Classes
class	ArchiveSerializer
	Interface for hierarchical archive serializers. More...
struct	ITensorBlob
	Type-erased interface for a serialized tensor blob. More...
class	ModelArchive
	ModelArchive provides high-level helpers for component serialization. More...
struct	PretrainedMetadata
	Metadata for pretrained model. More...
class	PretrainedModelReader
	Reader for Mila pretrained binary format. More...
class	SerializationMetadata
	Type-safe metadata container for component serialization. More...
class	Serializer
	Minimal base interface for model serialization backends. More...
struct	TensorBlob
	Concrete tensor blob owning a TensorBuffer-backed raw byte buffer. More...
struct	TensorBlobMetadata
	Metadata for a tensor blob in pretrained model format. More...
struct	TensorMetadata
	Metadata describing a tensor in serialized form. More...
class	ZipSerializer
	ZIP archive serializer built on miniz. More...

Typedefs
using	json = nlohmann::json
using	Mila::Dnn::Serialization::MetadataValue
	Type-safe metadata value supporting common serialization types.

Enumerations
enum class	DType : uint32_t { Float32 = 0 , Float16 = 1 , BFloat16 = 2 , Int32 = 3 }
enum class	Mila::Dnn::Serialization::OpenMode { Read , Write }
	Mode indicating whether the archive is used for reading or writing. More...
enum class	Mila::Dnn::Serialization::SerializationMode { Checkpoint , WeightsOnly , Architecture }
	Modes for serialization and deserialization. More...

Functions
constexpr std::string_view	Mila::Dnn::Serialization::archiveModeToString (OpenMode mode)
	Convert ArchiveMode to string.
TensorDataType	dtypeToTensorDataType (uint32_t dtype)
template<typename MR = Compute::CpuMemoryResource> requires isValidTensor<dtype_t::UINT8, MR>
TensorBlob< MR >	Mila::Dnn::Serialization::readTensorBlob (const ModelArchive &archive, const std::string &prefix, int device_id=0)
	Read tensor metadata and raw bytes from prefix into a typed blob.
std::string	Mila::Dnn::Serialization::serializationModeToString (SerializationMode mode)
	Convert a SerializationMode value to a human-readable string.
void	Mila::Dnn::Serialization::writeTensorBlob (ModelArchive &archive, const std::string &prefix, const TensorMetadata &meta, const void *data, size_t size)
	Write tensor metadata and raw bytes under the given prefix into archive.

Typedef Documentation

◆ json

using Mila::Dnn::Serialization::json = nlohmann::json

Examples: /__w/Mila/Mila/Mila/Src/Dnn/Serialization/ModelArchive.ixx.

◆ MetadataValue

using Mila::Dnn::Serialization::MetadataValue

export

Initial value:

 std::variant<
        std::string,
        int64_t,
        double,
        bool,
        std::vector<int64_t>,
        std::vector<double>,
        std::vector<std::string>
    >

Type-safe metadata value supporting common serialization types.

Encapsulates primitive types, strings, and structured data without exposing the underlying serialization format to user code.

Enumeration Type Documentation

◆ DType

enum class Mila::Dnn::Serialization::DType : uint32_t

strong

Enumerator
Float32
Float16
BFloat16
Int32

◆ OpenMode

enum class Mila::Dnn::Serialization::OpenMode

exportstrong

Mode indicating whether the archive is used for reading or writing.

Preconditions:

The underlying ModelSerializer should be opened for the matching operation (openForRead for Read, openForWrite for Write) before calling archive helpers.

Semantics:

Read : permits read operations (readJson, readBlob, file queries).
Write : permits write operations (writeJson, writeBlob, add data).

Example:

auto serializer = std::make_unique<ZipSerializer>( ... );
serializer->openForWrite(...);
ModelArchive archive( std::move(serializer), ArchiveMode::Write );
archive.writeJson("meta.json", j);

Enumerator
Read
Write

Examples: /__w/Mila/Mila/Mila/Src/Dnn/Serialization/ModelArchive.ixx.

