Support multiple root modules in namespace packages by enumerating them:
```toml
[tool.uv.build-backend]
module-name = ["foo", "bar"]
```
This allows applications with multiple root packages without migrating
to workspaces. Since those are regular module names (we iterate over
them an process each one like a single module names), it allows
combining dotted (namespace) names and regular names. It also
technically allows combining regular and stub modules, even though this
is even less recommends.
We don't recommend this structure (please use a workspace instead, or
structure everything in one root module), but it reduces the number of
cases that need `namespace = true`.
Fixes#14435Fixes#14438
---------
Co-authored-by: Zanie Blue <contact@zanie.dev>
## Summary
This PR intends to enable `--torch-backend=auto` to detect Intel GPUs
automatically:
- On Linux, detection is performed using the `lspci` command via
`Display controller` id.
- On Windows, ~~detection is done via a `powershell` query to
`Win32_VideoController`~~. Skip support for now—revisit once a better
solution is available.
Currently, Intel GPUs (XPU) do not rely on specific driver or toolkit
versions to distribute different PyTorch wheels.
## Test Plan
<!-- How was it tested? -->
On Linux:
~~On Windows:
~~
---------
Co-authored-by: Charlie Marsh <charlie.r.marsh@gmail.com>
Reverts:
- #14349
- #14346
- #14245
Retains the test cases. Includes a `find-links` test case.
Supersedes
- https://github.com/astral-sh/uv/pull/14387
- https://github.com/astral-sh/uv/pull/14503
We originally got a report at
https://github.com/astral-sh/uv/issues/13707 that inclusion of a
trailing slash on an index URL was causing lockfile churn despite having
no semantic meaning and resolved the issue by adding normalization that
stripped trailing slashes at parse time.
We then discovered that, while there are not semantic differences for
trailing slashes on Simple API index URLs, there are differences for
some flat (or find links) indexes. As reported in
https://github.com/astral-sh/uv/issues/14367, the change in
https://github.com/astral-sh/uv/pull/14245 caused a regression for at
least one user.
We attempted to fix the regression via a few approaches.
https://github.com/astral-sh/uv/pull/14387 attempted to differentiate
between Simple API and flat index URL parsing, but failed to account for
the `Deserialize` implementation, which always assumed Simple API-style
index URLs and incorrectly trimmed trailing slashes in various cases
where we deserialized the `IndexUrl` type from a file. I attempted to
resolve this by performing a larger refactor, but it ended up being
quite painful. In particular, the `Index` type was a blocker — we don't
know the `IndexUrl` variant until we've parsed the `IndexFormat` and
having a multi-stage deserializer is not appealing but adding a new
intermediate type (i.e., `RawIndex`) is painful due to the pervasiveness
of `Index`. Given that we've regressed behavior here and there's not a
straight-forward fix, we're reverting the normalization entirely.
https://github.com/astral-sh/uv/pull/14503 attempted to perform
normalization at compare-time, but that means we'd fail to invalidate
the lockfile when the a trailing slash was added or removed and given
that a trailing slash has semantic meaning for a find-links URL... we'd
have another correctness problem.
After this revert, we'll retain all index URLs verbatim. The downside to
this approach is that we'll be adding a bunch of trailing slashes back
to lockfiles that we previously normalized out, and we'll be reverting
our fix for users with inconsistent trailing slashes on their index
URLs. Users affected by the original motivating issue should use
consistent trailing slashes on their URLs, as they do frequently have
semantic meaning. We may want to revisit normalization and type aware
index URL parsing as part of a larger change.
Closes https://github.com/astral-sh/uv/issues/14367
<!--
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
We are using UV as a library and `installer()` returned `"pip\n"`. The
packages got installed by the pip package manager and not by UV. pip
seems to add a new line to the `INSTALLER` file and UV does not.
<!-- What's the purpose of the change? What does it do, and why? -->
## Test Plan
<!-- How was it tested? -->
When [updating](https://github.com/astral-sh/uv/pull/14475) to the
latest `reqwest` version, our fragment propagation test broke. That test
was partially testing the `reqwest` behavior, so this PR moves the
fragment test to directly test our logic for constructing redirect
requests.
<!--
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
<!-- What's the purpose of the change? What does it do, and why? -->
In pixi we overlay the PyPI packages over the conda packages and we
sometimes need to figure out what PyPI packages are involved in the
no-solution error. We could parse the error message, but this is pretty
error-prone, so it would be good to get access to more information. A
lot of information in this module is private and should probably stay
this way, but package names are easy enough to expose. This would help
us a lot!
I collect into a HashSet to remove duplication, and did not want to
expose a rustc_hash datastructure directly, thats's why I've chosen to
expose as an iterator :)
Let me know if any changes need to be done, and thanks!
---------
Co-authored-by: Zanie Blue <contact@zanie.dev>
Hey, are you okay with exposing the `ErrorTree` for library consumers?
We have a use case that needs more information on conflicts. We need the
tree-structure of the conflict and be able to traverse it in particular.
Signed-off-by: Simon Sure <ssure@palantir.com>
The uv build backend has gone through some feedback cycles, we expect no
more major configuration changes, and we're ready to take the next step:
The uv build backend in stable.
This PR stabilizes:
* Using `uv_build` as build backend
* The documentation of the uv build backend
* The direct build fast path, where uv doesn't use PEP 517 if you're
using `uv_build` in a compatible version.
* `uv build --list`, which is limited to `uv_build`.
It does not:
* Make `uv_build` the default on `uv init`
* Make `--package` the default on `uv init`
## Summary
If we fail to acquire a lock on an environment, uv shouldn't fail; we
should just warn. In some cases, users run uv with read-only permissions
for their projects, etc.
For now, I kept any locks acquired _in the cache_ as hard failures,
since we always need write-access to the cache.
Closes https://github.com/astral-sh/uv/issues/14411.
## Summary
The idea here is that if a user runs `uv pip compile --universal`, we
should ignore the patch version on the current interpreter. I think this
makes sense... `--universal` tries to resolve for all future versions,
so it seems a bit odd that we'd start at the _current_ patch version.
Closes https://github.com/astral-sh/uv/issues/14397.
Clap does not perform global validation, so flag that are declared as
overriding can be set at the same time:
https://github.com/clap-rs/clap/issues/6049. This would previously cause
a panic. We work around this by choosing the yes-value always and
writing a warning.
An alternative would be erroring when both are set, but it's unclear to
me if this may break things we want to support. (`UV_OFFLINE=1 cargo run
-q pip --no-offline install tqdm --no-cache` is already banned).
Fixes https://github.com/astral-sh/uv/pull/14299
**Test Plan**
```
$ cargo run -q pip --offline install --no-offline tqdm --no-cache
warning: Boolean flags on different levels are not correctly supported (https://github.com/clap-rs/clap/issues/6049)
× No solution found when resolving dependencies:
╰─▶ Because tqdm was not found in the cache and you require tqdm, we can conclude that your requirements are unsatisfiable.
hint: Packages were unavailable because the network was disabled. When the network is disabled, registry packages may only be read from the cache.
```
## Summary
The basic idea here is that we can (should) reuse a build environment
across resolution (`prepare_metadata_for_build_wheel`) and installation.
This also happens to solve the build-PyTorch-from-source problem, since
we use a consistent build environment between the invocations.
Since `SourceDistributionBuilder` is stateless, we instead store the
builds on `BuildContext`, and we key them by various properties: the
underlying interpreter, the configuration settings, etc. This just
ensures that if we build the same package twice within a process, we
don't accidentally reuse an incompatible build (virtual) environment.
(Note that still drop build environments at the end of the command, and
don't attempt to reuse them across processes.)
Closes#14269.
