feat: ship fp16kernels in Python wheels (#1148)

Same deal as https://github.com/lancedb/lance/pull/2098

.github/workflows/build_linux_wheel/action.yml

@@ -14,6 +14,10 @@ inputs:
     # Note: this does *not* mean the host is arm64, since we might be cross-compiling.
     required: false
     default: "false"
+  manylinux:
+    description: "The manylinux version to build for"
+    required: false
+    default: "2_17"
 runs:
   using: "composite"
   steps:
@@ -28,7 +32,7 @@ runs:
         command: build
         working-directory: python
         target: x86_64-unknown-linux-gnu
-        manylinux: "2_17"
+        manylinux: ${{ inputs.manylinux }}
         args: ${{ inputs.args }}
         before-script-linux: |
           set -e
@@ -43,7 +47,7 @@ runs:
         command: build
         working-directory: python
         target: aarch64-unknown-linux-gnu
-        manylinux: "2_24"
+        manylinux: ${{ inputs.manylinux }}
         args: ${{ inputs.args }}
         before-script-linux: |
           set -e

.github/workflows/pypi-publish.yml

@@ -6,13 +6,23 @@ on:
 
 jobs:
   linux:
+    name: Python ${{ matrix.config.platform }} manylinux${{ matrix.config.manylinux }}
     timeout-minutes: 60
     strategy:
       matrix:
         python-minor-version: ["8"]
-        platform:
-          - x86_64
-          - aarch64
+        config:
+          - platform: x86_64
+            manylinux: "2_17"
+            extra_args: ""
+          - platform: x86_64
+            manylinux: "2_28"
+            extra_args: "--features fp16kernels"
+          - platform: aarch64
+            manylinux: "2_24"
+            extra_args: ""
+          # We don't build fp16 kernels for aarch64, because it uses 
+          # cross compilation image, which doesn't have a new enough compiler.
     runs-on: "ubuntu-22.04"
     steps:
       - uses: actions/checkout@v4
@@ -26,8 +36,9 @@ jobs:
       - uses: ./.github/workflows/build_linux_wheel
         with:
           python-minor-version: ${{ matrix.python-minor-version }}
-          args: "--release --strip"
-          arm-build: ${{ matrix.platform == 'aarch64' }}
+          args: "--release --strip ${{ matrix.config.extra_args }}"
+          arm-build: ${{ matrix.config.platform == 'aarch64' }}
+          manylinux: ${{ matrix.config.manylinux }}
       - uses: ./.github/workflows/upload_wheel
         with:
           token: ${{ secrets.LANCEDB_PYPI_API_TOKEN }}
@@ -58,7 +69,7 @@ jobs:
       - uses: ./.github/workflows/build_mac_wheel
         with:
           python-minor-version: ${{ matrix.python-minor-version }}
-          args: "--release --strip --target ${{ matrix.config.target }}"
+          args: "--release --strip --target ${{ matrix.config.target }} --features fp16kernels"
       - uses: ./.github/workflows/upload_wheel
         with:
           python-minor-version: ${{ matrix.python-minor-version }}

.github/workflows/rust.yml

@@ -31,6 +31,10 @@ jobs:
       run:
         shell: bash
         working-directory: rust
+    env:
+      # Need up-to-date compilers for kernels
+      CC: gcc-12
+      CXX: g++-12
     steps:
     - uses: actions/checkout@v4
       with:
@@ -54,6 +58,10 @@ jobs:
       run:
         shell: bash
         working-directory: rust
+    env:
+      # Need up-to-date compilers for kernels
+      CC: gcc-12
+      CXX: g++-12
     steps:
     - uses: actions/checkout@v4
       with:

python/Cargo.toml

@@ -31,3 +31,6 @@ pyo3-build-config = { version = "0.20.3", features = [
     "extension-module",
     "abi3-py38",
 ] }
+
+[features]
+fp16kernels = ["lancedb/fp16kernels"]

rust/lancedb/Cargo.toml

@@ -50,3 +50,4 @@ walkdir = "2"
 [features]
 default = ["remote"]
 remote = ["dep:reqwest"]
+fp16kernels = ["lance-linalg/fp16kernels"]