ExternalVendorCode/whisper.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/whisper.cpp.git synced 2024-12-18 20:27:53 +00:00
Code Issues Actions 4 Packages Projects Releases Wiki Activity

ggml : sync (ggml-alloc, GPU, eps, etc.) (#1220)

Browse Source 
* ggml : sync (ggml-alloc, GPU, eps, etc.)

* ggml : fix build

* wasm : fix build
This commit is contained in:
Georgi Gerganov 2023-09-05 13:54:40 +03:00 committed by GitHub
parent 6780c98e19
commit 59a3d0cb57
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
17 changed files with 11254 additions and 4095 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
bindings/javascript/libwhisper.worker.js
View File

@ -1 +1 @@
"use strict";var Module={};var ENVIRONMENT_IS_NODE=typeof process=="object"&&typeof process.versions=="object"&&typeof process.versions.node=="string";if(ENVIRONMENT_IS_NODE){var nodeWorkerThreads=require("worker_threads");var parentPort=nodeWorkerThreads.parentPort;parentPort.on("message",data=>onmessage({data:data}));var fs=require("fs");Object.assign(global,{self:global,require:require,Module:Module,location:{href:__filename},Worker:nodeWorkerThreads.Worker,importScripts:function(f){(0,eval)(fs.readFileSync(f,"utf8")+"//# sourceURL="+f)},postMessage:function(msg){parentPort.postMessage(msg)},performance:global.performance||{now:function(){return Date.now()}}})}var initializedJS=false;var pendingNotifiedProxyingQueues=[];function threadPrintErr(){var text=Array.prototype.slice.call(arguments).join(" ");if(ENVIRONMENT_IS_NODE){fs.writeSync(2,text+"\n");return}console.error(text)}function threadAlert(){var text=Array.prototype.slice.call(arguments).join(" ");postMessage({cmd:"alert",text:text,threadId:Module["_pthread_self"]()})}var err=threadPrintErr;self.alert=threadAlert;Module["instantiateWasm"]=(info,receiveInstance)=>{var instance=new WebAssembly.Instance(Module["wasmModule"],info);receiveInstance(instance);Module["wasmModule"]=null;return instance.exports};self.onunhandledrejection=e=>{throw e.reason??e};self.onmessage=e=>{try{if(e.data.cmd==="load"){Module["wasmModule"]=e.data.wasmModule;for(const handler of e.data.handlers){Module[handler]=function(){postMessage({cmd:"callHandler",handler:handler,args:[...arguments]})}}Module["wasmMemory"]=e.data.wasmMemory;Module["buffer"]=Module["wasmMemory"].buffer;Module["ENVIRONMENT_IS_PTHREAD"]=true;if(typeof e.data.urlOrBlob=="string"){importScripts(e.data.urlOrBlob)}else{var objectUrl=URL.createObjectURL(e.data.urlOrBlob);importScripts(objectUrl);URL.revokeObjectURL(objectUrl)}whisper_factory(Module).then(function(instance){Module=instance})}else if(e.data.cmd==="run"){Module["__performance_now_clock_drift"]=performance.now()-e.data.time;Module["__emscripten_thread_init"](e.data.pthread_ptr,0,0,1);Module["establishStackSpace"]();Module["PThread"].receiveObjectTransfer(e.data);Module["PThread"].threadInitTLS();if(!initializedJS){Module["__embind_initialize_bindings"]();pendingNotifiedProxyingQueues.forEach(queue=>{Module["executeNotifiedProxyingQueue"](queue)});pendingNotifiedProxyingQueues=[];initializedJS=true}try{Module["invokeEntryPoint"](e.data.start_routine,e.data.arg)}catch(ex){if(ex!="unwind"){if(ex instanceof Module["ExitStatus"]){if(Module["keepRuntimeAlive"]()){}else{Module["__emscripten_thread_exit"](ex.status)}}else{throw ex}}}}else if(e.data.cmd==="cancel"){if(Module["_pthread_self"]()){Module["__emscripten_thread_exit"](-1)}}else if(e.data.target==="setimmediate"){}else if(e.data.cmd==="processProxyingQueue"){if(initializedJS){Module["executeNotifiedProxyingQueue"](e.data.queue)}else{pendingNotifiedProxyingQueues.push(e.data.queue)}}else if(e.data.cmd){err("worker.js received unknown command "+e.data.cmd);err(e.data)}}catch(ex){if(Module["__emscripten_thread_crashed"]){Module["__emscripten_thread_crashed"]()}throw ex}};
"use strict";var Module={};var ENVIRONMENT_IS_NODE=typeof process=="object"&&typeof process.versions=="object"&&typeof process.versions.node=="string";if(ENVIRONMENT_IS_NODE){var nodeWorkerThreads=require("worker_threads");var parentPort=nodeWorkerThreads.parentPort;parentPort.on("message",data=>onmessage({data:data}));var fs=require("fs");Object.assign(global,{self:global,require:require,Module:Module,location:{href:__filename},Worker:nodeWorkerThreads.Worker,importScripts:f=>(0,eval)(fs.readFileSync(f,"utf8")+"//# sourceURL="+f),postMessage:msg=>parentPort.postMessage(msg),performance:global.performance||{now:Date.now}})}var initializedJS=false;function threadPrintErr(){var text=Array.prototype.slice.call(arguments).join(" ");if(ENVIRONMENT_IS_NODE){fs.writeSync(2,text+"\n");return}console.error(text)}function threadAlert(){var text=Array.prototype.slice.call(arguments).join(" ");postMessage({cmd:"alert",text:text,threadId:Module["_pthread_self"]()})}var err=threadPrintErr;self.alert=threadAlert;Module["instantiateWasm"]=(info,receiveInstance)=>{var module=Module["wasmModule"];Module["wasmModule"]=null;var instance=new WebAssembly.Instance(module,info);return receiveInstance(instance)};self.onunhandledrejection=e=>{throw e.reason||e};function handleMessage(e){try{if(e.data.cmd==="load"){let messageQueue=[];self.onmessage=e=>messageQueue.push(e);self.startWorker=instance=>{Module=instance;postMessage({"cmd":"loaded"});for(let msg of messageQueue){handleMessage(msg)}self.onmessage=handleMessage};Module["wasmModule"]=e.data.wasmModule;for(const handler of e.data.handlers){Module[handler]=(...args)=>{postMessage({cmd:"callHandler",handler:handler,args:args})}}Module["wasmMemory"]=e.data.wasmMemory;Module["buffer"]=Module["wasmMemory"].buffer;Module["ENVIRONMENT_IS_PTHREAD"]=true;if(typeof e.data.urlOrBlob=="string"){importScripts(e.data.urlOrBlob)}else{var objectUrl=URL.createObjectURL(e.data.urlOrBlob);importScripts(objectUrl);URL.revokeObjectURL(objectUrl)}whisper_factory(Module)}else if(e.data.cmd==="run"){Module["__emscripten_thread_init"](e.data.pthread_ptr,0,0,1);Module["__emscripten_thread_mailbox_await"](e.data.pthread_ptr);Module["establishStackSpace"]();Module["PThread"].receiveObjectTransfer(e.data);Module["PThread"].threadInitTLS();if(!initializedJS){Module["__embind_initialize_bindings"]();initializedJS=true}try{Module["invokeEntryPoint"](e.data.start_routine,e.data.arg)}catch(ex){if(ex!="unwind"){throw ex}}}else if(e.data.cmd==="cancel"){if(Module["_pthread_self"]()){Module["__emscripten_thread_exit"](-1)}}else if(e.data.target==="setimmediate"){}else if(e.data.cmd==="checkMailbox"){if(initializedJS){Module["checkMailbox"]()}}else if(e.data.cmd){err(`worker.js received unknown command ${e.data.cmd}`);err(e.data)}}catch(ex){if(Module["__emscripten_thread_crashed"]){Module["__emscripten_thread_crashed"]()}throw ex}}self.onmessage=handleMessage;

12
bindings/javascript/whisper.js
View File

File diff suppressed because one or more lines are too long

94
examples/common.cpp
View File

@ -1,3 +1,5 @@
#define _USE_MATH_DEFINES // for M_PI
#include "common.h"
// third-party utilities
@ -13,53 +15,59 @@
#include <codecvt>
#include <sstream>
#ifndef M_PI
#define M_PI 3.14159265358979323846
#endif
#if defined(_MSC_VER)
#pragma warning(disable: 4244 4267) // possible loss of data
#endif
// Function to check if the next argument exists
std::string get_next_arg(int& i, int argc, char** argv, const std::string& flag, gpt_params& params) {
if (i + 1 < argc && argv[i + 1][0] != '-') {
return argv[++i];
} else {
fprintf(stderr, "error: %s requires one argument.\n", flag.c_str());
gpt_print_usage(argc, argv, params);
exit(0);
}
}
bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
for (int i = 1; i < argc; i++) {
std::string arg = argv[i];
if (arg == "-s" || arg == "--seed") {
params.seed = std::stoi(argv[++i]);
params.seed = std::stoi(get_next_arg(i, argc, argv, arg, params));
} else if (arg == "-t" || arg == "--threads") {
params.n_threads = std::stoi(argv[++i]);
params.n_threads = std::stoi(get_next_arg(i, argc, argv, arg, params));
} else if (arg == "-ngl" || arg == "--gpu-layers" || arg == "--n-gpu-layers") {
params.n_gpu_layers = std::stoi(get_next_arg(i, argc, argv, arg, params));
} else if (arg == "-p" || arg == "--prompt") {
params.prompt = argv[++i];
params.prompt = get_next_arg(i, argc, argv, arg, params);
} else if (arg == "-n" || arg == "--n_predict") {
params.n_predict = std::stoi(argv[++i]);
params.n_predict = std::stoi(get_next_arg(i, argc, argv, arg, params));
} else if (arg == "--top_k") {
params.top_k = std::max(1, std::stoi(argv[++i]));
params.top_k = std::stoi(get_next_arg(i, argc, argv, arg, params));
} else if (arg == "--top_p") {
params.top_p = std::stof(argv[++i]);
params.top_p = std::stof(get_next_arg(i, argc, argv, arg, params));
} else if (arg == "--temp") {
params.temp = std::stof(argv[++i]);
params.temp = std::stof(get_next_arg(i, argc, argv, arg, params));
} else if (arg == "--repeat-last-n") {
params.repeat_last_n = std::stof(argv[++i]);
params.repeat_last_n = std::stoi(get_next_arg(i, argc, argv, arg, params));
} else if (arg == "--repeat-penalty") {
params.repeat_penalty = std::stof(argv[++i]);
params.repeat_penalty = std::stof(get_next_arg(i, argc, argv, arg, params));
} else if (arg == "-b" || arg == "--batch_size") {
params.n_batch = std::stoi(argv[++i]);
params.n_batch= std::stoi(get_next_arg(i, argc, argv, arg, params));
} else if (arg == "-m" || arg == "--model") {
params.model = argv[++i];
params.model = get_next_arg(i, argc, argv, arg, params);
} else if (arg == "-i" || arg == "--interactive") {
params.interactive = true;
} else if (arg == "-ip" || arg == "--interactive-port") {
params.interactive = true;
params.interactive_port = std::stoi(argv[++i]);
params.interactive_port = std::stoi(get_next_arg(i, argc, argv, arg, params));
} else if (arg == "-h" || arg == "--help") {
gpt_print_usage(argc, argv, params);
exit(0);
} else if (arg == "-f" || arg == "--file") {
if (++i > argc) {
fprintf(stderr, "Invalid file param");
break;
}
get_next_arg(i, argc, argv, arg, params);
std::ifstream file(argv[i]);
if (!file) {
fprintf(stderr, "error: failed to open file '%s'\n", argv[i]);
@ -70,7 +78,7 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
params.prompt.pop_back();
}
} else if (arg == "-tt" || arg == "--token_test") {
params.token_test = argv[++i];
params.token_test = get_next_arg(i, argc, argv, arg, params);
}
else {
fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
@ -89,6 +97,7 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
fprintf(stderr, " -h, --help show this help message and exit\n");
fprintf(stderr, " -s SEED, --seed SEED RNG seed (default: -1)\n");
fprintf(stderr, " -t N, --threads N number of threads to use during computation (default: %d)\n", params.n_threads);
fprintf(stderr, " -ngl N, --gpu-layers N number of layers to offload to GPU on supported models (default: %d)\n", params.n_gpu_layers);
fprintf(stderr, " -p PROMPT, --prompt PROMPT\n");
fprintf(stderr, " prompt to start generation with (default: random)\n");
fprintf(stderr, " -f FNAME, --file FNAME\n");
@ -755,3 +764,46 @@ float similarity(const std::string & s0, const std::string & s1) {
return 1.0f - (dist / std::max(s0.size(), s1.size()));
}
bool sam_params_parse(int argc, char ** argv, sam_params & params) {
for (int i = 1; i < argc; i++) {
std::string arg = argv[i];
if (arg == "-s" || arg == "--seed") {
params.seed = std::stoi(argv[++i]);
} else if (arg == "-t" || arg == "--threads") {
params.n_threads = std::stoi(argv[++i]);
} else if (arg == "-m" || arg == "--model") {
params.model = argv[++i];
} else if (arg == "-i" || arg == "--inp") {
params.fname_inp = argv[++i];
} else if (arg == "-o" || arg == "--out") {
params.fname_out = argv[++i];
} else if (arg == "-h" || arg == "--help") {
sam_print_usage(argc, argv, params);
exit(0);
} else {
fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
sam_print_usage(argc, argv, params);
exit(0);
}
}
return true;
}
void sam_print_usage(int argc, char ** argv, const sam_params & params) {
fprintf(stderr, "usage: %s [options]\n", argv[0]);
fprintf(stderr, "\n");
fprintf(stderr, "options:\n");
fprintf(stderr, " -h, --help show this help message and exit\n");
fprintf(stderr, " -s SEED, --seed SEED RNG seed (default: -1)\n");
fprintf(stderr, " -t N, --threads N number of threads to use during computation (default: %d)\n", params.n_threads);
fprintf(stderr, " -m FNAME, --model FNAME\n");
fprintf(stderr, " model path (default: %s)\n", params.model.c_str());
fprintf(stderr, " -i FNAME, --inp FNAME\n");
fprintf(stderr, " input file (default: %s)\n", params.fname_inp.c_str());
fprintf(stderr, " -o FNAME, --out FNAME\n");
fprintf(stderr, " output file (default: %s)\n", params.fname_out.c_str());
fprintf(stderr, "\n");
}

21
examples/common.h
View File

@ -11,7 +11,7 @@
#define COMMON_SAMPLE_RATE 16000
//
// CLI argument parsing
// GPT CLI argument parsing
//
struct gpt_params {
@ -33,6 +33,8 @@ struct gpt_params {
bool interactive = false;
int32_t interactive_port = -1;
int32_t n_gpu_layers = 0;
};
bool gpt_params_parse(int argc, char ** argv, gpt_params & params);
@ -155,3 +157,20 @@ bool vad_simple(
// compute similarity between two strings using Levenshtein distance
float similarity(const std::string & s0, const std::string & s1);
//
// SAM argument parsing
//
struct sam_params {
int32_t seed = -1; // RNG seed
int32_t n_threads = std::min(4, (int32_t) std::thread::hardware_concurrency());
std::string model = "models/sam-vit-b/ggml-model-f16.bin"; // model path
std::string fname_inp = "img.jpg";
std::string fname_out = "img.out";
};
bool sam_params_parse(int argc, char ** argv, sam_params & params);
void sam_print_usage(int argc, char ** argv, const sam_params & params);

