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

increase initial buffer too

Browse source
This commit is contained in:
Concedo 2023-04-23 00:07:33 +08:00
parent d2f14b2b1f
commit b5d6284190
3 changed files with 3 additions and 3 deletions
Download patch file Download diff file Expand all files Collapse all files

2
otherarch/gpt2_v2.cpp
View file

@ -371,8 +371,8 @@ bool gpt2_eval(
const int n_vocab = hparams.n_vocab;
//todo: there is a bug that causes the buffer to oom and I cannot figure it out, hack to increase size for now
static size_t buf_size = 256u*1024*1024;
const size_t extra_buf = 64u*1024*1024;
static size_t buf_size = 256u*1024*1024 + extra_buf;
static void * buf = malloc(buf_size);
if (mem_per_token > 0 && mem_per_token*N*1.6 > buf_size) {

2
otherarch/gptj_v2.cpp
View file

@ -382,8 +382,8 @@ bool gptj_eval(
const int d_key = n_embd/n_head;
//todo: there is a bug that causes the buffer to oom and I cannot figure it out, hack to increase size for now
static size_t buf_size = 256u*1024*1024;
const size_t extra_buf = 64u*1024*1024;
static size_t buf_size = 256u*1024*1024 + extra_buf;
static void * buf = malloc(buf_size);
if (mem_per_token > 0 && mem_per_token*N*1.5 > buf_size) {

2
otherarch/neox.cpp
View file

@ -364,8 +364,8 @@ bool stablelm_eval(
const int n_vocab = hparams.n_vocab;
const int n_rot = hparams.n_rot;
static size_t buf_size = 256u*1024*1024;
const size_t extra_buf = 64u*1024*1024;
static size_t buf_size = 256u*1024*1024 + extra_buf;
static void * buf = malloc(buf_size);
if (mem_per_token > 0 && mem_per_token*N*1.5 > buf_size) {