Class Hierarchy

This inheritance list is sorted roughly, but not completely, alphabetically:

CMila::Dnn::AxisPartition	Information about axis partitioning of a tensor
CMila::Dnn::GemmaBlock< TDeviceType, TPrecision, kGlobal, TWeightQuant, TKvPolicy, kDelegatedFeedForward, kMixtureOfExperts >::BlockBuildContexts	The per-child build contexts and split-scratch geometry this block implies
CMila::Dnn::LlamaBlock< TDeviceType, TPrecision, TWeightQuant, TKvPolicy >::BlockBuildContexts	The per-child build contexts this block implies
CMila::Dnn::QwenAttentionBlock< TDeviceType, TPrecision, TWeightPlan, TKvPolicy >::BlockBuildContexts	The per-child build contexts and split-scratch geometry this block implies
CMila::Dnn::QwenDeltaNetBlock< TDeviceType, TPrecision, TWeightPlan >::BlockContexts
CMila::Data::BpeTrainer	Corpus accumulator and trainer for BPE vocabularies
CMila::Data::BpeVocabularyConfig	Configuration for the BPE vocabulary
CMila::Dnn::BufferedTokenStreamer< Sink, BufSize >	Buffers BufSize tokens before forwarding a contiguous span to Sink
CMila::Dnn::BuildContext	Build-time context for Component::build()
CMila::Data::CharTrainer	Character-level tokenizer trainer
CMila::Data::CharVocabularyConfig	Configuration for Character-level tokenizer training
CMila::Distribution::CleanOptions
CMila::Dnn::Component< TDeviceType, TPrecision >	Abstract base class for neural network components
CMila::Dnn::Activation< TDeviceType, TPrecision, ActivationType::Silu >
CMila::Dnn::AttentionOutputGate< TDeviceType, TPrecision, ActivationType::Sigmoid >
CMila::Dnn::AttentionOutputGate< TDeviceType, TPrecision, ActivationType::Silu >
CMila::Dnn::Lpe< TDeviceType, dtype_t::INT32, TPrecision >
CMila::Dnn::MixtureOfExperts< TDeviceType, TPrecision, ActivationType::Gelu, TWeightQuant >
CMila::Dnn::Swiglu< TDeviceType, TPrecision, TGate >
CMila::Dnn::Swiglu< TDeviceType, TPrecision, ActivationType::Gelu >
CMila::Dnn::Swiglu< TDeviceType, TPrecision, ActivationType::Silu >
CMila::Dnn::TokenEmbedding< TDeviceType, dtype_t::INT32, TPrecision, TableQuantizationPolicy >
CMila::Dnn::TokenEmbedding< TDeviceType, dtype_t::INT32, TPrecision >
CMila::Dnn::TokenEmbedding< TDeviceType, dtype_t::INT32, TPrecision, typename PrecisionPlan::EmbeddingTable, kEmbeddingTableResidency >
CMila::Dnn::Activation< TDeviceType, TPrecision, TFn >	Unified elementwise activation component
CMila::Dnn::AttentionOutputGate< TDeviceType, TPrecision, TGate >	Elementwise gate over two same-shaped inputs: out = TGate(gate) * value
CMila::Dnn::CausalConv1d< TDeviceType, TPrecision >
CMila::Dnn::CompositeComponent< TDeviceType, TPrecision >	A component that contains and manages child components
CMila::Dnn::GatedMLP< TDeviceType, TPrecision, ActivationType::Gelu, TWeightQuant >
CMila::Dnn::GemmaBlock< TDeviceType, TPrecision, false, TWeightQuantization, TKvCachePolicy, kDelegatedFeedForward, kMixtureOfExperts >
CMila::Dnn::GemmaBlock< TDeviceType, TPrecision, true, TWeightQuantization, NoKvCompression, kDelegatedFeedForward, kMixtureOfExperts >
CMila::Dnn::LlamaBlock< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy >
CMila::Dnn::QwenAttentionBlock< TDeviceType, TPrecision, TWeightPlan, TKvCachePolicy >
CMila::Dnn::QwenDeltaNetBlock< TDeviceType, TPrecision, TWeightPlan >
CMila::Dnn::GatedMLP< TDeviceType, TPrecision, TGate, TWeightQuantization >	Gated feed-forward (GatedMLP) composite component
CMila::Dnn::GemmaBlock< TDeviceType, TPrecision, kGlobal, TWeightQuant, TKvPolicy, kDelegatedFeedForward, kMixtureOfExperts >	One Gemma 4 decoder block; kGlobal selects the global (full-attention) geometry
CMila::Dnn::GptBlock< TDeviceType, TPrecision >	Transformer encoder block as a composite component
CMila::Dnn::LlamaBlock< TDeviceType, TPrecision, TWeightQuant, TKvPolicy >
CMila::Dnn::MLP< TDeviceType, TPrecision >	Dense feed-forward (MLP) composite component
CMila::Dnn::Network< TDeviceType, TPrecision >	Root composite network container
CMila::Dnn::LanguageModelNetwork< TDeviceType, TPrecision >	The type erasure boundary between a language model and its transformer
CMila::Dnn::GemmaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy, kDelegatedFeedForward, kMixtureOfExperts >	Gemma 4 transformer (decoder-only) for autoregressive inference
