When we generate conflict markers for each resolution after the
resolver runs, it turns out that generating them just from exclusion
rules is not sufficient.
For example, if `foo` and `bar` are declared as conflicting extras, then
we end up with the following forks:
A: extra != 'foo'
B: extra != 'bar'
C: extra != 'foo' and extra != 'bar'
Now let's take an example where these forks don't share the same version
for all packages. Consider a case where `idna==3.9` is in forks A and C,
but where `idna==3.10` is in fork B. If we combine the markers in forks
A and C through disjunction, we get the following:
idna==3.9: extra != 'foo' or (extra != 'foo' and extra != 'bar')
idna==3.10: extra != 'bar'
Which simplifies to:
idna==3.9: extra != 'foo'
idna==3.10: extra != 'bar'
But these are clearly not disjoint. Both dependencies could be selected,
for example, when neither `foo` nor `bar` are active. We can remedy this
by keeping around the inclusion rules for each fork:
A: extra != 'foo' and extra == 'bar'
B: extra != 'bar' and extra == 'foo'
C: extra != 'foo' and extra != 'bar'
And so for `idna`, we have:
idna==3.9: (extra != 'foo' and extra == 'bar') or (extra != 'foo' and extra != 'bar')
idna==3.10: extra != 'bar' and extra == 'foo'
Which simplifies to:
idna==3.9: extra != 'foo'
idna==3.10: extra != 'bar' and extra == 'foo'
And these *are* properly disjoint. There is no way for them both to be
active. This also correctly accounts for fork C where neither `foo` nor
`bar` are active, and yet, `idna==3.9` is still enabled but `idna==3.10`
is not. (In the [motivating example], this comes from `baz` being enabled.)
That is, this captures the idea that for `idna==3.10` to be installed,
there must actually be a specific extra that is enabled. That's what
makes it disjoint from `idna==3.9`.
We aren't quite done yet, because this does add *too many* conflict
markers to dependency edges that don't need it. In the next commit,
we'll add in our world knowledge to simplify these conflict markers.
[motivating example]: https://github.com/astral-sh/uv/issues/9289
Previously, we had copied the behavior of `try_markers` to return
`None` in the case where the marker was always true. I believe this
was done because it somewhat implies that there is no forking
happening. But I find this somewhat strange personally, and instead
flipped this around so that it still returns a marker in that case.
The one call site that is impacted by this is the resolution
graph construction. If we left it as-is, it would end up with
a list of one marker that is always true in some cases. And this
in turn results in writing an empty `resolution-markers` to the
lock file. Probably the output logic should be tweaked instead,
but we leave it alone for now.
This effectively combines a PEP 508 marker and an as-yet-specified
marker for expressing conflicts among extras and groups.
This just defines the type and threads it through most of the various
points in the code that previously used `MarkerTree` only. Some parts
do still continue to use `MarkerTree` specifically, e.g., when dealing
with non-universal resolution or exporting to `requirements.txt`.
This doesn't change any behavior.
This doesn't change any behavior. My guess is that this code was
a casualty of refactoring. But basically, it was doing redundant
case analysis and iterating over all resolutions (even though it's
in the branch that can only occur when there is only one
resolution).
This filtering is now redundant, since forking now avoids these
degenerate cases by construction.
The main change to forking that enables skipping over "always
false" forks is that forking now starts with the parent's markers
instead of starting with MarkerTree::TRUE and trying to combine
them with the parent's markers later. This in turn leads to
skipping over anything that "can't" happen when combined with the
parents markers. So we never hit the case of generating a fork
that, when combined with the parent's markers, results in a
marker that is always false. We just avoid it in the first place.
