Skip to content

[XPU] fix all_to_all with unequal splits when input/output is empty #72005

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
Apr 1, 2025
Merged

[XPU] fix all_to_all with unequal splits when input/output is empty #72005

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
ec6b2ab
[XPU] fix all_to_all with unequal splits when input/output is empty
dynamicheart Mar 31, 2025
25ba5ab
fix
dynamicheart Mar 31, 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
298 changes: 147 additions & 151 deletions paddle/fluid/distributed/collective/process_group_bkcl.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -466,74 +466,72 @@ std::shared_ptr<ProcessGroup::Task> ProcessGroupBKCL::AllToAll(

int64_t nranks = size_;

if (in_row_size > 0 && out_row_size > 0) {
std::vector<int64_t> in_numel_vec(nranks);
std::vector<int64_t> in_offset_vec(nranks);
std::vector<int64_t> out_numel_vec(nranks);
std::vector<int64_t> out_offset_vec(nranks);

int64_t in_offset = 0;
int64_t out_offset = 0;
for (int64_t i = 0; i < nranks; i++) {
int64_t in_numel = in_split_sizes[i] * in_row_size;
int64_t out_numel = out_split_sizes[i] * out_row_size;

in_numel_vec[i] = in_numel;
in_offset_vec[i] = in_offset;
in_offset += in_numel;

out_numel_vec[i] = out_numel;
out_offset_vec[i] = out_offset;
out_offset += out_numel;
}

PADDLE_ENFORCE_GE(
in_tensor.place().GetDeviceId(),
0,
common::errors::PreconditionNotMet(
"The all_to_all device id must greater or equal than 0."));
phi::XPUPlace place = in_tensor.place();
auto allocator = std::unique_ptr<phi::Allocator>(
new paddle::experimental::DefaultAllocator(place));
phi::DenseTensorMeta meta(phi::DataType::INT64, phi::DDim{nranks});

phi::DenseTensor in_size_tensor = {allocator.get(), meta};
phi::DenseTensor in_offset_tensor = {allocator.get(), meta};
phi::DenseTensor out_size_tensor = {allocator.get(), meta};
phi::DenseTensor out_offset_tensor = {allocator.get(), meta};

memory::Copy(place,
in_size_tensor.data(),
phi::CPUPlace(),
in_numel_vec.data(),
in_size_tensor.numel() * sizeof(int64_t));

memory::Copy(place,
in_offset_tensor.data(),
phi::CPUPlace(),
in_offset_vec.data(),
in_offset_tensor.numel() * sizeof(int64_t));

memory::Copy(place,
out_size_tensor.data(),
phi::CPUPlace(),
out_numel_vec.data(),
out_size_tensor.numel() * sizeof(int64_t));

memory::Copy(place,
out_offset_tensor.data(),
phi::CPUPlace(),
out_offset_vec.data(),
out_offset_tensor.numel() * sizeof(int64_t));

comm_context->AllToAllUnequalSplit(out_tensor,
in_tensor,
out_size_tensor,
out_offset_tensor,
in_size_tensor,
in_offset_tensor,
stream);
std::vector<int64_t> in_numel_vec(nranks);
std::vector<int64_t> in_offset_vec(nranks);
std::vector<int64_t> out_numel_vec(nranks);
std::vector<int64_t> out_offset_vec(nranks);

int64_t in_offset = 0;
int64_t out_offset = 0;
for (int64_t i = 0; i < nranks; i++) {
int64_t in_numel = in_split_sizes[i] * in_row_size;
int64_t out_numel = out_split_sizes[i] * out_row_size;

in_numel_vec[i] = in_numel;
in_offset_vec[i] = in_offset;
in_offset += in_numel;

out_numel_vec[i] = out_numel;
out_offset_vec[i] = out_offset;
out_offset += out_numel;
}

PADDLE_ENFORCE_GE(
in_tensor.place().GetDeviceId(),
0,
common::errors::PreconditionNotMet(
"The all_to_all device id must greater or equal than 0."));
phi::XPUPlace place = in_tensor.place();
auto allocator = std::unique_ptr<phi::Allocator>(
new paddle::experimental::DefaultAllocator(place));
phi::DenseTensorMeta meta(phi::DataType::INT64, phi::DDim{nranks});

phi::DenseTensor in_size_tensor = {allocator.get(), meta};
phi::DenseTensor in_offset_tensor = {allocator.get(), meta};
phi::DenseTensor out_size_tensor = {allocator.get(), meta};
phi::DenseTensor out_offset_tensor = {allocator.get(), meta};

memory::Copy(place,
in_size_tensor.data(),
phi::CPUPlace(),
in_numel_vec.data(),
in_size_tensor.numel() * sizeof(int64_t));

memory::Copy(place,
in_offset_tensor.data(),
phi::CPUPlace(),
in_offset_vec.data(),
in_offset_tensor.numel() * sizeof(int64_t));

