llama-cpp-pro 0.2.2

package/cpp/llama-mmap.cpp

@@ -0,0 +1,609 @@
+#include "llama-mmap.h"
+
+#include "llama-impl.h"
+
+#include "ggml.h"
+
+#include <cstring>
+#include <climits>
+#include <stdexcept>
+#include <cerrno>
+#include <algorithm>
+
+#ifdef __has_include
+    #if __has_include(<unistd.h>)
+        #include <unistd.h>
+        #if defined(_POSIX_MAPPED_FILES)
+            #include <sys/mman.h>
+            #include <fcntl.h>
+        #endif
+        #if defined(_POSIX_MEMLOCK_RANGE)
+            #include <sys/resource.h>
+        #endif
+    #endif
+#endif
+
+#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
+
+#if defined(__APPLE__)
+#include <TargetConditionals.h>
+#endif
+
+// TODO: consider moving to llama-impl.h if needed in more places
+#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;
+}
+#endif
+
+// llama_file
+
+struct llama_file::impl {
+#if defined(_WIN32)
+    HANDLE fp_win32;
+    std::string 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: %lx", error_code);
+        } else {
+            ret = lpMsgBuf;
+            LocalFree(lpMsgBuf);
+        }
+
+        return ret;
+    }
+
+    impl(const char * fname, const char * mode) {
+        fp = lm_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 tell() const {
+        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 seek(size_t offset, int whence) const {
+        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 read_raw(void * ptr, size_t len) const {
+        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;
+        }
+    }
+
+    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 {
+        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;
+        }
+    }
+
+    void write_u32(uint32_t val) const {
+        write_raw(&val, sizeof(val));
+    }
+
+    ~impl() {
+        if (fp) {
+            std::fclose(fp);
+        }
+    }
+#else
+    impl(const char * fname, const char * mode) {
+        fp = lm_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 tell() const {
+// TODO: this ifdef is never true?
+#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 seek(size_t offset, int whence) const {
+// TODO: this ifdef is never true?
+#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 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");
+        }
+    }
+
+    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 {
+        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)));
+        }
+    }
+
+    void write_u32(uint32_t val) const {
+        write_raw(&val, sizeof(val));
+    }
+
+    ~impl() {
+        if (fp) {
+            std::fclose(fp);
+        }
+    }
+#endif
+
+    FILE * fp;
+    size_t size;
+};
+
+llama_file::llama_file(const char * fname, const char * mode) : pimpl(std::make_unique<impl>(fname, mode)) {}
+llama_file::~llama_file() = default;
+
+size_t llama_file::tell() const { return pimpl->tell(); }
+size_t llama_file::size() const { return pimpl->size; }
+
+int llama_file::file_id() const {
+#ifdef _WIN32
+    return _fileno(pimpl->fp);
+#else
+#if defined(fileno)
+    return fileno(pimpl->fp);
+#else
+    return ::fileno(pimpl->fp);
+#endif
+#endif
+}
+
+void llama_file::seek(size_t offset, int whence) const { pimpl->seek(offset, whence); }
+void llama_file::read_raw(void * ptr, size_t len) const { pimpl->read_raw(ptr, len); }
+
+uint32_t llama_file::read_u32() const { return pimpl->read_u32(); }
+
+void llama_file::write_raw(const void * ptr, size_t len) const { pimpl->write_raw(ptr, len); }
+void llama_file::write_u32(uint32_t val) const { pimpl->write_u32(val); }
+
+// llama_mmap
+
+struct llama_mmap::impl {
+#ifdef _POSIX_MAPPED_FILES
+    std::vector<std::pair<size_t, size_t>> mapped_fragments;
+
+    impl(struct llama_file * file, size_t prefetch, bool numa) {
+        size = file->size();
+        int fd = file->file_id();
+        int flags = MAP_SHARED;
+        if (numa) { prefetch = 0; }
+#ifdef __linux__
+        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; }
+#endif
+        addr = mmap(NULL, file->size(), PROT_READ, flags, fd, 0);
+        if (addr == MAP_FAILED) {
+            throw std::runtime_error(format("mmap failed: %s", strerror(errno)));
+        }
+
+        if (prefetch > 0) {
+            if (madvise(addr, std::min(file->size(), prefetch), MADV_WILLNEED)) {
+                fprintf(stderr, "warning: madvise(.., MADV_WILLNEED) failed: %s\n",
+                        strerror(errno));
+            }
+        }
+        if (numa) {
+            if (madvise(addr, file->size(), MADV_RANDOM)) {
+                fprintf(stderr, "warning: madvise(.., MADV_RANDOM) failed: %s\n",
+                        strerror(errno));
+            }
+        }
+
+        mapped_fragments.emplace_back(0, file->size());
+    }
+
+    static void align_range(size_t * first, size_t * last, size_t page_size) {
+        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;
+
+        *last = *last & ~(page_size - 1);
+
+        if (*last <= *first) {
+            *last = *first;
+        }
+    }
+
+    void unmap_fragment(size_t first, size_t last) {
+#if defined(__EMSCRIPTEN__)
+        // HeapFS / mmapAlloc regions are not real POSIX mappings — munmap EINVAL.
