github/gpt4all
1
0
Fork 0
mirror of https://github.com/nomic-ai/gpt4all.git synced 2025-09-07 11:30:05 +00:00
Code Issues Packages Projects Releases Wiki Activity

Dlopen better implementation management (Version 2)

Browse Source
This commit is contained in:
niansa
2023-05-31 21:37:25 +02:00
committed by Adam Treat
parent 92407438c8
commit a3d08cdcd5
7 changed files with 80 additions and 47 deletions
Download Patch File Download Diff File Expand all files Collapse all files

59
gpt4all-backend/llmodel.cpp
View File

@@ -5,6 +5,7 @@
#include <vector>
#include <fstream>
#include <filesystem>
#include <cassert>
@@ -24,12 +25,27 @@ bool requires_avxonly() {
}
static Dlhandle *get_implementation(std::ifstream& f, const std::string& buildVariant) {
// Collect all model implementation libraries
// NOTE: allocated on heap so we leak intentionally on exit so we have a chance to clean up the
// individual models without the cleanup of the static list interfering
static auto* libs = new std::vector<Dlhandle>([] () {
std::vector<Dlhandle> fres;
LLModel::Implementation::Implementation(Dlhandle &&dlhandle_) : dlhandle(std::move(dlhandle_)) {
auto get_model_type = dlhandle.get<const char *()>("get_model_type");
assert(get_model_type);
modelType = get_model_type();
auto get_build_variant = dlhandle.get<const char *()>("get_build_variant");
assert(get_build_variant);
buildVariant = get_build_variant();
magicMatch = dlhandle.get<bool(std::ifstream&)>("magic_match");
assert(magicMatch);
construct_ = dlhandle.get<LLModel *()>("construct");
assert(construct_);
}
bool LLModel::Implementation::isImplementation(const Dlhandle &dl) {
return dl.get<bool(uint32_t)>("is_g4a_backend_model_implementation");
}
const std::vector<LLModel::Implementation> &LLModel::getImplementationList() {
static auto* libs = new std::vector<LLModel::Implementation>([] () {
std::vector<LLModel::Implementation> fres;
auto search_in_directory = [&](const std::filesystem::path& path) {
// Iterate over all libraries
@@ -41,9 +57,10 @@ static Dlhandle *get_implementation(std::ifstream& f, const std::string& buildVa
// Add to list if model implementation
try {
Dlhandle dl(p.string());
if (dl.get<bool(uint32_t)>("is_g4a_backend_model_implementation")) {
fres.emplace_back(std::move(dl));
if (!Implementation::isImplementation(dl)) {
continue;
}
fres.emplace_back(Implementation(std::move(dl)));
} catch (...) {}
}
};
@@ -54,21 +71,24 @@ static Dlhandle *get_implementation(std::ifstream& f, const std::string& buildVa
#endif
return fres;
}());
// Return static result
return *libs;
}
const LLModel::Implementation* LLModel::getImplementation(std::ifstream& f, const std::string& buildVariant) {
// Iterate over all libraries
for (auto& dl : *libs) {
for (const auto& i : getImplementationList()) {
f.seekg(0);
// Check that magic matches
auto magic_match = dl.get<bool(std::ifstream&)>("magic_match");
if (!magic_match || !magic_match(f)) {
if (!i.magicMatch(f)) {
continue;
}
// Check that build variant is correct
auto get_build_variant = dl.get<const char *()>("get_build_variant");
if (buildVariant != (get_build_variant?get_build_variant():"default")) {
if (buildVariant != i.buildVariant) {
continue;
}
// Looks like we're good to go, return this dlhandle
return &dl;
return &i;
}
// Nothing found, so return nothing
return nullptr;
@@ -87,14 +107,9 @@ LLModel *LLModel::construct(const std::string &modelPath, std::string buildVaria
std::ifstream f(modelPath, std::ios::binary);
if (!f) return nullptr;
// Get correct implementation
auto impl = get_implementation(f, buildVariant);
auto impl = getImplementation(f, buildVariant);
if (!impl) return nullptr;
f.close();
// Get inference constructor
auto constructor = impl->get<LLModel *()>("construct");
if (!constructor) return nullptr;
// Construct llmodel implementation
auto fres = constructor();
// Return final instance
return fres;
// Construct and return llmodel implementation
return impl->construct();
}