CMila::Dnn::GptTransformer< TDeviceType, TPrecision >	GPT-2 style transformer (decoder-only) for autoregressive token prediction
CMila::Dnn::LlamaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy >	LLaMA-style transformer (decoder-only) for autoregressive token prediction
CMila::Dnn::QwenTransformer< TDeviceType, TPrecision, TWeightPlan, TKvCachePolicy >	Qwen 3.8 transformer (decoder-only) for autoregressive inference
CMila::Dnn::QwenAttentionBlock< TDeviceType, TPrecision, TWeightPlan, TKvPolicy >	One Qwen 3.8 full-attention decoder block
CMila::Dnn::QwenDeltaNetBlock< TDeviceType, TPrecision, TWeightPlan >	One Gated DeltaNet decoder block
CMila::Dnn::Router< TDeviceType, TPrecision >	Routes each token to its top_k experts
CMila::Dnn::GatedDeltaRule< TDeviceType, TPrecision >
CMila::Dnn::Gelu< TDeviceType, TPrecision >	Gaussian Error Linear Unit (GELU) activation component
CMila::Dnn::LayerNorm< TDeviceType, TPrecision >	Device-templated Layer Normalization component
CMila::Dnn::Loss< TDeviceType, TPrecision >	Abstract base class for neural network loss functions
CMila::Dnn::MixtureOfExperts< TDeviceType, TPrecision, TGate, TWeightQuantization >	Stacked experts: gate_up_proj [E, 2I, H] and down_proj [E, H, I], checkpoint-named
CMila::Dnn::MultiHeadAttention< TDeviceType, TPrecision >	Multi-Head Attention module that accepts concatenated QKV input
CMila::Dnn::Residual< TDeviceType, TPrecision >	Device-templated Residual connection component
CMila::Dnn::RmsNorm< TDeviceType, TPrecision >	Device-templated RMS Normalization component
CMila::Dnn::Rope< TDeviceType, TPrecision >	Device-templated RoPE component
CMila::Dnn::Softmax< TDeviceType, TPrecision >	Softmax activation module (device-templated)
CMila::Dnn::SoftmaxCrossEntropy< TDeviceType, TPrecision >	Fused SoftmaxCrossEntropy loss module (device-templated)
CMila::Dnn::Swiglu< TDeviceType, TPrecision, TGate >	Gated-linear-unit (GLU-family) activation component
CMila::Dnn::Component< TDeviceType, dtype_t::FP32 >
CMila::Dnn::Lpe< TDeviceType, TIndex, TPrecision >	Encoder module for token and positional embeddings (device-templated)
CMila::Dnn::TokenEmbedding< TDeviceType, TIndex, TPrecision, TTableQuantization, TTableResidency >	Pure token embedding component (device-templated)
CMila::Dnn::Component< TDeviceType, TComputePrecision >
CMila::Dnn::GroupedQueryAttention< TDeviceType, TPrecision, TKvPolicy >
CMila::Dnn::Linear< TDeviceType, TPrecision, TWeightQuantization >
CMila::Dnn::Linear< TDeviceType, TPrecision >
CMila::Dnn::Linear< TDeviceType, TPrecision, TWeightQuant >
CMila::Dnn::Linear< TDeviceType, TPrecision, TableQuantizationPolicy >
CMila::Dnn::Linear< TDeviceType, TPrecision, typename PrecisionPlan::QkvProjection >
CMila::Dnn::Linear< TDeviceType, TPrecision, typename PrecisionPlan::OutputProjection >
CMila::Dnn::Linear< TDeviceType, TPrecision, typename PrecisionPlan::FeedForwardGateUp >
CMila::Dnn::Linear< TDeviceType, TPrecision, typename PrecisionPlan::FeedForwardDown >
CMila::Dnn::Linear< TDeviceType, TPrecision, typename PrecisionPlan::DeltaNetQueryKey >
CMila::Dnn::Linear< TDeviceType, TPrecision, typename PrecisionPlan::DeltaNetValueGateOutput >
CMila::Dnn::Linear< TDeviceType, TPrecision, typename PrecisionPlan::DeltaNetGating >
CMila::Dnn::Linear< TDeviceType, TPrecision, typename PrecisionPlan::LanguageModelHead >
CMila::Dnn::GroupedQueryAttention< TDeviceType, TComputePrecision, TKvPolicy >	Grouped-Query Attention module that accepts concatenated QKV input
CMila::Dnn::Linear< TDeviceType, TComputePrecision, TWeightQuant >	Device-templated fully connected (linear) component
CMila::Dnn::ComponentConfig	Abstract base for component configuration objects
CMila::Dnn::ActivationConfig	Configuration class for the elementwise Activation component
CMila::Dnn::AttentionOutputGateConfig
CMila::Dnn::CausalConv1dConfig
CMila::Dnn::CrossEntropyConfig	Configuration for fused SoftmaxCrossEntropy loss
CMila::Dnn::GatedDeltaRuleConfig
CMila::Dnn::GatedMLPConfig	Configuration class for the gated feed-forward (GatedMLP) block
CMila::Dnn::GeluConfig	Configuration class for GELU module
CMila::Dnn::GemmaConfig	Network-level configuration for Gemma 4 transformer networks
CMila::Dnn::GptBlockConfig	Configuration class for GPT transformer blocks
