vrr/koboldcpp
2
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2025-09-11 01:24:36 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

bug fixes for openblas

Browse source
This commit is contained in:
Concedo 2023-06-23 22:45:22 +08:00
parent 43c2891afa
commit f39a746089
5 changed files with 40 additions and 70 deletions
Download patch file Download diff file Expand all files Collapse all files

21
otherarch/neox_v3.cpp
View file

@ -401,7 +401,7 @@ bool gpt_neox_eval(
const std::vector<gpt_vocab::id> & embd_inp,
std::vector<float> & embd_w,
size_t & mem_per_token,
bool use_scratch=true) {
bool use_scratch) {
const int N = embd_inp.size();
const auto & hparams = model.hparams;
@ -418,21 +418,14 @@ bool gpt_neox_eval(
// use 2 scratch buffers
// TODO: very hacky solution - reimplement in a more elegant way
static size_t scr0_size = (n_ctx>1024?512u:256u)*1024*1024;
static void * scr0;
static size_t scr0_size = (n_embd>2400?512u:256u)*1024*1024;
static size_t scr1_size = (n_embd>2400?512u:256u)*1024*1024;
static size_t scr1_size = (n_ctx>1024?512u:256u)*1024*1024;
static void * scr1;
if(use_scratch)
{
scr0 = malloc(scr0_size);
scr1 = malloc(scr1_size);
}
static void * scr0 = malloc(scr0_size);
static void * scr1 = malloc(scr1_size);
size_t scratch_needed_mem = mem_per_token*N;
if (mem_per_token > 0 && scratch_needed_mem*1.1 > buf_size) {
const size_t buf_size_new = 64u*1024*1024 + 1.2*(scratch_needed_mem); // add 10% to account for ggml object overhead
if (mem_per_token > 0 && (mem_per_token*N*2 + 64u*1024*1024) > buf_size) {
const size_t buf_size_new = 360u*1024*1024 + 1.2*(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