Files
lancedb/.github/workflows/pypi-publish.yml
T
Tobias 7b6ee0d655 feat(wheels): publish lancedb-compat for pre-Haswell x86_64 hosts (#3327)
Tracks #3324. On x86_64 CPUs without AVX2 (Sandy Bridge / Ivy Bridge /
Westmere on Intel; Bulldozer / Piledriver / Steamroller on AMD), `import
lancedb` SIGILLs because the wheel bakes AVX2 + FMA into every compiled
function. Per [westonpace's
review](https://github.com/lancedb/lancedb/issues/3324#issuecomment-4328944354),
the default `lancedb` wheel stays fast; pre-Haswell users get a
separately-published `lancedb-compat` wheel.

## Summary

- Adds a `lancedb-compat` matrix entry to `pypi-publish.yml` that builds
with `RUSTFLAGS="-C target-cpu=x86-64-v2"` (Nehalem-class baseline).
Same Python API (`import lancedb` works) — files install to the same
namespace, so the two wheels conflict at install time and users pick
one. Same pattern as `psycopg2` / `psycopg2-binary` and `tensorflow` /
`tensorflow-cpu`.
- Generalizes `build_linux_wheel` and `upload_wheel` composites with
optional `package-name` and `rustflags` inputs (defaults preserve the
existing 4 `lancedb` matrix entries verbatim).
- Documents the choice in `python/README.md`: `pip install
lancedb-compat` for pre-Haswell hosts.

The default `.cargo/config.toml` baseline is unchanged.

## Sequencing

1. ~~lance-format/lance#6630 merges → runtime SIMD dispatch lands in
lance.~~ **Done — merged.**
2. lancedb's lance dep is bumped to a release that includes it (separate
PR / normal cadence).
3. This PR's `lancedb-compat` wheel build path starts producing a wheel
that runs on pre-Haswell hardware. **Maintainer setup**: register
`lancedb-compat` on PyPI and configure trusted publishing.

## Verified end-to-end on Sandy Bridge Xeon E5-2609

Verification was done locally against a fork-pinned lance dep that
includes the runtime dispatch implementation, using the same
`RUSTFLAGS="-C target-cpu=x86-64-v2"` flags this PR uses in CI:

```
$ RUSTFLAGS="-C target-cpu=x86-64-v2" maturin build --release
$ pip install ./target/wheels/lancedb-*.whl
$ python verify.py
PASS: import + simd dispatch + table create + vector search all work.
```

Pre-fix on the same CPU (default `pip install lancedb`): `Illegal
instruction (core dumped)`. Full reproducer (deps + clone + build +
verification):
https://gist.github.com/tobocop2/2e341358b55c143527416edfdb1e37df.
Fork-internal verification PR with the dep bump and full logs:
[`tobocop2/lancedb#2`](https://github.com/tobocop2/lancedb/pull/2).

## Benchmarks — no regressions on modern CPUs from the lance-side change

These are the numbers I ran for the lance PR, confirming the runtime
dispatch doesn't slow down the default (`target-cpu=haswell`) wheel that
existing users install. Criterion, one machine, one session, base → PR,
no `RUSTFLAGS` override. Full methodology, null experiments, and logs:
[lance-format/lance#6630 benchmark
comment](https://github.com/lance-format/lance/pull/6630#issuecomment-4933063394)
and the [logs
gist](https://gist.github.com/tobocop2/3c6d0f449cbd736aa2501f89a7fe56a2).

| benchmark | EPYC 7B13 (`avx2`, `fma`, no `avx512f`) | Xeon Cascade
Lake (`avx512f`) |
|---|---|---|
| `Cosine(f32, scalar)` *(control)* | +0.04% | +0.09% |
| `Cosine(f64, scalar)` | −0.34% | −1.94% |
| `Cosine(u8, SIMD)` | +2.30% | +3.63% |
| `Dot(f16, SIMD)` | −0.58% | +0.61% |
| `Dot(f32, SIMD)` | +0.34% | **−6.08%** |
| `Dot(f32, arrow_arity)` | +0.02% | −0.00% |
| `L2(f32, scalar)` | −0.10% | −0.02% |
| `L2(f32, simd)` (dim 1024) | +2.63% | −0.53% |
| **`L2(simd,f32x8)` (dim 8)** | **−45.9%** | **−25.1%** |
| `L2(u8, SIMD)` | +0.42% | −3.11% |
| `NormL2(f32, SIMD)` | −1.02% | −4.17% |
| `NormL2(f64, SIMD)` | +3.51% | −0.58% |

Nothing regresses beyond the noise floor. Dim 8 — the PQ sub-vector
width — improves 25–46%.

