Skip to content

refactor: control optimizer checkpoint state with pointers #202

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
kilinchange merged 1 commit into from
Aug 7, 2026
Merged

refactor: control optimizer checkpoint state with pointers #202

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
7 changes: 2 additions & 5 deletions example/gpt2/main.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -92,7 +92,6 @@ DEFINE_uint32(save_interval, 0, "save checkpoint every N steps; 0 disables savin
DEFINE_string(load, "", "checkpoint directory to resume from");
DEFINE_string(save, "", "root directory used to store checkpoints");
DEFINE_uint32(max_checkpoint_keep, 3, "max number of checkpoint steps to keep");
DEFINE_bool(save_optimizer_state, true, "whether optimizer state is persisted in checkpoints");
// precision check
DEFINE_string(
precision_check, "",
Expand Down Expand Up @@ -366,10 +365,9 @@ void Train(const nn::parallel::Rank &rank) {
const auto resume_result = ResumeFromCheckpoint({.resume_root = FLAGS_load,
.rank = rank,
.model = model,
.optimizer = optimizer,
.optimizer = nullptr,
.model_config = model_config,
.state = state,
.load_optimizer_state = false,
.lr_scheduler = scheduler});
start_step = resume_result.global_step;
size_t consumed_batches = resume_result.consumed_batches;
Expand Down Expand Up @@ -398,12 +396,11 @@ void Train(const nn::parallel::Rank &rank) {
.tp_size = tp_world_size,
.sp_size = sp_world_size,
.pp_size = pp_world_size,
.save_optimizer_state = FLAGS_save_optimizer_state,
.checkpoint_root_dir = FLAGS_save,
.max_checkpoint_keep = FLAGS_max_checkpoint_keep,
.rank = rank,
.model = *model,
.optimizer = *optimizer,
.optimizer = nullptr,
.lr_scheduler = scheduler.get(),
});
};
Expand Down
7 changes: 3 additions & 4 deletions example/llama3/main.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -90,6 +90,7 @@ DEFINE_uint32(save_interval, 0, "save checkpoint every N steps; 0 disables savin
DEFINE_string(load, "", "checkpoint directory to resume from");
DEFINE_string(save, "", "root directory used to store checkpoints");
DEFINE_uint32(max_checkpoint_keep, 3, "max number of checkpoint steps to keep");
DEFINE_bool(load_optimizer_state, true, "whether optimizer state is restored from checkpoints");
DEFINE_bool(save_optimizer_state, true, "whether optimizer state is persisted in checkpoints");

// precision check
Expand Down Expand Up @@ -345,10 +346,9 @@ void Train(const nn::parallel::Rank &rank) {
const auto resume_result = ResumeFromCheckpoint({.resume_root = FLAGS_load,
.rank = rank,
.model = model,
.optimizer = optimizer,
.optimizer = FLAGS_load_optimizer_state ? optimizer : nullptr,
.model_config = model_config,
.state = state,
.load_optimizer_state = true,
.lr_scheduler = scheduler});

start_step = resume_result.global_step;
Expand Down Expand Up @@ -378,12 +378,11 @@ void Train(const nn::parallel::Rank &rank) {
.tp_size = tp_world_size,
.sp_size = sp_world_size,
.pp_size = pp_world_size,
.save_optimizer_state = FLAGS_save_optimizer_state,
.checkpoint_root_dir = FLAGS_save,
.max_checkpoint_keep = FLAGS_max_checkpoint_keep,
.rank = rank,
.model = *model,
.optimizer = *optimizer,
.optimizer = FLAGS_save_optimizer_state ? optimizer.get() : nullptr,
.lr_scheduler = scheduler.get(),
});
};
Expand Down
4 changes: 2 additions & 2 deletions infini_train/include/checkpoint/checkpoint.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -32,10 +32,10 @@ struct TrainerState {
class Checkpoint {
public:
static void Save(const std::filesystem::path &checkpoint_dir, const nn::Module &model, const Optimizer *optimizer,
const TrainerState &state, bool save_optimizer_state, const LRScheduler *lr_scheduler);
const TrainerState &state, const LRScheduler *lr_scheduler);

static void Load(const std::filesystem::path &checkpoint_dir, nn::Module &model, Optimizer *optimizer,
TrainerState &state, bool load_optimizer_state, LRScheduler *lr_scheduler);
TrainerState &state, LRScheduler *lr_scheduler);

private:
static void SaveStateDict(const std::filesystem::path &path,
Expand Down
4 changes: 1 addition & 3 deletions infini_train/include/checkpoint/checkpoint_manager.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -29,7 +29,6 @@ struct ResumeFromCheckpointArgs {
std::shared_ptr<Optimizer> optimizer;
const nn::TransformerConfig &model_config;
TrainerState &state;
bool load_optimizer_state;
std::shared_ptr<LRScheduler> lr_scheduler = nullptr;
};

