GB4198

Graph-Break Type

Short name describing what triggered the graph break

Attempted to copy.deepcopy a tensor

Context

Values or code snippet captured at the break point

copy.deepcopy({self})

Explanation

Explanation of why the graph break was triggered

Dynamo does not support copy.deepcopy() on tensors.

Hints

Hints on how to resolve the graph break

• Avoid calling copy.deepcopy() on tensors inside compiled regions.

Additional Information

This graph break happens when copy.deepcopy() is called on a tensor inside a compiled region. This often shows up indirectly, for example when code snapshots state by deep-copying a container that holds tensors.

Example code that causes the graph break is:

import copy
import torch


@torch.compile(fullgraph=True)
def fn(x):
    saved = copy.deepcopy(x)
    return x + saved


fn(torch.randn(3))

Use tensor.clone() instead. clone() is differentiable and is captured into the graph, so it gives you an independent copy without breaking:

import torch


@torch.compile(fullgraph=True)
def fn(x):
    saved = x.clone()
    return x + saved


fn(torch.randn(3))

If you are deep-copying a container (e.g. a list or dict of tensors), replace the copy.deepcopy with a comprehension that clones each tensor:

import torch


@torch.compile(fullgraph=True)
def fn(xs):
    saved = [t.clone() for t in xs]
    return [a + b for a, b in zip(xs, saved)]


fn([torch.randn(3), torch.randn(3)])

Click here to add Additional Info

Back to Registry