Skip to content

[PyTorch][MOE] Tentative Fix For Replacing from_blob with empty for experts receiving zero tokens #2134

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
merged 2 commits into from
Aug 28, 2025
Merged

[PyTorch][MOE] Tentative Fix For Replacing from_blob with empty for experts receiving zero tokens #2134

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
cf086d2
use torch empty for empty shape instead of from_blob
zhongbozhu Aug 27, 2025
82ecf60
Merge branch 'main' into zhongbo/fix_bulk_alloc
ksivaman Aug 28, 2025
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
14 changes: 4 additions & 10 deletions transformer_engine/pytorch/csrc/extensions/cast.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -205,11 +205,8 @@ std::tuple<std::vector<py::object>, std::vector<TensorWrapper>> bulk_allocate_fp
auto make_torch_view = [](std::shared_ptr<at::Tensor> &buffer, const std::vector<size_t> &shape,
size_t offset, at::ScalarType dtype) -> at::Tensor {
std::vector<int64_t> shape_int64(shape.begin(), shape.end());
// in the case where full buffer is empty because local rank receives no tokens for all the experts
// then the data_ptr is nullptr, we need to return an empty tensor instead of calling from_blob
// but in the case where some experts receive tokens, some not, we want to leverage from_blob
// as much as possible to avoid CPU overhead
if (buffer->data_ptr<uint8_t>() == nullptr) {
bool is_empty_shape = product(shape) == 0;
if (buffer->data_ptr<uint8_t>() == nullptr || is_empty_shape) {
return at::empty(shape_int64, at::device(at::kCUDA).dtype(dtype));
}
return at::from_blob(
Expand Down Expand Up @@ -359,11 +356,8 @@ std::tuple<std::vector<py::object>, std::vector<TensorWrapper>> bulk_allocate_mx
auto make_torch_view = [](std::shared_ptr<at::Tensor> &buffer, const std::vector<size_t> &shape,
size_t offset, at::ScalarType dtype) -> at::Tensor {
std::vector<int64_t> shape_int64(shape.begin(), shape.end());
// in the case where full buffer is empty because local rank receives no tokens for all the experts
// then the data_ptr is nullptr, we need to return an empty tensor instead of calling from_blob
// but in the case where some experts receive tokens, some not, we want to leverage from_blob
// as much as possible to avoid CPU overhead
if (buffer->data_ptr<uint8_t>() == nullptr) {
bool is_empty_shape = product(shape) == 0;
if (buffer->data_ptr<uint8_t>() == nullptr || is_empty_shape) {
return at::empty(shape_int64, at::device(at::kCUDA).dtype(dtype));
}
return at::from_blob(
Expand Down