15
examples/talk.wasm/gpt-2.cpp
View File

@ -191,9 +191,9 @@ bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab &
// create the ggml context
{
struct ggml_init_params params = {
.mem_size = ctx_size,
.mem_buffer = NULL,
.no_alloc = false,
/*.mem_size =*/ ctx_size,
/*.mem_buffer =*/ NULL,
/*.no_alloc =*/ false,
};
model.ctx = ggml_init(params);
@ -420,7 +420,6 @@ bool gpt2_eval(
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));
@ -442,7 +441,7 @@ bool gpt2_eval(
// norm
{
// [ 768, N]
cur = ggml_norm(ctx0, inpL);
cur = ggml_norm(ctx0, inpL, 1e-5f);
// cur = ln_1_g*cur + ln_1_b
// [ 768, N]
@ -589,7 +588,7 @@ bool gpt2_eval(
{
// norm
{
cur = ggml_norm(ctx0, inpFF);
cur = ggml_norm(ctx0, inpFF, 1e-5f);
// cur = ln_2_g*cur + ln_2_b
// [ 768, N]
@ -644,7 +643,7 @@ bool gpt2_eval(
// norm
{
// [ 768, N]
inpL = ggml_norm(ctx0, inpL);
inpL = ggml_norm(ctx0, inpL, 1e-5f);
// inpL = ln_f_g*inpL + ln_f_b
// [ 768, N]
@ -665,7 +664,7 @@ bool gpt2_eval(
// run the computation
ggml_build_forward_expand (&gf, inpL);
ggml_graph_compute (ctx0, &gf);
ggml_graph_compute_with_ctx(ctx0, &gf, n_threads);
//if (n_past%100 == 0) {
// ggml_graph_print (&gf);

10
examples/talk/gpt-2.cpp
View File

@ -379,6 +379,7 @@ bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab &
// - embd_inp: the embeddings of the tokens in the context
// - embd_w: the predicted logits for the next token
//
// TODO: sync latest version from ggml repo
bool gpt2_eval(
const gpt2_model & model,
const int n_threads,
@ -420,7 +421,6 @@ bool gpt2_eval(
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));
@ -442,7 +442,7 @@ bool gpt2_eval(
// norm
{
// [ 768, N]
cur = ggml_norm(ctx0, inpL);
cur = ggml_norm(ctx0, inpL, 1e-5f);
// cur = ln_1_g*cur + ln_1_b
// [ 768, N]
@ -589,7 +589,7 @@ bool gpt2_eval(
{
// norm
{
cur = ggml_norm(ctx0, inpFF);
cur = ggml_norm(ctx0, inpFF, 1e-5f);
// cur = ln_2_g*cur + ln_2_b
// [ 768, N]
@ -644,7 +644,7 @@ bool gpt2_eval(
// norm
{
// [ 768, N]
inpL = ggml_norm(ctx0, inpL);
inpL = ggml_norm(ctx0, inpL, 1e-5f);
// inpL = ln_f_g*inpL + ln_f_b
// [ 768, N]
@ -665,7 +665,7 @@ bool gpt2_eval(
// run the computation
ggml_build_forward_expand (&gf, inpL);
ggml_graph_compute (ctx0, &gf);
ggml_graph_compute_with_ctx(ctx0, &gf, n_threads);
//if (n_past%100 == 0) {
// ggml_graph_print (&gf);

594
ggml-alloc.c Normal file
View File

@ -0,0 +1,594 @@
#include "ggml-alloc.h"
#include "ggml.h"
#include <assert.h>
#include <stdarg.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#define UNUSED(x) (void)(x)
#define MAX(a, b) ((a) > (b) ? (a) : (b))
#define GGML_MAX_CONCUR (2*GGML_MAX_NODES)
//#define GGML_ALLOCATOR_DEBUG
//#define AT_PRINTF printf
#define AT_PRINTF(...) ((void)0)
struct hash_node {
struct ggml_tensor * t;
int n_children;
int n_views;
};
static size_t hash(void * p) {
return (size_t)p % GGML_GRAPH_HASHTABLE_SIZE;
}
static struct hash_node * hash_get(struct hash_node hash_table[], struct ggml_tensor * t) {
size_t h = hash(t);
// linear probing
size_t i = h;
while (hash_table[i].t != NULL) {
if (hash_table[i].t == t) {
return &hash_table[i];
}
i = (i + 1) % GGML_GRAPH_HASHTABLE_SIZE;
if (i == h) {
// hash table is full
GGML_ASSERT(false);
}
}
hash_table[i].t = t;
return &hash_table[i];
}
// TODO: GGML_PAD ?
static size_t aligned_offset(const void * buffer, size_t offset, size_t alignment) {
assert(alignment && !(alignment & (alignment - 1))); // power of 2
size_t align = (alignment - (((uintptr_t)buffer + offset) % alignment)) % alignment;
return offset + align;
}
struct free_block {
void * addr;
size_t size;
};
#define MAX_FREE_BLOCKS 128
struct ggml_allocr {
void * data;
size_t size;
size_t alignment;
int n_free_blocks;
struct free_block free_blocks[MAX_FREE_BLOCKS];
struct hash_node hash_table[GGML_GRAPH_HASHTABLE_SIZE];
size_t max_size;
bool measure;
int parse_seq[GGML_MAX_CONCUR];
int parse_seq_len;
#ifdef GGML_ALLOCATOR_DEBUG
struct ggml_tensor * allocated_tensors[1024];
#endif
};
#ifdef GGML_ALLOCATOR_DEBUG
static void add_allocated_tensor(struct ggml_allocr * alloc, struct ggml_tensor * tensor) {
for (int i = 0; i < 1024; i++) {
if (alloc->allocated_tensors[i] == NULL) {
alloc->allocated_tensors[i] = tensor;
return;
}
}
GGML_ASSERT(!"out of allocated_tensors");
}
static void remove_allocated_tensor(struct ggml_allocr * alloc, struct ggml_tensor * tensor) {
for (int i = 0; i < 1024; i++) {
if (alloc->allocated_tensors[i] == tensor ||
(alloc->allocated_tensors[i] != NULL && alloc->allocated_tensors[i]->data == tensor->data)) {
alloc->allocated_tensors[i] = NULL;
return;
}
}
printf("tried to free tensor %s not found\n", tensor->name);
GGML_ASSERT(!"tensor not found");
}
#endif
static size_t ggml_allocator_get_alloc_size(struct ggml_allocr * alloc, struct ggml_tensor * tensor) {
return ggml_nbytes(tensor);
UNUSED(alloc);
}
void ggml_allocr_alloc(struct ggml_allocr * alloc, struct ggml_tensor * tensor) {
size_t size = ggml_allocator_get_alloc_size(alloc, tensor);
size = aligned_offset(NULL, size, alloc->alignment);
AT_PRINTF("%s: allocating %s (%zu bytes) - ", __func__, tensor->name, size);
size_t max_avail = 0;
// find the best fitting free block besides the last block
int best_fit_block = -1;
size_t best_fit_size = SIZE_MAX;
for (int i = 0; i < alloc->n_free_blocks - 1; i++) {
struct free_block * block = &alloc->free_blocks[i];
max_avail = MAX(max_avail, block->size);
if (block->size >= size && block->size <= best_fit_size) {
best_fit_block = i;
best_fit_size = block->size;
}
}
AT_PRINTF("block %d\n", best_fit_block);
if (best_fit_block == -1) {
// the last block is our last resort
struct free_block * block = &alloc->free_blocks[alloc->n_free_blocks - 1];
if (block->size >= size) {
best_fit_block = alloc->n_free_blocks - 1;
max_avail = MAX(max_avail, block->size);
} else {
fprintf(stderr, "%s: not enough space in the buffer (needed %zu, largest block available %zu)\n",
__func__, size, max_avail);
GGML_ASSERT(!"not enough space in the buffer");
return;
}
}
struct free_block * block = &alloc->free_blocks[best_fit_block];
void * addr = block->addr;
block->addr = (char*)block->addr + size;
block->size -= size;
if (block->size == 0) {
// remove block if empty
alloc->n_free_blocks--;
for (int j = best_fit_block; j < alloc->n_free_blocks; j++) {
alloc->free_blocks[j] = alloc->free_blocks[j+1];
}
}
tensor->data = addr;
#ifdef GGML_ALLOCATOR_DEBUG
add_allocated_tensor(alloc, tensor);
size_t cur_max = (char*)addr - (char*)alloc->data + size;
if (cur_max > alloc->max_size) {
printf("max_size = %.2f MB: tensors: ", cur_max / 1024.0 / 1024.0);
for (int i = 0; i < 1024; i++) {
if (alloc->allocated_tensors[i]) {
printf("%s (%.2f MB) ", alloc->allocated_tensors[i]->name, ggml_nbytes(alloc->allocated_tensors[i]) / 1024.0 / 1024.0);
}
}
printf("\n");
}
#endif
alloc->max_size = MAX(alloc->max_size, (char*)addr - (char*)alloc->data + size);
}
// this is a very naive implementation, but for our case the number of free blocks should be very small
static void ggml_allocator_free_tensor(struct ggml_allocr * alloc, struct ggml_tensor * tensor) {
void * ptr = tensor->data;
if (ptr < alloc->data || (char*)ptr >= (char*)alloc->data + alloc->max_size) {
// the tensor was not allocated in this buffer
// this can happen because the graph allocator will try to free weights and other tensors from different buffers
// the easiest way to deal with this is just to ignore it
return;
}
size_t size = ggml_allocator_get_alloc_size(alloc, tensor);
size = aligned_offset(NULL, size, alloc->alignment);
AT_PRINTF("%s: freeing %s (%zu bytes) - n_free_blocks = %d\n", __func__, tensor->name, size, alloc->n_free_blocks);
#ifdef GGML_ALLOCATOR_DEBUG
remove_allocated_tensor(alloc, tensor);
#endif
// see if we can merge with an existing block
for (int i = 0; i < alloc->n_free_blocks; i++) {
struct free_block * block = &alloc->free_blocks[i];
// check if ptr is at the end of the block
if ((char*)block->addr + block->size == ptr) {
block->size += size;
// check if we can merge with the next block
if (i < alloc->n_free_blocks - 1 && (char*)block->addr + block->size == alloc->free_blocks[i+1].addr) {
block->size += alloc->free_blocks[i+1].size;
alloc->n_free_blocks--;
for (int j = i+1; j < alloc->n_free_blocks; j++) {
alloc->free_blocks[j] = alloc->free_blocks[j+1];
}
}
return;
}
// check if ptr is at the beginning of the block
if ((char*)ptr + size == block->addr) {
block->addr = ptr;
block->size += size;
// check if we can merge with the previous block
if (i > 0 && (char*)alloc->free_blocks[i-1].addr + alloc->free_blocks[i-1].size == block->addr) {
alloc->free_blocks[i-1].size += block->size;
alloc->n_free_blocks--;
for (int j = i; j < alloc->n_free_blocks; j++) {
alloc->free_blocks[j] = alloc->free_blocks[j+1];
}
}
return;
}
}
// otherwise, add a new block
GGML_ASSERT(alloc->n_free_blocks < MAX_FREE_BLOCKS && "out of free blocks");
// insert the new block in the correct position to keep the array sorted by address (to make merging blocks faster)
int insert_pos = 0;
while (insert_pos < alloc->n_free_blocks && alloc->free_blocks[insert_pos].addr < ptr) {
insert_pos++;
}
// shift all blocks from insert_pos onward to make room for the new block
for (int i = alloc->n_free_blocks; i > insert_pos; i--) {
alloc->free_blocks[i] = alloc->free_blocks[i-1];
}
// insert the new block
alloc->free_blocks[insert_pos].addr = ptr;
alloc->free_blocks[insert_pos].size = size;
alloc->n_free_blocks++;
}
void ggml_allocr_set_parse_seq(struct ggml_allocr * alloc, const int * list, int n) {
for (int i = 0; i < n; i++) {
alloc->parse_seq[i] = list[i];
}
alloc->parse_seq_len = n;
}
void ggml_allocr_reset(struct ggml_allocr * alloc) {
alloc->n_free_blocks = 1;
size_t align_offset = aligned_offset(alloc->data, 0, alloc->alignment);
alloc->free_blocks[0].addr = (char *)alloc->data + align_offset;
alloc->free_blocks[0].size = alloc->size - align_offset;
}
struct ggml_allocr * ggml_allocr_new(void * data, size_t size, size_t alignment) {
struct ggml_allocr * alloc = (struct ggml_allocr *)malloc(sizeof(struct ggml_allocr) /* + n_free_blocks * sizeof(struct free_block) */);
*alloc = (struct ggml_allocr){
/*.data = */ data,
/*.size = */ size,
/*.alignment = */ alignment,
/*.n_free_blocks = */ 0,
/*.free_blocks = */ {{0}},
/*.hash_table = */ {{0}},
/*.max_size = */ 0,
/*.measure = */ false,
/*.parse_seq = */ {0},
/*.parse_seq_len = */ 0,
#ifdef GGML_ALLOCATOR_DEBUG
/*.allocated_tensors = */ {0},
#endif
};
ggml_allocr_reset(alloc);
return alloc;
}
// address and size of the buffer when measuring
// it needs to be large enough to fit all the tensors, but it cannot overlap with other existing buffers
static void * const MEASURE_BASE_ADDR = (void *) 0x1000;
static const size_t MEASURE_MAX_SIZE = 1ULL<<40; // 1 TB
struct ggml_allocr * ggml_allocr_new_measure(size_t alignment) {
struct ggml_allocr * alloc = (struct ggml_allocr *)malloc(sizeof(struct ggml_allocr) /* + n_free_blocks * sizeof(struct free_block) */);
*alloc = (struct ggml_allocr){
/*.data = */ MEASURE_BASE_ADDR,
/*.size = */ MEASURE_MAX_SIZE,
/*.alignment = */ alignment,
/*.n_free_blocks = */ 0,
/*.free_blocks = */ {{0}},
/*.hash_table = */ {{0}},
/*.max_size = */ 0,
/*.measure = */ true,
/*.parse_seq = */ {0},
/*.parse_seq_len = */ 0,
#ifdef GGML_ALLOCATOR_DEBUG
/*.allocated_tensors = */ {0},
#endif
};
ggml_allocr_reset(alloc);
return alloc;
}
void ggml_allocr_free(struct ggml_allocr * alloc) {
free(alloc);
}
bool ggml_allocr_is_measure(struct ggml_allocr * alloc) {
return alloc->measure;
}
//////////// compute graph allocator
static bool ggml_is_view(struct ggml_tensor * t) {
return t->op == GGML_OP_RESHAPE || t->op == GGML_OP_VIEW || t->op == GGML_OP_TRANSPOSE ||
t->op == GGML_OP_PERMUTE || t->op == GGML_OP_CPY;
}
static bool ggml_are_same_layout(const struct ggml_tensor * a, const struct ggml_tensor * b) {
if (a->type != b->type) {
return false;
}
for (int i = 0; i < GGML_MAX_DIMS; i++) {
if (a->ne[i] != b->ne[i]) {
return false;
}
if (a->nb[i] != b->nb[i]) {
return false;
}
}
return true;
}
static struct ggml_tensor * get_view_parent(struct ggml_tensor * t) {
switch (t->op) {
case GGML_OP_PERMUTE:
case GGML_OP_RESHAPE:
case GGML_OP_TRANSPOSE:
case GGML_OP_VIEW:
return t->src[0];
case GGML_OP_CPY:
return t->src[1];
default:
return NULL;
}
}
static struct ggml_tensor * get_view_source(struct ggml_tensor * t) {
struct ggml_tensor * parent = t;
do {
parent = get_view_parent(parent);
} while (ggml_is_view(parent));
return parent;
}
static bool ggml_op_can_inplace(enum ggml_op op) {
switch (op) {
case GGML_OP_SCALE:
case GGML_OP_DIAG_MASK_ZERO:
case GGML_OP_DIAG_MASK_INF:
case GGML_OP_ADD:
case GGML_OP_ADD1:
case GGML_OP_ACC:
case GGML_OP_SUB:
case GGML_OP_MUL:
case GGML_OP_DIV:
case GGML_OP_SQR:
case GGML_OP_SQRT:
case GGML_OP_LOG:
case GGML_OP_UNARY:
case GGML_OP_ROPE:
case GGML_OP_RMS_NORM:
case GGML_OP_SET:
case GGML_OP_SOFT_MAX:
case GGML_OP_CONT:
case GGML_OP_ADD_REL_POS:
return true;
default:
return false;
}
}
static void allocate_node(struct ggml_allocr * alloc, struct ggml_tensor * node) {
struct hash_node * ht = alloc->hash_table;
if (node->data == NULL) {
if (ggml_is_view(node)) {
size_t offset;
switch(node->op) {
case GGML_OP_VIEW:
memcpy(&offset, node->op_params, sizeof(size_t));
node->data = (char *) node->src[0]->data + offset;
break;
case GGML_OP_PERMUTE:
case GGML_OP_RESHAPE:
case GGML_OP_TRANSPOSE:
node->data = node->src[0]->data;
break;
case GGML_OP_CPY:
node->data = node->src[1]->data;
break;
default:
GGML_ASSERT(!"unknown view op");
break;
}
} else {
// see if we can reuse a parent's buffer (inplace)
if (ggml_op_can_inplace(node->op)) {
for (int i = 0; i < GGML_MAX_SRC; i++) {
struct ggml_tensor * parent = node->src[i];
if (parent == NULL) {
break;
}
// if the node's data is external, then we cannot re-use it
if ((char *) parent->data < (char *) alloc->data ||
(char *) parent->data >= ((char *) alloc->data + alloc->size)) {
AT_PRINTF("not reusing parent %s for %s as %p is external\n", parent->name, node->name, parent->data);
continue;
}
struct hash_node * p_hn = hash_get(ht, parent);
if (parent->data != NULL && p_hn->n_children == 1 && p_hn->n_views == 0 && ggml_are_same_layout(node, parent)) {
if (ggml_is_view(parent)) {
struct ggml_tensor * view_src = get_view_source(parent);
struct hash_node * view_src_hn = hash_get(ht, view_src);
if (view_src_hn->n_views == 1 && view_src_hn->n_children == 0 && view_src->data == parent->data) {
// TODO: the offset of the view parent must be kept to ensure that the op doesn't overwrite
// the parent's data that it will need later (same layout requirement). the problem is that then
// we cannot free the tensor because the original address of the allocation is lost.
// adding a view_src pointer to the tensor would solve this and simplify the code dealing with views
// for now, we only reuse the parent's data if the offset is zero (view_src->data == parent->data)
AT_PRINTF("reusing view parent %s (%s) for %s\n", parent->name, view_src->name, node->name);
node->data = parent->data;
return;
}
}
else {
AT_PRINTF("reusing parent %s for %s\n", parent->name, node->name);
node->data = parent->data;
return;
}
}
}
}
ggml_allocr_alloc(alloc, node);
}
}
}
static size_t ggml_allocator_alloc_graph_tensors_n(
struct ggml_allocr * alloc,
struct ggml_cgraph ** graphs, int n_graphs,
struct ggml_tensor *** inputs, struct ggml_tensor *** outputs) {
// reset hash table
struct hash_node * ht = alloc->hash_table;
memset(ht, 0, sizeof(struct hash_node) * GGML_GRAPH_HASHTABLE_SIZE);
// count number of children and views
for (int g = 0; g < n_graphs; g++) {
struct ggml_cgraph * gf = graphs[g];
for (int i = 0; i < gf->n_nodes; i++) {
struct ggml_tensor * node = gf->nodes[i];
if (ggml_is_view(node)) {
struct ggml_tensor * view_src = get_view_source(node);
hash_get(ht, view_src)->n_views += 1;
}
for (int j = 0; j < GGML_MAX_SRC; j++) {
struct ggml_tensor * parent = node->src[j];
if (parent == NULL) {
break;
}
hash_get(ht, parent)->n_children += 1;
}
}
}
// allocate tensors
for (int g = 0; g < n_graphs; g++) {
struct ggml_cgraph * gf = graphs[g];
AT_PRINTF("####### graph %d/%d\n", g, n_graphs);
// graph inputs are allocated first to ensure that they are not overwritten by each other
if (inputs != NULL && inputs[g] != NULL) {
for (int i = 0; inputs[g][i] != NULL; i++) {
struct ggml_tensor * input = inputs[g][i];
AT_PRINTF("input: %s\n", input->name);
allocate_node(alloc, input);
}
}
// if we have parse_seq then we allocate nodes following the list, and we only free nodes at barriers
int last_barrier_pos = 0;
int n_nodes = alloc->parse_seq_len ? alloc->parse_seq_len : gf->n_nodes;
for (int ind = 0; ind < n_nodes; ind++) {
// allocate a node if there is no parse_seq or this is not a barrier
if ((alloc->parse_seq_len==0) || alloc->parse_seq[ind] != -1) {
int i = alloc->parse_seq_len ? alloc->parse_seq[ind] : ind;
struct ggml_tensor * node = gf->nodes[i];
// allocate parents (leafs)
for (int j = 0; j < GGML_MAX_SRC; j++) {
struct ggml_tensor * parent = node->src[j];
if (parent == NULL) {
break;
}
allocate_node(alloc, parent);
}
// allocate node
allocate_node(alloc, node);
AT_PRINTF("exec: %s (%s) <= ", ggml_op_name(node->op), node->name);
for (int j = 0; j < GGML_MAX_SRC; j++) {
struct ggml_tensor * parent = node->src[j];
if (parent == NULL) {
break;
}
AT_PRINTF("%s", parent->name);
if (j < GGML_MAX_SRC - 1 && node->src[j + 1] != NULL) {
AT_PRINTF(", ");
}
}
AT_PRINTF("\n");
}
// update parents
// update immediately if there is no parse_seq
// update only at barriers if there is parse_seq
if ((alloc->parse_seq_len==0) || alloc->parse_seq[ind] == -1) {
int update_start = alloc->parse_seq_len ? last_barrier_pos : ind;
int update_end = alloc->parse_seq_len ? ind : ind + 1;
for (int i = update_start; i < update_end; i++) {
int node_i = alloc->parse_seq_len ? alloc->parse_seq[i] : i;
struct ggml_tensor * node = gf->nodes[node_i];
for (int j = 0; j < GGML_MAX_SRC; j++) {
struct ggml_tensor * parent = node->src[j];
if (parent == NULL) {
break;
}
struct hash_node * p_hn = hash_get(ht, parent);
p_hn->n_children -= 1;
//AT_PRINTF("parent %s: %d children, %d views\n", parent->name, parent->n_children, parent->n_views);
if (p_hn->n_children == 0 && p_hn->n_views == 0) {
if (ggml_is_view(parent)) {
struct ggml_tensor * view_src = get_view_source(parent);
struct hash_node * view_src_hn = hash_get(ht, view_src);
view_src_hn->n_views -= 1;
AT_PRINTF("view_src %s: %d children, %d views\n", view_src->name, view_src_hn->n_children, view_src_hn->n_views);
if (view_src_hn->n_views == 0 && view_src_hn->n_children == 0 && view_src->data != node->data) {
ggml_allocator_free_tensor(alloc, view_src);
}
}
else {
if (parent->data != node->data) {
ggml_allocator_free_tensor(alloc, parent);
}
}
}
}
}
AT_PRINTF("\n");
if (alloc->parse_seq_len) {
last_barrier_pos = ind + 1;
}
}
}
// free graph outputs here that wouldn't be freed otherwise because they have no children
if (outputs != NULL && outputs[g] != NULL) {
for (int i = 0; outputs[g][i] != NULL; i++) {
struct ggml_tensor * output = outputs[g][i];
AT_PRINTF("output: %s\n", output->name);
ggml_allocator_free_tensor(alloc, output);
}
}
}
return alloc->max_size;
}
size_t ggml_allocr_alloc_graph(struct ggml_allocr * alloc, struct ggml_cgraph * graph) {
return ggml_allocator_alloc_graph_tensors_n(alloc, &graph, 1, NULL, NULL);
}