---

To be transparent: this isn't my domain of expertise and the lance-side
implementation is AI-generated. I verified it works end-to-end on the
failing hardware. Happy to roll in feedback.
2026-07-16 10:54:42 -07:00

266 lines
9.3 KiB
YAML

name: PyPI Publish
on:
push:
tags:
- 'python-v*'
pull_request:
# This should trigger a dry run (we skip the final publish step)
paths:
- .github/workflows/pypi-publish.yml
- .github/workflows/build_linux_wheel/action.yml
- .github/workflows/build_mac_wheel/action.yml
- .github/workflows/build_windows_wheel/action.yml
- Cargo.toml # Change in dependency frequently breaks builds
- Cargo.lock
env:
PIP_EXTRA_INDEX_URL: "https://pypi.fury.io/lance-format/ https://pypi.fury.io/lancedb/"
permissions:
contents: read
jobs:
linux:
name: Python ${{ matrix.config.package_name }} ${{ matrix.config.platform }} manylinux${{ matrix.config.manylinux }}
timeout-minutes: 60
strategy:
matrix:
config:
- platform: x86_64
manylinux: "2_28"
extra_args: "--features fp16kernels"
runner: ubuntu-22.04
package_name: "lancedb"
rustflags: ""
# For successful fat LTO builds, we need a large runner to avoid OOM errors.
- platform: aarch64
manylinux: "2_28"
extra_args: "--features fp16kernels"
runner: ubuntu-2404-8x-arm64
package_name: "lancedb"
rustflags: ""
# `lancedb-compat`: pre-Haswell-friendly variant for x86_64 hosts
# without AVX2 (Sandy Bridge / Ivy Bridge / Westmere on Intel,
# Bulldozer / Piledriver / Steamroller on AMD). Compiled at the
# `x86-64-v2` baseline; runtime SIMD dispatch in lance-linalg
# picks the appropriate tier (scalar / AVX / AVX+FMA / AVX2+FMA
# / AVX-512) at load time. Same import as `lancedb` -- conflicts
# at install time, so users pick one.
- platform: x86_64
manylinux: "2_28"
extra_args: ""
runner: ubuntu-22.04
package_name: "lancedb-compat"
rustflags: "-Ctarget-cpu=x86-64-v2"
runs-on: ${{ matrix.config.runner }}
steps:
- uses: actions/checkout@v6
with:
fetch-depth: 0
lfs: true
- name: Set up Python
uses: actions/setup-python@v6
with:
python-version: "3.10"
- uses: ./.github/workflows/build_linux_wheel
with:
python-minor-version: 10
args: "--release --strip ${{ matrix.config.extra_args }}"
arm-build: ${{ matrix.config.platform == 'aarch64' }}
manylinux: ${{ matrix.config.manylinux }}
package-name: ${{ matrix.config.package_name }}
rustflags: ${{ matrix.config.rustflags }}
- uses: actions/upload-artifact@v7
if: startsWith(github.ref, 'refs/tags/python-v')
with:
name: wheels-linux-${{ matrix.config.package_name }}-${{ matrix.config.platform }}-${{ matrix.config.manylinux }}
path: target/wheels/*.whl
if-no-files-found: error
mac:
timeout-minutes: 90
runs-on: ${{ matrix.config.runner }}
strategy:
matrix:
config:
- target: aarch64-apple-darwin
runner: warp-macos-14-arm64-6x
env:
MACOSX_DEPLOYMENT_TARGET: 10.15
steps:
- uses: actions/checkout@v6
with:
fetch-depth: 0
lfs: true
- name: Set up Python
uses: actions/setup-python@v6
with:
python-version: "3.13"
- uses: ./.github/workflows/build_mac_wheel
with:
python-minor-version: 10
args: "--release --strip --target ${{ matrix.config.target }} --features fp16kernels"
- uses: actions/upload-artifact@v7
if: startsWith(github.ref, 'refs/tags/python-v')
with:
name: wheels-mac-${{ matrix.config.target }}
path: target/wheels/lancedb-*.whl
if-no-files-found: error
windows:
timeout-minutes: 90
runs-on: windows-latest
env:
# link.exe is single-threaded and the long pole on Windows builds. Use
# rustc's bundled lld-link instead.
CARGO_TARGET_X86_64_PC_WINDOWS_MSVC_LINKER: rust-lld
steps:
- uses: actions/checkout@v6
with:
fetch-depth: 0
lfs: true
- name: Set up Python
uses: actions/setup-python@v6
with:
python-version: "3.13"
- uses: ./.github/workflows/build_windows_wheel
with:
python-minor-version: 10
args: "--release --strip"
- uses: actions/upload-artifact@v7
