ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-02-23 22:54:10 +00:00
Code Issues Packages Projects Releases Wiki Activity

llama: factor out model loader

Browse Source
This commit is contained in:
Iwan Kawrakow
2025-08-13 12:12:46 +03:00
parent 2ac615507f
commit 6b6d25bfbf
9 changed files with 2309 additions and 2088 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
src/CMakeLists.txt
View File

@@ -17,6 +17,8 @@ add_library(llama
llama-vocab.cpp
llama-grammar.cpp
llama-sampling.cpp
llama-mmap.cpp
llama-model-loader.cpp
unicode.h
unicode.cpp
unicode-data.cpp

287
src/llama-arch.h Normal file
View File

@@ -0,0 +1,287 @@
#pragma once
#include <string>
enum llm_arch {
LLM_ARCH_LLAMA,
LLM_ARCH_LLAMA4,
LLM_ARCH_DECI,
LLM_ARCH_FALCON,
LLM_ARCH_BAICHUAN,
LLM_ARCH_GROK,
LLM_ARCH_GPT2,
LLM_ARCH_GPTJ,
LLM_ARCH_GPTNEOX,
LLM_ARCH_MPT,
LLM_ARCH_STARCODER,
LLM_ARCH_REFACT,
LLM_ARCH_BERT,
LLM_ARCH_NOMIC_BERT,
LLM_ARCH_JINA_BERT_V2,
LLM_ARCH_BLOOM,
LLM_ARCH_STABLELM,
LLM_ARCH_QWEN,
LLM_ARCH_QWEN2,
LLM_ARCH_QWEN2MOE,
LLM_ARCH_QWEN3,
LLM_ARCH_QWEN3MOE,
LLM_ARCH_PHI2,
LLM_ARCH_PHI3,
LLM_ARCH_PLAMO,
LLM_ARCH_CODESHELL,
LLM_ARCH_ORION,
LLM_ARCH_INTERNLM2,
LLM_ARCH_MINICPM,
LLM_ARCH_GEMMA,
LLM_ARCH_GEMMA2,
LLM_ARCH_GEMMA3,
LLM_ARCH_STARCODER2,
LLM_ARCH_MAMBA,
LLM_ARCH_XVERSE,
LLM_ARCH_COMMAND_R,
LLM_ARCH_DBRX,
LLM_ARCH_OLMO,
LLM_ARCH_OPENELM,
LLM_ARCH_ARCTIC,
LLM_ARCH_DEEPSEEK2,
LLM_ARCH_CHATGLM,
LLM_ARCH_GLM4,
LLM_ARCH_GLM4_MOE,
LLM_ARCH_BITNET,
LLM_ARCH_BITNET_25,
LLM_ARCH_BITNET_B158,
LLM_ARCH_T5,
LLM_ARCH_T5ENCODER,
LLM_ARCH_JAIS,
LLM_ARCH_GRANITE,
LLM_ARCH_GRANITE_MOE,
LLM_ARCH_COHERE2,
LLM_ARCH_DOTS1,
LLM_ARCH_HUNYUAN_MOE,
LLM_ARCH_OPENAI_MOE,
LLM_ARCH_UNKNOWN,
};
enum llm_kv {
LLM_KV_GENERAL_TYPE,
LLM_KV_GENERAL_ARCHITECTURE,
LLM_KV_GENERAL_QUANTIZATION_VERSION,
LLM_KV_GENERAL_ALIGNMENT,
LLM_KV_GENERAL_NAME,
LLM_KV_GENERAL_AUTHOR,
LLM_KV_GENERAL_VERSION,
LLM_KV_GENERAL_URL,
LLM_KV_GENERAL_DESCRIPTION,
LLM_KV_GENERAL_LICENSE,
LLM_KV_GENERAL_SOURCE_URL,
LLM_KV_GENERAL_SOURCE_HF_REPO,
LLM_KV_VOCAB_SIZE,
LLM_KV_CONTEXT_LENGTH,
LLM_KV_EMBEDDING_LENGTH,
LLM_KV_BLOCK_COUNT,
LLM_KV_LEADING_DENSE_BLOCK_COUNT,
LLM_KV_FEED_FORWARD_LENGTH,
LLM_KV_EXPERT_FEED_FORWARD_LENGTH,
LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH,
LLM_KV_USE_PARALLEL_RESIDUAL,
LLM_KV_TENSOR_DATA_LAYOUT,
LLM_KV_EXPERT_COUNT,
LLM_KV_EXPERT_USED_COUNT,
LLM_KV_EXPERT_SHARED_COUNT,
LLM_KV_EXPERT_WEIGHTS_SCALE,
LLM_KV_EXPERT_WEIGHTS_NORM,
LLM_KV_EXPERT_GATING_FUNC,
LLM_KV_NEXTN_PREDICT_LAYERS,
LLM_KV_POOLING_TYPE,
LLM_KV_LOGIT_SCALE,
LLM_KV_DECODER_START_TOKEN_ID,
LLM_KV_ATTN_LOGIT_SOFTCAPPING,
LLM_KV_FINAL_LOGIT_SOFTCAPPING,
LLM_KV_SWIN_NORM,
LLM_KV_RESCALE_EVERY_N_LAYERS,
LLM_KV_TIME_MIX_EXTRA_DIM,
LLM_KV_TIME_DECAY_EXTRA_DIM,
LLM_KV_RESIDUAL_SCALE,
LLM_KV_EMBEDDING_SCALE,
LLM_KV_TOKEN_SHIFT_COUNT,
LLM_KV_INTERLEAVE_MOE_LAYER_STEP,
LLM_KV_ATTENTION_HEAD_COUNT,
LLM_KV_ATTENTION_HEAD_COUNT_KV,
LLM_KV_ATTENTION_MAX_ALIBI_BIAS,
LLM_KV_ATTENTION_CLAMP_KQV,
LLM_KV_ATTENTION_KEY_LENGTH,
LLM_KV_ATTENTION_VALUE_LENGTH,
LLM_KV_ATTENTION_LAYERNORM_EPS,
LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,
LLM_KV_ATTENTION_CAUSAL,
LLM_KV_ATTENTION_Q_LORA_RANK,
LLM_KV_ATTENTION_KV_LORA_RANK,
LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT,
LLM_KV_ATTENTION_SLIDING_WINDOW,
LLM_KV_ATTENTION_SCALE,
LLM_KV_ROPE_DIMENSION_COUNT,
LLM_KV_ROPE_FREQ_BASE,
LLM_KV_ROPE_SCALE_LINEAR,
LLM_KV_ROPE_SCALING_TYPE,
LLM_KV_ROPE_SCALING_FACTOR,
LLM_KV_ROPE_SCALING_ATTN_FACTOR,
LLM_KV_ROPE_SCALING_ORIG_CTX_LEN,
LLM_KV_ROPE_SCALING_FINETUNED,
LLM_KV_ROPE_SCALING_YARN_LOG_MUL,
LLM_KV_SPLIT_NO,
LLM_KV_SPLIT_COUNT,
LLM_KV_SPLIT_TENSORS_COUNT,
LLM_KV_SSM_INNER_SIZE,