Expand All @@ -51,12 +50,11 @@ struct SaveCheckpointArgs {
int tp_size = 1;
int sp_size = 1;
int pp_size = 1;
bool save_optimizer_state = true;
std::filesystem::path checkpoint_root_dir;
size_t max_checkpoint_keep = 0;
const nn::parallel::Rank &rank;
const nn::Module &model;
const Optimizer &optimizer;
const Optimizer *optimizer = nullptr;
const LRScheduler *lr_scheduler = nullptr;
};

Expand Down
10 changes: 4 additions & 6 deletions infini_train/src/checkpoint/checkpoint.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -182,16 +182,15 @@ template <typename T> T ExtractNumberField(const std::string &content, const std
} // namespace

void Checkpoint::Save(const std::filesystem::path &checkpoint_dir, const nn::Module &model, const Optimizer *optimizer,
const TrainerState &state, bool save_optimizer_state, const LRScheduler *lr_scheduler) {
const TrainerState &state, const LRScheduler *lr_scheduler) {
std::filesystem::create_directories(checkpoint_dir);
LOG(INFO) << "[CKPT] Save begin: dir=" << checkpoint_dir << ", global_step=" << state.global_step;

const auto model_path = checkpoint_dir / ("model.ckpt");

SaveStateDict(model_path, model.StateDict());

if (save_optimizer_state) {
CHECK(optimizer != nullptr) << "Optimizer pointer is null, cannot save optimizer state.";
if (optimizer != nullptr) {
auto opt_state = optimizer->StateDict();
if (!opt_state.empty()) {
const auto opt_path = checkpoint_dir / "optimizer.ckpt";
Expand All @@ -208,14 +207,13 @@ void Checkpoint::Save(const std::filesystem::path &checkpoint_dir, const nn::Mod
}

void Checkpoint::Load(const std::filesystem::path &checkpoint_dir, nn::Module &model, Optimizer *optimizer,
TrainerState &state, bool load_optimizer_state, LRScheduler *lr_scheduler) {
TrainerState &state, LRScheduler *lr_scheduler) {
const auto model_path = checkpoint_dir / "model.ckpt";
LOG(INFO) << "[CKPT] Loading model: " << model_path;

model.LoadStateDict(LoadStateDict(model_path));

if (load_optimizer_state) {
CHECK(optimizer != nullptr) << "Optimizer pointer is null, cannot load optimizer state.";
if (optimizer != nullptr) {
const auto opt_path = checkpoint_dir / "optimizer.ckpt";
if (std::filesystem::exists(opt_path)) {
LOG(INFO) << "[CKPT] Loading optimizer: " << opt_path;
Expand Down
5 changes: 2 additions & 3 deletions infini_train/src/checkpoint/checkpoint_manager.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -39,8 +39,7 @@ ResumeFromCheckpointResult ResumeFromCheckpoint(const ResumeFromCheckpointArgs &
}
}

Checkpoint::Load(resume_dir, *args.model, args.optimizer.get(), args.state, args.load_optimizer_state,
args.lr_scheduler.get());
Checkpoint::Load(resume_dir, *args.model, args.optimizer.get(), args.state, args.lr_scheduler.get());

result.global_step = static_cast<int>(args.state.global_step);

Expand Down Expand Up @@ -89,7 +88,7 @@ void SaveCheckpoint(const SaveCheckpointArgs &args) {
state.sp_size = args.sp_size;
state.pp_size = args.pp_size;

Checkpoint::Save(args.save_dir, args.model, &args.optimizer, state, args.save_optimizer_state, args.lr_scheduler);
Checkpoint::Save(args.save_dir, args.model, args.optimizer, state, args.lr_scheduler);

const auto ckpt_end = std::chrono::high_resolution_clock::now();
const double ckpt_ms = std::chrono::duration<double, std::milli>(ckpt_end - ckpt_start).count();
Expand Down
2 changes: 0 additions & 2 deletions scripts/test_config.json
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -839,7 +839,6 @@
"pipeline_parallel": 2,
"virtual_pipeline_parallel": 2,
"save": "@CKPT_ROOT_DIR@/3d_ddp2_tp2_pp2_no_resume",
"save_optimizer_state": true,
"max_checkpoint_keep": 5
}
},
Expand All @@ -857,7 +856,6 @@
"virtual_pipeline_parallel": 2,
"load": "@CKPT_ROOT_DIR@/3d_ddp2_tp2_pp2_no_resume/checkpoint_step_000030",
"save": "@CKPT_ROOT_DIR@/3d_tp2_pp2_resume",
"save_optimizer_state": true,
"max_checkpoint_keep": 5
}
}
Expand Down
4 changes: 2 additions & 2 deletions tests/checkpoint/test_checkpoint_serialization.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -30,7 +30,7 @@ TEST_P(CheckpointSerializationTest, SaveAndLoadModelFP32) {

