Attribute class for extra information into computations

Parameters ---------- graph_def : tf.GraphDef The target GraphDef to be partitioned. supported_list : Set[str] = _SEGMENT_SUP

The configuration for torch blade |Config| .. |Config| replace:: :class:`.Config` Example:: import blade import torch_blad

All nodes are assumed to be either in no branch, then branch, else branch, or both branches (such as merge nodes). The code below relies on Else and T

Parameters ---------- graph : SimpleGraph The original main graph. nodes_idx : Set[int] Index of nodes that belong t

Indicates how a FunctionDef is associated with a graph node (e.g. the node is a function call, or the node has function attrs).

Instances of DeviceId represent TensorFlow devices as integers. This helps avoid having to manipulate device names as strings when auto-clustering.

Convenience class for rowwise dropout as described in subsection 1.11.6.

FunctionalizeCond groups all the state used by functionalizing conditionals of the given graph together.

A wrapper Graph to networkx Graph that meets our usages

Reprensets a symbolic expression of symbolicDims.

This module includes original float op and fake quantized op (i.e. observed module). Mean square error is used to analyze the quantization pr

This class is used to collect the calibration data for each fusion group. Basically, the whole process will be divided into the follo

Convenience class for columnwise dropout as described in subsection 1.11.6.

Evoformer

a small class for calculating MD5 hashes of strings or byte arrays it is not meant to be fast or secure usage: 1) feed it blocks of uchars with updat

A class that encapsulates all the states for a random number generator using the philox_4x32_10 algorithm. Each invocation returns a 128-bit random bi

Implements Algorithm 16.

Implements Algorithm 14.

Implements Algorithm 13.

Implements Algorithm 12.

Implements Algorithm 11.

cpu execution engine only.