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

Allocate a small amount of extra context for GGUF to deal with KV fragmentation causing issues in some scenarios.

Browse source
This commit is contained in:
Concedo 2023-11-28 20:55:14 +08:00
parent d2ef458b02
commit ba5c33319b
2 changed files with 5 additions and 4 deletions
Download patch file Download diff file Expand all files Collapse all files

4
gpttype_adapter.cpp
View file

@ -883,7 +883,7 @@ ModelLoadResult gpttype_load_model(const load_model_inputs inputs, FileFormat in
{
llama_model_params model_params = llama_model_default_params();
llama_context_params llama_ctx_params = llama_context_default_params();
llama_ctx_params.n_ctx = clamped_max_context_length;
llama_ctx_params.n_ctx = clamped_max_context_length + 64; //add some extra context to deal with KV fragmentation
//llama_ctx_paran_parts = -1;
llama_ctx_params.seed = -1;
llama_ctx_params.f16_kv = inputs.f16_kv;
@ -1808,7 +1808,7 @@ generation_outputs gpttype_generate(const generation_inputs inputs, generation_o
if (!evalres)
{
fprintf(stderr, "\nFailed to predict! Check your context buffer sizes!\n");
fprintf(stderr, "\nFailed to predict at %d! Check your context buffer sizes!\n",n_past);
snprintf(output.text, sizeof(output.text), "%s", "");
output.status = 0;
generation_finished = true;

5
koboldcpp.py
View file

@ -2164,13 +2164,14 @@ def main(launch_args,start_server=True):
if args.port_param!=defaultport:
args.port = args.port_param
print(f"Starting Kobold HTTP Server on port {args.port} at http://localhost:{args.port}/api")
print(f"Starting OpenAI Compatible Endpoint on port {args.port} at http://localhost:{args.port}/v1")
epurl = ""
if args.host=="":
epurl = f"http://localhost:{args.port}"
else:
epurl = f"http://{args.host}:{args.port}"
print(f"Starting Kobold HTTP Server on port {args.port} at {epurl}/api/")
print(f"Starting OpenAI Compatible Endpoint on port {args.port} at {epurl}/v1/")
if args.launch:
try: