[IntelNav] ci: install minimal HIP set, avoid rccl (needs kernel driver) #5

Workflow file for this run

.github/workflows/intelnav-release.yml at 06f62ae

	# IntelNav release — builds the IntelNav-patched libllama for each
	# (backend x OS x arch) the Rust FFI crate (`intelnav-ggml`) expects to
	# find under ~/.cache/intelnav/libllama/. Triggered on tag push matching
	# `intelnav-v*` (so upstream ggml-org's own releases don't fire this),
	# and manually via workflow_dispatch for dev iteration.
	#
	# Each job produces:
	# libllama-<backend>-<os>-<arch>.tar.gz
	# containing:
	# bin/libllama.so* (or .dylib / .dll on other OSes)
	# bin/libggml.so*
	# bin/libggml-base.so*
	# bin/libggml-cpu*.so (one or more, per CPU-variant build)
	# bin/libggml-<backend>.so (ROCm / CUDA / Vulkan / SYCL)
	# include/llama.h
	# include/ggml.h
	# include/ggml-*.h
	# LICENSE
	# INTELNAV_SHA (git rev of the fork — the FFI crate
	# reads this to key its cache path)
	#
	# On tag push, all artifacts are uploaded to the matching GitHub Release.
	# On workflow_dispatch, they stay as job artifacts for inspection.
	#
	# Scope for the first cut: Linux x86_64 only, with CPU + Vulkan + ROCm.
	# Each is a free runner and none require proprietary SDKs beyond what
	# `apt` and the ROCm/Vulkan repos provide. CUDA, Metal, SYCL, Windows,
	# and macOS arm64 are follow-ups — deliberately left out so this
	# workflow actually ships and starts producing artifacts this week.

	name: IntelNav release

	on:
	workflow_dispatch:
	inputs:
	make_release:
	description: "Upload artifacts to a GitHub Release (requires a tag)."
	type: boolean
	default: false
	push:
	tags:
	- 'intelnav-v*'

	concurrency:
	group: ${{ github.workflow }}-${{ github.head_ref && github.ref \|\| github.run_id }}
	cancel-in-progress: true

	env:
	# Common cmake args across all backends. Keep tools/server/examples
	# OFF — consumers of intelnav-ggml link libllama + libggml* as a
	# library and don't need the CLI tools.
	CMAKE_COMMON: >-
	-DLLAMA_BUILD_EXAMPLES=OFF
	-DLLAMA_BUILD_TESTS=OFF
	-DLLAMA_BUILD_TOOLS=OFF
	-DLLAMA_BUILD_SERVER=OFF
	-DCMAKE_BUILD_TYPE=Release
	-DBUILD_SHARED_LIBS=ON
	# ccache key suffix used by every job. Bump to bust.
	CCACHE_GEN: v1

	jobs:
	# ---------------------------------------------------------------------
	# linux-x64 CPU — universal fallback. Ships with GGML_CPU_ALL_VARIANTS
	# so runtime dispatch picks SSE / AVX2 / AVX-512 as available.
	# ---------------------------------------------------------------------
	linux-x64-cpu:
	runs-on: ubuntu-22.04
	steps:
	- uses: actions/checkout@v6
	with: { fetch-depth: 0 }

	- uses: ggml-org/ccache-action@v1.2.21
	with:
	key: intelnav-linux-x64-cpu-${{ env.CCACHE_GEN }}
	evict-old-files: 1d

	- name: Configure
	run: \|
	cmake -B build \
	${{ env.CMAKE_COMMON }} \
	-DGGML_NATIVE=OFF \
	-DGGML_CPU_ALL_VARIANTS=ON \
	-DGGML_BACKEND_DL=ON

	- name: Build
	run: cmake --build build -j $(nproc) --target llama

	- name: Package
	run: .github/actions/intelnav-pack.sh linux-x64-cpu
	shell: bash

	- uses: actions/upload-artifact@v6
	with:
	name: libllama-linux-x64-cpu
	path: dist/*.tar.gz

	# ---------------------------------------------------------------------
	# linux-x64 Vulkan — the "any GPU on Linux" backend. Covers AMD
	# (without ROCm), Intel Arc, NVIDIA (without CUDA), and integrated
	# graphics on every modern distro.
	# ---------------------------------------------------------------------
	linux-x64-vulkan:
	runs-on: ubuntu-22.04
	steps:
	- uses: actions/checkout@v6
	with: { fetch-depth: 0 }

