Add A10g workflow to gain access to A10G GPU

.ci/torchbench/check-ssh.sh

@@ -0,0 +1,12 @@
+#!/bin/bash
+set -eou pipefail
+
+echo "Holding runner for 2 hours until all ssh sessions have logged out"
+for _ in $(seq 1440); do
+    # Break if no ssh session exists anymore
+    if [ "$(who)" = "" ]; then
+      break
+    fi
+    echo "."
+    sleep 5
+done

.github/workflows/linux-test-a10g.yml

@@ -0,0 +1,43 @@
+name: TorchBench PR Test on A10G
+on:
+  workflow_dispatch:
+  pull_request:
+
+jobs:
+  linux-test-a10g:
+    # Don't run on forked repos
+    # Only run on PR labeled 'with-ssh'
+    if: github.repository_owner == 'pytorch' && contains(github.event.pull_request.labels.*.name, 'with-ssh')
+    runs-on: linux.g5.4xlarge.nvidia.gpu
+    timeout-minutes: 240
+    environment: docker-s3-upload
+    env:
+      CONDA_ENV: "pr-test-cuda"
+      TEST_CONFIG: "cuda"
+      HUGGING_FACE_HUB_TOKEN: ${{ secrets.HUGGING_FACE_HUB_TOKEN }}
+    steps:
+      - name: Checkout TorchBench
+        uses: actions/checkout@v3
+      - name: Setup SSH (Click me for login details)
+        uses: pytorch/test-infra/.github/actions/setup-ssh@main
+        with:
+          github-secret: ${{ secrets.TORCHBENCH_ACCESS_TOKEN }}
+      - name: Install Conda
+        run: |
+          bash ./.ci/torchbench/install-conda.sh
+      - name: Install TorchBench
+        run: |
+          bash ./.ci/torchbench/install.sh
+      - name: Wait for SSH session to end
+        if: always()
+        run: |
+          bash ./.ci/torchbench/check-ssh.sh
+      - name: Clean up Conda env
+        if: always()
+        run: |
+          . ${HOME}/miniconda3/etc/profile.d/conda.sh
+          conda remove -n "${CONDA_ENV}" --all
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.sha }}-${{ github.event_name == 'workflow_dispatch' }}
+  cancel-in-progress: true