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

fix perplexity - it's memory needs dont grow, so we skip it

Browse source
This commit is contained in:
Green Sky 2023-03-23 20:50:09 +01:00
parent 424281a4fb
commit 2d262ea9f0
No known key found for this signature in database
2 changed files with 2 additions and 2 deletions
Download patch file Download diff file Expand all files Collapse all files

2
llama.cpp
View file

@ -632,7 +632,7 @@ static bool llama_eval_internal(
auto & mem_at_token1 = lctx.mem_at_token1; auto & mem_at_token1 = lctx.mem_at_token1;
// TODO: fix this hardcoded size // TODO: fix this hardcoded size
static size_t buf_size = size_t(n_ctx)*1024*1024; static size_t buf_size = size_t(n_ctx)*size_t(N)*128*1024;
static void * buf = malloc(buf_size); static void * buf = malloc(buf_size);
const size_t C0 = mem_at_token0; // ~base const size_t C0 = mem_at_token0; // ~base

2
main.cpp
View file

@ -219,7 +219,7 @@ int main(int argc, char ** argv) {
// (fill in mem_at_token0 and mem_at_token1) // (fill in mem_at_token0 and mem_at_token1)
// TODO: better way to do that // TODO: better way to do that
// TODO(Green-Sky): move to internal and detect first time usage // TODO(Green-Sky): move to internal and detect first time usage
{ if (!params.perplexity) { // perplexity does not grow over time
// we make 2 evals, of batchsize to take 2 measurements, to determine base and growth // we make 2 evals, of batchsize to take 2 measurements, to determine base and growth
std::vector<llama_token> tmp(params.n_batch*2, 2); std::vector<llama_token> tmp(params.n_batch*2, 2);
tmp[0] = llama_token_bos(); tmp[0] = llama_token_bos();