OperationTraits.Cpu.ixx File Reference

OperationTraits specializations for all CPU operation backends. More...

import Dnn.Quantization.Weight.Policies;
import Compute.CpuSoftmaxOp;
import Compute.CpuResidualOp;
import Compute.CpuEncoderOp;
import Compute.CpuGeluOp;
import Compute.CpuLinearOp;
import Compute.CpuAttention;
import Compute.OperationTraits.Template;

Classes
struct	Mila::Dnn::Compute::OperationTraits< OperationType::GeluOp, DeviceType::Cpu, TensorDataType::FP32, void >
struct	Mila::Dnn::Compute::OperationTraits< OperationType::LinearOp, DeviceType::Cpu, TensorDataType::FP32, NoWeightQuant >
struct	Mila::Dnn::Compute::OperationTraits< OperationType::LpeOp, DeviceType::Cpu, TensorDataType::FP32, void >
struct	Mila::Dnn::Compute::OperationTraits< OperationType::MultiHeadAttentionOp, DeviceType::Cpu, TensorDataType::FP32, void >
struct	Mila::Dnn::Compute::OperationTraits< OperationType::ResidualOp, DeviceType::Cpu, TensorDataType::FP32, void >
struct	Mila::Dnn::Compute::OperationTraits< OperationType::SoftmaxOp, DeviceType::Cpu, TensorDataType::FP32, void >

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

Detailed Description

OperationTraits specializations for all CPU operation backends.

This partition module is the single registration point for every (OperationType, Cpu, TPrecision, TPolicy) -> concrete op mapping.

CPU ops are currently concrete (non-templated) FP32-only implementations. BF16 CPU paths are not a current Mila target.

Migration status: LinearOp complete (NoWeightQuant; quantized policies are CUDA-only) GeluOp complete ResidualOp complete SoftmaxOp complete MultiHeadAttentionOp complete LpeOp complete CrossEntropyOp pending (CpuSoftmaxCrossEntropyOp not yet wired into CMake) SamplingOp pending