uv Cannot install the CPU version of torch

Cannot install the CPU version of torch

Open Nishikoh opened this issue 6 months ago • 21 comments

I tried to install the CPU version of torch but could not.

uv pip install torch==2.1.0+cpu --find-links https://download.pytorch.org/whl/torch_stable.html
# and next command gives the same result.
uv pip install torch==2.1.0 --index-url https://download.pytorch.org/whl/cpu

  × No solution found when resolving dependencies:
  ╰─▶ Because there is no version of torch==2.1.0+cpu and you require torch==2.1.0+cpu, we can conclude that the requirements are unsatisfiable.

uv version: v0.1.2 Python 3.11.7 Ubuntu 20.4 X86_64 Architecture

By the way, the install of the GPU version of torch is successful.

uv pip install torch==2.1.0
# success

Feb 16 '24 14:02 Nishikoh

Thanks! Will take a look.

Feb 16 '24 14:02 charliermarsh

Here's verbose output on my machine

❯ echo "torch==2.1.0+cpu" | uv pip compile --find-links https://download.pytorch.org/whl/torch_stable.html -
  × No solution found when resolving dependencies:
  ╰─▶ Because there is no version of torch==2.1.0+cpu and you require torch==2.1.0+cpu, we can conclude that the requirements are unsatisfiable.
❯ echo "torch==2.1.0+cpu" | uv pip compile --find-links https://download.pytorch.org/whl/torch_stable.html - -v
    0.000491s DEBUG uv_interpreter::virtual_env Found a virtualenv through VIRTUAL_ENV at: /Users/mz/eng/src/astral-sh/uv/.venv
    0.000781s DEBUG uv_interpreter::interpreter Using cached markers for: /Users/mz/eng/src/astral-sh/uv/.venv/bin/python
    0.000791s DEBUG uv::commands::pip_compile Using Python 3.9.6 interpreter at /Users/mz/eng/src/astral-sh/uv/.venv/bin/python for builds
 uv_client::cached_client::get_serde 
   uv_client::cached_client::get_cacheable 
     uv_client::cached_client::read_and_parse_cache file=/Users/mz/Library/Caches/uv/flat-index-v0/html/76f6661a0037bb05.msgpack
        0.010378s   0ms  WARN uv_client::cached_client Broken cache entry at /Users/mz/Library/Caches/uv/flat-index-v0/html/76f6661a0037bb05.msgpack, removing: failed to open file `/Users/mz/Library/Caches/uv/flat-index-v0/html/76f6661a0037bb05.msgpack`
        0.010766s   0ms DEBUG uv_client::cached_client No cache entry for: https://download.pytorch.org/whl/torch_stable.html
     uv_client::cached_client::fresh_request url="https://download.pytorch.org/whl/torch_stable.html"
     uv_client::cached_client::new_cache file=/Users/mz/Library/Caches/uv/flat-index-v0/html/76f6661a0037bb05.msgpack
     uv_client::flat_index::parse_flat_index_html url=https://download.pytorch.org/whl/torch_stable.html
       uv_client::html::parse url=Url { scheme: "https", cannot_be_a_base: false, username: "", password: None, host: Some(Domain("download.pytorch.org")), port: None, path: "/whl/torch_stable.html", query: None, fragment: None }
    0.318613s DEBUG uv_client::flat_index Found 2243 packages in `--find-links` entry: https://download.pytorch.org/whl/torch_stable.html
 uv_client::flat_index::from_entries 
 uv_resolver::resolver::solve 
      0.322186s   1ms DEBUG uv_resolver::resolver Solving with target Python version 3.9.6
   uv_resolver::resolver::choose_version package=root
   uv_resolver::resolver::get_dependencies package=root, version=0a0.dev0
        0.322468s   0ms DEBUG uv_resolver::resolver Adding direct dependency: torch==2.1.0+cpu
   uv_resolver::resolver::choose_version package=torch
     uv_resolver::resolver::package_wait package_name=torch
 uv_resolver::resolver::process_request request=Versions torch
   uv_client::registry_client::simple_api package=torch
     uv_client::cached_client::get_cacheable 
       uv_client::cached_client::read_and_parse_cache file=/Users/mz/Library/Caches/uv/simple-v0/pypi/torch.rkyv
 uv_resolver::resolver::process_request request=Prefetch torch ==2.1.0+cpu
          0.323335s   0ms DEBUG uv_client::cached_client Found fresh response for: https://pypi.org/simple/torch
 uv_resolver::version_map::from_metadata 
        0.324165s   1ms DEBUG uv_resolver::resolver Searching for a compatible version of torch (==2.1.0+cpu)
      0.324187s   3ms DEBUG uv_resolver::resolver No compatible version found for: torch
  × No solution found when resolving dependencies:
  ╰─▶ Because there is no version of torch==2.1.0+cpu and you require torch==2.1.0+cpu, we can conclude that the requirements are unsatisfiable.

Feb 16 '24 15:02 zanieb

You can see these versions are available at https://download.pytorch.org/whl/torch_stable.html

Adding a debug log and it looks like we're filtering these out before candidate selection

