gpt4all/gpt4all-backend/gptj.cpp

1077 lines
35 KiB
C++

#define GPTJ_H_I_KNOW_WHAT_I_AM_DOING_WHEN_INCLUDING_THIS_FILE
#include "gptj_impl.h"
#include "utils.h"
#include <cassert>
#include <cmath>
#include <cstdio>
#include <cstring>
#include <fstream>
#include <map>
#include <string>
#include <vector>
#include <iostream>
#if defined(_WIN32) && defined(_MSC_VER)
#define WIN32_LEAN_AND_MEAN
#ifndef NOMINMAX
#define NOMINMAX
#endif
#include <windows.h>
#include <io.h>
#include <stdio.h>
#else
#include <unistd.h>
#endif
#include <sstream>
#include <unordered_set>
#include <ggml.h>
namespace {
const char *modelType_ = "MPT";
static const size_t MB = 1024*1024;
}
// default hparams (GPT-J 6B)
struct gptj_hparams {
int32_t n_vocab = 50400;
int32_t n_ctx = 2048;
int32_t n_embd = 4096;
int32_t n_head = 16;
int32_t n_layer = 28;
int32_t n_rot = 64;
int32_t f16 = 1;
};
struct gptj_layer {
// normalization
struct ggml_tensor * ln_1_g;
struct ggml_tensor * ln_1_b;
// attention
struct ggml_tensor * c_attn_q_proj_w;
struct ggml_tensor * c_attn_k_proj_w;
struct ggml_tensor * c_attn_v_proj_w;
struct ggml_tensor * c_attn_proj_w;
// ff
struct ggml_tensor * c_mlp_fc_w;
struct ggml_tensor * c_mlp_fc_b;
struct ggml_tensor * c_mlp_proj_w;
struct ggml_tensor * c_mlp_proj_b;
};
struct gptj_buffer {
uint8_t * addr = NULL;
size_t size = 0;
void resize(size_t size) {
delete[] addr;
addr = new uint8_t[size];
this->size = size;
}
~gptj_buffer() {
fflush(stdout);
delete[] addr;
}
};
struct gptj_kv_cache {
struct ggml_tensor * k;
struct ggml_tensor * v;
struct ggml_context * ctx = NULL;
gptj_buffer buf;
int n; // number of tokens currently in the cache
~gptj_kv_cache() {
if (ctx) {
ggml_free(ctx);
}
}
};
struct gptj_model {
gptj_hparams hparams;
// normalization
struct ggml_tensor * ln_f_g;
struct ggml_tensor * ln_f_b;
struct ggml_tensor * wte; // position embedding
struct ggml_tensor * lmh_g; // language model head
struct ggml_tensor * lmh_b; // language model bias
std::vector<gptj_layer> layers;
// key + value memory
struct gptj_kv_cache kv_self;
//
struct ggml_context * ctx;
std::map<std::string, struct ggml_tensor *> tensors;
gptj_buffer buf;
~gptj_model() {
if (ctx) {
ggml_free(ctx);
}
}
};
static bool kv_cache_init(
const struct gptj_hparams & hparams,
struct gptj_kv_cache & cache,
ggml_type wtype,
int n_ctx) {
const int n_embd = hparams.n_embd;
const int n_layer = hparams.n_layer;
const int64_t n_mem = (int64_t)n_layer*n_ctx;
const int64_t n_elements = n_embd*n_mem;
cache.buf.resize(2u*n_elements*ggml_type_size(wtype) + 2u*MB);
struct ggml_init_params params;
params.mem_size = cache.buf.size;
params.mem_buffer = cache.buf.addr;
params.no_alloc = false;
cache.ctx = ggml_init(params);
if (!cache.ctx) {
fprintf(stderr, "%s: failed to allocate memory for kv cache\n", __func__);
return false;
}
cache.k = ggml_new_tensor_1d(cache.ctx, wtype, n_elements);
cache.v = ggml_new_tensor_1d(cache.ctx, wtype, n_elements);
return true;
}
// load the model's weights from a stream
bool gptj_model_load(const std::string &fname, std::istream &fin, gptj_model & model, gpt_vocab & vocab) {
printf("%s: loading model from '%s' - please wait ...\n", __func__, fname.c_str());
// verify magic
{
uint32_t magic;
fin.read((char *) &magic, sizeof(magic));
if (magic != 0x67676d6c) {
fprintf(stderr, "%s: invalid model file '%s' (bad magic)\n", __func__, fname.c_str());
return false;
}
}
// load hparams
{
auto & hparams = model.hparams;
fin.read((char *) &hparams.n_vocab, sizeof(hparams.n_vocab));