26
ggml-alloc.h Normal file
View File

@ -0,0 +1,26 @@
#pragma once
#include "ggml.h"
#ifdef __cplusplus
extern "C" {
#endif
GGML_API struct ggml_allocr * ggml_allocr_new(void * data, size_t size, size_t alignment);
GGML_API struct ggml_allocr * ggml_allocr_new_measure(size_t alignment);
// tell the allocator to parse nodes following the order described in the list
// you should call this if your graph are optimized to execute out-of-order
GGML_API void ggml_allocr_set_parse_seq(struct ggml_allocr * alloc, const int * list, int n);
GGML_API void ggml_allocr_free(struct ggml_allocr * alloc);
GGML_API bool ggml_allocr_is_measure(struct ggml_allocr * alloc);
GGML_API void ggml_allocr_reset(struct ggml_allocr * alloc);
GGML_API void ggml_allocr_alloc(struct ggml_allocr * alloc, struct ggml_tensor * tensor);
GGML_API size_t ggml_allocr_alloc_graph(struct ggml_allocr * alloc, struct ggml_cgraph * graph);
#ifdef __cplusplus
}
#endif

4196
ggml-cuda.cu
View File

File diff suppressed because it is too large Load Diff

46
ggml-cuda.h
View File

@ -2,34 +2,44 @@
#include "ggml.h"
#ifdef GGML_USE_HIPBLAS
#define GGML_CUDA_NAME "ROCm"
#define GGML_CUBLAS_NAME "hipBLAS"
#else
#define GGML_CUDA_NAME "CUDA"
#define GGML_CUBLAS_NAME "cuBLAS"
#endif
#ifdef __cplusplus
extern "C" {
#endif
#define GGML_CUDA_MAX_DEVICES 16
void ggml_init_cublas(void);
void ggml_cuda_set_tensor_split(const float * tensor_split);
GGML_API void ggml_init_cublas(void);
GGML_API void * ggml_cuda_host_malloc(size_t size);
GGML_API void ggml_cuda_host_free(void * ptr);
void ggml_cuda_mul(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
bool ggml_cuda_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
size_t ggml_cuda_mul_mat_get_wsize(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
void ggml_cuda_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst, void * wdata, size_t wsize);
GGML_API bool ggml_cuda_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
GGML_API void ggml_cuda_set_tensor_split(const float * tensor_split);
GGML_API void ggml_cuda_transform_tensor(void * data, struct ggml_tensor * tensor);
GGML_API void ggml_cuda_free_data(struct ggml_tensor * tensor);
// TODO: export these with GGML_API
void * ggml_cuda_host_malloc(size_t size);
void ggml_cuda_host_free(void * ptr);
GGML_API void ggml_cuda_assign_buffers(struct ggml_tensor * tensor);
GGML_API void ggml_cuda_assign_buffers_no_scratch(struct ggml_tensor * tensor);
GGML_API void ggml_cuda_assign_buffers_force_inplace(struct ggml_tensor * tensor);
void ggml_cuda_transform_tensor(void * data, struct ggml_tensor * tensor);
GGML_API void ggml_cuda_assign_buffers_no_alloc(struct ggml_tensor * tensor);
GGML_API void ggml_cuda_assign_scratch_offset(struct ggml_tensor * tensor, size_t offset);
void ggml_cuda_free_data(struct ggml_tensor * tensor);
void ggml_cuda_assign_buffers(struct ggml_tensor * tensor);
void ggml_cuda_assign_buffers_no_scratch(struct ggml_tensor * tensor);
void ggml_cuda_assign_buffers_force_inplace(struct ggml_tensor * tensor);
void ggml_cuda_set_main_device(int main_device);
void ggml_cuda_set_scratch_size(size_t scratch_size);
void ggml_cuda_free_scratch(void);
bool ggml_cuda_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * tensor);
GGML_API void ggml_cuda_set_main_device(int main_device);
GGML_API void ggml_cuda_set_mul_mat_q(bool mul_mat_q);
GGML_API void ggml_cuda_set_scratch_size(size_t scratch_size);
GGML_API void ggml_cuda_free_scratch(void);
GGML_API bool ggml_cuda_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * tensor);
GGML_API int ggml_cuda_get_device_count(void);
GGML_API void ggml_cuda_get_device_description(int device, char * description, size_t description_size);
#ifdef __cplusplus
}

20
ggml-metal.h
View File

@ -24,6 +24,7 @@
// max memory buffers that can be mapped to the device
#define GGML_METAL_MAX_BUFFERS 16
#define GGML_METAL_MAX_COMMAND_BUFFERS 32
struct ggml_tensor;
struct ggml_cgraph;
@ -34,9 +35,16 @@ extern "C" {
struct ggml_metal_context;
struct ggml_metal_context * ggml_metal_init(void);
// number of command buffers to use
struct ggml_metal_context * ggml_metal_init(int n_cb);
void ggml_metal_free(struct ggml_metal_context * ctx);
void * ggml_metal_host_malloc(size_t n);
void ggml_metal_host_free (void * data);
// set the number of command buffers to use
void ggml_metal_set_n_cb(struct ggml_metal_context * ctx, int n_cb);
// creates a mapping between a host memory buffer and a device memory buffer
// - make sure to map all buffers used in the graph before calling ggml_metal_graph_compute
// - the mapping is used during computation to determine the arguments of the compute kernels
@ -57,6 +65,16 @@ void ggml_metal_set_tensor(struct ggml_metal_context * ctx, struct ggml_tensor *
// get data from the device into host memory
void ggml_metal_get_tensor(struct ggml_metal_context * ctx, struct ggml_tensor * t);
// try to find operations that can be run concurrently in the graph
// you should run it again if the topology of your graph changes
void ggml_metal_graph_find_concurrency(struct ggml_metal_context * ctx, struct ggml_cgraph * gf, bool check_mem);
// if the graph has been optimized for concurrently dispatch, return length of the concur_list if optimized
int ggml_metal_if_optimized(struct ggml_metal_context * ctx);
// output the concur_list for ggml_alloc
int * ggml_metal_get_concur_list(struct ggml_metal_context * ctx);
// same as ggml_graph_compute but uses Metal
// creates gf->n_threads command buffers in parallel
void ggml_metal_graph_compute(struct ggml_metal_context * ctx, struct ggml_cgraph * gf);

