fix(dataloader): track distributed progress by global batch

example/gpt2/main.cc

@@ -383,13 +383,20 @@ void Train(const nn::parallel::Rank &rank) {
     start_step = resume_result.global_step;
     size_t consumed_train_samples = resume_result.consumed_train_samples;
 
-    // TODO(jym): Replace with Sampler abstraction when available.
-    // Skip dataloader to resume from the correct batch position.
-    if (consumed_train_samples > 0) {
-        const size_t num_skips
-            = DataLoaderBatchesToSkip(consumed_train_samples, train_loader_batch_size, ddp_world_size);
-        for (size_t i = 0; i < num_skips; ++i) { ++train_iter; }
-    }
+    const size_t consumed_dataloader_steps
+        = DataLoaderStepsToSkip(consumed_train_samples, train_loader_batch_size, ddp_world_size);
+    train_iter.SeekDataLoaderStep(consumed_dataloader_steps % train_loader.NumDataLoaderSteps());
+    auto next_train_batch = [&]() {
+        auto batch = *train_iter;
+        // if we are trying to overfit a single batch, we reset the loader here by commenting out the line below
+        // TODO(dcj): support dataloader.reset() later
+        ++train_iter;
+        if (train_iter == train_loader.end()) {
+            train_iter = train_loader.begin();
+        }
+        consumed_train_samples += train_loader_batch_size * ddp_world_size;
+        return batch;
+    };
 
     auto save_checkpoint = [&](const std::filesystem::path &save_dir, int64_t global_step) {
         SaveCheckpoint({
@@ -463,11 +470,7 @@ void Train(const nn::parallel::Rank &rank) {
                 infini_train::AutocastGuard autocast_guard(device.type(), dtype);
 
                 // (bs, seq_len), (bs, seq_len)
-                auto [x, y] = *train_iter;
-                // if we are trying to overfit a single batch, we reset the loader here by commenting out the line below
-                // TODO(dcj): support dataloader.reset() later
-                ++train_iter;
-                consumed_train_samples += static_cast<size_t>(FLAGS_batch_size) * ddp_world_size;
+                auto [x, y] = next_train_batch();
                 x = std::make_shared<Tensor>(x->To(device));
                 y = std::make_shared<Tensor>(y->To(device));
 
@@ -499,11 +502,7 @@ void Train(const nn::parallel::Rank &rank) {
                 scheduler->Step();
             }
         } else {
-            auto [x, y] = *train_iter;
-            // if we are trying to overfit a single batch, we reset the loader here by commenting out the line below
-            // TODO(dcj): support dataloader.reset() later
-            ++train_iter;
-            consumed_train_samples += train_loader_batch_size * ddp_world_size;
+            auto [x, y] = next_train_batch();
             x = std::make_shared<Tensor>(x->To(device));
             y = std::make_shared<Tensor>(y->To(device));
 

example/llama3/main.cc

@@ -365,13 +365,20 @@ void Train(const nn::parallel::Rank &rank) {
     start_step = resume_result.global_step;
     size_t consumed_train_samples = resume_result.consumed_train_samples;
 
-    // TODO(jym): Replace with Sampler abstraction when available.
-    // Skip dataloader to resume from the correct batch position.
-    if (consumed_train_samples > 0) {
-        const size_t num_skips
-            = DataLoaderBatchesToSkip(consumed_train_samples, train_loader_batch_size, ddp_world_size);
-        for (size_t i = 0; i < num_skips; ++i) { ++train_iter; }
-    }
+    const size_t consumed_dataloader_steps
+        = DataLoaderStepsToSkip(consumed_train_samples, train_loader_batch_size, ddp_world_size);
+    train_iter.SeekDataLoaderStep(consumed_dataloader_steps % train_loader.NumDataLoaderSteps());
+    auto next_train_batch = [&]() {
+        auto batch = *train_iter;
+        // if we are trying to overfit a single batch, we reset the loader here by commenting out the line below
+        // TODO(dcj): support dataloader.reset() later
+        ++train_iter;
+        if (train_iter == train_loader.end()) {
+            train_iter = train_loader.begin();
+        }
+        consumed_train_samples += train_loader_batch_size * ddp_world_size;
+        return batch;
+    };
 
     auto save_checkpoint = [&](const std::filesystem::path &save_dir, int64_t global_step) {
         SaveCheckpoint({
@@ -443,11 +450,7 @@ void Train(const nn::parallel::Rank &rank) {
                 infini_train::AutocastGuard autocast_guard(device.type(), dtype);
 