fin.read((char *) &hparams.n_ctx, sizeof(hparams.n_ctx));
fin.read((char *) &hparams.n_embd, sizeof(hparams.n_embd));
fin.read((char *) &hparams.n_head, sizeof(hparams.n_head));
fin.read((char *) &hparams.n_layer, sizeof(hparams.n_layer));
fin.read((char *) &hparams.n_rot, sizeof(hparams.n_rot));
fin.read((char *) &hparams.f16, sizeof(hparams.f16));
printf("%s: n_vocab = %d\n", __func__, hparams.n_vocab);
printf("%s: n_ctx = %d\n", __func__, hparams.n_ctx);
printf("%s: n_embd = %d\n", __func__, hparams.n_embd);
printf("%s: n_head = %d\n", __func__, hparams.n_head);
printf("%s: n_layer = %d\n", __func__, hparams.n_layer);
printf("%s: n_rot = %d\n", __func__, hparams.n_rot);
printf("%s: f16 = %d\n", __func__, hparams.f16);
}
// load vocab
{
int32_t n_vocab = 0;
fin.read((char *) &n_vocab, sizeof(n_vocab));
if (n_vocab != model.hparams.n_vocab) {
fprintf(stderr, "%s: invalid model file '%s' (bad vocab size %d != %d)\n",
__func__, fname.c_str(), n_vocab, model.hparams.n_vocab);
return false;
}
std::string word;
for (int i = 0; i < n_vocab; i++) {
uint32_t len;
fin.read((char *) &len, sizeof(len));
word.resize(len);
fin.read((char *) word.data(), len);
vocab.token_to_id[word] = i;
vocab.id_to_token[i] = word;
}
}
// for the big tensors, we have the option to store the data in 16-bit floats or quantized
// in order to save memory and also to speed up the computation
ggml_type wtype = GGML_TYPE_COUNT;
switch (model.hparams.f16) {
case 0: wtype = GGML_TYPE_F32; break;
case 1: wtype = GGML_TYPE_F16; break;
case 2: wtype = GGML_TYPE_Q4_0; break;
case 3: wtype = GGML_TYPE_Q4_1; break;
case 5: wtype = GGML_TYPE_Q4_2; break;
default:
{
fprintf(stderr, "%s: invalid model file '%s' (bad f16 value %d)\n",
__func__, fname.c_str(), model.hparams.f16);
return false;
}
}
auto & ctx = model.ctx;
size_t ctx_size = 0;
{
const auto & hparams = model.hparams;
const int n_embd = hparams.n_embd;
const int n_layer = hparams.n_layer;
const int n_ctx = hparams.n_ctx;
const int n_vocab = hparams.n_vocab;
ctx_size += n_embd*ggml_type_sizef(GGML_TYPE_F32); // ln_f_g
ctx_size += n_embd*ggml_type_sizef(GGML_TYPE_F32); // ln_f_b
ctx_size += n_embd*n_vocab*ggml_type_sizef(wtype); // wte
ctx_size += n_embd*n_vocab*ggml_type_sizef(wtype); // lmh_g
ctx_size += n_vocab*ggml_type_sizef(GGML_TYPE_F32); // lmh_b
ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_1_g
ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_1_b
ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype)); // c_attn_q_proj_w
ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype)); // c_attn_k_proj_w
ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype)); // c_attn_v_proj_w
ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype)); // c_attn_proj_w
ctx_size += n_layer*(4*n_embd*n_embd*ggml_type_sizef(wtype)); // c_mlp_fc_w
ctx_size += n_layer*( 4*n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_mlp_fc_b
ctx_size += n_layer*(4*n_embd*n_embd*ggml_type_sizef(wtype)); // c_mlp_proj_w
ctx_size += n_layer*( n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_mlp_proj_b
ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(GGML_TYPE_F32); // memory_k
ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(GGML_TYPE_F32); // memory_v
ctx_size += (5 + 10*n_layer)*256; // object overhead
printf("%s: ggml ctx size = %6.2f MB\n", __func__, ctx_size/(1024.0*1024.0));
}
// create the ggml context
{
struct ggml_init_params params = {
.mem_size = ctx_size,
.mem_buffer = NULL,
.no_alloc = false
};
model.ctx = ggml_init(params);
if (!model.ctx) {
