PR #180580: doc: clarify that torch.normal does not support integer dtypes

• Repository: pytorch/pytorch
• Author: Abineshabee
• State: open | merged: False
• Link: https://github.com/pytorch/pytorch/pull/180580

Description (problem / solution / changelog)

Description

The documentation page for torch.normal does not mention that integer dtypes (e.g., torch.long, torch.int32) are not supported. This causes a confusing NotImplementedError on both CPU and CUDA with no explanation in the docs.

Error Reproduced

CPU: torch.normal(mean=0, std=1, size=(5,), dtype=torch.long) NotImplementedError: "normal_kernel_cpu" not implemented for 'Long'

CUDA: torch.normal(mean=0, std=1, size=(5,), dtype=torch.long, device='cuda') NotImplementedError: "normal_kernel_cuda" not implemented for 'Long'

What This PR Does

Added a .. note:: to the torch.normal docstring in torch/_torch_docs.py clarifying:

• Only floating-point dtypes are supported
• torch.set_default_dtype() does NOT affect integer tensor creation
• torch.tensor(1) always returns int64 — use torch.tensor(1.0) instead

How to Fix (for users)

Wrong - raises NotImplementedError

torch.normal(mean=torch.tensor(1), std=1, size=(5,))

Correct

torch.normal(mean=torch.tensor(1.0), std=1, size=(5,))

Linked Issue

Fixes #180560

Changes

• torch/_torch_docs.py — added dtype note to torch.normal docstring

Changed files

• torch/_torch_docs.py (modified, +2/-2)