Paul Adenot
4dd522c652
This is the result of running `./mach vendor third_party/llama.cpp/moz.yaml --patch none && ./mach vendor third_party/llama.cpp/moz.yaml --patch only`. This vendors the CMakeLists.txt files, and parses them, generating sources.mozbuild. Some custom code runs after that, but things have been stable for 3 or 4 updates in the last few weeks. This runs in a separate process (Inference) that has a similar sandbox to a content process. Only CPU inference is enabled, hardware acceleration will come at a later stage, and after sandboxing work. A few patches were needed, most notably to work around the lack of exceptions in Firefox, and to not use functions from libstdc++ that are too recent. Differential Revision: https://phabricator.services.mozilla.com/D253058
54 lines
2.4 KiB
Diff
54 lines
2.4 KiB
Diff
diff --git a/src/llama-model-loader.cpp b/src/llama-model-loader.cpp
|
|
index 1a90a1eb88..6f5f3d2868 100644
|
|
--- a/src/llama-model-loader.cpp
|
|
+++ b/src/llama-model-loader.cpp
|
|
@@ -5,7 +5,6 @@
|
|
#include <array>
|
|
#include <cinttypes>
|
|
#include <cstring>
|
|
-#include <future>
|
|
|
|
#include "moz-overrides.h"
|
|
|
|
@@ -926,7 +925,7 @@
|
|
GGML_ASSERT(size_data != 0 && "call init_mappings() first");
|
|
|
|
std::vector<no_init<uint8_t>> read_buf;
|
|
- std::vector<std::future<std::pair<ggml_tensor *, bool>>> validation_result;
|
|
+ std::vector<std::pair<ggml_tensor *, bool>> validation_result;
|
|
|
|
// 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.
|
|
@@ -1041,9 +1040,7 @@
|
|
uint8_t * data = (uint8_t *) mapping->addr() + weight->offs;
|
|
|
|
if (check_tensors) {
|
|
- validation_result.emplace_back(std::async(std::launch::async, [cur, data, n_size] {
|
|
- return std::make_pair(cur, ggml_validate_row_data(cur->type, data, n_size));
|
|
- }));
|
|
+ validation_result.push_back(std::make_pair(cur, ggml_validate_row_data(cur->type, data, n_size)));
|
|
}
|
|
|
|
GGML_ASSERT(buf_mmap || cur->data); // either we have a buffer to allocate the tensor in, or it is already allocated
|
|
@@ -1066,9 +1063,7 @@
|
|
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));
|
|
- }));
|
|
+ validation_result.push_back(std::make_pair(cur, ggml_validate_row_data(cur->type, cur->data, n_size)));
|
|
}
|
|
} else {
|
|
// If upload_backend is valid load the tensor in chunks to pinned memory and upload the buffers asynchronously to the GPU.
|
|
@@ -1116,8 +1111,7 @@
|
|
|
|
// check validation results
|
|
bool validation_failed = false;
|
|
- for (auto & future : validation_result) {
|
|
- auto result = future.get();
|
|
+ for (const auto & result : validation_result) {
|
|
if (!result.second) {
|
|
LLAMA_LOG_ERROR("%s: tensor '%s' has invalid data\n", __func__, ggml_get_name(result.first));
|
|
validation_failed = true;
|