575
ggml-metal.m
View File

@ -5,7 +5,11 @@
#import <Foundation/Foundation.h>
#import <Metal/Metal.h>
#import <MetalPerformanceShaders/MetalPerformanceShaders.h>
#undef MIN
#undef MAX
#define MIN(a, b) ((a) < (b) ? (a) : (b))
#define MAX(a, b) ((a) > (b) ? (a) : (b))
#ifdef GGML_METAL_NDEBUG
#define metal_printf(...)
@ -15,6 +19,8 @@
#define UNUSED(x) (void)(x)
#define GGML_MAX_CONCUR (2*GGML_MAX_NODES)
struct ggml_metal_buffer {
const char * name;
@ -25,21 +31,30 @@ struct ggml_metal_buffer {
};
struct ggml_metal_context {
float * logits;
int n_cb;
id<MTLDevice> device;
id<MTLCommandQueue> queue;
id<MTLLibrary> library;
id<MTLCommandBuffer> command_buffers [GGML_METAL_MAX_COMMAND_BUFFERS];
id<MTLComputeCommandEncoder> command_encoders[GGML_METAL_MAX_COMMAND_BUFFERS];
dispatch_queue_t d_queue;
int n_buffers;
struct ggml_metal_buffer buffers[GGML_METAL_MAX_BUFFERS];
int concur_list[GGML_MAX_CONCUR];
int concur_list_len;
// custom kernels
#define GGML_METAL_DECL_KERNEL(name) \
id<MTLFunction> function_##name; \
id<MTLComputePipelineState> pipeline_##name
GGML_METAL_DECL_KERNEL(add);
GGML_METAL_DECL_KERNEL(add_row); // TODO: avoid this extra kernel, instead extend the "add" kernel to support broadcast
GGML_METAL_DECL_KERNEL(mul);
GGML_METAL_DECL_KERNEL(mul_row); // TODO: avoid this extra kernel, instead extend the "mul" kernel to support broadcast
GGML_METAL_DECL_KERNEL(scale);
@ -51,6 +66,7 @@ struct ggml_metal_context {
GGML_METAL_DECL_KERNEL(get_rows_f16);
GGML_METAL_DECL_KERNEL(get_rows_q4_0);
GGML_METAL_DECL_KERNEL(get_rows_q4_1);
GGML_METAL_DECL_KERNEL(get_rows_q8_0);
GGML_METAL_DECL_KERNEL(get_rows_q2_K);
GGML_METAL_DECL_KERNEL(get_rows_q3_K);
GGML_METAL_DECL_KERNEL(get_rows_q4_K);
@ -61,11 +77,21 @@ struct ggml_metal_context {
GGML_METAL_DECL_KERNEL(mul_mat_f16_f32);
GGML_METAL_DECL_KERNEL(mul_mat_q4_0_f32);
GGML_METAL_DECL_KERNEL(mul_mat_q4_1_f32);
GGML_METAL_DECL_KERNEL(mul_mat_q8_0_f32);
GGML_METAL_DECL_KERNEL(mul_mat_q2_K_f32);
GGML_METAL_DECL_KERNEL(mul_mat_q3_K_f32);
GGML_METAL_DECL_KERNEL(mul_mat_q4_K_f32);
GGML_METAL_DECL_KERNEL(mul_mat_q5_K_f32);
GGML_METAL_DECL_KERNEL(mul_mat_q6_K_f32);
GGML_METAL_DECL_KERNEL(mul_mm_f16_f32);
GGML_METAL_DECL_KERNEL(mul_mm_q4_0_f32);
GGML_METAL_DECL_KERNEL(mul_mm_q4_1_f32);
GGML_METAL_DECL_KERNEL(mul_mm_q8_0_f32);
GGML_METAL_DECL_KERNEL(mul_mm_q2_K_f32);
GGML_METAL_DECL_KERNEL(mul_mm_q3_K_f32);
GGML_METAL_DECL_KERNEL(mul_mm_q4_K_f32);
GGML_METAL_DECL_KERNEL(mul_mm_q5_K_f32);
GGML_METAL_DECL_KERNEL(mul_mm_q6_K_f32);
GGML_METAL_DECL_KERNEL(rope);
GGML_METAL_DECL_KERNEL(alibi_f32);
GGML_METAL_DECL_KERNEL(cpy_f32_f16);
@ -86,22 +112,18 @@ static NSString * const msl_library_source = @"see metal.metal";
@implementation GGMLMetalClass
@end
struct ggml_metal_context * ggml_metal_init(void) {
struct ggml_metal_context * ggml_metal_init(int n_cb) {
fprintf(stderr, "%s: allocating\n", __func__);
struct ggml_metal_context * ctx = malloc(sizeof(struct ggml_metal_context));
ctx->n_cb = MIN(n_cb, GGML_METAL_MAX_BUFFERS);
ctx->device = MTLCreateSystemDefaultDevice();
ctx->queue = [ctx->device newCommandQueue];
ctx->n_buffers = 0;
ctx->concur_list_len = 0;
// determine if we can use MPS
if (MPSSupportsMTLDevice(ctx->device)) {
fprintf(stderr, "%s: using MPS\n", __func__);
} else {
fprintf(stderr, "%s: not using MPS\n", __func__);
GGML_ASSERT(false && "MPS not supported");
}
ctx->d_queue = dispatch_queue_create("llama.cpp", DISPATCH_QUEUE_CONCURRENT);
#if 0
// compile from source string and show compile log
@ -111,7 +133,7 @@ struct ggml_metal_context * ggml_metal_init(void) {
ctx->library = [ctx->device newLibraryWithSource:msl_library_source options:nil error:&error];
if (error) {
fprintf(stderr, "%s: error: %s\n", __func__, [[error description] UTF8String]);
exit(1);
return NULL;
}
}
#else
@ -129,7 +151,7 @@ struct ggml_metal_context * ggml_metal_init(void) {
NSString * src = [NSString stringWithContentsOfFile:path encoding:NSUTF8StringEncoding error:&error];
if (error) {
fprintf(stderr, "%s: error: %s\n", __func__, [[error description] UTF8String]);
exit(1);
return NULL;
}
#ifdef GGML_QKK_64
@ -141,19 +163,27 @@ struct ggml_metal_context * ggml_metal_init(void) {
#endif
if (error) {
fprintf(stderr, "%s: error: %s\n", __func__, [[error description] UTF8String]);
exit(1);
return NULL;
}
}
#endif
// load kernels
{
NSError * error = nil;
#define GGML_METAL_ADD_KERNEL(name) \
ctx->function_##name = [ctx->library newFunctionWithName:@"kernel_"#name]; \
ctx->pipeline_##name = [ctx->device newComputePipelineStateWithFunction:ctx->function_##name error:nil]; \
fprintf(stderr, "%s: loaded %-32s %16p\n", __func__, "kernel_"#name, (void *) ctx->pipeline_##name);
ctx->pipeline_##name = [ctx->device newComputePipelineStateWithFunction:ctx->function_##name error:&error]; \
fprintf(stderr, "%s: loaded %-32s %16p | th_max = %4d | th_width = %4d\n", __func__, "kernel_"#name, (void *) ctx->pipeline_##name, \
(int) ctx->pipeline_##name.maxTotalThreadsPerThreadgroup, \
(int) ctx->pipeline_##name.threadExecutionWidth); \
if (error) { \
fprintf(stderr, "%s: load pipeline error: %s\n", __func__, [[error description] UTF8String]); \
return NULL; \
}
GGML_METAL_ADD_KERNEL(add);
GGML_METAL_ADD_KERNEL(add_row);
GGML_METAL_ADD_KERNEL(mul);
GGML_METAL_ADD_KERNEL(mul_row);
GGML_METAL_ADD_KERNEL(scale);
@ -165,6 +195,7 @@ struct ggml_metal_context * ggml_metal_init(void) {
GGML_METAL_ADD_KERNEL(get_rows_f16);
GGML_METAL_ADD_KERNEL(get_rows_q4_0);
GGML_METAL_ADD_KERNEL(get_rows_q4_1);
GGML_METAL_ADD_KERNEL(get_rows_q8_0);
GGML_METAL_ADD_KERNEL(get_rows_q2_K);
GGML_METAL_ADD_KERNEL(get_rows_q3_K);
GGML_METAL_ADD_KERNEL(get_rows_q4_K);
@ -175,11 +206,21 @@ struct ggml_metal_context * ggml_metal_init(void) {
GGML_METAL_ADD_KERNEL(mul_mat_f16_f32);
GGML_METAL_ADD_KERNEL(mul_mat_q4_0_f32);
GGML_METAL_ADD_KERNEL(mul_mat_q4_1_f32);
GGML_METAL_ADD_KERNEL(mul_mat_q8_0_f32);
GGML_METAL_ADD_KERNEL(mul_mat_q2_K_f32);
GGML_METAL_ADD_KERNEL(mul_mat_q3_K_f32);
GGML_METAL_ADD_KERNEL(mul_mat_q4_K_f32);
GGML_METAL_ADD_KERNEL(mul_mat_q5_K_f32);
GGML_METAL_ADD_KERNEL(mul_mat_q6_K_f32);
GGML_METAL_ADD_KERNEL(mul_mm_f16_f32);
GGML_METAL_ADD_KERNEL(mul_mm_q4_0_f32);
GGML_METAL_ADD_KERNEL(mul_mm_q8_0_f32);
GGML_METAL_ADD_KERNEL(mul_mm_q4_1_f32);
GGML_METAL_ADD_KERNEL(mul_mm_q2_K_f32);
GGML_METAL_ADD_KERNEL(mul_mm_q3_K_f32);
GGML_METAL_ADD_KERNEL(mul_mm_q4_K_f32);
GGML_METAL_ADD_KERNEL(mul_mm_q5_K_f32);
GGML_METAL_ADD_KERNEL(mul_mm_q6_K_f32);
GGML_METAL_ADD_KERNEL(rope);
GGML_METAL_ADD_KERNEL(alibi_f32);
GGML_METAL_ADD_KERNEL(cpy_f32_f16);
@ -202,12 +243,97 @@ struct ggml_metal_context * ggml_metal_init(void) {
void ggml_metal_free(struct ggml_metal_context * ctx) {
fprintf(stderr, "%s: deallocating\n", __func__);
#define GGML_METAL_DEL_KERNEL(name) \
[ctx->function_##name release]; \
[ctx->pipeline_##name release];
GGML_METAL_DEL_KERNEL(add);
GGML_METAL_DEL_KERNEL(add_row);
GGML_METAL_DEL_KERNEL(mul);
GGML_METAL_DEL_KERNEL(mul_row);
GGML_METAL_DEL_KERNEL(scale);
GGML_METAL_DEL_KERNEL(silu);
GGML_METAL_DEL_KERNEL(relu);
GGML_METAL_DEL_KERNEL(gelu);
GGML_METAL_DEL_KERNEL(soft_max);
GGML_METAL_DEL_KERNEL(diag_mask_inf);
GGML_METAL_DEL_KERNEL(get_rows_f16);
GGML_METAL_DEL_KERNEL(get_rows_q4_0);
GGML_METAL_DEL_KERNEL(get_rows_q4_1);
GGML_METAL_DEL_KERNEL(get_rows_q8_0);
GGML_METAL_DEL_KERNEL(get_rows_q2_K);
GGML_METAL_DEL_KERNEL(get_rows_q3_K);
GGML_METAL_DEL_KERNEL(get_rows_q4_K);
GGML_METAL_DEL_KERNEL(get_rows_q5_K);
GGML_METAL_DEL_KERNEL(get_rows_q6_K);
GGML_METAL_DEL_KERNEL(rms_norm);
GGML_METAL_DEL_KERNEL(norm);
GGML_METAL_DEL_KERNEL(mul_mat_f16_f32);
GGML_METAL_DEL_KERNEL(mul_mat_q4_0_f32);
GGML_METAL_DEL_KERNEL(mul_mat_q4_1_f32);
GGML_METAL_DEL_KERNEL(mul_mat_q8_0_f32);
GGML_METAL_DEL_KERNEL(mul_mat_q2_K_f32);
GGML_METAL_DEL_KERNEL(mul_mat_q3_K_f32);
GGML_METAL_DEL_KERNEL(mul_mat_q4_K_f32);
GGML_METAL_DEL_KERNEL(mul_mat_q5_K_f32);
GGML_METAL_DEL_KERNEL(mul_mat_q6_K_f32);
GGML_METAL_DEL_KERNEL(mul_mm_f16_f32);
GGML_METAL_DEL_KERNEL(mul_mm_q4_0_f32);
GGML_METAL_DEL_KERNEL(mul_mm_q8_0_f32);
GGML_METAL_DEL_KERNEL(mul_mm_q4_1_f32);
GGML_METAL_DEL_KERNEL(mul_mm_q2_K_f32);
GGML_METAL_DEL_KERNEL(mul_mm_q3_K_f32);
GGML_METAL_DEL_KERNEL(mul_mm_q4_K_f32);
GGML_METAL_DEL_KERNEL(mul_mm_q5_K_f32);
GGML_METAL_DEL_KERNEL(mul_mm_q6_K_f32);
GGML_METAL_DEL_KERNEL(rope);
GGML_METAL_DEL_KERNEL(alibi_f32);
GGML_METAL_DEL_KERNEL(cpy_f32_f16);
GGML_METAL_DEL_KERNEL(cpy_f32_f32);
GGML_METAL_DEL_KERNEL(cpy_f16_f16);
#undef GGML_METAL_DEL_KERNEL
for (int i = 0; i < ctx->n_buffers; ++i) {
[ctx->buffers[i].metal release];
}
[ctx->library release];
[ctx->queue release];
[ctx->device release];
dispatch_release(ctx->d_queue);
free(ctx);
}
void * ggml_metal_host_malloc(size_t n) {
void * data = NULL;
const int result = posix_memalign((void **) &data, getpagesize(), n);
if (result != 0) {
fprintf(stderr, "%s: error: posix_memalign failed\n", __func__);
return NULL;
}
return data;
}
void ggml_metal_host_free(void * data) {
free(data);
}
void ggml_metal_set_n_cb(struct ggml_metal_context * ctx, int n_cb) {
ctx->n_cb = MIN(n_cb, GGML_METAL_MAX_BUFFERS);
}
int ggml_metal_if_optimized(struct ggml_metal_context * ctx) {
return ctx->concur_list_len;
}
int * ggml_metal_get_concur_list(struct ggml_metal_context * ctx) {
return ctx->concur_list;
}
// finds the Metal buffer that contains the tensor data on the GPU device
// the assumption is that there is 1-to-1 mapping between the host and device memory buffers, so we can find the
// Metal buffer based on the host memory pointer
@ -346,48 +472,154 @@ void ggml_metal_get_tensor(
memcpy(t->data, (void *) ((uint8_t *) id_src.contents + offs), ggml_nbytes(t));
}
void ggml_metal_graph_find_concurrency(
struct ggml_metal_context * ctx,
struct ggml_cgraph * gf, bool check_mem) {
int search_depth = gf->n_nodes; //we only find concurrency in this range to avoid wasting too much time
int nodes_unused[GGML_MAX_CONCUR];
for (int i = 0; i < GGML_MAX_CONCUR; i++) { ctx->concur_list[i] = 0; }
for (int i = 0; i < gf->n_nodes; i++) { nodes_unused[i] = 1; }
ctx->concur_list_len = 0;
int n_left = gf->n_nodes;
int n_start = 0; // all nodes before n_start at nodes_unused array have been sorted and store back to ctx->concur_list
int level_pos = 0; // at ctx->concur_list, the last layer (level) ends at level_pos
while (n_left > 0) {
// number of nodes at a layer (that can be issued concurrently)
int concurrency = 0;
for (int i = n_start; i < ((n_start + search_depth > gf->n_nodes) ? gf->n_nodes : n_start + search_depth); i++) {
if (nodes_unused[i]) {
// if the requirements for gf->nodes[i] are satisfied
int exe_flag = 1;
// scan all srcs
for (int src_ind = 0; src_ind < GGML_MAX_SRC; src_ind++) {
struct ggml_tensor * src_cur = gf->nodes[i]->src[src_ind];
if (src_cur) {
// if is leaf nodes it's satisfied.
// TODO: ggml_is_leaf()
if (src_cur->op == GGML_OP_NONE && src_cur->grad == NULL) {
continue;
}
// otherwise this src should be the output from previous nodes.
int is_found = 0;
// scan 2*search_depth back because we inserted barrier.
//for (int j = ((level_pos - 2*search_depth) < 0 ? 0 : (level_pos - 2*search_depth)); j < level_pos; j++) {
for (int j = MAX(0, level_pos - 2*search_depth); j < level_pos; j++) {
if (ctx->concur_list[j] >= 0 && gf->nodes[ctx->concur_list[j]] == src_cur) {
is_found = 1;
break;
}
}
if (is_found == 0) {
exe_flag = 0;
break;
}
}
}
if (exe_flag && check_mem) {
// check if nodes[i]'s data will be overwritten by a node before nodes[i].
// if node[5] and node[3] write to the same memory region, then we can't issue node[5] before node[3]
int64_t data_start = (int64_t) gf->nodes[i]->data;
int64_t length = (int64_t) ggml_nbytes(gf->nodes[i]);
for (int j = n_start; j < i; j++) {
if (nodes_unused[j] && gf->nodes[j]->op != GGML_OP_RESHAPE \
&& gf->nodes[j]->op != GGML_OP_VIEW \
&& gf->nodes[j]->op != GGML_OP_TRANSPOSE \
&& gf->nodes[j]->op != GGML_OP_PERMUTE) {
if (((int64_t)gf->nodes[j]->data) >= data_start + length || \
((int64_t)gf->nodes[j]->data) + (int64_t) ggml_nbytes(gf->nodes[j]) <= data_start) {
continue;
}
exe_flag = 0;
}
}
}
if (exe_flag) {
ctx->concur_list[level_pos + concurrency] = i;
nodes_unused[i] = 0;
concurrency++;
ctx->concur_list_len++;
}
}
}
n_left -= concurrency;
// adding a barrier different layer
ctx->concur_list[level_pos + concurrency] = -1;
ctx->concur_list_len++;
// jump all sorted nodes at nodes_bak
while (!nodes_unused[n_start]) {
n_start++;
}
level_pos += concurrency + 1;
}
if (ctx->concur_list_len > GGML_MAX_CONCUR) {
fprintf(stderr, "%s: too many elements for metal ctx->concur_list!\n", __func__);
}
}
void ggml_metal_graph_compute(
struct ggml_metal_context * ctx,
struct ggml_cgraph * gf) {
metal_printf("%s: evaluating graph\n", __func__);
@autoreleasepool {
// if there is ctx->concur_list, dispatch concurrently
// else fallback to serial dispatch
MTLComputePassDescriptor * edesc = MTLComputePassDescriptor.computePassDescriptor;
const bool has_concur = ctx->concur_list_len && ctx->concur_list_len <= GGML_MAX_CONCUR;
const int n_nodes = has_concur ? ctx->concur_list_len : gf->n_nodes;
edesc.dispatchType = has_concur ? MTLDispatchTypeConcurrent : MTLDispatchTypeSerial;
// create multiple command buffers and enqueue them
// then, we encode the graph into the command buffers in parallel
const int n_cb = gf->n_threads;
NSMutableArray * command_buffers = [NSMutableArray arrayWithCapacity:n_cb];
const int n_cb = ctx->n_cb;
for (int i = 0; i < n_cb; ++i) {
command_buffers[i] = [ctx->queue commandBuffer];
ctx->command_buffers[i] = [ctx->queue commandBuffer];
// enqueue the command buffers in order to specify their execution order
[command_buffers[i] enqueue];
[ctx->command_buffers[i] enqueue];
ctx->command_encoders[i] = [ctx->command_buffers[i] computeCommandEncoderWithDescriptor: edesc];
}
// TODO: is this the best way to start threads?
dispatch_queue_t queue = dispatch_queue_create("llama.cpp", DISPATCH_QUEUE_CONCURRENT);
for (int cb_idx = 0; cb_idx < n_cb; ++cb_idx) {
const int n_nodes_per_cb = (gf->n_nodes + n_cb - 1) / n_cb;
const int n_nodes_per_cb = (n_nodes + n_cb - 1) / n_cb;
dispatch_async(queue, ^{
dispatch_async(ctx->d_queue, ^{
size_t offs_src0 = 0;
size_t offs_src1 = 0;
size_t offs_dst = 0;
id<MTLCommandBuffer> command_buffer = command_buffers[cb_idx];
id<MTLComputeCommandEncoder> encoder = nil;
id<MTLCommandBuffer> command_buffer = ctx->command_buffers[cb_idx];
id<MTLComputeCommandEncoder> encoder = ctx->command_encoders[cb_idx];
const int node_start = (cb_idx + 0) * n_nodes_per_cb;
const int node_end = (cb_idx == n_cb - 1) ? gf->n_nodes : (cb_idx + 1) * n_nodes_per_cb;
const int node_end = MIN((cb_idx == n_cb - 1) ? n_nodes : (cb_idx + 1) * n_nodes_per_cb, n_nodes);
for (int ind = node_start; ind < node_end; ++ind) {
const int i = has_concur ? ctx->concur_list[ind] : ind;
if (i == -1) {
[encoder memoryBarrierWithScope:MTLBarrierScopeBuffers];
continue;
}
for (int i = node_start; i < node_end; ++i) {
metal_printf("%s: encoding node %3d, op = %8s\n", __func__, i, ggml_op_name(gf->nodes[i]->op));
struct ggml_tensor * src0 = gf->nodes[i]->src0;
struct ggml_tensor * src1 = gf->nodes[i]->src1;
struct ggml_tensor * src0 = gf->nodes[i]->src[0];
struct ggml_tensor * src1 = gf->nodes[i]->src[1];
struct ggml_tensor * dst = gf->nodes[i];
const int64_t ne00 = src0 ? src0->ne[0] : 0;
@ -443,6 +675,7 @@ void ggml_metal_graph_compute(
//}
switch (dst->op) {
case GGML_OP_NONE:
case GGML_OP_RESHAPE:
case GGML_OP_VIEW:
case GGML_OP_TRANSPOSE:
@ -452,14 +685,16 @@ void ggml_metal_graph_compute(
} break;
case GGML_OP_ADD:
{
if (encoder == nil) {
encoder = [command_buffer computeCommandEncoder];
}
if (ggml_nelements(src1) == ne10) {
// src1 is a row
[encoder setComputePipelineState:ctx->pipeline_add_row];
} else {
[encoder setComputePipelineState:ctx->pipeline_add];
}
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
[encoder setBuffer:id_dst offset:offs_dst atIndex:2];
[encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
const int64_t n = ggml_nelements(dst);
@ -467,10 +702,6 @@ void ggml_metal_graph_compute(
} break;
case GGML_OP_MUL:
{
if (encoder == nil) {
encoder = [command_buffer computeCommandEncoder];