fprintf(stderr, "%s: ggml_init() failed\n", __func__);
return false;
}
}
// prepare memory for the weights
{
const auto & hparams = model.hparams;
const int n_embd = hparams.n_embd;
const int n_layer = hparams.n_layer;
const int n_vocab = hparams.n_vocab;
model.layers.resize(n_layer);
model.wte = ggml_new_tensor_2d(ctx, wtype, n_embd, n_vocab);
model.ln_f_g = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
model.ln_f_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
model.lmh_g = ggml_new_tensor_2d(ctx, wtype, n_embd, n_vocab);
model.lmh_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_vocab);
// map by name
model.tensors["transformer.wte.weight"] = model.wte;
model.tensors["transformer.ln_f.weight"] = model.ln_f_g;
model.tensors["transformer.ln_f.bias"] = model.ln_f_b;
model.tensors["lm_head.weight"] = model.lmh_g;
model.tensors["lm_head.bias"] = model.lmh_b;
for (int i = 0; i < n_layer; ++i) {
auto & layer = model.layers[i];
layer.ln_1_g = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
layer.ln_1_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
layer.c_attn_q_proj_w = ggml_new_tensor_2d(ctx, wtype, n_embd, n_embd);
layer.c_attn_k_proj_w = ggml_new_tensor_2d(ctx, wtype, n_embd, n_embd);
layer.c_attn_v_proj_w = ggml_new_tensor_2d(ctx, wtype, n_embd, n_embd);
layer.c_attn_proj_w = ggml_new_tensor_2d(ctx, wtype, n_embd, n_embd);
layer.c_mlp_fc_w = ggml_new_tensor_2d(ctx, wtype, n_embd, 4*n_embd);
layer.c_mlp_fc_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 4*n_embd);
layer.c_mlp_proj_w = ggml_new_tensor_2d(ctx, wtype, 4*n_embd, n_embd);
layer.c_mlp_proj_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
// map by name
model.tensors["transformer.h." + std::to_string(i) + ".ln_1.weight"] = layer.ln_1_g;
model.tensors["transformer.h." + std::to_string(i) + ".ln_1.bias"] = layer.ln_1_b;
model.tensors["transformer.h." + std::to_string(i) + ".attn.q_proj.weight"] = layer.c_attn_q_proj_w;
model.tensors["transformer.h." + std::to_string(i) + ".attn.k_proj.weight"] = layer.c_attn_k_proj_w;
model.tensors["transformer.h." + std::to_string(i) + ".attn.v_proj.weight"] = layer.c_attn_v_proj_w;
model.tensors["transformer.h." + std::to_string(i) + ".attn.out_proj.weight"] = layer.c_attn_proj_w;
model.tensors["transformer.h." + std::to_string(i) + ".mlp.fc_in.weight"] = layer.c_mlp_fc_w;
model.tensors["transformer.h." + std::to_string(i) + ".mlp.fc_in.bias"] = layer.c_mlp_fc_b;
model.tensors["transformer.h." + std::to_string(i) + ".mlp.fc_out.weight"] = layer.c_mlp_proj_w;
model.tensors["transformer.h." + std::to_string(i) + ".mlp.fc_out.bias"] = layer.c_mlp_proj_b;
}
}
// key + value memory
{
const auto & hparams = model.hparams;
if (!kv_cache_init(hparams, model.kv_self, GGML_TYPE_F16, model.hparams.n_ctx)) {
fprintf(stderr, "%s: kv_cache_init() failed for self-attention cache\n", __func__);
ggml_free(ctx);
return false;
}
const size_t memory_size = ggml_nbytes(model.kv_self.k) + ggml_nbytes(model.kv_self.v);
printf("%s: kv self size = %7.2f MB\n", __func__, memory_size / 1024.0 / 1024.0);
}
// load weights
{
int n_tensors = 0;
size_t total_size = 0;
printf("%s: ", __func__);
while (true) {
int32_t n_dims;
int32_t length;
int32_t ftype;
fin.read(reinterpret_cast<char *>(&n_dims), sizeof(n_dims));
fin.read(reinterpret_cast<char *>(&length), sizeof(length));
fin.read(reinterpret_cast<char *>(&ftype), sizeof(ftype));
if (fin.eof()) {
break;
}
int32_t nelements = 1;
int32_t ne[2] = { 1, 1 };
for (int i = 0; i < n_dims; ++i) {