LLM_KV_SSM_CONV_KERNEL,
LLM_KV_SSM_STATE_SIZE,
LLM_KV_SSM_TIME_STEP_RANK,
LLM_KV_TOKENIZER_MODEL,
LLM_KV_TOKENIZER_PRE,
LLM_KV_TOKENIZER_LIST,
LLM_KV_TOKENIZER_TOKEN_TYPE,
LLM_KV_TOKENIZER_TOKEN_TYPE_COUNT,
LLM_KV_TOKENIZER_SCORES,
LLM_KV_TOKENIZER_MERGES,
LLM_KV_TOKENIZER_BOS_ID,
LLM_KV_TOKENIZER_EOS_ID,
LLM_KV_TOKENIZER_UNK_ID,
LLM_KV_TOKENIZER_SEP_ID,
LLM_KV_TOKENIZER_PAD_ID,
LLM_KV_TOKENIZER_CLS_ID,
LLM_KV_TOKENIZER_MASK_ID,
LLM_KV_TOKENIZER_ADD_BOS,
LLM_KV_TOKENIZER_ADD_EOS,
LLM_KV_TOKENIZER_ADD_PREFIX,
LLM_KV_TOKENIZER_REMOVE_EXTRA_WS,
LLM_KV_TOKENIZER_PRECOMPILED_CHARSMAP,
LLM_KV_TOKENIZER_HF_JSON,
LLM_KV_TOKENIZER_RWKV,
LLM_KV_TOKENIZER_CHAT_TEMPLATE,
LLM_KV_TOKENIZER_CHAT_TEMPLATE_N,
LLM_KV_TOKENIZER_FIM_PRE_ID,
LLM_KV_TOKENIZER_FIM_SUF_ID,
LLM_KV_TOKENIZER_FIM_MID_ID,
LLM_KV_TOKENIZER_FIM_PAD_ID,
LLM_KV_TOKENIZER_FIM_REP_ID,
LLM_KV_TOKENIZER_FIM_SEP_ID,
LLM_KV_TOKENIZER_PREFIX_ID,
LLM_KV_TOKENIZER_SUFFIX_ID,
LLM_KV_TOKENIZER_MIDDLE_ID,
LLM_KV_TOKENIZER_EOT_ID,
LLM_KV_TOKENIZER_EOM_ID,
LLM_KV_ADAPTER_TYPE,
LLM_KV_ADAPTER_LORA_ALPHA,
};
struct LLM_KV {
LLM_KV(llm_arch arch, const char* suffix = nullptr);
llm_arch arch;
const char* suffix;
std::string operator()(llm_kv kv) const;
};
enum llm_tensor {
LLM_TENSOR_TOKEN_EMBD,
LLM_TENSOR_TOKEN_EMBD_NORM,
LLM_TENSOR_TOKEN_TYPES,
LLM_TENSOR_POS_EMBD,
LLM_TENSOR_OUTPUT,
LLM_TENSOR_OUTPUT_NORM,
LLM_TENSOR_ROPE_FREQS,
LLM_TENSOR_ROPE_FACTORS_LONG,
LLM_TENSOR_ROPE_FACTORS_SHORT,
LLM_TENSOR_ATTN_Q,
LLM_TENSOR_ATTN_K,
LLM_TENSOR_ATTN_V,
LLM_TENSOR_ATTN_QKV,
LLM_TENSOR_ATTN_OUT,
LLM_TENSOR_ATTN_NORM,
LLM_TENSOR_ATTN_NORM_2,
LLM_TENSOR_ATTN_OUT_NORM,
LLM_TENSOR_ATTN_POST_NORM,
LLM_TENSOR_ATTN_ROT_EMBD,
LLM_TENSOR_ATTN_SINKS,
LLM_TENSOR_FFN_GATE_INP,
LLM_TENSOR_FFN_GATE_INP_SHEXP,
LLM_TENSOR_FFN_NORM,
LLM_TENSOR_FFN_POST_NORM,
LLM_TENSOR_FFN_GATE,
LLM_TENSOR_FFN_DOWN,
LLM_TENSOR_FFN_UP,
LLM_TENSOR_FFN_ACT,
LLM_TENSOR_FFN_DOWN_EXP, // split experts for backward compatibility
LLM_TENSOR_FFN_GATE_EXP,
LLM_TENSOR_FFN_UP_EXP,
LLM_TENSOR_FFN_NORM_EXPS,
LLM_TENSOR_FFN_DOWN_EXPS, // merged experts
LLM_TENSOR_FFN_GATE_EXPS,
LLM_TENSOR_FFN_UP_EXPS,
LLM_TENSOR_FFN_DOWN_SHEXP,
LLM_TENSOR_FFN_GATE_SHEXP,
LLM_TENSOR_FFN_UP_SHEXP,
LLM_TENSOR_FFN_EXP_PROBS_B,
LLM_TENSOR_ATTN_Q_NORM,
LLM_TENSOR_ATTN_K_NORM,
LLM_TENSOR_LAYER_OUT_NORM,
LLM_TENSOR_SSM_IN,
LLM_TENSOR_SSM_CONV1D,
LLM_TENSOR_SSM_X,
LLM_TENSOR_SSM_DT,
LLM_TENSOR_SSM_A,
LLM_TENSOR_SSM_D,
LLM_TENSOR_SSM_OUT,
LLM_TENSOR_ATTN_Q_A,
LLM_TENSOR_ATTN_Q_B,
LLM_TENSOR_ATTN_KV_A_MQA,
LLM_TENSOR_ATTN_KV_B,
LLM_TENSOR_ATTN_K_B,
LLM_TENSOR_ATTN_V_B,
LLM_TENSOR_ATTN_Q_A_NORM,
LLM_TENSOR_ATTN_KV_A_NORM,
LLM_TENSOR_ATTN_SUB_NORM,
LLM_TENSOR_FFN_SUB_NORM,
LLM_TENSOR_DEC_ATTN_NORM,
LLM_TENSOR_DEC_ATTN_Q,
LLM_TENSOR_DEC_ATTN_K,
LLM_TENSOR_DEC_ATTN_V,
LLM_TENSOR_DEC_ATTN_OUT,
LLM_TENSOR_DEC_ATTN_REL_B,
LLM_TENSOR_DEC_CROSS_ATTN_NORM,
LLM_TENSOR_DEC_CROSS_ATTN_Q,
LLM_TENSOR_DEC_CROSS_ATTN_K,
LLM_TENSOR_DEC_CROSS_ATTN_V,
LLM_TENSOR_DEC_CROSS_ATTN_OUT,
LLM_TENSOR_DEC_CROSS_ATTN_REL_B,
LLM_TENSOR_DEC_FFN_NORM,
LLM_TENSOR_DEC_FFN_GATE,
LLM_TENSOR_DEC_FFN_DOWN,
LLM_TENSOR_DEC_FFN_UP,
LLM_TENSOR_DEC_OUTPUT_NORM,
LLM_TENSOR_ENC_ATTN_NORM,
LLM_TENSOR_ENC_ATTN_Q,
LLM_TENSOR_ENC_ATTN_K,
LLM_TENSOR_ENC_ATTN_V,
LLM_TENSOR_ENC_ATTN_OUT,
LLM_TENSOR_ENC_ATTN_REL_B,
LLM_TENSOR_ENC_FFN_NORM,
LLM_TENSOR_ENC_FFN_GATE,
LLM_TENSOR_ENC_FFN_DOWN,
LLM_TENSOR_ENC_FFN_UP,
LLM_TENSOR_ENC_OUTPUT_NORM,
LLM_TENSOR_NEXTN_EH_PROJ,
LLM_TENSOR_NEXTN_EMBED_TOKENS,
LLM_TENSOR_NEXTN_ENORM,
LLM_TENSOR_NEXTN_HNORM,
LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD,
LLM_TENSOR_NEXTN_SHARED_HEAD_NORM,
};
llm_arch llm_arch_from_string(const std::string & name);