	- name: Install Vulkan SDK prereqs
	run: \|
	sudo apt-get update
	sudo apt-get install -y --no-install-recommends \
	build-essential curl tar xz-utils libxcb-xinput0 libxcb-xinerama0 libxcb-cursor-dev libssl-dev

	# Pull the LunarG SDK rather than scraping apt packages that move
	# around between Ubuntu LTSes. Matches upstream `build-vulkan.yml`.
	- name: Install LunarG Vulkan SDK
	run: \|
	set -e
	ver="$(curl -fsSL https://vulkan.lunarg.com/sdk/latest/linux.txt)"
	echo "Vulkan SDK version: $ver"
	curl -fsSL "https://sdk.lunarg.com/sdk/download/${ver}/linux/vulkan_sdk.tar.xz" -o /tmp/vulkan.tar.xz
	mkdir -p "$HOME/vulkan_sdk"
	tar -xJf /tmp/vulkan.tar.xz -C "$HOME/vulkan_sdk" --strip-components=1
	echo "VULKAN_SDK=$HOME/vulkan_sdk/x86_64" >> "$GITHUB_ENV"
	echo "$HOME/vulkan_sdk/x86_64/bin" >> "$GITHUB_PATH"
	echo "LD_LIBRARY_PATH=$HOME/vulkan_sdk/x86_64/lib${LD_LIBRARY_PATH:+:$LD_LIBRARY_PATH}" >> "$GITHUB_ENV"

	- uses: ggml-org/ccache-action@v1.2.21
	with:
	key: intelnav-linux-x64-vulkan-${{ env.CCACHE_GEN }}
	evict-old-files: 1d

	- name: Configure
	run: \|
	cmake -B build \
	${{ env.CMAKE_COMMON }} \
	-DGGML_NATIVE=OFF \
	-DGGML_VULKAN=ON \
	-DGGML_BACKEND_DL=ON

	- name: Build
	run: cmake --build build -j $(nproc) --target llama

	- name: Package
	run: .github/actions/intelnav-pack.sh linux-x64-vulkan
	shell: bash

	- uses: actions/upload-artifact@v6
	with:
	name: libllama-linux-x64-vulkan
	path: dist/*.tar.gz

	# ---------------------------------------------------------------------
	# linux-x64 ROCm — AMD Radeon (gfx9+) acceleration, the headline
	# IntelNav win. The user's RX 6600 (gfx1032) is covered by setting
	# AMDGPU_TARGETS to include gfx1030 (handles gfx1031/1032 via override).
	# ---------------------------------------------------------------------
	linux-x64-rocm:
	runs-on: ubuntu-22.04
	steps:
	- uses: actions/checkout@v6
	with: { fetch-depth: 0 }

	- name: Install ROCm
	run: \|
	# The GH jammy runner ships pre-existing rocm-* bits from the
	# Ubuntu archive (e.g. rocm-cmake 5.0) which conflict with
	# Radeon's own 6.x packages. Purge first, then pin Radeon's
	# repo with apt preferences so the full 6.x stack resolves
	# as a single coherent set.
	sudo apt-get update
	sudo apt-get install -y --no-install-recommends wget gnupg
	sudo apt-get purge -y 'rocm-' 'hip-' 'miopen-' 'hsa-' 'hipblas' 'rocblas' 'rocfft*' \|\| true
	sudo apt-get autoremove -y \|\| true

	sudo mkdir -p /etc/apt/keyrings
	wget -qO - https://repo.radeon.com/rocm/rocm.gpg.key \
	\| sudo gpg --dearmor -o /etc/apt/keyrings/rocm.gpg
	# ROCm 6.3+ is required by ggml-cuda's vendor header — it
	# references `__hip_fp8_e4m3` which older HIP runtimes don't
	# define. Pin to a current point release on the 6.4 line.
	echo "deb [arch=amd64 signed-by=/etc/apt/keyrings/rocm.gpg] https://repo.radeon.com/rocm/apt/6.4 jammy main" \
	\| sudo tee /etc/apt/sources.list.d/rocm.list
	printf 'Package: *\nPin: origin "repo.radeon.com"\nPin-Priority: 600\n' \
	\| sudo tee /etc/apt/preferences.d/rocm-pin-600 >/dev/null

