From 90a478bd4d97683cd18643cee35e0521fd42dc28 Mon Sep 17 00:00:00 2001
From: Olivier Chafik <ochafik@google.com>
Date: Wed, 8 Jan 2025 17:50:26 +0000
Subject: [PATCH] Package linux cuda releases for various caps

---
 .github/workflows/build.yml | 85 ++++++++++++++++++++++++++++++++++++-
 ggml/src/kompute            |  1 +
 2 files changed, 84 insertions(+), 2 deletions(-)
 create mode 160000 ggml/src/kompute

diff --git a/.github/workflows/build.yml b/.github/workflows/build.yml
index a377eff38..df9abf304 100644
--- a/.github/workflows/build.yml
+++ b/.github/workflows/build.yml
@@ -809,7 +809,52 @@ jobs:
 
   ubuntu-latest-cmake-cuda:
     runs-on: ubuntu-latest
-    container: nvidia/cuda:12.6.2-devel-ubuntu24.04
+
+    strategy:
+      matrix:
+        cuda:
+          # Colab and lightning.ai currently use CUDA 12.2 (test w/ `nvidia-smi | grep "CUDA Version: "`)
+          # Capabilities of GPUs are listed on https://developer.nvidia.com/cuda-gpus, can test w/ `nvidia-smi --query-gpu=compute_cap --format=csv`
+          # See available containers at https://hub.docker.com/r/nvidia/cuda/tags
+          - version: 12.2
+            container: nvidia/cuda:12.2.2-devel-ubuntu24.04
+            cap: 7.5
+            arch: 75-real
+            example: 'T4'
+            package: true
+          - version: 12.2
+            container: nvidia/cuda:12.2.2-devel-ubuntu24.04
+            cap: 8.0
+            arch: 80-real
+            example: 'A100'
+            package: true
+          - version: 12.2
+            container: nvidia/cuda:12.2.2-devel-ubuntu24.04
+            cap: 8.6
+            arch: 86-real
+            example: 'A10'
+            package: true
+          - version: 12.2
+            container: nvidia/cuda:12.2.2-devel-ubuntu24.04
+            cap: 8.9
+            arch: 89-real
+            example: 'L4, L40S'
+          - version: 12.2
+            container: nvidia/cuda:12.2.2-devel-ubuntu24.04
+            cap: 9.0
+            arch: 90-real
+            example: 'H100'
+            package: true
+          # Build only, don't package.
+          - version: 12.6
+            container: nvidia/cuda:12.6.2-devel-ubuntu24.04
+            cap: 8.9
+            arch: 89-real
+            package: false
+
+    container: nvidia/cuda:${{ matrix.cuda.version }}-devel-ubuntu24.04
+
+    name: Linux CUDA ${{ matrix.cuda.version }} Cap ${{ matrix.cuda.cap }} (e.g. ${{ matrix.cuda_cap.example_gpu }})
 
     steps:
         - name: Clone
@@ -825,9 +870,45 @@ jobs:
 
         - name: Build with CMake
           run: |
-            cmake -S . -B build -G Ninja -DCMAKE_BUILD_TYPE=Release -DGGML_NATIVE=OFF -DGGML_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES=89-real -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined -DLLAMA_FATAL_WARNINGS=ON
+            cmake -S . -B build -G Ninja \
+              -DCMAKE_BUILD_TYPE=Release \
+              -DGGML_NATIVE=OFF \
+              -DGGML_CUDA=ON \
+              -DCMAKE_CUDA_ARCHITECTURES=${{ matrix.cuda.arch }} \
+              -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined \
+              -DLLAMA_CURL=ON \
+              -DLLAMA_FATAL_WARNINGS=ON
             cmake --build build
 
+        - name: Determine tag name
+          if: ${{ matrix.cuda.package }}
+          id: tag
+          shell: bash
+          run: |
+            BUILD_NUMBER="$(git rev-list --count HEAD)"
+            SHORT_HASH="$(git rev-parse --short=7 HEAD)"
+            if [[ "${{ env.BRANCH_NAME }}" == "master" ]]; then
+              echo "name=b${BUILD_NUMBER}" >> $GITHUB_OUTPUT
+            else
+              SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-')
+              echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT
+            fi
+            echo "cuda_name=cuda-${{ matrix.cuda.short_version }}-cap-${{ matrix.cuda.cap }}" >> $GITHUB_OUTPUT
+
+        - name: Pack artifacts
+          id: pack_artifacts
+          if: ${{ matrix.cuda.package && (( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true') }}
+          run: |
+            cp LICENSE ./build/bin/
+            zip -r llama-${{ steps.tag.outputs.name }}-bin-ubuntu-x64-${{ steps.tag.outputs.cuda_name }}.zip ./build/bin/*
+
+        - name: Upload artifacts
+          if: ${{ matrix.cuda.package && (( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true') }}
+          uses: actions/upload-artifact@v4
+          with:
+            path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-x64-${{ steps.tag.outputs.cuda_name }}.zip
+            name: llama-bin-ubuntu-x64.zip
+
   windows-2019-cmake-cuda:
     runs-on: windows-2019
 
diff --git a/ggml/src/kompute b/ggml/src/kompute
new file mode 160000
index 000000000..4565194ed
--- /dev/null
+++ b/ggml/src/kompute
@@ -0,0 +1 @@
+Subproject commit 4565194ed7c32d1d2efa32ceab4d3c6cae006306