vrr/koboldcpp
2
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2026-05-22 11:16:08 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 1

fix t5 scale as well

Browse source
This commit is contained in:
Concedo 2025-10-12 20:35:46 +08:00
parent e0ba01c65e
commit 54db35cd7a
1 changed files with 3 additions and 1 deletions
Download patch file Download diff file Expand all files Collapse all files

4
otherarch/sdcpp/t5.hpp
View file

@ -504,7 +504,9 @@ public:
T5DenseGatedActDense(int64_t model_dim, int64_t ff_dim) {
blocks["wi_0"] = std::shared_ptr<GGMLBlock>(new Linear(model_dim, ff_dim, false));
blocks["wi_1"] = std::shared_ptr<GGMLBlock>(new Linear(model_dim, ff_dim, false));
blocks["wo"] = std::shared_ptr<GGMLBlock>(new Linear(ff_dim, model_dim, false));
float scale = 1.f / 32.f;
// The purpose of the scale here is to prevent NaN issues on some backends(CUDA, ...).
blocks["wo"] = std::shared_ptr<GGMLBlock>(new Linear(ff_dim, model_dim, false, false, false, scale));
}
struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {