Skip to content

[BIT] Fix paddle.unstack api for big tensor #72931

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
May 30, 2025
Merged

[BIT] Fix paddle.unstack api for big tensor #72931

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
1980035
fix unstack big tensor
cszdrg May 26, 2025
7a5873c
Merge branch 'develop' of https://github.com/cszdrg/Paddle into develop
cszdrg May 29, 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
4 changes: 2 additions & 2 deletions paddle/phi/infermeta/backward.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1747,8 +1747,8 @@ void UnStackGradInferMeta(const std::vector<const MetaTensor*>& out_grad,
rank));
if (axis < 0) axis += (rank + 1);

auto vec = common::vectorize<int>(input_dims[0]);
vec.insert(vec.begin() + axis, static_cast<int>(input_dims.size()));
auto vec = common::vectorize<int64_t>(input_dims[0]);
vec.insert(vec.begin() + axis, static_cast<int64_t>(input_dims.size()));
x_grad->set_dims(common::make_ddim(vec));
x_grad->set_dtype(out_grad[0]->dtype());
}
Expand Down
2 changes: 1 addition & 1 deletion paddle/phi/infermeta/unary.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -6075,7 +6075,7 @@ void UnStackInferMeta(const MetaTensor& x,
x_dim[axis],
num));
}
auto vec = common::vectorize<int>(x_dim);
auto vec = common::vectorize<int64_t>(x_dim);
vec.erase(vec.begin() + axis);
for (size_t i = 0; i < output_count; i++) {
outs[i]->set_dims(common::make_ddim(vec));
Expand Down
11 changes: 7 additions & 4 deletions paddle/phi/kernels/funcs/stack_and_unstack.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -210,7 +210,7 @@ void LaunchUnStackKernel(const Context& ctx,
constexpr int kWarpSize = 32;
constexpr int kMaxOut = 16;

int tid_x = 0, tid_y = 0, bid_x = 0, bid_y = 1;
int64_t tid_x = 0, tid_y = 0, bid_x = 0, bid_y = 1;
if (split_dim < kMaxOut) {
tid_y = split_dim;
tid_x =
Expand All @@ -219,10 +219,13 @@ void LaunchUnStackKernel(const Context& ctx,
} else {
tid_y = kMaxOut;
tid_x = kWarpSize;
bid_y = backends::gpu::DivUp<int>(split_dim, kMaxOut);
bid_y = backends::gpu::DivUp<int64_t>(split_dim, kMaxOut);
}
int tile_x_num = backends::gpu::DivUp<int>(out_row, tid_x);
bid_x = std::min(tile_x_num, backends::gpu::kMultiDimslimit);
int64_t tile_x_num = backends::gpu::DivUp<int64_t>(out_row, tid_x);
if (tile_x_num < static_cast<int64_t>(backends::gpu::kMultiDimslimit))
bid_x = tile_x_num;
else
bid_x = backends::gpu::kMultiDimslimit;
dim3 blocks(tid_x, tid_y, 1);
dim3 grids(bid_x, bid_y, 1);

Expand Down
12 changes: 6 additions & 6 deletions paddle/phi/kernels/impl/unstack_grad_kernel_impl.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -28,13 +28,13 @@ void UnStackGradKernel(const Context &dev_ctx,
DenseTensor *x_grad) {
if (axis < 0) axis += (x[0]->dims().size() + 1);

int n = static_cast<int>(x.size());
int64_t n = static_cast<int64_t>(x.size());
auto *x_grad_data = dev_ctx.template Alloc<T>(x_grad);
std::vector<const T *> x_datas(n);
for (int i = 0; i < n; i++) x_datas[i] = x[i]->data<T>();
for (int64_t i = 0; i < n; i++) x_datas[i] = x[i]->data<T>();

int pre = 1;
int post = 1;
int64_t pre = 1;
int64_t post = 1;
auto &dim = x[0]->dims();
for (auto i = 0; i < axis; ++i) pre *= dim[i];
for (auto i = axis; i < dim.size(); ++i) post *= dim[i];
Expand All @@ -56,8 +56,8 @@ void UnStackGradKernel(const Context &dev_ctx,

size_t x_offset = 0;
size_t y_offset = 0;
for (int i = 0; i < pre; i++) {
for (int j = 0; j < n; j++) {
for (int64_t i = 0; i < pre; i++) {
for (int64_t j = 0; j < n; j++) {
std::memcpy(
x_grad_data + y_offset, x_data_arr[j] + x_offset, post * sizeof(T));
y_offset += post;
Expand Down
Loading