50
src/llama-impl.h
View File

@@ -10,6 +10,11 @@
#define LLAMA_API_INTERNAL
#include "llama.h"
#include <stdexcept>
#include <climits>
#include <cstdarg>
#include <vector>
#include <cinttypes>
#include <cstring>
#ifdef __GNUC__
#ifdef __MINGW32__
@@ -166,3 +171,48 @@ struct ring_buffer {
size_t pos = 0;
std::vector<T> data;
};
LLAMA_ATTRIBUTE_FORMAT(1, 2)
static std::string format(const char * fmt, ...) {
va_list ap;
va_list ap2;
va_start(ap, fmt);
va_copy(ap2, ap);
int size = vsnprintf(NULL, 0, fmt, ap);
GGML_ASSERT(size >= 0 && size < INT_MAX); // NOLINT
std::vector<char> buf(size + 1);
int size2 = vsnprintf(buf.data(), size + 1, fmt, ap2);
GGML_ASSERT(size2 == size);
va_end(ap2);
va_end(ap);
return std::string(buf.data(), size);
}
static std::string llama_format_tensor_shape(const std::vector<int64_t> & ne) {
char buf[256];
snprintf(buf, sizeof(buf), "%5" PRId64, ne.at(0));
for (size_t i = 1; i < ne.size(); i++) {
snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf), ", %5" PRId64, ne.at(i));
}
return buf;
}
static std::string llama_format_tensor_shape(const struct ggml_tensor * t) {
char buf[256];
snprintf(buf, sizeof(buf), "%5" PRId64, t->ne[0]);
for (int i = 1; i < GGML_MAX_DIMS; i++) {
snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf), ", %5" PRId64, t->ne[i]);
}
return buf;
}
template <typename T>
struct no_init {
T value;
no_init() { /* do nothing */ }
};
struct gguf_context;
std::string gguf_kv_to_str(const gguf_context * ctx_gguf, int i);

