AI/gpt4all
1
0
Fork 0
mirror of https://github.com/nomic-ai/gpt4all.git synced 2024-09-19 23:35:41 +00:00
Code Issues Packages Projects Releases Wiki Activity

Use ggml scratch bufs for mpt and gptj models (#1104)

Browse Source 
* backend/gptj: use scratch buffers

reduces total memory required and makes eval buf not grow with n_past

* backend/mpt: use scratch bufs

* fix format-related compile warnings
This commit is contained in:
Aaron Miller 2023-06-30 10:53:45 -07:00 committed by GitHub
parent 70cbff70cc
commit 40a3faeb05
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
2 changed files with 44 additions and 24 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

35
gpt4all-backend/gptj.cpp
View File

@ -5,6 +5,7 @@
#include "llmodel_shared.h"
#include <cassert>
#include <cinttypes>
#include <cmath>
#include <cstdio>
#include <cstring>
@ -85,7 +86,9 @@ struct gptj_model {
struct ggml_context * ctx;
std::map<std::string, struct ggml_tensor *> tensors;
llm_buffer buf;
llm_buffer eval_buf;
llm_buffer scr0_buf;
llm_buffer scr1_buf;
~gptj_model() {
if (ctx) {
@ -393,7 +396,7 @@ bool gptj_model_load(const std::string &fname, std::istream &fin, gptj_model & m
}
if (tensor->ne[0] != ne[0] || tensor->ne[1] != ne[1]) {
fprintf(stderr, "%s: tensor '%s' has wrong shape in model file: got [%lu, %lu], expected [%d, %d]\n",
fprintf(stderr, "%s: tensor '%s' has wrong shape in model file: got [%" PRId64 ", %" PRId64 "], expected [%d, %d]\n",
__func__, name.data(), tensor->ne[0], tensor->ne[1], ne[0], ne[1]);
return false;
}
@ -438,6 +441,9 @@ bool gptj_model_load(const std::string &fname, std::istream &fin, gptj_model & m
printf("%s: model size = %8.2f MB / num tensors = %d\n", __func__, total_size/1024.0/1024.0, n_tensors);
}
model.scr0_buf.resize(256u * 1024 * 1024);
model.scr1_buf.resize(256u * 1024 * 1024);
return true;
}
@ -484,24 +490,24 @@ bool gptj_eval(
const int n_rot = hparams.n_rot;
const size_t init_buf_size = 1024_MiB;
if (!model.buf.addr || model.buf.size < init_buf_size)
model.buf.resize(init_buf_size);
if (!model.eval_buf.addr || model.eval_buf.size < init_buf_size)
model.eval_buf.resize(init_buf_size);
if (mem_per_token > 0 && mem_per_token*N > model.buf.size) {
if (mem_per_token > 0 && mem_per_token*N > model.eval_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__, model.buf.size, buf_size_new);
printf("\n%s: reallocating buffer from %zu to %zu bytes\n", __func__, model.eval_buf.size, buf_size_new);
// reallocate
model.buf.resize(buf_size_new);
if (model.buf.addr == nullptr) {
fprintf(stderr, "%s: failed to allocate %zu bytes\n", __func__, model.buf.size);
model.eval_buf.resize(buf_size_new);
if (model.eval_buf.addr == nullptr) {
fprintf(stderr, "%s: failed to allocate %zu bytes\n", __func__, model.eval_buf.size);
return false;
}
}
struct ggml_init_params params = {
.mem_size = model.buf.size,
.mem_buffer = model.buf.addr,
.mem_size = model.eval_buf.size,
.mem_buffer = model.eval_buf.addr,
.no_alloc = false
};
@ -517,7 +523,7 @@ bool gptj_eval(
for (int il = 0; il < n_layer; ++il) {
struct ggml_tensor * cur;
ggml_set_scratch(ctx0, {0, model.scr0_buf.size, model.scr0_buf.addr, });
// norm
{
cur = ggml_norm(ctx0, inpL);
@ -612,6 +618,7 @@ bool gptj_eval(
struct ggml_tensor * inpFF = cur;
ggml_set_scratch(ctx0, {0, model.scr1_buf.size, model.scr1_buf.addr, });
// feed-forward network
// this is independent of the self-attention result, so it could be done in parallel to the self-attention
{
@ -645,6 +652,8 @@ bool gptj_eval(
inpL = ggml_add(ctx0, cur, inpL);
}
ggml_set_scratch(ctx0, {0, model.scr0_buf.size, model.scr0_buf.addr, });
// norm
{
inpL = ggml_norm(ctx0, inpL);
@ -657,6 +666,8 @@ bool gptj_eval(
ggml_repeat(ctx0, model.ln_f_b, inpL));
}
ggml_set_scratch(ctx0, { 0, 0, nullptr, });
// lm_head
{
inpL = ggml_mul_mat(ctx0, model.lmh_g, inpL);