fin.read(reinterpret_cast<char *>(&ne[i]), sizeof(ne[i]));
nelements *= ne[i];
}
std::string name(length, 0);
fin.read(&name[0], length);
if (model.tensors.find(name.data()) == model.tensors.end()) {
fprintf(stderr, "%s: unknown tensor '%s' in model file\n", __func__, name.data());
return false;
}
auto tensor = model.tensors[name.data()];
if (ggml_nelements(tensor) != nelements) {
fprintf(stderr, "%s: tensor '%s' has wrong size in model file\n", __func__, name.data());
return false;
}
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",
__func__, name.data(), tensor->ne[0], tensor->ne[1], ne[0], ne[1]);
return false;
}
if (0) {
static const char * ftype_str[] = { "f32", "f16", "q4_0", "q4_1", };
printf("%24s - [%5d, %5d], type = %6s, %6.2f MB, %9zu bytes\n", name.data(), ne[0], ne[1], ftype_str[ftype], ggml_nbytes(tensor)/1024.0/1024.0, ggml_nbytes(tensor));
}
size_t bpe = 0;
switch (ftype) {
case 0: bpe = ggml_type_size(GGML_TYPE_F32); break;
case 1: bpe = ggml_type_size(GGML_TYPE_F16); break;
case 2: bpe = ggml_type_size(GGML_TYPE_Q4_0); assert(ne[0] % 64 == 0); break;
case 3: bpe = ggml_type_size(GGML_TYPE_Q4_1); assert(ne[0] % 64 == 0); break;
default:
{
fprintf(stderr, "%s: unknown ftype %d in model file\n", __func__, ftype);
return false;
}
};
if ((nelements*bpe)/ggml_blck_size(tensor->type) != ggml_nbytes(tensor)) {
fprintf(stderr, "%s: tensor '%s' has wrong size in model file: got %zu, expected %zu\n",
__func__, name.data(), ggml_nbytes(tensor), nelements*bpe);
return false;
}
fin.read(reinterpret_cast<char *>(tensor->data), ggml_nbytes(tensor));
//printf("%42s - [%5d, %5d], type = %6s, %6.2f MB\n", name.data(), ne[0], ne[1], ftype == 0 ? "float" : "f16", ggml_nbytes(tensor)/1024.0/1024.0);
total_size += ggml_nbytes(tensor);
if (++n_tensors % 8 == 0) {
printf(".");
fflush(stdout);
}
}
printf(" done\n");
printf("%s: model size = %8.2f MB / num tensors = %d\n", __func__, total_size/1024.0/1024.0, n_tensors);
}
return true;
}
// load the model's weights from a file path
bool gptj_model_load(const std::string & fname, gptj_model & model, gpt_vocab & vocab) {
auto fin = std::ifstream(fname, std::ios::binary);
if (!fin) {
fprintf(stderr, "%s: failed to open '%s'\n", __func__, fname.c_str());
return false;
}
bool loaded = gptj_model_load(fname, fin, model, vocab);
fin.close();
return loaded;
}
// evaluate the transformer
//
// - model: the model
// - n_threads: number of threads to use
// - n_past: the context size so far
// - embd_inp: the embeddings of the tokens in the context
// - embd_w: the predicted logits for the next token
//
// The GPT-J model requires about 16MB of memory per input token.
//
bool gptj_eval(
gptj_model & model,
const int n_threads,
const int n_past,
const std::vector<gpt_vocab::id> & embd_inp,
std::vector<float> & embd_w,
size_t & mem_per_token) {
const int N = embd_inp.size();
const auto & hparams = model.hparams;
const int n_embd = hparams.n_embd;
const int n_layer = hparams.n_layer;
const int n_ctx = hparams.n_ctx;
const int n_head = hparams.n_head;
const int n_vocab = hparams.n_vocab;
const int n_rot = hparams.n_rot;
const size_t init_buf_size = 1024u*MB;
if (!model.buf.addr || model.buf.size < init_buf_size)
model.buf.resize(init_buf_size);
if (mem_per_token > 0 && mem_per_token*N > model.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);
return false;
}
}
struct ggml_init_params params = {
.mem_size = model.buf.size,
.mem_buffer = model.buf.addr,
.no_alloc = false
};
struct ggml_context * ctx0 = ggml_init(params);
struct ggml_cgraph gf = {};
