ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-02-27 16:44:21 +00:00
Code Issues Packages Projects Releases Wiki Activity

Adding SWIGLU unary op (#65)

Browse Source 
* Adding GGML_UNARY_OP_SWIGLU

This commit implements the ggml op and CPU compute
forward. I see ~3-4% speedup of PP-512 for Phi-3.5-mini.

* GGML_UNARY_OP_SWIGLU: CUDA implementation

I observe ~12% speedup for PP-512(Phi-3.5-mini).

* GGML_UNARY_OP_SWIGLU: Metal implementation

We get ~2% speedup for PP-512(Phi-3.5-mini).

* GGML_UNARY_OP_SWIGLU: minor improvement on Metal

* GGML_UNARY_OP_SWIGLU: cleanup

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
This commit is contained in:
Kawrakow
2024-09-28 13:37:25 +03:00
committed by GitHub
parent 1f61e91862
commit 737514fd81
8 changed files with 217 additions and 11 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
src/llama.cpp
View File

@@ -8111,16 +8111,8 @@ static struct ggml_tensor * llm_build_ffn(
} break;
case LLM_FFN_SWIGLU:
{
// Project to 4h. If using swiglu double the output width, see https://arxiv.org/pdf/2002.05202.pdf
int64_t split_point = cur->ne[0] / 2;
struct ggml_tensor * x0 = ggml_cont(ctx, ggml_view_2d(ctx, cur, split_point, cur->ne[1], cur->nb[1], 0));
struct ggml_tensor * x1 = ggml_cont(ctx, ggml_view_2d(ctx, cur, split_point, cur->ne[1], cur->nb[1], split_point * ggml_element_size(cur)));
x0 = ggml_silu(ctx, x0);
cb(cur, "ffn_silu", il);
cur = ggml_mul(ctx, x0, x1);
cb(cur, "ffn_mul", il);
cur = ggml_swiglu(ctx, cur);
cb(cur, "ffn_swiglu", il);
} break;
}