}
if (ggml_nelements(src1) == ne10) {
// src1 is a row
[encoder setComputePipelineState:ctx->pipeline_mul_row];
@ -488,10 +719,6 @@ void ggml_metal_graph_compute(
} break;
case GGML_OP_SCALE:
{
if (encoder == nil) {
encoder = [command_buffer computeCommandEncoder];
}
const float scale = *(const float *) src1->data;
[encoder setComputePipelineState:ctx->pipeline_scale];
@ -503,12 +730,10 @@ void ggml_metal_graph_compute(
[encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
} break;
case GGML_OP_SILU:
case GGML_OP_UNARY:
switch (ggml_get_unary_op(gf->nodes[i])) {
case GGML_UNARY_OP_SILU:
{
if (encoder == nil) {
encoder = [command_buffer computeCommandEncoder];
}
[encoder setComputePipelineState:ctx->pipeline_silu];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
@ -517,12 +742,8 @@ void ggml_metal_graph_compute(
[encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
} break;
case GGML_OP_RELU:
case GGML_UNARY_OP_RELU:
{
if (encoder == nil) {
encoder = [command_buffer computeCommandEncoder];
}
[encoder setComputePipelineState:ctx->pipeline_relu];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
@ -531,12 +752,8 @@ void ggml_metal_graph_compute(
[encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
} break;
case GGML_OP_GELU:
case GGML_UNARY_OP_GELU:
{
if (encoder == nil) {
encoder = [command_buffer computeCommandEncoder];
}
[encoder setComputePipelineState:ctx->pipeline_gelu];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
@ -545,12 +762,14 @@ void ggml_metal_graph_compute(
[encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
} break;
default:
{
fprintf(stderr, "%s: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(dst->op));
GGML_ASSERT(false);
}
} break;
case GGML_OP_SOFT_MAX:
{
if (encoder == nil) {
encoder = [command_buffer computeCommandEncoder];
}
const int nth = 32;
[encoder setComputePipelineState:ctx->pipeline_soft_max];
@ -565,11 +784,7 @@ void ggml_metal_graph_compute(
} break;
case GGML_OP_DIAG_MASK_INF:
{
if (encoder == nil) {
encoder = [command_buffer computeCommandEncoder];
}
const int n_past = ((int32_t *)(src1->data))[0];
const int n_past = ((int32_t *)(dst->op_params))[0];
[encoder setComputePipelineState:ctx->pipeline_diag_mask_inf];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
@ -585,53 +800,44 @@ void ggml_metal_graph_compute(
// TODO: needs to be updated after PR: https://github.com/ggerganov/ggml/pull/224
GGML_ASSERT(ne00 == ne10);
GGML_ASSERT(ne02 == ne12);
// GGML_ASSERT(ne02 == ne12); // Should be checked on individual data types until broadcast is implemented everywhere
uint gqa = ne12/ne02;
GGML_ASSERT(ne03 == ne13);
// for now the matrix-matrix multiplication kernel only works on A14+/M1+ SoCs
// AMD GPU and older A-chips will reuse matrix-vector multiplication kernel
if (ggml_is_contiguous(src0) &&
ggml_is_contiguous(src1) &&
(src0t == GGML_TYPE_F32 || src0t == GGML_TYPE_F16) && ne11 > 1) {
if (encoder != nil) {
[encoder endEncoding];
encoder = nil;
}
MPSDataType src0dt = src0t == GGML_TYPE_F32 ? MPSDataTypeFloat32 : MPSDataTypeFloat16;
MPSDataType src1dt = src1t == GGML_TYPE_F32 ? MPSDataTypeFloat32 : MPSDataTypeFloat16;
// for F32 x F32 we use MPS
MPSMatrixDescriptor * desc0 = [MPSMatrixDescriptor
matrixDescriptorWithRows:ne01 columns:ne00 rowBytes:src0->nb[1] dataType:src0dt];
MPSMatrixDescriptor * desc1 = [MPSMatrixDescriptor
matrixDescriptorWithRows:ne11 columns:ne10 rowBytes:src1->nb[1] dataType:src1dt];
MPSMatrixDescriptor * desc = [MPSMatrixDescriptor
matrixDescriptorWithRows:ne1 columns:ne0 rowBytes:dst->nb[1] dataType:MPSDataTypeFloat32];
MPSMatrixMultiplication * mul = [[MPSMatrixMultiplication alloc]
initWithDevice:ctx->device transposeLeft:false transposeRight:true
resultRows:ne11 resultColumns:ne01 interiorColumns:ne00 alpha:1.0 beta:0.0];
// we need to do ne02 multiplications
// TODO: is there a way to do this in parallel - currently very slow ..
// TODO: might be possible to offload part of the computation to ANE using Accelerate's CBLAS
for (int64_t i02 = 0; i02 < ne02; ++i02) {
size_t offs_src0_cur = offs_src0 + i02*nb02;
size_t offs_src1_cur = offs_src1 + i02*nb12;
size_t offs_dst_cur = offs_dst + i02*nb2;
MPSMatrix * mat_src0 = [[MPSMatrix alloc] initWithBuffer:id_src0 offset:offs_src0_cur descriptor:desc0];
MPSMatrix * mat_src1 = [[MPSMatrix alloc] initWithBuffer:id_src1 offset:offs_src1_cur descriptor:desc1];
MPSMatrix * mat_dst = [[MPSMatrix alloc] initWithBuffer:id_dst offset:offs_dst_cur descriptor:desc ];
[mul encodeToCommandBuffer:command_buffer leftMatrix:mat_src1 rightMatrix:mat_src0 resultMatrix:mat_dst];
src1t == GGML_TYPE_F32 &&
[ctx->device supportsFamily:MTLGPUFamilyApple7] &&
ne00%32 == 0 &&
ne11 > 1) {
switch (src0->type) {
case GGML_TYPE_F16: [encoder setComputePipelineState:ctx->pipeline_mul_mm_f16_f32]; break;
case GGML_TYPE_Q4_0: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q4_0_f32]; break;
case GGML_TYPE_Q4_1: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q4_1_f32]; break;
case GGML_TYPE_Q8_0: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q8_0_f32]; break;
case GGML_TYPE_Q2_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q2_K_f32]; break;
case GGML_TYPE_Q3_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q3_K_f32]; break;
case GGML_TYPE_Q4_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q4_K_f32]; break;
case GGML_TYPE_Q5_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q5_K_f32]; break;
case GGML_TYPE_Q6_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q6_K_f32]; break;
default: GGML_ASSERT(false && "MUL MAT-MAT not implemented");
}
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
[encoder setBuffer:id_dst offset:offs_dst atIndex:2];
[encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
[encoder setBytes:&ne02 length:sizeof(ne02) atIndex:4];
[encoder setBytes:&nb01 length:sizeof(nb01) atIndex:5];
[encoder setBytes:&nb02 length:sizeof(nb02) atIndex:6];
[encoder setBytes:&ne12 length:sizeof(ne12) atIndex:7];
[encoder setBytes:&ne0 length:sizeof(ne0) atIndex:8];
[encoder setBytes:&ne1 length:sizeof(ne1) atIndex:9];
[encoder setBytes:&gqa length:sizeof(gqa) atIndex:10];
[encoder setThreadgroupMemoryLength:8192 atIndex:0];
[encoder dispatchThreadgroups:MTLSizeMake( (ne11+31)/32, (ne01+63) / 64, ne12) threadsPerThreadgroup:MTLSizeMake(128, 1, 1)];
} else {
if (encoder == nil) {
encoder = [command_buffer computeCommandEncoder];
}
int nth0 = 32;
int nth1 = 1;
@ -639,8 +845,6 @@ void ggml_metal_graph_compute(
switch (src0t) {
case GGML_TYPE_F16:
{
GGML_ASSERT(ne02 == ne12);
nth0 = 64;
nth1 = 1;
[encoder setComputePipelineState:ctx->pipeline_mul_mat_f16_f32];
@ -663,13 +867,22 @@ void ggml_metal_graph_compute(
nth1 = 8;
[encoder setComputePipelineState:ctx->pipeline_mul_mat_q4_1_f32];
} break;
case GGML_TYPE_Q8_0:
{
GGML_ASSERT(ne02 == 1);
GGML_ASSERT(ne12 == 1);
nth0 = 8;
nth1 = 8;
[encoder setComputePipelineState:ctx->pipeline_mul_mat_q8_0_f32];
} break;
case GGML_TYPE_Q2_K:
{
GGML_ASSERT(ne02 == 1);
GGML_ASSERT(ne12 == 1);
nth0 = 4;
nth1 = 16;
nth0 = 2;
nth1 = 32;
[encoder setComputePipelineState:ctx->pipeline_mul_mat_q2_K_f32];
} break;
case GGML_TYPE_Q3_K:
@ -677,8 +890,8 @@ void ggml_metal_graph_compute(
GGML_ASSERT(ne02 == 1);
GGML_ASSERT(ne12 == 1);
nth0 = 4;
nth1 = 16;
nth0 = 2;
nth1 = 32;
[encoder setComputePipelineState:ctx->pipeline_mul_mat_q3_K_f32];
} break;
case GGML_TYPE_Q4_K:
@ -686,8 +899,8 @@ void ggml_metal_graph_compute(
GGML_ASSERT(ne02 == 1);
GGML_ASSERT(ne12 == 1);
nth0 = 4;
nth1 = 16;
nth0 = 2;
nth1 = 32;
[encoder setComputePipelineState:ctx->pipeline_mul_mat_q4_K_f32];
} break;
case GGML_TYPE_Q5_K:
@ -695,8 +908,8 @@ void ggml_metal_graph_compute(
GGML_ASSERT(ne02 == 1);
GGML_ASSERT(ne12 == 1);
nth0 = 4;
nth1 = 16;
nth0 = 2;
nth1 = 32;
[encoder setComputePipelineState:ctx->pipeline_mul_mat_q5_K_f32];
} break;
case GGML_TYPE_Q6_K:
@ -704,8 +917,8 @@ void ggml_metal_graph_compute(
GGML_ASSERT(ne02 == 1);
GGML_ASSERT(ne12 == 1);
nth0 = 4;
nth1 = 16;
nth0 = 2;
nth1 = 32;
[encoder setComputePipelineState:ctx->pipeline_mul_mat_q6_K_f32];
} break;
default:
@ -720,28 +933,36 @@ void ggml_metal_graph_compute(
[encoder setBuffer:id_dst offset:offs_dst atIndex:2];
[encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
[encoder setBytes:&ne01 length:sizeof(ne01) atIndex:4];
[encoder setBytes:&nb00 length:sizeof(nb00) atIndex:5];
[encoder setBytes:&nb01 length:sizeof(nb01) atIndex:6];
[encoder setBytes:&nb02 length:sizeof(nb02) atIndex:7];
[encoder setBytes:&ne10 length:sizeof(ne10) atIndex:8];
[encoder setBytes:&ne11 length:sizeof(ne11) atIndex:9];
[encoder setBytes:&nb10 length:sizeof(nb10) atIndex:10];
[encoder setBytes:&nb11 length:sizeof(nb11) atIndex:11];
[encoder setBytes:&nb12 length:sizeof(nb12) atIndex:12];
[encoder setBytes:&ne0 length:sizeof(ne0) atIndex:13];
[encoder setBytes:&ne1 length:sizeof(ne1) atIndex:14];
[encoder setBytes:&ne02 length:sizeof(ne02) atIndex:5];
[encoder setBytes:&nb00 length:sizeof(nb00) atIndex:6];
[encoder setBytes:&nb01 length:sizeof(nb01) atIndex:7];
[encoder setBytes:&nb02 length:sizeof(nb02) atIndex:8];
[encoder setBytes:&ne10 length:sizeof(ne10) atIndex:9];
[encoder setBytes:&ne11 length:sizeof(ne11) atIndex:10];
[encoder setBytes:&ne12 length:sizeof(ne12) atIndex:11];
[encoder setBytes:&nb10 length:sizeof(nb10) atIndex:12];
[encoder setBytes:&nb11 length:sizeof(nb11) atIndex:13];
[encoder setBytes:&nb12 length:sizeof(nb12) atIndex:14];
[encoder setBytes:&ne0 length:sizeof(ne0) atIndex:15];
[encoder setBytes:&ne1 length:sizeof(ne1) atIndex:16];
[encoder setBytes:&gqa length:sizeof(gqa) atIndex:17];
if (src0t == GGML_TYPE_Q4_0 || src0t == GGML_TYPE_Q4_1) {
[encoder setThreadgroupMemoryLength:nth0*nth1*sizeof(float) atIndex:0];
[encoder dispatchThreadgroups:MTLSizeMake(ne01, ne11, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
if (src0t == GGML_TYPE_Q4_0 || src0t == GGML_TYPE_Q4_1 || src0t == GGML_TYPE_Q8_0 ||
src0t == GGML_TYPE_Q2_K || src0t == GGML_TYPE_Q4_K) {
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
}
else if (src0t == GGML_TYPE_Q2_K ||
src0t == GGML_TYPE_Q3_K ||
src0t == GGML_TYPE_Q4_K ||
src0t == GGML_TYPE_Q5_K ||
src0t == GGML_TYPE_Q6_K) {
[encoder setThreadgroupMemoryLength:nth0*nth1*sizeof(float) atIndex:0];
[encoder dispatchThreadgroups:MTLSizeMake(ne01, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
else if (src0t == GGML_TYPE_Q3_K) {
#ifdef GGML_QKK_64
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 1)/2, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
#else
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
#endif
}
else if (src0t == GGML_TYPE_Q5_K) {
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
}
else if (src0t == GGML_TYPE_Q6_K) {
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 1)/2, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
} else {
[encoder setThreadgroupMemoryLength:nth0*sizeof(float) atIndex:0];
[encoder dispatchThreadgroups:MTLSizeMake(ne01, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
@ -750,14 +971,11 @@ void ggml_metal_graph_compute(
} break;
case GGML_OP_GET_ROWS:
{
if (encoder == nil) {
encoder = [command_buffer computeCommandEncoder];
}
switch (src0->type) {
case GGML_TYPE_F16: [encoder setComputePipelineState:ctx->pipeline_get_rows_f16]; break;
case GGML_TYPE_Q4_0: [encoder setComputePipelineState:ctx->pipeline_get_rows_q4_0]; break;
case GGML_TYPE_Q4_1: [encoder setComputePipelineState:ctx->pipeline_get_rows_q4_1]; break;
case GGML_TYPE_Q8_0: [encoder setComputePipelineState:ctx->pipeline_get_rows_q8_0]; break;
case GGML_TYPE_Q2_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q2_K]; break;
case GGML_TYPE_Q3_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q3_K]; break;
case GGML_TYPE_Q4_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q4_K]; break;
@ -779,13 +997,10 @@ void ggml_metal_graph_compute(
} break;
case GGML_OP_RMS_NORM:
{
if (encoder == nil) {
encoder = [command_buffer computeCommandEncoder];
}
float eps;
memcpy(&eps, dst->op_params, sizeof(float));
const float eps = 1e-6f;
const int nth = 256;
const int nth = 512;
[encoder setComputePipelineState:ctx->pipeline_rms_norm];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
@ -793,7 +1008,7 @@ void ggml_metal_graph_compute(
[encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:2];
[encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:3];
[encoder setBytes:&eps length:sizeof( float) atIndex:4];
[encoder setThreadgroupMemoryLength:nth*sizeof(float) atIndex:0];
[encoder setThreadgroupMemoryLength:nth/32*sizeof(float) atIndex:0];
const int64_t nrows = ggml_nrows(src0);
@ -801,11 +1016,8 @@ void ggml_metal_graph_compute(
} break;
case GGML_OP_NORM:
{
if (encoder == nil) {
encoder = [command_buffer computeCommandEncoder];
}
const float eps = 1e-5f;
float eps;
memcpy(&eps, dst->op_params, sizeof(float));
const int nth = 256;
@ -823,15 +1035,12 @@ void ggml_metal_graph_compute(
} break;
case GGML_OP_ALIBI:
{
if (encoder == nil) {
encoder = [command_buffer computeCommandEncoder];
}
GGML_ASSERT((src0t == GGML_TYPE_F32));
const int n_past = ((int32_t *) src1->data)[0]; UNUSED(n_past);
const int n_head = ((int32_t *) src1->data)[1];
const float max_bias = ((float *) src1->data)[2];
const int n_past = ((int32_t *) dst->op_params)[0]; UNUSED(n_past);
const int n_head = ((int32_t *) dst->op_params)[1];
float max_bias;
memcpy(&max_bias, (int32_t *) dst->op_params + 2, sizeof(float));
if (__builtin_popcount(n_head) != 1) {
GGML_ASSERT(false && "only power-of-two n_head implemented");
@ -860,19 +1069,21 @@ void ggml_metal_graph_compute(
[encoder setBytes:&nb2 length:sizeof(uint64_t) atIndex:16];
[encoder setBytes:&nb3 length:sizeof(uint64_t) atIndex:17];
[encoder setBytes:&m0 length:sizeof( float) atIndex:18];
const int nth = 32;
[encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
} break;
case GGML_OP_ROPE:
{
if (encoder == nil) {
encoder = [command_buffer computeCommandEncoder];
}
const int n_past = ((int32_t *) dst->op_params)[0];
const int n_dims = ((int32_t *) dst->op_params)[1];
const int mode = ((int32_t *) dst->op_params)[2];
const int n_dims = ((int32_t *) src1->data)[1];
const int mode = ((int32_t *) src1->data)[2];
const int n_past = ((int32_t *)(src1->data))[0];
float freq_base;
float freq_scale;
memcpy(&freq_base, (int32_t *) dst->op_params + 4, sizeof(float));
memcpy(&freq_scale, (int32_t *) dst->op_params + 5, sizeof(float));
[encoder setComputePipelineState:ctx->pipeline_rope];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
@ -896,15 +1107,15 @@ void ggml_metal_graph_compute(
[encoder setBytes:&n_past length:sizeof( int) atIndex:18];
[encoder setBytes:&n_dims length:sizeof( int) atIndex:19];
[encoder setBytes:&mode length:sizeof( int) atIndex:20];
[encoder setBytes:&freq_base length:sizeof(float) atIndex:21];
[encoder setBytes:&freq_scale length:sizeof(float) atIndex:22];
[encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
} break;
case GGML_OP_DUP:
case GGML_OP_CPY:
case GGML_OP_CONT:
{
if (encoder == nil) {
encoder = [command_buffer computeCommandEncoder];
}
const int nth = 32;
switch (src0t) {
@ -949,10 +1160,12 @@ void ggml_metal_graph_compute(
[encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
} break;
default:
{
fprintf(stderr, "%s: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(dst->op));
GGML_ASSERT(false);
}
}
}
if (encoder != nil) {
[encoder endEncoding];
@ -964,17 +1177,19 @@ void ggml_metal_graph_compute(
}
// wait for all threads to finish
dispatch_barrier_sync(queue, ^{});
[command_buffers[n_cb - 1] waitUntilCompleted];
dispatch_barrier_sync(ctx->d_queue, ^{});
// check status of command buffers
// needed to detect if the device ran out-of-memory for example (#1881)
for (int i = 0; i < n_cb; i++) {
MTLCommandBufferStatus status = (MTLCommandBufferStatus) [command_buffers[i] status];
[ctx->command_buffers[i] waitUntilCompleted];
MTLCommandBufferStatus status = (MTLCommandBufferStatus) [ctx->command_buffers[i] status];
if (status != MTLCommandBufferStatusCompleted) {
fprintf(stderr, "%s: command buffer %d failed with status %lu\n", __func__, i, status);
GGML_ASSERT(false);
}
}
}
}

