vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

ggml : add GGML_PAD_REFLECT_1D operation (ggml/1034)

Browse source 
* ggml_pad_reflect_1d defined in header

* implemented on CPU

* called the forward pass

* impl Metal kernel

* added Metal kernel

* added OP_PAD_REFLECT_1D in test-backend-ops.cpp

* add test-pad-reflect-1d test case

* test case support multiple backend
This commit is contained in:
PAB 2024-12-03 20:20:04 +01:00 • committed by Georgi Gerganov
parent d405804be8
commit c2082d93a8
No known key found for this signature in database
GPG key ID: 449E073F9DC10735
6 changed files with 192 additions and 2 deletions
Download patch file Download diff file Expand all files Collapse all files

8
ggml/include/ggml.h
View file

@ -499,6 +499,7 @@ extern "C" {
GGML_OP_POOL_2D_BACK,
GGML_OP_UPSCALE, // nearest interpolate
GGML_OP_PAD,
GGML_OP_PAD_REFLECT_1D,
GGML_OP_ARANGE,
GGML_OP_TIMESTEP_EMBEDDING,
GGML_OP_ARGSORT,
@ -1695,6 +1696,13 @@ extern "C" {
int p2,
int p3);
// pad each dimension with reflection: [a, b, c, d] -> [b, a, b, c, d, c]
GGML_API struct ggml_tensor * ggml_pad_reflect_1d(
struct ggml_context * ctx,
struct ggml_tensor * a,
int p0,
int p1);
// Ref: https://github.com/CompVis/stable-diffusion/blob/main/ldm/modules/diffusionmodules/util.py#L151
// timesteps: [N,]
// return: [N, dim]