## Summary
The issue here is fairly complex. Consider the following:
```toml
[project]
name = "project"
version = "0.1.0"
requires-python = ">=3.12.0"
dependencies = []
[project.optional-dependencies]
cpu = [
"torch>=2.5.1",
"torchvision>=0.20.1",
]
cu124 = [
"torch>=2.5.1",
"torchvision>=0.20.1",
]
[tool.uv]
conflicts = [
[
{ extra = "cpu" },
{ extra = "cu124" },
],
]
[tool.uv.sources]
torch = [
{ index = "pytorch-cpu", extra = "cpu", marker = "platform_system != 'Darwin'" },
]
torchvision = [
{ index = "pytorch-cpu", extra = "cpu", marker = "platform_system != 'Darwin'" },
]
[[tool.uv.index]]
name = "pytorch-cpu"
url = "https://download.pytorch.org/whl/cpu"
explicit = true
```
When solving this project, we first pick a PyTorch version from PyPI, to
solve the `cu124` extra, selecting `2.5.1`.
Later, we try to solve the `cpu` extra. In solving that extra, we look
at the PyTorch CPU index. Ideally, we'd select `2.5.1+cpu`... But
`2.5.1` is already a preference. So we choose that.
Now, we only respect preferences for explicit indexes if they came from
the same index.
Closes https://github.com/astral-sh/uv/issues/9295.
## Summary
The reqwest middleware doesn't retry errors that occur "after" the
request completes -- but in some cases, these do include spurious errors
that we want to retry. See https://github.com/astral-sh/uv/issues/8144
for examples. This PR adds a second retry layer during the response
_handler_, which should help with some of the spurious failures we see
in the linked issue.
Closes https://github.com/astral-sh/uv/issues/8144.
## Summary
This PR enables something like the "final boss" of PyTorch setups --
explicit support for CPU vs. GPU-enabled variants via extras:
```toml
[project]
name = "project"
version = "0.1.0"
requires-python = ">=3.13.0"
dependencies = []
[project.optional-dependencies]
cpu = [
"torch==2.5.1+cpu",
]
gpu = [
"torch==2.5.1",
]
[tool.uv.sources]
torch = [
{ index = "torch-cpu", extra = "cpu" },
{ index = "torch-gpu", extra = "gpu" },
]
[[tool.uv.index]]
name = "torch-cpu"
url = "https://download.pytorch.org/whl/cpu"
explicit = true
[[tool.uv.index]]
name = "torch-gpu"
url = "https://download.pytorch.org/whl/cu124"
explicit = true
[tool.uv]
conflicts = [
[
{ extra = "cpu" },
{ extra = "gpu" },
],
]
```
It builds atop the conflicting extras work to allow sources to be marked
as specific to a dedicated extra being enabled or disabled.
As part of this work, sources now have an `extra` field. If a source has
an `extra`, it means that the source is only applied to the requirement
when defined within that optional group. For example, `{ index =
"torch-cpu", extra = "cpu" }` above only applies to
`"torch==2.5.1+cpu"`.
The `extra` field does _not_ mean that the source is "enabled" when the
extra is activated. For example, this wouldn't work:
```toml
[project]
name = "project"
version = "0.1.0"
requires-python = ">=3.13.0"
dependencies = ["torch"]
[tool.uv.sources]
torch = [
{ index = "torch-cpu", extra = "cpu" },
{ index = "torch-gpu", extra = "gpu" },
]
[[tool.uv.index]]
name = "torch-cpu"
url = "https://download.pytorch.org/whl/cpu"
explicit = true
[[tool.uv.index]]
name = "torch-gpu"
url = "https://download.pytorch.org/whl/cu124"
explicit = true
```
In this case, the sources would effectively be ignored. Extras are
really confusing... but I think this is correct? We don't want enabling
or disabling extras to affect resolution information that's _outside_ of
the relevant optional group.
## Summary
These were moved as part of a broader refactor to create a single
integration test module. That "single integration test module" did
indeed have a big impact on compile times, which is great! But we aren't
seeing any benefit from moving these tests into their own files (despite
the claim in [this blog
post](https://matklad.github.io/2021/02/27/delete-cargo-integration-tests.html),
I see the same compilation pattern regardless of where the tests are
located). Plus, we don't have many of these, and same-file tests is such
a strong Rust convention.
