`torch.cuda.amp.autocast` / `GradScaler` usage across FunASR emits FutureWarning on torch ≥2.4 — migrate to `torch.amp`

[xyf5432]

Summary

FunASR uses the deprecated torch.cuda.amp APIs in 37 files. The
hot paths that actually call them:

File	Line
funasr/models/llm_asr/model.py	8 from torch.cuda.amp import autocast; 7 call sites (586, 621, 886, 1162, 1206, 1554 with torch.cuda.amp.autocast(...))
funasr/train_utils/trainer_ds.py	9 from torch.cuda.amp import autocast, GradScaler; 35 with torch.cuda.amp.autocast(enabled=True, dtype=dtype, ...)
funasr/models/bat/model.py	27 from torch.cuda.amp import autocast (guard only checks torch < 1.6, not the deprecation)
…34 more files reference torch.cuda.amp	see below

Verified on torch 2.11.0:

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

autocast has been deprecated since torch 2.4 and GradScaler since
torch 2.3. Both still work but print a FutureWarning on every call —
the current import guard in bat/model.py only covers torch < 1.6 and
does not protect against this.

Suggested fix

The replacement APIs (torch.amp.autocast('cuda', ...) since 2.0,
torch.amp.GradScaler('cuda', ...) since 2.3) are available in all
current torch versions. If older torch versions must remain supported,
gate on the version — 2.3 is the boundary because the import below needs
both names, and torch < 2.3's torch.cuda.amp is not yet deprecated:

if LooseVersion(torch.__version__) >= LooseVersion("2.3"):
    from torch.amp import autocast, GradScaler
else:
    from torch.cuda.amp import autocast, GradScaler

Installation has no torch version constraint

FunASR is a single-package monorepo (one setup.py → funasr; runtime/
holds platform deployments). The install path does not bound torch:

• setup.py install_requires does not list torch at all — pip
  never installs or constrains it.
• README.md:38-44
  instructs pip install torch torchaudio (latest PyPI = torch 2.11),
  then pip install funasr.

So on a fresh install with current PyPI, torch 2.11 is the documented
outcome and the deprecation warnings fire on every autocast/GradScaler
use.

(Also observed: the Triton GPU deployment image
runtime/triton_gpu/Dockerfile/Dockerfile.server
pins pip3 install torch==2.4.1 torchaudio==2.4.1 — pinned to match the
prebuilt kaldifeat CUDA wheel. torch 2.4.1 is itself the first release
where torch.cuda.amp.autocast is deprecated, so even that pinned
deployment ships the deprecated API.)

Files referencing torch.cuda.amp (37 total)

funasr/bin/train.py, funasr/bin/train_ds.py, funasr/models/campplus/model.py,
funasr/models/data2vec/data2vec.py, funasr/models/llm_asr/model.py,
funasr/models/xvector/e2e_sv.py, funasr/models/sanm_kws/model.py,
funasr/models/fsmn_kws/model.py, funasr/models/e_paraformer/pif_predictor.py,
funasr/models/lcbnet/model.py, funasr/models/scama/model.py,
funasr/models/emotion2vec/model.py, funasr/train_utils/trainer.py,
funasr/models/uniasr/model.py, funasr/models/fsmn_kws_mt/model.py,
funasr/train_utils/trainer_ds.py, funasr/models/transducer/model.py,
funasr/models/mfcca/e2e_asr_mfcca.py, funasr/models/sa_asr/e2e_sa_asr.py,
plus 18 more under funasr/models/ and funasr/train_utils/.

[xyf5432]

A note on impact, in case it informs the discussion:

The FutureWarning is not visible to users of two entry points —
funasr/utils/export_utils.py:7 and funasr/bin/realtime_ws.py:32 —
because both call warnings.filterwarnings("ignore") at module level,
which silences all warnings including this one. On the other entry points
(training / inference CLIs, model loading) the warning does show up,
though Python's default filter dedups by message + source line, so users
see it once per distinct call site per process rather than on every call.

The warning noise is not the real problem though — the removal is. The
deprecation notice says these APIs "will be removed in a future release"
(no version pinned; both still exist and work in torch 2.11). When the
removal lands, every from torch.cuda.amp import ... site in this repo
(37 files) raises ImportError at import time — a hard failure that no
warnings.filterwarnings("ignore") can suppress, and one that breaks
every model/trainer entry point at once. Migrating to torch.amp is
about compatibility with that removal, not just warning hygiene.