◆ SerializationMode

enum class Mila::Dnn::Serialization::SerializationMode

exportstrong

Modes for serialization and deserialization.

Selects which parts of a network are saved or restored:

Checkpoint: full state (architecture + weights + optimizer state)
WeightsOnly: only parameter tensors
Architecture: only network topology and metadata

Enumerator
Checkpoint
WeightsOnly
Architecture

Examples: /__w/Mila/Mila/Mila/Src/Dnn/Components/Activations/Gelu/Gelu.ixx, and /__w/Mila/Mila/Mila/Src/Dnn/Components/Normalization/Softmax.ixx.

Function Documentation

◆ archiveModeToString()

std::string_view Mila::Dnn::Serialization::archiveModeToString ( OpenMode mode )

constexprexport

Convert ArchiveMode to string.

Examples: /__w/Mila/Mila/Mila/Src/Dnn/Serialization/ModelArchive.ixx.

Here is the caller graph for this function:

◆ dtypeToTensorDataType()

TensorDataType Mila::Dnn::Serialization::dtypeToTensorDataType ( uint32_t dtype )

inline

Here is the caller graph for this function:

◆ readTensorBlob()

template<typename MR = Compute::CpuMemoryResource>
requires isValidTensor<dtype_t::UINT8, MR>

TensorBlob< MR > Mila::Dnn::Serialization::readTensorBlob	(	const ModelArchive &	archive,
		const std::string &	prefix,
		int	device_id = 0 )

export

Read tensor metadata and raw bytes from prefix into a typed blob.

Allocates a TensorBuffer<UINT8, MR> sized to total_bytes and reads directly into it via readBlobInto — no intermediate vector copy. When MR is CudaPinnedMemoryResource the returned blob carries a pinned host pointer ready for direct DMA in copyFromBlob.

Template Parameters

MR	Memory resource for the blob data buffer. Defaults to CpuMemoryResource.

Parameters

archive	ModelArchive to read from
prefix	Path prefix for tensor files (e.g., "tensors/weight")
device_id	Device index passed to the memory resource constructor.

Returns: TensorBlob<MR> owning the metadata and raw byte buffer.

Exceptions

std::runtime_error if read operations fail or size mismatch detected

Here is the call graph for this function:

◆ serializationModeToString()

std::string Mila::Dnn::Serialization::serializationModeToString ( SerializationMode mode )

export

Convert a SerializationMode value to a human-readable string.

Parameters

mode	The serialization mode to convert.

Returns: A short string identifying the mode ("Checkpoint", "WeightsOnly", or "Architecture").

Here is the caller graph for this function:

◆ writeTensorBlob()

void Mila::Dnn::Serialization::writeTensorBlob	(	ModelArchive &	archive,
		const std::string &	prefix,
		const TensorMetadata &	meta,
		const void *	data,
		size_t	size )

inlineexport

Write tensor metadata and raw bytes under the given prefix into archive.

Writes: prefix + "/meta.json" — TensorMetadata as SerializationMetadata prefix + "/data.bin" — raw tensor bytes

Parameters

archive	ModelArchive to write to
prefix	Path prefix for tensor files (e.g., "tensors/weight")
meta	Tensor metadata
data	Pointer to raw tensor bytes
size	Number of bytes to write

Exceptions

std::runtime_error if write operations fail

Here is the call graph for this function:

Here is the caller graph for this function:

Classes

Typedefs

Enumerations

Functions

Typedef Documentation

◆ json

◆ MetadataValue

Enumeration Type Documentation

◆ DType

◆ OpenMode

◆ SerializationMode

Function Documentation

◆ archiveModeToString()

◆ dtypeToTensorDataType()

◆ readTensorBlob()

◆ serializationModeToString()

◆ writeTensorBlob()