                 // (bs, seq_len), (bs, seq_len)
-                auto [x, y] = *train_iter;
-                // if we are trying to overfit a single batch, we reset the loader here by commenting out the line below
-                // TODO(dcj): support dataloader.reset() later
-                ++train_iter;
-                consumed_train_samples += static_cast<size_t>(FLAGS_batch_size) * ddp_world_size;
+                auto [x, y] = next_train_batch();
                 x = std::make_shared<Tensor>(x->To(device));
                 y = std::make_shared<Tensor>(y->To(device));
 
@@ -478,11 +481,7 @@ void Train(const nn::parallel::Rank &rank) {
                 scheduler->Step();
             }
         } else {
-            auto [x, y] = *train_iter;
-            // if we are trying to overfit a single batch, we reset the loader here by commenting out the line below
-            // TODO(dcj): support dataloader.reset() later
-            ++train_iter;
-            consumed_train_samples += train_loader_batch_size * ddp_world_size;
+            auto [x, y] = next_train_batch();
             x = std::make_shared<Tensor>(x->To(device));
             y = std::make_shared<Tensor>(y->To(device));
 

infini_train/include/checkpoint/checkpoint_manager.h

@@ -62,4 +62,4 @@ ResumeFromCheckpointResult ResumeFromCheckpoint(const ResumeFromCheckpointArgs &
 
 void SaveCheckpoint(const SaveCheckpointArgs &args);
 
-size_t DataLoaderBatchesToSkip(size_t consumed_train_samples, size_t local_batch_size, size_t ddp_world_size);
+size_t DataLoaderStepsToSkip(size_t consumed_train_samples, size_t local_batch_size, size_t ddp_world_size);

infini_train/include/dataloader.h

@@ -12,9 +12,6 @@ class Tensor;
 namespace infini_train {
 class DataLoaderIterator {
 public:
-    DataLoaderIterator(const Dataset &dataset, size_t batch_size, size_t batch_idx, size_t max_batch_idx,
-                       size_t ddp_rank = 0, size_t ddp_world_size = 1);
-
     std::pair<std::shared_ptr<Tensor>, std::shared_ptr<Tensor>> operator*() const;
 
     DataLoaderIterator &operator++();
@@ -24,11 +21,20 @@ class DataLoaderIterator {
     friend bool operator!=(const DataLoaderIterator &lhs, const DataLoaderIterator &rhs);
     friend bool operator==(const DataLoaderIterator &lhs, const DataLoaderIterator &rhs);
 
+    size_t DataLoaderStep() const;
+    DataLoaderIterator &SeekDataLoaderStep(size_t dataloader_step);
+
 private:
+    friend class DataLoader;
+    friend class DistributedDataLoader;
+
+    DataLoaderIterator(const Dataset &dataset, size_t batch_size, size_t dataloader_step, size_t num_dataloader_steps,
+                       size_t ddp_rank, size_t ddp_world_size);
+
     const Dataset *dataset_ = nullptr; // not owned
     size_t batch_size_ = 0;
-    size_t batch_idx_ = 0;
-    size_t max_batch_idx_ = 0;
+    size_t dataloader_step_ = 0;
+    size_t num_dataloader_steps_ = 0;
     size_t ddp_rank_ = 0;
     size_t ddp_world_size_ = 1;
 };
@@ -40,10 +46,12 @@ class DataLoader {
     virtual DataLoaderIterator begin() const;
     virtual DataLoaderIterator end() const;
 
+    size_t NumDataLoaderSteps() const;
+
 protected:
     std::shared_ptr<Dataset> dataset_;
     size_t batch_size_ = 0;
-    size_t max_batch_idx_ = 0;
+    size_t num_dataloader_steps_ = 0;
 };
 
 class DistributedDataLoader : public DataLoader {

infini_train/src/checkpoint/checkpoint_manager.cc

@@ -120,15 +120,15 @@ void SaveCheckpoint(const SaveCheckpointArgs &args) {
     }
 }
 
