vrr/koboldcpp
2
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2025-09-11 01:24:36 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

bugfixes and support for persistent states

Browse source
This commit is contained in:
Concedo 2023-03-20 00:59:45 +08:00
parent f952b7c613
commit 356c1b87ba
3 changed files with 66 additions and 22 deletions
Download patch file Download diff file Expand all files Collapse all files

25
llama_for_kobold.py
View file

@ -21,7 +21,8 @@ class generation_inputs(ctypes.Structure):
("top_k", ctypes.c_int),
("top_p", ctypes.c_float),
("rep_pen", ctypes.c_float),
("rep_pen_range", ctypes.c_int)]
("rep_pen_range", ctypes.c_int),
("reset_state", ctypes.c_bool)]
class generation_outputs(ctypes.Structure):
_fields_ = [("status", ctypes.c_int),
@ -45,7 +46,7 @@ def load_model(model_filename,batch_size=8,max_context_length=512,threads=4,n_pa
ret = handle.load_model(inputs)
return ret
def generate(prompt,max_length=20,temperature=0.8,top_k=100,top_p=0.85,rep_pen=1.1,rep_pen_range=128,seed=-1):
def generate(prompt,max_length=20,temperature=0.8,top_k=100,top_p=0.85,rep_pen=1.1,rep_pen_range=128,seed=-1,reset_state=True):
inputs = generation_inputs()
outputs = generation_outputs()
inputs.prompt = prompt.encode("UTF-8")
@ -56,6 +57,7 @@ def generate(prompt,max_length=20,temperature=0.8,top_k=100,top_p=0.85,rep_pen=1
inputs.rep_pen = rep_pen
inputs.rep_pen_range = rep_pen_range
inputs.seed = seed
inputs.reset_state = reset_state
ret = handle.generate(inputs,outputs)
if(ret.status==1):
return ret.text.decode("UTF-8")
@ -75,6 +77,7 @@ maxctx = 1024
maxlen = 256
modelbusy = False
port = 5001
last_context = ""
class ServerRequestHandler(http.server.BaseHTTPRequestHandler):
@ -120,7 +123,8 @@ class ServerRequestHandler(http.server.BaseHTTPRequestHandler):
content_length = int(self.headers['Content-Length'])
body = self.rfile.read(content_length)
if self.path.endswith('/api/v1/generate/') or self.path.endswith('/api/latest/generate/'):
global modelbusy
global modelbusy
global last_context
if modelbusy:
self.send_response(503)
self.end_headers()
@ -140,17 +144,26 @@ class ServerRequestHandler(http.server.BaseHTTPRequestHandler):
return
print("\nInput: " + json.dumps(genparams))
fresh_state = True
fullprompt = genparams.get('prompt', "")
newprompt = fullprompt
if last_context!="" and newprompt.startswith(last_context):
fresh_state = False
newprompt = newprompt[len(last_context):]
#print("trimmed: " + newprompt)
recvtxt = generate(
prompt=genparams.get('prompt', ""),
prompt=newprompt,
max_length=genparams.get('max_length', 50),
temperature=genparams.get('temperature', 0.8),
top_k=genparams.get('top_k', 100),
top_p=genparams.get('top_p', 0.85),
rep_pen=genparams.get('rep_pen', 1.1),
rep_pen_range=genparams.get('rep_pen_range', 128),
seed=-1
seed=-1,
reset_state=fresh_state
)
print("\nOutput: " + recvtxt)
last_context = fullprompt + recvtxt
res = {"results": [{"text": recvtxt}]}
self.send_response(200)
self.end_headers()
@ -241,7 +254,7 @@ if __name__ == '__main__':
mdl_nparts += 1
modelname = os.path.abspath(sys.argv[1])
print("Loading model: " + modelname)
loadok = load_model(modelname,128,maxctx,4,mdl_nparts)
loadok = load_model(modelname,24,maxctx,4,mdl_nparts)
print("Load Model OK: " + str(loadok))
if loadok: