): move the cpu and cuda extras to torch 2.13.0 (cu130)

[Pfannkuchensack]

Summary

The cpu and cuda extras were still on torch 2.7.1 (cu128) while rocm and xpu were already on 2.13.0. This moves them to 2.13.0+cpu / 2.13.0+cu130, so every backend on Windows and Linux x86_64 now runs the same torch. cu128 stops at torch 2.11, which is why the CUDA build moves to cu130. macOS (capped below 2.8) and linux-aarch64 stay on 2.7.1.

The bump broke three things. This PR fixes all three:

• Seeds changed. torch 2.13 changed its CPU float16/bfloat16 randn/rand for tensors of 16+ elements; float32 is unchanged. SD1.5/SDXL (noise node), SD3, FLUX.1, FLUX.2, CogView4 and the Z-Image seed variance enhancer drew their seeded noise in half precision. The same seed gave a different image after the bump, and a different image on macOS (2.7.1) than on Windows/Linux. A new setting noise_dtype (default float32, optional float16) now controls the draw, through TorchDevice.choose_noise_dtype. After the fix, 2.7.1 and 2.13 give the same composition: 31–36 dB instead of 11–12 dB.
• xformers stopped working. PyPI's xformers 0.0.35 is built for torch 2.10+cu128 and cannot load its extension next to cu130 torch, so every memory_efficient_attention call raised. InvokeAI prefers xformers on compute capability ≤ 7 when it is installed, and the launcher installs it for 20xx cards. xformers now comes from PyTorch's cu130 index, where it is built for CUDA 13.
• Dropped GPUs and old drivers failed silently. cu130 needs compute capability 7.5+ (Turing and newer) and an R580+ driver. With an older driver, torch has no CUDA and Invoke quietly runs on the CPU. On Maxwell/Pascal/Volta, every generation fails with "no kernel image". A new startup check logs an error for both cases, and the docs name the floor, the symptoms and the ways out.

Before → after, RTX 4090, Windows, warm run, same code, fp8_compute on, model cache emptied before each case:

Model (1024², Wan 512²)	2.7.1	2.13
SDXL, 25 steps	5.0 s	4.5 s
FLUX.1 dev int8, 20 steps	20.7 s	18.7 s
FLUX.2 Klein 4B fp8, 4 steps (_scaled_mm)	2.5 s	2.0 s
Z-Image Turbo nvfp4, 8 steps	6.5 s	6.6 s
Krea-2 Turbo nvfp4, 8 steps	12.1 s	11.1 s
Wan 2.2 TI2V-5B GGUF Q4_K_M, 20 steps	12.1 s	10.6 s
RealESRGAN 4× upscale	4.0 s	4.0 s

This supports the claim "no warm regression for models that fit in 24 GB on this machine", not more (one warm sample per case).

Related Issues / Discussions

• ROCm 10 for Linux and Windows is the next step and waits for ROCm on Windows: keep VRAM resident and bound attention and VAE memory #304. Upstream feat(deps): support ROCm on Windows through the rocm extra InvokeAI#9606 covers the Windows side for v6.
• Launcher follow-ups belong in invoke-ai/launcher: the wording of its "older Nvidia" option, since pre-Turing cards are no longer supported by the cuda extra.

QA Instructions

Run on Windows with an RTX 4090 and driver 610.47 (reports CUDA 13.3), using a venv synced from the lock (uv sync --locked --extra cuda --extra test):