557
src/llama-mmap.cpp Normal file
View File

@@ -0,0 +1,557 @@
#include "llama-mmap.h"
#include "llama-impl.h"
#include "ggml.h"
#include <stdexcept>
#include <cstring>
#if defined(_WIN32)
static std::string llama_format_win_err(DWORD err) {
LPSTR buf;
size_t size = FormatMessageA(FORMAT_MESSAGE_ALLOCATE_BUFFER | FORMAT_MESSAGE_FROM_SYSTEM | FORMAT_MESSAGE_IGNORE_INSERTS,
NULL, err, MAKELANGID(LANG_NEUTRAL, SUBLANG_DEFAULT), (LPSTR)&buf, 0, NULL);
if (!size) {
return "FormatMessageA failed";
}
std::string ret(buf, size);
LocalFree(buf);
return ret;
}
std::string llama_file::GetErrorMessageWin32(DWORD error_code) const {
std::string ret;
LPSTR lpMsgBuf = NULL;
DWORD bufLen = FormatMessageA(FORMAT_MESSAGE_ALLOCATE_BUFFER | FORMAT_MESSAGE_FROM_SYSTEM | FORMAT_MESSAGE_IGNORE_INSERTS,
NULL, error_code, MAKELANGID(LANG_NEUTRAL, SUBLANG_DEFAULT), (LPSTR)&lpMsgBuf, 0, NULL);
if (!bufLen) {
ret = format("Win32 error code: %s", error_code);
} else {
ret = lpMsgBuf;
LocalFree(lpMsgBuf);
}
return ret;
}
llama_file::llama_file(const char * fname, const char * mode) {
fp = ggml_fopen(fname, mode);
if (fp == NULL) {
throw std::runtime_error(format("failed to open %s: %s", fname, strerror(errno)));
}
fp_win32 = (HANDLE) _get_osfhandle(_fileno(fp));
seek(0, SEEK_END);
size = tell();
seek(0, SEEK_SET);
}
size_t llama_file::tell() const {
// SetFilePointerEx returns the current position when seeking relative 0 bytes
LARGE_INTEGER li;
li.QuadPart = 0;
BOOL ret = SetFilePointerEx(fp_win32, li, &li, FILE_CURRENT);
if (!ret) {
throw std::runtime_error(format("read error: %s", GetErrorMessageWin32(GetLastError()).c_str()));
}
return li.QuadPart;
}
void llama_file::seek(size_t offset, int whence) const {
// no need to convert SEEK_* to FILE_*. The enums are the same.
// Still, keep static asserts to avoid failures in the future.
static_assert(SEEK_SET == FILE_BEGIN, "SEEK_SET != FILE_BEGIN");
static_assert(SEEK_CUR == FILE_CURRENT, "SEEK_CUR != FILE_CURRENT");
static_assert(SEEK_END == FILE_END, "SEEK_END != FILE_END");
LARGE_INTEGER li;
li.QuadPart = offset;
BOOL ret = SetFilePointerEx(fp_win32, li, NULL, whence);
if (!ret) {
throw std::runtime_error(format("read error: %s", GetErrorMessageWin32(GetLastError()).c_str()));
}
}
void llama_file::read_raw(void * ptr, size_t len) const {
// On Win32 ReadFile is significant faster than fread which is again significant faster than std::fstream. Thus
// use the Win32 API to do file io instead of the C/C++ library functions.
// There are conditions under which ReadFile cannot read chunks >64MB.
// Thus split the operation into smaller chunks if len exceeds this limit.
size_t bytes_read = 0;
while (bytes_read < len) {
size_t chunk_size = std::min<size_t>(len - bytes_read, 64*1024*1024);
DWORD chunk_read = 0;
BOOL result = ReadFile(fp_win32, reinterpret_cast<char*>(ptr) + bytes_read, chunk_size, &chunk_read, NULL);
if (!result) {
throw std::runtime_error(format("read error: %s", GetErrorMessageWin32(GetLastError()).c_str()));
}
if (chunk_read < chunk_size || chunk_read == 0) {
throw std::runtime_error("unexpectedly reached end of file");
}
bytes_read += chunk_read;
} ;
}
void llama_file::write_raw(const void * ptr, size_t len) const {
// There are conditions under which WriteFile cannot write chunks >64MB.
// Thus split the operation into smaller chunks if len exceeds this limit.
size_t bytes_written = 0;
while (bytes_written < len) {
size_t chunk_size = std::min<size_t>(len - bytes_written, 64*1024*1024);
DWORD chunk_written = 0;
BOOL result = WriteFile(fp_win32, reinterpret_cast<char const*>(ptr) + bytes_written, chunk_size, &chunk_written, NULL);
if (!result) {
throw std::runtime_error(format("write error: %s", GetErrorMessageWin32(GetLastError()).c_str()));
}
if (chunk_written < chunk_size || chunk_written == 0) {
throw std::runtime_error("unexpectedly failed to write bytes");
}
bytes_written += chunk_written;
}
}
llama_file::~llama_file() {
if (fp) {
std::fclose(fp);
}
}
#else
llama_file::llama_file(const char * fname, const char * mode) {
fp = ggml_fopen(fname, mode);
if (fp == NULL) {
throw std::runtime_error(format("failed to open %s: %s", fname, strerror(errno)));
}
seek(0, SEEK_END);
size = tell();
seek(0, SEEK_SET);
}
size_t llama_file::tell() const {
#ifdef _WIN32
__int64 ret = _ftelli64(fp);
#else
long ret = std::ftell(fp);
#endif
if (ret == -1) {
throw std::runtime_error(format("ftell error: %s", strerror(errno)));
}
return (size_t) ret;
}
void llama_file::seek(size_t offset, int whence) const {
#ifdef _WIN32
int ret = _fseeki64(fp, (__int64) offset, whence);
#else
int ret = std::fseek(fp, (long) offset, whence);
#endif
if (ret != 0) {
throw std::runtime_error(format("seek error: %s", strerror(errno)));
}
}
void llama_file::read_raw(void * ptr, size_t len) const {
if (len == 0) {
return;
}
errno = 0;
std::size_t ret = std::fread(ptr, len, 1, fp);
if (ferror(fp)) {
throw std::runtime_error(format("read error: %s", strerror(errno)));
}
if (ret != 1) {
throw std::runtime_error("unexpectedly reached end of file");
}
}
void llama_file::write_raw(const void * ptr, size_t len) const {
if (len == 0) {
return;
}
errno = 0;
size_t ret = std::fwrite(ptr, len, 1, fp);
if (ret != 1) {
throw std::runtime_error(format("write error: %s", strerror(errno)));
}
}
llama_file::~llama_file() {
if (fp) {
std::fclose(fp);
}
}
#endif
using llama_files = std::vector<std::unique_ptr<llama_file>>;
#ifdef _POSIX_MAPPED_FILES
