When a package is included under a platform-specific marker, we know
that wheels that mismatch this marker can never be installed, so we drop
them from the lockfile.
In transformers, we have:
* `tensorflow-text`: `tensorflow-macos; python_full_version >= '3.13'
and platform_machine == 'arm64' and platform_system == 'Darwin'`
* `tensorflow-macos`: `tensorflow-cpu-aws; (python_full_version < '3.10'
and platform_machine == 'aarch64' and platform_system == 'Linux') or
(python_full_version >= '3.13' and platform_machine == 'aarch64' and
platform_system == 'Linux') or (python_full_version >= '3.13' and
platform_machine == 'arm64' and platform_system == 'Linux')`
* `tensorflow-macos`: `tensorflow-intel; python_full_version >= '3.13'
and platform_system == 'Windows'`
This means that `tensorflow-cpu-aws` and `tensorflow-intel` can never be
installed, and we can drop them from the lockfile.
## Summary
I'm not convinced that the behavior is correct as-implemented. When the
user passes a `--python >=3.8` or we discover a `requires-python` from
the workspace, we're currently writing that request out to
`.python-version`. I would probably rather that we write the resolved
patch version?
Closes https://github.com/astral-sh/uv/issues/6821.
The key change here is to use raw strings so that we don't need to
double-escape things like `\d`. And in particular, we rely on the fact
that `"\n"` and `r"\n"` are precisely equivalent when fed to
`Regex::new` in the `regex` crate.
Previously we were using `[+-~]`, but this includes the full range of
characters from `+` to `~`. Incidentally, this does include `-`. We
instead rewrite this as `[-+~]`, which probably matches the intent.
Unlike the previous update, this message is specifically referring to a
fork's markers inside the resolver. We probably *could* massage the
message to be simplified with respect to requires-python, but it's not
obvious to me that that is the right thing to do.
It is not clear whether this update is correct or not. Moreover, it's
not clear whether the status quo is correct or not. The problem is that
`transformers` is so big that it's very hard to understand what the
right output is without a deeper investigation.
One thing that is interesting is if fork prioritization is removed in
this PR *and* on `main`, then the differences in this ecosystem test go
away.
We've decided for now to move forward with this update even though we're
uncertain because this PR fixes a few outstanding correctness issues.
This update changes the error message to one that is worse than the
status quo, but it is still correct because `datasets >= 2.19` doesn't
actually exist given our `EXCLUDE_NEWER` in tests at present.
The underlying cause here seems to be in how PubGrub deals with
reporting incompatibilities. Namely, when it has `foo < 1` and
`foo >= 1`, it reports an incompatibility immediately before looking for
versions. But when it has `foo < 1` and `foo >= 1 ; marker`, then
because they aren't both pubgrub "packages," it starts requesting
versions first and hits the "not available" error path instead of the
"incompatible" error path.
Since this is more of an underlying issue with how we setup
`PubGrubPackage` and our interaction with pubgrub, we ended up deciding
to move forward here with the regression since this PR is fixing a
correctness issue. In particular, if one changes the `requires-python`
to `>=3.8`, then both `main` and this PR produce similarly bad error
messages.
This commit refactors how deal with `requires-python` so that instead of
simplifying markers of dependencies inside the resolver, we do it at the
edges of our system. When writing markers to output, we simplify when
there's an obvious `requires-python` context. And when reading markers
as input, we complexity markers with the relevant `requires-python`
constraint.
When I first wrote this routine, it was intended to only emit a trace
for the final "unioned" resolution. But we actually moved that semantic
operation to the construction of the resolution *graph*. So there is no
unioned `Resolution` any more.
But this is still useful to see. So I changed this to just emit a trace
of *every* resolution right before constructing the graph.
It might be nice to also emit a trace of the unioned graph too. Or
perhaps we should do that instead if this proves too noisy. (Although
this is only emitted at TRACE level.)
These are regression tests for #6269, #6412 and #6836. In this commit,
their test outputs are all wrong. We'll update these snapshots after
fixing the underlying bug by refactoring how `requires-python`
simplification works.
- Respect `UV_PROJECT_ENVIRONMENT` when in project root
- Add `--no-project` and `--no-workspace` to opt-out of above and
`requires-python` detection
- Rename `[NAME]` to `[PATH]` in CLI
Allows configuration of the (currently hard-coded) path to the virtual
environment in projects using the `UV_PROJECT_ENVIRONMENT` environment
variable.
If empty, we'll ignore it. If a relative path, it will be resolved
relative to the workspace root. If an absolute path, we'll use that.
This feature targets use in Docker images and CI. The variable is
intended to be set once in an isolated system and used for all uv
operations.
We do not expose a CLI option or configuration file setting — we may
pursue those later but I see them as lower priority. I think a
system-level environment variable addresses the most pressing use-cases
here.
This doesn't special-case the system environment. Which means that you
can use this to write to the system Python environment. I would
generally strongly recommend against doing so. The insightful comment
from @edmorley at
https://github.com/astral-sh/uv/issues/5229#issuecomment-2312702902
provides some context on why. More generally, `uv sync` will remove
packages from the environment by default. This means that if the system
environment contains any packages relevant to the operation of the
system (that are not dependencies of your project), `uv sync` will break
it. I'd only use this in Docker or CI, if anywhere. Virtual environments
have lots of benefits, and it's only [one line to "activate"
them](https://docs.astral.sh/uv/guides/integration/docker/#using-the-environment).
