[tensor-shapes] torch.maximum rejects an inline torch.where result with the same variadic shape

Abierto
#5,709 0 comentarios 0 reacciones 0 asignados Ver en GitHub

Nadie ha tomado este issue todavía.

Evaluación

Dificultad
4/5
Tiempo estimado
3-5 días
Aptitud para principiantes
30/100
Tipo de issue
Error
Claridad
Bastante claro
Estado de actividad
Activo
Stack tecnológico
python, pytorch, rust
Área
devtools

Línea de trabajo

Reproduce first via the sandbox link or the repro.py commands: the inline torch.where(...) and the named selected form must both report Tensor[S]. Then find the Rust checker's call-argument inference where a generic function's type-var tuple is solved — the named form solves it when binding selected, the inline form does so mid-argument. Done means pyrefly check passes on repro.py and the type-checker tests stay green. The payload names no source files, so start from call inference.

Escrito por el modelo de indexación a partir del texto del issue.

Descripción

pytorch tensor-shapes typechecking
Describe the Bug

Hej,

With Pyrefly and the PyTorch shape stubs at 1.5.0, I get a false positive when passing a torch.where result directly to torch.maximum. Giving that same result a local name makes the error disappear. Both forms should return Tensor[S]: the boolean mask and the two real-valued tensors all have shape S.

Reproduction

Save this as repro.py in an empty directory.

from __future__ import annotations

from typing import TYPE_CHECKING

import torch

if TYPE_CHECKING:
    from shape_extensions import IntTuple
    from torch import Tensor


def example[S: IntTuple](mask: Tensor[S], x: Tensor[S], y: Tensor[S]) -> Tensor[S]:
    return torch.maximum(torch.where(mask, x, y), y)

Run the following:

touch pyrefly.toml
uv venv --python 3.13.18 .venv
uv pip install --python .venv/bin/python --index https://download.pytorch.org/whl/cpu 'torch==2.14.0 cpu'
uv pip install --python .venv/bin/python pyrefly==1.5.0 pyrefly-torch-stubs==1.5.0
.venv/bin/pyrefly check --python-interpreter-path .venv/bin/python --output-format min-text repro.py

The checker exits with status 1 and underlines torch.where(mask, x, y).

ERROR repro.py:13:26-49: Argument `Tensor[S]` is not assignable to parameter `input` with type `Tensor[S]` in function `torch.maximum` [bad-argument-type]

Replacing the return statement with these two lines passes. reveal_type(selected) and reveal_type(torch.maximum(selected, y)) both report Tensor[S].

    selected = torch.where(mask, x, y)
    return torch.maximum(selected, y)

The inline form also passes when the rank is fixed, including Tensor[[N, M]] with two symbolic extents. A direct maximum on incompatible shapes (2, 3) and (3, 2) correctly fails. Calling the original function with same-shaped CPU tensors succeeds and gives the expected elementwise values.

Sandbox Link

https://pyrefly.org/sandbox/?project=v2.zVdda9swFP0rgr44I3U3xl4CG4VQtjC6lSUvwzNGdZTY1HZCbC_pYP995-pKluU66x72UEqJLV3dz3PPlXuUeVD7w65HmEmyaYEAqohtbHJNmkY21ODT6ur73U0y_3Qz_7z48rFHCeCXNDN19mRMQrWqOpN7lahToypKYEc1i6pZtejYnqjW1xmFPMOPG0hxg0fLWXc0DkpZP8yMaLSMp-LkvT323jRquzeLA-4GbTgs5Skv2zLgt2OmDkobgFaomtC_c-eYN1lS7FJZ_F-PQDCAj1oDF2f8-Ivn9jCL_SuW3xrosa79Yw7JCzHf4XC-zRoRpBNxqxop7grZoE1LTIdFlYZAzlrIDaCWSwAzFLeLFag5RUQq_FFdiG9q3aYIhYubKbFv70HdGZbehO_C14wN9F36ILcYQkQBTQaqA4M1eWpLDl3nZr2dl0biOaxNwReQpllLdaThKpKlduJ9D5A8UWxlaHvmxPghnokwhGMMBpN_K-vMfUXUh8EqdERxHBiCr_Zt4zChVSMsOtatOiVgmj5oumACvds3N7F4Qm3xd1OoUlXN5ZG43XgbCg4przZAGCaG-JlLseUrEsb_iWhC1GAwSU8ha2KtJvZrl3_dExqq6a5a50QnXQAoFYO9ofgjXp2SEpvGMUURm5pbdcPk8vaCEji-NZZ8bNnUP_XTNxUbNX6JnD27f7ZYZv-VlXNa_LGP_fGq-v70wBs4L7yio-qcT3TzsBeuesPkpXZ42N1D64JGdoJfd8WzNTQx-ou4FfYakoNvDrKqKZxoFQcbqjPnuRO9Hsjx4W1Ll7XEZbveq1RfQnGbogOw1ZnVGjvP-j0335UoktJpQe3xbLKxVnV6yO-Ro_tHYTpOFvkvLGjLIz3ncT5lJUxGvbQOmkE1Hp87UqgNMOnoSqPBLYzFRlb04Agm02By-QGXN8MJbJkYkvzrEhRoK0b3xJ9ez8UwBmNS_nIRfOYTzX6YDXHezQQCu31ZppkqJflvUujt1rz9tC3qYChj7_38NVSpIz6G0gKZ0bd_QrFxKxrYZrYeqvM6bBx_gy4ZapiKK1-zp_jJlba7u_7-Aw

(Only applicable for extension issues) IDE Information

No response

Lenguaje dominante
Rust
Estrellas
7.2k
Forks
578
Métricas de merge de PR
Sin PR fusionados en 30 d

Preparar el entorno

Primeros pasos

  1. Lee el issue completo y luego la guía de contribución del proyecto.
  2. Comenta en el issue que vas a ocuparte — evita que dos personas hagan lo mismo.
  3. Haz un fork del repositorio y trabaja en una rama.
  4. Abre un pull request que haga referencia al número del issue.

Más de facebook/pyrefly

Todos los issues de facebook/pyrefly

Issues similares

Más issues de Rust

Recibe los nuevos issues en tu correo

Un resumen breve de issues de GitHub para principiantes.