starred/ollama
1
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2026-04-24 09:46:01 +02:00
Code Issues Packages Projects Releases Wiki Activity

Update vendor ggml code to a5bb8ba4 (#13832)

Browse Source 
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
Co-authored-by: Gabe Goodhart <ghart@us.ibm.com>
Co-authored-by: Shalini Salomi Bodapati <Shalini.Salomi.Bodapati@ibm.com>
This commit is contained in:
Jeffrey Morgan
2026-02-02 17:31:59 -08:00
committed by GitHub
parent 8f4a008139
commit ef00199fb4
241 changed files with 21271 additions and 5074 deletions
Download Patch File Download Diff File Expand all files Collapse all files

112
llama/llama.cpp/src/llama-model-loader.cpp vendored
View File

@@ -2,6 +2,7 @@
#include "ggml.h"
#include <algorithm>
#include <array>
#include <cinttypes>
#include <cstring>
@@ -344,6 +345,7 @@ namespace GGUFMeta {
GGUFMeta::GKV<GGUFMeta::ArrayInfo>::get_kv(ctx, kid);
switch (arr_info.gt) {
case GGUF_TYPE_BOOL:
case GGUF_TYPE_UINT32:
case GGUF_TYPE_INT32: GGML_ASSERT((std::is_same<T, int32_t>::value) ||
(std::is_same<T, uint32_t>::value)); break;
@@ -365,7 +367,13 @@ namespace GGUFMeta {
result[i] = value;
}
} else {
std::copy((const T*)arr_info.data, (const T *)arr_info.data + arr_info.length, result.begin());
if (arr_info.gt == GGUF_TYPE_BOOL) {
std::transform((const bool *)arr_info.data, (const bool *)arr_info.data + arr_info.length, result.begin(), [](bool x) {
return static_cast<T>(x);
});
} else {
std::copy((const T*)arr_info.data, (const T *)arr_info.data + arr_info.length, result.begin());
}
}
return true;
@@ -462,6 +470,29 @@ namespace GGUFMeta {
return get_key_or_arr(llm_kv(kid), result, n, required);
}
bool llama_model_loader::get_key_or_arr(enum llm_kv kid, uint32_t & result, bool required) {
const std::string key = llm_kv(kid);
const int id = gguf_find_key(meta.get(), key.c_str());
if (id < 0) {
if (required) {
throw std::runtime_error(format("key not found in model: %s", key.c_str()));
}
return false;
}
// throw and error if type is an array
if (gguf_get_kv_type(meta.get(), id) == GGUF_TYPE_ARRAY) {
if (required) {
throw std::runtime_error(format("expected scalar, found array for key: %s", key.c_str()));
}
return false;
}
return get_key(key, result, required);
}
// TODO: this is not very clever - figure out something better
template bool llama_model_loader::get_key_or_arr<std::array<int, 4>>(enum llm_kv kid, std::array<int, 4> & result, uint32_t n, bool required);
template bool llama_model_loader::get_key_or_arr<std::array<uint32_t, 512>>(enum llm_kv kid, std::array<uint32_t, 512> & result, uint32_t n, bool required);
@@ -472,6 +503,7 @@ 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,
@@ -504,9 +536,23 @@ 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"));
files.emplace_back(new llama_file(fname.c_str(), "rb", use_direct_io));
contexts.emplace_back(ctx);
if (use_mmap && use_direct_io) {
if (files.back()->has_direct_io()) {
// Disable mmap, as DirectIO is available
use_mmap = false;
LLAMA_LOG_WARN("%s: direct I/O is enabled, disabling mmap\n", __func__);
} else {
// Disable DirectIO and reopen file using std::fopen for mmap
use_direct_io = false;
files.pop_back();
files.emplace_back(new llama_file(fname.c_str(), "rb", false));
LLAMA_LOG_WARN("%s: direct I/O is not available, using 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.
@@ -572,7 +618,7 @@ llama_model_loader::llama_model_loader(
}
}
files.emplace_back(new llama_file(fname_split, "rb"));
files.emplace_back(new llama_file(fname_split, "rb", use_direct_io));
contexts.emplace_back(ctx);
// Save tensors data offset info of the shard.