-size_t DataLoaderBatchesToSkip(size_t consumed_train_samples, size_t local_batch_size, size_t ddp_world_size) {
+size_t DataLoaderStepsToSkip(size_t consumed_train_samples, size_t local_batch_size, size_t ddp_world_size) {
     CHECK_GT(local_batch_size, 0);
     CHECK_GT(ddp_world_size, 0);
     CHECK_LE(local_batch_size, std::numeric_limits<size_t>::max() / ddp_world_size)
         << "Data loader batch size overflows size_t";
-    const size_t global_loader_batch_size = local_batch_size * ddp_world_size;
-    CHECK_EQ(consumed_train_samples % global_loader_batch_size, 0)
+    const size_t samples_per_dataloader_step = local_batch_size * ddp_world_size;
+    CHECK_EQ(consumed_train_samples % samples_per_dataloader_step, 0)
         << "consumed_train_samples=" << consumed_train_samples
         << " does not align with current local_batch_size=" << local_batch_size
         << " and ddp_world_size=" << ddp_world_size;
-    return consumed_train_samples / global_loader_batch_size;
+    return consumed_train_samples / samples_per_dataloader_step;
 }

infini_train/src/dataloader.cc

@@ -3,6 +3,7 @@
 #include <algorithm>
 #include <cstddef>
 #include <cstring>
+#include <functional>
 #include <numeric>
 #include <utility>
 
@@ -13,8 +14,14 @@
 
 namespace infini_train {
 namespace {
+size_t CheckedCeilDiv(size_t numerator, size_t denominator) {
+    CHECK_GT(denominator, 0);
+    return (numerator + denominator - 1) / denominator;
+}
+
 // TODO(dcj): Use official stack implementation later.
 std::shared_ptr<Tensor> Stack(const std::vector<std::shared_ptr<Tensor>> &tensors) {
+    CHECK(!tensors.empty()) << "Cannot stack an empty batch. Check DataLoader iterator end handling.";
     const int batch_size = tensors.size();
     const auto &dims = tensors[0]->Dims();
     const int stacked_dim = std::accumulate(dims.begin(), dims.end(), 1, std::multiplies<int64_t>());
@@ -35,21 +42,31 @@ std::shared_ptr<Tensor> Stack(const std::vector<std::shared_ptr<Tensor>> &tensor
 }
 } // namespace
 
-DataLoaderIterator::DataLoaderIterator(const Dataset &dataset, size_t batch_size, size_t batch_idx,
-                                       size_t max_batch_idx, size_t ddp_rank, size_t ddp_world_size)
-    : dataset_(&dataset), batch_size_(batch_size), batch_idx_(batch_idx), max_batch_idx_(max_batch_idx),
-      ddp_rank_(ddp_rank), ddp_world_size_(ddp_world_size){};
+DataLoaderIterator::DataLoaderIterator(const Dataset &dataset, size_t batch_size, size_t dataloader_step,
+                                       size_t num_dataloader_steps, size_t ddp_rank, size_t ddp_world_size)
+    : dataset_(&dataset), batch_size_(batch_size), dataloader_step_(dataloader_step),
+      num_dataloader_steps_(num_dataloader_steps), ddp_rank_(ddp_rank), ddp_world_size_(ddp_world_size){};
 
 std::pair<std::shared_ptr<Tensor>, std::shared_ptr<Tensor>> DataLoaderIterator::operator*() const {
     /*
       0,         1,            ..., x,                  ...
       [0, bs-1], [bs, 2*bs-1], ..., [x*bs, (x+1)*bs-1], ...
                                     ^
-                                    batch_idx
+                                    dataloader_step
     */
     std::vector<std::shared_ptr<Tensor>> data_vec;
     std::vector<std::shared_ptr<Tensor>> label_vec;
-    for (int idx = batch_idx_ * batch_size_; idx < (batch_idx_ + 1) * batch_size_ && idx < dataset_->Size(); ++idx) {
+    CHECK_LT(dataloader_step_, num_dataloader_steps_)
+        << "Cannot dereference DataLoader end iterator. dataloader_step=" << dataloader_step_
+        << ", num_dataloader_steps=" << num_dataloader_steps_ << ", ddp_rank=" << ddp_rank_
+        << ", ddp_world_size=" << ddp_world_size_;
+    const size_t start_idx = (dataloader_step_ * ddp_world_size_ + ddp_rank_) * batch_size_;
+    CHECK_LT(start_idx, dataset_->Size())
+        << "DataLoader batch starts past dataset end. dataloader_step=" << dataloader_step_
+        << ", start_idx=" << start_idx << ", dataset_size=" << dataset_->Size() << ", batch_size=" << batch_size_
+        << ", ddp_rank=" << ddp_rank_ << ", ddp_world_size=" << ddp_world_size_;
+    const size_t end_idx = std::min(start_idx + batch_size_, dataset_->Size());
+    for (size_t idx = start_idx; idx < end_idx; ++idx) {
         auto &&[data, label] = dataset_->operator[](idx);
         data_vec.push_back(std::move(data));
         label_vec.push_back(std::move(label));
@@ -58,7 +75,7 @@ std::pair<std::shared_ptr<Tensor>, std::shared_ptr<Tensor>> DataLoaderIterator::
 }
 