gf.n_threads = n_threads;
struct ggml_tensor * embd = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
memcpy(embd->data, embd_inp.data(), N*ggml_element_size(embd));
// wte
struct ggml_tensor * inpL = ggml_get_rows(ctx0, model.wte, embd);
for (int il = 0; il < n_layer; ++il) {
struct ggml_tensor * cur;
// norm
{
cur = ggml_norm(ctx0, inpL);
// cur = ln_1_g*cur + ln_1_b
cur = ggml_add(ctx0,
ggml_mul(ctx0,
ggml_repeat(ctx0, model.layers[il].ln_1_g, cur),
cur),
ggml_repeat(ctx0, model.layers[il].ln_1_b, cur));
}
struct ggml_tensor * inpSA = cur;
// self-attention
{
struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].c_attn_q_proj_w, cur);
struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].c_attn_k_proj_w, cur);
struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].c_attn_v_proj_w, cur);
// store key and value to memory
{
struct ggml_tensor * k = ggml_view_1d(ctx0, model.kv_self.k, N*n_embd, (ggml_element_size(model.kv_self.k)*n_embd)*(il*n_ctx + n_past));
struct ggml_tensor * v = ggml_view_1d(ctx0, model.kv_self.v, N*n_embd, (ggml_element_size(model.kv_self.v)*n_embd)*(il*n_ctx + n_past));
ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Kcur, k));
ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Vcur, v));
}
// Q = Qcur.contiguous().view(n_embd/n_head, n_head, N).permute(0, 2, 1, 3)
struct ggml_tensor * Q =
ggml_permute(ctx0,
ggml_rope(ctx0,
ggml_cpy(ctx0,
Qcur,
ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_embd/n_head, n_head, N)),
n_past, n_rot, 0),
0, 2, 1, 3);
// K = Kmem.view(n_embd/n_head, n_head, n_past + N).permute(0, 2, 1, 3)
struct ggml_tensor * K =
ggml_permute(ctx0,
ggml_rope(ctx0,
ggml_reshape_3d(ctx0,
ggml_view_1d(ctx0, model.kv_self.k, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(model.kv_self.k)*n_embd),
n_embd/n_head, n_head, n_past + N),
n_past, n_rot, 1),
0, 2, 1, 3);
// K * Q
struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q);
// KQ_scaled = KQ / sqrt(n_embd/n_head)
struct ggml_tensor * KQ_scaled =
ggml_scale(ctx0,
KQ,
ggml_new_f32(ctx0, 1.0f/sqrt(float(n_embd)/n_head))
);
// KQ_masked = mask_past(KQ_scaled)
struct ggml_tensor * KQ_masked = ggml_diag_mask_inf(ctx0, KQ_scaled, n_past);
// KQ = soft_max(KQ_masked)
struct ggml_tensor * KQ_soft_max = ggml_soft_max(ctx0, KQ_masked);
// V_trans = Vmem.view(n_embd/n_head, n_head, n_past + N).permute(1, 2, 0, 3).contiguous()
struct ggml_tensor * V_trans =
ggml_cpy(ctx0,
ggml_permute(ctx0,
ggml_reshape_3d(ctx0,
ggml_view_1d(ctx0, model.kv_self.v, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(model.kv_self.v)*n_embd),
n_embd/n_head, n_head, n_past + N),
1, 2, 0, 3),
ggml_new_tensor_3d(ctx0, model.kv_self.v->type, n_past + N, n_embd/n_head, n_head));
// KQV = transpose(V) * KQ_soft_max
struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V_trans, KQ_soft_max);
// KQV_merged = KQV.permute(0, 2, 1, 3)
struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
// cur = KQV_merged.contiguous().view(n_embd, N)
cur = ggml_cpy(ctx0,
KQV_merged,
ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, N));
// projection (no bias)
cur = ggml_mul_mat(ctx0,
model.layers[il].c_attn_proj_w,
cur);
}
struct ggml_tensor * inpFF = cur;
// feed-forward network
// this is independent of the self-attention result, so it could be done in parallel to the self-attention
{
// note here we pass inpSA instead of cur
cur = ggml_mul_mat(ctx0,
model.layers[il].c_mlp_fc_w,
inpSA);
cur = ggml_add(ctx0,
ggml_repeat(ctx0, model.layers[il].c_mlp_fc_b, cur),
cur);
// GELU activation