CMila::Dnn::GptConfig	Network-level configuration for GPT-style transformer networks
CMila::Dnn::GqaConfig	Configuration class for the Grouped-Query Attention module
CMila::Dnn::LayerNormConfig
CMila::Dnn::LinearConfig	Configuration object for a Linear (fully connected) layer
CMila::Dnn::LlamaConfig	Network-level configuration for LLaMA-style transformer networks
CMila::Dnn::LpeConfig	Configuration class for the Learned Positional Encoder
CMila::Dnn::MLPConfig	Configuration class for the Multi-Layer Perceptron (MLP) block
CMila::Dnn::MixtureOfExpertsConfig
CMila::Dnn::MultiHeadAttentionConfig	Configuration class for Attention module
CMila::Dnn::Optimizers::AdamWConfig	Configuration for AdamW optimizer
CMila::Dnn::QwenConfig	Network-level configuration for Qwen 3.8 transformer networks
CMila::Dnn::ResidualConfig	Configuration class for Residual connection component
CMila::Dnn::RmsNormConfig
CMila::Dnn::RopeConfig
CMila::Dnn::RouterConfig
CMila::Dnn::SoftmaxConfig	Configuration class for Softmax module
CMila::Dnn::SwigluConfig
CMila::Dnn::TokenEmbeddingConfig	Configuration for the TokenEmbedding component
CMila::Dnn::ComponentFactory	Factory for reconstructing components from serialized archives
CComputeDevice
CMila::Dnn::Compute::MetalDevice	Class representing a Metal compute device instance
CMila::Dnn::Compute::VulkanDevice	Class representing a Vulkan compute device instance
CMila::Dnn::Compute::ComputePassMask	A set of compute passes, small enough to test on the publication path
CMila::Dnn::Compute::CpuDeviceRegistrar	CPU device plugin for device-agnostic registration
CMila::Dnn::CpuTensorDataTypeTraits	CPU-specific traits for abstract tensor data types
CMila::Dnn::Compute::Cuda::CudaDataTypeTraits< TDataType >	Compile-time mapping from TensorDataType -> cudaDataType_t
CMila::Dnn::Compute::Cuda::CudaDataTypeTraits< TensorDataType::BF16 >
CMila::Dnn::Compute::Cuda::CudaDataTypeTraits< TensorDataType::FP16 >
CMila::Dnn::Compute::Cuda::CudaDataTypeTraits< TensorDataType::FP32 >
CMila::Dnn::Compute::Cuda::CudaDataTypeTraits< TensorDataType::FP8_E4M3 >
CMila::Dnn::Compute::Cuda::CudaDataTypeTraits< TensorDataType::FP8_E5M2 >
CMila::Dnn::Compute::Cuda::CudaDataTypeTraits< TensorDataType::INT32 >
CMila::Dnn::Compute::Cuda::CudaDataTypeTraits< TensorDataType::INT8 >
CMila::Dnn::Compute::CudaDeviceProps	Wrapper for CUDA device properties with cached values
CMila::Dnn::Compute::CudaDeviceRegistrar	CUDA device registrar for device-agnostic registration
CMila::Dnn::Compute::CudaTimer	GPU-accurate interval timer using a CUDA event pair
CMila::Data::DataLoader< TInputDataType, TTargetDataType, TMemoryResource >	Device-agnostic data loader interface using abstract tensor data types
CMila::Data::DataLoader< TensorDataType::INT32, TensorDataType::INT32, TMemoryResource >
CMila::Data::TokenSequenceLoader< TMemoryResource >	Token sequence loader for autoregressive language models
CMila::Dnn::Decoder< TDeviceType, TPrecision >
CMila::Dnn::DeploymentFootprint	What a deployment would cost, and how it would prefill
CMila::Dnn::Compute::Device	Abstract interface for compute device implementations
CMila::Dnn::Compute::CpuDevice	Class representing a CPU compute device
CMila::Dnn::Compute::CudaDevice	Class representing a CUDA compute device instance
CMila::Dnn::Compute::DeviceAccessible
CMila::Dnn::Compute::DeviceConstructionKey	Construction key for device factories
CMila::Dnn::Compute::DeviceId	Lightweight identifier for a compute device
CMila::Dnn::Compute::DeviceMemoryInfo	What a device currently has free, and what it has in total
CMila::Dnn::Compute::DeviceRegistrar	Device-agnostic registrar for automatic device discovery and registration
CMila::Dnn::Compute::DeviceRegistry	Registry of discovered compute devices with lazy instantiation
CMila::Dnn::Compute::DeviceTypeTraits< TDevice >
CMila::Dnn::Compute::DeviceTypeTraits< DeviceType::Cpu >	DeviceTypeTraits specialization for the CPU device
CMila::Dnn::Compute::DeviceTypeTraits< DeviceType::Cuda >	DeviceTypeTraits specialization for the CUDA device
CMila::Dnn::Compute::ExecutionContextTraits< TDeviceType >
CMila::Dnn::Compute::ExecutionContextTraits< DeviceType::Cpu >
CMila::Dnn::Compute::ExecutionContextTraits< DeviceType::Cuda >
Cstd::false_type
