## Summary
Now, after running `pip-install`, we validate that the set of installed
packages is consistent -- that is, that we don't have any packages that
are missing dependencies, or incompatible versions of installed
dependencies.
## Summary
At present, when performing a `pip-install`, we first do a resolution,
then take the set of requirements and basically run them through our
`pip-sync`, which itself includes re-resolving the dependencies to get a
specific `Dist` for each package. (E.g., the set of requirements might
say `flask==3.0.0`, but the installer needs a specific _wheel_ or source
distribution to install.)
This PR removes this second resolution by exposing the set of pinned
packages from the resolution. The main challenge here is that we have an
optimization in the resolver such that we let the resolver read metadata
from an incompatible wheel as long as a source distribution exists for a
given package. This lets us avoid building source distributions in the
resolver under the assumption that we'll be able to install the package
later on, if needed. As such, the resolver now needs to track the
resolution and installation filenames separately.
## Summary
At present, we have two separate phases within the installation pipeline
related to populating wheels into the cache. The first phase downloads
the distribution, and then builds any source distributions into wheels;
the second phase unzips all the built wheels into the cache.
This PR merges those two phases into one, such that we seamlessly
download, build, and unzip wheels in one pass. This is more efficient,
since we can start unzipping while we build. It also ensures that if the
install _fails_ partway through, we don't end up with a bunch of
downloaded wheels that we never had a chance to unzip. The code is also
much simpler.
The main downside is that the user-facing feedback isn't as granular,
since we only have one phase and one progress bar for what was
originally three distinct phases.
Closes https://github.com/astral-sh/puffin/issues/571.
## Test Plan
I ran the benchmark script on two separate requirements files, and saw a
7% and 31% speedup respectively:
```text
+ TARGET=./scripts/benchmarks/requirements.txt
+ hyperfine --runs 100 --warmup 10 --prepare 'virtualenv --clear .venv' './target/release/main pip-sync ./scripts/benchmarks/requirements.txt --no-cache' --prepare 'virtualenv --clear .venv' './target/release/puffin pip-sync ./scripts/benchmarks/requirements.txt --no-cache'
Benchmark 1: ./target/release/main pip-sync ./scripts/benchmarks/requirements.txt --no-cache
Time (mean ± σ): 269.4 ms ± 33.0 ms [User: 42.4 ms, System: 117.5 ms]
Range (min … max): 221.7 ms … 446.7 ms 100 runs
Benchmark 2: ./target/release/puffin pip-sync ./scripts/benchmarks/requirements.txt --no-cache
Time (mean ± σ): 250.6 ms ± 28.3 ms [User: 41.5 ms, System: 127.4 ms]
Range (min … max): 207.6 ms … 336.4 ms 100 runs
Summary
'./target/release/puffin pip-sync ./scripts/benchmarks/requirements.txt --no-cache' ran
1.07 ± 0.18 times faster than './target/release/main pip-sync ./scripts/benchmarks/requirements.txt --no-cache'
```
```text
+ TARGET=./scripts/benchmarks/requirements-large.txt
+ hyperfine --runs 100 --warmup 10 --prepare 'virtualenv --clear .venv' './target/release/main pip-sync ./scripts/benchmarks/requirements-large.txt --no-cache' --prepare 'virtualenv --clear .venv' './target/release/puffin pip-sync ./scripts/benchmarks/requirements-large.txt --no-cache'
Benchmark 1: ./target/release/main pip-sync ./scripts/benchmarks/requirements-large.txt --no-cache
Time (mean ± σ): 5.053 s ± 0.354 s [User: 1.413 s, System: 6.710 s]
Range (min … max): 4.584 s … 6.333 s 100 runs
Benchmark 2: ./target/release/puffin pip-sync ./scripts/benchmarks/requirements-large.txt --no-cache
Time (mean ± σ): 3.845 s ± 0.225 s [User: 1.364 s, System: 6.970 s]
Range (min … max): 3.482 s … 4.715 s 100 runs
Summary
'./target/release/puffin pip-sync ./scripts/benchmarks/requirements-large.txt --no-cache' ran
```
Currently, `dbg!` is hard to read because versions are verbose, showing
all optional fields, and we have a lot of versions. Changing debug
formatting to displaying the version number (which can be losslessly
converted to the struct and back) makes this more readable.