• Lock and pins: uv lock --locked (uv 0.11.28), python scripts/check_pins.py and scripts/check_aarch64_lock.py pass.
  • uv export per extra × platform: cpu/cuda/xpu/rocm and the no-extra resolution give 2.13.0 on win32 and linux x86_64. darwin and linux-aarch64 give 2.7.1.
  • Only torch, torchvision, triton, the NVIDIA/CUDA runtime wheels and the xformers source moved.
  • torch.cuda.get_arch_list(): sm_75, sm_80, sm_86, sm_90, sm_100, sm_120.
• Full suite: pytest -n 6 gives 11,674 passed, 184 skipped, 9 xfailed. This was run with HF_ENDPOINT unset: on this machine a local HF proxy makes 9 download/install tests fail, and they all pass without it.
• CUDA-dependent tests on the 4090 pass: fp8 _scaled_mm, allocator, attention probe, nvfp4, bitsandbytes, Krea-2 SDPA, FLUX.2 working memory. test_sdpa_scope confirms torch's sdpa_kernel still leaks across threads in 2.13.
• -m slow: 21 passed. The 3 errors in tests/backend/ip_adapter/test_ip_adapter.py come from a missing model_installer fixture and happen on main too.
• New tests on both torch 2.7.1 and 2.13: tests/backend/util/test_noise_dtype.py and tests/app/util/test_cuda_build_compatibility.py.
  • Golden float32 values, identical on both versions.
  • Mutation runs: ignoring the setting fails 11 tests; always drawing in float32 fails 6.
• E2E before/after (table above): all images inspected and clean.
  • With the final code, 2.7.1 vs 2.13 give SDXL 35.8 dB, FLUX.1 32.6 dB, upscale 34.3 dB, and Klein fp8 without fp8_compute (fp8 storage path) 31.2 dB.
  • On 2.13 the float32 SDXL image is bit-identical to the former float16-draw image.
• Working-memory calibrators (scripts/calibrate_flux2_working_memory.py, calibrate_qwen_vae_working_memory.py) on both versions: every estimate is still an upper bound on the path CUDA actually dispatches.
• xformers from the cu130 index (synced from the lock): memory_efficient_attention runs; max abs difference vs SDPA is 0.0. Its build metadata lists CUDA 13.0 and arch 7.5–12.1. uv pip compile --torch-backend cu130 also resolves xformers from whl/cu130, so manual installs get the right build.
• Startup check on this machine: driver probe 13030, no error logged.
• Docs: pnpm build and check-deploy-output pass. docs/src/generated/settings.json, openapi.json and schema.ts are regenerated.

Not tested:

• Linux, and the Docker cuda image.
• Turing, Pascal or older cards, and drivers older than R580. The startup check is covered with faked device queries only.
• The --torch-backend=cu126 fallback for older GPUs. The docs say so.
• NF4 (no model installed here).
• Qwen-Image timings are not comparable: its ~29 GiB transformer does not fit in 24 GB, and the offload-bound runs swing about 4× in both directions on both versions. It served as a functional check only.

Review

Resolved:

• The Z-Image seed variance enhancer's CPU bf16 torch.rand draw was also affected; it now draws through choose_noise_dtype.
• PyPI xformers was incompatible with cu130.
• There was no guidance or failure message for dropped GPUs and old drivers.
• The noise_dtype description overstated what float16 restores.
• A golden-value test now catches a future change to torch's float32 draw.
• Stale aarch64 comments were corrected, and the startup check now parses suffixed arch entries (sm_90a).

Remaining risks and limitations:

• Qwen-Image VAE CUDA constants. Measured decode peaks are 2770 on 2.13 (2734 on 2.7.1) against the shipped 2900 (4.7 % headroom). Encode peaks are 1557 against 1600 on both versions (2.8 %, same as before this PR). ROCm on Windows: keep VRAM resident and bound attention and VAE memory #304 rewrites exactly those lines, so the adjustment follows after it.
• Golden-value test on macOS. Its tolerance (1e-4) allows for the ARM normal_fill path; it was measured on x86 only.
• Pre-existing, unrelated flake. test_flux2_working_memory.py::…forced_math_forward[1-4096-512] measures 0 B when earlier CUDA tests ran in the same process. It fails the same way on 2.7.1.

Compatibility / Rollout

• GPU floor. Nvidia GPUs now need compute capability 7.5+ and a driver from the R580 series or newer. This covers Docker hosts running the cuda image as well. Maxwell, Pascal and Volta (GTX 9xx/10xx, Titan V, P40/P100/V100) are no longer supported by the cuda extra; the release notes should say so.
• Seeds. On Windows and Linux, images made before this update cannot be reproduced exactly for SD1.5/SDXL, SD3, FLUX.1, FLUX.2, CogView4 and Z-Image seed variance. On macOS, noise_dtype: float16 keeps the previous images. A FAQ entry explains this.
• New config setting. noise_dtype (float32 | float16, default float32) is additive and needs no migration. openapi.json and schema.ts are regenerated.
• pins.json. The cuda URLs moved to cu130. They are only read by launcher installs of releases before 6.14; v7 syncs --frozen from the lock.
• Lock. [tool.uv] constraint-dependencies holds win32/linux x86_64 resolutions to torch 2.13.0. It is not part of the published metadata, so installing the released package with --torch-backend keeps the open range. uv pip install inside a checkout is held to it.

Checklist

• The PR has a short but descriptive title, suitable for a changelog
• Meaningful regression coverage added / updated where needed; obsolete tests/code removed
• Persisted-state and API changes include required migrations / compatibility validation
• Relevant performance/efficiency opportunities considered; material claims have evidence
• Material review findings resolved and relevant checks rerun
• Documentation added / updated (if applicable)
• Updated What's New copy (if doing a release after this PR)

[joshistoast]

conflictssssssss