 DataLoaderIterator &DataLoaderIterator::operator++() {
-    batch_idx_ = std::min(batch_idx_ + ddp_world_size_, max_batch_idx_);
+    dataloader_step_ = std::min(dataloader_step_ + 1, num_dataloader_steps_);
     return *this;
 }
 
@@ -68,34 +85,64 @@ DataLoaderIterator DataLoaderIterator::operator++(int) {
     return tmp;
 }
 
-bool operator<(const DataLoaderIterator &lhs, const DataLoaderIterator &rhs) { return lhs.batch_idx_ < rhs.batch_idx_; }
+bool operator<(const DataLoaderIterator &lhs, const DataLoaderIterator &rhs) {
+    return lhs.dataloader_step_ < rhs.dataloader_step_;
+}
 
 bool operator!=(const DataLoaderIterator &lhs, const DataLoaderIterator &rhs) {
-    return lhs.batch_idx_ != rhs.batch_idx_;
+    return lhs.dataloader_step_ != rhs.dataloader_step_;
 }
 
 bool operator==(const DataLoaderIterator &lhs, const DataLoaderIterator &rhs) {
-    return lhs.batch_idx_ == rhs.batch_idx_;
+    return lhs.dataloader_step_ == rhs.dataloader_step_;
+}
+
+size_t DataLoaderIterator::DataLoaderStep() const { return dataloader_step_; }
+
+DataLoaderIterator &DataLoaderIterator::SeekDataLoaderStep(size_t dataloader_step) {
+    CHECK_LE(dataloader_step, num_dataloader_steps_)
+        << "Cannot seek past DataLoader end. dataloader_step=" << dataloader_step
+        << ", num_dataloader_steps=" << num_dataloader_steps_;
+    dataloader_step_ = dataloader_step;
+    return *this;
 }
 
 DataLoader::DataLoader(const std::shared_ptr<Dataset> &dataset, size_t batch_size)
-    : dataset_(dataset), batch_size_(batch_size), max_batch_idx_((dataset_->Size() + batch_size_ - 1) / batch_size_) {}
+    : dataset_(dataset), batch_size_(batch_size) {
+    CHECK(dataset_ != nullptr) << "DataLoader dataset must not be null";
+    CHECK_GT(batch_size_, 0) << "DataLoader batch_size must be greater than zero";
+    num_dataloader_steps_ = CheckedCeilDiv(dataset_->Size(), batch_size_);
+}
 
-DataLoaderIterator DataLoader::begin() const { return DataLoaderIterator(*dataset_, batch_size_, 0, max_batch_idx_); }
+DataLoaderIterator DataLoader::begin() const {
+    return DataLoaderIterator(*dataset_, batch_size_, 0, num_dataloader_steps_, 0, 1);
+}
 
 DataLoaderIterator DataLoader::end() const {
-    return DataLoaderIterator(*dataset_, batch_size_, max_batch_idx_, max_batch_idx_);
+    return DataLoaderIterator(*dataset_, batch_size_, num_dataloader_steps_, num_dataloader_steps_, 0, 1);
 }
 
