PretrainedReader.ixx File Reference

Reader for Mila pretrained binary format. More...

#include <filesystem>
#include <fstream>
#include <string>
#include <unordered_map>
#include <vector>
#include <cstdint>
#include <stdexcept>
#include <format>
#include <algorithm>
import Compute.CpuMemoryResource;
import Compute.Device;
import Dnn.TensorTypes;
import Dnn.TensorDataTypeTraits;
import Dnn.TensorBuffer;
import Dnn.TensorDataType;
import Serialization.Tensor;
import Serialization.OpenMode;
import Dnn.ITensor;
import Dnn.Tensor;
import Serialization.Serializer;

Classes
struct	Mila::Dnn::Serialization::PretrainedMetadata
	Metadata for pretrained model. More...
class	Mila::Dnn::Serialization::PretrainedModelReader
	Reader for Mila pretrained binary format. More...
struct	Mila::Dnn::Serialization::TensorBlobMetadata
	Metadata for a tensor blob in pretrained model format. More...

Namespaces
namespace	Mila
	Mila main API namespace.
namespace	Mila::Dnn
namespace	Mila::Dnn::Serialization

Enumerations
enum class	Mila::Dnn::Serialization::DType : uint32_t { Mila::Dnn::Serialization::Float32 = 0 , Mila::Dnn::Serialization::Float16 = 1 , Mila::Dnn::Serialization::BFloat16 = 2 , Mila::Dnn::Serialization::Int32 = 3 }

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

Detailed Description

Reader for Mila pretrained binary format.

Provides direct access to pretrained model weights stored in Mila's flat binary format. Used by fromPretrained() factory methods.

TODO (Alpha.5 Phase 6): Replace the per-tensor fstream read loop with CreateFileMapping/MapViewOfFile. The current approach issues one read per tensor blob (224+ for Llama 3.1 8B), capping throughput at ~2 GB/s against a PCIe 4.0 NVMe floor of ~7 GB/s (~8s load vs ~2s target). TensorBlob::data() should return a pointer into the mapped view; the ITensorBlob interface is stable.