CudaLinearOp.Plans.ixx File Reference

cuBLASLt plan builders for CudaLinearOp forward and backward passes. More...

#include <cublasLt.h>
#include <cuda_fp16.h>
#include <cuda_fp8.h>
#include <format>
#include <stdexcept>
#include "Kernels/Linear.cuh"
import Logging.Logger;
import Compute.CublasLtPlan;

Namespaces
namespace	Mila
	Mila main API namespace.
namespace	Mila::Dnn
namespace	Mila::Dnn::Compute
namespace	Mila::Dnn::Compute::Cuda
namespace	Mila::Dnn::Compute::Cuda::Linear
namespace	Mila::Dnn::Compute::Cuda::Linear::Detail

Typedefs
template<typename TComputePrecision>
using	Mila::Dnn::Compute::Cuda::Linear::Detail::CublasLtMatMulPlan = Mila::Dnn::Compute::Cuda::CublasLtMatMulPlan<TComputePrecision>

Functions
template<typename TComputePrecision>
CublasLtMatMulPlan< TComputePrecision >	Mila::Dnn::Compute::Cuda::Linear::Detail::build_backward_input_plan (cublasLtHandle_t handle, int batch_size, int in_features, int out_features, cudaDataType_t cuda_data_type, cublasComputeType_t compute_type, cudaDataType_t scale_type)
	Build cuBLASLt plan for backward input gradient computation.
template<typename TComputePrecision>
CublasLtMatMulPlan< TComputePrecision >	Mila::Dnn::Compute::Cuda::Linear::Detail::build_backward_weight_plan (cublasLtHandle_t handle, int batch_size, int in_features, int out_features, cudaDataType_t cuda_data_type, cublasComputeType_t compute_type, cudaDataType_t scale_type)
	Build cuBLASLt plan for backward weight gradient computation.
template<typename TComputePrecision>
CublasLtMatMulPlan< TComputePrecision >	Mila::Dnn::Compute::Cuda::Linear::Detail::build_forward_plan (cublasLtHandle_t handle, int batch_size, int in_features, int out_features, bool has_bias, cudaDataType_t cuda_data_type, cublasComputeType_t compute_type, cudaDataType_t scale_type)
template<typename TComputePrecision>
void	Mila::Dnn::Compute::Cuda::Linear::Detail::compute_bias_gradient (TComputePrecision *bias_grad, const TComputePrecision *output_grad, int batch_size, int out_features, cudaStream_t stream)
	Compute bias gradient via reduction sum across batch dimension.

Detailed Description

cuBLASLt plan builders for CudaLinearOp forward and backward passes.