AI/turbopilot
1
0
Fork 0
mirror of https://github.com/ravenscroftj/turbopilot.git synced 2024-10-01 01:06:01 -04:00
Code Issues Packages Projects Releases Wiki Activity

Merge pull request #67 from ravenscroftj/fix/mac-memory-usage

Browse Source 
Temporary fix for stablecode and starcoder on mac
This commit is contained in:
James Ravenscroft 2023-08-28 11:55:55 +01:00 committed by GitHub
commit eaeb52fcb0
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
2 changed files with 48 additions and 23 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

29
src/gptneox.cpp
View File

@ -84,20 +84,21 @@ bool gpt_neox_eval(
const int n_vocab = hparams.n_vocab;
const int n_rot = hparams.n_rot;
static size_t buf_size = 256u*1024*1024;
static size_t buf_size = 512u*1024*1024;
static void * buf = malloc(buf_size);
// use 2 scratch buffers
// TODO: very hacky solution - reimplement in a more elegant way
static size_t scr0_size = 256u*1024*1024;
static size_t scr0_size = 512*1024*1024;
static void * scr0 = malloc(scr0_size);
static size_t scr1_size = 256u*1024*1024;
static size_t scr1_size = 512*1024*1024;
static void * scr1 = malloc(scr1_size);
if (mem_per_token > 0 && mem_per_token*N > buf_size) {
if (mem_per_token > 0 && (mem_per_token*N) > buf_size) {
const size_t buf_size_new = 1.1*(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);
spdlog::debug("{}: reallocating buffer from {} to {} bytes\n", __func__, buf_size, buf_size_new);
// reallocate
@ -107,8 +108,6 @@ bool gpt_neox_eval(
fprintf(stderr, "%s: failed to allocate %zu bytes\n", __func__, buf_size);
return false;
}
spdlog::debug("{}: reallocating context buffer {} -> now {} bytes of tokens in prompt = {}", __func__, buf_size, buf_size_new);
}
struct ggml_init_params params = {
@ -303,12 +302,14 @@ bool gpt_neox_eval(
embd_w.resize(n_vocab);
memcpy(embd_w.data(), (float *) ggml_get_data(inpL) + (n_vocab*(N-1)), sizeof(float)*n_vocab);
spdlog::debug("used_mem = {}\n", ggml_used_mem(ctx0));
if (mem_per_token == 0) {
mem_per_token = ggml_used_mem(ctx0)/N;
mem_per_token = ggml_used_mem(ctx0) / N; //* 4;
spdlog::debug("Set mem_per_token={} / {} * {} = {}", ggml_used_mem(ctx0), N, 4, mem_per_token);
}
spdlog::debug("used_mem = {}\n", ggml_used_mem(ctx0));
//printf("used_mem = %zu\n", ggml_used_mem(ctx0));
ggml_free(ctx0);
@ -685,7 +686,13 @@ std::stringstream GPTNEOXModel::predict_impl(std::string prompt, int max_length,
std::vector<float> logits;
gpt_neox_eval((*model), config.n_threads, 0, { 0, 1, 2, 3 }, logits, mem_per_token);
std::vector<gpt_vocab::id> test = {};
for(int i=0;i<64;i++){
test.push_back(i);
}
gpt_neox_eval((*model), config.n_threads, 0, test, logits, mem_per_token);
const int64_t t_start_us = ggml_time_us();

28
src/starcoder.cpp
View File

@ -37,7 +37,7 @@ bool starcoder_eval(
const int n_head = hparams.n_head;
const int n_vocab = hparams.n_vocab;
static size_t buf_size = 256u*1024*1024;
static size_t buf_size = 512u*1024*1024;
static void * buf = malloc(buf_size);
// use 2 scratch buffers
@ -48,8 +48,10 @@ bool starcoder_eval(
static size_t scr1_size = 512u*1024*1024;
static void * scr1 = malloc(scr1_size);
if (mem_per_token > 0 && mem_per_token*N > buf_size) {
const size_t buf_size_new = 1.1*(mem_per_token*N); // add 10% to account for ggml object overhead
if (mem_per_token > 0 && 2*mem_per_token*N > buf_size) {
const size_t buf_size_new = 2*(mem_per_token*N); // add 10% to account for ggml object overhead
if(buf_size_new > buf_size){
spdlog::debug("{}: reallocating buffer from {} to {} bytes\n", __func__, buf_size, buf_size_new);
// reallocate
@ -61,12 +63,15 @@ bool starcoder_eval(
}
}
}
struct ggml_init_params params = {
/*.mem_size =*/ buf_size,
/*.mem_buffer =*/ buf,
/*.no_alloc =*/ false,
};
struct ggml_context * ctx0 = ggml_init(params);
struct ggml_cgraph gf = {};
@ -338,7 +343,9 @@ bool starcoder_eval(
if (mem_per_token == 0) {
mem_per_token = ggml_used_mem(ctx0)/N;
}
//printf("used_mem = %zu MB\n", ggml_used_mem(ctx0)/(1024*1024));
spdlog::debug("{}: used mem buf={} bytes", __func__, ggml_used_mem(ctx0));
ggml_free(ctx0);
@ -743,11 +750,22 @@ std::stringstream StarcoderModel::predict_impl(std::string prompt, int max_lengt
size_t mem_per_token = 0;
std::vector<float> logits;
std::vector<gpt_vocab::id> test = {};
for(int i=0;i<64;i++){
test.push_back(i);
}
spdlog::debug("{}: calculate required memory per token", __func__);
starcoder_eval((*model), config.n_threads, 0, test, logits, mem_per_token);
spdlog::debug("{}: mem_per_token={}", __func__, mem_per_token);
spdlog::debug("{}: total mem needed for prompt = {}*{}={}", __func__, embd_inp.size(), mem_per_token, embd_inp.size()*mem_per_token);
starcoder_eval((*model), config.n_threads, 0, { 0, 1, 2, 3 }, logits, mem_per_token);
for (int i = embd.size(); i < embd_inp.size() + n_predict; i++) {
// predict
spdlog::debug("{}: process token #{}: ", __func__, i);
if (embd.size() > 0) {
const int64_t t_start_us = ggml_time_us();