2098
ggml-metal.metal
View File

File diff suppressed because it is too large Load Diff

6
ggml-opencl.cpp
View File

@ -656,10 +656,14 @@ __kernel void dequantize_mul_mat_vec_q6_K(__global const struct block_q6_K * xx,
\n#if K_QUANTS_PER_ITERATION == 1\n
const int l0 = K_QUANTS_PER_ITERATION*in; // 0...15
const int is = 0;
\n#else\n
const int l0 = 4 * in; // 0, 4, 8, ..., 28
const int is = in / 4;
\n#endif\n
const int ql_offset = 64*im + l0;
const int qh_offset = 32*im + l0;
const int s_offset = 8*im + is;
@ -1376,7 +1380,7 @@ static void ggml_cl_mul_f32(const ggml_tensor * src0, const ggml_tensor * src1,
const int64_t ne00 = src0->ne[0];
const int64_t ne01 = src0->ne[1];
const int64_t ne02 = src0->ne[2];
const int64_t ne03 = src0->ne[2];
const int64_t ne03 = src0->ne[3];
const int64_t ne0 = ne00 * ne01 * ne02 * ne03;
const int64_t ne10 = src1->ne[0];
const int64_t ne11 = src1->ne[1];

6042
ggml.c
View File

File diff suppressed because it is too large Load Diff

606
ggml.h
View File

@ -65,7 +65,7 @@
// ggml_set_f32(a, 3.0f);
// ggml_set_f32(b, 4.0f);
//
// ggml_graph_compute(ctx0, &gf);
// ggml_graph_compute_with_ctx(ctx, &gf, n_threads);
//
// printf("f = %f\n", ggml_get_f32_1d(f, 0));
//
@ -130,13 +130,16 @@
// The data of the tensor is accessed via the "data" pointer. For example:
//
// {
// struct ggml_tensor * a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, 2, 3);
// const int nx = 2;
// const int ny = 3;
//
// // a[1, 2] = 1.0f;
// *(float *) ((char *) a->data + 2*a->nb[1] + 1*a->nb[0]) = 1.0f;
// struct ggml_tensor * a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, nx, ny);
//
// // a[2, 0] = 2.0f;
// *(float *) ((char *) a->data + 0*a->nb[1] + 2*a->nb[0]) = 2.0f;
// for (int y = 0; y < ny; y++) {
// for (int x = 0; x < nx; x++) {
// *(float *) ((char *) a->data + y*a->nb[1] + x*a->nb[0]) = x + y;
// }
// }
//
// ...
// }
@ -183,6 +186,15 @@
# define GGML_API
#endif
// TODO: support for clang
#ifdef __GNUC__
# define GGML_DEPRECATED(func, hint) func __attribute__((deprecated(hint)))
#elif defined(_MSC_VER)
# define GGML_DEPRECATED(func, hint) __declspec(deprecated(hint)) func
#else
# define GGML_DEPRECATED(func, hint) func
#endif
#include <stdint.h>
#include <stddef.h>
#include <stdbool.h>
@ -197,12 +209,29 @@
#define GGML_MAX_NODES 4096
#define GGML_MAX_PARAMS 256
#define GGML_MAX_CONTEXTS 64
#define GGML_MAX_OPT 4
#define GGML_MAX_NAME 48
#define GGML_MAX_SRC 6
#define GGML_MAX_NAME 64
#define GGML_MAX_OP_PARAMS 32
#define GGML_DEFAULT_N_THREADS 4
#if UINTPTR_MAX == 0xFFFFFFFF
#define GGML_MEM_ALIGN 4
#else
#define GGML_MEM_ALIGN 16
#endif
#define GGML_EXIT_SUCCESS 0
#define GGML_EXIT_ABORTED 1
#define GGUF_MAGIC 0x46554747 // "GGUF"
#define GGUF_VERSION 2
#define GGUF_DEFAULT_ALIGNMENT 32
#define GGML_UNUSED(x) (void)(x)
#define GGML_PAD(x, n) (((x) + (n) - 1) & ~((n) - 1))
#define GGML_ASSERT(x) \
do { \
if (!(x)) { \
@ -239,8 +268,9 @@
extern "C" {
#endif
#ifdef __ARM_NEON
// we use the built-in 16-bit float type
#if defined(__ARM_NEON) && defined(__CUDACC__)
typedef half ggml_fp16_t;
#elif defined(__ARM_NEON)
typedef __fp16 ggml_fp16_t;
#else
typedef uint16_t ggml_fp16_t;
@ -250,8 +280,8 @@ extern "C" {
GGML_API float ggml_fp16_to_fp32(ggml_fp16_t x);
GGML_API ggml_fp16_t ggml_fp32_to_fp16(float x);
GGML_API void ggml_fp16_to_fp32_row(const ggml_fp16_t * x, float * y, size_t n);
GGML_API void ggml_fp32_to_fp16_row(const float * x, ggml_fp16_t * y, size_t n);
GGML_API void ggml_fp16_to_fp32_row(const ggml_fp16_t * x, float * y, int n);
GGML_API void ggml_fp32_to_fp16_row(const float * x, ggml_fp16_t * y, int n);
struct ggml_object;
struct ggml_context;
@ -324,20 +354,12 @@ extern "C" {
GGML_OP_ARGMAX,
GGML_OP_REPEAT,
GGML_OP_REPEAT_BACK,
GGML_OP_ABS,
GGML_OP_SGN,
GGML_OP_NEG,
GGML_OP_STEP,
GGML_OP_TANH,
GGML_OP_ELU,
GGML_OP_RELU,
GGML_OP_GELU,
GGML_OP_GELU_QUICK,
GGML_OP_SILU,
GGML_OP_CONCAT,
GGML_OP_SILU_BACK,
GGML_OP_NORM, // normalize
GGML_OP_RMS_NORM,
GGML_OP_RMS_NORM_BACK,
GGML_OP_GROUP_NORM,
GGML_OP_MUL_MAT,
GGML_OP_OUT_PROD,
@ -363,16 +385,29 @@ extern "C" {
GGML_OP_CLAMP,
GGML_OP_CONV_1D,
GGML_OP_CONV_2D,
GGML_OP_CONV_TRANSPOSE_2D,
GGML_OP_POOL_1D,
GGML_OP_POOL_2D,
GGML_OP_UPSCALE, // nearest interpolate
GGML_OP_FLASH_ATTN,
GGML_OP_FLASH_FF,
GGML_OP_FLASH_ATTN_BACK,
GGML_OP_WIN_PART,
GGML_OP_WIN_UNPART,
GGML_OP_GET_REL_POS,
GGML_OP_ADD_REL_POS,
GGML_OP_UNARY,
GGML_OP_MAP_UNARY,
GGML_OP_MAP_BINARY,
GGML_OP_MAP_CUSTOM1_F32,
GGML_OP_MAP_CUSTOM2_F32,
GGML_OP_MAP_CUSTOM3_F32,
GGML_OP_MAP_CUSTOM1,
GGML_OP_MAP_CUSTOM2,
GGML_OP_MAP_CUSTOM3,
@ -383,6 +418,24 @@ extern "C" {
GGML_OP_COUNT,
};
enum ggml_unary_op {
GGML_UNARY_OP_ABS,
GGML_UNARY_OP_SGN,
GGML_UNARY_OP_NEG,
GGML_UNARY_OP_STEP,
GGML_UNARY_OP_TANH,
GGML_UNARY_OP_ELU,
GGML_UNARY_OP_RELU,
GGML_UNARY_OP_GELU,
GGML_UNARY_OP_GELU_QUICK,
GGML_UNARY_OP_SILU,
};
enum ggml_object_type {
GGML_OBJECT_TENSOR,
GGML_OBJECT_GRAPH,
GGML_OBJECT_WORK_BUFFER
};
// ggml object
struct ggml_object {
@ -391,7 +444,9 @@ extern "C" {
struct ggml_object * next;
char padding[8];
enum ggml_object_type type;
char padding[4];
};
static const size_t GGML_OBJECT_SIZE = sizeof(struct ggml_object);
@ -411,15 +466,13 @@ extern "C" {
// compute data
enum ggml_op op;
// op params - allocated as int32_t for alignment
int32_t op_params[GGML_MAX_OP_PARAMS / sizeof(int32_t)];
bool is_param;
struct ggml_tensor * grad;
struct ggml_tensor * src0;
struct ggml_tensor * src1;
struct ggml_tensor * opt[GGML_MAX_OPT];
// thread scheduling
int n_tasks;
struct ggml_tensor * src[GGML_MAX_SRC];
// performance
int perf_runs;
@ -437,25 +490,46 @@ extern "C" {
static const size_t GGML_TENSOR_SIZE = sizeof(struct ggml_tensor);
// the compute plan that needs to be prepared for ggml_graph_compute()
// since https://github.com/ggerganov/ggml/issues/287
struct ggml_cplan {
size_t work_size; // size of work buffer, calculated by `ggml_graph_plan()`
uint8_t * work_data; // work buffer, to be allocated by caller before calling to `ggml_graph_compute()`
int n_threads;
// the `n_tasks` of nodes, 1:1 mapping to cgraph nodes
int n_tasks[GGML_MAX_NODES];
// abort ggml_graph_compute when true
bool (*abort_callback)(void * data);
void * abort_callback_data;
};
// next prime after GGML_MAX_NODES
// #define GGML_GRAPH_HASHTABLE_SIZE 4099
// next prime after GGML_MAX_NODES * 2 (nodes + leafs)
#define GGML_GRAPH_HASHTABLE_SIZE 8273
// computation graph
struct ggml_cgraph {
int n_nodes;
int n_leafs;
int n_threads;
size_t work_size;
struct ggml_tensor * work;
struct ggml_tensor * nodes[GGML_MAX_NODES];
struct ggml_tensor * grads[GGML_MAX_NODES];
struct ggml_tensor * leafs[GGML_MAX_NODES];
void * visited_hash_table[GGML_GRAPH_HASHTABLE_SIZE];
// performance
int perf_runs;
int64_t perf_cycles;
int64_t perf_time_us;
};
static const size_t GGML_GRAPH_SIZE = sizeof(struct ggml_cgraph);
// scratch buffer
struct ggml_scratch {
size_t offs;
@ -509,6 +583,7 @@ extern "C" {
GGML_API int64_t ggml_nelements (const struct ggml_tensor * tensor);
GGML_API int64_t ggml_nrows (const struct ggml_tensor * tensor);
GGML_API size_t ggml_nbytes (const struct ggml_tensor * tensor);
GGML_API size_t ggml_nbytes_pad (const struct ggml_tensor * tensor); // same as ggml_nbytes() but padded to GGML_MEM_ALIGN
GGML_API size_t ggml_nbytes_split(const struct ggml_tensor * tensor, int nrows_split);
GGML_API int ggml_blck_size (enum ggml_type type);
@ -517,6 +592,7 @@ extern "C" {
GGML_API const char * ggml_type_name(enum ggml_type type);
GGML_API const char * ggml_op_name (enum ggml_op op);
GGML_API const char * ggml_op_symbol(enum ggml_op op);
GGML_API size_t ggml_element_size(const struct ggml_tensor * tensor);
@ -529,6 +605,8 @@ extern "C" {
GGML_API bool ggml_is_contiguous(const struct ggml_tensor * tensor);
GGML_API bool ggml_is_permuted (const struct ggml_tensor * tensor);
GGML_API bool ggml_are_same_shape(const struct ggml_tensor * t0, const struct ggml_tensor * t1);
// use this to compute the memory overhead of a tensor
GGML_API size_t ggml_tensor_overhead(void);
@ -540,6 +618,7 @@ extern "C" {
GGML_API size_t ggml_used_mem(const struct ggml_context * ctx);
GGML_API size_t ggml_set_scratch (struct ggml_context * ctx, struct ggml_scratch scratch);
GGML_API bool ggml_get_no_alloc(struct ggml_context * ctx);
GGML_API void ggml_set_no_alloc(struct ggml_context * ctx, bool no_alloc);
GGML_API void * ggml_get_mem_buffer (const struct ggml_context * ctx);
@ -599,6 +678,8 @@ extern "C" {
GGML_API void * ggml_get_data (const struct ggml_tensor * tensor);
GGML_API float * ggml_get_data_f32(const struct ggml_tensor * tensor);
GGML_API enum ggml_unary_op ggml_get_unary_op(const struct ggml_tensor * tensor);
GGML_API const char * ggml_get_name (const struct ggml_tensor * tensor);
GGML_API struct ggml_tensor * ggml_set_name ( struct ggml_tensor * tensor, const char * name);
GGML_API struct ggml_tensor * ggml_format_name( struct ggml_tensor * tensor, const char * fmt, ...);
@ -611,6 +692,11 @@ extern "C" {
struct ggml_context * ctx,
struct ggml_tensor * a);
// in-place, returns view(a)
GGML_API struct ggml_tensor * ggml_dup_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a);
GGML_API struct ggml_tensor * ggml_add(
struct ggml_context * ctx,
struct ggml_tensor * a,
@ -735,6 +821,13 @@ extern "C" {
struct ggml_tensor * a,
struct ggml_tensor * b);
// concat a and b on dim 2
// used in stable-diffusion
GGML_API struct ggml_tensor * ggml_concat(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b);
GGML_API struct ggml_tensor * ggml_abs(
struct ggml_context * ctx,
struct ggml_tensor * a);
@ -824,25 +917,42 @@ extern "C" {
struct ggml_tensor * b);
// normalize along rows
// TODO: eps is hardcoded to 1e-5 for now
GGML_API struct ggml_tensor * ggml_norm(
struct ggml_context * ctx,
struct ggml_tensor * a);
struct ggml_tensor * a,
float eps);
GGML_API struct ggml_tensor * ggml_norm_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a);
struct ggml_tensor * a,
float eps);
GGML_API struct ggml_tensor * ggml_rms_norm(
struct ggml_context * ctx,
struct ggml_tensor * a);
struct ggml_tensor * a,
float eps);
GGML_API struct ggml_tensor * ggml_rms_norm_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a);
struct ggml_tensor * a,
float eps);
// group normalize along ne0*ne1*n_groups
// used in stable-diffusion
// TODO: eps is hardcoded to 1e-6 for now
GGML_API struct ggml_tensor * ggml_group_norm(
struct ggml_context * ctx,
struct ggml_tensor * a,
int n_groups);
GGML_API struct ggml_tensor * ggml_group_norm_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a,
int n_groups);
// a - x
// b - dy
// TODO: update with configurable eps
GGML_API struct ggml_tensor * ggml_rms_norm_back(
struct ggml_context * ctx,
struct ggml_tensor * a,
@ -934,11 +1044,22 @@ extern "C" {
struct ggml_tensor * a,
struct ggml_tensor * b);
// a -> b, in-place, return view(b)
GGML_API struct ggml_tensor * ggml_cpy_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b);
// make contiguous
GGML_API struct ggml_tensor * ggml_cont(
struct ggml_context * ctx,
struct ggml_tensor * a);
// make contiguous, in-place
GGML_API struct ggml_tensor * ggml_cont_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a);
// return view(a), b specifies the new shape
// TODO: when we start computing gradient, make a copy instead of view
GGML_API struct ggml_tensor * ggml_reshape(
@ -1107,6 +1228,37 @@ extern "C" {
int mode,
int n_ctx);
// custom RoPE
GGML_API struct ggml_tensor * ggml_rope_custom(
struct ggml_context * ctx,
struct ggml_tensor * a,
int n_past,
int n_dims,
int mode,
int n_ctx,
float freq_base,
float freq_scale);
// in-place, returns view(a)
GGML_API struct ggml_tensor * ggml_rope_custom_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a,
int n_past,
int n_dims,
int mode,
int n_ctx,
float freq_base,
float freq_scale);
// xPos RoPE, in-place, returns view(a)
GGML_API struct ggml_tensor * ggml_rope_xpos_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a,
int n_past,
int n_dims,
float base,
bool down);
// rotary position embedding backward, i.e compute dx from dy
// a - dy
GGML_API struct ggml_tensor * ggml_rope_back(
@ -1114,7 +1266,12 @@ extern "C" {
struct ggml_tensor * a,
int n_past,
int n_dims,
int mode);
int mode,
int n_ctx,
float freq_base,
float freq_scale,
float xpos_base,
bool xpos_down);
// alibi position embedding
// in-place, returns view(a)
@ -1141,6 +1298,15 @@ extern "C" {
int p0, // padding
int d0); // dilation
// conv_1d with padding = half
// alias for ggml_conv_1d(a, b, s, a->ne[0]/2, d)
GGML_API struct ggml_tensor* ggml_conv_1d_ph(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
int s,
int d);
GGML_API struct ggml_tensor * ggml_conv_2d(
struct ggml_context * ctx,
struct ggml_tensor * a,
@ -1152,14 +1318,70 @@ extern "C" {
int d0,
int d1);
// conv_1d with padding = half
// alias for ggml_conv_1d(a, b, s, a->ne[0]/2, d)
GGML_API struct ggml_tensor* ggml_conv_1d_ph(
// kernel size is a->ne[0] x a->ne[1]
// stride is equal to kernel size
// padding is zero
// example:
// a: 16 16 3 768
// b: 1024 1024 3 1
// res: 64 64 768 1
// used in sam
GGML_API struct ggml_tensor * ggml_conv_2d_sk_p0(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b);
// kernel size is a->ne[0] x a->ne[1]
// stride is 1
// padding is half
// example:
// a: 3 3 256 256
// b: 64 64 256 1
// res: 64 64 256 1
// used in sam
GGML_API struct ggml_tensor * ggml_conv_2d_s1_ph(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b);
GGML_API struct ggml_tensor * ggml_conv_transpose_2d_p0(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
int s,
int d);
int stride);
enum ggml_op_pool {
GGML_OP_POOL_MAX,
GGML_OP_POOL_AVG,
GGML_OP_POOL_COUNT,
};
GGML_API struct ggml_tensor * ggml_pool_1d(
struct ggml_context * ctx,
struct ggml_tensor * a,
enum ggml_op_pool op,
int k0, // kernel size
int s0, // stride
int p0); // padding
GGML_API struct ggml_tensor * ggml_pool_2d(
struct ggml_context * ctx,
struct ggml_tensor * a,
enum ggml_op_pool op,
int k0,
int k1,
int s0,
int s1,
int p0,
int p1);
// nearest interpolate
// used in stable-diffusion
GGML_API struct ggml_tensor * ggml_upscale(
struct ggml_context * ctx,
struct ggml_tensor * a,
int scale_factor);
