| addComponent(ComponentPtr component) | Mila::Dnn::CompositeComponent< TDeviceType, TPrecision > | inline |
| AttentionType typedef (defined in Mila::Dnn::GemmaBlock< TDeviceType, TPrecision, kGlobal, TWeightQuant, TKvPolicy >) | Mila::Dnn::GemmaBlock< TDeviceType, TPrecision, kGlobal, TWeightQuant, TKvPolicy > | |
| build(const BuildContext &context) final | Mila::Dnn::Component< TDeviceType, TPrecision > | inlinevirtual |
| build_context_ | Mila::Dnn::Component< TDeviceType, TPrecision > | protected |
| childCount() const noexcept | Mila::Dnn::CompositeComponent< TDeviceType, TPrecision > | inline |
| clearComponents() | Mila::Dnn::CompositeComponent< TDeviceType, TPrecision > | inline |
| Component(const std::string &name) | Mila::Dnn::Component< TDeviceType, TPrecision > | inlineexplicit |
| ComponentBase typedef (defined in Mila::Dnn::CompositeComponent< TDeviceType, TPrecision >) | Mila::Dnn::CompositeComponent< TDeviceType, TPrecision > | |
| ComponentPtr typedef (defined in Mila::Dnn::CompositeComponent< TDeviceType, TPrecision >) | Mila::Dnn::CompositeComponent< TDeviceType, TPrecision > | |
| CompositeComponent(const std::string &name) | Mila::Dnn::CompositeComponent< TDeviceType, TPrecision > | inlineexplicit |
| CompositeComponent(const CompositeComponent &)=delete (defined in Mila::Dnn::CompositeComponent< TDeviceType, TPrecision >) | Mila::Dnn::CompositeComponent< TDeviceType, TPrecision > | |
| CompositeComponent(CompositeComponent &&) noexcept=default (defined in Mila::Dnn::CompositeComponent< TDeviceType, TPrecision >) | Mila::Dnn::CompositeComponent< TDeviceType, TPrecision > | |
| CompositeComponentBase typedef (defined in Mila::Dnn::GemmaBlock< TDeviceType, TPrecision, kGlobal, TWeightQuant, TKvPolicy >) | Mila::Dnn::GemmaBlock< TDeviceType, TPrecision, kGlobal, TWeightQuant, TKvPolicy > | |
| decode(const TensorType &input, dim_t position) override | Mila::Dnn::GemmaBlock< TDeviceType, TPrecision, kGlobal, TWeightQuant, TKvPolicy > | inlinevirtual |
| findComponent(const std::string &path) const | Mila::Dnn::CompositeComponent< TDeviceType, TPrecision > | inline |
| GeGLUType typedef (defined in Mila::Dnn::GemmaBlock< TDeviceType, TPrecision, kGlobal, TWeightQuant, TKvPolicy >) | Mila::Dnn::GemmaBlock< TDeviceType, TPrecision, kGlobal, TWeightQuant, TKvPolicy > | |
| GemmaBlock(const std::string &name, const GemmaConfig &config, std::optional< DeviceId > device_id=std::nullopt) (defined in Mila::Dnn::GemmaBlock< TDeviceType, TPrecision, kGlobal, TWeightQuant, TKvPolicy >) | Mila::Dnn::GemmaBlock< TDeviceType, TPrecision, kGlobal, TWeightQuant, TKvPolicy > | inlineexplicit |
| getComponent(const std::string &name) const | Mila::Dnn::CompositeComponent< TDeviceType, TPrecision > | inline |
| getComponentAs(const std::string &name) const | Mila::Dnn::CompositeComponent< TDeviceType, TPrecision > | inlineprotected |
| getComponents() const | Mila::Dnn::CompositeComponent< TDeviceType, TPrecision > | inline |
| getDeviceId() const override | Mila::Dnn::CompositeComponent< TDeviceType, TPrecision > | inlinevirtual |
| getDeviceType() | Mila::Dnn::Component< TDeviceType, TPrecision > | inlinestatic |
| getExecutionContext() const | Mila::Dnn::Component< TDeviceType, TPrecision > | inlineprotected |
| getGradients() const override | Mila::Dnn::CompositeComponent< TDeviceType, TPrecision > | inlinevirtual |
| getMemoryStats() const override | Mila::Dnn::GemmaBlock< TDeviceType, TPrecision, kGlobal, TWeightQuant, TKvPolicy > | inlinevirtual |