33
gpt4all-backend/mpt.cpp
View File

@ -5,6 +5,7 @@
#include "llmodel_shared.h"
#include <cassert>
#include <cinttypes>
#include <cmath>
#include <cstdio>
#include <cstring>
@ -80,7 +81,9 @@ struct mpt_model {
std::map<std::string, struct ggml_tensor *> tensors;
llm_buffer buf;
llm_buffer eval_buf;
llm_buffer scr0_buf;
llm_buffer scr1_buf;
~mpt_model() {
if (ctx) {
@ -370,8 +373,8 @@ bool mpt_model_load(const std::string &fname, std::istream &fin, mpt_model & mod
}
if (tensor->ne[0] != ne[0] || tensor->ne[1] != ne[1]) {
fprintf(stderr, "%s: tensor '%s' has wrong shape in model file: got [%d, %d], expected [%d, %d]\n",
__func__, name.data(), (int) tensor->ne[0], (int) tensor->ne[1], ne[0], ne[1]);
fprintf(stderr, "%s: tensor '%s' has wrong shape in model file: got [%" PRId64 ", %" PRId64 "], expected [%d, %d]\n",
__func__, name.data(), tensor->ne[0], tensor->ne[1], ne[0], ne[1]);
return false;
}
@ -403,6 +406,9 @@ bool mpt_model_load(const std::string &fname, std::istream &fin, mpt_model & mod
printf("%s: model size = %8.2f MB / num tensors = %d\n", __func__, total_size/1024.0/1024.0, n_tensors);
}
model.scr0_buf.resize(256u * 1024 * 1024);
model.scr1_buf.resize(256u * 1024 * 1024);
return true;
}
@ -438,24 +444,24 @@ bool mpt_eval(
const int n_vocab = hparams.n_vocab;
const size_t init_buf_size = 1024_MiB;
if (!model.buf.addr || model.buf.size < init_buf_size)
model.buf.resize(init_buf_size);
if (!model.eval_buf.addr || model.eval_buf.size < init_buf_size)
model.eval_buf.resize(init_buf_size);
if (mem_per_token > 0 && mem_per_token*N > model.buf.size) {
if (mem_per_token > 0 && mem_per_token*N > model.eval_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__, model.buf.size, buf_size_new);
// reallocate
model.buf.resize(buf_size_new);
if (model.buf.addr == nullptr) {
fprintf(stderr, "%s: failed to allocate %zu bytes\n", __func__, model.buf.size);
model.eval_buf.resize(buf_size_new);
if (model.eval_buf.addr == nullptr) {
fprintf(stderr, "%s: failed to allocate %zu bytes\n", __func__, model.eval_buf.size);
return false;
}
}
struct ggml_init_params params = {
.mem_size = model.buf.size,
.mem_buffer = model.buf.addr,
.mem_size = model.eval_buf.size,
.mem_buffer = model.eval_buf.addr,
.no_alloc = false
};
@ -470,6 +476,7 @@ bool mpt_eval(
struct ggml_tensor * inpL = ggml_get_rows(ctx0, model.wte, embd);
for (int il = 0; il < n_layer; ++il) {
ggml_set_scratch(ctx0, {0, model.scr0_buf.size, model.scr0_buf.addr, });
struct ggml_tensor * inpSA = inpL;
struct ggml_tensor * cur = inpSA;
@ -561,7 +568,7 @@ bool mpt_eval(
cur);
}
ggml_set_scratch(ctx0, {0, model.scr1_buf.size, model.scr1_buf.addr, });
// residual
struct ggml_tensor * resSA = ggml_add(ctx0, cur, inpSA);
// feed-forward network
@ -586,6 +593,7 @@ bool mpt_eval(
// self-attention + FF
inpL = ggml_add(ctx0, cur, resSA);
}
ggml_set_scratch(ctx0, {0, model.scr0_buf.size, model.scr0_buf.addr, });
struct ggml_tensor * out = inpL;
// -> logits
@ -594,6 +602,7 @@ bool mpt_eval(
out = ggml_mul(ctx0,
ggml_repeat(ctx0, model.norm_f_w, out),
out);
ggml_set_scratch(ctx0, { 0, 0, nullptr, });
out = ggml_mul_mat(ctx0, model.wte, out);
}