@@ -716,6 +762,7 @@ 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;
}
@@ -935,7 +982,15 @@ bool llama_model_loader::load_all_data(
// 4 staging buffers for async uploads, each sized 1MB seems to be a good default for single NVMe drives.
// NVMe raid configurations might require more / larger buffers.
constexpr size_t n_buffers = 4;
constexpr size_t buffer_size = 1 * 1024 * 1024; // 1MB
size_t alignment = 1;
for (const auto & file : files) {
alignment = std::max(file->read_alignment(), alignment);
}
// Buffer size: balance between memory usage and I/O efficiency
// 64MB works well for NVMe drives
const size_t buffer_size = alignment != 1 ? 64 * 1024 * 1024 + 2 * alignment : 1 * 1024 * 1024;
std::vector<ggml_backend_buffer_t> host_buffers;
std::vector<ggml_backend_event_t> events;
@@ -985,6 +1040,7 @@ bool llama_model_loader::load_all_data(
// If the backend is supported, create pinned memory buffers and events for synchronisation.
for (size_t idx = 0; idx < n_buffers; ++idx) {
auto * buf = ggml_backend_buft_alloc_buffer(host_buft, buffer_size);
if (!buf) {
LLAMA_LOG_DEBUG("%s: failed to allocate host buffer for async uploads for device %s\n", func,
ggml_backend_dev_name(dev));
@@ -1066,6 +1122,7 @@ bool llama_model_loader::load_all_data(
}
} else {
const auto & file = files.at(weight->idx);
if (ggml_backend_buffer_is_host(cur->buffer)) {
file->seek(weight->offs, SEEK_SET);
file->read_raw(cur->data, n_size);
@@ -1077,19 +1134,54 @@ bool llama_model_loader::load_all_data(
} else {
// If upload_backend is valid load the tensor in chunks to pinned memory and upload the buffers asynchronously to the GPU.
if (upload_backend) {
file->seek(weight->offs, SEEK_SET);
size_t offset = weight->offs;
alignment = file->read_alignment();
size_t aligned_offset = offset & ~(alignment - 1);
size_t offset_from_alignment = offset - aligned_offset;
file->seek(aligned_offset, SEEK_SET);
// Calculate aligned read boundaries
size_t read_start = aligned_offset;
size_t read_end = (offset + n_size + alignment - 1) & ~(alignment - 1);
size_t bytes_read = 0;
size_t data_read = 0; // Actual tensor data copied (excluding padding)
while (bytes_read < n_size) {
size_t read_iteration = std::min<size_t>(buffer_size, n_size - bytes_read);
while (bytes_read < read_end - read_start) {
size_t read_size = std::min<size_t>(buffer_size, read_end - read_start - bytes_read);
// Align the destination pointer within the pinned buffer
uintptr_t ptr_dest_aligned = (reinterpret_cast<uintptr_t>(host_ptrs[buffer_idx]) + alignment - 1) & ~(alignment - 1);
// Wait for previous upload to complete before reusing buffer
ggml_backend_event_synchronize(events[buffer_idx]);
file->read_raw(host_ptrs[buffer_idx], read_iteration);
ggml_backend_tensor_set_async(upload_backend, cur, host_ptrs[buffer_idx], bytes_read, read_iteration);
// Read aligned chunk from file
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;
size_t data_to_copy = read_size;
// Skip alignment padding at start of first chunk
if (bytes_read == 0) {
ptr_data += offset_from_alignment;
data_to_copy -= offset_from_alignment;
}
// Trim alignment padding at end of last chunk
if (aligned_offset + bytes_read + read_size > offset + n_size) {
data_to_copy -= (read_end - (offset + n_size));
}
// Async upload actual data to GPU
ggml_backend_tensor_set_async(upload_backend, cur,
reinterpret_cast<void *>(ptr_data), data_read, data_to_copy);
ggml_backend_event_record(events[buffer_idx], upload_backend);
bytes_read += read_iteration;
data_read += data_to_copy;
bytes_read += read_size;
++buffer_idx;
buffer_idx %= n_buffers;
}