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

just revert back to the working commit

Browse source
This commit is contained in:
Concedo 2023-04-23 00:35:42 +08:00
parent c60fb5ef4b
commit d41490c27b
3 changed files with 8 additions and 11 deletions
Download patch file Download diff file Expand all files Collapse all files

5
otherarch/gpt2_v2.cpp
View file

@ -371,12 +371,11 @@ 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
const size_t extra_buf = 64u*1024*1024;
static size_t buf_size = 256u*1024*1024 + extra_buf;
static size_t buf_size = 512u*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 = extra_buf + 2*(mem_per_token*N); // add 10% to account for ggml object overhead
const size_t buf_size_new = 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

7
otherarch/gptj_v2.cpp
View file

@ -382,12 +382,11 @@ 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
const size_t extra_buf = 64u*1024*1024;
static size_t buf_size = 256u*1024*1024 + extra_buf;
static size_t buf_size = 512u*1024*1024;
static void * buf = malloc(buf_size);
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
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
//printf("\n%s: reallocating buffer from %zu to %zu bytes\n", __func__, buf_size, buf_size_new);
// reallocate

7
otherarch/neox.cpp
View file

@ -364,12 +364,11 @@ bool stablelm_eval(
const int n_vocab = hparams.n_vocab;
const int n_rot = hparams.n_rot;
const size_t extra_buf = 64u*1024*1024;
static size_t buf_size = 256u*1024*1024 + extra_buf;
static size_t buf_size = 512u*1024*1024;
static void * buf = malloc(buf_size);
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
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
//printf("\n%s: reallocating buffer from %zu to %zu bytes\n", __func__, buf_size, buf_size_new);
// reallocate