Remove shared_ptr wrapping the AlignedMemory of models.

browsermt · Jul 30, 2023 · eef17fc · eef17fc
1 parent 4081ec0
commit eef17fc
Show file tree

Hide file tree

Showing 5 changed files with 7 additions and 7 deletions.
diff --git a/src/translator/byte_array_util.cpp b/src/translator/byte_array_util.cpp
@@ -91,14 +91,14 @@ AlignedMemory loadFileToMemory(const std::string& path, size_t alignment) {
   return alignedMemory;
 }
 
-std::vector<std::shared_ptr<AlignedMemory>> getModelMemoryFromConfig(marian::Ptr<marian::Options> options) {
+std::vector<AlignedMemory> getModelMemoryFromConfig(marian::Ptr<marian::Options> options) {
   auto models = options->get<std::vector<std::string>>("models");
 
-  std::vector<std::shared_ptr<AlignedMemory>> modelMemories(models.size());
+  std::vector<AlignedMemory> modelMemories(models.size());
   for (size_t i = 0; i < models.size(); ++i) {
     const auto model = models[i];
     if (marian::io::isBin(model)) {
-      modelMemories[i] = std::make_shared<AlignedMemory>(loadFileToMemory(model, 256));
+      modelMemories[i] = loadFileToMemory(model, 256);
     } else if (marian::io::isNpz(model)) {
       // if any of the models are npz format, we revert to loading from file for all models.
       LOG(debug, "Encountered an npz file {}; will use file loading for {} models", model, models.size());

diff --git a/src/translator/byte_array_util.h b/src/translator/byte_array_util.h
@@ -5,7 +5,7 @@ namespace marian {
 namespace bergamot {
 
 AlignedMemory loadFileToMemory(const std::string& path, size_t alignment);
-std::vector<std::shared_ptr<AlignedMemory>> getModelMemoryFromConfig(marian::Ptr<marian::Options> options);
+std::vector<AlignedMemory> getModelMemoryFromConfig(marian::Ptr<marian::Options> options);
 AlignedMemory getQualityEstimatorModel(const marian::Ptr<marian::Options>& options);
 AlignedMemory getQualityEstimatorModel(MemoryBundle& memoryBundle, const marian::Ptr<marian::Options>& options);
 AlignedMemory getShortlistMemoryFromConfig(marian::Ptr<marian::Options> options);

diff --git a/src/translator/definitions.h b/src/translator/definitions.h
@@ -19,7 +19,7 @@ typedef AlignedVector<char> AlignedMemory;
 /// Memory bundle for all byte-arrays.
 /// Can be a set/subset of model, shortlist, vocabs and ssplitPrefixFile bytes.
 struct MemoryBundle {
-  std::vector<std::shared_ptr<AlignedMemory>> models{};  ///< Byte-array of model (aligned to 256)
+  std::vector<AlignedMemory> models{};  ///< Byte-array of model (each element is aligned to 256)
   AlignedMemory shortlist{};                             ///< Byte-array of shortlist (aligned to 64)
 
   /// Vector of vocabulary memories (aligned to 64).

diff --git a/src/translator/translation_model.cpp b/src/translator/translation_model.cpp
@@ -66,7 +66,7 @@ void TranslationModel::loadBackend(size_t idx) {
     const std::vector<const void *> container = std::invoke([&]() {
       std::vector<const void *> model_ptrs(memory_.models.size());
       for (size_t i = 0; i < memory_.models.size(); ++i) {
-        const AlignedMemory &model = *memory_.models[i];
+        const AlignedMemory &model = memory_.models[i];
 
         ABORT_IF(model.size() == 0 || model.begin() == nullptr, "The provided memory is empty. Cannot load the model.");
         ABORT_IF(

diff --git a/wasm/bindings/service_bindings.cpp b/wasm/bindings/service_bindings.cpp
@@ -48,7 +48,7 @@ MemoryBundle prepareMemoryBundle(AlignedMemory* modelMemory, AlignedMemory* shor
                                  std::vector<AlignedMemory*> uniqueVocabsMemories,
                                  AlignedMemory* qualityEstimatorMemory) {
   MemoryBundle memoryBundle;
-  memoryBundle.models.emplace_back(std::make_shared<AlignedMemory>(std::move(*modelMemory)));
+  memoryBundle.models.emplace_back(std::move(*modelMemory));
   memoryBundle.shortlist = std::move(*shortlistMemory);
   memoryBundle.vocabs = std::move(prepareVocabsSmartMemories(uniqueVocabsMemories));
   if (qualityEstimatorMemory != nullptr) {