From 276e5b781155e3bbe6834472c58f03dfe62efabe Mon Sep 17 00:00:00 2001 From: Slaren <2141330+slaren@users.noreply.github.com> Date: Wed, 29 Mar 2023 08:31:26 +0200 Subject: [PATCH] Unmap the file in llama_free --- llama.cpp | 41 +++++++++++++++++++++++++++++++---------- 1 file changed, 31 insertions(+), 10 deletions(-) diff --git a/llama.cpp b/llama.cpp index 096735c..0c220e4 100644 --- a/llama.cpp +++ b/llama.cpp @@ -149,6 +149,10 @@ struct llama_model { // the model memory buffer std::vector buf; + // model memory mapped file + void * mm_addr; + size_t mm_length; + // tensors int n_loaded; std::unordered_map tensors; @@ -296,22 +300,32 @@ struct llama_context_params llama_context_default_params() { // model loading // -static void * mmap_file(const char* fname) { +static void mmap_file(const char* fname, void * &mm_addr, size_t &mm_length) { #if defined(MAP_FAILED) - // POSIX mmap + // POSIX int fd = open(fname, O_RDONLY); - size_t len = lseek(fd, 0, SEEK_END); - void * mm_addr = mmap(NULL, len, PROT_READ, MAP_SHARED, fd, 0); + mm_length = lseek(fd, 0, SEEK_END); + mm_addr = mmap(NULL, mm_length, PROT_READ, MAP_SHARED, fd, 0); + close(fd); if (mm_addr == MAP_FAILED) { perror("mmap failed"); mm_addr = NULL; + mm_length = 0; } - close(fd); - return mm_addr; #else // TODO: windows support (void)(fname); // suppress warnings - return NULL; +#endif +} + +static void munmap_file(void * addr, size_t length) { +#if defined(MAP_FAILED) + // POSIX + munmap(addr, length); +#else + // TODO: windows support + (void)(addr); // suppress warnings + (void)(length); #endif } @@ -480,12 +494,15 @@ static bool llama_model_load( bool use_mmap = (n_parts == 1); // try to memory map the model file - void* mm_addr = NULL; + void * mm_addr = NULL; if (use_mmap) { - mm_addr = mmap_file(fname.c_str()); - if (mm_addr == NULL) { + mmap_file(fname.c_str(), model.mm_addr, model.mm_length); + if (model.mm_addr == NULL) { use_mmap = false; } + else { + mm_addr = model.mm_addr; + } } auto & ctx = model.ctx; @@ -1750,6 +1767,10 @@ void llama_free(struct llama_context * ctx) { ggml_free(ctx->model.ctx); } + if (ctx->model.mm_addr) { + munmap_file(ctx->model.mm_addr, ctx->model.mm_length); + } + delete ctx; }