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

hotfix to revert the auto ctx scaling first, i didnt do it properly

Browse source
This commit is contained in:
Concedo 2023-08-31 00:58:52 +08:00
parent 5cd0309610
commit b6914ebd04
2 changed files with 6 additions and 6 deletions
Download patch file Download diff file Expand all files Collapse all files

10
gpttype_adapter.cpp
View file

@ -435,11 +435,11 @@ ModelLoadResult gpttype_load_model(const load_model_inputs inputs, FileFormat in
{ {
//approximate NTK aware ctx //approximate NTK aware ctx
auto effectivenctx = params.n_ctx; auto effectivenctx = params.n_ctx;
if((file_format == FileFormat::GGUF_LLAMA || file_format==FileFormat::GGUF_FALCON) && llama_ctx_v4->model.hparams.n_ctx_train>2048) // if((file_format == FileFormat::GGUF_LLAMA || file_format==FileFormat::GGUF_FALCON) && llama_ctx_v4->model.hparams.n_ctx_train>2048)
{ // {
float factor = llama_ctx_v4->model.hparams.n_ctx_train/2048; // float factor = llama_ctx_v4->model.hparams.n_ctx_train/2048;
effectivenctx = effectivenctx/factor; // effectivenctx = effectivenctx/factor;
} // }
rope_freq_base = (effectivenctx <= 3072 ? 26000.0f : (effectivenctx <= 4096 ? 32000.0f : (effectivenctx <= 6144 ? 54000.0f : (effectivenctx <= 8192 ? 82684.0f : (effectivenctx <= 12288 ? 140000.0f : 200000.0f))))); rope_freq_base = (effectivenctx <= 3072 ? 26000.0f : (effectivenctx <= 4096 ? 32000.0f : (effectivenctx <= 6144 ? 54000.0f : (effectivenctx <= 8192 ? 82684.0f : (effectivenctx <= 12288 ? 140000.0f : 200000.0f)))));
} }

2
koboldcpp.py
View file

@ -324,7 +324,7 @@ maxhordectx = 1024
maxhordelen = 256 maxhordelen = 256
modelbusy = threading.Lock() modelbusy = threading.Lock()
defaultport = 5001 defaultport = 5001
KcppVersion = "1.42" KcppVersion = "1.42.1"
showdebug = True showdebug = True
showsamplerwarning = True showsamplerwarning = True
showmaxctxwarning = True showmaxctxwarning = True