See e.g.
https://gist.github.com/konstin/38c0f32b109dffa73b3aa0ab86b9662b
**Before**
```text
version: Version {
epoch: 0,
release: [
1,
2,
3,
],
pre: None,
post: None,
dev: None,
local: None,
},
```
**After**
```text
version: "1.2.3",
```
I saw warnings when we were e.g. unzipping wheel and setuptools in two
tasks at the same time. We now keep track of in flight unzips.
This introduces a `OnceMap` abstraction which we also use in the
resolver.
Extends #517 with a suggestion from @konstin to parse the `SimpleJson`
into an intermediate type `SimpleMetadata(BTreeMap<Version,
VersionFiles>)` before converting to a `VersionMap`. This reduces the
number of times we need to parse the response. Additionally, we cache
the parsed response now instead of `SimpleJson`.
`VersionFiles` stores two vectors with
`WheelFilename`/`SourceDistFilename` and `File` tuples. These can be
iterated over together or separately. A new enum `DistFilename` was
added to capture the `SourceDistFilename` and `WheelFilename` variants
allowing iteration over both vectors.
I'm working off of @konstin's commit here to implement arbitrary unsat
test cases for the resolver.
The entirety of the resolver's io are two functions: Get the version map
for a package (PEP 440 version -> distribution) and get the metadata for
a distribution. A new trait `ResolverProvider` abstracts these two away and
allows replacing the real network requests e.g. with stored responses
(https://github.com/pradyunsg/pip-resolver-benchmarks/blob/main/scenarios/pyrax_198.json).
---------
Co-authored-by: konsti <konstin@mailbox.org>
## Summary
This enables users to rely on yanked versions via explicit `==` markers,
which is necessary in some projects (and, in my opinion, reasonable).
Closes#551.
It turns out that it's not uncommon to use timestamps as patch versions
(e.g., `20230628214621`). I believe this is the ISO 8601 "basic format".
These can't be represented by a `u32`, so I think it makes sense to just
bump to `u64` to remove this limitation.
## Summary
This PR modifies the behavior of our `--python-version` override in two
ways:
1. First, we always use the "real" interpreter in the source
distribution builder. I think this is correct. We don't need to use the
fake markers for recursive builds, because all we care about is the
top-level resolution, and we already assume that a single source
distribution will always return the same metadata regardless of its
build environment.
2. Second, we require that source distributions are compatible with
_both_ the "real" interpreter version and the marker environment. This
ensures that we don't try to build source distributions that are
compatible with our interpreter, but incompatible with the target
version.
Closes https://github.com/astral-sh/puffin/issues/407.
This removes the last usage of cacache by replacing it with a custom,
flat json caching keyed by the digest of the executable path.
A step towards #478. I've made `CachedByTimestamp<T>` generic over `T`
but intentionally not moved it to `puffin-cache` yet.
This is mostly a mechanical refactor that moves 80% of our code to the
same cache abstraction.
It introduces cache `Cache`, which abstracts away the path of the cache
and the temp dir drop and is passed throughout the codebase. To get a
specific cache bucket, you need to requests your `CacheBucket` from
`Cache`. `CacheBucket` is the centralizes the names of all cache
buckets, moving them away from the string constants spread throughout
the crates.
Specifically for working with the `CachedClient`, there is a
`CacheEntry`. I'm not sure yet if that is a strict improvement over
`cache_dir: PathBuf, cache_file: String`, i may have to rotate that
later.
The interpreter cache moved into `interpreter-v0`.
We can use the `CacheBucket` page to document the cache structure in
each bucket:
## Summary and motivation
For a given source dist, we store the metadata of each wheel built
through it in `built-wheel-metadata-v0/pypi/<source dist
filename>/metadata.json`. During resolution, we check the cache status
of the source dist. If it is fresh, we check `metadata.json` for a
matching wheel. If there is one we use that metadata, if there isn't, we
build one. If the source is stale, we build a wheel and override
`metadata.json` with that single wheel. This PR thereby ties the local
built wheel metadata cache to the freshness of the remote source dist.
This functionality is available through `SourceDistCachedBuilder`.
`puffin_installer::Builder`, `puffin_installer::Downloader` and
`Fetcher` are removed, instead there are now `FetchAndBuild` which calls
into the also new `SourceDistCachedBuilder`. `FetchAndBuild` is the new
main high-level abstraction: It spawns parallel fetching/building, for
wheel metadata it calls into the registry client, for wheel files it
fetches them, for source dists it calls `SourceDistCachedBuilder`. It
handles locks around builds, and newly added also inter-process file
locking for git operations.
