Skip to content

[PyTorch] Fix GIL/refcount abort in Comm+GEMM overlap and NCCL-EP init bindings #3203

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
denera merged 1 commit into from
Jul 10, 2026
Merged

[PyTorch] Fix GIL/refcount abort in Comm+GEMM overlap and NCCL-EP init bindings #3203

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
11 changes: 8 additions & 3 deletions transformer_engine/pytorch/csrc/extensions/ep.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -145,7 +145,13 @@ void ep_initialize(uintptr_t comm_ptr, const std::string& group_name, int64_t nu
.max_token_dtype = static_cast<NVTEDType>(GetTransformerEngineDType(torch_dtype)),
.zero_copy = zero_copy ? 1 : 0,
};
nvte_ep_initialize(static_cast<void*>(ep_comm), &cfg);
// Release the GIL only around the native init. It must stay held while pybind11 casts
// the ``max_token_dtype`` object above and destroys the by-value ``pybind11::object``
// parameter on return; releasing it across those trips pybind11's dec_ref GIL assertion.
{
pybind11::gil_scoped_release nogil;
nvte_ep_initialize(static_cast<void*>(ep_comm), &cfg);
}
g_zero_copy_enabled.store(zero_copy, std::memory_order_relaxed);
g_ep_initialized = true;
g_ep_group_name = group_name;
Expand Down Expand Up @@ -366,8 +372,7 @@ void register_ep_bindings(pybind11::module_& m) {
"Initialize the EP backend; borrows torch's NCCL comm pointed to by ``comm_ptr``.",
py::arg("comm_ptr"), py::arg("group_name"), py::arg("num_experts"),
py::arg("max_tokens_per_rank"), py::arg("max_recv_tokens_per_rank"), py::arg("hidden_dim"),
py::arg("max_num_sms") = 0, py::arg("max_token_dtype"), py::arg("zero_copy") = false,
py::call_guard<py::gil_scoped_release>());
py::arg("max_num_sms") = 0, py::arg("max_token_dtype"), py::arg("zero_copy") = false);
m.def("ep_finalize", &ep_finalize, "Tear down the EP backend. Idempotent.",
py::call_guard<py::gil_scoped_release>());
m.def("ep_get_zero_copy", &ep_get_zero_copy, "Return the current EP zero-copy toggle state.");
Expand Down
21 changes: 13 additions & 8 deletions transformer_engine/pytorch/csrc/extensions/pybind.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -716,6 +716,9 @@ PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
int num_max_streams, int comm_cga_size, int gemm_priority, int comm_priority,
int num_comm_sm, bool set_sm_margin, bool atomic_gemm,
bool rs_overlap_first_gemm) {
// Release the GIL only around the native construction (blocking collectives) to avoid
// tripping pybind11's inc_ref/dec_ref GIL assertions.
py::gil_scoped_release nogil;
if (use_cublasmp) {
return std::make_shared<CommOverlap>(helper, helper->mylocal, tp_size, comm_type,
buffer_shape, buffer_dtype, num_comm_sm,
Expand All @@ -726,8 +729,7 @@ PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
comm_cga_size, gemm_priority, comm_priority, num_comm_sm, set_sm_margin,
atomic_gemm, rs_overlap_first_gemm);
}),
py::call_guard<py::gil_scoped_release>(), py::arg("buffer_shape"),
py::arg("buffer_dtype"), py::arg("helper"), py::arg("tp_size"),
py::arg("buffer_shape"), py::arg("buffer_dtype"), py::arg("helper"), py::arg("tp_size"),
py::arg("use_cublasmp") = false,
py::arg("comm_type") = transformer_engine::CommOverlapType::RS,
py::arg("num_splits") = 4, py::arg("num_max_streams") = NVTE_COMM_OVERLAP_MAX_STREAMS,
Expand All @@ -751,6 +753,9 @@ PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
int comm_cga_size, int gemm_priority, int comm_priority, int num_comm_sm,
bool set_sm_margin, bool atomic_gemm, bool use_ce, bool aggregate,
bool use_cublasmp) {
// Release the GIL only around the native construction (blocking collectives) to avoid
// tripping pybind11's inc_ref/dec_ref GIL assertions.
py::gil_scoped_release nogil;
if (use_cublasmp) {
return std::make_shared<CommOverlapP2P>(helper, helper->mylocal, tp_size, comm_type,
buffer_shape, buffer_dtype, num_comm_sm,
Expand All @@ -761,12 +766,12 @@ PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
gemm_priority, comm_priority, num_comm_sm,
set_sm_margin, atomic_gemm, use_ce, aggregate);
}),
py::call_guard<py::gil_scoped_release>(), py::arg("buffer_shape"),
py::arg("buffer_dtype"), py::arg("helper"), py::arg("tp_size"), py::arg("comm_type"),
py::arg("num_max_streams") = NVTE_COMM_OVERLAP_MAX_STREAMS, py::arg("comm_cga_size") = 1,
py::arg("gemm_priority") = 0, py::arg("comm_priority") = 0, py::arg("num_comm_sm") = 1,
py::arg("set_sm_margin") = false, py::arg("atomic_gemm") = false,
py::arg("use_ce") = true, py::arg("aggregate") = false, py::arg("use_cublasmp") = false)
py::arg("buffer_shape"), py::arg("buffer_dtype"), py::arg("helper"), py::arg("tp_size"),
py::arg("comm_type"), py::arg("num_max_streams") = NVTE_COMM_OVERLAP_MAX_STREAMS,
py::arg("comm_cga_size") = 1, py::arg("gemm_priority") = 0, py::arg("comm_priority") = 0,
py::arg("num_comm_sm") = 1, py::arg("set_sm_margin") = false,
py::arg("atomic_gemm") = false, py::arg("use_ce") = true, py::arg("aggregate") = false,
py::arg("use_cublasmp") = false)
.def("copy_into_buffer",
static_cast<void (CommOverlapP2P::*)(const at::Tensor &, bool)>(
&CommOverlapP2P::copy_into_buffer),
Expand Down
Loading