[Fix] Deprecate torch.cuda.amp API

[MGAMZ]

This is a sub-PR of #1665

Brief

According to PyTorch:

torch.cuda.amp.GradScaler(args...) is deprecated. Please use torch.amp.GradScaler("cuda", args...) instead.

This includes two related replacement:

1. amp_optimizer_wrapper
2. test_optimizer_wrapper

PyTest Result

pytest tests/test_optim/test_optimizer/test_optimizer_wrapper.py

========================================================== test session starts ===========================================================
platform linux -- Python 3.13.9, pytest-8.4.1, pluggy-1.6.0
rootdir: /home/mgam/mgam_repos/mmengine
configfile: pytest.ini
plugins: anyio-4.9.0, hydra-core-1.3.2
collected 37 items                                                                                                                       

tests/test_optim/test_optimizer/test_optimizer_wrapper.py sssssssssssssssssss..................                                    [100%]

============================================================ warnings summary ============================================================
mmengine/utils/misc.py:477
  /home/mgam/mgam_repos/mmengine/mmengine/utils/misc.py:477: DeprecationWarning: 'maxsplit' is passed as positional argument
    summary_and_body = re.split(pattern, docstring, 1)

../../miniforge3/envs/pt/lib/python3.13/site-packages/pydantic/typing.py:400
  /home/mgam/miniforge3/envs/pt/lib/python3.13/site-packages/pydantic/typing.py:400: DeprecationWarning: Failing to pass a value to the 'type_params' parameter of 'typing._eval_type' is deprecated, as it leads to incorrect behaviour when calling typing._eval_type on a stringified annotation that references a PEP 695 type parameter. It will be disallowed in Python 3.15.
    value = _eval_type(value, base_globals, None)

tests/test_optim/test_optimizer/test_optimizer_wrapper.py::TestAmpOptimWrapper::test_optim_context_4
  /home/mgam/mgam_repos/mmengine/mmengine/runner/amp.py:119: DeprecationWarning: torch.get_autocast_gpu_dtype() is deprecated. Please use torch.get_autocast_dtype('cuda') instead. (Triggered internally at /pytorch/torch/csrc/autograd/init.cpp:887.)
    dtype = torch.get_autocast_gpu_dtype()

-- Docs: https://docs.pytest.org/en/stable/how-to/capture-warnings.html
========================================= 18 passed, 19 skipped, 3 warnings in 86.75s (0:01:26) ==========================================

[Copilot]

Pull Request Overview

This PR updates the codebase to use the new torch.amp.GradScaler API instead of the deprecated torch.cuda.amp.GradScaler API, following PyTorch's deprecation guidance.

Key Changes:

• Replaced deprecated torch.cuda.amp.GradScaler imports with torch.amp.GradScaler using a partial function to set the device to 'cuda'
• Applied this change to both production code (amp_optimizer_wrapper.py) and test code (test_optimizer_wrapper.py)

Reviewed Changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 3 comments.

File	Description
mmengine/optim/optimizer/amp_optimizer_wrapper.py	Updated GradScaler import from deprecated torch.cuda.amp to torch.amp with device parameter
tests/test_optim/test_optimizer/test_optimizer_wrapper.py	Updated GradScaler import in tests to use the new torch.amp API

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

The blank line after from functools import partial creates inconsistent import grouping. Move the functools import to be with other standard library imports (os, unittest) before the blank line that separates standard library imports from third-party imports.

Suggested change

	from functools import partial
	import unittest
	from unittest import TestCase
	from unittest.mock import MagicMock
	import unittest
	from unittest import TestCase
	from unittest.mock import MagicMock
	from functools import partial

[Copilot]

[nitpick] Creating a module-level variable GradScaler through partial assignment makes the code less maintainable and harder to understand. Consider either: (1) using amp_GradScaler('cuda', ...) directly at call sites, or (2) creating a proper wrapper function with a docstring explaining the device binding.

Suggested change

	GradScaler = partial(amp_GradScaler, device='cuda')
	def get_cuda_grad_scaler(*args, **kwargs):
	"""Return a torch.amp.GradScaler instance bound to the 'cuda' device.
	Args:
	*args: Positional arguments for torch.amp.GradScaler.
	**kwargs: Keyword arguments for torch.amp.GradScaler.
	Returns:
	amp_GradScaler: An instance of GradScaler with device='cuda'.
	"""
	return amp_GradScaler(*args, device='cuda', **kwargs)

[Copilot]

[nitpick] Creating a module-level variable GradScaler through partial assignment makes the code less maintainable and harder to understand. Consider either: (1) using amp_GradScaler('cuda', ...) directly at call sites, or (2) creating a proper wrapper function with a docstring explaining the device binding.

Suggested change

	GradScaler = partial(amp_GradScaler, device='cuda')
	def get_grad_scaler(*args, **kwargs):
	"""Create a torch.amp.GradScaler instance bound to device='cuda'.
	Args:
	*args: Positional arguments passed to torch.amp.GradScaler.
	**kwargs: Keyword arguments passed to torch.amp.GradScaler.
	Returns:
	amp_GradScaler: An instance of torch.amp.GradScaler with device='cuda'.
	"""
	return amp_GradScaler(*args, device='cuda', **kwargs)

[MGAMZ]

@HAOCHENYE This one is ready to be reviewed.

The copilot review suggested a different implementation which also LGTM. You may choose one to merge~~

[HAOCHENYE]

This workaround LGTM, I'll merge it after fixing the lint