Struct CompileSpec

• Defined in File torch_tensorrt.h

Struct Documentation

struct torch_tensorrt::torchscript::CompileSpec

Settings data structure for Torch-TensorRT TorchScript compilation

Public Functions

TORCHTRT_API CompileSpec(std::vector<std::vector<int64_t>> fixed_sizes)

Construct a new Compile Spec object Convienence constructor to set fixed input size from vectors describing size of input tensors. Each entry in the vector represents a input and should be provided in call order.

This constructor should be use as a convience in the case that all inputs are static sized and you are okay with default input dtype and formats (FP32 for FP32 and INT8 weights, FP16 for FP16 weights, contiguous)

Parameters

fixed_sizes –

TORCHTRT_API CompileSpec(std::vector<c10::ArrayRef<int64_t>> fixed_sizes)

Construct a new Compile Spec object Convienence constructor to set fixed input size from c10::ArrayRef’s (the output of tensor.sizes()) describing size of input tensors. Each entry in the vector represents a input and should be provided in call order.

This constructor should be use as a convience in the case that all inputs are static sized and you are okay with default input dtype and formats (FP32 for FP32 and INT8 weights, FP16 for FP16 weights, contiguous)

Parameters

fixed_sizes –

TORCHTRT_API CompileSpec(std::vector<Input> inputs)

Construct a new Compile Spec object from input ranges. Each entry in the vector represents a input and should be provided in call order.

Use this constructor to define inputs with dynamic shape, specific input types or tensor formats

Parameters

inputs –

TORCHTRT_API CompileSpec(torch::jit::IValue input_signature)

Construct a new Compile Spec object from IValue which represents the nesting of input tensors for a module.

Parameters

input_signature –

Public Members

GraphInputs graph_inputs

Specifications for inputs to the engine, can store a IValue which has stored complex Input or a flatened Input.

std::set<DataType> enabled_precisions = {DataType::kFloat}

The set of precisions TensorRT is allowed to use for kernels during compilation.

bool disable_tf32 = false

Prevent Float32 layers from using TF32 data format

TF32 computes inner products by rounding the inputs to 10-bit mantissas before multiplying, but accumulates the sum using 23-bit mantissas. This is the behavior of FP32 layers by default.

bool sparse_weights = false

Enable sparsity for weights of conv and FC layers

bool refit = false

Build a refitable engine

bool debug = false

Build a debugable engine

bool truncate_long_and_double = false

Truncate long/double type to int/float type

bool allow_shape_tensors = false

Allow shape tensors (from IShape layer) in the graph

Device device

Target Device

EngineCapability capability = EngineCapability::kSTANDARD

Sets the restrictions for the engine (CUDA Safety)

uint64_t num_avg_timing_iters = 1

Number of averaging timing iterations used to select kernels

uint64_t workspace_size = 0

Maximum size of workspace given to TensorRT

uint64_t dla_sram_size = 1048576

Fast software managed RAM used by DLA to communicate within a layer.

uint64_t dla_local_dram_size = 1073741824

Host RAM used by DLA to share intermediate tensor data across operations

uint64_t dla_global_dram_size = 536870912

host RAM used by DLA to store weights and metadata for execution

nvinfer1::IInt8Calibrator *ptq_calibrator = nullptr

Calibration dataloaders for each input for post training quantizatiom

bool require_full_compilation = false

Require the full module be compiled to TensorRT instead of potentially running unsupported operations in PyTorch

uint64_t min_block_size = 3

Minimum number of contiguous supported operators to compile a subgraph to TensorRT

std::vector<std::string> torch_executed_ops

List of aten operators that must be run in PyTorch. An error will be thrown if this list is not empty but require_full_compilation is True

std::vector<std::string> torch_executed_modules

List of modules that must be run in PyTorch. An error will be thrown if this list is not empty but require_full_compilation is True