Skip to content

[core] feat: improve variant resolution #420

[core] feat: improve variant resolution

[core] feat: improve variant resolution #420

Workflow file for this run

name: "Build and test kernel"
on:
push:
branches: [main]
pull_request:
branches: [main]
types: [opened, synchronize, reopened] # trigger on PRs
workflow_dispatch:
jobs:
build:
name: Build kernels (${{ matrix.arch }})
strategy:
matrix:
include:
- arch: x86_64-linux
runner: aws-highmemory-32-plus-nix
- arch: aarch64-linux
runner: aws-r8g-8xl-plus-nix
runs-on:
group: ${{ matrix.runner }}
steps:
- uses: actions/checkout@v6
- uses: DeterminateSystems/nix-installer-action@main
with:
extra-conf: |
max-jobs = 4
cores = 12
sandbox-fallback = false
- uses: cachix/cachix-action@v16
with:
name: huggingface
authToken: "${{ secrets.CACHIX_AUTH_TOKEN }}"
env:
USER: runner
- name: Nix info
run: nix-shell -p nix-info --run "nix-info -m"
- name: Build relu kernel
run: ( cd builder/examples/relu && nix build .\#redistributable.torch29-cxx11-cu126-${{ matrix.arch }} )
- name: Copy relu kernel
run: cp -rL builder/examples/relu/result relu-kernel
- name: Build relu-tvm-ffi kernel
run: ( cd builder/examples/relu-tvm-ffi && nix build .\#redistributable.tvm-ffi01-cu126-${{ matrix.arch }} )
- name: Copy relu-tvm-ffi kernel
run: cp -rL builder/examples/relu-tvm-ffi/result relu-tvm-ffi-kernel
- name: Build extra-data kernel
run: ( cd builder/examples/extra-data && nix build .\#redistributable.torch29-cxx11-cu126-${{ matrix.arch }} )
- name: Copy extra-data kernel
run: cp -rL builder/examples/extra-data/result extra-data
- name: Build relu kernel (CPU)
run: ( cd builder/examples/relu && nix build .\#redistributable.torch29-cxx11-cpu-${{ matrix.arch }} )
- name: Copy relu kernel (CPU)
run: cp -rL builder/examples/relu/result relu-kernel-cpu
- name: Build cutlass GEMM kernel
run: ( cd builder/examples/cutlass-gemm && nix build .\#redistributable.torch29-cxx11-cu126-${{ matrix.arch }} )
- name: Copy cutlass GEMM kernel
run: cp -rL builder/examples/cutlass-gemm/result cutlass-gemm-kernel
- name: Build relu-backprop-compile kernel
run: ( cd builder/examples/relu-backprop-compile && nix build .\#redistributable.torch29-cxx11-cu126-${{ matrix.arch }} )
- name: Copy relu-backprop-compile kernel
run: cp -rL builder/examples/relu-backprop-compile/result relu-backprop-compile-kernel
# Just test that we build with the extra torchVersions argument.
- name: Build relu kernel (specific Torch version)
run: ( cd builder/examples/relu-specific-torch && nix build . )
- name: Build relu kernel (compiler flags)
run: ( cd builder/examples/relu-compiler-flags && nix build .\#redistributable.torch29-cxx11-cu126-${{ matrix.arch }} )
- name: Test that we can build a test shell (e.g. that gcc corresponds to CUDA-required)
run: ( cd builder/examples/relu && nix build .#devShells.${{ matrix.arch }}.test )
- name: Build silu-and-mul kernel
run: ( cd builder/examples/silu-and-mul && nix build .\#redistributable.torch-cuda )
- name: Copy silu-and-mul kernel
run: cp -rL builder/examples/silu-and-mul/result silu-and-mul-kernel
- name: Upload kernel artifacts
uses: actions/upload-artifact@v6
with:
name: built-kernels-${{ matrix.arch }}
path: |
activation-kernel
cutlass-gemm-kernel
extra-data
relu-kernel
relu-tvm-ffi-kernel
relu-kernel-cpu
relu-backprop-compile-kernel
silu-and-mul-kernel
test:
name: Test kernels
needs: build
runs-on:
group: aws-g6-12xlarge-plus
steps:
- uses: actions/checkout@v6
- name: Download kernel artifacts
uses: actions/download-artifact@v7
with:
name: built-kernels-x86_64-linux
path: .
- name: Set up Docker Buildx
uses: docker/setup-buildx-action@v3
- name: Build Docker image
uses: docker/build-push-action@v6
with:
context: .
file: builder/tests/Dockerfile.test-kernel
platforms: linux/amd64
load: true
push: false
tags: kernel-builder:latest
- name: Run Tests
run: |
docker run --gpus all kernel-builder:latest