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

added stop reason in the perf endpoint

Browse source
This commit is contained in:
Concedo 2023-07-24 11:55:35 +08:00
parent 910744e2c0
commit 280abaf029
4 changed files with 18 additions and 1 deletions
Download patch file Download diff file Expand all files Collapse all files

3
expose.cpp
View file

@ -229,6 +229,9 @@ extern "C"
int get_last_token_count() { int get_last_token_count() {
return last_token_count; return last_token_count;
} }
int get_last_stop_reason() {
return (int)last_stop_reason;
}
const char* get_pending_output() { const char* get_pending_output() {
return gpttype_get_pending_output().c_str(); return gpttype_get_pending_output().c_str();

8
expose.h
View file

@ -14,6 +14,13 @@ enum samplers
KCPP_SAMPLER_REP_PEN=6, KCPP_SAMPLER_REP_PEN=6,
KCPP_SAMPLER_MAX KCPP_SAMPLER_MAX
}; };
enum stop_reason
{
INVALID=-1,
OUT_OF_TOKENS=0,
EOS_TOKEN=1,
CUSTOM_STOPPER=2,
};
struct load_model_inputs struct load_model_inputs
{ {
const int threads; const int threads;
@ -76,3 +83,4 @@ extern bool generation_finished;
extern float last_eval_time; extern float last_eval_time;
extern float last_process_time; extern float last_process_time;
extern int last_token_count; extern int last_token_count;
extern stop_reason last_stop_reason;

4
gpttype_adapter.cpp
View file

@ -36,6 +36,7 @@ bool generation_finished;
float last_process_time = 0; float last_process_time = 0;
float last_eval_time = 0; float last_eval_time = 0;
int last_token_count = 0; int last_token_count = 0;
stop_reason last_stop_reason = stop_reason::INVALID;
std::vector<std::string> generated_tokens; std::vector<std::string> generated_tokens;
//return val: 0=fail, 1=(original ggml, alpaca), 2=(ggmf), 3=(ggjt) //return val: 0=fail, 1=(original ggml, alpaca), 2=(ggmf), 3=(ggjt)
@ -871,6 +872,7 @@ const std::string & gpttype_get_pending_output()
generation_outputs gpttype_generate(const generation_inputs inputs, generation_outputs &output) generation_outputs gpttype_generate(const generation_inputs inputs, generation_outputs &output)
{ {
concat_output = ""; concat_output = "";
last_stop_reason = stop_reason::OUT_OF_TOKENS;
stop_sequence.clear(); stop_sequence.clear();
for(int x=0;x<stop_token_max;++x) for(int x=0;x<stop_token_max;++x)
{ {
@ -1433,6 +1435,7 @@ generation_outputs gpttype_generate(const generation_inputs inputs, generation_o
stopper_unused_tokens = remaining_tokens; stopper_unused_tokens = remaining_tokens;
printf("\n(EOS token triggered!)"); printf("\n(EOS token triggered!)");
remaining_tokens = 0; remaining_tokens = 0;
last_stop_reason = stop_reason::EOS_TOKEN;
} }
for (const auto &matched : stop_sequence) for (const auto &matched : stop_sequence)
@ -1445,6 +1448,7 @@ generation_outputs gpttype_generate(const generation_inputs inputs, generation_o
{ {
printf("\n(Stop sequence triggered: <%s>)", matched.c_str()); printf("\n(Stop sequence triggered: <%s>)", matched.c_str());
} }
last_stop_reason = stop_reason::CUSTOM_STOPPER;
break; break;
} }
} }

4
koboldcpp.py
View file

@ -165,6 +165,7 @@ def init_library():
handle.get_last_eval_time.restype = ctypes.c_float handle.get_last_eval_time.restype = ctypes.c_float
handle.get_last_process_time.restype = ctypes.c_float handle.get_last_process_time.restype = ctypes.c_float
handle.get_last_token_count.restype = ctypes.c_int handle.get_last_token_count.restype = ctypes.c_int
handle.get_last_stop_reason.restype = ctypes.c_int
handle.abort_generate.restype = ctypes.c_bool handle.abort_generate.restype = ctypes.c_bool
handle.get_pending_output.restype = ctypes.c_char_p handle.get_pending_output.restype = ctypes.c_char_p
@ -470,7 +471,8 @@ class ServerRequestHandler(http.server.SimpleHTTPRequestHandler):
lastp = handle.get_last_process_time() lastp = handle.get_last_process_time()
laste = handle.get_last_eval_time() laste = handle.get_last_eval_time()
lastc = handle.get_last_token_count() lastc = handle.get_last_token_count()
response_body = (json.dumps({"last_process":lastp,"last_eval":laste,"last_token_count":lastc}).encode()) stopreason = handle.get_last_stop_reason()
response_body = (json.dumps({"last_process":lastp,"last_eval":laste,"last_token_count":lastc, "stop_reason":stopreason}).encode())
if response_body is None: if response_body is None:
self.send_response(404) self.send_response(404)