+        LM_GGML_UNUSED(first);
+        LM_GGML_UNUSED(last);
+        return;
+#endif
+        int page_size = sysconf(_SC_PAGESIZE);
+        align_range(&first, &last, page_size);
+        size_t len = last - first;
+
+        if (len == 0) {
+            return;
+        }
+
+        LM_GGML_ASSERT(first % page_size == 0);
+        LM_GGML_ASSERT(last % page_size == 0);
+        LM_GGML_ASSERT(last > first);
+
+        void * next_page_start = (uint8_t *) addr + first;
+
+        if (munmap(next_page_start, len)) {
+            LLAMA_LOG_WARN("warning: munmap failed: %s\n", strerror(errno));
+        }
+
+        std::vector<std::pair<size_t, size_t>> new_mapped_fragments;
+        for (const auto & frag : mapped_fragments) {
+            if (frag.first < first && frag.second > last) {
+                new_mapped_fragments.emplace_back(frag.first, first);
+                new_mapped_fragments.emplace_back(last, frag.second);
+            } else if (frag.first < first && frag.second > first) {
+                new_mapped_fragments.emplace_back(frag.first, first);
+            } else if (frag.first < last && frag.second > last) {
+                new_mapped_fragments.emplace_back(last, frag.second);
+            } else if (frag.first >= first && frag.second <= last) {
+            } else {
+                new_mapped_fragments.push_back(frag);
+            }
+        }
+        mapped_fragments = std::move(new_mapped_fragments);
+    }
+
+    ~impl() {
+#if !defined(__EMSCRIPTEN__)
+        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));
+            }
+        }
+#endif
+    }
+#elif defined(_WIN32)
+    impl(struct llama_file * file, size_t prefetch, bool numa) {
+        LM_GGML_UNUSED(numa);
+
+        size = file->size();
+
+        HANDLE hFile = (HANDLE) _get_osfhandle(file->file_id());
+
+        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
+            BOOL (WINAPI *pPrefetchVirtualMemory) (HANDLE, ULONG_PTR, PWIN32_MEMORY_RANGE_ENTRY, ULONG);
+            HMODULE hKernel32 = GetModuleHandleW(L"kernel32.dll");
+
+            pPrefetchVirtualMemory = (decltype(pPrefetchVirtualMemory))(void *) GetProcAddress(hKernel32, "PrefetchVirtualMemory");
+
+            if (pPrefetchVirtualMemory) {
+                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
+            LLAMA_LOG_DEBUG("skipping PrefetchVirtualMemory because _WIN32_WINNT < 0x602\n");
+#endif
+        }
+    }
+
+    void unmap_fragment(size_t first, size_t last) {
+        LM_GGML_UNUSED(first);
+        LM_GGML_UNUSED(last);
+    }
+
+    ~impl() {
+        if (!UnmapViewOfFile(addr)) {
+            LLAMA_LOG_WARN("warning: UnmapViewOfFile failed: %s\n",
+                    llama_format_win_err(GetLastError()).c_str());
+        }
+    }
+#else
+    impl(struct llama_file * file, size_t prefetch, bool numa) {
+        LM_GGML_UNUSED(file);
+        LM_GGML_UNUSED(prefetch);
+        LM_GGML_UNUSED(numa);
+
+        throw std::runtime_error("mmap not supported");
+    }
+
+    void unmap_fragment(size_t first, size_t last) {
+        LM_GGML_UNUSED(first);
+        LM_GGML_UNUSED(last);
+
+        throw std::runtime_error("mmap not supported");
+    }
+#endif
+
+    void * addr;
+    size_t size;
+};
+
+llama_mmap::llama_mmap(struct llama_file * file, size_t prefetch, bool numa) : pimpl(std::make_unique<impl>(file, prefetch, numa)) {}
+llama_mmap::~llama_mmap() = default;
+
+size_t llama_mmap::size() const { return pimpl->size; }
+void * llama_mmap::addr() const { return pimpl->addr; }
+
+void llama_mmap::unmap_fragment(size_t first, size_t last) { pimpl->unmap_fragment(first, last); }
+
+#if defined(_POSIX_MEMLOCK_RANGE) || defined(_WIN32) || defined(__EMSCRIPTEN__)
+// Emscripten: mmap() works on MEMFS/HeapFS-backed files when HeapFS patches stream_ops.mmap.