CMila::Dnn::dependent_false< T >
CMila::Distribution::FetchReport
CMila::Dnn::functor_of< TFn >	Compile-time map from an elementwise ActivationType to its functor
CMila::Dnn::functor_of< ActivationType::Gelu >
CMila::Dnn::functor_of< ActivationType::LeakyRelu >
CMila::Dnn::functor_of< ActivationType::Mish >
CMila::Dnn::functor_of< ActivationType::None >
CMila::Dnn::functor_of< ActivationType::Relu >
CMila::Dnn::functor_of< ActivationType::Sigmoid >
CMila::Dnn::functor_of< ActivationType::Silu >
CMila::Dnn::functor_of< ActivationType::Tanh >
CMila::Dnn::GemmaBlockWorkspace< TDeviceType, TPrecision >	Transformer-owned shared activation workspace for GemmaBlock (pooling)
CMila::Dnn::GemmaBlockWorkspaceWidths	Max-geometry slot widths, shared by the allocation and the transformer's footprint
CMila::Dnn::GenerateParams	Per-call inputs to LanguageModel::generate
CMila::Dnn::Compute::GqaState	Non-owning pointers to shared transient GQA scratch buffers
CMila::Dnn::Compute::GqaWorkspace< TDeviceType, TPrecision >	The GQA transient the attention layers of one stack share, owned as one unit
Cstd::hash< Mila::Dnn::Compute::DeviceId >	Hash specialization for DeviceId
CMila::Dnn::Compute::HostAccessible
CMila::Distribution::HttpClient	GET a URL, streaming the body to a sink, over whichever transport it is given
CMila::Distribution::HttpFetch
CMila::Distribution::HttpHeader
CMila::Distribution::HttpRequest
CMila::Distribution::HttpResponse
CMila::Distribution::HttpResult
CMila::Distribution::HubModel	One repository as a hub reports it, before any manifest is fetched
CMila::Dnn::Compute::IExecutionContext	Type-erased execution context interface
CMila::Dnn::Compute::CpuExecutionContext	CPU execution context specialization
CMila::Dnn::Compute::CudaExecutionContext	CUDA execution context specialization
CMila::Dnn::Compute::ExecutionContext< DeviceType::Metal >	Metal execution context specialization
CMila::Dnn::Compute::ExecutionContext< DeviceType::Vulkan >	Vulkan execution context specialization
CMila::Distribution::IHttpTransport	Performs one HTTP GET
CMila::Distribution::CurlHttpTransport	Libcurl as an IHttpTransport
CMila::Distribution::NullHttpTransport	The transport of a build compiled without one
CMila::Dnn::Compute::IKvCacheLifecycle	Capability interface for KV-cache state management
CMila::Dnn::Compute::IKvInference	Compute interface for attention operations that maintain a KV cache
CMila::Dnn::Compute::IPackedKvInference	KV-cache inference interface for packed-QKV MHA backends
CMila::Distribution::IModelHub	A remote that serves manifests and files
CMila::Distribution::HuggingFaceHub	The HuggingFace hub
CMila::Distribution::InstallOptions	How a package is to be installed
CMila::Dnn::Compute::IPositionalDecode	Capability interface for position-dependent unary operations
CMila::Dnn::Compute::IPositionalPairedOp	Capability interface for position-dependent paired operations
CMila::Dnn::ITensor	Abstract interface providing essential tensor information and data access
CMila::Dnn::Tensor< TInputDataType, TMemoryResource >
CMila::Dnn::Tensor< TTargetDataType, TMemoryResource >
CMila::Dnn::Tensor< TensorDataType::INT32, TMemoryResource >
CMila::Dnn::Tensor< TPrecision, MR >
CMila::Dnn::Tensor< TComputePrecision, MR >
CMila::Dnn::Tensor< kCacheDtype, MR >
CMila::Dnn::Tensor< TensorDataType::FP32, CpuMemoryResource >
CMila::Dnn::Tensor< TensorDataType::FP32, MR >
CMila::Dnn::Tensor< TIndex, MR >
CMila::Dnn::Tensor< kTableDtype, TableMemoryResource >
CMila::Dnn::Tensor< TTableQuantization::kScaleDtype, MR >
CMila::Dnn::Tensor< kWeightDtype, MR >
CMila::Dnn::Tensor< TWeightQuant::kScaleDtype, MR >
CMila::Dnn::Tensor< TensorDataType::UINT8, MR >
CMila::Dnn::Tensor< TensorDataType::INT32, MR >
CMila::Dnn::Tensor< dtype_t::INT32, MR >
CMila::Dnn::Tensor< TDataType, Compute::CpuMemoryResource >
CMila::Dnn::Tensor< dtype_t::INT32, Mila::Dnn::Compute::CpuMemoryResource >
CMila::Dnn::Tensor< TensorDataType::FP32, Mila::Dnn::Compute::CpuMemoryResource >
CMila::Dnn::Tensor< TensorDataType::INT32, Mila::Dnn::Compute::CpuMemoryResource >
CMila::Dnn::Tensor< TDataType, TMemoryResource >	Device-aware N-dimensional tensor
CMila::Dnn::Serialization::ITensorBlob	Type-erased interface for a serialized tensor blob
