thek0tyara/llama-cpp-turboquant
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions 4

llama : add use_direct_io flag for model loading (#18166)

Browse source 
* Adding --direct-io flag for model loading

* Fixing read_raw() calls

* Fixing Windows read_raw_at

* Changing type off_t to size_t for windows and Renaming functions

* disable direct io when mmap is explicitly enabled

* Use read_raw_unsafe when upload_backend is available, not functional on some devices with Vulkan and SYCL

* Fallback to std::fread in case O_DIRECT fails due to bad address

* Windows: remove const keywords and unused functions

* Update src/llama-mmap.cpp

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

---------

Co-authored-by: jtischbein <jtischbein@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
This commit is contained in:
Julius Tischbein 2026-01-08 07:35:30 +01:00 • committed by GitHub
parent 568371a726
commit 2038101bd9
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
12 changed files with 118 additions and 53 deletions
Download patch file Download diff file Expand all files Collapse all files

22
src/llama-model-loader.cpp
View file

@ -495,6 +495,7 @@ llama_model_loader::llama_model_loader(
const std::string & fname,
std::vector<std::string> & splits,
bool use_mmap,
bool use_direct_io,
bool check_tensors,
bool no_alloc,
const llama_model_kv_override * param_overrides_p,
@ -527,9 +528,17 @@ llama_model_loader::llama_model_loader(
get_key(llm_kv(LLM_KV_GENERAL_ARCHITECTURE), arch_name, false);
llm_kv = LLM_KV(llm_arch_from_string(arch_name));
files.emplace_back(new llama_file(fname.c_str(), "rb", !use_mmap));
files.emplace_back(new llama_file(fname.c_str(), "rb", use_direct_io));
contexts.emplace_back(ctx);
use_direct_io = use_direct_io && files.back()->has_direct_io();
// Disable mmap in case Direct I/O is enabled and available
if (use_direct_io && use_mmap) {
use_mmap = false;
LLAMA_LOG_WARN("%s: direct I/O is enabled, disabling mmap\n", __func__);
}
// Save tensors data offset of the main file.
// For subsidiary files, `meta` tensor data offset must not be used,
// so we build a unified tensors index for weights.
@ -595,7 +604,7 @@ llama_model_loader::llama_model_loader(
}
}
files.emplace_back(new llama_file(fname_split, "rb", !use_mmap));
files.emplace_back(new llama_file(fname_split, "rb", use_direct_io));
contexts.emplace_back(ctx);
// Save tensors data offset info of the shard.
@ -739,6 +748,7 @@ llama_model_loader::llama_model_loader(
}
this->use_mmap = use_mmap;
this->use_direct_io = use_direct_io;
this->check_tensors = check_tensors;
this->no_alloc = no_alloc;
}
@ -1100,7 +1110,8 @@ bool llama_model_loader::load_all_data(
const auto & file = files.at(weight->idx);
if (ggml_backend_buffer_is_host(cur->buffer)) {
file->read_raw_at(cur->data, n_size, weight->offs);
file->seek(weight->offs, SEEK_SET);
file->read_raw(cur->data, n_size);
if (check_tensors) {
validation_result.emplace_back(std::async(std::launch::async, [cur, n_size] {
return std::make_pair(cur, ggml_validate_row_data(cur->type, cur->data, n_size));
@ -1132,7 +1143,7 @@ bool llama_model_loader::load_all_data(
ggml_backend_event_synchronize(events[buffer_idx]);
// Read aligned chunk from file
file->read_raw(reinterpret_cast<void *>(ptr_dest_aligned), read_size);
file->read_raw_unsafe(reinterpret_cast<void *>(ptr_dest_aligned), read_size);
// Calculate actual data portion (excluding alignment padding)
uintptr_t ptr_data = ptr_dest_aligned;
@ -1162,7 +1173,8 @@ bool llama_model_loader::load_all_data(
}
} else {
read_buf.resize(n_size);
file->read_raw_at(read_buf.data(), n_size, weight->offs);
file->seek(weight->offs, SEEK_SET);
file->read_raw(read_buf.data(), n_size);
ggml_backend_tensor_set(cur, read_buf.data(), 0, n_size);
if (check_tensors && !ggml_validate_row_data(cur->type, read_buf.data(), n_size)) {
throw std::runtime_error(format("tensor '%s' has invalid data", ggml_get_name(cur)));