frozenleaves/pytorch
1
0
Fork 0
mirror of https://github.com/pytorch/pytorch.git synced 2025-11-06 00:54:56 +08:00
Code Issues Packages Projects Releases Wiki Activity

[PyTorch] Avoid initializing storage for empty Optionals

Browse Source 
Pull Request resolved: https://github.com/pytorch/pytorch/pull/78947

We don't need to initialize for the non-constexpr case ever, or in the constexpr case after C++20.

Differential Revision: [D36519379](https://our.internmc.facebook.com/intern/diff/D36519379/)

Approved by: https://github.com/ezyang, https://github.com/malfet
This commit is contained in:
Scott Wolchok
2022-06-07 12:55:21 -07:00
committed by PyTorch MergeBot
parent 65fd0cdddb
commit cf3ce329b5
2 changed files with 21 additions and 15 deletions
Download Patch File Download Diff File Expand all files Collapse all files

19
torch/csrc/distributed/rpc/rpc_agent.cpp
View File

@ -5,20 +5,13 @@ namespace torch {
namespace distributed {
namespace rpc {
namespace {
// WorkerInfo needs to be registered exactly once. Since the op registration
// happens in libtorch_python we wrap the class registration in a helper to make
// sure that if there's multiple copies of Python such as used in torch::deploy
// we only ever register it once.
static std::once_flag workerInfoFlag;
static c10::optional<torch::class_<WorkerInfo>> workerInfo;
} // namespace
RegisterWorkerInfoOnce::RegisterWorkerInfoOnce() {
std::call_once(workerInfoFlag, []() {
workerInfo = torch::class_<WorkerInfo>("dist_rpc", "WorkerInfo")
.def(torch::init<std::string, int64_t>());
});
// WorkerInfo needs to be registered exactly once. Since the op registration
// happens in libtorch_python we wrap the class registration in a helper to make
// sure that if there's multiple copies of Python such as used in torch::deploy
// we only ever register it once.
static auto workerInfo = torch::class_<WorkerInfo>("dist_rpc", "WorkerInfo")
.def(torch::init<std::string, int64_t>());
}
constexpr size_t WorkerInfo::MAX_NAME_LEN;