CMila::Dnn::Serialization::TensorBlob< MR >	Concrete tensor blob owning a TensorBuffer-backed raw byte buffer
CMila::Dnn::Serialization::TensorBlobView	Non-owning ITensorBlob view over externally-owned bytes
CMila::Dnn::ITransformerBlock< TDeviceType, TPrecision >	Polymorphic inference interface for one transformer block
CMila::Dnn::GemmaBlock< TDeviceType, TPrecision, false, TWeightQuantization, TKvCachePolicy, kDelegatedFeedForward, kMixtureOfExperts >
CMila::Dnn::GemmaBlock< TDeviceType, TPrecision, true, TWeightQuantization, NoKvCompression, kDelegatedFeedForward, kMixtureOfExperts >
CMila::Dnn::QwenAttentionBlock< TDeviceType, TPrecision, TWeightPlan, TKvCachePolicy >
CMila::Dnn::QwenDeltaNetBlock< TDeviceType, TPrecision, TWeightPlan >
CMila::Dnn::GemmaBlock< TDeviceType, TPrecision, kGlobal, TWeightQuant, TKvPolicy, kDelegatedFeedForward, kMixtureOfExperts >	One Gemma 4 decoder block; kGlobal selects the global (full-attention) geometry
CMila::Dnn::QwenAttentionBlock< TDeviceType, TPrecision, TWeightPlan, TKvPolicy >	One Qwen 3.8 full-attention decoder block
CMila::Dnn::QwenDeltaNetBlock< TDeviceType, TPrecision, TWeightPlan >	One Gated DeltaNet decoder block
CMila::Dnn::LanguageModelConfig< TDerived >	CRTP base configuration for all deployable Mila language models
CMila::Dnn::LanguageModelConfig< GemmaModelConfig >
CMila::Dnn::GemmaModelConfig	Deployment configuration for Gemma language models
CMila::Dnn::LanguageModelConfig< GptModelConfig >
CMila::Dnn::GptModelConfig	Deployment configuration for Gpt language models
CMila::Dnn::LanguageModelConfig< LlamaModelConfig >
CMila::Dnn::LlamaModelConfig	Deployment configuration for Llama language models
CMila::Dnn::LanguageModelConfig< QwenModelConfig >
CMila::Dnn::QwenModelConfig	Deployment configuration for Qwen 3.8 language models
CMila::Dnn::LearningRateScheduler	Abstract base for learning-rate schedulers
CMila::Dnn::ConstantLRScheduler	Constant learning-rate scheduler
CMila::Dnn::CosineLRScheduler	Cosine annealing scheduler
CMila::Dnn::LinearLRScheduler	Linear decay scheduler
CMila::Logging::Logger	Abstract logging interface and static facade
CMila::Logging::ConsoleSink	Thread-safe logging sink that writes formatted records to the console
CMila::Logging::FileSink	Thread-safe logging sink that writes formatted records to a file
CMila::Logging::NullSink	A logging sink that silently discards all records
Cstd::pmr::memory_resource
CMila::Dnn::Compute::MemoryResource	Clean memory resource abstraction for device-specific memory allocation
CMila::Dnn::Compute::CpuMemoryResource	CPU memory resource for host-accessible memory allocation
CMila::Dnn::Compute::CudaDeviceMemoryResource	CUDA device memory resource for GPU-accessible memory allocation
CMila::Dnn::Compute::CudaManagedMemoryResource	CUDA managed memory resource for unified host/device accessible memory
CMila::Dnn::Compute::CudaPinnedMemoryResource	CUDA pinned memory resource for fast host/device transfer memory
CMila::Dnn::Compute::MetalMemoryResource	Stub implementation for non-Apple platforms
CMila::Dnn::Compute::TrackedMemoryResource	A memory resource wrapper that tracks allocation and deallocation statistics
CMila::Dnn::Compute::VulkanMemoryResource	Stub implementation for platforms without Vulkan support
CMila::Dnn::Compute::MemoryAllocationStats	Process-wide allocator counters for all TrackedMemoryResource instances
CMila::Dnn::Compute::MemoryResourceTraits< TMemoryResource >	Memory resource traits for compile-time dispatch optimization
CMila::Dnn::Compute::MemoryResourceTraits< CpuMemoryResource >	CPU-specific memory resource traits providing detailed CPU backend characteristics
CMila::Dnn::Compute::MemoryResourceTraits< CudaDeviceMemoryResource >	CUDA device memory resource traits providing detailed GPU backend characteristics
CMila::Dnn::Compute::MemoryResourceTraits< CudaManagedMemoryResource >	CUDA managed memory resource traits providing unified memory characteristics
CMila::Dnn::Compute::MemoryResourceTraits< CudaPinnedMemoryResource >	CUDA pinned memory resource traits providing fast transfer characteristics
CMila::Dnn::MemoryStats	Memory allocation breakdown for a single component
CMila::Dnn::Compute::MetalDevicePlugin	Metal device plugin for device-agnostic registration