| getName() const | Mila::Dnn::Component< TDeviceType, TPrecision > | inline |
| getParameterNames() const override | Mila::Dnn::GemmaBlock< TDeviceType, TPrecision, kGlobal, TWeightQuant, TKvPolicy > | inlinevirtual |
| getParameters() const override | Mila::Dnn::CompositeComponent< TDeviceType, TPrecision > | inlinevirtual |
| getPrecision() noexcept | Mila::Dnn::Component< TDeviceType, TPrecision > | inlinestatic |
| getRequiredMemory(const BuildContext &context) const override | Mila::Dnn::GemmaBlock< TDeviceType, TPrecision, kGlobal, TWeightQuant, TKvPolicy > | inlinevirtual |
| getTrainingMode() const noexcept | Mila::Dnn::Component< TDeviceType, TPrecision > | inline |
| getType() const override | Mila::Dnn::GemmaBlock< TDeviceType, TPrecision, kGlobal, TWeightQuant, TKvPolicy > | inlinevirtual |
| hasChildren() const noexcept | Mila::Dnn::CompositeComponent< TDeviceType, TPrecision > | inline |
| hasComponent(const std::string &name) const | Mila::Dnn::CompositeComponent< TDeviceType, TPrecision > | inline |
| hasExecutionContext() const noexcept | Mila::Dnn::Component< TDeviceType, TPrecision > | inlineprotected |
| headDim() const noexcept (defined in Mila::Dnn::GemmaBlock< TDeviceType, TPrecision, kGlobal, TWeightQuant, TKvPolicy >) | Mila::Dnn::GemmaBlock< TDeviceType, TPrecision, kGlobal, TWeightQuant, TKvPolicy > | inline |
| HostStagingMemoryResource typedef | Mila::Dnn::Component< TDeviceType, TPrecision > | protected |
| installSharedWorkspace(const GemmaBlockWorkspace< TDeviceType, TPrecision > &workspace) | Mila::Dnn::GemmaBlock< TDeviceType, TPrecision, kGlobal, TWeightQuant, TKvPolicy > | inline |
| isBuilt() const final | Mila::Dnn::Component< TDeviceType, TPrecision > | inlinevirtual |
| isGlobal() noexcept (defined in Mila::Dnn::GemmaBlock< TDeviceType, TPrecision, kGlobal, TWeightQuant, TKvPolicy >) | Mila::Dnn::GemmaBlock< TDeviceType, TPrecision, kGlobal, TWeightQuant, TKvPolicy > | inlinestatic |
| keyEqualsValue() const noexcept (defined in Mila::Dnn::GemmaBlock< TDeviceType, TPrecision, kGlobal, TWeightQuant, TKvPolicy >) | Mila::Dnn::GemmaBlock< TDeviceType, TPrecision, kGlobal, TWeightQuant, TKvPolicy > | inline |
| kvProjWidth() const noexcept (defined in Mila::Dnn::GemmaBlock< TDeviceType, TPrecision, kGlobal, TWeightQuant, TKvPolicy >) | Mila::Dnn::GemmaBlock< TDeviceType, TPrecision, kGlobal, TWeightQuant, TKvPolicy > | inline |
| LinearType typedef (defined in Mila::Dnn::GemmaBlock< TDeviceType, TPrecision, kGlobal, TWeightQuant, TKvPolicy >) | Mila::Dnn::GemmaBlock< TDeviceType, TPrecision, kGlobal, TWeightQuant, TKvPolicy > | |
| load_(ModelArchive &archive, SerializationMode mode) override | Mila::Dnn::GemmaBlock< TDeviceType, TPrecision, kGlobal, TWeightQuant, TKvPolicy > | inlinevirtual |
| loadParameter(const std::string &name, const ITensorBlob &blob) override | Mila::Dnn::GemmaBlock< TDeviceType, TPrecision, kGlobal, TWeightQuant, TKvPolicy > | inlinevirtual |
| loadParameterFromBlob(const std::string ¶m_name, const Serialization::ITensorBlob &blob, Tensor< TParameterPrecision, TMemoryResource > &target, const shape_t &expected_shape) | Mila::Dnn::Component< TDeviceType, TPrecision > | inlineprotected |
| MR typedef (defined in Mila::Dnn::GemmaBlock< TDeviceType, TPrecision, kGlobal, TWeightQuant, TKvPolicy >) | Mila::Dnn::GemmaBlock< TDeviceType, TPrecision, kGlobal, TWeightQuant, TKvPolicy > | |
| numKVHeads() const noexcept (defined in Mila::Dnn::GemmaBlock< TDeviceType, TPrecision, kGlobal, TWeightQuant, TKvPolicy >) | Mila::Dnn::GemmaBlock< TDeviceType, TPrecision, kGlobal, TWeightQuant, TKvPolicy > | inline |