        0.382322s   4ms DEBUG uv_resolver::resolver Searching for a compatible version of torch (==2.1.0+cpu)
        0.382340s   4ms DEBUG uv_resolver::candidate_selector Checking 2.2.0
        0.382353s   4ms DEBUG uv_resolver::candidate_selector Checking 2.1.2
        0.382366s   4ms DEBUG uv_resolver::candidate_selector Checking 2.1.1
        0.382378s   4ms DEBUG uv_resolver::candidate_selector Checking 2.1.0
        0.382391s   4ms DEBUG uv_resolver::candidate_selector Checking 2.0.1
        0.382404s   4ms DEBUG uv_resolver::candidate_selector Checking 2.0.0
        0.382417s   4ms DEBUG uv_resolver::candidate_selector Checking 1.13.1
        0.382429s   4ms DEBUG uv_resolver::candidate_selector Checking 1.13.0
        0.382442s   4ms DEBUG uv_resolver::candidate_selector Checking 1.12.1
        0.382454s   4ms DEBUG uv_resolver::candidate_selector Checking 1.12.0
        0.382466s   4ms DEBUG uv_resolver::candidate_selector Checking 1.11.0
        0.382479s   4ms DEBUG uv_resolver::candidate_selector Checking 1.10.2
        0.382491s   4ms DEBUG uv_resolver::candidate_selector Checking 1.10.1
        0.382503s   4ms DEBUG uv_resolver::candidate_selector Checking 1.10.0
        0.382516s   4ms DEBUG uv_resolver::candidate_selector Checking 1.9.1
        0.382528s   4ms DEBUG uv_resolver::candidate_selector Checking 1.9.0
        0.382540s   4ms DEBUG uv_resolver::candidate_selector Checking 1.8.1
        0.382552s   4ms DEBUG uv_resolver::candidate_selector Checking 1.8.0
        0.382564s   4ms DEBUG uv_resolver::candidate_selector Checking 1.7.1
        0.382577s   4ms DEBUG uv_resolver::candidate_selector Checking 1.7.0
        0.382589s   4ms DEBUG uv_resolver::candidate_selector Checking 1.6.0
        0.382601s   4ms DEBUG uv_resolver::candidate_selector Checking 1.5.1
        0.382613s   4ms DEBUG uv_resolver::candidate_selector Checking 1.5.0
        0.382625s   4ms DEBUG uv_resolver::candidate_selector Checking 1.4.0
        0.382638s   4ms DEBUG uv_resolver::candidate_selector Checking 1.3.1
        0.382650s   4ms DEBUG uv_resolver::candidate_selector Checking 1.3.0.post2
        0.382663s   4ms DEBUG uv_resolver::candidate_selector Checking 1.3.0
        0.382675s   4ms DEBUG uv_resolver::candidate_selector Checking 1.2.0
        0.382687s   4ms DEBUG uv_resolver::candidate_selector Checking 1.1.0.post2
        0.382699s   4ms DEBUG uv_resolver::candidate_selector Checking 1.1.0
        0.382712s   4ms DEBUG uv_resolver::candidate_selector Checking 1.0.1.post2
        0.382725s   4ms DEBUG uv_resolver::candidate_selector Checking 1.0.1
        0.382737s   4ms DEBUG uv_resolver::candidate_selector Checking 1.0.0
        0.382749s   4ms DEBUG uv_resolver::candidate_selector Checking 0.4.1.post2
        0.382762s   4ms DEBUG uv_resolver::candidate_selector Checking 0.4.1
        0.382774s   4ms DEBUG uv_resolver::candidate_selector Checking 0.4.0
        0.382786s   4ms DEBUG uv_resolver::candidate_selector Checking 0.3.1
        0.382798s   4ms DEBUG uv_resolver::candidate_selector Checking 0.3.0.post4
        0.382812s   4ms DEBUG uv_resolver::candidate_selector Checking 0.3.0.post3
        0.382825s   4ms DEBUG uv_resolver::candidate_selector Checking 0.3.0.post2
        0.382841s   4ms DEBUG uv_resolver::candidate_selector Checking 0.3.0
        0.382854s   4ms DEBUG uv_resolver::candidate_selector Checking 0.2.0.post3
        0.382867s   4ms DEBUG uv_resolver::candidate_selector Checking 0.2.0.post2
        0.382881s   4ms DEBUG uv_resolver::candidate_selector Checking 0.2.0.post1
        0.382894s   4ms DEBUG uv_resolver::candidate_selector Checking 0.1.12.post2
        0.382907s   5ms DEBUG uv_resolver::candidate_selector Checking 0.1.12.post1
        0.382921s   5ms DEBUG uv_resolver::candidate_selector Checking 0.1.11.post5
        0.382933s   5ms DEBUG uv_resolver::candidate_selector Checking 0.1.11.post4
        0.382946s   5ms DEBUG uv_resolver::candidate_selector Checking 0.1.10.post2
        0.382960s   5ms DEBUG uv_resolver::candidate_selector Checking 0.1.10.post1
        0.382974s   5ms DEBUG uv_resolver::candidate_selector Checking 0.1.9.post2
        0.382987s   5ms DEBUG uv_resolver::candidate_selector Checking 0.1.9.post1
        0.383001s   5ms DEBUG uv_resolver::candidate_selector Checking 0.1.8.post1
        0.383015s   5ms DEBUG uv_resolver::candidate_selector Checking 0.1.7.post2
        0.383029s   5ms DEBUG uv_resolver::candidate_selector Checking 0.1.6.post22
        0.383042s   5ms DEBUG uv_resolver::candidate_selector Checking 0.1.6.post20
        0.383082s   5ms DEBUG uv_resolver::candidate_selector Checking 0.1.6.post17

Feb 16 '24 15:02 zanieb

Ya there are definitely versions that should be installable available but I'm not getting any luck on Linux x86_64, worked fine on an M1 MacBook though.

uv pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu

Feb 16 '24 20:02 justinh-rahb

Oops, I guess I didn't provide enough information about my environment. I use the following environment.

Python 3.11.7 Ubuntu 20.4 X86_64 Architecture

Feb 16 '24 21:02 Nishikoh

There is a way to install torch by typing the URL directly. I tried that too, but it failed.