cur = ggml_gelu(ctx0, cur);
// projection
// cur = proj_w*cur + proj_b
cur = ggml_mul_mat(ctx0,
model.layers[il].c_mlp_proj_w,
cur);
cur = ggml_add(ctx0,
ggml_repeat(ctx0, model.layers[il].c_mlp_proj_b, cur),
cur);
}
// self-attention + FF
cur = ggml_add(ctx0, cur, inpFF);
// input for next layer
inpL = ggml_add(ctx0, cur, inpL);
}
// norm
{
inpL = ggml_norm(ctx0, inpL);
// inpL = ln_f_g*inpL + ln_f_b
inpL = ggml_add(ctx0,
ggml_mul(ctx0,
ggml_repeat(ctx0, model.ln_f_g, inpL),
inpL),
ggml_repeat(ctx0, model.ln_f_b, inpL));
}
// lm_head
{
inpL = ggml_mul_mat(ctx0, model.lmh_g, inpL);
inpL = ggml_add(ctx0,
ggml_repeat(ctx0, model.lmh_b, inpL),
inpL);
}
// logits -> probs
//inpL = ggml_soft_max(ctx0, inpL);
// run the computation
ggml_build_forward_expand(&gf, inpL);
ggml_graph_compute (ctx0, &gf);
//if (n_past%100 == 0) {
// ggml_graph_print (&gf);
// ggml_graph_dump_dot(&gf, NULL, "gpt-2.dot");
//}
//embd_w.resize(n_vocab*N);
//memcpy(embd_w.data(), ggml_get_data(inpL), sizeof(float)*n_vocab*N);
// return result for just the last token
embd_w.resize(n_vocab);
memcpy(embd_w.data(), (float *) ggml_get_data(inpL) + (n_vocab*(N-1)), sizeof(float)*n_vocab);
if (mem_per_token == 0) {
mem_per_token = ggml_used_mem(ctx0)/N;
}
//printf("used_mem = %zu\n", ggml_used_mem(ctx0));
ggml_free(ctx0);
return true;
}
#define GPTJ_MAX_RNG_STATE 64*1024
size_t gptj_get_state_size(const gptj_model &model)
{
// we don't know size of rng until we actually serialize it. so reserve more than enough memory for its serialized state.
// for reference, std::mt19937(1337) serializes to 6701 bytes.
const size_t s_rng_size = sizeof(size_t);
const size_t s_rng = GPTJ_MAX_RNG_STATE;
const size_t s_kv_size = sizeof(size_t);
const size_t s_kv_ntok = sizeof(int);
const size_t s_kv = model.kv_self.buf.size;
const size_t s_total = (
+ s_rng_size
+ s_rng
+ s_kv_size
+ s_kv_ntok
+ s_kv
);
fflush(stdout);
return s_total;
}
size_t gptj_copy_state_data(const gptj_model &model, const std::mt19937 &rng, uint8_t *dest)
{
uint8_t * out = dest;
fflush(stdout);
// copy rng
{
std::stringstream rng_ss;
rng_ss << rng;
const size_t rng_size = rng_ss.str().size();
char rng_buf[GPTJ_MAX_RNG_STATE];
memset(&rng_buf[0], 0, GPTJ_MAX_RNG_STATE);
memcpy(&rng_buf[0], rng_ss.str().data(), rng_ss.str().size());
memcpy(out, &rng_size, sizeof(rng_size)); out += sizeof(rng_size);
memcpy(out, &rng_buf[0], GPTJ_MAX_RNG_STATE); out += GPTJ_MAX_RNG_STATE;
}
// copy kv cache
{
const size_t kv_size = model.kv_self.buf.size;
const int kv_ntok = model.kv_self.n;
memcpy(out, &kv_size, sizeof(kv_size)); out += sizeof(kv_size);
memcpy(out, &kv_ntok, sizeof(kv_ntok)); out += sizeof(kv_ntok);
if (kv_size) {
memcpy(out, model.kv_self.buf.addr, kv_size); out += kv_size;
}
}
const size_t written = out - dest;
assert(written == gptj_get_state_size(model));
fflush(stdout);
return written;
}
size_t gptj_set_state_data(gptj_model *model, std::mt19937 *rng, const uint8_t *src)
{
const uint8_t * in = src;
// set rng
{
size_t rng_size;
char rng_buf[GPTJ_MAX_RNG_STATE];
memcpy(&rng_size, in, sizeof(rng_size)); in += sizeof(rng_size);
memcpy(&rng_buf[0], in, GPTJ_MAX_RNG_STATE); in += GPTJ_MAX_RNG_STATE;
std::stringstream rng_ss;
rng_ss.str(std::string(&rng_buf[0], rng_size));
rng_ss >> *rng;
assert(rng_ss.fail() == false);
}
// set kv cache
{
size_t kv_size;
int kv_ntok;
memcpy(&kv_size, in, sizeof(kv_size)); in += sizeof(kv_size);
memcpy(&kv_ntok, in, sizeof(kv_ntok)); in += sizeof(kv_ntok);
if (kv_size) {
assert(model->kv_self.buf.size == kv_size);