If/when we see https://github.com/astral-sh/uv/issues/14171 again, this
should clarify whether our retry logic was skipped (i.e. a transient
error wasn't correctly identified as transient), or whether we exhausted
our retries. Previously, if you ran a local example fileserver as in
https://github.com/astral-sh/uv/issues/14171#issuecomment-3014580701 and
then you tried to install Python from it, you'd get:
```
$ export UV_TEST_NO_CLI_PROGRESS=1
$ uv python install 3.8.20 --mirror http://localhost:8000 2>&1 | cat
error: Failed to install cpython-3.8.20-linux-x86_64-gnu
Caused by: Failed to extract archive: cpython-3.8.20-20241002-x86_64-unknown-linux-gnu-install_only_stripped.tar.gz
Caused by: failed to unpack `/home/jacko/.local/share/uv/python/.temp/.tmpS4sHHZ/python/lib/libpython3.8.so.1.0`
Caused by: failed to unpack `python/lib/libpython3.8.so.1.0` into `/home/jacko/.local/share/uv/python/.temp/.tmpS4sHHZ/python/lib/libpython3.8.so.1.0`
Caused by: error decoding response body
Caused by: request or response body error
Caused by: error reading a body from connection
Caused by: Connection reset by peer (os error 104)
```
With this change you get:
```
error: Failed to install cpython-3.8.20-linux-x86_64-gnu
Caused by: Request failed after 3 retries
Caused by: Failed to extract archive: cpython-3.8.20-20241002-x86_64-unknown-linux-gnu-install_only_stripped.tar.gz
Caused by: failed to unpack `/home/jacko/.local/share/uv/python/.temp/.tmp4Ia24w/python/lib/libpython3.8.so.1.0`
Caused by: failed to unpack `python/lib/libpython3.8.so.1.0` into `/home/jacko/.local/share/uv/python/.temp/.tmp4Ia24w/python/lib/libpython3.8.so.1.0`
Caused by: error decoding response body
Caused by: request or response body error
Caused by: error reading a body from connection
Caused by: Connection reset by peer (os error 104)
```
At the same time, I'm updating the way we handle the retry count to
avoid nested retry loops exceeding the intended number of attempts, as I
mentioned at
https://github.com/astral-sh/uv/issues/14069#issuecomment-3020634281.
It's not clear to me whether we actually want this part of the change,
and I need feedback here.
## Summary
This PR ensures that we avoid cleaning up build directories until the
end of a resolve-and-install cycle. It's not bulletproof (since we could
still run into issues with `uv lock` followed by `uv sync` whereby a
build directory gets cleaned up that's still referenced in the `build`
artifacts), but it at least gets PyTorch building without error with `uv
pip install .`, which is a case that's been reported several times.
Closes https://github.com/astral-sh/uv/issues/14269.
The marker display code assumes that all versions are normalized, in
that all trailing zeroes are stripped. This is not the case for
tilde-equals and equals-star versions, where the trailing zeroes (before
the `.*`) are semantically relevant. This would cause path
dependent-behavior where we would get a different marker string
depending on whether a version with or without a trailing zero was added
to the cache first.
To handle both equals-star and tilde-equals when converting
`python_version` to `python_full_version` markers, we have to merge the
version normalization (i.e. trimming the trailing zeroes) and the
conversion both to `python_full_version` and to `Ranges`, while special
casing equals-star and tilde-equals.
To avoid churn in lockfiles, we only trim in the conversion to `Ranges`
for markers, but keep using untrimmed versions for requires-python.
(Note that this behavior is technically also path dependent, as versions
with and without trailing zeroes have the same Hash and Eq. E.q.,
`requires-python == ">= 3.10.0"` and `requires-python == ">= 3.10"` in
the same workspace could lead to either value in `uv.lock`, and which
one it is could change if we make unrelated (performance) changes.
Always trimming however definitely changes lockfiles, a churn I wouldn't
do outside another breaking or lockfile-changing change.) Nevertheless,
there is a change for users who have `requires-python = "~= 3.12.0"` in
their `pyproject.toml`, as this now hits the correct normalization path.
Fixes#14231Fixes#14270
## Summary
There's a good example of the downside of using verbatim URLs here:
https://github.com/astral-sh/uv/pull/14197#discussion_r2163599625 (we
show two relative paths that point to the same directory, but it's not
clear from the error message).
The diff:
```
2 2 │ ----- stdout -----
3 3 │
4 4 │ ----- stderr -----
5 5 │ error: Requirements contain conflicting URLs for package `library` in all marker environments:
6 │-- ../../library
7 │-- ./library
6 │+- file://[TEMP_DIR]/library
7 │+- file://[TEMP_DIR]/library (editable)
```
## Summary
<!-- What's the purpose of the change? What does it do, and why? -->
As explained in the [`codspeed-rust` v3 release
notes](https://github.com/CodSpeedHQ/codspeed-rust/releases/tag/v3.0.0),
the `v3` of the compatibility layers is now required to work with the
latest version(`v3`) of `cargo-codspeed`.
and prefer emulated x64 windows in its stead.
