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

add an extra buffer to mem allocations

Browse source
This commit is contained in:
Concedo 2023-04-23 00:04:32 +08:00
parent 7c60441d71
commit d2f14b2b1f
3 changed files with 8 additions and 5 deletions
Download patch file Download diff file Expand all files Collapse all files

3
otherarch/gpt2_v2.cpp
View file

@ -372,10 +372,11 @@ bool gpt2_eval(
//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 void * buf = malloc(buf_size);
if (mem_per_token > 0 && mem_per_token*N*1.6 > buf_size) {
const size_t buf_size_new = 2*(mem_per_token*N); // add 10% to account for ggml object overhead
const size_t buf_size_new = extra_buf + 2*(mem_per_token*N); // add 10% to account for ggml object overhead
//printf("\n%s: reallocating buffer from %zu to %zu bytes\n", __func__, buf_size, buf_size_new);
// reallocate

5
otherarch/gptj_v2.cpp
View file

@ -383,10 +383,11 @@ bool gptj_eval(
//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 void * buf = malloc(buf_size);
if (mem_per_token > 0 && mem_per_token*N*1.4 > buf_size) {
const size_t buf_size_new = 1.6*(mem_per_token*N); // add 10% to account for ggml object overhead
if (mem_per_token > 0 && mem_per_token*N*1.5 > buf_size) {
const size_t buf_size_new = extra_buf + 1.6*(mem_per_token*N); // add 10% to account for ggml object overhead
//printf("\n%s: reallocating buffer from %zu to %zu bytes\n", __func__, buf_size, buf_size_new);
// reallocate

5
otherarch/neox.cpp
View file

@ -365,10 +365,11 @@ bool stablelm_eval(
const int n_rot = hparams.n_rot;
static size_t buf_size = 256u*1024*1024;
const size_t extra_buf = 64u*1024*1024;
static void * buf = malloc(buf_size);
if (mem_per_token > 0 && mem_per_token*N*1.2 > buf_size) {
const size_t buf_size_new = 1.4*(mem_per_token*N); // add 10% to account for ggml object overhead
if (mem_per_token > 0 && mem_per_token*N*1.5 > buf_size) {
const size_t buf_size_new = extra_buf + 1.6*(mem_per_token*N); // add 10% to account for ggml object overhead
//printf("\n%s: reallocating buffer from %zu to %zu bytes\n", __func__, buf_size, buf_size_new);
// reallocate