Fetching and building source distributions now happens in parallel in
`pip-sync`, i.e. we don't have to wait for the largest wheel to be
downloaded to start building source distributions.
In a follow-up PR, I'll also clear built wheels when they've become
stale.
Another effect is that in a fully cached resolution, we need neither zip
reading nor email parsing.
Closes#473
## Source dist cache structure
Entries by supported sources:
* `<build wheel metadata cache>/pypi/foo-1.0.0.zip/metadata.json`
* `<build wheel metadata
cache>/<sha256(index-url)>/foo-1.0.0.zip/metadata.json`
* `<build wheel metadata
cache>/url/<sha256(url)>/foo-1.0.0.zip/metadata.json`
But the url filename does not need to be a valid source dist filename
(<https://github.com/search?q=path%3A**%2Frequirements.txt+master.zip&type=code>),
so it could also be the following and we have to take any string as
filename:
* `<build wheel metadata
cache>/url/<sha256(url)>/master.zip/metadata.json`
Example:
```text
# git source dist
pydantic-extra-types @ git+https://github.com/pydantic/pydantic-extra-types.git
# pypi source dist
django_allauth==0.51.0
# url source dist
werkzeug @ ff1904eb5e/werkzeug-3.0.1.tar.gz
```
will be stored as
```text
built-wheel-metadata-v0
├── git
│ └── 5c56bc1c58c34c11
│ └── 843b753e9e8cb74e83cac55598719b39a4d5ef1f
│ └── metadata.json
├── pypi
│ └── django-allauth-0.51.0.tar.gz
│ └── metadata.json
└── url
└── 6781bd6440ae72c2
└── werkzeug-3.0.1.tar.gz
└── metadata.json
```
The inside of a `metadata.json`:
```json
{
"data": {
"django_allauth-0.51.0-py3-none-any.whl": {
"metadata-version": "2.1",
"name": "django-allauth",
"version": "0.51.0",
...
}
}
}
```
Preparing for #235, some refactoring to `puffin_interpreter`.
* Added a dedicated error type instead of anyhow
* `InterpreterInfo` -> `Interpreter`
* `detect_virtual_env` now returns an option so it can be chained for
#235
## Summary
This PR adds support for local path dependencies. The approach mostly
just falls out of our existing approach and infrastructure for Git and
URL dependencies.
Closes https://github.com/astral-sh/puffin/issues/436. (We'll open a
separate issue for editable installs.)
## Test Plan
Added `pip-compile` tests that pre-download a wheel or source
distribution, then install it via local path.
## Summary
A variety of small refactors to the distribution types crate to (1)
return `Result` if we find an invalid wheel, rather than treating it as
a source distribution with a `.whl` suffix, and (2) DRY up some repeated
code around URLs.
A consistent cache structure for remote wheel metadata:
* `<wheel metadata cache>/pypi/foo-1.0.0-py3-none-any.json`
* `<wheel metadata
cache>/<digest(index-url)>/foo-1.0.0-py3-none-any.json`
* `<wheel metadata cache>/url/<digest(url)>/foo-1.0.0-py3-none-any.json`
The source dist caching will use a similar structure (#468).
## Summary
This PR unifies the behavior that lived in the resolver's `distribution`
crates with the behaviors that were spread between the various structs
in the installer crate into a single `Fetcher` struct that is intended
to manage all interactions with distributions. Specifically, the
interface of this struct is such that it can access distribution
metadata, download distributions, return those downloads, etc., all with
a common cache.
Overall, this is mostly just DRYing up code that was repeated between
the two crates, and putting it behind a reasonable shared interface.
## Summary
This crate only contains types, and I want to introduce a new crate for
all _operations_ on distributions, so this feels like a more natural
name given we also have `pypi-types`.
## Summary
This is a refactor to address a TODO in the build context whereby we
aren't respecting the resolution options in recursive resolutions. Now,
the options are split out from the resolution _manifest_, and shared
across the build context tree.
This script can compare different requirements between pip(-compile) and
puffin across python versions, with debug and release builds.
Examples:
```shell
scripts/compare_with_pip/compare_with_pip.py
scripts/compare_with_pip/compare_with_pip.py -p 3.10
scripts/compare_with_pip/compare_with_pip.py --release -p 3.9 --target 'transformers[deepspeed-testing,dev-tensorflow]'
```
It found a bunch of fixed bugs, e.g. the lack of yanked package handling
and source dist handling, as well as #423, which is currently most of
the output.