llama_mmap::llama_mmap(struct llama_file * file, size_t prefetch, bool numa, [[maybe_unused]] bool use_thp) {
size = file->size;
int fd = fileno(file->fp);
int flags = MAP_SHARED;
// prefetch/readahead impairs performance on NUMA systems
if (numa) { prefetch = 0; }
#ifdef __linux__
// advise the kernel to read the file sequentially (increases readahead)
if (posix_fadvise(fd, 0, 0, POSIX_FADV_SEQUENTIAL)) {
LLAMA_LOG_WARN("warning: posix_fadvise(.., POSIX_FADV_SEQUENTIAL) failed: %s\n",
strerror(errno));
}
if (prefetch) { flags |= MAP_POPULATE; }
if (use_thp) {
size_t huge = get_default_huge_page_size();
auto size = huge*((file->size + huge - 1)/huge);
addr = mmap(nullptr, size, PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANONYMOUS | MAP_HUGETLB, -1, 0);
if (addr != MAP_FAILED) {
printf("%s: using THP with page size %zu MiB ", __func__, huge/(1024*1024));
fflush(stdout);
size_t tot = 0;
while (tot < file->size) {
auto n_read = pread(fd, static_cast<char*>(addr) + tot, file->size - tot, tot);
if (n_read < 0) throw std::runtime_error(format("Reading into mapped huge pages failed at %zu (%s)", tot, strerror(errno)));
printf("."); fflush(stdout);
tot += n_read;
}
printf(" done\n");
mapped_fragments.emplace_back(0, file->size);
mapped_page_size = huge;
return;
}
else {
fprintf(stderr, "%s: mmap with huge page size %zu MiB failed (%s)\n", __func__, huge/(1024*1024), strerror(errno));
}
}
#endif
addr = mmap(NULL, file->size, PROT_READ, flags, fd, 0);
if (addr == MAP_FAILED) { // NOLINT
throw std::runtime_error(format("mmap failed: %s", strerror(errno)));
}
if (prefetch > 0) {
// advise the kernel to preload the mapped memory
if (posix_madvise(addr, std::min(file->size, prefetch), POSIX_MADV_WILLNEED)) {
LLAMA_LOG_WARN("warning: posix_madvise(.., POSIX_MADV_WILLNEED) failed: %s\n",
strerror(errno));
}
}
if (numa) {
// advise the kernel not to use readahead
// (because the next page might not belong on the same node)
if (posix_madvise(addr, file->size, POSIX_MADV_RANDOM)) {
LLAMA_LOG_WARN("warning: posix_madvise(.., POSIX_MADV_RANDOM) failed: %s\n",
strerror(errno));
}
}
// initialize list of mapped_fragments
mapped_fragments.emplace_back(0, file->size);
}
static void llama_mmap::align_range(size_t * first, size_t * last, size_t page_size) {
// align first to the next page
size_t offset_in_page = *first & (page_size - 1);
size_t offset_to_page = offset_in_page == 0 ? 0 : page_size - offset_in_page;
*first += offset_to_page;
// align last to the previous page
*last = *last & ~(page_size - 1);
if (*last <= *first) {
*last = *first;
}
}
// partially unmap the file in the range [first, last)
void llama_mmap::unmap_fragment(size_t first, size_t last) {
// note: this function must not be called multiple times with overlapping ranges
// otherwise, there is a risk of invalidating addresses that have been repurposed for other mappings
int page_size = mapped_page_size > 0 ? mapped_page_size : sysconf(_SC_PAGESIZE);
align_range(&first, &last, page_size);
size_t len = last - first;
if (len == 0) {
return;
}
GGML_ASSERT(first % page_size == 0);
GGML_ASSERT(last % page_size == 0);
GGML_ASSERT(last > first);
void * next_page_start = (uint8_t *) addr + first;
// unmap the range
if (munmap(next_page_start, len)) {
LLAMA_LOG_WARN("warning: munmap failed: %s\n", strerror(errno));
}
// update the list of mapped fragments to avoid unmapping the same range again in the destructor
std::vector<std::pair<size_t, size_t>> new_mapped_fragments;
for (const auto & frag : mapped_fragments) {
if (frag.first < first && frag.second > last) {
// the range is in the middle of the fragment, split it
new_mapped_fragments.emplace_back(frag.first, first);
new_mapped_fragments.emplace_back(last, frag.second);
} else if (frag.first < first && frag.second > first) {
// the range starts in the middle of the fragment
new_mapped_fragments.emplace_back(frag.first, first);
} else if (frag.first < last && frag.second > last) {
// the range ends in the middle of the fragment
new_mapped_fragments.emplace_back(last, frag.second);
} else if (frag.first >= first && frag.second <= last) {
// the range covers the entire fragment
} else {
// the range is outside the fragment
new_mapped_fragments.push_back(frag);
}
}
mapped_fragments = std::move(new_mapped_fragments);
}
#ifdef __linux__
static int llama_mmap::get_default_huge_page_size() {
int pg_size = 2048;
std::ifstream in("/proc/meminfo");
if (in) {
std::string line;
while (true) {
std::getline(in, line);
if (in.fail()) break;
if (auto pos = line.find("Hugepagesize:"); pos != std::string::npos) {
std::istringstream str(line.data() + pos + 13);
int aux;
str >> aux;
if (!str.fail()) pg_size = aux;
break;
}
}
}
return pg_size * 1024;
}
#endif
llama_mmap::~llama_mmap() {
for (const auto & frag : mapped_fragments) {
if (munmap((char *) addr + frag.first, frag.second - frag.first)) {
LLAMA_LOG_WARN("warning: munmap failed: %s\n", strerror(errno));
}
}
}
#elif defined(_WIN32)
llama_mmap::llama_mmap(struct llama_file * file, size_t prefetch, bool numa, [[maybe_unused]] bool use_thp) {
GGML_UNUSED(numa);
size = file->size;
HANDLE hFile = (HANDLE) _get_osfhandle(_fileno(file->fp));
HANDLE hMapping = CreateFileMappingA(hFile, NULL, PAGE_READONLY, 0, 0, NULL);
if (hMapping == NULL) {
DWORD error = GetLastError();
throw std::runtime_error(format("CreateFileMappingA failed: %s", llama_format_win_err(error).c_str()));
}
addr = MapViewOfFile(hMapping, FILE_MAP_READ, 0, 0, 0);
