Archives
/
gpt4all
mirror of https://github.com/nomic-ai/gpt4all
2
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

metal replit (#931)

Browse Source 
metal+replit

makes replit work with Metal and removes its use of `mem_per_token`
in favor of fixed size scratch buffers (closer to llama.cpp)
pull/977/head
Aaron Miller 1 year ago committed by GitHub
parent a9b33c3d10
commit f71d8efc71
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
3 changed files with 102 additions and 32 deletions
Show Stats Download Patch File Download Diff File

8
gpt4all-backend/CMakeLists.txt
Unescape Escape View File

@ -97,6 +97,10 @@ foreach(BUILD_VARIANT IN LISTS BUILD_VARIANTS)
LLAMA_VERSIONS=>=3 LLAMA_DATE=999999)
prepare_target(llamamodel-mainline llama-mainline)
add_library(replit-mainline-${BUILD_VARIANT} SHARED
replit.cpp utils.h utils.cpp llmodel_shared.cpp)
prepare_target(replit-mainline llama-mainline)
if (NOT LLAMA_METAL)
add_library(llamamodel-230519-${BUILD_VARIANT} SHARED
llamamodel.cpp llmodel_shared.cpp)
@ -116,10 +120,6 @@ foreach(BUILD_VARIANT IN LISTS BUILD_VARIANTS)
add_library(mpt-${BUILD_VARIANT} SHARED
mpt.cpp utils.h utils.cpp llmodel_shared.cpp)
prepare_target(mpt ggml-230511)
add_library(replit-${BUILD_VARIANT} SHARED
replit.cpp utils.h utils.cpp llmodel_shared.cpp)
prepare_target(replit ggml-230511)
endif()
endforeach()

2
gpt4all-backend/llama.cpp-mainline

@ -1 +1 @@
Subproject commit 74a6d922f12ccfe16b0c265f43be8978c6f25e98
Subproject commit 4458a8eaf443e7fa0e764682d22213fa4fef90c3