if: startsWith(github.ref, 'refs/tags/python-v')
with:
name: wheels-windows
path: target/wheels/lancedb-*.whl
if-no-files-found: error
publish:
name: Publish wheels
if: startsWith(github.ref, 'refs/tags/python-v')
needs: [linux, mac, windows]
runs-on: ubuntu-latest
permissions:
id-token: write
contents: read
steps:
- uses: actions/checkout@v6
- name: Download wheel artifacts
uses: actions/download-artifact@v8
with:
pattern: wheels-*
path: target/wheels
merge-multiple: true
- name: List wheels
run: ls -la target/wheels
- name: Choose repo
id: choose_repo
run: |
if [[ ${{ github.ref }} == *beta* ]]; then
echo "repo=fury" >> $GITHUB_OUTPUT
else
echo "repo=pypi" >> $GITHUB_OUTPUT
fi
- name: Publish to Fury
if: steps.choose_repo.outputs.repo == 'fury'
env:
FURY_TOKEN: ${{ secrets.FURY_TOKEN }}
run: |
shopt -s nullglob
WHEELS=(target/wheels/*.whl)
if [[ ${#WHEELS[@]} -eq 0 ]]; then
echo "No wheels found in target/wheels/" >&2
exit 1
fi
for WHEEL in "${WHEELS[@]}"; do
echo "Uploading $WHEEL to Fury"
curl -f -F package=@"$WHEEL" "https://$FURY_TOKEN@push.fury.io/lancedb/"
done
# NOTE: pypa/gh-action-pypi-publish must be invoked directly from a
# workflow file, not from inside a composite action. When called from a
# composite, `github.action_repository` is empty (actions/runner#2473)
# and the action falls back to `github.repository`, producing a bogus
# `docker://ghcr.io/<repo>:<ref>` image reference that GHA tries to pull.
- name: Publish to PyPI
if: steps.choose_repo.outputs.repo == 'pypi'
uses: pypa/gh-action-pypi-publish@release/v1
with:
packages-dir: target/wheels/
gh-release:
if: startsWith(github.ref, 'refs/tags/python-v')
runs-on: ubuntu-latest
permissions:
contents: write
steps:
- uses: actions/checkout@v6
with:
fetch-depth: 0
lfs: true
- name: Extract version
id: extract_version
env:
GITHUB_REF: ${{ github.ref }}
run: |
set -e
echo "Extracting tag and version from $GITHUB_REF"
if [[ $GITHUB_REF =~ refs/tags/python-v(.*) ]]; then
VERSION=${BASH_REMATCH[1]}
TAG=python-v$VERSION
echo "tag=$TAG" >> $GITHUB_OUTPUT
echo "version=$VERSION" >> $GITHUB_OUTPUT
else
echo "Failed to extract version from $GITHUB_REF"
exit 1
fi
echo "Extracted version $VERSION from $GITHUB_REF"
if [[ $VERSION =~ beta ]]; then
echo "This is a beta release"
# Get last release (that is not this one)
FROM_TAG=$(git tag --sort='version:refname' \
| grep ^python-v \
| grep -vF "$TAG" \
| python ci/semver_sort.py python-v \
| tail -n 1)
else
echo "This is a stable release"
# Get last stable tag (ignore betas)
FROM_TAG=$(git tag --sort='version:refname' \
| grep ^python-v \
| grep -vF "$TAG" \
| grep -v beta \
| python ci/semver_sort.py python-v \
| tail -n 1)
fi
echo "Found from tag $FROM_TAG"
echo "from_tag=$FROM_TAG" >> $GITHUB_OUTPUT
- name: Create Python Release Notes
id: python_release_notes
uses: mikepenz/release-changelog-builder-action@v4
with:
configuration: .github/release_notes.json
toTag: ${{ steps.extract_version.outputs.tag }}
fromTag: ${{ steps.extract_version.outputs.from_tag }}
env:
GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
- name: Create Python GH release
uses: softprops/action-gh-release@v2
with:
prerelease: ${{ contains('beta', github.ref) }}
tag_name: ${{ steps.extract_version.outputs.tag }}
token: ${{ secrets.GITHUB_TOKEN }}
generate_release_notes: false
name: Python LanceDB v${{ steps.extract_version.outputs.version }}
body: ${{ steps.python_release_notes.outputs.changelog }}
report-failure:
name: Report Workflow Failure
runs-on: ubuntu-latest
needs: [linux, mac, windows, publish]
permissions:
contents: read
issues: write
if: always() && failure() && startsWith(github.ref, 'refs/tags/python-v')
steps:
- uses: actions/checkout@v6
- uses: ./.github/actions/create-failure-issue
with:
job-results: ${{ toJSON(needs) }}
workflow-name: ${{ github.workflow }}