vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

try to scale custom roped models

Browse source
This commit is contained in:
Concedo 2023-11-19 16:24:13 +08:00
parent 22c56f9221
commit cf646fa809
1 changed files with 13 additions and 1 deletions
Download patch file Download diff file Expand all files Collapse all files

14
gpttype_adapter.cpp
View file

@ -941,7 +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)
{
printf("Automatic RoPE Scaling: Using model internal values.\n");
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
{