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

merge CLBlast improvements - GPU dequant

Browse source
This commit is contained in:
Concedo 2023-04-16 01:17:40 +08:00
parent 3e992eabb4
commit ad5676810a
4 changed files with 13 additions and 5 deletions
Download patch file Download diff file Expand all files Collapse all files

1
expose.h
View file

@ -11,6 +11,7 @@ struct load_model_inputs
const bool use_mmap; const bool use_mmap;
const bool use_smartcontext; const bool use_smartcontext;
const int clblast_info = 0; const int clblast_info = 0;
const int blasbatchsize = 512;
}; };
struct generation_inputs struct generation_inputs
{ {

4
gpttype_adapter.cpp
View file

@ -30,6 +30,7 @@ static int n_past = 0;
static int n_threads = 4; static int n_threads = 4;
static int n_batch = 8; static int n_batch = 8;
static bool useSmartContext = false; static bool useSmartContext = false;
static int blasbatchsize = 512;
static std::string modelname; static std::string modelname;
static std::vector<gpt_vocab::id> last_n_tokens; static std::vector<gpt_vocab::id> last_n_tokens;
static std::vector<gpt_vocab::id> current_context_tokens; static std::vector<gpt_vocab::id> current_context_tokens;
@ -53,6 +54,7 @@ ModelLoadResult gpttype_load_model(const load_model_inputs inputs, FileFormat in
n_batch = params.n_batch = inputs.batch_size; n_batch = params.n_batch = inputs.batch_size;
modelname = params.model = inputs.model_filename; modelname = params.model = inputs.model_filename;
useSmartContext = inputs.use_smartcontext; useSmartContext = inputs.use_smartcontext;
blasbatchsize = inputs.blasbatchsize;
params.memory_f16 = inputs.f16_kv; params.memory_f16 = inputs.f16_kv;
params.n_ctx = inputs.max_context_length; params.n_ctx = inputs.max_context_length;
model_v1.hparams.n_ctx = model_v2.hparams.n_ctx = model_gpt2_v1.hparams.n_ctx = model_gpt2_v2.hparams.n_ctx = params.n_ctx; model_v1.hparams.n_ctx = model_v2.hparams.n_ctx = model_gpt2_v1.hparams.n_ctx = model_gpt2_v2.hparams.n_ctx = params.n_ctx;
@ -208,7 +210,7 @@ generation_outputs gpttype_generate(const generation_inputs inputs, generation_o
int original_threads = params.n_threads; int original_threads = params.n_threads;
if (blasmode) if (blasmode)
{ {
params.n_batch = 512; //received reports of 1024 and above crashing on some models params.n_batch = blasbatchsize; //received reports of 1024 and above crashing on some models
params.n_threads = 1; params.n_threads = 1;
} }

