tensorflow /

TensorFlow

Live map
View repository
Snapshotmaster · 45bf59d

Two execution paths from Python tensors: immediate eager operations and optimized `tf.function` graphs.

Click a building or route
Choose a path

Follow a Python call immediately through the eager runtime into a registered device kernel and back as concrete tensor handles.

System map8 components · 9 connections
130%
TensorFlow architectureThe Python API starts at upper left. Eager calls take the short center route to kernels; decorated functions travel through tracing, function storage, Grappler, and the graph executor before joining the same kernel and device layer.op name, eager tensors…operation inputs, attr…Tensor buffers, alloca…output tensor handles …Pythontensor APIEager bridgeDevice managerOpKernelregistryGraph executortf.functiontracerFunctionlibraryGrappleroptimizer
Connections
BuildingsEntryServiceComputeStorageRuntimeToolingExternal

Terms, in plain English

Eager executionRunning each TensorFlow operation immediately so Python receives concrete tensor results step by step.

How to study this repo

  1. 1
    Compare eager with traced calls

    Run a tiny function normally and under tf.function, then inspect tracing count and the ConcreteFunction graph operations.

  2. 2
    Follow one OpKernel

    Choose a simple operation, find its kernel registration, and connect its input tensors and device context to op_kernel.cc.

  3. 3
    Observe a graph rewrite

    Enable graph optimizer diagnostics and identify one Grappler change before following the optimized node into executor.cc.