| onBuilding(const BuildContext &context) override | Mila::Dnn::GemmaBlock< TDeviceType, TPrecision, kGlobal, TWeightQuant, TKvPolicy > | inlineprotectedvirtual |
| onExecutionContextSet() override | Mila::Dnn::CompositeComponent< TDeviceType, TPrecision > | inlineprotectedvirtual |
| onTrainingModeChanging(TrainingMode training_mode) override | Mila::Dnn::GemmaBlock< TDeviceType, TPrecision, kGlobal, TWeightQuant, TKvPolicy > | inlineprotectedvirtual |
| operator=(const CompositeComponent &)=delete (defined in Mila::Dnn::CompositeComponent< TDeviceType, TPrecision >) | Mila::Dnn::CompositeComponent< TDeviceType, TPrecision > | |
| operator=(CompositeComponent &&) noexcept=default (defined in Mila::Dnn::CompositeComponent< TDeviceType, TPrecision >) | Mila::Dnn::CompositeComponent< TDeviceType, TPrecision > | |
| optimize() | Mila::Dnn::CompositeComponent< TDeviceType, TPrecision > | inlineprotectedvirtual |
| packedQKVWidth() const noexcept (defined in Mila::Dnn::GemmaBlock< TDeviceType, TPrecision, kGlobal, TWeightQuant, TKvPolicy >) | Mila::Dnn::GemmaBlock< TDeviceType, TPrecision, kGlobal, TWeightQuant, TKvPolicy > | inline |
| parameterCount() const override | Mila::Dnn::CompositeComponent< TDeviceType, TPrecision > | inlinevirtual |
| prefill(const TensorType &input, dim_t position_offset) override | Mila::Dnn::GemmaBlock< TDeviceType, TPrecision, kGlobal, TWeightQuant, TKvPolicy > | inlinevirtual |
| qProjWidth() const noexcept (defined in Mila::Dnn::GemmaBlock< TDeviceType, TPrecision, kGlobal, TWeightQuant, TKvPolicy >) | Mila::Dnn::GemmaBlock< TDeviceType, TPrecision, kGlobal, TWeightQuant, TKvPolicy > | inline |
| removeComponent(const std::string &name) | Mila::Dnn::CompositeComponent< TDeviceType, TPrecision > | inline |
| requireSerializableParameters() const override | Mila::Dnn::CompositeComponent< TDeviceType, TPrecision > | inlineprotectedvirtual |
| resetKVCache() override | Mila::Dnn::GemmaBlock< TDeviceType, TPrecision, kGlobal, TWeightQuant, TKvPolicy > | inlinevirtual |
| ResidualType typedef (defined in Mila::Dnn::GemmaBlock< TDeviceType, TPrecision, kGlobal, TWeightQuant, TKvPolicy >) | Mila::Dnn::GemmaBlock< TDeviceType, TPrecision, kGlobal, TWeightQuant, TKvPolicy > | |
| resolveBlockBuildContexts(const BuildContext &context) const (defined in Mila::Dnn::GemmaBlock< TDeviceType, TPrecision, kGlobal, TWeightQuant, TKvPolicy >) | Mila::Dnn::GemmaBlock< TDeviceType, TPrecision, kGlobal, TWeightQuant, TKvPolicy > | inlineprotected |
| rewindKvCache(dim_t position) override | Mila::Dnn::GemmaBlock< TDeviceType, TPrecision, kGlobal, TWeightQuant, TKvPolicy > | inlinevirtual |
| RmsNormType typedef (defined in Mila::Dnn::GemmaBlock< TDeviceType, TPrecision, kGlobal, TWeightQuant, TKvPolicy >) | Mila::Dnn::GemmaBlock< TDeviceType, TPrecision, kGlobal, TWeightQuant, TKvPolicy > | |
| ropeTheta() const noexcept (defined in Mila::Dnn::GemmaBlock< TDeviceType, TPrecision, kGlobal, TWeightQuant, TKvPolicy >) | Mila::Dnn::GemmaBlock< TDeviceType, TPrecision, kGlobal, TWeightQuant, TKvPolicy > | inline |
| RopeType typedef (defined in Mila::Dnn::GemmaBlock< TDeviceType, TPrecision, kGlobal, TWeightQuant, TKvPolicy >) | Mila::Dnn::GemmaBlock< TDeviceType, TPrecision, kGlobal, TWeightQuant, TKvPolicy > | |
| rotaryDim() const noexcept (defined in Mila::Dnn::GemmaBlock< TDeviceType, TPrecision, kGlobal, TWeightQuant, TKvPolicy >) | Mila::Dnn::GemmaBlock< TDeviceType, TPrecision, kGlobal, TWeightQuant, TKvPolicy > | inline |
