vrr/koboldcpp
2
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2025-09-13 02:19:41 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

switched to NTK aware scaling

Browse source
This commit is contained in:
Concedo 2023-07-02 17:25:08 +08:00
parent e19483ca0f
commit e17c8497cf
4 changed files with 26 additions and 25 deletions
Download patch file Download diff file Expand all files Collapse all files

8
ggml.h
View file

@ -201,12 +201,6 @@
#define GGML_MAX_NAME 48
#define GGML_DEFAULT_N_THREADS 4
// Maximum training context of the model in use
// For the LLaMA models this is normally 2048, but somehow "stepping out" by 128 gives better results (tested at 7B and 13B)
#ifndef GGML_TRAINING_CTX
#define GGML_TRAINING_CTX 2176
#endif
#define GGML_ASSERT(x) \
do { \
if (!(x)) { \
@ -510,6 +504,8 @@ extern "C" {
// use this to compute the memory overhead of a tensor
GGML_API size_t ggml_tensor_overhead(void);
GGML_API float get_theta_scale(int n_dims,int n_past,int n_ctx);
// main
GGML_API struct ggml_context * ggml_init(struct ggml_init_params params);