You're on Linux (manylinux_2_39_x86_64), but torch (v2.5.1) only has wheels for the following platform: linux_aarch64

[sglbl]

Summary

When I try to install requirements (that has pytorch with cuda) from pyproject.toml, I got this error.

$ uv sync --extra gpu
error: Distribution `torch==2.5.1 @ registry+https://download.pytorch.org/whl/cu124` can't be installed 
because it doesn't have a source distribution or wheel for the current platform
hint: You're on Linux (`manylinux_2_39_x86_64`), 
but `torch` (v2.5.1) only has wheels for the following platform: `linux_aarch64`

No one on other machine has touched to this pyproject toml / uv.lock file. I solved the issue by removing uv.lock.
The problem is all the time I have this kind of problems with uv.lock, when I remove it it works.
But since I installed torch==2.5.1 before and uv added this as "torch>=2.5.1" , after removing uv.lock it insalls torch==2.6.0+cu124.

Other than this many_linux and linux uncompatibility bug, is there a way to add a specific version to pyproject.toml with uv (without editing toml file by hand)?

Platform

Linux Ubuntu 24.04

Version

uv 0.5.29

Python version

Python 3.10.12

[charliermarsh]

You mean like uv add torch==2.6.0? Or uv add torch==2.6.0+cu124?

(Unfortunately I can't help further without a complete reproduction, like a pyproject.toml and a series of commands that reproduces the issue.)

[zanieb]

Note the manylinux_2_39_x86_64 tag is x86_64 while the other is linux_aarch64 which is arm64 — that's the incompatibility.

[ldng]

Same issue.

Archlinux | glibc-2.41+r6+gcf88351b685d-1 | linux-6.13.2.arch1-1 | Arch: X86_64

Minimal pyproject.toml to reproduce :

[project]
name = "aarch64-issue"
version = "0.1.0"
description = "Add your description here"
readme = "README.md"
requires-python = "==3.10.*"
dependencies = [
    "torch",
]


[[tool.uv.index]]
name = "pytorch-rocm"
url = "https://download.pytorch.org/whl/rocm6.2"
explicit = true

[tool.uv.sources]
torch = [
  { index = "pytorch-rocm", marker = "sys_platform == 'linux'" },
]

$ uv sync --verbose

DEBUG uv 0.5.30 (ddbc6e315 2025-02-10)
DEBUG Found project root: `/tmp/aarch64_issue`
DEBUG No workspace root found, using project root
DEBUG Acquired lock for `/tmp/aarch64_issue`
DEBUG Reading Python requests from version file at `/tmp/aarch64_issue/.python-version`
DEBUG Using Python request `3.10` from version file at `.python-version`
DEBUG Checking for Python environment at `.venv`
DEBUG Searching for Python 3.10 in managed installations or search path
DEBUG Searching for managed installations at `~/.local/share/uv/python`
DEBUG Skipping incompatible managed installation `cpython-3.14.0a4-linux-x86_64-gnu`
DEBUG Skipping incompatible managed installation `cpython-3.12.8-linux-x86_64-gnu`
DEBUG Skipping incompatible managed installation `cpython-3.11.11-linux-x86_64-gnu`
DEBUG Found managed installation `cpython-3.10.16-linux-x86_64-gnu`
DEBUG Found `cpython-3.10.16-linux-x86_64-gnu` at `~/.local/share/uv/python/cpython-3.10.16-linux-x86_64-gnu/bin/python3.10` (managed installations)
Using CPython 3.10.16
Creating virtual environment at: .venv
DEBUG Assessing Python executable as base candidate: ~/.local/share/uv/python/cpython-3.10.16-linux-x86_64-gnu/bin/python3.10
DEBUG Using base executable for virtual environment: ~/.local/share/uv/python/cpython-3.10.16-linux-x86_64-gnu/bin/python3.10
DEBUG Released lock at `/tmp/uv-6ac5999c7feb2517.lock`
DEBUG Using request timeout of 30s
DEBUG Found static `pyproject.toml` for: aarch64-issue @ file:///tmp/aarch64_issue
DEBUG No workspace root found, using project root
DEBUG Existing `uv.lock` satisfies workspace requirements
Resolved 10 packages in 0.34ms
error: Distribution `torch==2.0.1 @ registry+https://download.pytorch.org/whl/rocm6.2` can't be installed because it doesn't have a source distribution or wheel for the current platform

hint: You're on Linux (`manylinux_2_40_x86_64`), but `torch` (v2.0.1) only has wheels for the following platform: `manylinux2014_aarch64`

[charliermarsh]

You need this:

[project]
name = "aarch64-issue"
version = "0.1.0"
description = "Add your description here"
readme = "README.md"
requires-python = "==3.10.*"
dependencies = [
    "torch>=2.5.1",
    "pytorch-triton-rocm",
]


[[tool.uv.index]]
name = "pytorch-rocm"
url = "https://download.pytorch.org/whl/rocm6.2"
explicit = true

[tool.uv.sources]
torch = [
  { index = "pytorch-rocm", marker = "sys_platform == 'linux'" },
]
pytorch-triton-rocm = [
  { index = "pytorch-rocm", marker = "sys_platform == 'linux'" },
]

You're using an explicit index, but pytorch-triton-rocm doesn't exist on PyPI at the version you need, so the resolver is backtracking.

[charliermarsh]

(I don't think your issue is related? OP is asking about the cu124 index.)

[sglbl]

This is my toml file.

[project]
name = "ex"
version = "0.1.0"
description = "example Project"
readme = "README.md"
requires-python = ">=3.10"
dependencies = [
    "aiohttp>=3.11.11",
    "fastapi>=0.115.6",
    "gradio==5.6.0",
    "gradio-log>=0.0.8",
    "haystack-ai==2.4.0",
    "lingua-language-detector==2.0.2",
    "loguru>=0.7.3",
    "msal>=1.31.1",
    "munch>=4.0.0",
    "ollama-haystack==0.0.7",
    "openai>=1.59.4",
    "openpyxl>=3.1.5",
    "psycopg2>=2.9.10",
    "pypdf>=5.1.0",
    "pytest>=8.3.4",
    "python-dotenv>=1.0.1",
    "qdrant-haystack>=7.0.0",
    "ring>=0.10.1",
    "sentence-transformers>=3.3.1",
    "sqlalchemy==2.0.31",
    "sqlitedict>=2.1.0",
    "tabulate>=0.9.0",
    "tiktoken>=0.8.0",
    "transformers>=4.47.1",
    "unidecode>=1.3.8",
    "xlsxwriter>=3.2.1",
]


[dependency-groups]
dev = [
    "jupyter>=1.1.1",
    "pypdf2>=3.0.1",
]

[project.optional-dependencies]
cpu = [
  "torch==2.5.1",
]
gpu = [
  "torch==2.5.1",
]

[tool.uv]
conflicts = [ # type:ignore
  [
    { extra = "cpu" },
    { extra = "gpu" },
  ],
]

[tool.uv.sources]
torch = [
  { index = "pytorch-cpu", extra = "cpu", marker = "platform_system != 'Darwin'" },
  { index = "pytorch-gpu", extra = "gpu" },
]

[[tool.uv.index]]
name = "pytorch-cpu"
url = "https://download.pytorch.org/whl/cpu"
explicit = true

[[tool.uv.index]]
name = "pytorch-gpu"
url = "https://download.pytorch.org/whl/cu124"
explicit = true

[tool.pytest.ini_options]
pythonpath = "." # Add the project root path to the pythonpath [to be able to import the 'src' module]
filterwarnings = ["ignore::DeprecationWarning", "ignore::FutureWarning", "ignore::PendingDeprecationWarning"]

On the first install everything is fine (always). But when I add a new package and change from gpu to cpu version makes some changes on uv.lock file and then new installations create problem. So I need to remove uv.lock. (Unfortunately I don't remember the exact order of reproducing the issue or if it was because of an exact package(such as lingua..)) But both of my computers that I use for this project is has x86_64 so it shouldn't even have linux_aarch64:

~$ uname -m
x86_64

@charliermarsh My second question was: when I use uv add torch, it will install last version (for example 2.6.0+cu124) and add >=2.6.0+cu124 to toml. But if I remove the lock and some months later if I reinstall from my toml, it might install a uncompatible newer version which I don't want. So is there a way to add it to toml without specifying the version? (This question is not that important for me, just a side question)

[ldng]

(I don't think your issue is related? OP is asking about the cu124 index.)

I'll take your word for it. Still, the "aarch64" hint is confusing in both cases.

[zanieb]

@ldng The hint is just a fact about what wheels are available for that version...

You're on x86-64, only aarch64 wheels are available for the version in the lockfile — those are incompatible architectures.

The fix is to adjust your dependencies so we do not lock a version that cannot be used on your current platform.

[zanieb]

I'm happy to improve the hint though, do you have a suggestion that would clarify things for you?

[sglbl]

@charliermarsh If you cannot reproduce, if you want I can close the issue until somebody else encounters this issue and opens it again

[charliermarsh]

I think this could plausibly be improved by #11546 which I just merged. If you encounter it again and have a reproduction, I'm always happy to re-open.