GB0107

Graph-Break Type

Short name describing what triggered the graph break

Tensor.random_ op

Context

Values or code snippet captured at the break point

Tensor.{name}(args={args}, kwargs={kwargs})

Explanation

Explanation of why the graph break was triggered

This is currently not supported.

Hints

Hints on how to resolve the graph break

• Use the out-of-place version of this op
• It may be possible to write Dynamo tracing rules for this code. Please report an issue to PyTorch if you encounter this graph break often and it is causing performance issues.

Additional Information

This graph break happens when the in-place Tensor.random_() method is called inside a compiled region. random_() is not currently supported by Dynamo, in part because it decomposes to an aten op whose name collides with a Python keyword.

Example code that causes the graph break is:

import torch


@torch.compile(fullgraph=True)
def fn(x):
    noise = torch.zeros_like(x)
    noise.random_(0, 10)
    return x + noise


fn(torch.randn(4))

Use an out-of-place factory function that returns a new random tensor instead of filling one in place. torch.randint_like covers the integer case that random_ is normally used for:

import torch


@torch.compile(fullgraph=True)
def fn(x):
    noise = torch.randint_like(x, 0, 10)
    return x + noise


fn(torch.randn(4))

This break is specific to Tensor.random_; other in-place RNG methods such as Tensor.uniform_ and Tensor.normal_ are supported and do not need to be rewritten.

If this is causing problems for you, please comment on https://github.com/pytorch/pytorch/issues/156614

Click here to add Additional Info

Back to Registry