DWORD error = GetLastError();
CloseHandle(hMapping);
if (addr == NULL) {
throw std::runtime_error(format("MapViewOfFile failed: %s", llama_format_win_err(error).c_str()));
}
if (prefetch > 0) {
#if _WIN32_WINNT >= 0x602
// PrefetchVirtualMemory is only present on Windows 8 and above, so we dynamically load it
BOOL (WINAPI *pPrefetchVirtualMemory) (HANDLE, ULONG_PTR, PWIN32_MEMORY_RANGE_ENTRY, ULONG);
HMODULE hKernel32 = GetModuleHandleW(L"kernel32.dll");
// may fail on pre-Windows 8 systems
pPrefetchVirtualMemory = reinterpret_cast<decltype(pPrefetchVirtualMemory)> (GetProcAddress(hKernel32, "PrefetchVirtualMemory"));
if (pPrefetchVirtualMemory) {
// advise the kernel to preload the mapped memory
WIN32_MEMORY_RANGE_ENTRY range;
range.VirtualAddress = addr;
range.NumberOfBytes = (SIZE_T) std::min(size, prefetch);
if (!pPrefetchVirtualMemory(GetCurrentProcess(), 1, &range, 0)) {
LLAMA_LOG_WARN("warning: PrefetchVirtualMemory failed: %s\n",
llama_format_win_err(GetLastError()).c_str());
}
}
#else
throw std::runtime_error("PrefetchVirtualMemory unavailable");
#endif
}
}
void llama_mmap::unmap_fragment(size_t first, size_t last) {
// not supported
GGML_UNUSED(first);
GGML_UNUSED(last);
}
llama_mmap::~llama_mmap() {
if (!UnmapViewOfFile(addr)) {
LLAMA_LOG_WARN("warning: UnmapViewOfFile failed: %s\n",
llama_format_win_err(GetLastError()).c_str());
}
}
#else
llama_mmap::llama_mmap(struct llama_file * file, size_t prefetch, bool numa, bool use_thp) {
GGML_UNUSED(file);
GGML_UNUSED(prefetch);
GGML_UNUSED(numa);
GGML_UNUSED(use_thp);
throw std::runtime_error("mmap not supported");
}
void llama_mmap::unmap_fragment(size_t first, size_t last) {
GGML_UNUSED(first);
GGML_UNUSED(last);
throw std::runtime_error("mmap not supported");
}
#endif
using llama_mmaps = std::vector<std::unique_ptr<llama_mmap>>;
void llama_mlock::init(void * ptr) {
GGML_ASSERT(addr == NULL && size == 0); // NOLINT
addr = ptr;
}
void llama_mlock::grow_to(size_t target_size) {
GGML_ASSERT(addr);
if (failed_already) {
return;
}
size_t granularity = lock_granularity();
target_size = (target_size + granularity - 1) & ~(granularity - 1);
if (target_size > size) {
if (raw_lock((uint8_t *) addr + size, target_size - size)) {
size = target_size;
} else {
failed_already = true;
}
}
}
#ifdef _POSIX_MEMLOCK_RANGE
static size_t llama_m::lock_granularity() {
return (size_t) sysconf(_SC_PAGESIZE);
}
#ifdef __APPLE__
#define MLOCK_SUGGESTION \
"Try increasing the sysctl values 'vm.user_wire_limit' and 'vm.global_user_wire_limit' and/or " \
"decreasing 'vm.global_no_user_wire_amount'. Also try increasing RLIMIT_MEMLOCK (ulimit -l).\n"
#else
#define MLOCK_SUGGESTION \
"Try increasing RLIMIT_MEMLOCK ('ulimit -l' as root).\n"
#endif
bool llama_mlock::raw_lock(const void * addr, size_t size) const {
if (!mlock(addr, size)) {
return true;
}
char* errmsg = std::strerror(errno);
bool suggest = (errno == ENOMEM);
// Check if the resource limit is fine after all
struct rlimit lock_limit;
if (suggest && getrlimit(RLIMIT_MEMLOCK, &lock_limit)) {
suggest = false;
}
if (suggest && (lock_limit.rlim_max > lock_limit.rlim_cur + size)) {
suggest = false;
}
LLAMA_LOG_WARN("warning: failed to mlock %zu-byte buffer (after previously locking %zu bytes): %s\n%s",
size, this->size, errmsg, suggest ? MLOCK_SUGGESTION : "");
return false;
}
#undef MLOCK_SUGGESTION
void llama_mlock::raw_unlock(void * addr, size_t size) {
if (munlock(addr, size)) {
LLAMA_LOG_WARN("warning: failed to munlock buffer: %s\n", std::strerror(errno));
}
}
#elif defined(_WIN32)
size_t llama_mlock::lock_granularity() {
SYSTEM_INFO si;
GetSystemInfo(&si);
return (size_t) si.dwPageSize;
}
bool llama_mlock::raw_lock(void * ptr, size_t len) const {
for (int tries = 1; ; tries++) {
if (VirtualLock(ptr, len)) {
return true;
}
if (tries == 2) {
LLAMA_LOG_WARN("warning: failed to VirtualLock %zu-byte buffer (after previously locking %zu bytes): %s\n",
len, size, llama_format_win_err(GetLastError()).c_str());
return false;
}
// It failed but this was only the first try; increase the working
// set size and try again.
SIZE_T min_ws_size, max_ws_size;
if (!GetProcessWorkingSetSize(GetCurrentProcess(), &min_ws_size, &max_ws_size)) {
LLAMA_LOG_WARN("warning: GetProcessWorkingSetSize failed: %s\n",
llama_format_win_err(GetLastError()).c_str());
return false;
}
// Per MSDN: "The maximum number of pages that a process can lock
// is equal to the number of pages in its minimum working set minus
// a small overhead."
// Hopefully a megabyte is enough overhead:
size_t increment = len + 1048576;
// The minimum must be <= the maximum, so we need to increase both:
min_ws_size += increment;
max_ws_size += increment;
if (!SetProcessWorkingSetSize(GetCurrentProcess(), min_ws_size, max_ws_size)) {
LLAMA_LOG_WARN("warning: SetProcessWorkingSetSize failed: %s\n",
llama_format_win_err(GetLastError()).c_str());
return false;
}
}
}
void llama_mlock::raw_unlock(void * ptr, size_t len) {
if (!VirtualUnlock(ptr, len)) {
LLAMA_LOG_WARN("warning: failed to VirtualUnlock buffer: %s\n",
llama_format_win_err(GetLastError()).c_str());
}
}
#else
size_t llama_mlock::lock_granularity() {
return (size_t) 65536;
}
bool llama_mlock::raw_lock([[maybe_unused]] void * addr, [[maybe_unused]] size_t len) const {
LLAMA_LOG_WARN("warning: mlock not supported on this system\n");
return false;
}
void llama_mlock::raw_unlock([[maybe_unused]] void * addr, [[maybe_unused]] size_t len) {}
#endif