CMila::Data::MilaFileHeader	Common file header for Mila data files
CMila::Dnn::Model< TDeviceType, TPrecision >
CMila::Dnn::LanguageModel< TDeviceType, TPrecision >
CMila::Dnn::GemmaModel< TDeviceType, TPrecision >	Gemma 4 compatible inference model
CMila::Dnn::GptModel< TDeviceType, TPrecision >	GPT inference model
CMila::Dnn::LlamaModel< TDeviceType, TPrecision >	LLaMA 3 compatible inference model
CMila::Dnn::QwenModel< TDeviceType, TPrecision >	Qwen 3.8 compatible inference model
CMila::Dnn::Serialization::ModelArchive	ModelArchive provides high-level helpers for component serialization
CMila::Dnn::ModelConfig	Abstract base configuration for all deployable Mila models
CMila::Distribution::ModelCoordinate	Where a repository lives on a hub
CMila::Distribution::ModelFile	One file composing a model, as the manifest declares it
CMila::Distribution::ModelManifest	A model's published description
CMila::Distribution::ModelPackage	Everything a published or installable model is, as a directory
CMila::Distribution::ModelRecord	An installed model: what the manifest published, plus how this copy came to be here
CMila::Distribution::ModelResolver	Pulls a published model into the store
CMila::Distribution::ModelStore	The local store of installed models
CMila::Dnn::MultiAxisPartition	Multi-axis partition for normalization over trailing dimensions
CMila::Dnn::NetworkFactory	Factory registry for Network deserialization
CMila::Dnn::Quant::KvCache::NoKvCompression	Identity policy - no compression
CMila::Dnn::Quant::Weight::NoWeightQuant
CMila::Profiling::NvtxRange
CMila::Dnn::Compute::ObservableStage	A stage a component publishes, and the passes it is published on
CMila::Dnn::Compute::Operation< TDeviceType, TComputePrecision >
CMila::Dnn::Compute::OperationTraits< TOp, TDeviceType, TPrecision, TPolicy >	Primary traits template for unified compile-time operation dispatch
CMila::Dnn::Optimizer< TDeviceType, TPrecision >	Abstract base class for parameter optimizers
CMila::Dnn::Optimizers::AdamWOptimizer< TDeviceType, TPrecision >	Device-agnostic AdamW optimizer
CMila::Distribution::PackageRequest	One model's files, as a caller hands them to the packager
CMila::Distribution::PackageValidation	What validation found
CMila::Dnn::Quant::Weight::PerChannelFp8< TStorage >
CMila::Dnn::Quant::KvCache::PerChannelKvFp8< TStorage >	Symmetric per-head per-token FP8 KV cache compression policy
CMila::Dnn::Quant::Weight::PerGroupCodebook2< kGroupSize >
CMila::Dnn::Quant::Weight::PerGroupCodebook3< kGroupSize >
CMila::Dnn::Quant::Weight::PerGroupFp4< kGroupSize >
CMila::Dnn::Quant::Weight::PerGroupInt4< kGroupSize >
CMila::Dnn::PrefillChunking	How a transformer would chunk a prefill pass at a given context length
CMila::Dnn::QwenAttentionBlockWorkspace< TDeviceType, TPrecision >	Transformer-owned shared activation workspace for QwenAttentionBlock (pooling)
CMila::Dnn::QwenDeltaNetBlockWorkspace< TDeviceType, TPrecision >	Transformer-owned shared activation workspace for QwenDeltaNetBlock (pooling)
CMila::Dnn::QwenOraclePrecisionPlan	The Section 5 oracle: FP4 across the body, everything else as the plan has it
CMila::Dnn::QwenPrecisionPlan	The section 5 allocation: 25.62 B resident weights at 2.90 bits average, 8.65 GiB
CMila::Dnn::QwenPlanDetail::QwenPrecisionPlanSelector< T, kIsBarePolicy >
CMila::Dnn::QwenPlanDetail::QwenPrecisionPlanSelector< T, true >
CMila::Core::RandomGenerator	Singleton class providing centralized random number generation
CMila::Distribution::RemovalReport
CMila::Dnn::Router< TDeviceType, TPrecision >::Routing	One call's result. Both views are router-owned and valid until the next forward()
CMila::Dnn::Serialization::SafeTensorsWriter	Writes a safetensors file: u64 header length, JSON header, packed data
CMila::Dnn::Sampler< TDeviceType, TPrecision >	Abstract base for token samplers
CMila::Dnn::TokenSampler< TDeviceType, TPrecision >	The standard token sampler: temperature / top-k / top-p multinomial
CMila::Dnn::SamplingConfig	Model-fixed configuration for TokenSampler
CMila::Dnn::SamplingParams	Per-call sampling knobs consumed by the TokenSampler
CMila::Dnn::Serialization::ModelArchive::ScopedScope
CMila::Dnn::SequenceLogLikelihood	Teacher-forced log-likelihood of a token sequence under a model
