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

server: fix core dump when input prompt larger than prompt context (n_ctx)

Browse source
This commit is contained in:
ydlme 2023-11-10 22:53:25 +08:00
parent df9d1293de
commit f034effa22
1 changed files with 7 additions and 0 deletions
Download patch file Download diff file Expand all files Collapse all files

7
examples/server/server.cpp
View file

@ -1560,6 +1560,13 @@ struct llama_server_context
if (!slot.params.cache_prompt) if (!slot.params.cache_prompt)
{ {
llama_sampling_reset(slot.ctx_sampling); llama_sampling_reset(slot.ctx_sampling);
// if input prompt is too big, truncate it
if (slot.num_prompt_tokens >= slot.n_ctx)
{
slot.num_prompt_tokens = slot.n_ctx - 1;
prompt_tokens = std::vector<llama_token>(prompt_tokens.end() - slot.num_prompt_tokens, prompt_tokens.end());
slot.truncated = true;
}
slot.n_past = 0; slot.n_past = 0;
slot.num_prompt_tokens_processed = slot.num_prompt_tokens; slot.num_prompt_tokens_processed = slot.num_prompt_tokens;