GGML_API struct ggml_tensor * ggml_flash_attn(
struct ggml_context * ctx,
@ -1204,6 +1426,37 @@ extern "C" {
int h0,
int w);
GGML_API struct ggml_tensor * ggml_unary(
struct ggml_context * ctx,
struct ggml_tensor * a,
enum ggml_unary_op op);
GGML_API struct ggml_tensor * ggml_unary_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a,
enum ggml_unary_op op);
// used in sam
GGML_API struct ggml_tensor * ggml_get_rel_pos(
struct ggml_context * ctx,
struct ggml_tensor * a,
int qh,
int kh);
// used in sam
GGML_API struct ggml_tensor * ggml_add_rel_pos(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * pw,
struct ggml_tensor * ph);
GGML_API struct ggml_tensor * ggml_add_rel_pos_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * pw,
struct ggml_tensor * ph);
// custom operators
typedef void (*ggml_unary_op_f32_t) (const int, float *, const float *);
@ -1213,63 +1466,129 @@ extern "C" {
typedef void (*ggml_custom2_op_f32_t)(struct ggml_tensor *, const struct ggml_tensor *, const struct ggml_tensor *);
typedef void (*ggml_custom3_op_f32_t)(struct ggml_tensor *, const struct ggml_tensor *, const struct ggml_tensor *, const struct ggml_tensor *);
GGML_API struct ggml_tensor * ggml_map_unary_f32(
GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_unary_f32(
struct ggml_context * ctx,
struct ggml_tensor * a,
ggml_unary_op_f32_t fun);
ggml_unary_op_f32_t fun),
"use ggml_map_custom1 instead");
GGML_API struct ggml_tensor * ggml_map_unary_inplace_f32(
GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_unary_inplace_f32(
struct ggml_context * ctx,
struct ggml_tensor * a,
ggml_unary_op_f32_t fun);
ggml_unary_op_f32_t fun),
"use ggml_map_custom1_inplace instead");
GGML_API struct ggml_tensor * ggml_map_binary_f32(
GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_binary_f32(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
ggml_binary_op_f32_t fun);
ggml_binary_op_f32_t fun),
"use ggml_map_custom2 instead");
GGML_API struct ggml_tensor * ggml_map_binary_inplace_f32(
GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_binary_inplace_f32(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
ggml_binary_op_f32_t fun);
ggml_binary_op_f32_t fun),
"use ggml_map_custom2_inplace instead");
GGML_API struct ggml_tensor * ggml_map_custom1_f32(
GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom1_f32(
struct ggml_context * ctx,
struct ggml_tensor * a,
ggml_custom1_op_f32_t fun);
ggml_custom1_op_f32_t fun),
"use ggml_map_custom1 instead");
GGML_API struct ggml_tensor * ggml_map_custom1_inplace_f32(
GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom1_inplace_f32(
struct ggml_context * ctx,
struct ggml_tensor * a,
ggml_custom1_op_f32_t fun);
ggml_custom1_op_f32_t fun),
"use ggml_map_custom1_inplace instead");
GGML_API struct ggml_tensor * ggml_map_custom2_f32(
GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom2_f32(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
ggml_custom2_op_f32_t fun);
ggml_custom2_op_f32_t fun),
"use ggml_map_custom2 instead");
GGML_API struct ggml_tensor * ggml_map_custom2_inplace_f32(
GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom2_inplace_f32(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
ggml_custom2_op_f32_t fun);
ggml_custom2_op_f32_t fun),
"use ggml_map_custom2_inplace instead");
GGML_API struct ggml_tensor * ggml_map_custom3_f32(
GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom3_f32(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
struct ggml_tensor * c,
ggml_custom3_op_f32_t fun);
ggml_custom3_op_f32_t fun),
"use ggml_map_custom3 instead");
GGML_API struct ggml_tensor * ggml_map_custom3_inplace_f32(
GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom3_inplace_f32(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
struct ggml_tensor * c,
ggml_custom3_op_f32_t fun);
ggml_custom3_op_f32_t fun),
"use ggml_map_custom3_inplace instead");
// custom operators v2
typedef void (*ggml_custom1_op_t)(struct ggml_tensor * dst , const struct ggml_tensor * a, int ith, int nth, void * userdata);
typedef void (*ggml_custom2_op_t)(struct ggml_tensor * dst , const struct ggml_tensor * a, const struct ggml_tensor * b, int ith, int nth, void * userdata);
typedef void (*ggml_custom3_op_t)(struct ggml_tensor * dst , const struct ggml_tensor * a, const struct ggml_tensor * b, const struct ggml_tensor * c, int ith, int nth, void * userdata);
#define GGML_N_TASKS_MAX -1
GGML_API struct ggml_tensor * ggml_map_custom1(
struct ggml_context * ctx,
struct ggml_tensor * a,
ggml_custom1_op_t fun,
int n_tasks,
void * userdata);
GGML_API struct ggml_tensor * ggml_map_custom1_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a,
ggml_custom1_op_t fun,
int n_tasks,
void * userdata);
GGML_API struct ggml_tensor * ggml_map_custom2(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
ggml_custom2_op_t fun,
int n_tasks,
void * userdata);
GGML_API struct ggml_tensor * ggml_map_custom2_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
ggml_custom2_op_t fun,
int n_tasks,
void * userdata);
GGML_API struct ggml_tensor * ggml_map_custom3(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
struct ggml_tensor * c,
ggml_custom3_op_t fun,
int n_tasks,
void * userdata);
GGML_API struct ggml_tensor * ggml_map_custom3_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
struct ggml_tensor * c,
ggml_custom3_op_t fun,
int n_tasks,
void * userdata);
// loss function
@ -1292,14 +1611,27 @@ extern "C" {
struct ggml_context * ctx,
struct ggml_tensor * tensor);
GGML_API void ggml_build_forward_expand(struct ggml_cgraph * cgraph, struct ggml_tensor * tensor);
GGML_API struct ggml_cgraph ggml_build_forward (struct ggml_tensor * tensor);
GGML_API struct ggml_cgraph ggml_build_backward(struct ggml_context * ctx, struct ggml_cgraph * gf, bool keep);
GGML_API void ggml_graph_compute(struct ggml_context * ctx, struct ggml_cgraph * cgraph);
// graph allocation in a context
GGML_API struct ggml_cgraph * ggml_new_graph (struct ggml_context * ctx);
GGML_API struct ggml_cgraph * ggml_build_forward_ctx(struct ggml_context * ctx, struct ggml_tensor * tensor);
GGML_API size_t ggml_graph_overhead(void);
// ggml_graph_plan() has to be called before ggml_graph_compute()
// when plan.work_size > 0, caller must allocate memory for plan.work_data
GGML_API struct ggml_cplan ggml_graph_plan (struct ggml_cgraph * cgraph, int n_threads /*= GGML_DEFAULT_N_THREADS*/);
GGML_API int ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cplan * cplan);
GGML_API void ggml_graph_reset (struct ggml_cgraph * cgraph);
// same as ggml_graph_compute() but the work data is allocated as a part of the context
// note: the drawback of this API is that you must have ensured that the context has enough memory for the work data
GGML_API void ggml_graph_compute_with_ctx(struct ggml_context * ctx, struct ggml_cgraph * cgraph, int n_threads);
GGML_API struct ggml_tensor * ggml_graph_get_tensor(struct ggml_cgraph * cgraph, const char * name);
GGML_API void ggml_graph_export(const struct ggml_cgraph * cgraph, const char * fname);
@ -1488,6 +1820,127 @@ extern "C" {
GGML_API size_t ggml_quantize_chunk(enum ggml_type type, const float * src, void * dst, int start, int n, int64_t * hist);
//
// gguf
//
enum gguf_type {
GGUF_TYPE_UINT8 = 0,
GGUF_TYPE_INT8 = 1,
GGUF_TYPE_UINT16 = 2,
GGUF_TYPE_INT16 = 3,
GGUF_TYPE_UINT32 = 4,
GGUF_TYPE_INT32 = 5,
GGUF_TYPE_FLOAT32 = 6,
GGUF_TYPE_BOOL = 7,
GGUF_TYPE_STRING = 8,
GGUF_TYPE_ARRAY = 9,
GGUF_TYPE_UINT64 = 10,
GGUF_TYPE_INT64 = 11,
GGUF_TYPE_FLOAT64 = 12,
GGUF_TYPE_COUNT, // marks the end of the enum
};
struct gguf_context;
struct gguf_init_params {
bool no_alloc;
// if not NULL, create a ggml_context and allocate the tensor data in it
struct ggml_context ** ctx;
};
GGML_API struct gguf_context * gguf_init_empty(void);
GGML_API struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_params params);
//GGML_API struct gguf_context * gguf_init_from_buffer(..);
GGML_API void gguf_free(struct gguf_context * ctx);
GGML_API const char * gguf_type_name(enum gguf_type type);
GGML_API int gguf_get_version (struct gguf_context * ctx);
GGML_API size_t gguf_get_alignment (struct gguf_context * ctx);
GGML_API size_t gguf_get_data_offset(struct gguf_context * ctx);
GGML_API void * gguf_get_data (struct gguf_context * ctx);
GGML_API int gguf_get_n_kv(struct gguf_context * ctx);
GGML_API int gguf_find_key(struct gguf_context * ctx, const char * key);
GGML_API const char * gguf_get_key (struct gguf_context * ctx, int i);
GGML_API enum gguf_type gguf_get_kv_type (struct gguf_context * ctx, int i);
GGML_API enum gguf_type gguf_get_arr_type(struct gguf_context * ctx, int i);
// results are undefined if the wrong type is used for the key
GGML_API uint8_t gguf_get_val_u8 (struct gguf_context * ctx, int i);
GGML_API int8_t gguf_get_val_i8 (struct gguf_context * ctx, int i);
GGML_API uint16_t gguf_get_val_u16 (struct gguf_context * ctx, int i);
GGML_API int16_t gguf_get_val_i16 (struct gguf_context * ctx, int i);
GGML_API uint32_t gguf_get_val_u32 (struct gguf_context * ctx, int i);
GGML_API int32_t gguf_get_val_i32 (struct gguf_context * ctx, int i);
GGML_API float gguf_get_val_f32 (struct gguf_context * ctx, int i);
GGML_API uint64_t gguf_get_val_u64 (struct gguf_context * ctx, int i);
GGML_API int64_t gguf_get_val_i64 (struct gguf_context * ctx, int i);
GGML_API double gguf_get_val_f64 (struct gguf_context * ctx, int i);
GGML_API bool gguf_get_val_bool(struct gguf_context * ctx, int i);
GGML_API const char * gguf_get_val_str (struct gguf_context * ctx, int i);
GGML_API int gguf_get_arr_n (struct gguf_context * ctx, int i);
GGML_API const void * gguf_get_arr_data(struct gguf_context * ctx, int i);
GGML_API const char * gguf_get_arr_str (struct gguf_context * ctx, int key_id, int i);
GGML_API int gguf_get_n_tensors (struct gguf_context * ctx);
GGML_API int gguf_find_tensor (struct gguf_context * ctx, const char * name);
GGML_API size_t gguf_get_tensor_offset(struct gguf_context * ctx, int i);
GGML_API char * gguf_get_tensor_name (struct gguf_context * ctx, int i);
// overrides existing values or adds a new one
GGML_API void gguf_set_val_u8 (struct gguf_context * ctx, const char * key, uint8_t val);
GGML_API void gguf_set_val_i8 (struct gguf_context * ctx, const char * key, int8_t val);
GGML_API void gguf_set_val_u16 (struct gguf_context * ctx, const char * key, uint16_t val);
GGML_API void gguf_set_val_i16 (struct gguf_context * ctx, const char * key, int16_t val);
GGML_API void gguf_set_val_u32 (struct gguf_context * ctx, const char * key, uint32_t val);
GGML_API void gguf_set_val_i32 (struct gguf_context * ctx, const char * key, int32_t val);
GGML_API void gguf_set_val_f32 (struct gguf_context * ctx, const char * key, float val);
GGML_API void gguf_set_val_u64 (struct gguf_context * ctx, const char * key, uint64_t val);
GGML_API void gguf_set_val_i64 (struct gguf_context * ctx, const char * key, int64_t val);
GGML_API void gguf_set_val_f64 (struct gguf_context * ctx, const char * key, double val);
GGML_API void gguf_set_val_bool(struct gguf_context * ctx, const char * key, bool val);
GGML_API void gguf_set_val_str (struct gguf_context * ctx, const char * key, const char * val);
GGML_API void gguf_set_arr_data(struct gguf_context * ctx, const char * key, enum gguf_type type, const void * data, int n);
GGML_API void gguf_set_arr_str (struct gguf_context * ctx, const char * key, const char ** data, int n);
// set or add KV pairs from another context
GGML_API void gguf_set_kv(struct gguf_context * ctx, struct gguf_context * src);
// manage tensor info
GGML_API void gguf_add_tensor(struct gguf_context * ctx, const struct ggml_tensor * tensor);
GGML_API void gguf_set_tensor_type(struct gguf_context * ctx, const char * name, enum ggml_type type);
GGML_API void gguf_set_tensor_data(struct gguf_context * ctx, const char * name, const void * data, size_t size);
// writing gguf files can be done in 2 ways:
//
// - write the entire gguf_context to a binary file in a single pass:
//
// gguf_write_to_file(ctx, fname);
//
// - first prepare a file with a placeholder for the meta data, write the tensor data, then write the meta data:
//
// FILE * f = fopen(fname, "wb");
// fseek(f, gguf_get_meta_size(ctx), SEEK_SET);
// fwrite(f, ...);
// void * data = gguf_meta_get_meta_data(ctx);
// fseek(f, 0, SEEK_SET);
// fwrite(f, data, gguf_get_meta_size(ctx));
// free(data);
// fclose(f);
//
// write the entire context to a binary file
GGML_API void gguf_write_to_file(struct gguf_context * ctx, const char * fname, bool only_meta);
// get the size in bytes of the meta data (header, kv pairs, tensor info) including padding
GGML_API size_t gguf_get_meta_size(struct gguf_context * ctx);
GGML_API void gguf_get_meta_data(struct gguf_context * ctx, void * data);
//
// system info
//
@ -1521,20 +1974,23 @@ extern "C" {
#else
#define GGML_RESTRICT restrict
#endif
typedef void (*dequantize_row_q_t)(const void * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
typedef void (*quantize_row_q_t) (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k);
typedef void (*vec_dot_q_t) (const int n, float * GGML_RESTRICT s, const void * GGML_RESTRICT x, const void * GGML_RESTRICT y);
typedef void (*ggml_to_float_t) (const void * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
typedef void (*ggml_from_float_t)(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k);
typedef void (*ggml_vec_dot_t) (const int n, float * GGML_RESTRICT s, const void * GGML_RESTRICT x, const void * GGML_RESTRICT y);
typedef struct {
dequantize_row_q_t dequantize_row_q;
quantize_row_q_t quantize_row_q;
quantize_row_q_t quantize_row_q_reference;
quantize_row_q_t quantize_row_q_dot;
vec_dot_q_t vec_dot_q;
const char * type_name;
int blck_size;
size_t type_size;
bool is_quantized;
ggml_to_float_t to_float;
ggml_from_float_t from_float;
ggml_from_float_t from_float_reference;
ggml_vec_dot_t vec_dot;
enum ggml_type vec_dot_type;
} quantize_fns_t;
} ggml_type_traits_t;
quantize_fns_t ggml_internal_get_quantize_fn(size_t i);
ggml_type_traits_t ggml_internal_get_type_traits(enum ggml_type type);
#ifdef __cplusplus
}