175
src/llama-mmap.h Normal file
View File

@@ -0,0 +1,175 @@
#pragma once
#include <cstdio>
#include <cstddef>
#include <cstdint>
#include <vector>
#include <memory>
#if defined(_WIN32)
#define WIN32_LEAN_AND_MEAN
#ifndef NOMINMAX
#define NOMINMAX
#endif
#include <windows.h>
#ifndef PATH_MAX
#define PATH_MAX MAX_PATH
#endif
#include <io.h>
#endif
struct llama_file {
#if defined(_WIN32)
// use FILE * so we don't have to re-open the file to mmap
FILE * fp;
HANDLE fp_win32;
size_t size;
private:
std::string GetErrorMessageWin32(DWORD error_code) const;
public:
llama_file(const char * fname, const char * mode);
size_t tell() const;
void seek(size_t offset, int whence) const;
void read_raw(void * ptr, size_t len) const;
uint32_t read_u32() const {
uint32_t val;
read_raw(&val, sizeof(val));
return val;
}
void write_raw(const void * ptr, size_t len) const;
void write_u32(std::uint32_t val) const {
write_raw(&val, sizeof(val));
}
~llama_file();
#else
// use FILE * so we don't have to re-open the file to mmap
FILE * fp;
size_t size;
llama_file(const char * fname, const char * mode);
size_t tell() const;
void seek(size_t offset, int whence) const;
void read_raw(void * ptr, size_t len) const;
uint32_t read_u32() const {
uint32_t ret;
read_raw(&ret, sizeof(ret));
return ret;
}
void write_raw(const void * ptr, size_t len) const;
void write_u32(std::uint32_t val) const {
write_raw(&val, sizeof(val));
}
~llama_file();
#endif
};
using llama_files = std::vector<std::unique_ptr<llama_file>>;
struct llama_mmap {
void * addr;
size_t size;
size_t mapped_page_size = 0;
llama_mmap(const llama_mmap &) = delete;
#ifdef _POSIX_MAPPED_FILES
static constexpr bool SUPPORTED = true;
// list of mapped fragments (first_offset, last_offset)
std::vector<std::pair<size_t, size_t>> mapped_fragments;
llama_mmap(struct llama_file * file, size_t prefetch = (size_t) -1 /* -1 = max value */, bool numa = false, bool use_thp = false);
static void align_range(size_t * first, size_t * last, size_t page_size) {
// align first to the next page
size_t offset_in_page = *first & (page_size - 1);
size_t offset_to_page = offset_in_page == 0 ? 0 : page_size - offset_in_page;
*first += offset_to_page;
// align last to the previous page
*last = *last & ~(page_size - 1);
if (*last <= *first) {
*last = *first;
}
}
// partially unmap the file in the range [first, last)
void unmap_fragment(size_t first, size_t last);
#ifdef __linux__
static int get_default_huge_page_size();
#endif
~llama_mmap();
#elif defined(_WIN32)
static constexpr bool SUPPORTED = true;
llama_mmap(struct llama_file * file, size_t prefetch = (size_t) -1, bool numa = false, bool use_thp = false);
void unmap_fragment(size_t first, size_t last);
~llama_mmap();
#else
static constexpr bool SUPPORTED = false;
llama_mmap(struct llama_file * file, size_t prefetch = -1, bool numa = false, bool use_thp = false);
void unmap_fragment(size_t first, size_t last);
#endif
};
using llama_mmaps = std::vector<std::unique_ptr<llama_mmap>>;
// Represents some region of memory being locked using mlock or VirtualLock;
// will automatically unlock on destruction.
struct llama_mlock {
void * addr = NULL;
size_t size = 0;
bool failed_already = false;
llama_mlock() {}
llama_mlock(const llama_mlock &) = delete;
~llama_mlock() {
if (size) {
raw_unlock(addr, size);
}
}
void init(void * ptr);
void grow_to(size_t target_size);
static size_t lock_granularity();
bool raw_lock(void * ptr, size_t len) const;
static void raw_unlock(void * ptr, size_t len);
#ifdef _POSIX_MEMLOCK_RANGE
static constexpr bool SUPPORTED = true;
#elif defined(_WIN32)
static constexpr bool SUPPORTED = true;
#else
static constexpr bool SUPPORTED = false;
#endif
};
using llama_mlocks = std::vector<std::unique_ptr<llama_mlock>>;

