Skip to content

ggml : add ggml_repeat_4d #13824

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 1 commit into from
May 27, 2025
Merged

ggml : add ggml_repeat_4d #13824

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
9 changes: 9 additions & 0 deletions ggml/include/ggml.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -935,6 +935,15 @@ extern "C" {
struct ggml_tensor * a,
struct ggml_tensor * b);

// repeat a to the specified shape
GGML_API struct ggml_tensor * ggml_repeat_4d(
struct ggml_context * ctx,
struct ggml_tensor * a,
int64_t ne0,
int64_t ne1,
int64_t ne2,
int64_t ne3);

// sums repetitions in a into shape of b
GGML_API struct ggml_tensor * ggml_repeat_back(
struct ggml_context * ctx,
Expand Down
20 changes: 20 additions & 0 deletions ggml/src/ggml.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2312,6 +2312,26 @@ struct ggml_tensor * ggml_repeat(
return result;
}

struct ggml_tensor * ggml_repeat_4d(
struct ggml_context * ctx,
struct ggml_tensor * a,
int64_t ne0, int64_t ne1, int64_t ne2, int64_t ne3) {
const bool can_repeat = ggml_is_empty(a) || (
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

What is the purpose of repeating empty tensors?

Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This is to match the behavior of ggml_can_repeat, I think it prevents division by 0 (ne%0) in the conditions below.

In practice, I think it can be useful for one particular use case: in llama.cpp, when we do a warmup, IIRC we have 0 output tokens, so the embedding tensor right before output head has ne[1] == 0

(ne0 % a->ne[0] == 0) &&
(ne1 % a->ne[1] == 0) &&
(ne2 % a->ne[2] == 0) &&
(ne3 % a->ne[3] == 0)
);
GGML_ASSERT(can_repeat);

struct ggml_tensor * result = ggml_new_tensor_4d(ctx, a->type, ne0, ne1, ne2, ne3);

result->op = GGML_OP_REPEAT;
result->src[0] = a;

return result;
}

// ggml_repeat_back

struct ggml_tensor * ggml_repeat_back(
Expand Down
Loading