Additional CI runners (#1639)

* Improvements for testing suite

* Add workflow for macOS arm64 CPU tests

* Update tests.yml

* Update tests.yml

Use new L4 and CPU runners for testing.

* Update tests.yml

.github/workflows/tests.yml

@@ -77,14 +77,14 @@ jobs:
         if: startsWith(matrix.os, 'windows')
         uses: ilammy/msvc-dev-cmd@v1.13.0 # to use cl
 
-      # We're running on T4 only for now, so we only target sm75.
+      # Target just the architectures used for our runners.
       - name: Build C++ / CUDA
         run: bash .github/scripts/build-cuda.sh
         env:
           build_os: ${{ matrix.os }}
           build_arch: ${{ matrix.arch }}
           cuda_version: ${{ matrix.cuda_version }}
-          cuda_targets: "75"
+          cuda_targets: "75;89"
 
       - name: Upload build artifact
         uses: actions/upload-artifact@v4
@@ -104,13 +104,14 @@ jobs:
         include:
           - os: ubuntu-22.04
             arch: x86_64
+            runner: banb-aws-general-8-plus-use1-public-80
           - os: ubuntu-22.04-arm
             arch: aarch64
           - os: windows-2025
             arch: x86_64
           - os: macos-15
             arch: arm64
-    runs-on: ${{ matrix.os }}
+    runs-on: ${{ matrix.runner || matrix.os }}
     env:
       BNB_TEST_DEVICE: cpu
     steps:
@@ -148,6 +149,7 @@ jobs:
       matrix:
         os: [ubuntu-22.04, windows-2025]
         arch: [x86_64]
+        gpu: [T4, L4]
         cuda_version: ["11.8.0", "12.8.1"]
         include:
           - cuda_version: "11.8.0"
@@ -156,13 +158,29 @@ jobs:
           - cuda_version: "12.8.1"
             torch_version: "2.7.0"
             pypi_index: "https://download.pytorch.org/whl/cu128"
+
+          # L4 runners
+          - os: ubuntu-22.04
+            gpu: L4
+            runner: bandb-aws-g6-4xlarge-plus-use1-public-80
+
+          # T4 runners
+          - os: ubuntu-22.04
+            gpu: T4
+            runner: CUDA-Linux-x64
+          - os: windows-2025
+            gpu: T4
+            runner: CUDA-Windows-x64
         exclude:
           # Our current T4 Windows runner has a driver too old (471.11)
           # and cannot support CUDA 12+. Skip for now.
           - os: windows-2025
             cuda_version: "12.8.1"
-    runs-on:
-      labels: ${{ contains(matrix.os, 'windows') && 'CUDA-Windows-x64' || 'CUDA-Linux-x64' }}
+
+          # No Windows L4 runners.
+          - os: windows-2025
+            gpu: L4
+    runs-on: ${{ matrix.runner }}
     env:
       BNB_TEST_DEVICE: cuda
     steps: