happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

llama: fix leaked buffers for mmap + split files (#16765)

This commit is contained in:
Johannes Gäßler 2025-10-27 09:17:31 +01:00 committed by GitHub
parent 75cbdd3fce
commit 945501f5ea
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 16 additions and 10 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

24
src/llama-model.cpp
View File

@ -15,7 +15,6 @@
#include <algorithm> #include <algorithm>
#include <cassert> #include <cassert>
#include <cmath>
#include <cfloat> #include <cfloat>
#include <cstring> #include <cstring>
#include <cmath> #include <cmath>
@ -438,7 +437,7 @@ struct llama_model::impl {
llama_mlocks mlock_mmaps; llama_mlocks mlock_mmaps;
// contexts where the model tensors metadata is stored as well ass the corresponding buffers: // contexts where the model tensors metadata is stored as well ass the corresponding buffers:
std::vector<std::pair<ggml_context_ptr, ggml_backend_buffer_ptr>> ctxs_bufs; std::vector<std::pair<ggml_context_ptr, std::vector<ggml_backend_buffer_ptr>>> ctxs_bufs;
buft_list_t cpu_buft_list; buft_list_t cpu_buft_list;
std::map<ggml_backend_dev_t, buft_list_t> gpu_buft_list; std::map<ggml_backend_dev_t, buft_list_t> gpu_buft_list;
@ -6186,7 +6185,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
bool buffer_from_host_ptr_supported = props.caps.buffer_from_host_ptr; bool buffer_from_host_ptr_supported = props.caps.buffer_from_host_ptr;
bool is_default_buft = buft == ggml_backend_dev_buffer_type(dev); bool is_default_buft = buft == ggml_backend_dev_buffer_type(dev);
ggml_backend_buffer_t buf = nullptr; std::vector<ggml_backend_buffer_ptr> bufs;
if (ml.use_mmap && use_mmap_buffer && buffer_from_host_ptr_supported && is_default_buft) { if (ml.use_mmap && use_mmap_buffer && buffer_from_host_ptr_supported && is_default_buft) {
for (uint32_t idx = 0; idx < ml.files.size(); idx++) { for (uint32_t idx = 0; idx < ml.files.size(); idx++) {
// only the mmap region containing the tensors in the model is mapped to the backend buffer // only the mmap region containing the tensors in the model is mapped to the backend buffer
@ -6199,15 +6198,16 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
continue; continue;
} }
const size_t max_size = ggml_get_max_tensor_size(ctx); const size_t max_size = ggml_get_max_tensor_size(ctx);
buf = ggml_backend_dev_buffer_from_host_ptr(dev, (char *) addr + first, last - first, max_size); ggml_backend_buffer_t buf = ggml_backend_dev_buffer_from_host_ptr(dev, (char *) addr + first, last - first, max_size);
if (buf == nullptr) { if (buf == nullptr) {
throw std::runtime_error(format("unable to allocate %s buffer", ggml_backend_buft_name(buft))); throw std::runtime_error(format("unable to allocate %s buffer", ggml_backend_buft_name(buft)));
} }
bufs.emplace_back(buf);
buf_map.emplace(idx, buf); buf_map.emplace(idx, buf);
} }
} }
else { else {
buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx, buft); ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx, buft);
if (buf == nullptr) { if (buf == nullptr) {
throw std::runtime_error(format("unable to allocate %s buffer", ggml_backend_buft_name(buft))); throw std::runtime_error(format("unable to allocate %s buffer", ggml_backend_buft_name(buft)));
} }
@ -6217,11 +6217,12 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
mlock_buf->init (ggml_backend_buffer_get_base(buf)); mlock_buf->init (ggml_backend_buffer_get_base(buf));
mlock_buf->grow_to(ggml_backend_buffer_get_size(buf)); mlock_buf->grow_to(ggml_backend_buffer_get_size(buf));
} }
bufs.emplace_back(buf);
for (uint32_t idx = 0; idx < ml.files.size(); idx++) { for (uint32_t idx = 0; idx < ml.files.size(); idx++) {
buf_map.emplace(idx, buf); buf_map.emplace(idx, buf);
} }
} }
pimpl->ctxs_bufs.emplace_back(std::move(ctx_ptr), buf); pimpl->ctxs_bufs.emplace_back(std::move(ctx_ptr), std::move(bufs));
for (auto & buf : buf_map) { for (auto & buf : buf_map) {
// indicate that this buffer contains weights // indicate that this buffer contains weights
@ -6247,8 +6248,11 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
} }
// print memory requirements per buffer type // print memory requirements per buffer type
for (auto & [_, buf] : pimpl->ctxs_bufs) { for (auto & [_, bufs] : pimpl->ctxs_bufs) {
LLAMA_LOG_INFO("%s: %12s model buffer size = %8.2f MiB\n", __func__, ggml_backend_buffer_name(buf.get()), ggml_backend_buffer_get_size(buf.get()) / 1024.0 / 1024.0); for (auto & buf: bufs) {
LLAMA_LOG_INFO("%s: %12s model buffer size = %8.2f MiB\n",
__func__, ggml_backend_buffer_name(buf.get()), ggml_backend_buffer_get_size(buf.get()) / 1024.0 / 1024.0);
}
} }
// populate tensors_by_name // populate tensors_by_name
@ -6300,9 +6304,11 @@ size_t llama_model::n_devices() const {
std::map<ggml_backend_buffer_type_t, size_t> llama_model::memory_breakdown() const { std::map<ggml_backend_buffer_type_t, size_t> llama_model::memory_breakdown() const {
std::map<ggml_backend_buffer_type_t, size_t> ret; std::map<ggml_backend_buffer_type_t, size_t> ret;
for (const auto & [_, buf] : pimpl->ctxs_bufs) { for (const auto & [_, bufs] : pimpl->ctxs_bufs) {
for (const auto & buf : bufs) {
ret[ggml_backend_buffer_get_type(buf.get())] += ggml_backend_buffer_get_size(buf.get()); ret[ggml_backend_buffer_get_type(buf.get())] += ggml_backend_buffer_get_size(buf.get());
} }
}
return ret; return ret;
} }