| addComponent(ComponentPtr component) | Mila::Dnn::CompositeComponent< TDeviceType, TPrecision > | inline |
| backward(const TokenIndexType &, const TensorType &) override (defined in Mila::Dnn::GemmaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy, kDelegatedFeedForward, kMixtureOfExperts >) | Mila::Dnn::GemmaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy, kDelegatedFeedForward, kMixtureOfExperts > | inline |
| Mila::Dnn::LanguageModelNetwork::backward(const TokenIndexType &input, const TensorType &output_grad)=0 | Mila::Dnn::LanguageModelNetwork< TDeviceType, TPrecision > | pure virtual |
| 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 > | |
| componentPaths() const | Mila::Dnn::CompositeComponent< TDeviceType, TPrecision > | inline |
| ComponentPtr typedef (defined in Mila::Dnn::GemmaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy, kDelegatedFeedForward, kMixtureOfExperts >) | Mila::Dnn::GemmaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy, kDelegatedFeedForward, kMixtureOfExperts > | |
| CompositeBase typedef (defined in Mila::Dnn::Network< TDeviceType, TPrecision >) | Mila::Dnn::Network< 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 > | |
| createOptimizer(const TConfig &config) | Mila::Dnn::Network< TDeviceType, TPrecision > | inline |
| decode(const TokenIndexType &input, dim_t position) override (defined in Mila::Dnn::GemmaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy, kDelegatedFeedForward, kMixtureOfExperts >) | Mila::Dnn::GemmaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy, kDelegatedFeedForward, kMixtureOfExperts > | inline |
| Mila::Dnn::LanguageModelNetwork::decode(const TokenIndexType &input, dim_t position)=0 | Mila::Dnn::LanguageModelNetwork< TDeviceType, TPrecision > | pure virtual |
| findComponent(const std::string &path) const | Mila::Dnn::CompositeComponent< TDeviceType, TPrecision > | inline |
| forward(const TokenIndexType &) override (defined in Mila::Dnn::GemmaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy, kDelegatedFeedForward, kMixtureOfExperts >) | Mila::Dnn::GemmaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy, kDelegatedFeedForward, kMixtureOfExperts > | inline |
| Mila::Dnn::LanguageModelNetwork::forward(const TokenIndexType &input)=0 | Mila::Dnn::LanguageModelNetwork< TDeviceType, TPrecision > | pure virtual |
| GemmaTransformer(const std::string &name, const GemmaConfig &config, DeviceId device_id) (defined in Mila::Dnn::GemmaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy, kDelegatedFeedForward, kMixtureOfExperts >) | Mila::Dnn::GemmaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy, kDelegatedFeedForward, kMixtureOfExperts > | 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 noexcept override | Mila::Dnn::Network< TDeviceType, TPrecision > | inlinevirtual |
| getDeviceType() | Mila::Dnn::Component< TDeviceType, TPrecision > | inlinestatic |
| getExecutionContext() const | Mila::Dnn::Component< TDeviceType, TPrecision > | inline |
| getGradients() const override | Mila::Dnn::CompositeComponent< TDeviceType, TPrecision > | inlinevirtual |
| getMemoryStats() const override | Mila::Dnn::GemmaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy, kDelegatedFeedForward, kMixtureOfExperts > | inlinevirtual |
| getModelType() const (defined in Mila::Dnn::GemmaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy, kDelegatedFeedForward, kMixtureOfExperts >) | Mila::Dnn::GemmaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy, kDelegatedFeedForward, kMixtureOfExperts > | inline |
| getName() const | Mila::Dnn::Component< TDeviceType, TPrecision > | inline |
| getObservableStages() const | Mila::Dnn::Component< TDeviceType, TPrecision > | inlinevirtual |
| getObservedPasses() const noexcept (defined in Mila::Dnn::Component< TDeviceType, TPrecision >) | Mila::Dnn::Component< TDeviceType, TPrecision > | inline |
| getOutputs() const | Mila::Dnn::Component< TDeviceType, TPrecision > | inlinevirtual |
| getParameterNames() const | Mila::Dnn::Component< TDeviceType, TPrecision > | 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::GemmaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy, kDelegatedFeedForward, kMixtureOfExperts > | inlinevirtual |
| getRuntimeMode() const noexcept (defined in Mila::Dnn::Component< TDeviceType, TPrecision >) | Mila::Dnn::Component< TDeviceType, TPrecision > | inline |
| getTrainingMode() const noexcept | Mila::Dnn::Component< TDeviceType, TPrecision > | inline |
| getType() const override | Mila::Dnn::Network< TDeviceType, TPrecision > | inlinevirtual |
| GlobalBlockType typedef (defined in Mila::Dnn::GemmaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy, kDelegatedFeedForward, kMixtureOfExperts >) | Mila::Dnn::GemmaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy, kDelegatedFeedForward, kMixtureOfExperts > | |