+const bool llama_mmap::SUPPORTED  = true;
+#else
+const bool llama_mmap::SUPPORTED  = false;
+#endif
+
+// llama_mlock
+
+struct llama_mlock::impl {
+#ifdef _POSIX_MEMLOCK_RANGE
+    static size_t lock_granularity() {
+        return (size_t) sysconf(_SC_PAGESIZE);
+    }
+
+    bool raw_lock(const void * addr, size_t size) const {
+        if (!mlock(addr, size)) {
+            return true;
+        }
+
+#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
+
+        char* errmsg = std::strerror(errno);
+        bool suggest = (errno == ENOMEM);
+#if defined(TARGET_OS_VISION) || defined(TARGET_OS_TV) || defined(_AIX)
+        // visionOS/tvOS dont't support RLIMIT_MEMLOCK
+        // Skip resource limit checks on visionOS/tvOS
+        suggest = false;
+#else
+        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;
+        }
+#endif
+
+        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;
+    }
+
+    static void 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)
+    static size_t lock_granularity() {
+        SYSTEM_INFO si;
+        GetSystemInfo(&si);
+        return (size_t) si.dwPageSize;
+    }
+
+    bool 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;
+            }
+
+            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;
+            }
+            size_t increment = len + 1048576;
+            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;
+            }
+        }
+    }
+
+    static void 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
+    static size_t lock_granularity() {
+        return (size_t) 65536;
+    }
+
+    bool raw_lock(const void * addr, size_t len) const {
+        LLAMA_LOG_WARN("warning: mlock not supported on this system\n");
+        return false;
+    }
+
+    static void raw_unlock(const void * addr, size_t len) {}
+#endif
+
+    impl() : addr(NULL), size(0), failed_already(false) {}
+
+    void init(void * ptr) {
+        LM_GGML_ASSERT(addr == NULL && size == 0);
+        addr = ptr;
+    }
+
+    void grow_to(size_t target_size) {
+        LM_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;
+            }
+        }
+    }
+
+    void * addr;
+    size_t size;
+
+    bool failed_already;
+};
+
+llama_mlock::llama_mlock() : pimpl(std::make_unique<impl>()) {}
+llama_mlock::~llama_mlock() = default;
+
+void llama_mlock::init(void * ptr) { pimpl->init(ptr); }
+void llama_mlock::grow_to(size_t target_size) { pimpl->grow_to(target_size); }
+
+#if defined(_POSIX_MEMLOCK_RANGE) || defined(_WIN32)
+const bool llama_mlock::SUPPORTED = true;
+#else
+const bool llama_mlock::SUPPORTED = false;
+#endif
+
+size_t llama_path_max() {
+    return PATH_MAX;
+}

package/cpp/llama-mmap.h

@@ -0,0 +1,68 @@
+#pragma once
+
+#include <cstdint>
+#include <memory>
+#include <vector>
+
+struct llama_file;
+struct llama_mmap;
+struct llama_mlock;
+
+using llama_files  = std::vector<std::unique_ptr<llama_file>>;
+using llama_mmaps  = std::vector<std::unique_ptr<llama_mmap>>;
+using llama_mlocks = std::vector<std::unique_ptr<llama_mlock>>;
+
+struct llama_file {
+    llama_file(const char * fname, const char * mode);
+    ~llama_file();
+
+    size_t tell() const;
+    size_t size() const;
+
+    int file_id() const; // fileno overload
+
+    void seek(size_t offset, int whence) const;
+
+    void read_raw(void * ptr, size_t len) const;
+    uint32_t read_u32() const;
+
+    void write_raw(const void * ptr, size_t len) const;
+    void write_u32(uint32_t val) const;
+
+private:
+    struct impl;
+    std::unique_ptr<impl> pimpl;
+};
+
+struct llama_mmap {
+    llama_mmap(const llama_mmap &) = delete;
+    llama_mmap(struct llama_file * file, size_t prefetch = (size_t) -1, bool numa = false);
+    ~llama_mmap();
+
+    size_t size() const;
+    void * addr() const;
+
+    void unmap_fragment(size_t first, size_t last);
+
+    static const bool SUPPORTED;
+
+private:
+    struct impl;
+    std::unique_ptr<impl> pimpl;
+};
+
+struct llama_mlock {
+    llama_mlock();
+    ~llama_mlock();
+
+    void init(void * ptr);
+    void grow_to(size_t target_size);
+
+    static const bool SUPPORTED;
+
+private:
+    struct impl;
+    std::unique_ptr<impl> pimpl;
+};
+
+size_t llama_path_max();