vrr/koboldcpp
2
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2025-09-10 09:04:36 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

KCPP Fetches AMD ROCm Memory without a stick, CC_TURING Gets the Boot, koboldcpp_hipblas.dll Talks To The Hand, and hipBLAS Compiler Finds Its Independence! (#517)

Browse source 
* AMD ROCm memory fetching and max mem setting

* Update .gitignore with koboldcpp_hipblas.dll

* Update CMakeLists.txt remove CC_TURING for AMD

* separate hipBLAS compiler, update MMV_Y, move CXX/CC print

separate hipBLAS compiler, update MMV_Y value, move the section that prints CXX and CC compiler name
This commit is contained in:
YellowRoseCx 2023-11-05 08:23:18 -06:00 • committed by GitHub
parent a62468ec4c
commit e2e5fe56a8
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
4 changed files with 33 additions and 26 deletions
Download patch file Download diff file Expand all files Collapse all files

29
koboldcpp.py
View file

@ -1086,6 +1086,7 @@ def show_new_gui():
from subprocess import run, CalledProcessError
FetchedCUdevices = []
FetchedCUdeviceMem = []
AMDgpu = None
try: # Get OpenCL GPU names on windows using a special binary. overwrite at known index if found.
basepath = os.path.abspath(os.path.dirname(__file__))
output = run([((os.path.join(basepath, "winclinfo.exe")) if os.name == 'nt' else "clinfo"),"--json"], capture_output=True, text=True, check=True, encoding='utf-8').stdout
@ -1119,29 +1120,41 @@ def show_new_gui():
try: # Get AMD ROCm GPU names
output = run(['rocminfo'], capture_output=True, text=True, check=True, encoding='utf-8').stdout
device_name = None
for line in output.splitlines():
for line in output.splitlines(): # read through the output line by line
line = line.strip()
if line.startswith("Marketing Name:"): device_name = line.split(":", 1)[1].strip()
elif line.startswith("Device Type:") and "GPU" in line and device_name is not None: FetchedCUdevices.append(device_name)
if line.startswith("Marketing Name:"): device_name = line.split(":", 1)[1].strip() # if we find a named device, temporarily save the name
elif line.startswith("Device Type:") and "GPU" in line and device_name is not None: # if the following Device Type is a GPU (not a CPU) then add it to devices list
FetchedCUdevices.append(device_name)
AMDgpu = True
elif line.startswith("Device Type:") and "GPU" not in line: device_name = None
if FetchedCUdevices:
getamdvram = run(['rocm-smi', '--showmeminfo', 'vram', '--csv'], capture_output=True, text=True, check=True, encoding='utf-8').stdout # fetch VRAM of devices
FetchedCUdeviceMem = [line.split(",")[1].strip() for line in getamdvram.splitlines()[1:] if line.strip()]
except Exception as e:
pass
for idx in range(0,4):
if(len(FetchedCUdevices)>idx):
CUDevicesNames[idx] = FetchedCUdevices[idx]
MaxMemory[0] = max(int(FetchedCUdeviceMem[idx])*1024*1024,MaxMemory[0])
pass
if AMDgpu:
MaxMemory[0] = max(int(FetchedCUdeviceMem[idx]),MaxMemory[0])
else:
MaxMemory[0] = max(int(FetchedCUdeviceMem[idx])*1024*1024,MaxMemory[0])
pass
#autopick cublas if suitable
global exitcounter
if exitcounter < 100 and MaxMemory[0]>3500000000 and CUDevicesNames[0]!="" and "Use CuBLAS" in runopts and runopts_var.get()=="Use OpenBLAS":
runopts_var.set("Use CuBLAS")
pass
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 "Use CuBLAS" in runopts:
runopts_var.set("Use CuBLAS")
pass
elif "Use hipBLAS (ROCM)" in runopts:
runopts_var.set("Use hipBLAS (ROCM)")
changed_gpu_choice_var()
return
def autoset_gpu_layers(filepath): #shitty algo to determine how many layers to use
try:
global gui_layers_untouched