`torch.cuda.amp` in `__main__` self-test blocks and the IPEX hijack — deprecated since torch 2.3/2.4, unguarded on the main install path

Summary

17 references to the deprecated torch.cuda.amp API remain at HEAD (37a1cbbc). torch.cuda.amp.GradScaler has been deprecated since torch 2.3 and torch.cuda.amp.autocast since torch 2.4 — both still work but emit a FutureWarning, and the deprecation notice says they "will be removed in a future release". When that removal lands, every site below fails at import/run time.

Two groups:

1. __main__ self-test blocks — 8 references in 4 files, no version guard

All 8 are inside if __name__ == "__main__": blocks (at sdxl_original_unet.py:1241, sdxl_original_control_net.py:199, control_net_lllite.py:374, control_net_lllite_for_train.py:414). None of the 4 files has any version guard (LooseVersion / torch.__version__ / hasattr(torch.amp, ...) / try-except) around them. These are developer-facing smoke tests — the README and docs never mention running them — but they are the documented way to sanity-check those two UNet / ControlNet implementations.

2. library/ipex/hijacks.py — 9 references in the IPEX monkey-patch, hasattr-guarded

Lines 448-461 of ipex_hijacks(): custom_fwd/custom_bwd aliasing and torch.cuda.amp = torch.xpu.amp / torch.cuda.amp = torch.amp module swaps, guarded by hasattr(torch.xpu, "amp"), hasattr(torch.amp, "custom_fwd"), etc. This runs only on the Intel GPU (IPEX) initialization path — the guarded aliasing actually replaces torch.cuda.amp with the torch.amp implementation, so the deprecation concern here is lower.

No torch version constraint

• requirements.txt has no torch entry; torch is pulled in only indirectly via diffusers[torch]==0.32.1 (requirements.txt:3). setup.py has no install_requires.
• README.md:211 states this explicitly: "The file does not contain requirements for PyTorch. Because the version of PyTorch depends on the environment, it is not included in the file."
• README.md:213: "The scripts are tested with PyTorch 2.6.0. PyTorch 2.6.0 or later is required." and README.md:215 recommends PyTorch 2.8.0 for RTX 50-series — both of these already deprecate torch.cuda.amp.GradScaler/autocast (since 2.3/2.4), so anyone on the documented install path gets the FutureWarning when running one of the self-test blocks, and would break outright once torch removes the module.

Suggested fix

Low-risk, 8 lines: swap the self-test blocks to the device-agnostic APIs.

scaler = torch.amp.GradScaler("cuda", enabled=True)
with torch.amp.autocast("cuda", enabled=True, dtype=torch.bfloat16):

torch.amp.GradScaler exists since torch 2.3, torch.amp.autocast since torch 2.0, so this works on the torch ≥ 2.6.0 floor the README already declares — no version-branching needed. The ipex/hijacks.py aliasing is arguably fine as-is (it already targets torch.amp), but worth a comment.

Low priority: the main training path (train_network.py, sdxl_train.py, train_util.py) has zero torch.cuda.amp references; this only affects the undocumented self-test blocks and the IPEX path.

Reference

• PyTorch AMP docs — official deprecation notice

• huggingface/lerobot#3167 — same migration in LeRobot, merged

• #1649 — closed; the only amp-related issue, and it is about a FutureWarning coming from torch's own torch.utils.checkpoint.py, not this repo's code