1057
src/llama-model-loader.cpp Normal file
View File

File diff suppressed because it is too large Load Diff

169
src/llama-model-loader.h Normal file
View File

@@ -0,0 +1,169 @@
#pragma once
#include "llama.h"
#include "llama-impl.h"
#include "llama-mmap.h"
#include "llama-arch.h"
#include <cstdint>
#include <cstddef>
#include <stdexcept>
#include <unordered_map>
#include <vector>
enum llama_fver {
GGUF_FILE_VERSION_V1 = 1,
GGUF_FILE_VERSION_V2 = 2,
GGUF_FILE_VERSION_V3 = 3,
};
static const char * llama_file_version_name(llama_fver version) {
switch (version) {
case GGUF_FILE_VERSION_V1: return "GGUF V1 (support until nov 2023)";
case GGUF_FILE_VERSION_V2: return "GGUF V2";
case GGUF_FILE_VERSION_V3: return "GGUF V3 (latest)";
}
return "unknown";
}
using llama_buf_map = std::unordered_map<uint32_t, ggml_backend_buffer_t>;
struct llama_model_loader {
int n_kv = 0;
int n_tensors = 0;
int n_created = 0;
int64_t n_elements = 0;
size_t n_bytes = 0;
bool use_mmap = false;
bool check_tensors;
bool repack_tensors = false;
bool use_thp = false;
llama_files files;
llama_ftype ftype;
llama_fver fver;
llama_mmaps mappings;
// Holds information on a model weight
struct llama_tensor_weight {
uint16_t idx; // source file index
size_t offs; // tensor data offset in the original file
ggml_tensor * tensor;
llama_tensor_weight(const llama_file * file, uint16_t idx, const char * name, const struct gguf_context * gguf_ctx, ggml_tensor * tensor) : idx(idx), tensor(tensor) {
const int tensor_idx = gguf_find_tensor(gguf_ctx, name);
offs = gguf_get_data_offset(gguf_ctx) + gguf_get_tensor_offset(gguf_ctx, tensor_idx);
if (offs + ggml_nbytes(tensor) < offs || offs + ggml_nbytes(tensor) > file->size) {
throw std::runtime_error(format("tensor '%s' data is not within the file bounds, model is corrupted or incomplete", name));
}
}
};
std::vector<llama_tensor_weight> weights;
std::unordered_map<std::string, struct llama_model_kv_override> kv_overrides;
const llama_model_tensor_buft_override * tensor_buft_overrides;
gguf_context * meta = NULL;
std::vector<ggml_context *> contexts;
std::string arch_name;
LLM_KV llm_kv = LLM_KV(LLM_ARCH_UNKNOWN);
llama_model_loader(const std::string & fname, bool use_mmap, bool check_tensors, bool repack_tensors, bool use_thp,
const llama_model_kv_override * param_overrides_p,
const llama_model_tensor_buft_override * param_tensor_buft_overrides_p);
~llama_model_loader();
template<typename T>
typename std::enable_if<std::is_integral<T>::value, bool>::type
get_arr_n(const std::string & key, T & result, const bool required = true);
template<typename T>
typename std::enable_if<std::is_integral<T>::value, bool>::type
get_arr_n(const enum llm_kv kid, T & result, const bool required = true);
template<typename T>
bool get_arr(const std::string & key, std::vector<T> & result, const bool required = true);
template<typename T, size_t N_MAX>
bool get_arr(const std::string & key, std::array<T, N_MAX> & result, const bool required = true);
template<typename T>
bool get_arr(const enum llm_kv kid, T & result, const bool required = true);
template<typename T>
bool get_key(const std::string & key, T & result, const bool required = true);
template<typename T>
bool get_key(const enum llm_kv kid, T & result, const bool required = true);
// get array of n <= N_MAX elements, or a single element repeated n times
template<typename T, size_t N_MAX>
bool get_key_or_arr(const std::string & key, std::array<T, N_MAX> & result, uint32_t n, const bool required = true);
template<typename T>
bool get_key_or_arr(const enum llm_kv kid, T & result, uint32_t n, const bool required = true);
const std::string& get_arch_name() const { return arch_name; }
enum llm_arch get_arch() const { return llm_kv.arch; }
const char * get_tensor_name(int i) const;
const llama_tensor_weight * get_weight(const char * name) const;
const llama_tensor_weight * get_weight(int i) const {
return get_weight(get_tensor_name(i));
}
const llama_tensor_weight & require_weight(const char * name) const;
struct ggml_tensor * get_tensor_meta(const char * name) const;
struct ggml_tensor * require_tensor_meta(const char * name) const;
struct ggml_tensor * get_tensor_meta(int i) const {
return get_tensor_meta(get_tensor_name(i));
}
struct ggml_tensor * create_tensor_for(struct ggml_context * ctx, const struct ggml_tensor * cur, bool duplicated);
const struct ggml_tensor * check_tensor_dims(const std::string & name, const std::vector<int64_t> & ne, bool required) const;
static const int TENSOR_NOT_REQUIRED = 1 << 0;
static const int TENSOR_DUPLICATED = 1 << 1;
static const int TENSOR_SKIP = 1 << 2;
struct ggml_tensor * create_tensor(struct ggml_context * ctx, const std::string & name, const std::vector<int64_t> & ne, int flags = 0);
struct ggml_tensor * create_tensor_as_view(struct ggml_context * ctx, struct ggml_tensor * base,
const std::string & name, const std::vector<int64_t> & ne, size_t offset, bool required = true);
void done_getting_tensors() const;
void init_mappings(bool prefetch = true, llama_mlocks * mlock_mmaps = nullptr, bool use_thp = false);
void get_mapping_range(size_t * first, size_t * last, void ** addr, int idx, ggml_context * ctx) const;
// for backwards compatibility, does not support ggml-backend
void load_data_for(struct ggml_tensor * cur) const;
size_t size_done = 0;
size_t size_data = 0;
std::vector<std::pair<size_t, size_t>> mmaps_used;
// Returns false if cancelled by progress_callback
bool load_all_data(
struct ggml_context * ctx,
llama_buf_map & bufs_mmap,
llama_mlocks * lmlocks,
llama_progress_callback progress_callback,
void * progress_callback_user_data);
};

16
src/llama-vocab.cpp
View File

@@ -16,22 +16,6 @@
// helpers
//
LLAMA_ATTRIBUTE_FORMAT(1, 2)
static std::string format(const char * fmt, ...) {
va_list ap;
va_list ap2;
va_start(ap, fmt);
va_copy(ap2, ap);
int size = vsnprintf(NULL, 0, fmt, ap);
GGML_ASSERT(size >= 0 && size < INT_MAX); // NOLINT
std::vector<char> buf(size + 1);
int size2 = vsnprintf(buf.data(), size + 1, fmt, ap2);
GGML_ASSERT(size2 == size);
va_end(ap2);
va_end(ap);
return std::string(buf.data(), size);
}
struct naive_trie {
naive_trie() : has_value(false), value(0) {
}

2084
src/llama.cpp
View File

File diff suppressed because it is too large Load Diff