+size_t DataLoader::NumDataLoaderSteps() const { return num_dataloader_steps_; }
+
 DistributedDataLoader::DistributedDataLoader(const std::shared_ptr<Dataset> &dataset, size_t batch_size,
                                              size_t ddp_rank, size_t ddp_world_size)
-    : DataLoader(dataset, batch_size), ddp_rank_(ddp_rank), ddp_world_size_(ddp_world_size) {}
+    : DataLoader(dataset, batch_size), ddp_rank_(ddp_rank), ddp_world_size_(ddp_world_size) {
+    CHECK_GT(ddp_world_size_, 0);
+    CHECK_LT(ddp_rank_, ddp_world_size_);
+    const size_t samples_per_dataloader_step = ddp_world_size_ * batch_size_;
+    CHECK_GE(dataset_->Size(), samples_per_dataloader_step)
+        << "DistributedDataLoader needs enough samples for one DataLoader step. dataset_size=" << dataset_->Size()
+        << ", samples_per_dataloader_step=" << samples_per_dataloader_step << " (" << batch_size_ << " per rank * "
+        << ddp_world_size_ << " ranks). Reduce batch size/world size or use a larger dataset.";
+    num_dataloader_steps_ = dataset_->Size() / samples_per_dataloader_step;
+}
 
 DataLoaderIterator DistributedDataLoader::begin() const {
-    return DataLoaderIterator(*dataset_, batch_size_, ddp_rank_, max_batch_idx_, ddp_rank_, ddp_world_size_);
+    return DataLoaderIterator(*dataset_, batch_size_, 0, num_dataloader_steps_, ddp_rank_, ddp_world_size_);
 }
 
 DataLoaderIterator DistributedDataLoader::end() const {
-    return DataLoaderIterator(*dataset_, batch_size_, max_batch_idx_, max_batch_idx_, ddp_rank_, ddp_world_size_);
+    return DataLoaderIterator(*dataset_, batch_size_, num_dataloader_steps_, num_dataloader_steps_, ddp_rank_,
+                              ddp_world_size_);
 }
 } // namespace infini_train

tests/CMakeLists.txt

@@ -12,6 +12,9 @@ add_subdirectory(distributed)
 # Module tests
 add_subdirectory(module)
 
+# DataLoader tests
+add_subdirectory(dataloader)
+
 # Tensor tests
 add_subdirectory(tensor)
 

tests/checkpoint/test_trainer_state.cc

@@ -105,18 +105,18 @@ TEST_P(TrainerStateTest, RoundTrip) {
     std::filesystem::remove_all(dir);
 }
 
-TEST_P(TrainerStateTest, DataLoaderSkipUsesCurrentBatchConfiguration) {
-    EXPECT_EQ(DataLoaderBatchesToSkip(/*consumed_train_samples=*/400, /*local_batch_size=*/4,
-                                      /*ddp_world_size=*/2),
+TEST_P(TrainerStateTest, DataLoaderStepsToSkipUsesCurrentBatchConfiguration) {
+    EXPECT_EQ(DataLoaderStepsToSkip(/*consumed_train_samples=*/400, /*local_batch_size=*/4,
+                                    /*ddp_world_size=*/2),
               50);
-    EXPECT_EQ(DataLoaderBatchesToSkip(/*consumed_train_samples=*/400, /*local_batch_size=*/8,
-                                      /*ddp_world_size=*/2),
+    EXPECT_EQ(DataLoaderStepsToSkip(/*consumed_train_samples=*/400, /*local_batch_size=*/8,
+                                    /*ddp_world_size=*/2),
               25);
 }
 
-TEST_P(TrainerStateTest, DataLoaderSkipRejectsUnalignedConfiguration) {
-    EXPECT_DEATH(DataLoaderBatchesToSkip(/*consumed_train_samples=*/400, /*local_batch_size=*/6,
-                                         /*ddp_world_size=*/4),
+TEST_P(TrainerStateTest, DataLoaderStepsToSkipRejectsUnalignedConfiguration) {
+    EXPECT_DEATH(DataLoaderStepsToSkip(/*consumed_train_samples=*/400, /*local_batch_size=*/6,
+                                       /*ddp_world_size=*/4),
                  "does not align");
 }
 

tests/dataloader/CMakeLists.txt

@@ -0,0 +1,8 @@
+# ==========================================================================
+# DataLoader tests
+# ==========================================================================
+
+infini_train_add_test(test_dataloader
+  SOURCES test_dataloader.cc
+  LABELS cpu
+)