$ uv pip install torch@https://download.pytorch.org/whl/cpu/torch-2.1.0%2Bcpu-cp311-cp311-linux_x86_64.whl#sha256=5954924ce74bc7e6a6c811e3fa4bdda9936d9889f6369fd068420c444bfd1cae

error: The wheel filename "torch-2.1.0%2Bcpu-cp311-cp311-linux_x86_64.whl" has an invalid version part: after parsing 2.1.0, found "%2Bcpu" after it, which is not part of a valid version

Feb 16 '24 22:02 Nishikoh

Converting the string "%2Bcpu" to "+cpu" and executing it produces another error.

$ uv pip install torch@https://download.pytorch.org/whl/cpu/torch-2.1.0+cpu-cp311-cp311-linux_x86_64.whl#sha256=5954924ce74bc7e6a6c811e3fa4bdda9936d9889f6369fd068420c444bfd1cae

error: Failed to download: torch @ https://download.pytorch.org/whl/cpu/torch-2.1.0+cpu-cp311-cp311-linux_x86_64.whl#sha256=5954924ce74bc7e6a6c811e3fa4bdda9936d9889f6369fd068420c444bfd1cae
  Caused by: HTTP status client error (403 Forbidden) for url (https://download.pytorch.org/whl/cpu/torch-2.1.0+cpu-cp311-cp311-linux_x86_64.whl#sha256=5954924ce74bc7e6a6c811e3fa4bdda9936d9889f6369fd068420c444bfd1cae)

Feb 16 '24 22:02 Nishikoh

I think part of the problem here is solved by https://github.com/astral-sh/uv/pull/1544 (but perhaps not all of it). Do you mind trying again once that's out later today?

Feb 16 '24 22:02 charliermarsh

Thank you for working on the problem!

I think part of the problem here is solved by https://github.com/astral-sh/uv/pull/1544 (but perhaps not all of it). Do you mind trying again once that's out later today?

I'm actually going on a two-day trip starting today. So I won't be able to check it until after that. Sorry about that. If anyone is tracking this issue, I would be grateful if you could check on my behalf.

Feb 16 '24 22:02 Nishikoh

I'm actually going on a two-day trip starting today. So I won't be able to check it until after that. Sorry about that. If anyone is tracking this issue, I would be grateful if you could check on my behalf.

I got you fam 👌

Feb 16 '24 22:02 justinh-rahb

Do you mind giving this at try (at your convenience) with v0.1.3?

Feb 17 '24 01:02 charliermarsh

Just using the uv tool on my bare Linux machine is working now, but I'm still having a heck of a time making things work in Docker containers. Keeps trying to install six from the PyTorch index, which of course fails because there is no such package there. I did try to explicitly set the version of torch as well, no luck.

Verbose log: https://pastebin.com/SQh4cLQS

Feb 17 '24 04:02 justinh-rahb

I have tried three different INSTALL methods and have confirmed that they all work 👏

uv version: v0.1.4

$ uv pip install torch@https://download.pytorch.org/whl/cpu/torch-2.1.0%2Bcpu-cp311-cp311-linux_x86_64.whl#sha256=5954924ce74bc7e6a6c811e3fa4bdda9936d9889f6369fd068420c444bfd1cae

Resolved 9 packages in 4.46s
Downloaded 1 package in 4.10s
Installed 9 packages in 149ms
 + filelock==3.13.1
 + fsspec==2024.2.0
 + jinja2==3.1.3
 + markupsafe==2.1.5
 + mpmath==1.3.0
 + networkx==3.2.1
 + sympy==1.12
 + torch==2.1.0+cpu (from https://download.pytorch.org/whl/cpu/torch-2.1.0%2Bcpu-cp311-cp311-linux_x86_64.whl#sha256=5954924ce74bc7e6a6c811e3fa4bdda9936d9889f6369fd068420c444bfd1cae)
 + typing-extensions==4.9.0

The next two methods worked in the same way 👍

$ uv pip install torch==2.1.0+cpu --index-url https://download.pytorch.org/whl/cpu
$ uv pip install torch==2.1.0 --index-url https://download.pytorch.org/whl/cpu

Great project and a big thank you to the maintainers!

edit

The following command failed

$ uv pip install torch==2.1.0 --index-url https://download.pytorch.org/whl/cpu

The following command succeeds

uv pip install torch==2.1.0+cpu --find-links https://download.pytorch.org/whl/torch_stable.html

Feb 19 '24 02:02 Nishikoh

Verbose log: https://pastebin.com/SQh4cLQS

I tried to check this log on my end, but also failed. I tried explicitly specifying the version and cpu, but no luck.

$ uv pip install torchvision==0.16.1+cpu torchaudio==2.1.1+cpu torch==2.1.1+cpu --index-url https://download.pytorch.org/whl/cpu

  x No solution found when resolving dependencies:
  `-> Because torchvision==0.16.1+cpu depends on torch==2.1.1 and you require torch==2.1.1+cpu, we can conclude that root==0a0.dev0 and torchvision==0.16.1+cpu are incompatible.
      And because you require torchvision==0.16.1+cpu, we can conclude that the requirements are unsatisfiable.

Feb 19 '24 02:02 Nishikoh

I have a similar problem when trying to use uv pip compile:

In pyproject.toml I have the +cpu dependencies specifed, and they are available on our internal Artifactory (installation works without an issue), but the compilation fails for some reason...

EDIT: Using v0.1.5 btw.

EDIT2: Identical error and message happen when trying to install a package with torch+cpu dependencies, although "direct" installation of +cpu dependencies (uv pip install ...) works without issues.

Feb 19 '24 14:02 baskervilski

Just to clarify, uv pip install works but uv pip compile does not? What's the exact compile invocation and input file?

