FutureWarning: torch.cuda.amp.autocast is deprecated (use torch.amp.autocast)

[chipi]

Description

When using thinc (e.g. via spaCy with PyTorch-backed components), PyTorch emits a FutureWarning because thinc still uses the deprecated torch.cuda.amp.autocast() API.

Warning:

FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast(...)` instead.

Location: thinc/shims/pytorch.py

• Line 114: with torch.cuda.amp.autocast(self._mixed_precision): (in predict)
• Line 128: with torch.cuda.amp.autocast(self._mixed_precision): (in begin_update)

Expected behavior

No deprecation warning. PyTorch recommends the device-agnostic API:

• Old (deprecated): torch.cuda.amp.autocast(...)
• New: torch.amp.autocast(device_type="cuda", ...) or torch.autocast("cuda", ...)

See PyTorch amp docs.

Suggested fix

Replace:

with torch.cuda.amp.autocast(self._mixed_precision):

with the new API, e.g. (depending on PyTorch version compatibility):

with torch.amp.autocast(device_type="cuda", dtype=self._mixed_precision or torch.float16):

or the equivalent that preserves the current _mixed_precision behaviour (e.g. dtype=torch.float16 when mixed precision is enabled).

Environment

• thinc: 8.2.5
• spacy: 3.7.5
• PyTorch: 2.x (emits the deprecation)

Triggered when running pipelines that use spaCy with PyTorch (e.g. NER, transformers-backed components). Cosmetic only (no functional impact), but clutters logs and CI output.