Example output:
https://gist.github.com/konstin/9ccf8dc7c2dcca737bf705429ced4892#443 should be merged first
Extends #424 with support for URL dependency incompatibilities.
Requires changes to `miette` to prevent URLs from being word wrapped;
accepted upstream in https://github.com/zkat/miette/pull/321
We're willing to use platform-incompatible wheels during resolution, to
quicken access to metadata... But we should avoid choosing an
incompatible wheel if the package lacks a source distribution since, in
that case, we definitely won't be able to install it.
Closes https://github.com/astral-sh/puffin/issues/439.
Always¹ clear the temporary directories we create.
* Clear source dist downloads: Previously, the temporary directories
would remain in the cache dir, now they are cleared properly
* Clear wheel file downloads: Delete the `.whl` file, we only need to
cache the unpacked wheel
* Consistent handling of cache arguments: Abstract the handling for CLI
cache args away, again making sure we remove the `--no-cache` temp dir.
There are no more `into_path()` calls that persist `TempDir`s that i
could find.
¹Assuming drop is run, and deleting the directory doesn't silently
error.
Addresses
https://github.com/astral-sh/puffin/issues/309#issuecomment-1792648969
Similar to #338 this throws an error when merging versions results in an
empty set. Instead of propagating that error, we capture it and return a
new dependency type of `Unusable`. Unusable dependencies are a new
incompatibility kind which includes an arbitrary "reason" string that we
present to the user. Adding a new incompatibility kind requires changes
to the vendored pubgrub crate.
We could use this same incompatibility kind for conflicting urls as in
#284 which should allow the solver to backtrack to another valid version
instead of failing (see #425).
Unlike #383 this does not require changes to PubGrub's package mapping
model. I think in the long run we'll want PubGrub to accept multiple
versions per package to solve this specific issue, but we're interested
in it being merged upstream first. This pull request is just using the
issue as a simple case to explore adding a new incompatibility type.
We may or may not be able convince them to add this new incompatibility
type upstream. As discussed in
https://github.com/pubgrub-rs/pubgrub/issues/152, we may want a more
general incompatibility kind instead which can be used for arbitrary
problems. An upstream pull request has been opened for discussion at
https://github.com/pubgrub-rs/pubgrub/pull/153.
Related to:
- https://github.com/pubgrub-rs/pubgrub/issues/152
- #338
- #383
---------
Co-authored-by: konsti <konstin@mailbox.org>
A fork will let us stay up to date with the upstream while replaying our
work on top of it.
I expect a similar workflow to the RustPython-Parser fork we maintained,
except that I wrote an automation to create tags for each commit on the
fork (https://github.com/zanieb/pubgrub/pull/2) so we do not need to
manually tag and document each commit.
To update with the upstream:
- Rebase our fork's `main` branch on top of the latest changes in
upstream's `dev` branch
- Force push, overwriting our `main` branch history
- Change the commit hash here to the last commit on `main` in our fork
Since we automatically tag each commit on the fork, we should never lose
the commits that are dropped from `main` during rebase.
This works by filtering out files with a more recent upload time, so if
the index you use does not provide upload times, the results might be
inaccurate. pypi provides upload times for all files. This is, the field
is non-nullable in the warehouse schema, but the simple API PEP does not
know this field.
If you have only pypi dependencies, this means deterministic,
reproducible(!) resolution. We could try doing the same for git repos
but it doesn't seem worth the effort, i'd recommend pinning commits
since git histories are arbitrarily malleable and also if you care about
reproducibility and such you such not use git dependencies but a custom
index.
Timestamps are given either as RFC 3339 timestamps such as
`2006-12-02T02:07:43Z` or as UTC dates in the same format such as
`2006-12-02`. Dates are interpreted as including this day, i.e. until
midnight UTC that day. Date only is required to make this ergonomic and
midnight seems like an ergonomic choice.