Feb 19 '24 15:02 zanieb

So.

uv pip install torch==2.1.1+cpu => works 🟢
uv pip install -e . (where the local project has torch==2.1.1+cpu as a dependency) => fails. 🔴
uv pip compile -o requirements.txt pyproject.toml (where the local project has torch==2.1.1+cpu as a dependency) => fails. 🔴

Error message (eviaea-api is my project): 305975352-db46cf10-25bd-4f27-a541-399bba7c210c

Feb 19 '24 17:02 baskervilski

Ok, found a workaround(?), using the --override option:

I created the overrides.txt file with the following content

# overrides.txt
torch==2.1.1+cpu
torchvision==0.16.1+cpu

Then called: uv pip compile pyproject.toml -o requirements.txt --override overrides.txt and uv pip install -e . --override overrides.txt

For now I can live with this.

Feb 22 '24 13:02 baskervilski

Thanks @baskervilski - that makes sense to me actually (that it works), I believe the bug here relates to our handling of local versions which we’ll need to fix.

Feb 22 '24 13:02 charliermarsh

OK, so I'm looking into this (and still catching up conceptually with how #1855 ties into this) and thought it might be helpful to just start sharing what I'm seeing. At least initially, I suspect I'm not sharing anything that @charliermarsh doesn't already know. :)

So what I did here was to try and isolate where the problem is coming from. I was able to get one of the commands shared above failing for me (I'm on Linux x86-64):

$ uv pip install torch==2.1.1 --index-url https://download.pytorch.org/whl/cpu
  x No solution found when resolving dependencies:
  `-> Because torch==2.1.1 is unusable because no wheels are available with a matching platform and you require torch==2.1.1, we can conclude that the requirements are unsatisfiable.

And specifically, pip install with the same command works fine:

$ pip install torch==2.1.1 --index-url https://download.pytorch.org/whl/cpu
Looking in indexes: https://download.pytorch.org/whl/cpu
Collecting torch==2.1.1
  Using cached https://download.pytorch.org/whl/cpu/torch-2.1.1%2Bcpu-cp311-cp311-linux_x86_64.whl (184.9 MB)
Collecting filelock (from torch==2.1.1)
  Using cached https://download.pytorch.org/whl/filelock-3.9.0-py3-none-any.whl (9.7 kB)
Collecting typing-extensions (from torch==2.1.1)
  Using cached https://download.pytorch.org/whl/typing_extensions-4.8.0-py3-none-any.whl (31 kB)
Collecting sympy (from torch==2.1.1)
  Using cached https://download.pytorch.org/whl/sympy-1.12-py3-none-any.whl (5.7 MB)
Collecting networkx (from torch==2.1.1)
  Using cached https://download.pytorch.org/whl/networkx-3.2.1-py3-none-any.whl (1.6 MB)
Collecting jinja2 (from torch==2.1.1)
  Using cached https://download.pytorch.org/whl/Jinja2-3.1.2-py3-none-any.whl (133 kB)
Collecting fsspec (from torch==2.1.1)
  Using cached https://download.pytorch.org/whl/fsspec-2023.4.0-py3-none-any.whl (153 kB)
Collecting MarkupSafe>=2.0 (from jinja2->torch==2.1.1)
  Using cached https://download.pytorch.org/whl/MarkupSafe-2.1.3-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (28 kB)
Collecting mpmath>=0.19 (from sympy->torch==2.1.1)
  Using cached https://download.pytorch.org/whl/mpmath-1.3.0-py3-none-any.whl (536 kB)
Installing collected packages: mpmath, typing-extensions, sympy, networkx, MarkupSafe, fsspec, filelock, jinja2, torch
Successfully installed MarkupSafe-2.1.3 filelock-3.9.0 fsspec-2023.4.0 jinja2-3.1.2 mpmath-1.3.0 networkx-3.2.1 sympy-1.12 torch-2.1.1+cpu typing-extensions-4.8.0

[notice] A new release of pip is available: 23.2.1 -> 24.0
[notice] To update, run: pip install --upgrade pip

So I started taking a look at why uv doesn't see torch 2.1.1. To start with, I wanted to see with what our VersionMap was being populated with. So I printed out debug info for wheels with versions that contain 2.1.1 in them. I got this:

"2.1.1": PackageName("torch"), wheel version: "2.1.1" python tags: cp310, abi tags: cp310, platform tags: manylinux_2_17_aarch64, manylinux2014_aarch64
    filename: "torch-2.1.1-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl"
    requires_python: None
    url: /whl/cpu/torch-2.1.1-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl#sha256=84fefd63356416c0cd20578637ccdbb82164993400ed17b57c951dd6376dcee8
"2.1.1": PackageName("torch"), wheel version: "2.1.1" python tags: cp310, abi tags: none, platform tags: macosx_10_9_x86_64
    filename: "torch-2.1.1-cp310-none-macosx_10_9_x86_64.whl"
    requires_python: None
    url: /whl/cpu/torch-2.1.1-cp310-none-macosx_10_9_x86_64.whl#sha256=1e1e5faddd43a8f2c0e0e22beacd1e235a2e447794d807483c94a9e31b54a758
"2.1.1": PackageName("torch"), wheel version: "2.1.1" python tags: cp310, abi tags: none, platform tags: macosx_11_0_arm64
    filename: "torch-2.1.1-cp310-none-macosx_11_0_arm64.whl"
    requires_python: None
    url: /whl/cpu/torch-2.1.1-cp310-none-macosx_11_0_arm64.whl#sha256=e76bf3c5c354874f1da465c852a2fb60ee6cbce306e935337885760f080f9baa
"2.1.1": PackageName("torch"), wheel version: "2.1.1" python tags: cp311, abi tags: cp311, platform tags: manylinux_2_17_aarch64, manylinux2014_aarch64
    filename: "torch-2.1.1-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl"
    requires_python: None
    url: /whl/cpu/torch-2.1.1-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl#sha256=61b51b33c61737c287058b0c3061e6a9d3c363863e4a094f804bc486888a188a
"2.1.1": PackageName("torch"), wheel version: "2.1.1" python tags: cp311, abi tags: none, platform tags: macosx_10_9_x86_64
    filename: "torch-2.1.1-cp311-none-macosx_10_9_x86_64.whl"
    requires_python: None
    url: /whl/cpu/torch-2.1.1-cp311-none-macosx_10_9_x86_64.whl#sha256=a70593806f1d7e6b53657d96810518da0f88ef2608c98a402955765b8c79d52c
"2.1.1": PackageName("torch"), wheel version: "2.1.1" python tags: cp311, abi tags: none, platform tags: macosx_11_0_arm64
    filename: "torch-2.1.1-cp311-none-macosx_11_0_arm64.whl"
    requires_python: None
    url: /whl/cpu/torch-2.1.1-cp311-none-macosx_11_0_arm64.whl#sha256=e312f7e82e49565f7667b0bbf9559ab0c597063d93044740781c02acd5a87978
"2.1.1": PackageName("torch"), wheel version: "2.1.1" python tags: cp38, abi tags: cp38, platform tags: manylinux_2_17_aarch64, manylinux2014_aarch64
    filename: "torch-2.1.1-cp38-cp38-manylinux_2_17_aarch64.manylinux2014_aarch64.whl"
    requires_python: None
    url: /whl/cpu/torch-2.1.1-cp38-cp38-manylinux_2_17_aarch64.manylinux2014_aarch64.whl#sha256=9ca0fcbf3d5ba644d6a8572c83a9abbdf5f7ff575bc38529ef6c185a3a71bde9
"2.1.1": PackageName("torch"), wheel version: "2.1.1" python tags: cp38, abi tags: none, platform tags: macosx_10_9_x86_64
    filename: "torch-2.1.1-cp38-none-macosx_10_9_x86_64.whl"
    requires_python: None
    url: /whl/cpu/torch-2.1.1-cp38-none-macosx_10_9_x86_64.whl#sha256=d56b032176458e2af4709627bbd2c20fe2917eff8cd087a7fe313acccf5ce2f1
"2.1.1": PackageName("torch"), wheel version: "2.1.1" python tags: cp38, abi tags: none, platform tags: macosx_11_0_arm64
    filename: "torch-2.1.1-cp38-none-macosx_11_0_arm64.whl"
    requires_python: None
    url: /whl/cpu/torch-2.1.1-cp38-none-macosx_11_0_arm64.whl#sha256=29e3b90a8c281f6660804a939d1f4218604c80162e521e1e6d8c8557325902a0
"2.1.1": PackageName("torch"), wheel version: "2.1.1" python tags: cp39, abi tags: cp39, platform tags: manylinux_2_17_aarch64, manylinux2014_aarch64
    filename: "torch-2.1.1-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl"
    requires_python: None
    url: /whl/cpu/torch-2.1.1-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl#sha256=b31230bd058424e56dba7f899280dbc6ac8b9948e43902e0c84a44666b1ec151
"2.1.1": PackageName("torch"), wheel version: "2.1.1" python tags: cp39, abi tags: none, platform tags: macosx_10_9_x86_64
    filename: "torch-2.1.1-cp39-none-macosx_10_9_x86_64.whl"
    requires_python: None
    url: /whl/cpu/torch-2.1.1-cp39-none-macosx_10_9_x86_64.whl#sha256=715b50d8c1de5da5524a68287eb000f73e026e74d5f6b12bc450ef6995fcf5f9
"2.1.1": PackageName("torch"), wheel version: "2.1.1" python tags: cp39, abi tags: none, platform tags: macosx_11_0_arm64
    filename: "torch-2.1.1-cp39-none-macosx_11_0_arm64.whl"
    requires_python: None
    url: /whl/cpu/torch-2.1.1-cp39-none-macosx_11_0_arm64.whl#sha256=db67e8725c76f4c7f4f02e7551bb16e81ba1a1912867bc35d7bb96d2be8c78b4