34
whisper.cpp
View File

@ -441,6 +441,7 @@ struct whisper_hparams {
int32_t n_text_layer = 4;
int32_t n_mels = 80;
int32_t ftype = 1;
float eps = 1e-5f;
};
// audio encoding layer
@ -1578,7 +1579,7 @@ static bool whisper_encode_internal(
{
wstate.use_buf(ctx0, 0);
cur = ggml_norm(ctx0, inpL);
cur = ggml_norm(ctx0, inpL, hparams.eps);
// cur = ln_0_w*cur + ln_0_b
cur = ggml_add(ctx0,
@ -1725,7 +1726,7 @@ static bool whisper_encode_internal(
{
wstate.use_buf(ctx0, 0);
cur = ggml_norm(ctx0, inpFF);
cur = ggml_norm(ctx0, inpFF, hparams.eps);
wstate.use_buf(ctx0, 1);
@ -1788,7 +1789,7 @@ static bool whisper_encode_internal(
{
wstate.use_buf(ctx0, 0);
cur = ggml_norm(ctx0, cur);
cur = ggml_norm(ctx0, cur, hparams.eps);
wstate.use_buf(ctx0, 1);
@ -1805,10 +1806,9 @@ static bool whisper_encode_internal(
// run the computation
{
struct ggml_cgraph gf = {};
gf.n_threads = n_threads;
ggml_build_forward_expand (&gf, cur);
ggml_graph_compute(ctx0, &gf);
ggml_graph_compute_with_ctx(ctx0, &gf, n_threads);
//ggml_graph_print(&gf);
}
@ -1851,12 +1851,11 @@ static bool whisper_encode_internal(
// pre-compute cross-attention memory
{
struct ggml_cgraph gf = {};
gf.n_threads = n_threads;
// TODO: hack to disconnect the encoded features from the previous graph
cur->op = GGML_OP_NONE;
cur->src0 = nullptr;
cur->src1 = nullptr;
cur->src[0] = nullptr;
cur->src[1] = nullptr;
for (int il = 0; il < model.hparams.n_text_layer; ++il) {
auto& layer = model.layers_decoder[il];
@ -1894,7 +1893,7 @@ static bool whisper_encode_internal(
ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Vcross, v));
}
ggml_graph_compute(ctx0, &gf);
ggml_graph_compute_with_ctx(ctx0, &gf, n_threads);
//ggml_graph_print(&gf);
}
@ -1965,7 +1964,6 @@ static bool whisper_decode_internal(
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, tokens, N*ggml_element_size(embd));
@ -1992,7 +1990,7 @@ static bool whisper_decode_internal(
{
wstate.use_buf(ctx0, 0);
cur = ggml_norm(ctx0, inpL);
cur = ggml_norm(ctx0, inpL, hparams.eps);
// cur = ln_0_w*cur + ln_0_b
cur = ggml_add(ctx0,
@ -2119,7 +2117,7 @@ static bool whisper_decode_internal(
{
wstate.use_buf(ctx0, 0);
cur = ggml_norm(ctx0, inpCA); // note: we use inpCA here
cur = ggml_norm(ctx0, inpCA, hparams.eps); // note: we use inpCA here
// cur = ln_0_w*cur + ln_0_b
cur = ggml_add(ctx0,
@ -2229,7 +2227,7 @@ static bool whisper_decode_internal(
{
wstate.use_buf(ctx0, 0);
cur = ggml_norm(ctx0, inpFF);
cur = ggml_norm(ctx0, inpFF, hparams.eps);
wstate.use_buf(ctx0, 1);
@ -2284,7 +2282,7 @@ static bool whisper_decode_internal(
{
wstate.use_buf(ctx0, 0);
cur = ggml_norm(ctx0, cur);
cur = ggml_norm(ctx0, cur, hparams.eps);
wstate.use_buf(ctx0, 1);
@ -2309,7 +2307,7 @@ static bool whisper_decode_internal(
// run the computation
{
ggml_build_forward_expand (&gf, logits);
ggml_graph_compute (ctx0, &gf);
ggml_graph_compute_with_ctx(ctx0, &gf, n_threads);
}
// extract logits for all N tokens
@ -5165,17 +5163,15 @@ WHISPER_API const char * whisper_bench_ggml_mul_mat_str(int n_threads) {
struct ggml_cgraph gf = ggml_build_forward(c);
gf.n_threads = n_threads;
double tsum = 0.0;
// heat-up
ggml_graph_compute(ctx0, &gf);
ggml_graph_compute_with_ctx(ctx0, &gf, n_threads);
for (int i = 0; i < n_max; ++i) {
const int64_t t0 = ggml_time_us();
ggml_graph_compute(ctx0, &gf);
ggml_graph_compute_with_ctx(ctx0, &gf, n_threads);
const int64_t t1 = ggml_time_us();