	# Install a minimal HIP + BLAS set rather than the `rocm-dev`
	# / `rocm-libs` metas — those pull in `rccl`, which 6.4
	# hard-depends on AMD kernel driver libs (libdrm-amdgpu-*)
	# that don't exist on GH-hosted runners. Our build only
	# needs hipcc + HIP headers + hipBLAS + rocBLAS.
	sudo apt-get update
	sudo apt-get install -y --no-install-recommends \
	hipcc hip-dev rocm-cmake rocm-device-libs \
	hipblas-dev rocblas-dev comgr hsa-rocr-dev \
	rocminfo rocm-llvm
	echo "/opt/rocm/bin" >> $GITHUB_PATH
	echo "/opt/rocm/llvm/bin" >> $GITHUB_PATH

	- uses: ggml-org/ccache-action@v1.2.21
	with:
	key: intelnav-linux-x64-rocm-${{ env.CCACHE_GEN }}
	evict-old-files: 1d

	- name: Configure
	env:
	# gfx1030 bytecode is binary-compatible with gfx1031/1032 via
	# HSA_OVERRIDE_GFX_VERSION=10.3.0 at run time — one artifact
	# covers the RX 6600 / 6700 / 6800 family. Add more archs
	# here as we get hardware to test on.
	AMDGPU_TARGETS: "gfx900;gfx906;gfx908;gfx90a;gfx1030;gfx1100"
	run: \|
	cmake -B build \
	${{ env.CMAKE_COMMON }} \
	-DGGML_NATIVE=OFF \
	-DGGML_HIP=ON \
	-DCMAKE_C_COMPILER=/opt/rocm/llvm/bin/clang \
	-DCMAKE_CXX_COMPILER=/opt/rocm/llvm/bin/clang++ \
	-DAMDGPU_TARGETS=$AMDGPU_TARGETS \
	-DGGML_BACKEND_DL=ON

	- name: Build
	run: cmake --build build -j $(nproc) --target llama

	- name: Package
	run: .github/actions/intelnav-pack.sh linux-x64-rocm
	shell: bash

	- uses: actions/upload-artifact@v6
	with:
	name: libllama-linux-x64-rocm
	path: dist/*.tar.gz

	# ---------------------------------------------------------------------
	# linux-x64 CUDA — NVIDIA acceleration. Uses Jimver's cuda-toolkit
	# action to install a pinned CUDA version on the hosted runner
	# (~3 GB download cached per runner). Test runs against this
	# artifact require an NVIDIA GPU, so the intelnav repo exercises it
	# only on self-hosted `[gpu, nvidia]` runners — but we publish the
	# build unconditionally so those runners have something to pull.
	# ---------------------------------------------------------------------
	linux-x64-cuda:
	runs-on: ubuntu-22.04
	steps:
	- uses: actions/checkout@v6
	with: { fetch-depth: 0 }

	# Install CUDA 12.6 from NVIDIA's own apt repo directly — the
	# Jimver action's sub-package name for cublas drifted between
	# CUDA 11 (cuda-cublas-) and CUDA 12 (libcublas-), and pinning
	# the real package names keeps us independent of that.
	- name: Install CUDA 12.6
	run: \|
	set -e
	sudo apt-get update
	sudo apt-get install -y --no-install-recommends ca-certificates gnupg curl
	sudo mkdir -p /etc/apt/keyrings
	curl -fsSL https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/x86_64/3bf863cc.pub \
	\| sudo gpg --dearmor -o /etc/apt/keyrings/nvidia-cuda.gpg
	echo "deb [arch=amd64 signed-by=/etc/apt/keyrings/nvidia-cuda.gpg] https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/x86_64/ /" \
	\| sudo tee /etc/apt/sources.list.d/nvidia-cuda.list
	sudo apt-get update
	sudo apt-get install -y --no-install-recommends \
	cuda-nvcc-12-6 cuda-cudart-dev-12-6 libcublas-dev-12-6 libcublas-12-6
	echo "/usr/local/cuda-12.6/bin" >> $GITHUB_PATH
	echo "CUDA_PATH=/usr/local/cuda-12.6" >> $GITHUB_ENV

	- uses: ggml-org/ccache-action@v1.2.21
	with:
	key: intelnav-linux-x64-cuda-${{ env.CCACHE_GEN }}
	evict-old-files: 1d

