How does TensorFlow execute computations internally in graph mode vs eager mode?

Graph mode builds a static computation graph, while eager mode executes operations immediately.

In graph mode, TensorFlow constructs a Directed Acyclic Graph (DAG) where nodes represent operations and edges represent tensors. This graph is optimized before execution (fusion, pruning, device placement). In eager mode, each operation is executed immediately in Python, providing better debugging but lower optimization potential. Graph mode allows XLA compilation, distributed execution, and kernel fusion, making it significantly more efficient for production workloads.