| save_(ModelArchive &archive, SerializationMode mode) const override (defined in Mila::Dnn::GemmaBlock< TDeviceType, TPrecision, kGlobal, TWeightQuant, TKvPolicy >) | Mila::Dnn::GemmaBlock< TDeviceType, TPrecision, kGlobal, TWeightQuant, TKvPolicy > | inlinevirtual |
| saveFlatTensors(Serialization::SafeTensorsWriter &writer, const std::string &prefix, Serialization::TensorSavePass pass) const override | Mila::Dnn::GemmaBlock< TDeviceType, TPrecision, kGlobal, TWeightQuant, TKvPolicy > | inlinevirtual |
| saveParameterToArchive(ModelArchive &archive, const std::string ¶meter_name, const Tensor< TParameterPrecision, TMemoryResource > ¶meter) const | Mila::Dnn::Component< TDeviceType, TPrecision > | inlineprotected |
| saveParameterToWriter(Serialization::SafeTensorsWriter &writer, const std::string &flat_name, const Tensor< TParameterPrecision, TMemoryResource > ¶meter, Serialization::TensorSavePass pass) const | Mila::Dnn::Component< TDeviceType, TPrecision > | inlineprotected |
| setExecutionContext(IExecutionContext *context) | Mila::Dnn::Component< TDeviceType, TPrecision > | inlineprotected |
| setState(const GqaState &state) override | Mila::Dnn::GemmaBlock< TDeviceType, TPrecision, kGlobal, TWeightQuant, TKvPolicy > | inlinevirtual |
| setTrainingMode(TrainingMode mode) | Mila::Dnn::Component< TDeviceType, TPrecision > | inline |
| setUseFlashDecode(bool enabled) (defined in Mila::Dnn::GemmaBlock< TDeviceType, TPrecision, kGlobal, TWeightQuant, TKvPolicy >) | Mila::Dnn::GemmaBlock< TDeviceType, TPrecision, kGlobal, TWeightQuant, TKvPolicy > | inline |
| setUseFlashPrefill(bool enabled) (defined in Mila::Dnn::GemmaBlock< TDeviceType, TPrecision, kGlobal, TWeightQuant, TKvPolicy >) | Mila::Dnn::GemmaBlock< TDeviceType, TPrecision, kGlobal, TWeightQuant, TKvPolicy > | inline |
| supportsKVCache() const noexcept override | Mila::Dnn::GemmaBlock< TDeviceType, TPrecision, kGlobal, TWeightQuant, TKvPolicy > | inlinevirtual |
| synchronize() override | Mila::Dnn::CompositeComponent< TDeviceType, TPrecision > | inlinevirtual |
| TensorType typedef (defined in Mila::Dnn::GemmaBlock< TDeviceType, TPrecision, kGlobal, TWeightQuant, TKvPolicy >) | Mila::Dnn::GemmaBlock< TDeviceType, TPrecision, kGlobal, TWeightQuant, TKvPolicy > | |
| toString() const override | Mila::Dnn::CompositeComponent< TDeviceType, TPrecision > | inlinevirtual |
| tryFindComponent(const std::string &path) const | Mila::Dnn::CompositeComponent< TDeviceType, TPrecision > | inline |
| window() const noexcept (defined in Mila::Dnn::GemmaBlock< TDeviceType, TPrecision, kGlobal, TWeightQuant, TKvPolicy >) | Mila::Dnn::GemmaBlock< TDeviceType, TPrecision, kGlobal, TWeightQuant, TKvPolicy > | inline |
| zeroGradients() | Mila::Dnn::Component< TDeviceType, TPrecision > | inlinevirtual |
| ~Component()=default (defined in Mila::Dnn::Component< TDeviceType, TPrecision >) | Mila::Dnn::Component< TDeviceType, TPrecision > | virtual |
| ~CompositeComponent()=default (defined in Mila::Dnn::CompositeComponent< TDeviceType, TPrecision >) | Mila::Dnn::CompositeComponent< TDeviceType, TPrecision > | virtual |
| ~GemmaBlock() override=default (defined in Mila::Dnn::GemmaBlock< TDeviceType, TPrecision, kGlobal, TWeightQuant, TKvPolicy >) | Mila::Dnn::GemmaBlock< TDeviceType, TPrecision, kGlobal, TWeightQuant, TKvPolicy > | |
| ~IDecoderLayer()=default (defined in Mila::Dnn::IDecoderLayer< TDeviceType, TPrecision >) | Mila::Dnn::IDecoderLayer< TDeviceType, TPrecision > | virtual |