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

Add flag for lowvram directly into cublas launch param

Browse source 
Merge remote-tracking branch 'yellowrose/pr/open/LostRuins/koboldcpp/lowvram' into concedo_experimental

# Conflicts:
#	koboldcpp.py
This commit is contained in:
Concedo 2023-06-29 17:07:31 +08:00
commit 4b3a1282f0
3 changed files with 7 additions and 3 deletions
Download patch file Download diff file Expand all files Collapse all files

1
expose.h
View file

@ -8,6 +8,7 @@ struct load_model_inputs
const int max_context_length; const int max_context_length;
const int batch_size; const int batch_size;
const bool f16_kv; const bool f16_kv;
const bool low_vram;
const char * executable_path; const char * executable_path;
const char * model_filename; const char * model_filename;
const char * lora_filename; const char * lora_filename;

1
gpttype_adapter.cpp
View file

@ -377,6 +377,7 @@ ModelLoadResult gpttype_load_model(const load_model_inputs inputs, FileFormat in
//llama_ctx_paran_parts = -1; //llama_ctx_paran_parts = -1;
llama_ctx_params.seed = -1; llama_ctx_params.seed = -1;
llama_ctx_params.f16_kv = inputs.f16_kv; llama_ctx_params.f16_kv = inputs.f16_kv;
llama_ctx_params.low_vram = inputs.low_vram;
llama_ctx_params.logits_all = false; llama_ctx_params.logits_all = false;
llama_ctx_params.use_mmap = inputs.use_mmap; llama_ctx_params.use_mmap = inputs.use_mmap;
llama_ctx_params.use_mlock = inputs.use_mlock; llama_ctx_params.use_mlock = inputs.use_mlock;

8
koboldcpp.py
View file

@ -16,6 +16,7 @@ class load_model_inputs(ctypes.Structure):
("max_context_length", ctypes.c_int), ("max_context_length", ctypes.c_int),
("batch_size", ctypes.c_int), ("batch_size", ctypes.c_int),
("f16_kv", ctypes.c_bool), ("f16_kv", ctypes.c_bool),
("low_vram", ctypes.c_bool),
("executable_path", ctypes.c_char_p), ("executable_path", ctypes.c_char_p),
("model_filename", ctypes.c_char_p), ("model_filename", ctypes.c_char_p),
("lora_filename", ctypes.c_char_p), ("lora_filename", ctypes.c_char_p),
@ -89,7 +90,7 @@ def init_library():
use_clblast = False #uses CLBlast instead use_clblast = False #uses CLBlast instead
use_cublas = False #uses cublas instead use_cublas = False #uses cublas instead
use_noavx2 = False #uses openblas with no avx2 instructions use_noavx2 = False #uses openblas with no avx2 instructions
print(args.usecublas)
if args.noavx2: if args.noavx2:
use_noavx2 = True use_noavx2 = True
if not file_exists(lib_openblas_noavx2) or (os.name=='nt' and not file_exists("libopenblas.dll")): if not file_exists(lib_openblas_noavx2) or (os.name=='nt' and not file_exists("libopenblas.dll")):
@ -105,7 +106,7 @@ def init_library():
else: else:
print("Attempting to use CLBlast library for faster prompt ingestion. A compatible clblast will be required.") print("Attempting to use CLBlast library for faster prompt ingestion. A compatible clblast will be required.")
use_clblast = True use_clblast = True
elif args.usecublas: elif (args.usecublas and args.usecublas!=""):
if not file_exists(lib_cublas): if not file_exists(lib_cublas):
print("Warning: CuBLAS library file not found. Non-BLAS library will be used.") print("Warning: CuBLAS library file not found. Non-BLAS library will be used.")
else: else:
@ -160,6 +161,7 @@ def load_model(model_filename):
inputs.batch_size = 8 inputs.batch_size = 8
inputs.max_context_length = maxctx #initial value to use for ctx, can be overwritten inputs.max_context_length = maxctx #initial value to use for ctx, can be overwritten
inputs.threads = args.threads inputs.threads = args.threads
inputs.low_vram = (True if args.usecublas=="lowvram" else False)
inputs.blasthreads = args.blasthreads inputs.blasthreads = args.blasthreads
inputs.f16_kv = True inputs.f16_kv = True
inputs.use_mmap = (not args.nommap) inputs.use_mmap = (not args.nommap)
@ -874,7 +876,7 @@ if __name__ == '__main__':
compatgroup = parser.add_mutually_exclusive_group() compatgroup = parser.add_mutually_exclusive_group()
compatgroup.add_argument("--noblas", help="Do not use OpenBLAS for accelerated prompt ingestion", action='store_true') compatgroup.add_argument("--noblas", help="Do not use OpenBLAS for accelerated prompt ingestion", action='store_true')
compatgroup.add_argument("--useclblast", help="Use CLBlast for GPU Acceleration. Must specify exactly 2 arguments, platform ID and device ID (e.g. --useclblast 1 0).", type=int, choices=range(0,9), nargs=2) compatgroup.add_argument("--useclblast", help="Use CLBlast for GPU Acceleration. Must specify exactly 2 arguments, platform ID and device ID (e.g. --useclblast 1 0).", type=int, choices=range(0,9), nargs=2)
compatgroup.add_argument("--usecublas", help="Use CuBLAS for GPU Acceleration. Requires Nvidia GPU.", action='store_true') compatgroup.add_argument("--usecublas", help="Use CuBLAS for GPU Acceleration. Requires Nvidia GPU. Select lowvram to not allocate VRAM scratch buffer.", default='', const='normal', nargs='?', choices=['normal', 'lowvram'])
parser.add_argument("--gpulayers", help="Set number of layers to offload to GPU when using GPU. Requires GPU.",metavar=('[GPU layers]'), type=int, default=0) parser.add_argument("--gpulayers", help="Set number of layers to offload to GPU when using GPU. Requires GPU.",metavar=('[GPU layers]'), type=int, default=0)
args = parser.parse_args() args = parser.parse_args()
main(args) main(args)