memory::Copy(place,
out_size_tensor.data(),
phi::CPUPlace(),
out_numel_vec.data(),
out_size_tensor.numel() * sizeof(int64_t));

memory::Copy(place,
out_offset_tensor.data(),
phi::CPUPlace(),
out_offset_vec.data(),
out_offset_tensor.numel() * sizeof(int64_t));

comm_context->AllToAllUnequalSplit(out_tensor,
in_tensor,
out_size_tensor,
out_offset_tensor,
in_size_tensor,
in_offset_tensor,
stream);
}
},
in_tensor,
Expand Down Expand Up @@ -614,95 +612,93 @@ std::shared_ptr<ProcessGroup::Task> ProcessGroupBKCL::AllToAll(
out_numel_sum += (*out_tensors)[i].numel();
}

if (in_numel_sum > 0 || out_numel_sum > 0) {
std::vector<int64_t> in_numel_vec(nranks);
std::vector<int64_t> in_offset_vec(nranks);
std::vector<int64_t> out_numel_vec(nranks);
std::vector<int64_t> out_offset_vec(nranks);

int64_t in_offset = 0;
int64_t out_offset = 0;
for (int64_t i = 0; i < nranks; i++) {
int64_t in_numel = in_tensors[i].numel();
int64_t out_numel = (*out_tensors)[i].numel();

in_numel_vec[i] = in_numel;
in_offset_vec[i] = in_offset;
in_offset += in_numel;

out_numel_vec[i] = out_numel;
out_offset_vec[i] = out_offset;
out_offset += out_numel;
}

PADDLE_ENFORCE_GE(
in_tensors[0].place().GetDeviceId(),
0,
common::errors::PreconditionNotMet(
"The all_to_all device id must greater or equal than 0."));
phi::XPUPlace place = in_tensors[0].place();
auto allocator = std::unique_ptr<phi::Allocator>(
new paddle::experimental::DefaultAllocator(place));

phi::DenseTensorMeta concated_in_tensor_meta(in_tensors[0].dtype(),
phi::DDim{in_numel_sum});
phi::DenseTensorMeta concated_out_tensor_meta(
(*out_tensors)[0].dtype(), phi::DDim{out_numel_sum});
phi::DenseTensorMeta split_meta(phi::DataType::INT64,
phi::DDim{nranks});

phi::DenseTensor concated_in_tensor = {allocator.get(),
concated_in_tensor_meta};
phi::DenseTensor concated_out_tensor = {allocator.get(),
concated_out_tensor_meta};
phi::DenseTensor in_size_tensor = {allocator.get(), split_meta};
phi::DenseTensor in_offset_tensor = {allocator.get(), split_meta};
phi::DenseTensor out_size_tensor = {allocator.get(), split_meta};
phi::DenseTensor out_offset_tensor = {allocator.get(), split_meta};

if (in_numel_sum > 0) {
ConcatTensorByNumel(*GetDeviceContext(place, use_calc_stream),
in_tensors,
&concated_in_tensor);
}
std::vector<int64_t> in_numel_vec(nranks);
std::vector<int64_t> in_offset_vec(nranks);
std::vector<int64_t> out_numel_vec(nranks);
std::vector<int64_t> out_offset_vec(nranks);

memory::Copy(place,
in_size_tensor.data(),
phi::CPUPlace(),
in_numel_vec.data(),
in_size_tensor.numel() * sizeof(int64_t));

memory::Copy(place,
in_offset_tensor.data(),
phi::CPUPlace(),
in_offset_vec.data(),
in_offset_tensor.numel() * sizeof(int64_t));
int64_t in_offset = 0;
int64_t out_offset = 0;
for (int64_t i = 0; i < nranks; i++) {
int64_t in_numel = in_tensors[i].numel();
int64_t out_numel = (*out_tensors)[i].numel();

memory::Copy(place,
out_size_tensor.data(),
phi::CPUPlace(),
out_numel_vec.data(),
out_size_tensor.numel() * sizeof(int64_t));
in_numel_vec[i] = in_numel;
in_offset_vec[i] = in_offset;
in_offset += in_numel;

memory::Copy(place,
out_offset_tensor.data(),
phi::CPUPlace(),
out_offset_vec.data(),
out_offset_tensor.numel() * sizeof(int64_t));
out_numel_vec[i] = out_numel;
out_offset_vec[i] = out_offset;
out_offset += out_numel;
}

comm_context->AllToAllUnequalSplit(&concated_out_tensor,
concated_in_tensor,
out_size_tensor,
out_offset_tensor,
in_size_tensor,
in_offset_tensor,
stream);
PADDLE_ENFORCE_GE(
in_tensors[0].place().GetDeviceId(),
0,
common::errors::PreconditionNotMet(
"The all_to_all device id must greater or equal than 0."));
phi::XPUPlace place = in_tensors[0].place();
auto allocator = std::unique_ptr<phi::Allocator>(
new paddle::experimental::DefaultAllocator(place));