If you are considering using this feature to use Docker bind mounts for
developing in containers, I would highly recommend reading our [Docker
container development
documentation](https://docs.astral.sh/uv/guides/integration/docker/#developing-in-a-container)
first. If the solutions there do not work for you, please open an issue
describing your use-case and why.
We do not read `VIRTUAL_ENV` and do not have plans to at this time.
Reading `VIRTUAL_ENV` is high-risk, because users can easily leave an
environment active and use the uv project interface today. Reading
`VIRTUAL_ENV` would be a breaking change. Additionally, uv is
intentionally moving away from the concept of "active environments" and
I don't think syncing to an "active" environment is the right behavior
while managing projects. I plan to add a warning if `VIRTUAL_ENV` is
set, to avoid confusion in this area (see
https://github.com/astral-sh/uv/pull/6864).
This does not directly enable centrally managed virtual environments. If
you set `UV_PROJECT_ENVIRONMENT` to an absolute path and use it across
multiple projects, they will clobber each other's environments. However,
you could use this with something like `direnv` to achieve "centrally
managed" environments. I intend to build a prototype of this eventually.
See #1495 for more details on this use-case.
Lots of discussion about this feature in:
- https://github.com/astral-sh/rye/issues/371
- https://github.com/astral-sh/rye/pull/1222
- https://github.com/astral-sh/rye/issues/1211
- https://github.com/astral-sh/uv/issues/5229
- https://github.com/astral-sh/uv/issues/6669
- https://github.com/astral-sh/uv/issues/6612
Follow-ups:
- #6835
- https://github.com/astral-sh/uv/pull/6864
- Document this in the project concept documentation (can probably
re-use some of this post)
Closes https://github.com/astral-sh/uv/issues/6669
Closes https://github.com/astral-sh/uv/issues/5229
Closes https://github.com/astral-sh/uv/issues/6612
## Summary
Http headers are supposed to be case-insensitive (RFC 2616), but there
are some implementations that don't normalize them.
I noticed it while migrating to `uv`, calls to an internal registry
failed. A man in the middle server helped me to find that `pip` uses
Title-Case while `uv pip` uses lowercase.
## Test Plan
I tested `uv` with the same server and now it works fine.
<!--
Thank you for contributing to uv! To help us out with reviewing, please
consider the following:
- Does this pull request include a summary of the change? (See below.)
- Does this pull request include a descriptive title?
- Does this pull request include references to any relevant issues?
-->
## Summary
Separate exceptions for different timeouts to make it easier to debug
issues like #6105.
<!-- What's the purpose of the change? What does it do, and why? -->
## Test Plan
<!-- How was it tested? -->
Not tested at all.
## Summary
Closes#6319.
## Test Plan
I tested with `file:///mirror`, `file://localhost/mirror`, and
`http://mirror` to confirm that it was working as expected.
``` shell-session
/private/tmp/mirror-local 07:08:18
:) tree mirror
mirror/
└── 20240814/
└── cpython-3.12.5+20240814-aarch64-apple-darwin-install_only_stripped.tar.gz
```
<img width="626" alt="image"
src="https://github.com/user-attachments/assets/9c04224d-305c-47ee-a524-4a6abeb79da4">
## Summary
Right now, we have slightly different `requires-python` semantics for
`-p 3.11` vs. `-p 3.11 --universal`, and slightly different (wrong)
semantics for how we compare against the _installed_ Python version
(which doesn't ignore upper bounds, but should).
This PR rips it all out and replaces it with consistent semantics across
`uv lock`, `uv pip compile -p 3.11`, and `uv pip compile -p 3.11
--universal`. We now always ignore upper bounds.
Closes https://github.com/astral-sh/uv/issues/6859.
Closes https://github.com/astral-sh/uv/issues/5045.
Forward an error for missing temp directories:
```
$ env TMPDIR=.tmp uv-debug pip install httpx
error: No such file or directory (os error 2) at path "/home/konsti/projects/uv/.tmp/.tmpgIBhhh"
```
Fixes#6878
## Summary
A few of these should use `absolute` instead of `canonicalize`; and
apparently we no longer need to strip the `CANONICAL_CWD` to get tests
passing.
## Summary
This PR addresses an issue on Windows where `std::fs::canonicalize` can
fail or panic when resolving paths on mapped network drives. By
replacing it with `Simplified::simple_canonicalize`, we aim to improve
the robustness and cross-platform compatibility of path resolution.
### Changes
* Updated `CANONICAL_CWD` in `path.rs` to use
`Simplified::simple_canonicalize` instead of `std::fs::canonicalize`.
### Why
* `std::fs::canonicalize` has known issues with resolving paths on
mapped network drives on Windows, which can lead to panics or incorrect
path resolution.
* `Simplified::simple_canonicalize` internally uses
`dunce::canonicalize`, which handles these cases more gracefully,
ensuring better stability and reliability.
## Test Plan
Since `simple_canonicalize` has already been tested in a prior PR, this
change is expected to work without introducing any new issues. No
additional tests are necessary beyond ensuring existing tests pass,
which will confirm the correctness of the change.
## Summary
- The change relates to #6635 is to include compiled python files (.pyc)
in the uv run command.
- After this change `uv run foo.pyc` should spawn `python foo.pyc`.
## Test Plan
- There is a test that uses TestContext to compile and run a simple
python file that prints "Hello World".
- I built the project locally and tried the same with a simple python
file that I had compiled.