## Summary
I was wrongly using `.name()` to detect if a package was "not root", but
in `pip compile`, the root can have a name -- so we were failing to find
the derivation chain.
## Summary
This PR adds context to our error messages to explain _why_ a given
package was included, if we fail to download or build it.
It's quite a large change, but it motivated some good refactors and
improvements along the way.
Closes https://github.com/astral-sh/uv/issues/8962.
## Summary
This PR should not contain any user-visible changes, but the goal is to
refactor the `Resolution` type to retain a dependency graph. We want to
be able to explain _why_ a given package was excluded on error (see:
https://github.com/astral-sh/uv/issues/8962), which in turn requires
that at install time, we can go back and figure out the dependency
chain. At present, `Resolution` is just a map from package name to
distribution; this PR remodels it as a graph in which each node is a
package, and the edges contain markers plus extras or dependency groups.
## Summary
As discussed in Discord... This struct has evolved to include a lot of
information apart from the `petgraph::Graph`. And I want to add a graph
to the simplified `Resolution` type. So I think this name makes more
sense.
Surprisingly, this seems to be all that's necessary.
Previously, we were only extracting an extra from a
PubGrubPackage to test for conflicts. But now we extract
either an extra or a group. The surrounding code all
remains the same.
We do need to add some extra checking for groups
specifically, but I believe that's it.
This adds support for providing conflicting group names in addition to
extra names to `Conflicts`.
This merely makes "room" for it in the types while keeping everything
working. We'll add proper support for it in the next commit.
Note that one interesting trick we do here is depend directly on
`hashbrown` so that we can make use of its `Equivalent` trait. This in
turn lets us use things like `ConflictItemRef` as a lookup key for a
hashset that contains `ConflictItem`. This mirrors using a `&str` as a
lookup key for a hashset that contains `String`, but works for arbitrary
types. `std` doesn't support this, but `hashbrown` does. This trick in
turn lets us simplify some of our data structures.
This also rejiggers some of the serde-interaction with the conflicting
types. We now use a wire type to represent our conflicting items for
more flexibility. i.e., Support `extra` XOR `group` fields.
Since this is intended to support _both_ groups and extras, it doesn't
make sense to just name it for groups. And since there isn't really a
word that encapsulates both "extra" and "group," we just fall back to
the super general "conflicts."
We'll rename the variables and other things in the next commit.
## Summary
I need this for the derivation chain work
(https://github.com/astral-sh/uv/issues/8962), but it just seems
generally useful. You can't always get a version from a `Dist` (it could
be URL-based!), but when we create a `ResolvedDist`, we _do_ know the
version (and not just the URL). This PR preserves it.
This PR adds support for conflicting extras. For example, consider
some optional dependencies like this:
```toml
[project.optional-dependencies]
project1 = ["numpy==1.26.3"]
project2 = ["numpy==1.26.4"]
```
These dependency specifications are not compatible with one another.
And if you ask uv to lock these, you'll get an unresolvable error.
With this PR, you can now add this to your `pyproject.toml` to get
around this:
```toml
[tool.uv]
conflicting-groups = [
[
{ package = "project", extra = "project1" },
{ package = "project", extra = "project2" },
],
]
```
This will make the universal resolver create additional forks
internally that keep the dependencies from the `project1` and
`project2` extras separate. And we make all of this work by reporting
an error at **install** time if one tries to install with two or more
extras that have been declared as conflicting. (If we didn't do this,
it would be possible to try and install two different versions of the
same package into the same environment.)
This PR does *not* add support for conflicting **groups**, but it is
intended to add support in a follow-up PR.
Closes#6981Fixes#8024
Ref #6729, Ref #6830
This should also hopefully unblock
https://github.com/dagster-io/dagster/pull/23814, but in my testing, I
did run into other problems (specifically, with `pywin`). But it does
resolve the problem with incompatible dependencies in two different
extras once you declare `test-airflow-1` and `test-airflow-2` as
conflicting for `dagster-airflow`.