"2.1.1+cpu": PackageName("torch"), wheel version: "2.1.1+cpu" python tags: cp310, abi tags: cp310, platform tags: linux_x86_64
    filename: "torch-2.1.1+cpu-cp310-cp310-linux_x86_64.whl"
    requires_python: None
    url: /whl/cpu/torch-2.1.1%2Bcpu-cp310-cp310-linux_x86_64.whl#sha256=16e6cdb1991ff1f15f469b12fffb5b20f00b1ecff8c1073c23c7760fba90fcbc
"2.1.1+cpu": PackageName("torch"), wheel version: "2.1.1+cpu" python tags: cp310, abi tags: cp310, platform tags: win_amd64
    filename: "torch-2.1.1+cpu-cp310-cp310-win_amd64.whl"
    requires_python: None
    url: /whl/cpu/torch-2.1.1%2Bcpu-cp310-cp310-win_amd64.whl#sha256=8e4e47aa4a004d2f5edd516b296da6fba07981ae9d1fc0da862abf651dc07d96
"2.1.1+cpu": PackageName("torch"), wheel version: "2.1.1+cpu" python tags: cp311, abi tags: cp311, platform tags: linux_x86_64
    filename: "torch-2.1.1+cpu-cp311-cp311-linux_x86_64.whl"
    requires_python: None
    url: /whl/cpu/torch-2.1.1%2Bcpu-cp311-cp311-linux_x86_64.whl#sha256=d83b13cb17544f9851cc31fed197865eae0c0f5d32df9d8d6d8535df7d2e5109
"2.1.1+cpu": PackageName("torch"), wheel version: "2.1.1+cpu" python tags: cp311, abi tags: cp311, platform tags: win_amd64
    filename: "torch-2.1.1+cpu-cp311-cp311-win_amd64.whl"
    requires_python: None
    url: /whl/cpu/torch-2.1.1%2Bcpu-cp311-cp311-win_amd64.whl#sha256=23be0cb945970443c97d4f9ea61ed03b27f924d835de689dd4134f30966c13f7
"2.1.1+cpu": PackageName("torch"), wheel version: "2.1.1+cpu" python tags: cp38, abi tags: cp38, platform tags: linux_x86_64
    filename: "torch-2.1.1+cpu-cp38-cp38-linux_x86_64.whl"
    requires_python: None
    url: /whl/cpu/torch-2.1.1%2Bcpu-cp38-cp38-linux_x86_64.whl#sha256=9399a8dfe4833bb544e7a0c332c47db1e389a06b4ae5ac9b3b167d863adc95d9
"2.1.1+cpu": PackageName("torch"), wheel version: "2.1.1+cpu" python tags: cp38, abi tags: cp38, platform tags: win_amd64
    filename: "torch-2.1.1+cpu-cp38-cp38-win_amd64.whl"
    requires_python: None
    url: /whl/cpu/torch-2.1.1%2Bcpu-cp38-cp38-win_amd64.whl#sha256=d8e19bb465e1fa15f3231ff57bbf0a673e5dcaca39eee4f58a4be7832b80c9da
"2.1.1+cpu": PackageName("torch"), wheel version: "2.1.1+cpu" python tags: cp39, abi tags: cp39, platform tags: linux_x86_64
    filename: "torch-2.1.1+cpu-cp39-cp39-linux_x86_64.whl"
    requires_python: None
    url: /whl/cpu/torch-2.1.1%2Bcpu-cp39-cp39-linux_x86_64.whl#sha256=b2cc98815251f8a2d102c2f8f4afe8304c2df61ce9c237198032c7903d97fdbb
"2.1.1+cpu": PackageName("torch"), wheel version: "2.1.1+cpu" python tags: cp39, abi tags: cp39, platform tags: win_amd64
    filename: "torch-2.1.1+cpu-cp39-cp39-win_amd64.whl"
    requires_python: None
    url: /whl/cpu/torch-2.1.1%2Bcpu-cp39-cp39-win_amd64.whl#sha256=b09d431c8e53511b6c3624f1d79cce9cd28f4c27ca20c5a49e4dcc1f9e7377e5

In the 2.1.1 case, notice that none of them match my particular platform. They're all either macOS (either x86-64 or aarch64) or Linux (only aarch64). But in the 2.1.1+cpu case, there are wheels that match my platform.

So the next thing I wanted to do was check to see whether the 2.1.1+cpu version was ever being requested, and if so, whether any of its wheels were found to be compatible:

lazy init for: "2.1.1"
/whl/cpu/torch-2.1.1-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl#sha256=84fefd63356416c0cd20578637ccdbb82164993400ed17b57c951dd6376dcee8 :: Incompatible(Tag(Abi))
/whl/cpu/torch-2.1.1-cp310-none-macosx_10_9_x86_64.whl#sha256=1e1e5faddd43a8f2c0e0e22beacd1e235a2e447794d807483c94a9e31b54a758 :: Incompatible(Tag(Abi))
/whl/cpu/torch-2.1.1-cp310-none-macosx_11_0_arm64.whl#sha256=e76bf3c5c354874f1da465c852a2fb60ee6cbce306e935337885760f080f9baa :: Incompatible(Tag(Abi))
/whl/cpu/torch-2.1.1-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl#sha256=61b51b33c61737c287058b0c3061e6a9d3c363863e4a094f804bc486888a188a :: Incompatible(Tag(Platform))
/whl/cpu/torch-2.1.1-cp311-none-macosx_10_9_x86_64.whl#sha256=a70593806f1d7e6b53657d96810518da0f88ef2608c98a402955765b8c79d52c :: Incompatible(Tag(Platform))
/whl/cpu/torch-2.1.1-cp311-none-macosx_11_0_arm64.whl#sha256=e312f7e82e49565f7667b0bbf9559ab0c597063d93044740781c02acd5a87978 :: Incompatible(Tag(Platform))
/whl/cpu/torch-2.1.1-cp38-cp38-manylinux_2_17_aarch64.manylinux2014_aarch64.whl#sha256=9ca0fcbf3d5ba644d6a8572c83a9abbdf5f7ff575bc38529ef6c185a3a71bde9 :: Incompatible(Tag(Abi))
/whl/cpu/torch-2.1.1-cp38-none-macosx_10_9_x86_64.whl#sha256=d56b032176458e2af4709627bbd2c20fe2917eff8cd087a7fe313acccf5ce2f1 :: Incompatible(Tag(Abi))
/whl/cpu/torch-2.1.1-cp38-none-macosx_11_0_arm64.whl#sha256=29e3b90a8c281f6660804a939d1f4218604c80162e521e1e6d8c8557325902a0 :: Incompatible(Tag(Abi))
/whl/cpu/torch-2.1.1-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl#sha256=b31230bd058424e56dba7f899280dbc6ac8b9948e43902e0c84a44666b1ec151 :: Incompatible(Tag(Abi))
/whl/cpu/torch-2.1.1-cp39-none-macosx_10_9_x86_64.whl#sha256=715b50d8c1de5da5524a68287eb000f73e026e74d5f6b12bc450ef6995fcf5f9 :: Incompatible(Tag(Abi))
/whl/cpu/torch-2.1.1-cp39-none-macosx_11_0_arm64.whl#sha256=db67e8725c76f4c7f4f02e7551bb16e81ba1a1912867bc35d7bb96d2be8c78b4 :: Incompatible(Tag(Abi))

So 2.1.1 is requested (by the resolver), and as expected, none of its wheels are considered compatible. But crucially, 2.1.1+cpu is never even requested in the first place. My next step is to understand why that is.