9
koboldcpp.py
View file

@ -17,7 +17,8 @@ class load_model_inputs(ctypes.Structure):
("n_parts_overwrite", ctypes.c_int), ("n_parts_overwrite", ctypes.c_int),
("use_mmap", ctypes.c_bool), ("use_mmap", ctypes.c_bool),
("use_smartcontext", ctypes.c_bool), ("use_smartcontext", ctypes.c_bool),
("clblast_info", ctypes.c_int)] ("clblast_info", ctypes.c_int),
("blasbatchsize", ctypes.c_int)]
class generation_inputs(ctypes.Structure): class generation_inputs(ctypes.Structure):
_fields_ = [("seed", ctypes.c_int), _fields_ = [("seed", ctypes.c_int),
@ -66,7 +67,7 @@ def init_library():
handle.generate.argtypes = [generation_inputs, ctypes.c_wchar_p] #apparently needed for osx to work. i duno why they need to interpret it that way but whatever handle.generate.argtypes = [generation_inputs, ctypes.c_wchar_p] #apparently needed for osx to work. i duno why they need to interpret it that way but whatever
handle.generate.restype = generation_outputs handle.generate.restype = generation_outputs
def load_model(model_filename,batch_size=8,max_context_length=512,n_parts_overwrite=-1,threads=6,use_mmap=False,use_smartcontext=False): def load_model(model_filename,batch_size=8,max_context_length=512,n_parts_overwrite=-1,threads=6,use_mmap=False,use_smartcontext=False,blasbatchsize=512):
inputs = load_model_inputs() inputs = load_model_inputs()
inputs.model_filename = model_filename.encode("UTF-8") inputs.model_filename = model_filename.encode("UTF-8")
inputs.batch_size = batch_size inputs.batch_size = batch_size
@ -76,6 +77,7 @@ def load_model(model_filename,batch_size=8,max_context_length=512,n_parts_overwr
inputs.f16_kv = True inputs.f16_kv = True
inputs.use_mmap = use_mmap inputs.use_mmap = use_mmap
inputs.use_smartcontext = use_smartcontext inputs.use_smartcontext = use_smartcontext
inputs.blasbatchsize = blasbatchsize
clblastids = 0 clblastids = 0
if args.useclblast: if args.useclblast:
clblastids = 100 + int(args.useclblast[0])*10 + int(args.useclblast[1]) clblastids = 100 + int(args.useclblast[0])*10 + int(args.useclblast[1])
@ -388,7 +390,7 @@ def main(args):
mdl_nparts = sum(1 for n in range(1, 9) if os.path.exists(f"{ggml_selected_file}.{n}")) + 1 mdl_nparts = sum(1 for n in range(1, 9) if os.path.exists(f"{ggml_selected_file}.{n}")) + 1
modelname = os.path.abspath(ggml_selected_file) modelname = os.path.abspath(ggml_selected_file)
print(f"Loading model: {modelname} \n[Parts: {mdl_nparts}, Threads: {args.threads}, SmartContext: {args.smartcontext}]") print(f"Loading model: {modelname} \n[Parts: {mdl_nparts}, Threads: {args.threads}, SmartContext: {args.smartcontext}]")
loadok = load_model(modelname,8,maxctx,mdl_nparts,args.threads,(not args.nommap),args.smartcontext) loadok = load_model(modelname,8,maxctx,mdl_nparts,args.threads,(not args.nommap),args.smartcontext,args.blasbatchsize)
print("Load Model OK: " + str(loadok)) print("Load Model OK: " + str(loadok))
if not loadok: if not loadok:
@ -435,6 +437,7 @@ if __name__ == '__main__':
default_threads = (physical_core_limit if physical_core_limit<=3 else max(3,physical_core_limit-1)) default_threads = (physical_core_limit if physical_core_limit<=3 else max(3,physical_core_limit-1))
parser.add_argument("--threads", help="Use a custom number of threads if specified. Otherwise, uses an amount based on CPU cores", type=int, default=default_threads) parser.add_argument("--threads", help="Use a custom number of threads if specified. Otherwise, uses an amount based on CPU cores", type=int, default=default_threads)
parser.add_argument("--psutil_set_threads", help="Experimental flag. If set, uses psutils to determine thread count based on physical cores.", action='store_true') parser.add_argument("--psutil_set_threads", help="Experimental flag. If set, uses psutils to determine thread count based on physical cores.", action='store_true')
parser.add_argument("--blasbatchsize", help="Sets the batch size used in BLAS processing (default 512)", type=int,choices=[128,256,512,1024], default=512)
parser.add_argument("--stream", help="Uses pseudo streaming", action='store_true') parser.add_argument("--stream", help="Uses pseudo streaming", action='store_true')
parser.add_argument("--smartcontext", help="Reserving a portion of context to try processing less frequently.", action='store_true') parser.add_argument("--smartcontext", help="Reserving a portion of context to try processing less frequently.", action='store_true')
parser.add_argument("--nommap", help="If set, do not use mmap to load newer models", action='store_true') parser.add_argument("--nommap", help="If set, do not use mmap to load newer models", action='store_true')

4
llama_adapter.cpp
View file

@ -28,6 +28,7 @@ static int n_past = 0;
static int n_threads = 4; static int n_threads = 4;
static int n_batch = 8; static int n_batch = 8;
static bool useSmartContext = false; static bool useSmartContext = false;
static int blasbatchsize = 512;
static std::string modelname; static std::string modelname;
static llama_context *ctx; static llama_context *ctx;
static std::vector<llama_token> last_n_tokens; static std::vector<llama_token> last_n_tokens;
@ -44,6 +45,7 @@ bool llama_load_model(const load_model_inputs inputs, FileFormat in_file_format)
n_batch = inputs.batch_size; n_batch = inputs.batch_size;
modelname = inputs.model_filename; modelname = inputs.model_filename;
useSmartContext = inputs.use_smartcontext; useSmartContext = inputs.use_smartcontext;
blasbatchsize = inputs.blasbatchsize;
ctx_params.n_ctx = inputs.max_context_length; ctx_params.n_ctx = inputs.max_context_length;
ctx_params.n_parts = -1;//inputs.n_parts_overwrite; ctx_params.n_parts = -1;//inputs.n_parts_overwrite;
@ -143,7 +145,7 @@ generation_outputs llama_generate(const generation_inputs inputs, generation_out
int original_threads = params.n_threads; int original_threads = params.n_threads;
if (blasmode) if (blasmode)
{ {
params.n_batch = 512; //received reports of 1024 and above crashing on some models params.n_batch = blasbatchsize; //received reports of 1024 and above crashing on some models
params.n_threads = 1; params.n_threads = 1;
} }