phi::DenseTensorMeta concated_in_tensor_meta(in_tensors[0].dtype(),
phi::DDim{in_numel_sum});
phi::DenseTensorMeta concated_out_tensor_meta((*out_tensors)[0].dtype(),
phi::DDim{out_numel_sum});
phi::DenseTensorMeta split_meta(phi::DataType::INT64,
phi::DDim{nranks});

phi::DenseTensor concated_in_tensor = {allocator.get(),
concated_in_tensor_meta};
phi::DenseTensor concated_out_tensor = {allocator.get(),
concated_out_tensor_meta};
phi::DenseTensor in_size_tensor = {allocator.get(), split_meta};
phi::DenseTensor in_offset_tensor = {allocator.get(), split_meta};
phi::DenseTensor out_size_tensor = {allocator.get(), split_meta};
phi::DenseTensor out_offset_tensor = {allocator.get(), split_meta};

if (in_numel_sum > 0) {
ConcatTensorByNumel(*GetDeviceContext(place, use_calc_stream),
in_tensors,
&concated_in_tensor);
}

if (out_numel_sum > 0) {
SplitTensorByNumel(*GetDeviceContext(place, use_calc_stream),
concated_out_tensor,
out_tensors);
}
memory::Copy(place,
in_size_tensor.data(),
phi::CPUPlace(),
in_numel_vec.data(),
in_size_tensor.numel() * sizeof(int64_t));

memory::Copy(place,
in_offset_tensor.data(),
phi::CPUPlace(),
in_offset_vec.data(),
in_offset_tensor.numel() * sizeof(int64_t));

memory::Copy(place,
out_size_tensor.data(),
phi::CPUPlace(),
out_numel_vec.data(),
out_size_tensor.numel() * sizeof(int64_t));

memory::Copy(place,
out_offset_tensor.data(),
phi::CPUPlace(),
out_offset_vec.data(),
out_offset_tensor.numel() * sizeof(int64_t));

comm_context->AllToAllUnequalSplit(&concated_out_tensor,
concated_in_tensor,
out_size_tensor,
out_offset_tensor,
in_size_tensor,
in_offset_tensor,
stream);

if (out_numel_sum > 0) {
SplitTensorByNumel(*GetDeviceContext(place, use_calc_stream),
concated_out_tensor,
out_tensors);
}
},
in_tensors,
Expand Down
29 changes: 20 additions & 9 deletions paddle/fluid/distributed/collective/process_group_kernel_utils.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -29,6 +29,10 @@ struct ConcatDenseTensorByNumel {
void operator()(const DeviceContext &context,
const std::vector<phi::DenseTensor> &in,
phi::DenseTensor *out) {
if (out->numel() == 0) {
return;
}

auto out_dims = common::vectorize(out->dims());
auto flattened_out_dims = {out->numel()};
std::vector<phi::DenseTensor> in_flatten;
Expand All @@ -39,11 +43,12 @@ struct ConcatDenseTensorByNumel {

int64_t in_numel_sum = 0;
for (auto &tensor : in) {
phi::DenseTensor tensor_flatten(tensor.Holder(), tensor.meta());
tensor_flatten.Resize({tensor.numel()});
in_flatten.push_back(tensor_flatten);

in_numel_sum += tensor.numel();
if (tensor.numel() > 0) {
phi::DenseTensor tensor_flatten(tensor.Holder(), tensor.meta());
tensor_flatten.Resize({tensor.numel()});
in_flatten.push_back(tensor_flatten);
in_numel_sum += tensor.numel();
}
}
PADDLE_ENFORCE_EQ(
out->numel(),
Expand Down Expand Up @@ -105,6 +110,10 @@ struct SplitDenseTensorByNumel {
void operator()(const DeviceContext &context,
const phi::DenseTensor &in,
std::vector<phi::DenseTensor> *out) {
if (in.numel() == 0) {
return;
}

phi::DenseTensor in_flatten(in.Holder(), in.meta());
in_flatten.Resize({in.numel()});

Expand All @@ -115,10 +124,12 @@ struct SplitDenseTensorByNumel {
int64_t out_numel_sum = 0;

for (auto &tensor : *out) {
phi::DenseTensor tensor_flatten(tensor.Holder(), tensor.meta());
tensor_flatten.Resize({tensor.numel()});
out_flatten.push_back(tensor_flatten);
out_numel_sum += tensor.numel();
if (tensor.numel() > 0) {
phi::DenseTensor tensor_flatten(tensor.Holder(), tensor.meta());
tensor_flatten.Resize({tensor.numel()});
out_flatten.push_back(tensor_flatten);
out_numel_sum += tensor.numel();
}
}
for (auto &tensor : out_flatten) {
shape_refer.push_back(&tensor);
Expand Down
Loading
Loading