auto opt1 = std::make_shared<optimizers::Adam>(model1->Parameters(), 0.01);
TrainerState saved{.global_step = 42, .consumed_batches = 100};
Checkpoint::Save(dir, *model1, opt1.get(), saved, /*save_optimizer_state=*/true, nullptr);
Checkpoint::Save(dir, *model1, opt1.get(), saved, nullptr);

auto model2 = std::make_shared<nn::Linear>(3, 2, true, GetDevice());
auto q1 = std::make_shared<Tensor>(std::vector<int64_t>{2, 3}, DataType::kFLOAT32, GetDevice());
Expand All @@ -42,7 +42,7 @@ TEST_P(CheckpointSerializationTest, SaveAndLoadModelFP32) {
auto opt2 = std::make_shared<optimizers::Adam>(model2->Parameters(), 0.01);

TrainerState loaded;
Checkpoint::Load(dir, *model2, opt2.get(), loaded, /*load_optimizer_state=*/true, nullptr);
Checkpoint::Load(dir, *model2, opt2.get(), loaded, nullptr);

EXPECT_EQ(loaded.global_step, 42);
EXPECT_EQ(loaded.consumed_batches, 100);
Expand Down
6 changes: 3 additions & 3 deletions tests/checkpoint/test_lr_scheduler_state.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -60,15 +60,15 @@ TEST_P(LRSchedulerCheckpointTest, SaveAndLoadLRSchedulerState) {
StepTimes(sched1, 3);

TrainerState saved{.global_step = 3, .consumed_batches = 12};
Checkpoint::Save(dir, *model1, opt1.get(), saved, /*save_optimizer_state=*/false, sched1.get());
Checkpoint::Save(dir, *model1, nullptr, saved, sched1.get());
EXPECT_TRUE(std::filesystem::exists(dir / "lr_scheduler.ckpt"));

auto model2 = MakeModel(GetDevice());
auto opt2 = std::make_shared<optimizers::SGD>(model2->Parameters(), kBaseLR);
auto sched2 = CreateLRScheduler(opt2, MakeSchedulerConfig());

TrainerState loaded;
Checkpoint::Load(dir, *model2, opt2.get(), loaded, /*load_optimizer_state=*/false, sched2.get());
Checkpoint::Load(dir, *model2, nullptr, loaded, sched2.get());

EXPECT_EQ(loaded.global_step, 3);
EXPECT_EQ(loaded.consumed_batches, 12);
Expand All @@ -91,7 +91,7 @@ TEST_P(LRSchedulerCheckpointTest, SkipsLRSchedulerStateWhenSchedulerIsNull) {
auto opt1 = std::make_shared<optimizers::SGD>(model1->Parameters(), kBaseLR);

TrainerState saved{.global_step = 3};
Checkpoint::Save(dir, *model1, opt1.get(), saved, /*save_optimizer_state=*/false, nullptr);
Checkpoint::Save(dir, *model1, nullptr, saved, nullptr);
EXPECT_FALSE(std::filesystem::exists(dir / "lr_scheduler.ckpt"));

std::filesystem::remove_all(dir);
Expand Down
6 changes: 3 additions & 3 deletions tests/checkpoint/test_trainer_state.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -44,7 +44,7 @@ TEST_P(TrainerStateTest, TrainerStateFileCreated) {
*model->mutable_parameter("weight") = p;
auto opt = std::make_shared<optimizers::SGD>(model->Parameters(), 0.01);

Checkpoint::Save(dir, *model, opt.get(), saved, /*save_optimizer_state=*/true, nullptr);
Checkpoint::Save(dir, *model, opt.get(), saved, nullptr);

EXPECT_TRUE(std::filesystem::exists(dir / "trainer_state.json"));

Expand Down Expand Up @@ -80,7 +80,7 @@ TEST_P(TrainerStateTest, RoundTrip) {
*model1->mutable_parameter("weight") = p1;
auto opt1 = std::make_shared<optimizers::SGD>(model1->Parameters(), 0.01);

Checkpoint::Save(dir, *model1, opt1.get(), saved, /*save_optimizer_state=*/false, nullptr);
Checkpoint::Save(dir, *model1, nullptr, saved, nullptr);

auto model2 = std::make_shared<nn::Linear>(1, 3, true, GetDevice());
auto p2 = std::make_shared<Tensor>(std::vector<int64_t>{3}, DataType::kFLOAT32, GetDevice());
Expand All @@ -89,7 +89,7 @@ TEST_P(TrainerStateTest, RoundTrip) {
auto opt2 = std::make_shared<optimizers::SGD>(model2->Parameters(), 0.01);

TrainerState loaded;
Checkpoint::Load(dir, *model2, opt2.get(), loaded, /*load_optimizer_state=*/false, nullptr);
Checkpoint::Load(dir, *model2, nullptr, loaded, nullptr);

EXPECT_EQ(loaded.global_step, 99);
EXPECT_EQ(loaded.consumed_batches, 5000);
Expand Down
Loading