Set pytorch cuda arch list explicitly to get rid of warning.

[hzhangxyz]

Description

Set pytorch cuda arch list explicitly to get rid of warning.

The pytorch raises warning when compiling before:

.../site-packages/torch/utils/cpp_extension.py:2059: UserWarning: TORCH_CUDA_ARCH_LIST is not set, all archs for visible cards are included for compilation.
If this is not desired, please set os.environ['TORCH_CUDA_ARCH_LIST'].
  warnings.warn(

Checklist:

• I have read the CONTRIBUTING.md.

[Copilot]

Pull Request Overview

This PR adds logic to explicitly set TORCH_CUDA_ARCH_LIST based on available GPU capabilities to remove the warning when building PyTorch extensions.

• Introduce _set_torch_cuda_arch_list helper to compute and export the CUDA arch list.
• Invoke this helper at the start of _load_module.

Comments suppressed due to low confidence (1)

qmb/hamiltonian.py:20

• The new _set_torch_cuda_arch_list logic isn’t covered by existing tests. Adding unit tests to validate behavior for various torch.cuda scenarios (no GPUs, single GPU, multiple GPUs) would help prevent regressions.

    def _set_torch_cuda_arch_list(cls) -> None:

[Copilot]

If supported_sm is empty, calling max() will raise a ValueError. Consider guarding against an empty list (e.g. return early) to avoid a crash when no arch entries are found.

Suggested change

	supported_sm = [int(arch[3:]) for arch in torch.cuda.get_arch_list() if arch.startswith("sm_")]
	supported_sm = [int(arch[3:]) for arch in torch.cuda.get_arch_list() if arch.startswith("sm_")]
	if not supported_sm:
	return # No supported architectures, exit early

[hzhangxyz]

torch.cuda.is_available() 之前判断过，不会空

[Copilot]

Sorting the version strings lexicographically can misorder multi-digit majors (e.g. "10.2" vs "9.1"). Use a numeric key (e.g. sorted(arch_list, key=lambda v: tuple(map(int, v.split('.'))))) to ensure correct ascending order.

Suggested change

	os.environ["TORCH_CUDA_ARCH_LIST"] = ";".join(sorted(arch_list))
	os.environ["TORCH_CUDA_ARCH_LIST"] = ";".join(sorted(arch_list, key=lambda v: tuple(map(int, v.split('.')))))

[msg-bq]

这里copilot给的提示是对的吧，不像上一个没必要。师兄的版本号是str，字典序不妥，还得是像(sm // 10, sm % 10)那种拆开的叭。虽然排序无妨

[hzhangxyz]

确实，这里排序是为了每次运行他们的顺序是一致的，不然无法有效利用缓存。

另外，我这段代码抄的:

https://github.com/pytorch/pytorch/blob/0d01bafc34fc99a0b3e46cbf1ecfd8f97563efa6/torch/utils/cpp_extension.py#L2428-L2444