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

cleanup, up ver (+1 squashed commits)

Browse source 
Squashed commits:

[1ea303d6] cleanup , up ver (+1 squashed commits)

Squashed commits:

[79f09b22] cleanup
This commit is contained in:
Concedo 2023-11-05 22:29:36 +08:00
parent e2e5fe56a8
commit ea81eae189
2 changed files with 4 additions and 7 deletions
Download patch file Download diff file Expand all files Collapse all files

2
colab.ipynb
View file

@ -67,7 +67,7 @@
"!test -f koboldcpp_cublas.so && echo Build Skipped || make koboldcpp_cublas LLAMA_CUBLAS=1\r\n", "!test -f koboldcpp_cublas.so && echo Build Skipped || make koboldcpp_cublas LLAMA_CUBLAS=1\r\n",
"!cp koboldcpp_cublas.so koboldcpp_cublas.dat\r\n", "!cp koboldcpp_cublas.so koboldcpp_cublas.dat\r\n",
"!apt install aria2 -y\r\n", "!apt install aria2 -y\r\n",
"!aria2c -x 10 -o model.ggml --summary-interval=3 --download-result=full --allow-overwrite=true --file-allocation=none $Model\r\n", "!aria2c -x 10 -o model.ggml --summary-interval=5 --download-result=default --allow-overwrite=true --file-allocation=none $Model\r\n",
"!python koboldcpp.py model.ggml --usecublas 0 mmq --multiuser --gpulayers $Layers --contextsize $ContextSize --hordeconfig concedo 1 1 --remotetunnel\r\n" "!python koboldcpp.py model.ggml --usecublas 0 mmq --multiuser --gpulayers $Layers --contextsize $ContextSize --hordeconfig concedo 1 1 --remotetunnel\r\n"
] ]
} }

9
koboldcpp.py
View file

@ -379,7 +379,7 @@ maxhordelen = 256
modelbusy = threading.Lock() modelbusy = threading.Lock()
requestsinqueue = 0 requestsinqueue = 0
defaultport = 5001 defaultport = 5001
KcppVersion = "1.48" KcppVersion = "1.48.1"
showdebug = True showdebug = True
showsamplerwarning = True showsamplerwarning = True
showmaxctxwarning = True showmaxctxwarning = True
@ -1140,21 +1140,18 @@ def show_new_gui():
MaxMemory[0] = max(int(FetchedCUdeviceMem[idx]),MaxMemory[0]) MaxMemory[0] = max(int(FetchedCUdeviceMem[idx]),MaxMemory[0])
else: else:
MaxMemory[0] = max(int(FetchedCUdeviceMem[idx])*1024*1024,MaxMemory[0]) MaxMemory[0] = max(int(FetchedCUdeviceMem[idx])*1024*1024,MaxMemory[0])
pass
#autopick cublas if suitable #autopick cublas if suitable, requires at least 3.5GB VRAM to auto pick
global exitcounter global exitcounter
if exitcounter < 100 and MaxMemory[0]>3500000000 and CUDevicesNames[0]!="" and "Use CuBLAS" or "Use hipBLAS (ROCM)" in runopts and runopts_var.get()=="Use OpenBLAS": if exitcounter < 100 and MaxMemory[0]>3500000000 and CUDevicesNames[0]!="" and ("Use CuBLAS" in runopts or "Use hipBLAS (ROCM)" in runopts) and runopts_var.get()=="Use OpenBLAS":
if "Use CuBLAS" in runopts: if "Use CuBLAS" in runopts:
runopts_var.set("Use CuBLAS") runopts_var.set("Use CuBLAS")
pass
elif "Use hipBLAS (ROCM)" in runopts: elif "Use hipBLAS (ROCM)" in runopts:
runopts_var.set("Use hipBLAS (ROCM)") runopts_var.set("Use hipBLAS (ROCM)")
changed_gpu_choice_var() changed_gpu_choice_var()
return return
def autoset_gpu_layers(filepath): #shitty algo to determine how many layers to use def autoset_gpu_layers(filepath): #shitty algo to determine how many layers to use
try: try:
global gui_layers_untouched global gui_layers_untouched