void * k_data = model->kv_self.k->data; // remember data pointers
void * v_data = model->kv_self.v->data; // because their value is stored in buf and overwritten by memcpy
memcpy(model->kv_self.buf.addr, in, kv_size); in += kv_size;
model->kv_self.k->data = k_data; // restore correct data pointers
model->kv_self.v->data = v_data;
}
model->kv_self.n = kv_ntok;
}
const size_t nread = in - src;
assert(nread == gptj_get_state_size(*model));
fflush(stdout);
return nread;
}
struct GPTJPrivate {
const std::string modelPath;
bool modelLoaded;
gpt_vocab vocab;
gptj_model *model = nullptr;
int64_t n_threads = 0;
size_t mem_per_token = 0;
std::mt19937 rng;
};
GPTJ::GPTJ()
: d_ptr(new GPTJPrivate) {
d_ptr->model = new gptj_model;
d_ptr->modelLoaded = false;
}
bool GPTJ::loadModel(const std::string &modelPath) {
std::mt19937 rng(time(NULL));
d_ptr->rng = rng;
auto fin = std::ifstream(modelPath, std::ios::binary);
// load the model
if (!gptj_model_load(modelPath, fin, *d_ptr->model, d_ptr->vocab)) {
std::cerr << "GPT-J ERROR: failed to load model from " << modelPath;
return false;
}
d_ptr->n_threads = std::min(4, (int32_t) std::thread::hardware_concurrency());
d_ptr->modelLoaded = true;
fflush(stdout);
return true;
}
void GPTJ::setThreadCount(int32_t n_threads) {
d_ptr->n_threads = n_threads;
}
int32_t GPTJ::threadCount() const
{
return d_ptr->n_threads;
}
GPTJ::~GPTJ()
{
delete d_ptr->model;
}
bool GPTJ::isModelLoaded() const
{
return d_ptr->modelLoaded;
}
size_t GPTJ::stateSize() const
{
return gptj_get_state_size(*d_ptr->model);
}
size_t GPTJ::saveState(uint8_t *dest) const
{
return gptj_copy_state_data(*d_ptr->model, d_ptr->rng, dest);
}
size_t GPTJ::restoreState(const uint8_t *src)
{
return gptj_set_state_data(d_ptr->model, &d_ptr->rng, src);
}
void GPTJ::prompt(const std::string &prompt,
std::function<bool(int32_t)> promptCallback,
std::function<bool(int32_t, const std::string&)> responseCallback,
std::function<bool(bool)> recalculateCallback,
PromptContext &promptCtx) {
if (!isModelLoaded()) {
std::cerr << "GPT-J ERROR: prompt won't work with an unloaded model!\n";
return;
}
// tokenize the prompt
std::vector<gpt_vocab::id> embd_inp = ::gpt_tokenize(d_ptr->vocab, prompt);
// save the context size
promptCtx.n_ctx = d_ptr->model->hparams.n_ctx;
if ((int) embd_inp.size() > promptCtx.n_ctx - 4) {
responseCallback(-1, "ERROR: The prompt size exceeds the context window size and cannot be processed.");
std::cerr << "GPT-J ERROR: The prompt is" << embd_inp.size() <<
"tokens and the context window is" << promptCtx.n_ctx << "!\n";
return;
}
promptCtx.n_predict = std::min(promptCtx.n_predict, promptCtx.n_ctx - (int) embd_inp.size());
promptCtx.n_past = std::min(promptCtx.n_past, promptCtx.n_ctx);
// determine the required inference memory per token:
static bool initialized = false;
static std::vector<gpt_vocab::id> p_instruct;
static std::vector<gpt_vocab::id> r_instruct;
if (!initialized) {
gptj_eval(*d_ptr->model, d_ptr->n_threads, 0, { 0, 1, 2, 3 }, promptCtx.logits,
d_ptr->mem_per_token);
initialized = true;
}
// process the prompt in batches
size_t i = 0;
while (i < embd_inp.size()) {
size_t batch_end = std::min(i + promptCtx.n_batch, embd_inp.size());
std::vector<gpt_vocab::id> batch(embd_inp.begin() + i, embd_inp.begin() + batch_end);
// Check if the context has run out...
if (promptCtx.n_past + int32_t(batch.size()) > promptCtx.n_ctx) {
const int32_t erasePoint = promptCtx.n_ctx * promptCtx.contextErase;
// Erase the first percentage of context from the tokens...
