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

server : don't overfill the batch during infill (#10018)

Browse source 
ggml-ci
This commit is contained in:
Georgi Gerganov 2024-10-28 08:49:32 +02:00 • committed by GitHub
parent 8841ce3f43
commit 8125e6cbfc
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
2 changed files with 5 additions and 2 deletions
Download patch file Download diff file Expand all files Collapse all files

6
examples/server/utils.hpp
View file

@ -266,8 +266,10 @@ static llama_tokens format_infill(
}
// for now pick FIM context to fit in a batch (ratio prefix:suffix = 3:1, TODO: configurable?)
const int n_suffix_take = std::min<int>(tokens_suffix.size(), (n_batch/4));
const int n_prefix_take = std::min<int>(tokens_prefix.size(), 3*(n_batch/4) - 3);
const int n_prefix_take = std::min<int>(tokens_prefix.size(), 3*(n_batch/4));
const int n_suffix_take = std::min<int>(tokens_suffix.size(), std::max<int>(0, (n_batch/4) - (2 + tokens_prompt.size())));
SRV_DBG("n_prefix_take = %d, n_suffix_take = %d, total = %d\n", n_prefix_take, n_suffix_take, (n_prefix_take + n_suffix_take));
// fill the rest of the context with extra chunks
const int n_extra_take = std::min<int>(std::max<int>(0, n_ctx - (n_batch) - 2*n_predict), extra_tokens.size());