Public/koboldcpp
1
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2026-09-20 01:31:42 +02:00
Code Issues Packages Projects Releases Wiki Activity

the dark gods have been sated, and redpajama is integrated... but at what cost?

Browse Source
This commit is contained in:
Concedo
2023-05-08 20:58:00 +08:00
parent b9904c3093
commit 2f2eff6e13
4 changed files with 57 additions and 23 deletions
Download Patch File Download Diff File Expand all files Collapse all files
otherarch/neox.cpp
+20 -7
View File
@@ -345,7 +345,8 @@ bool stablelm_eval(
const int n_past,
const std::vector<gpt_vocab::id> & embd_inp,
std::vector<float> & embd_w,
size_t & mem_per_token) {
size_t & mem_per_token,
FileFormat file_format) {
const int N = embd_inp.size();
const auto & hparams = model.hparams;
@@ -494,6 +495,12 @@ bool stablelm_eval(
}
}
if(file_format==FileFormat::NEOX_3)
{
// layer input + Attn
cur = ggml_add(ctx0, cur, inpL);
}
struct ggml_tensor * inpFF = cur;
// feed-forward network
@@ -502,7 +509,7 @@ bool stablelm_eval(
// post attention layer norm
// note here we pass inpL instead of cur
{
cur = ggml_norm(ctx0, inpL);
cur = ggml_norm(ctx0, (file_format==FileFormat::NEOX_3?cur:inpL));
cur = ggml_add(ctx0,
ggml_mul(ctx0,
@@ -533,11 +540,17 @@ bool stablelm_eval(
cur);
}
// layer input + FF
cur = ggml_add(ctx0, cur, inpFF);
// input for next layer
inpL = ggml_add(ctx0, cur, inpL);
if (file_format == FileFormat::NEOX_3)
{
// layer input + FF
inpL = ggml_add(ctx0, cur, inpFF);
}
else
{
cur = ggml_add(ctx0, cur, inpFF);
// input for next layer
inpL = ggml_add(ctx0, cur, inpL);
}
}
// norm