CMila::Data::SerializationMetadata	Type-safe metadata container for component serialization
CMila::Dnn::Optimizers::SerializationMetadata	Type-safe metadata container for component serialization
CMila::Dnn::Serialization::SerializationMetadata	Type-safe metadata container for component serialization
CMila::Dnn::SerializationMetadata	Type-safe metadata container for component serialization
CMila::Dnn::Serialization::Serializer	Minimal base interface for model serialization backends
CMila::Dnn::Serialization::ArchiveSerializer	Interface for hierarchical archive serializers
CMila::Dnn::Serialization::ZipSerializer	ZIP archive serializer built on miniz
CMila::Distribution::Sha256	Incremental SHA-256
CMila::Dnn::Quant::KvCache::SlidingWindowKvCache	Bounded sliding-window KV cache (uncompressed ring buffer)
CMila::Data::SpecialTokens	Configuration for special tokens across all tokenizer types
CMila::Dnn::QwenDeltaNetBlock< TDeviceType, TPrecision, TWeightPlan >::StateSnapshot	Everything this block carries between chunks – all of it, in one value
CMila::Utils::StepLogger
CMila::Distribution::StoredModel	A record together with where its bytes actually are
CMila::Distribution::StoreUsage
CMila::Dnn::Serialization::TensorBlobMetadata	Metadata for a tensor blob in a weights file
CMila::Dnn::TensorBuffer< TDataType, TMemoryResource, TrackMemory >	Device-agnostic buffer for storing tensor data with abstract type system
CMila::Dnn::Compute::Cuda::TensorDataTypeMap< TDataType >	Compile-time mapping from abstract TensorDataType -> CUDA native device type
CMila::Dnn::TensorDataTypeMap< TElementType >	Primary template for mapping concrete C++ types to TensorDataType
CMila::Dnn::TensorDataTypeMap< __nv_fp8_e4m3 >
CMila::Dnn::TensorDataTypeMap< __nv_fp8_e5m2 >
CMila::Dnn::TensorDataTypeMap< float >	Concrete type mapping for float (FP32)
CMila::Dnn::TensorDataTypeMap< half >
CMila::Dnn::TensorDataTypeMap< nv_bfloat16 >
CMila::Dnn::TensorDataTypeMap< std::int16_t >	Concrete type mapping for 16-bit signed integer
CMila::Dnn::TensorDataTypeMap< std::int32_t >	Concrete type mapping for 32-bit signed integer
CMila::Dnn::TensorDataTypeMap< std::int8_t >	Concrete type mapping for 8-bit signed integer
CMila::Dnn::TensorDataTypeMap< std::uint16_t >	Concrete type mapping for 16-bit unsigned integer
CMila::Dnn::TensorDataTypeMap< std::uint32_t >	Concrete type mapping for 32-bit unsigned integer
CMila::Dnn::TensorDataTypeMap< std::uint8_t >	Concrete type mapping for 8-bit unsigned integer
CMila::Dnn::Compute::Cuda::TensorDataTypeMap< TensorDataType::BF16 >	Maps TensorDataType::BF16 to CUDA __nv_bfloat16
CMila::Dnn::Compute::Cuda::TensorDataTypeMap< TensorDataType::FP16 >	Maps TensorDataType::FP16 to CUDA __half
CMila::Dnn::Compute::Cuda::TensorDataTypeMap< TensorDataType::FP32 >	Maps TensorDataType::FP32 to CUDA float
CMila::Dnn::Compute::Cuda::TensorDataTypeMap< TensorDataType::FP4_E2M1 >	Maps TensorDataType::FP4_E2M1 to std::uint8_t
CMila::Dnn::Compute::Cuda::TensorDataTypeMap< TensorDataType::FP4_E3M0 >	Maps TensorDataType::FP4_E3M0 to std::uint8_t
CMila::Dnn::Compute::Cuda::TensorDataTypeMap< TensorDataType::FP8_E4M3 >	Maps TensorDataType::FP8_E4M3 to CUDA __nv_fp8_e4m3
CMila::Dnn::Compute::Cuda::TensorDataTypeMap< TensorDataType::FP8_E5M2 >	Maps TensorDataType::FP8_E5M2 to CUDA __nv_fp8_e5m2
CMila::Dnn::Compute::Cuda::TensorDataTypeMap< TensorDataType::INT16 >	Maps TensorDataType::INT16 to std::int16_t
CMila::Dnn::Compute::Cuda::TensorDataTypeMap< TensorDataType::INT32 >	Maps TensorDataType::INT32 to std::int32_t
CMila::Dnn::Compute::Cuda::TensorDataTypeMap< TensorDataType::INT8 >	Maps TensorDataType::INT8 to std::int8_t
CMila::Dnn::Compute::Cuda::TensorDataTypeMap< TensorDataType::UINT16 >	Maps TensorDataType::UINT16 to std::uint16_t
CMila::Dnn::Compute::Cuda::TensorDataTypeMap< TensorDataType::UINT32 >	Maps TensorDataType::UINT32 to std::uint32_t
CMila::Dnn::Compute::Cuda::TensorDataTypeMap< TensorDataType::UINT8 >	Maps TensorDataType::UINT8 to std::uint8_t
CMila::Dnn::TensorDataTypeTraits< TDataType >	Compile-time traits for TensorDataType enumeration values
CMila::Dnn::TensorDataTypeTraits< TensorDataType::BF16 >	Traits specialization for 16-bit brain floating point
CMila::Dnn::TensorDataTypeTraits< TensorDataType::FP16 >	Traits specialization for 16-bit half precision floating point