| 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 |
| HostStagingMemoryResource typedef | Mila::Dnn::Component< TDeviceType, TPrecision > | protected |
| isBuilt() const final | Mila::Dnn::Component< TDeviceType, TPrecision > | inlinevirtual |
| isInferenceMode() const noexcept (defined in Mila::Dnn::Component< TDeviceType, TPrecision >) | Mila::Dnn::Component< TDeviceType, TPrecision > | inline |
| isTrainingMode() const noexcept (defined in Mila::Dnn::Component< TDeviceType, TPrecision >) | Mila::Dnn::Component< TDeviceType, TPrecision > | inline |
| LanguageModelNetwork(const std::string &name) (defined in Mila::Dnn::LanguageModelNetwork< TDeviceType, TPrecision >) | Mila::Dnn::LanguageModelNetwork< TDeviceType, TPrecision > | inlineexplicit |
| LmHeadLinearType typedef (defined in Mila::Dnn::GemmaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy, kDelegatedFeedForward, kMixtureOfExperts >) | Mila::Dnn::GemmaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy, kDelegatedFeedForward, kMixtureOfExperts > | |
| load(ModelArchive &archive, SerializationMode mode) | Mila::Dnn::Network< TDeviceType, TPrecision > | inline |
| load_(ModelArchive &archive, SerializationMode mode) override | Mila::Dnn::Network< TDeviceType, TPrecision > | inlineprotectedvirtual |
| loadParameter(const std::string &, const Serialization::ITensorBlob &) | Mila::Dnn::Component< TDeviceType, TPrecision > | 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 |
| loadParameters(WeightsReader &reader) (defined in Mila::Dnn::GemmaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy, kDelegatedFeedForward, kMixtureOfExperts >) | Mila::Dnn::GemmaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy, kDelegatedFeedForward, kMixtureOfExperts > | inline |
| LocalBlockType typedef (defined in Mila::Dnn::GemmaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy, kDelegatedFeedForward, kMixtureOfExperts >) | Mila::Dnn::GemmaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy, kDelegatedFeedForward, kMixtureOfExperts > | |
| MR typedef (defined in Mila::Dnn::GemmaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy, kDelegatedFeedForward, kMixtureOfExperts >) | Mila::Dnn::GemmaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy, kDelegatedFeedForward, kMixtureOfExperts > | |
| Network(const std::string &name) | Mila::Dnn::Network< TDeviceType, TPrecision > | inlineexplicit |
| NetworkBase typedef (defined in Mila::Dnn::GemmaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy, kDelegatedFeedForward, kMixtureOfExperts >) | Mila::Dnn::GemmaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy, kDelegatedFeedForward, kMixtureOfExperts > | |
| observe(std::string_view pattern, Compute::ComputePassMask passes, Compute::ActivationObserver sink) | Mila::Dnn::CompositeComponent< TDeviceType, TPrecision > | inline |
| onBuilding(const BuildContext &context) override | Mila::Dnn::GemmaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy, kDelegatedFeedForward, kMixtureOfExperts > | inlineprotectedvirtual |
| onExecutionContextSet() override | Mila::Dnn::CompositeComponent< TDeviceType, TPrecision > | inlineprotectedvirtual |
| onTrainingModeChanging(TrainingMode training_mode) override | Mila::Dnn::GemmaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy, kDelegatedFeedForward, kMixtureOfExperts > | 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 |
| parameterCount() const override | Mila::Dnn::GemmaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy, kDelegatedFeedForward, kMixtureOfExperts > | inlinevirtual |
| prefill(const TokenIndexType &input) override (defined in Mila::Dnn::GemmaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy, kDelegatedFeedForward, kMixtureOfExperts >) | Mila::Dnn::GemmaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy, kDelegatedFeedForward, kMixtureOfExperts > | inline |
| Mila::Dnn::LanguageModelNetwork::prefill(const TokenIndexType &input)=0 | Mila::Dnn::LanguageModelNetwork< TDeviceType, TPrecision > | pure virtual |
| prefillChunking(int64_t B, int64_t T_ctx) const | Mila::Dnn::GemmaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy, kDelegatedFeedForward, kMixtureOfExperts > | inline |
| prefillFrom(const TokenIndexType &input, dim_t start_offset) override | Mila::Dnn::GemmaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy, kDelegatedFeedForward, kMixtureOfExperts > | inline |
| Mila::Dnn::LanguageModelNetwork::prefillFrom(const TokenIndexType &input, dim_t start_offset) | Mila::Dnn::LanguageModelNetwork< TDeviceType, TPrecision > | inlinevirtual |