NOTE: This PR doesn't make `conflicting-groups` public yet. And in a
follow-up PR, I plan to switch the name to `conflicts` instead of
`conflicting-groups`, since it will be able to accept conflicting extras
_and_ conflicting groups.
## Summary
We're inconsistent with these -- sometimes it's `Error::Fetch` and
sometimes it's `Error::Download`. The message says download, so let's
just use that?
## Summary
This got moved to `InstallTarget`! Must've been an oversight not to
delete. I verified that no code was changed here since the date that we
moved it to `InstallTarget`.
## Summary
Just as we don't enforce tag compliance, we shouldn't enforce
`--no-build` when validating the lockfile. If we end up building from
source, the distribution database will correctly error.
Closes https://github.com/astral-sh/uv/issues/9016.
## Summary
At time of writing, `markupsafe==3.0.2` exists on the PyTorch index, but
there's
only a single wheel:
`MarkupSafe-3.0.2-cp313-cp313-manylinux_2_17_x86_64.manylinux2014_x86_64.whl`
Meanwhile, there are a large number of wheels on PyPI for the same
version. If the
user is on Python 3.12, and we return the incompatible PyTorch wheel
without
considering the PyPI wheels, PubGrub will mark 3.0.2 as an incompatible
version,
even though there are compatible wheels on PyPI.
Closes https://github.com/astral-sh/uv/issues/8922.
## Summary
We were making some incorrect assumptions in the extra-merging code for
universal `pip compile`. This PR corrects those assumptions and adds a
bunch of additional tests.
Closes https://github.com/astral-sh/uv/issues/8915.
After https://github.com/astral-sh/uv/pull/8797, we have spec-compliant
handling for local version identifiers and can completely remove all the
special-casing around it.
Implement a full working version of local version semantics. The (AFAIA)
major move towards this was implemented in #2430. This added support
such that the version specifier `torch==2.1.0+cpu` would install
`torch@2.1.0+cpu` and consider `torch@2.1.0+cpu` a valid way to satisfy
the requirement `torch==2.1.0` in further dependency resolution.
In this feature, we more fully support local version semantics. Namely,
we now allow `torch==2.1.0` to install `torch@2.1.0+cpu` regardless of
whether `torch@2.1.0` (no local tag) actually exists.
We do this by adding an internal-only `Max` value to local versions that
compare greater to all other local versions. Then we can translate
`torch==2.1.0` into bounds: greater than 2.1.0 with no local tag and
less than 2.1.0 with the `Max` local tag.
Depends on https://github.com/astral-sh/packse/pull/227.
closes#6640
Could you suggest how I should test it?
(already tested locally)
---------
Co-authored-by: konstin <konstin@mailbox.org>
Co-authored-by: Charles Tapley Hoyt <cthoyt@gmail.com>
Co-authored-by: Charlie Marsh <charlie.r.marsh@gmail.com>
This updates the surrounding code to use the new ResolverEnvironment
type. In some cases, this simplifies caller code by removing case
analysis. There *shouldn't* be any behavior changes here. Some test
snapshots were updated to account for some minor tweaks to error
messages.
I didn't split this up into separate commits because it would have been
too difficult/costly.
This type is intended to replace `ResolverMarkers`. The main difference
between them is that this type encapsulates more decision making by
un-exporting the different cases. So instead of callers needing to do
explicit case analysis depending on the type of resolver environment,
callers instead use methods that know how to do the right thing. In the
next commit, there are at least a few cases where this greatly
simplifies case analysis on the caller side.
The motivation for this type is to centralize decision making about
forking. In particular, we want to expand forking to include conflicting
groups instead of just `MarkerTree`. So to a certain extent, the
refactor here is about removing bare use of `MarkerTree` in favor of a
more purpose built type that encapsulates the forking logic.
The encapsulation is not quite perfect here. I expect to improve on it a
bit once we add support for conflicting groups.
This is split off from the subsequent commit (that makes use of
`ResolverEnvironment`) so that it's a bit easier to review the addition
in isolation.
