A Tensor's Autograd Life
LifecycleA tensor's autograd states, from recorded to walked to released, plus the version-mismatch recovery loop.
8 states · 4 guided chapters
Key facts
What gets recorded
- grad_fn (XxxBackward), next_edges and the input versions
- A leaf gets AccumulateGrad as its grad_fn on first use
Ways back
- VersionMismatch clears once forward re-runs and builds a new graph
- Only a graph kept with retain_graph=True can be walked again
Guided chapters
- 01
Record to .grad
A leaf used in an op builds a graph; backward fills the grads, then the graph is released.
- 02
Keeping the graph
Only a graph kept with retain_graph=True can run backward again.
- 03
Version mismatch
An in-place change to a saved tensor stops the walk; re-running forward recovers.
- 04
Nothing recorded
With GradMode off (no_grad, inference_mode) the kernels skip node creation entirely.