This is preparatory work for shipping support for uv downloading and
installing aarch64 (arm64) windows Pythons. We've [had builds for this
platform ready for a
while](https://github.com/astral-sh/python-build-standalone/pull/387),
but have held back on shipping them due to a fundamental problem:
**The Python packaging ecosystem does not have strong support for
aarch64 windows**, e.g., not many projects build aarch64 wheels yet. The
net effect of this is that, if we handed you an aarch64 python
interpreter on windows, you would have to build a lot more sdists, and
there's a high chance you will simply fail to build that sdist and be
sad.
Yes unfortunately, in this case a non-native Python interpreter simply
*works better* than the native one... in terms of working at all, today.
Of course, if the native interpreter works for your project, it should
presumably have better performance and platform compatibility.
We do not want to stand in the way of progress, as ideally this
situation is a temporary state of affairs as the ecosystem grows to
support aarch64 windows. To enable progress, on aarch64 Windows builds
of uv:
* We will still use a native python interpreter, e.g., if it's at the
front of your `PATH` or the only installed version.
* If we are choosing between equally good interpreters that differ in
architecture, x64 will be preferred.
* If the aarch64 version is newer, we will prefer the aarch64 one.
* We will emit a diagnostic on installation, and show the python request
to pass to uv to force aarch64 windows to be used.
* Will be shipping [aarch64 Windows Python
downloads](https://github.com/astral-sh/python-build-standalone/pull/387)
* Will probably add some kind of global override setting/env-var to
disable this behaviour.
* Will be shipping this behaviour in
[astral-sh/setup-uv](https://github.com/astral-sh/setup-uv)
We're coordinating with Microsoft, GitHub (for the `setup-python`
action), and the CPython team (for the `python.org` installers), to
ensure we're aligned on this default and the timing of toggling to
prefer native distributions in the future.
See discussion in
- https://github.com/astral-sh/uv/issues/12906
---
This is an alternative to
* #13719
which uses sorting rather than filtering, as discussed in
* #13721
This fixes an obscure cache collision in Python interpreter queries,
which we believe to be the root cause of CI flakes we've been seeing
where a project environment is invalidated and recreated.
This work follows from the logs in [this CI
run](https://github.com/astral-sh/uv/actions/runs/15934322410/job/44950599993?pr=14326)
which captured one of the flakes with tracing enabled. There, we can see
that the project environment is invalidated because the Python
interpreter in the environment has a different version than expected:
```
DEBUG Checking for Python environment at `.venv`
TRACE Cached interpreter info for Python 3.12.9, skipping probing: .venv/bin/python3
DEBUG The interpreter in the project environment has different version (3.12.9) than it was created with (3.9.21)
```
(this message is updated to reflect #14329)
The flow is roughly:
- We create an environment with 3.12.9
- We query the environment, and cache the interpreter version for
`.venv/bin/python`
- We create an environment for 3.9.12, replacing the existing one
- We query the environment, and read the cached information
The Python cache entries are keyed by the absolute path to the
interpreter, and rely on the modification time (ctime, nsec resolution)
of the canonicalized path to determine if the cache entry should be
invalidated. The key is a hex representation of a u64 sea hasher output
— which is very unlikely to collide.
After an audit of the Python query caching logic, we determined that the
most likely cause of a collision in cache entries is that the
modification times of underlying interpreters are identical. This seems
pretty feasible, especially if the file system does not support
nanosecond precision — though it appears that the GitHub runners do
support it.
The fix here is to include the canonicalized path in the cache key,
which ensures we're looking at the modification time of the _same_
underlying interpreter.
This will "invalidate" all existing interpreter cache entries but that's
not a big deal.
This should also have the effect of reducing cache churn for
interpreters in virtual environments. Now, when you change Python
versions, we won't invalidate the previous cache entry so if you change
_back_ to the old version we can re-use our cached information.
It's a bit speculative, since we don't have a deterministic reproduction
in CI, but this is the strongest candidate given the logs and should
increase correctness regardless.
Closes https://github.com/astral-sh/uv/issues/14160
Closes https://github.com/astral-sh/uv/issues/13744
Closes https://github.com/astral-sh/uv/issues/13745
Once it's confirmed the flakes are resolved, we should revert
- https://github.com/astral-sh/uv/pull/14275
- #13817