Feb 22 '24 14:02 BurntSushi

Candidate selection believes 2.1.1+cpu does not satisfy ==2.1.1:

[andrew@duff i1497]$ RUST_LOG=trace uv pip install torch==2.1.1 --index-url https://download.pytorch.org/whl/cpu
⠙ Resolving dependencies...                                                                                                                                                                                                             SELECTING CANDIDATE FOR PackageName("torch") and range Range { segments: [(Included("2.1.1"), Included("2.1.1"))] }
select candidate: "2.2.1+cpu"
    version not in range
select candidate: "2.2.1"
    version not in range
select candidate: "2.2.0+cpu"
    version not in range
select candidate: "2.2.0"
    version not in range
select candidate: "2.1.2+cpu"
    version not in range
select candidate: "2.1.2"
    version not in range
select candidate: "2.1.1+cpu"
    version not in range
select candidate: "2.1.1"
lazy init for: "2.1.1"
SELECTING CANDIDATE FOR PackageName("torch") and range Range { segments: [(Included("2.1.1"), Included("2.1.1"))] }
select candidate: "2.2.1+cpu"
    version not in range
select candidate: "2.2.1"
    version not in range
select candidate: "2.2.0+cpu"
    version not in range
select candidate: "2.2.0"
    version not in range
select candidate: "2.1.2+cpu"
    version not in range
select candidate: "2.1.2"
    version not in range
select candidate: "2.1.1+cpu"
    version not in range
select candidate: "2.1.1"

(Again, I think @charliermarsh already figured this out. But I'm confirming.)

Feb 22 '24 15:02 BurntSushi

@Nishikoh I've noticed that while uv-main pip install 'torch==2.1.1' --index-url https://download.pytorch.org/whl/cpu fails for me, uv-main pip install 'torch==2.1.1+cpu' --index-url https://download.pytorch.org/whl/cpu succeeds. Can you confirm for me that the latter still fails for you?

Does it fail for anyone else?

Feb 27 '24 14:02 BurntSushi

@BurntSushi I ran the two commands and confirmed that the latter was successful.

# uv v0.1.11
root@bad8609641ae:/# uv pip install 'torch==2.1.1' --index-url https://download.pytorch.org/whl/cpu
  × No solution found when resolving dependencies:
  ╰─▶ Because torch==2.1.1 is unusable because no wheels are available with a matching platform and you require torch==2.1.1, we can
      conclude that the requirements are unsatisfiable.


root@bad8609641ae:/# uv pip install 'torch==2.1.1+cpu' --index-url https://download.pytorch.org/whl/cpu
Resolved 9 packages in 3.59s
Downloaded 9 packages in 9.92s
Installed 9 packages in 211ms
 + filelock==3.9.0
 + fsspec==2023.4.0
 + jinja2==3.1.2
 + markupsafe==2.1.3
 + mpmath==1.3.0
 + networkx==3.2.1
 + sympy==1.12
 + torch==2.1.1+cpu
 + typing-extensions==4.8.0

My past comment may have been wrong and confused you. https://github.com/astral-sh/uv/issues/1497#issuecomment-1951600402

Feb 27 '24 15:02 Nishikoh

This should work as of the latest release:

uv pip install torch==2.1.0+cpu --find-links https://download.pytorch.org/whl/torch_stable.html

Mar 20 '24 04:03 charliermarsh

This should work as of the latest release:

@charliermarsh do you have any plans to add support for the most common case? uv pip install torch==2.1.0 --index-url https://download.pytorch.org/whl/cpu

Mar 20 '24 11:03 johnnv1

uv pip install torch==2.1.0 --index-url https://download.pytorch.org/whl/cpu

This works for me! py3.10 and latest uv 0.1.35.

This should work as of the latest release:

uv pip install torch==2.1.0+cpu --find-links https://download.pytorch.org/whl/torch_stable.html

On the other hand, this raises:

uv pip install torch==2.1.0+cpu --find-links https://download.pytorch.org/whl/torch_stable.html
  × No solution found when resolving dependencies:
  ╰─▶ Because torch==2.1.0+cpu is unusable because no wheels are available with a matching Python implementation and you require torch==2.1.0+cpu, we can conclude that the requirements are unsatisfiable.

EDIT: for this last command I also tried with the latest stable uv, 0.1.39.

May 07 '24 08:05 baggiponte

Can you confirm that the environment you're in is using Python 3.10, and that you're on x86 Linux or Windows?

May 07 '24 13:05 charliermarsh

The problem here is the behavior for installing CPU-only torch is different on arm64/aarch64 and x86_64/amd64 systems.

amd64 requires us to explicitly provide torch==x.x.x+cpu
meanwhile for arm64 we don't need the moniker.

On amd64,

$ arch
x86_64


$ uv pip install torch==2.3.0 --index-url https://download.pytorch.org/whl/cpu
  × No solution found when resolving dependencies:
  ╰─▶ Because torch==2.3.0 is unusable because no wheels are available with a matching Python ABI and you require torch==2.3.0, we can conclude that the requirements are unsatisfiable.


# with +cpu moniker
$ uv pip install torch==2.3.0+cpu --index-url https://download.pytorch.org/whl/cpu
Resolved 9 packages in 4.33s
Downloaded 9 packages in 2.90s
Installed 9 packages in 114ms
 + filelock==3.13.1
 + fsspec==2024.2.0
 + jinja2==3.1.3
 + markupsafe==2.1.5
 + mpmath==1.3.0
 + networkx==3.2.1
 + sympy==1.12
 + torch==2.3.0+cpu
 + typing-extensions==4.9.0

On arm64, adding the +cpu fails & without it installs.

~ arch
arm64


# with +cpu moniker
~ uv pip install torch==2.3.0+cpu --index-url https://download.pytorch.org/whl/cpu
  × No solution found when resolving dependencies:
  ╰─▶ Because torch==2.3.0+cpu is unusable because no wheels are available with a matching Python ABI and you require torch==2.3.0+cpu, we can conclude that the requirements are
      unsatisfiable.