In action for `pandas`:
```console
$ target/debug/puffin pip-compile --exclude-newer 2023-11-16 target/pandas.in
Resolved 6 packages in 679ms
# This file was autogenerated by Puffin v0.0.1 via the following command:
# target/debug/puffin pip-compile --exclude-newer 2023-11-16 target/pandas.in
numpy==1.26.2
# via pandas
pandas==2.1.3
python-dateutil==2.8.2
# via pandas
pytz==2023.3.post1
# via pandas
six==1.16.0
# via python-dateutil
tzdata==2023.3
# via pandas
$ target/debug/puffin pip-compile --exclude-newer 2022-11-16 target/pandas.in
Resolved 5 packages in 655ms
# This file was autogenerated by Puffin v0.0.1 via the following command:
# target/debug/puffin pip-compile --exclude-newer 2022-11-16 target/pandas.in
numpy==1.23.4
# via pandas
pandas==1.5.1
python-dateutil==2.8.2
# via pandas
pytz==2022.6
# via pandas
six==1.16.0
# via python-dateutil
$ target/debug/puffin pip-compile --exclude-newer 2021-11-16 target/pandas.in
Resolved 5 packages in 594ms
# This file was autogenerated by Puffin v0.0.1 via the following command:
# target/debug/puffin pip-compile --exclude-newer 2021-11-16 target/pandas.in
numpy==1.21.4
# via pandas
pandas==1.3.4
python-dateutil==2.8.2
# via pandas
pytz==2021.3
# via pandas
six==1.16.0
# via python-dateutil
```
[PEP 691](https://peps.python.org/pep-0691/#project-detail) has slightly
different, more relaxed rules around file metadata. These changes are
now reflected in the `File` struct. This will make it easier to support
alternative indices.
I had expected that i need to introduce a separate type for that, so i'm
happy it's two `Option`s more and an alias.
Part of #412
Previously, we were assuming that `which <python>` return the path to
the python executable. This is not true when using pyenv shims, which
are bash scripts. Instead, we have to use `sys.executable`. Luckily,
we're already querying the python interpreter and can do it in that
pass.
We are also not allowed to cache the execution of the python interpreter
through the shim because pyenv might change the target. As a heuristic,
we check whether `sys.executable`, the real binary, is the same our
canonicalized `which` result.
---------
Co-authored-by: Zanie Blue <contact@zanie.dev>
## Summary
This PR implements logic to sort wheels by priority, where priority is
defined as preferring more "specific" wheels over less "specific"
wheels. For example, in the case of Black, my machine now selects
`black-23.11.0-cp311-cp311-macosx_11_0_arm64.whl`, whereas sorting by
lowest priority instead gives me `black-23.11.0-py3-none-any.whl`.
As part of this change, I've also modified the resolver to fallback to
using incompatible wheels when determining package metadata, if no
compatible wheels are available.
The `VersionMap` was also moved out of `resolver.rs` and into its own
file with a wrapper type, for clarity.
Closes https://github.com/astral-sh/puffin/issues/380.
Closes https://github.com/astral-sh/puffin/issues/421.
The pip compile test now explicitly set their python version and `puffin
venv` resolves e.g. `python3.12` correctly now. The venv creation is
moved to a shared method
Implement two behaviors for yanked versions:
* During `pip-compile`, yanked versions are filtered out entirely, we
currently treat them is if they don't exist. This is leads to confusing
error messages because a version that does exist seems to have suddenly
disappeared.
* During `pip-sync`, we warn when we fetch a remote distribution and it
has been yanked. We currently don't warn on cached or installed
distributions that have been yanked.
Filter out source dists and wheels whose `requires-python` from the
simple api is incompatible with the current python version.
This change showed an important problem: When we use a fake python
version for resolving, building source distributions breaks down because
we can only build with versions we actually have.
This change became surprisingly big. The tests now require python 3.7 to
be installed, but changing that would mean an even bigger change.
Fixes#388
This copies the allocator configuration used in the Ruff project. In
particular, this gives us an instant 10% win when resolving the top 1K
PyPI packages:
$ hyperfine \
"./target/profiling/puffin-dev-main resolve-many --cache-dir
cache-docker-no-build --no-build pypi_top_8k_flat.txt --limit 1000 2>
/dev/null" \
"./target/profiling/puffin-dev resolve-many --cache-dir
cache-docker-no-build --no-build pypi_top_8k_flat.txt --limit 1000 2>
/dev/null"
Benchmark 1: ./target/profiling/puffin-dev-main resolve-many --cache-dir
cache-docker-no-build --no-build pypi_top_8k_flat.txt --limit 1000 2>
/dev/null
Time (mean ± σ): 974.2 ms ± 26.4 ms [User: 17503.3 ms, System: 2205.3
ms]
Range (min … max): 943.5 ms … 1015.9 ms 10 runs
Benchmark 2: ./target/profiling/puffin-dev resolve-many --cache-dir
cache-docker-no-build --no-build pypi_top_8k_flat.txt --limit 1000 2>
/dev/null
Time (mean ± σ): 883.1 ms ± 23.3 ms [User: 14626.1 ms, System: 2542.2
ms]
Range (min … max): 849.5 ms … 916.9 ms 10 runs
Summary
'./target/profiling/puffin-dev resolve-many --cache-dir
cache-docker-no-build --no-build pypi_top_8k_flat.txt --limit 1000 2>
/dev/null' ran
1.10 ± 0.04 times faster than './target/profiling/puffin-dev-main
resolve-many --cache-dir cache-docker-no-build --no-build
pypi_top_8k_flat.txt --limit 1000 2> /dev/null'
I was moved to do this because I noticed `malloc`/`free` taking up a
fairly sizeable percentage of time during light profiling.