	- name: Configure
	run: \|
	# Narrow arch list keeps the CUDA build under the hosted
	# runner's 6-hour ceiling and within practical ccache reuse.
	# 86 covers RTX 30-series / A100-like; 89 covers RTX 40.
	# Users with older (Turing sm_75) or newer (Hopper sm_90)
	# hardware rebuild from source locally — add them back when
	# there's real CI demand.
	cmake -B build \
	${{ env.CMAKE_COMMON }} \
	-DGGML_NATIVE=OFF \
	-DGGML_CUDA=ON \
	-DCMAKE_CUDA_ARCHITECTURES="86;89" \
	-DGGML_BACKEND_DL=ON

	- name: Build
	run: cmake --build build -j $(nproc) --target llama

	- name: Package
	run: .github/actions/intelnav-pack.sh linux-x64-cuda
	shell: bash

	- uses: actions/upload-artifact@v6
	with:
	name: libllama-linux-x64-cuda
	path: dist/*.tar.gz

	# ---------------------------------------------------------------------
	# macOS arm64 Metal — Apple Silicon via Metal Performance Shaders.
	# macos-14 is the first free runner with an M1; anything older is
	# Intel and we'd build a different arch. Metal is on by default in
	# mainline llama.cpp; we just need to keep it explicit.
	# ---------------------------------------------------------------------
	macos-arm64-metal:
	runs-on: macos-14
	steps:
	- uses: actions/checkout@v6
	with: { fetch-depth: 0 }

	- uses: ggml-org/ccache-action@v1.2.21
	with:
	key: intelnav-macos-arm64-metal-${{ env.CCACHE_GEN }}
	evict-old-files: 1d

	- name: Configure
	run: \|
	cmake -B build \
	${{ env.CMAKE_COMMON }} \
	-DGGML_NATIVE=OFF \
	-DGGML_METAL=ON \
	-DGGML_METAL_EMBED_LIBRARY=ON \
	-DGGML_BACKEND_DL=ON

	- name: Build
	run: cmake --build build -j $(sysctl -n hw.ncpu) --target llama

	- name: Package
	run: .github/actions/intelnav-pack.sh macos-arm64-metal
	shell: bash

	- uses: actions/upload-artifact@v6
	with:
	name: libllama-macos-arm64-metal
	path: dist/*.tar.gz

	# ---------------------------------------------------------------------
	# Upload to Release. Only runs on tag push or when
	# workflow_dispatch.inputs.make_release is true.
	# ---------------------------------------------------------------------
	publish-release:
	needs: [linux-x64-cpu, linux-x64-vulkan, linux-x64-rocm, linux-x64-cuda, macos-arm64-metal]
	if: startsWith(github.ref, 'refs/tags/intelnav-v') \|\| inputs.make_release
	runs-on: ubuntu-22.04
	# `attestations: write` is required for GitHub's built-in
	# actions/attest-build-provenance to sign the artifacts with a
	# keyless sigstore identity scoped to this workflow run. Consumers
	# — and specifically task #14's runtime downloader — verify the
	# attestation before caching and dlopening the blob, closing the
	# supply-chain RCE vector flagged in docs/dev/M0_AUDIT.md S1.
	permissions:
	contents: write
	attestations: write
	id-token: write
	steps:
	- uses: actions/checkout@v6

	- name: Download artifacts
	uses: actions/download-artifact@v6
	with:
	path: artifacts

	- name: Collect artifacts
	id: collect
	run: \|
	mkdir -p dist
	find artifacts -name '*.tar.gz' -exec cp {} dist/ \;
	ls -la dist/
	# Emit a newline-separated list for the attestor below.
	{
	echo "paths<<EOF"
	ls dist/*.tar.gz
	echo "EOF"
	} >> "$GITHUB_OUTPUT"

	- name: Generate build-provenance attestation
	uses: actions/attest-build-provenance@v3
	with:
	subject-path: ${{ steps.collect.outputs.paths }}

	- name: Publish
	uses: softprops/action-gh-release@v2
	with:
	files: dist/*.tar.gz
	fail_on_unmatched_files: true
	generate_release_notes: true

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[IntelNav] ci: install minimal HIP set, avoid rccl (needs kernel driver) #5

Workflow file

[IntelNav] ci: install minimal HIP set, avoid rccl (needs kernel driver) #5

Uh oh!

Workflow file for this run