CMila::Dnn::TensorDataTypeTraits< TensorDataType::FP32 >	Traits specialization for 32-bit IEEE 754 floating point
CMila::Dnn::TensorDataTypeTraits< TensorDataType::FP4_E2M1 >	Traits specialization for 4-bit floating point with E2M1 format
CMila::Dnn::TensorDataTypeTraits< TensorDataType::FP4_E3M0 >	Traits specialization for 4-bit floating point with E3M0 format
CMila::Dnn::TensorDataTypeTraits< TensorDataType::FP8_E4M3 >	Traits specialization for 8-bit floating point with E4M3 format
CMila::Dnn::TensorDataTypeTraits< TensorDataType::FP8_E5M2 >	Traits specialization for 8-bit floating point with E5M2 format
CMila::Dnn::TensorDataTypeTraits< TensorDataType::INT16 >	Traits specialization for 16-bit signed integer
CMila::Dnn::TensorDataTypeTraits< TensorDataType::INT32 >	Traits specialization for 32-bit signed integer
CMila::Dnn::TensorDataTypeTraits< TensorDataType::INT8 >	Traits specialization for 8-bit signed integer
CMila::Dnn::TensorDataTypeTraits< TensorDataType::UINT16 >	Traits specialization for 16-bit unsigned integer
CMila::Dnn::TensorDataTypeTraits< TensorDataType::UINT32 >	Traits specialization for 32-bit unsigned integer
CMila::Dnn::TensorDataTypeTraits< TensorDataType::UINT8 >	Traits specialization for 8-bit unsigned integer
CMila::Dnn::TensorHostTypeMap< TDataType >	Maps abstract TensorDataType to host-compatible C++ type and TensorDataType
CMila::Dnn::TensorHostTypeMap< TensorDataType::BF16 >	Host type for 16-bit brain floating point
CMila::Dnn::TensorHostTypeMap< TensorDataType::FP16 >	Host type for 16-bit half precision floating point
CMila::Dnn::TensorHostTypeMap< TensorDataType::FP32 >	Host type for 32-bit IEEE 754 floating point
CMila::Dnn::TensorHostTypeMap< TensorDataType::FP8_E4M3 >	Host type for 8-bit floating point with E4M3 format
CMila::Dnn::TensorHostTypeMap< TensorDataType::FP8_E5M2 >	Host type for 8-bit floating point with E5M2 format
CMila::Dnn::TensorHostTypeMap< TensorDataType::INT16 >	Host type for 16-bit signed integer
CMila::Dnn::TensorHostTypeMap< TensorDataType::INT32 >	Host type for 32-bit signed integer
CMila::Dnn::TensorHostTypeMap< TensorDataType::INT8 >	Host type for 8-bit signed integer
CMila::Dnn::TensorHostTypeMap< TensorDataType::UINT16 >	Host type for 16-bit unsigned integer
CMila::Dnn::TensorHostTypeMap< TensorDataType::UINT32 >	Host type for 32-bit unsigned integer
CMila::Dnn::TensorHostTypeMap< TensorDataType::UINT8 >	Host type for 8-bit unsigned integer
CMila::Dnn::Serialization::TensorMetadata	Metadata describing a tensor in serialized form
CMila::Dnn::TensorOpsTraits< TDevice >	Device-to-backend map behind the TensorOps alias template
CMila::Dnn::TensorShape	Fixed-capacity inline shape descriptor for N-dimensional tensors
CMila::Data::Tokenizer
CMila::Data::BpeTokenizer	Unified BPE tokenizer targeting GPT-2, Llama 3.x, and Mistral model families
CMila::Data::CharTokenizer	Character-level tokenizer
CMila::Data::TokenizerTrainer	Abstract interface for training tokenizer vocabularies from text corpora
CMila::Data::TokenizerVocabulary	Generic tokenizer vocabulary interface
CMila::Data::BpeVocabulary	Unified Byte Pair Encoding (BPE) vocabulary
CMila::Data::CharVocabulary	Character vocabulary for tokenization
CMila::Data::TokenSequenceLoaderConfig	Configuration for StreamingSequenceLoader behavior
CMila::Dnn::Conversation::ToolCall	One call the model asked for, as a name and a JSON object of arguments
CMila::Data::TrainerFactory	Factory for creating tokenizer trainers and loading vocabularies
CMila::Dnn::Conversation::Turn	A single turn in a conversation
CMila::Dnn::Quant::Weight::UniformPrecisionPlan< TPolicy >	Assigns one policy to every role
CMila::Dnn::QwenUniformPrecisionPlan< NoWeightQuant >
CMila::Dnn::QwenUniformPrecisionPlan< T >
CMila::Dnn::QwenUniformPrecisionPlan< TPolicy >	One policy across every Qwen role, DeltaNet included
CMila::Dnn::UniqueIdGenerator	Thread-safe generator for unique tensor identifiers
CMila::Version	Semantic Version data
CMila::Dnn::VulkanTensorTraits	Vulkan-specific traits for abstract tensor data types
CMila::Dnn::Serialization::WeightsMetadata	The model configuration a weights file carries
CMila::Dnn::Serialization::WeightsReader	Reader for a Mila weights file