As is becoming a pattern, it will be easier to review this
commit-by-commit.
Ref #396 (wouldn't call this issue fixed)
-----
I did also try adding a `smallvec` optimization to the
`Version::release` field, but it didn't bare any fruit. I still think
there is more to explore since the results I observed don't quite line
up with what I expect. (So probably either my mental model is off or my
measurement process is flawed.) You can see that attempt with a little
more explanation here:
f9528b4ecd
In the course of adding the `smallvec` optimization, I also shrunk the
`Version` fields from a `usize` to a `u32`. They should at least be a
fixed size integer since version numbers aren't used to index memory,
and I shrunk it to `u32` since it seems reasonable to assume that all
version numbers will be smaller than `2^32`.
Before:
```
cargo run --bin puffin-dev -q -- resolve-cli "transformers[accelerate, agents, all, audio, codecarbon, deepspeed, deepspeed-testing, dev, dev-tensorflow, dev-torch, docs, docs_specific, flax, flax-speech, ftfy, integrations, ja, modelcreation, onnx, onnxruntime, optuna, quality, ray, retrieval, sagemaker, sentencepiece, serving, sigopt, sklearn, speech, testing, tf, tf-cpu, tf-speech, timm, tokenizers, torch, torch-speech, torch-vision, torchhub, video, vision]"
puffin-dev failed
Caused by: No solution found when resolving: transformers[accelerate,agents,all,audio,codecarbon,deepspeed,deepspeed-testing,dev,dev-tensorflow,dev-torch,docs,docs-specific,flax,flax-speech,ftfy,integrations,ja,modelcreation,onnx,onnxruntime,optuna,quality,ray,retrieval,sagemaker,sentencepiece,serving,sigopt,sklearn,speech,testing,tf,tf-cpu,tf-speech,timm,tokenizers,torch,torch-speech,torch-vision,torchhub,video,vision]
Caused by: Not a valid package or extra name: ".none". Names must start and end with a letter or digit and may only contain -, _, ., and alphanumeric characters
```
After:
```
cargo run --bin puffin-dev -q -- resolve-cli "transformers[accelerate, agents, all, audio, codecarbon, deepspeed, deepspeed-testing, dev, dev-tensorflow, dev-torch, docs, docs_specific, flax, flax-speech, ftfy, integrations, ja, modelcreation, onnx, onnxruntime, optuna, quality, ray, retrieval, sagemaker, sentencepiece, serving, sigopt, sklearn, speech, testing, tf, tf-cpu, tf-speech, timm, tokenizers, torch, torch-speech, torch-vision, torchhub, video, vision]"
puffin-dev failed
Caused by: No solution found when resolving: transformers[accelerate,agents,all,audio,codecarbon,deepspeed,deepspeed-testing,dev,dev-tensorflow,dev-torch,docs,docs-specific,flax,flax-speech,ftfy,integrations,ja,modelcreation,onnx,onnxruntime,optuna,quality,ray,retrieval,sagemaker,sentencepiece,serving,sigopt,sklearn,speech,testing,tf,tf-cpu,tf-speech,timm,tokenizers,torch,torch-speech,torch-vision,torchhub,video,vision]
Caused by: Couldn't parse metadata in fastapi-0.10.1-py3-none-any.whl (97ac91cb7c/fastapi-0.10.1-py3-none-any.whl)
Caused by: Not a valid package or extra name: ".none". Names must start and end with a letter or digit and may only contain -, _, ., and alphanumeric characters
```
## Summary
It looks like, when you install `pip`, it includes a bunch of
`__pycache__` directories in the RECORD file (although these directories
don't exist until you run `pip`). Our uninstaller assumed that the
RECORD file only contained _files_.
Closes https://github.com/astral-sh/puffin/issues/389.
