Skip to content

issue/1193 - Add a forward_ function to the Embedding module. #1226

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

Draft
pengcheng888 wants to merge 1 commit into
base: main
Choose a base branch
from
Draft

issue/1193 - Add a forward_ function to the Embedding module. #1226

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
2 changes: 2 additions & 0 deletions include/infinicore/nn/embedding.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -60,6 +60,8 @@ class Embedding : public Module {
* Input shape: [10] -> Output shape: [10, embedding_dim]
*/
Tensor forward(const Tensor &indices) const;

Tensor forward_(Tensor &output, const Tensor &indices) const;

// Module information
size_t num_embeddings() const { return num_embeddings_; }
Expand Down
25 changes: 17 additions & 8 deletions src/infinicore/nn/embedding.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -43,13 +43,6 @@ Embedding::Embedding(size_t num_embeddings,
}

Tensor Embedding::forward(const Tensor &indices) const {
// TODO: Implement on-device embedding for all devices, then remove the condition and the classic approach
auto device_type = device_.getType();
if (device_type == Device::Type::NVIDIA || device_type == Device::Type::ILUVATAR || device_type == Device::Type::METAX || device_type == Device::Type::MOORE || device_type == Device::Type::ALI || device_type == Device::Type::QY) {
// Use op::embedding which supports device-side input and batch dimension
return op::embedding(indices->contiguous()->to(device_), weight_);
}

// Get the shape of indices
auto indices_shape = indices->shape();

Expand All @@ -58,12 +51,28 @@ Tensor Embedding::forward(const Tensor &indices) const {
output_shape.push_back(embedding_dim_);

// Create output tensor on the same device as weight
auto out = Tensor::empty(output_shape, weight_->dtype(), weight_->device());
auto output_embeds = Tensor::empty(output_shape, weight_->dtype(), weight_->device());

return this->forward_(output_embeds, indices);
}

Tensor Embedding::forward_(Tensor &output_embeds, const Tensor &indices) const {
// TODO: Implement on-device embedding for all devices, then remove the condition and the classic approach
auto device_type = device_.getType();
if (device_type == Device::Type::NVIDIA || device_type == Device::Type::ILUVATAR || device_type == Device::Type::METAX || device_type == Device::Type::MOORE || device_type == Device::Type::ALI || device_type == Device::Type::QY) {
// Use op::embedding which supports device-side input and batch dimension
op::embedding_(output_embeds, indices->contiguous()->to(device_), weight_);
return output_embeds;
}

auto out = output_embeds;
// Flatten indices for sequential row copies
auto cpu_device = Device(Device::Type::CPU, 0);
auto indices_cpu = indices->to(cpu_device)->contiguous();

// Get the shape of indices
auto indices_shape = indices->shape();

// Calculate total number of lookups
size_t num_lookups = 1;
for (auto dim : indices_shape) {
Expand Down
Loading