std::cerr << "GPTJ: reached the end of the context window so resizing\n";
promptCtx.tokens.erase(promptCtx.tokens.begin(), promptCtx.tokens.begin() + erasePoint);
promptCtx.n_past = promptCtx.tokens.size();
recalculateContext(promptCtx, recalculateCallback);
assert(promptCtx.n_past + int32_t(batch.size()) <= promptCtx.n_ctx);
}
if (!evalTokens(promptCtx, batch)) {
std::cerr << "GPT-J ERROR: Failed to process prompt\n";
return;
}
size_t tokens = batch_end - i;
for (size_t t = 0; t < tokens; ++t) {
if (int32_t(promptCtx.tokens.size()) == promptCtx.n_ctx)
promptCtx.tokens.erase(promptCtx.tokens.begin());
promptCtx.tokens.push_back(batch.at(t));
if (!promptCallback(batch.at(t)))
return;
}
promptCtx.n_past += batch.size();
i = batch_end;
}
std::string cachedResponse;
std::vector<gpt_vocab::id> cachedTokens;
std::unordered_set<std::string> reversePrompts
= { "### Instruction", "### Prompt", "### Response", "### Human", "### Assistant", "### Context" };
// predict next tokens
for (int i = 0; i < promptCtx.n_predict; i++) {
// sample next token
const int n_vocab = d_ptr->model->hparams.n_vocab;
gpt_vocab::id id = 0;
{
const size_t n_prev_toks = std::min((size_t) promptCtx.repeat_last_n, promptCtx.tokens.size());
id = gpt_sample_top_k_top_p(n_vocab,
promptCtx.tokens.data() + promptCtx.tokens.size() - n_prev_toks,
n_prev_toks,
promptCtx.logits,
promptCtx.top_k, promptCtx.top_p, promptCtx.temp,
promptCtx.repeat_penalty,
d_ptr->rng);
}
// Check if the context has run out...
if (promptCtx.n_past + 1 > promptCtx.n_ctx) {
const int32_t erasePoint = promptCtx.n_ctx * promptCtx.contextErase;
// Erase the first percentage of context from the tokens...
std::cerr << "GPTJ: reached the end of the context window so resizing\n";
promptCtx.tokens.erase(promptCtx.tokens.begin(), promptCtx.tokens.begin() + erasePoint);
promptCtx.n_past = promptCtx.tokens.size();
recalculateContext(promptCtx, recalculateCallback);
assert(promptCtx.n_past + 1 <= promptCtx.n_ctx);
}
if (!evalTokens(promptCtx, { id })) {
std::cerr << "GPT-J ERROR: Failed to predict next token\n";
return;
}
promptCtx.n_past += 1;
// display text
if (id == 50256 /*end of text*/)
return;
const std::string str = d_ptr->vocab.id_to_token[id];
// Check if the provided str is part of our reverse prompts
bool foundPartialReversePrompt = false;
const std::string completed = cachedResponse + str;
if (reversePrompts.find(completed) != reversePrompts.end())
return;
// Check if it partially matches our reverse prompts and if so, cache
for (auto s : reversePrompts) {
if (s.compare(0, completed.size(), completed) == 0) {
foundPartialReversePrompt = true;
cachedResponse = completed;
break;
}
}
// Regardless the token gets added to our cache
cachedTokens.push_back(id);
// Continue if we have found a partial match
if (foundPartialReversePrompt)
continue;
// Empty the cache
for (auto t : cachedTokens) {
if (int32_t(promptCtx.tokens.size()) == promptCtx.n_ctx)
promptCtx.tokens.erase(promptCtx.tokens.begin());
promptCtx.tokens.push_back(t);
if (!responseCallback(t, d_ptr->vocab.id_to_token[t]))
return;
}
cachedTokens.clear();
}
}
bool GPTJ::evalTokens(PromptContext &ctx, const std::vector<int32_t> &tokens)
{
return gptj_eval(*d_ptr->model, d_ptr->n_threads, ctx.n_past, tokens, ctx.logits, d_ptr->mem_per_token);
}
#if defined(_WIN32)
#define DLL_EXPORT __declspec(dllexport)
#else
#define DLL_EXPORT __attribute__ ((visibility ("default")))
#endif
extern "C" {
DLL_EXPORT bool is_g4a_backend_model_implementation() {
return true;
}
DLL_EXPORT const char *get_model_type() {
return modelType_;
}
DLL_EXPORT const char *get_build_variant() {
return GGML_BUILD_VARIANT;
}
DLL_EXPORT bool magic_match(std::istream& f) {
uint32_t magic = 0;
f.read(reinterpret_cast<char*>(&magic), sizeof(magic));
return magic == 0x67676d6c;
}
DLL_EXPORT LLModel *construct() {
return new GPTJ;
}
}