vrr/koboldcpp
2
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2025-09-10 17:14:36 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

revert auto rope scaling for already-ropetuned models - just use their values

Browse source
This commit is contained in:
Concedo 2023-11-24 14:20:36 +08:00
parent 56a5fa7a60
commit eb42c73953
3 changed files with 70 additions and 46 deletions
Download patch file Download diff file Expand all files Collapse all files

24
gpttype_adapter.cpp
View file

@ -941,19 +941,19 @@ ModelLoadResult gpttype_load_model(const load_model_inputs inputs, FileFormat in
llamamodel->hparams.rope_freq_scale_train!=1.0f ||
llamamodel->hparams.rope_scaling_type_train==2)
{
float ropemultiplier = 1.0f;
if(llamamodel->hparams.rope_scaling_type_train!=2 &&
llamamodel->hparams.n_ctx_train > 2048 && clamped_max_context_length > llamamodel->hparams.n_ctx_train)
{
ropemultiplier = (float)llamamodel->hparams.n_ctx_train / (float)clamped_max_context_length;
llama_ctx_params.rope_freq_base = rope_freq_base = llamamodel->hparams.rope_freq_base_train;
llama_ctx_params.rope_freq_scale = rope_freq_scale = ropemultiplier * llamamodel->hparams.rope_freq_scale_train;
printf("Automatic RoPE Scaling: Using (scale:%.3f, base:%.1f).\n", rope_freq_scale, rope_freq_base);
}
else
{
// float ropemultiplier = 1.0f;
// if(llamamodel->hparams.rope_scaling_type_train!=2 &&
// llamamodel->hparams.n_ctx_train > 2048 && clamped_max_context_length > llamamodel->hparams.n_ctx_train)
// {
// ropemultiplier = (float)llamamodel->hparams.n_ctx_train / (float)clamped_max_context_length;
// llama_ctx_params.rope_freq_base = rope_freq_base = llamamodel->hparams.rope_freq_base_train;
// llama_ctx_params.rope_freq_scale = rope_freq_scale = ropemultiplier * llamamodel->hparams.rope_freq_scale_train;
// printf("Automatic RoPE Scaling: Using (scale:%.3f, base:%.1f).\n", rope_freq_scale, rope_freq_base);
// }
// else
// {
printf("Automatic RoPE Scaling: Using model internal value.\n");
}
//}
}
else
{