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

up ver, set the cuda pool malloc lookahead back to 5% instead of 2% (+1 squashed commits)

Browse source 
Squashed commits:

[e0f65278] up ver, set the cuda pool malloc lookahead back to 5% instead of 2%
This commit is contained in:
Concedo 2023-08-09 11:50:12 +08:00
parent 926d90fbab
commit 159ad9269d
2 changed files with 2 additions and 2 deletions
Download patch file Download diff file Expand all files Collapse all files

2
ggml-cuda.cu
View file

@ -4332,7 +4332,7 @@ static void * ggml_cuda_pool_malloc(size_t size, size_t * actual_size) {
} }
void * ptr; void * ptr;
size_t look_ahead_size = (size_t) (1.02 * size); size_t look_ahead_size = (size_t) (1.05 * size);
look_ahead_size = 256 * ((look_ahead_size + 255)/256); look_ahead_size = 256 * ((look_ahead_size + 255)/256);
CUDA_CHECK(cudaMalloc((void **) &ptr, look_ahead_size)); CUDA_CHECK(cudaMalloc((void **) &ptr, look_ahead_size));
*actual_size = look_ahead_size; *actual_size = look_ahead_size;

2
koboldcpp.py
View file

@ -303,7 +303,7 @@ maxhordectx = 1024
maxhordelen = 256 maxhordelen = 256
modelbusy = threading.Lock() modelbusy = threading.Lock()
defaultport = 5001 defaultport = 5001
KcppVersion = "1.39.1" KcppVersion = "1.39.2"
showdebug = True showdebug = True
showsamplerwarning = True showsamplerwarning = True
showmaxctxwarning = True showmaxctxwarning = True