| publish(Compute::ComputePass pass, std::string_view stage, const ITensor &value) const | Mila::Dnn::Component< TDeviceType, TPrecision > | inlineprotected |
| removeComponent(const std::string &name) | Mila::Dnn::CompositeComponent< TDeviceType, TPrecision > | inline |
| requireSerializableParameters() const override | Mila::Dnn::CompositeComponent< TDeviceType, TPrecision > | inlineprotectedvirtual |
| resetKvCache() (defined in Mila::Dnn::GemmaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy, kDelegatedFeedForward, kMixtureOfExperts >) | Mila::Dnn::GemmaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy, kDelegatedFeedForward, kMixtureOfExperts > | inline |
| rewindKvCache(dim_t position) override | Mila::Dnn::GemmaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy, kDelegatedFeedForward, kMixtureOfExperts > | inlinevirtual |
| RmsNormType typedef (defined in Mila::Dnn::GemmaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy, kDelegatedFeedForward, kMixtureOfExperts >) | Mila::Dnn::GemmaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy, kDelegatedFeedForward, kMixtureOfExperts > | |
| save(ModelArchive &archive, SerializationMode mode) const | Mila::Dnn::Network< TDeviceType, TPrecision > | inline |
| save_(ModelArchive &archive, SerializationMode) const override | Mila::Dnn::GemmaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy, kDelegatedFeedForward, kMixtureOfExperts > | inlineprotectedvirtual |
| saveFlatTensors(Serialization::SafeTensorsWriter &writer, const std::string &prefix, Serialization::TensorSavePass pass) const override | Mila::Dnn::CompositeComponent< TDeviceType, TPrecision > | 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 |
| scoreTokens(const TokenIndexType &input) | Mila::Dnn::LanguageModelNetwork< TDeviceType, TPrecision > | inlinevirtual |
| setExecutionContext(IExecutionContext *context) | Mila::Dnn::Component< TDeviceType, TPrecision > | inlineprotected |
| setObservedPasses(Compute::ComputePassMask passes) noexcept | Mila::Dnn::Component< TDeviceType, TPrecision > | inline |
| setTrainingMode(TrainingMode mode) | Mila::Dnn::Component< TDeviceType, TPrecision > | inline |
| stopObserving() | Mila::Dnn::CompositeComponent< TDeviceType, TPrecision > | inline |
| synchronize() override | Mila::Dnn::Network< TDeviceType, TPrecision > | inlinevirtual |
| TableQuantizationPolicy typedef (defined in Mila::Dnn::GemmaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy, kDelegatedFeedForward, kMixtureOfExperts >) | Mila::Dnn::GemmaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy, kDelegatedFeedForward, kMixtureOfExperts > | |
| TensorType typedef (defined in Mila::Dnn::GemmaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy, kDelegatedFeedForward, kMixtureOfExperts >) | Mila::Dnn::GemmaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy, kDelegatedFeedForward, kMixtureOfExperts > | |
| TokenEmbeddingType typedef (defined in Mila::Dnn::GemmaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy, kDelegatedFeedForward, kMixtureOfExperts >) | Mila::Dnn::GemmaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy, kDelegatedFeedForward, kMixtureOfExperts > | |
| TokenIndexType typedef (defined in Mila::Dnn::GemmaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy, kDelegatedFeedForward, kMixtureOfExperts >) | Mila::Dnn::GemmaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy, kDelegatedFeedForward, kMixtureOfExperts > | |
| toString() const override | Mila::Dnn::GemmaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy, kDelegatedFeedForward, kMixtureOfExperts > | inlinevirtual |
| TransformerBlockType typedef (defined in Mila::Dnn::GemmaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy, kDelegatedFeedForward, kMixtureOfExperts >) | Mila::Dnn::GemmaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy, kDelegatedFeedForward, kMixtureOfExperts > | |
| tryFindComponent(const std::string &path) const | Mila::Dnn::CompositeComponent< TDeviceType, TPrecision > | inline |
| verifyArchitectureCompatibility(const WeightsMetadata &metadata) | Mila::Dnn::Network< TDeviceType, TPrecision > | inlineprotected |
| 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 |
| ~GemmaTransformer() override=default (defined in Mila::Dnn::GemmaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy, kDelegatedFeedForward, kMixtureOfExperts >) | Mila::Dnn::GemmaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy, kDelegatedFeedForward, kMixtureOfExperts > | |
| ~LanguageModelNetwork() override=default (defined in Mila::Dnn::LanguageModelNetwork< TDeviceType, TPrecision >) | Mila::Dnn::LanguageModelNetwork< TDeviceType, TPrecision > | |
| ~Network() override=default (defined in Mila::Dnn::Network< TDeviceType, TPrecision >) | Mila::Dnn::Network< TDeviceType, TPrecision > | |
1.17.0