ci: refactor self-hosted runner configuration to use JSON format

- Convert runs-on array syntax to single runner value with fromjson()
- Update runner parameter defaults to JSON string format
- Add runner labels to build workflow calls
- Update test workflow runner configurations
- Remove unused test-utils.yml workflow

.github/workflows/_build_linux_arm_self_host.yml

@@ -42,9 +42,7 @@ on:
 jobs:
   build_wheels_self_hosted:
     name: Build wheels and Upload (Linux x86_64, self-hosted runner)
-    runs-on:
-      - ${{ inputs.runner }}
-      - arm64
+    runs-on: ${{ inputs.runner }}
     container:
       image: ${{ inputs.container-image }}
       options: --platform linux/arm64

.github/workflows/_build_linux_self_host.yml

@@ -27,7 +27,7 @@ on:
         description: "Runner type"
         required: false
         type: string
-        default: "self-hosted"
+        default: '["self-hosted"]'
       is-upload:
         description: "Whether to upload the release asset"
         required: false
@@ -48,9 +48,7 @@ jobs:
   build_wheels_self_hosted:
     if: ${{ inputs.use-container }}
     name: Build wheels and Upload (Linux x86_64, self-hosted runner)
-    runs-on:
-      - ${{ inputs.runner }}
-      - x64
+    runs-on: ${{ fromjson(inputs.runner) }}
     container:
       image: ${{ inputs.container-image }}
     defaults:
@@ -187,9 +185,7 @@ jobs:
   build_wheels_self_hosted_no_container:
     if: ${{ !inputs.use-container }}
     name: Build wheels and Upload (Linux x86_64, self-hosted runner)
-    runs-on:
-      - ${{ inputs.runner }}
-      - x64
+    runs-on: ${{ fromjson(inputs.runner) }}
     defaults:
       run:
         shell: bash

.github/workflows/build.yml

@@ -94,6 +94,7 @@ jobs:
       python-version: ${{ matrix.python-version }}
       torch-version: ${{ matrix.torch-version }}
       cuda-version: ${{ matrix.cuda-version }}
+      runner: '["self-hosted", "x64"]'
     secrets: inherit
 
   build_wheels_linux_arm64_self_hosted:
@@ -114,6 +115,7 @@ jobs:
       python-version: ${{ matrix.python-version }}
       torch-version: ${{ matrix.torch-version }}
       cuda-version: ${{ matrix.cuda-version }}
+      runner: '["self-hosted-arm", "arm64"]'
     secrets: inherit
 
   # #########################################################

.github/workflows/test-arm-self-hosted.yml

@@ -18,7 +18,7 @@ jobs:
         torch-version: ["2.9.1"]
         # https://developer.nvidia.com/cuda-toolkit-archive
         cuda-version: ["13.0.1"]
-        runner: ["self-hosted"]
+        runner: ['["self-hosted-arm"]']
     with:
       flash-attn-version: ${{ matrix.flash-attn-version }}
       python-version: ${{ matrix.python-version }}

.github/workflows/test-self-hosted-linux.yml

@@ -18,7 +18,7 @@ jobs:
         torch-version: ["2.9.1"]
         # https://developer.nvidia.com/cuda-toolkit-archive
         cuda-version: ["13.0.1"]
-        runner: ["openci-runner-beta"]
+        runner: ['["openci-runner-beta"]']
         use-container: [false]
     with:
       flash-attn-version: ${{ matrix.flash-attn-version }}

.github/workflows/test-utils.yml

@@ -1,72 +0,0 @@
-name: Test runner environment
-
-on:
-  workflow_dispatch:
-
-jobs:
-  create_releases:
-    name: Create Releases
-    strategy:
-      fail-fast: false
-      matrix:
-        python-version: ["3.13", "3.12"]
-        runner: ["ubuntu-22.04-arm", "ubuntu-22.04", "self-hosted"]
-    runs-on: ${{ matrix.runner }}
-    steps:
-      - name: Check environment
-        shell: bash
-        run: |
-          cat /etc/os-release || true
-          echo "-----------------------------"
-          cat /etc/lsb-release || true
-          echo "-----------------------------"
-          lscpu || true
-          echo "-----------------------------"
-          df -h || true
-          echo "-----------------------------"
-          free -h || true
-
-      - name: Determine MAX_JOBS and NVCC_THREADS
-        shell: bash
-        run: |
-          # Determine MAX_JOBS and NVCC_THREADS based on system resources
-          NUM_THREADS=$(nproc)
-          RAM_GB=$(free -g | awk '/^Mem:/{print $2}')
-          echo "System resources:"
-          echo "  CPU threads: $NUM_THREADS"
-          echo "  RAM: ${RAM_GB}GB"
-
-          # Determine MAX_JOBS and NVCC_THREADS based on system resources
-          if [[ -z "${MAX_JOBS:-}" && -z "${NVCC_THREADS:-}" ]]; then
-            # Calculate max product based on following constraints:
-            # - MAX_JOBS x NVCC_THREADS(<= 4) <= NUM_THREADS
-            # - 2.5GB x MAX_JOBS x NVCC_THREADS(<= 4) <= RAM_GB
-
-            # Set MAX_PRODUCT from RAM
-            MAX_PRODUCT_CPU=$NUM_THREADS
-            MAX_PRODUCT_RAM=$(awk -v ram="$RAM_GB" 'BEGIN {print int(ram / 2.5)}')
-            MAX_PRODUCT=$((MAX_PRODUCT_CPU < MAX_PRODUCT_RAM ? MAX_PRODUCT_CPU : MAX_PRODUCT_RAM))
-
-            # Set MAX_JOBS and NVCC_THREADS so that MAX_JOBS x NVCC_THREADS ≈ MAX_PRODUCT with NVCC_THREADS <= 4
-            BASE_THREADS=$(awk -v max="$MAX_PRODUCT" 'BEGIN {print int(sqrt(max))}')
-
-            if awk "BEGIN {exit !($RAM_GB <= 16)}"; then
-              # If RAM is less than 16GB, set NVCC_THREADS to 1 and MAX_JOBS to 2
-              NVCC_THREADS=1
-              MAX_JOBS=2
-            elif (( BASE_THREADS <= 4 )); then
-              NVCC_THREADS=$BASE_THREADS
-              MAX_JOBS=$BASE_THREADS
-            else
-              NVCC_THREADS=4
-              MAX_JOBS=$((MAX_PRODUCT / NVCC_THREADS))
-            fi
-
-            # Ensure minimum values of 1
-            MAX_JOBS=$((MAX_JOBS < 1 ? 1 : MAX_JOBS))
-            NVCC_THREADS=$((NVCC_THREADS < 1 ? 1 : NVCC_THREADS))
-          fi
-
-          echo "Build parallelism settings:"
-          echo "  MAX_JOBS: $MAX_JOBS"
-          echo "  NVCC_THREADS: $NVCC_THREADS"