124
gpt4all-backend/replit.cpp
Unescape Escape View File

@ -32,6 +32,9 @@
#include <vector>
#include <regex>
#include <ggml.h>
#ifdef GGML_USE_METAL
#include <ggml-metal.h>
#endif
/**
IMPORTANT: This model backend and convert script were developed for the original Huggingface
@ -226,6 +229,15 @@ struct replit_model {
struct replit_kv_cache kv_self;
struct ggml_context * ctx;
void * eval_buf;
size_t eval_buf_size;
void * scr0_buf;
size_t scr0_buf_size;
void * scr1_buf;
size_t scr1_buf_size;
#ifdef GGML_USE_METAL
struct ggml_metal_context * ctx_metal;
#endif
std::map<std::string, struct ggml_tensor *> tensors;
};
@ -304,7 +316,6 @@ bool replit_model_load(const std::string & fname, std::istream &fin, replit_mode
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",
@ -496,6 +507,32 @@ bool replit_model_load(const std::string & fname, std::istream &fin, replit_mode
printf("%s: model size = %8.2f MB / num tensors = %d\n", __func__, total_size / 1024.0 / 1024.0, n_tensors);
}
model.eval_buf_size = 256u * 1024 * 1024;
model.eval_buf = malloc(model.eval_buf_size);
model.scr0_buf_size = 256u * 1024 * 1024;
model.scr0_buf = malloc(model.scr0_buf_size);
model.scr1_buf_size = 256u * 1024 * 1024;
model.scr1_buf = malloc(model.scr1_buf_size);
#ifdef GGML_USE_METAL
model.ctx_metal = ggml_metal_init();
void* data_ptr = ggml_get_mem_buffer(model.ctx);
size_t data_size = ggml_get_mem_size(model.ctx);
#define GGML_CHECK_BUF(result) if (!(result)) { \
std::cerr << __func__ << ": failed to add buffer" << std::endl; \
ggml_free(model.ctx); \
return false; \
}
GGML_CHECK_BUF(ggml_metal_add_buffer(model.ctx_metal, "data", data_ptr, data_size));
GGML_CHECK_BUF(ggml_metal_add_buffer(model.ctx_metal, "kv", ggml_get_mem_buffer(model.kv_self.ctx),
ggml_get_mem_size(model.kv_self.ctx)));
GGML_CHECK_BUF(ggml_metal_add_buffer(model.ctx_metal, "eval", model.eval_buf, model.eval_buf_size));
GGML_CHECK_BUF(ggml_metal_add_buffer(model.ctx_metal, "scr0", model.scr0_buf, model.scr0_buf_size));
GGML_CHECK_BUF(ggml_metal_add_buffer(model.ctx_metal, "scr1", model.scr1_buf, model.scr1_buf_size));
#endif
return true;
}
@ -533,30 +570,12 @@ bool replit_eval(const replit_model & model, const int n_threads, const int n_pa
const int n_head = hparams.n_head;
const int n_vocab = hparams.n_vocab;
static size_t buf_size = 256u * 1024 * 1024;
static void * buf = malloc(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);
// reallocate
buf_size = buf_size_new;
buf = realloc(buf, buf_size);
if (buf == nullptr) {
fprintf(stderr, "%s: failed to allocate %zu bytes\n", __func__, buf_size);
return false;
}
}
struct ggml_init_params params = {
.mem_size = buf_size,
.mem_buffer = buf,
struct ggml_init_params eval_ctx_params = {
.mem_size = model.eval_buf_size,
.mem_buffer = model.eval_buf,
.no_alloc = false,
};
struct ggml_context * ctx0 = ggml_init(params);
struct ggml_context * ctx0 = ggml_init(eval_ctx_params);
struct ggml_cgraph gf = {.n_threads = n_threads};
struct ggml_tensor * embd = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
@ -565,7 +584,7 @@ bool replit_eval(const replit_model & model, const int n_threads, const int n_pa
struct ggml_tensor * inpL = ggml_get_rows(ctx0, model.wte_weight, embd);
for (int il = 0; il < n_layer; ++il) {
ggml_set_scratch(ctx0, {0, model.scr0_buf_size, model.scr0_buf, });
struct ggml_tensor * cur;
// a = self.ln_1(x)
@ -624,7 +643,7 @@ bool replit_eval(const replit_model & model, const int n_threads, const int n_pa
ggml_scale(ctx0, KQ, ggml_new_f32(ctx0, 1.0f / sqrt(float(n_embd) / n_head)));
// Alibi
struct ggml_tensor * KQ_scaled_alibi = ggml_alibi(ctx0, ggml_cont(ctx0, KQ_scaled), n_past, n_head);
struct ggml_tensor * KQ_scaled_alibi = ggml_alibi(ctx0, KQ_scaled, n_past, n_head, 8.0f);
// KQ_masked = mask_past(KQ_scaled)
struct ggml_tensor * KQ_masked = ggml_diag_mask_inf(ctx0, KQ_scaled_alibi, n_past);
@ -656,6 +675,7 @@ bool replit_eval(const replit_model & model, const int n_threads, const int n_pa
// projection
{ cur = ggml_mul_mat(ctx0, model.layers[il].c_attn_out_proj_weight, cur); }
}
ggml_set_scratch(ctx0, {0, model.scr1_buf_size, model.scr1_buf, });
inpL = ggml_add(ctx0, inpL, cur);
@ -682,7 +702,7 @@ bool replit_eval(const replit_model & model, const int n_threads, const int n_pa
// x = x + n
inpL = ggml_add(ctx0, inpL, cur);
}
ggml_set_scratch(ctx0, {0, model.scr0_buf_size, model.scr0_buf, });
// norm
{
inpL = ggml_norm(ctx0, inpL);
@ -690,6 +710,7 @@ bool replit_eval(const replit_model & model, const int n_threads, const int n_pa
inpL = ggml_mul(ctx0, ggml_repeat(ctx0, model.ln_f_weight, inpL), inpL);
}
ggml_set_scratch(ctx0, {0, 0, nullptr, });
// output embedding weight tied to input embedding
inpL = ggml_mul_mat(ctx0, model.wte_weight, inpL);
@ -698,7 +719,22 @@ bool replit_eval(const replit_model & model, const int n_threads, const int n_pa
// run the computation
ggml_build_forward_expand(&gf, inpL);
#ifdef GGML_USE_METAL
if (N == 1) {
// llama.cpp doesn't use metal for batch/prompt processing presently
// pending changes to the metal matmul kernel - only use it for generation (N=1)
ggml_metal_graph_compute(model.ctx_metal, &gf);
ggml_metal_get_tensor(model.ctx_metal, inpL);
} else {
// We need to sync the GPU KV cache with the CPU KV cache
ggml_metal_get_tensor(model.ctx_metal, model.kv_self.k);
ggml_metal_get_tensor(model.ctx_metal, model.kv_self.v);
ggml_graph_compute(ctx0, &gf);
}
#else
ggml_graph_compute(ctx0, &gf);
#endif
// std::cout << "Qcur" << std::endl;
// print_tensor(Qcur);
@ -882,6 +918,19 @@ int32_t Replit::threadCount() const
Replit::~Replit()
{
if(d_ptr->model->ctx) {
ggml_free(d_ptr->model->ctx);
d_ptr->model->ctx = nullptr;
}
if(d_ptr->model->eval_buf) {
free(d_ptr->model->eval_buf);
}
if(d_ptr->model->scr0_buf) {
free(d_ptr->model->scr0_buf);
}
if(d_ptr->model->scr1_buf) {
free(d_ptr->model->scr1_buf);
}
delete d_ptr->model;
}
@ -965,7 +1014,28 @@ DLL_EXPORT const char *get_build_variant() {
DLL_EXPORT bool magic_match(std::istream& f) {
uint32_t magic = 0;
f.read(reinterpret_cast<char*>(&magic), sizeof(magic));
return magic == 0x7265706c;
if (magic != 0x7265706c) return false;
#ifdef GGML_USE_METAL
off_t offset = sizeof(uint32_t) * 5; // n_vocab, n_ctx, n_embd, n_head, n_layer
f.seekg(offset, std::ios_base::cur);
uint32_t ftype;
f.read(reinterpret_cast<char*>(&ftype), sizeof(ftype)); // ftype
const int32_t qntvr = ftype / GGML_QNT_VERSION_FACTOR;
ftype %= GGML_QNT_VERSION_FACTOR;
switch (ftype) {
case 1: return true; // GGML_TYPE_F16
case 2: // GGML_TYPE_Q4_0
if (qntvr != GGML_QNT_VERSION)
{
std::cerr << "replit: not using metal (unsupported qnt ver)" << std::endl;
return false;
}
return true;
default: return false;
}
#else
return true;
#endif
}
DLL_EXPORT LLModel *construct() {

Write Preview
Loading…
Cancel
Save