# without the moniker
~ uv pip install torch==2.3.0 --index-url https://download.pytorch.org/whl/cpu
Resolved 9 packages in 3.58s
Installed 9 packages in 210ms
 + filelock==3.13.1
 + fsspec==2024.2.0
 + jinja2==3.1.3
 + markupsafe==2.1.5
 + mpmath==1.3.0
 + networkx==3.2.1
 + sympy==1.12
 + torch==2.3.0
 + typing-extensions==4.9.0

uv has known limitations around local version identifiers but unsure why pip is able to make this work the way pytorch has documented https://pytorch.org/get-started/previous-versions/#v222 but uv can't

May 07 '24 16:05 yashgorana

Can you confirm that the environment you're in is using Python 3.10, and that you're on x86 Linux or Windows?

Ops, sorry, I am on ARM macOS. You are right.

May 08 '24 11:05 baggiponte

Perhaps it's due to a lack of supported variants that they omitted the local identifier there? (here upstream PyTorch states they don't intend to publish aarch64 variants with GPU accel)

# Torch 2.3.0 + Python 3.12 Linux x86_64/aarch64 wheels

# +cpu
# https://download.pytorch.org/whl/cpu/torch/
torch-2.3.0+cpu-cp312-cp312-linux_x86_64.whl
# No local identifier for arm64:
torch-2.3.0-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl

# +cu121
# https://download.pytorch.org/whl/cu121/torch/
torch-2.3.0+cu121-cp312-cp312-linux_x86_64.whl

Reference

Using Docker from an AMD64 machine to test ARM64 environment:

# Base environment:
$ arch
x86_64
$ docker run --rm -it --platform linux/arm64 --workdir /tmp fedora:40 bash
$ arch
aarch64

# Install uv:
$ curl -LsSf https://astral.sh/uv/install.sh | sh && source $HOME/.cargo/env
# Create and enter venv:
$ uv venv && source .venv/bin/activate

# Install:
$ uv pip install \
    --extra-index-url https://download.pytorch.org/whl/cpu \
    torch torchvision torchaudio

Resolved 13 packages in 20.18s
Downloaded 13 packages in 12.26s
Installed 13 packages in 486ms
 + filelock==3.13.1
 + fsspec==2024.2.0
 + jinja2==3.1.3
 + markupsafe==2.1.5
 + mpmath==1.3.0
 + networkx==3.2.1
 + numpy==1.26.3
 + pillow==10.2.0
 + sympy==1.12
 + torch==2.3.0
 + torchaudio==2.3.0
 + torchvision==0.18.0
 + typing-extensions==4.9.0

You can kind of see why PyTorch has local identifier variants when comparing to PyPi?:

# ARM64 (PyPi instead of PyTorch)
# Slightly newer versions of the packages but otherwise equivalent
$ uv pip install torch torchvision torchaudio

Resolved 13 packages in 1.08s
Downloaded 13 packages in 23.73s
Installed 13 packages in 425ms
 + filelock==3.14.0
 + fsspec==2024.3.1
 + jinja2==3.1.4
 + markupsafe==2.1.5
 + mpmath==1.3.0
 + networkx==3.3
 + numpy==1.26.4
 + pillow==10.3.0
 + sympy==1.12
 + torch==2.3.0
 + torchaudio==2.3.0
 + torchvision==0.18.0
 + typing-extensions==4.11.0

# AMD64 (PyPi instead of PyTorch)
# Defaults to latest cuda variant (+cu121)
# Same nvidia specific deps as PyTorch +cu121, except for nvidia-nvjitlink-cu12 (Cuda 12.4 here instead of 12.1)
$ uv pip install torch torchvision torchaudio

Resolved 25 packages in 558ms
Downloaded 25 packages in 53.62s
Installed 25 packages in 189ms
 + filelock==3.14.0
 + fsspec==2024.3.1
 + jinja2==3.1.4
 + markupsafe==2.1.5
 + mpmath==1.3.0
 + networkx==3.3
 + numpy==1.26.4
 + nvidia-cublas-cu12==12.1.3.1
 + nvidia-cuda-cupti-cu12==12.1.105
 + nvidia-cuda-nvrtc-cu12==12.1.105
 + nvidia-cuda-runtime-cu12==12.1.105
 + nvidia-cudnn-cu12==8.9.2.26
 + nvidia-cufft-cu12==11.0.2.54
 + nvidia-curand-cu12==10.3.2.106
 + nvidia-cusolver-cu12==11.4.5.107
 + nvidia-cusparse-cu12==12.1.0.106
 + nvidia-nccl-cu12==2.20.5
 + nvidia-nvjitlink-cu12==12.4.127
 + nvidia-nvtx-cu12==12.1.105
 + pillow==10.3.0
 + sympy==1.12
 + torch==2.3.0
 + torchaudio==2.3.0
 + torchvision==0.18.0
 + typing-extensions==4.11.0

Resolve platform packaging inconsistency upstream

Maybe open an issue about it upstream since it seems to be more of a packaging concern unrelated to uv?:

https://github.com/pytorch/pytorch/issues/110791#issuecomment-1753334468

please do not hesitate to open a new issue (or a PR) if you want to propose dropping +cpu suffix for Linux and Windows wheels, as nobody seems to be relying on the old installation method anyway.

Which makes sense that CPU should be the default, with local identifiers only used for actual variants (eg: different cuda version support). Might introduce some friction to any existing CI perhaps, but you can see similar with nvidia's own Docker images they publish where the naming convention isn't always consistent 😅 (at least it broke a PyTorch docker build CI task recently)

May 09 '24 08:05 polarathene

uv uv copied to clipboard

Cannot install the CPU version of torch

Reference

Resolve platform packaging inconsistency upstream

uv
uv copied to clipboard