CI: retry docker pulls in workflow image downloads

.github/scripts/docker_pull_with_retry.sh

@@ -0,0 +1,28 @@
+#!/bin/bash
+
+set -euo pipefail
+
+if [ "$#" -ne 1 ]; then
+  echo "Usage: $0 <image>" >&2
+  exit 2
+fi
+
+IMAGE="$1"
+MAX_ATTEMPTS="${DOCKER_PULL_MAX_ATTEMPTS:-3}"
+RETRY_DELAY_SECONDS="${DOCKER_PULL_RETRY_DELAY_SECONDS:-10}"
+
+for attempt in $(seq 1 "${MAX_ATTEMPTS}"); do
+  echo "Pulling Docker image '${IMAGE}' (attempt ${attempt}/${MAX_ATTEMPTS})"
+  if docker pull "${IMAGE}"; then
+    echo "Docker pull succeeded for '${IMAGE}' on attempt ${attempt}"
+    exit 0
+  fi
+
+  if [ "${attempt}" -lt "${MAX_ATTEMPTS}" ]; then
+    echo "Docker pull failed for '${IMAGE}' on attempt ${attempt}; retrying in ${RETRY_DELAY_SECONDS}s..."
+    sleep "${RETRY_DELAY_SECONDS}"
+  fi
+done
+
+echo "Docker pull failed for '${IMAGE}' after ${MAX_ATTEMPTS} attempts" >&2
+exit 1

.github/workflows/atom-test.yaml

@@ -21,6 +21,8 @@ env:
   ATOM_BRANCH: "main"
   ATOM_REPOSITORY_URL: "ROCm/ATOM"
   BASE_IMAGE: "rocm/atom-dev:latest"
+  DOCKER_PULL_MAX_ATTEMPTS: "3"
+  DOCKER_PULL_RETRY_DELAY_SECONDS: "10"
   GITHUB_REPO_URL: ${{ github.event.pull_request.head.repo.clone_url || 'https://github.com/ROCm/Aiter.git' }}
   GITHUB_COMMIT_SHA: ${{ github.event.pull_request.head.sha || github.event.head_commit.id }}
 
@@ -95,7 +97,18 @@ jobs:
       - name: Download the ATOM base image
         if: matrix.run_on_pr == true || github.event_name != 'pull_request'
         run: |
-          docker pull ${{ env.BASE_IMAGE }}
+          for attempt in $(seq 1 "${DOCKER_PULL_MAX_ATTEMPTS}"); do
+            if docker pull "${{ env.BASE_IMAGE }}"; then
+              echo "Docker pull succeeded on attempt ${attempt}"
+              exit 0
+            fi
+            echo "Docker pull attempt ${attempt} failed"
+            if [ "${attempt}" -lt "${DOCKER_PULL_MAX_ATTEMPTS}" ]; then
+              sleep "${DOCKER_PULL_RETRY_DELAY_SECONDS}"
+            fi
+          done
+          echo "Docker pull failed after ${DOCKER_PULL_MAX_ATTEMPTS} attempts"
+          exit 1
 
       - name: Generate Dockerfile
         if: matrix.run_on_pr == true || github.event_name != 'pull_request'

.github/workflows/flash_attention_integration.yaml

@@ -34,6 +34,8 @@ env:
   FA_REPOSITORY_URL: https://github.com/Dao-AILab/flash-attention.git
   BASE_IMAGE: rocm/pytorch:latest@sha256:683765a52c61341e1674fe730ab3be861a444a45a36c0a8caae7653a08a0e208
   AITER_SUBMODULE_PATH: third_party/aiter
+  DOCKER_PULL_MAX_ATTEMPTS: "3"
+  DOCKER_PULL_RETRY_DELAY_SECONDS: "10"
 
 jobs:
   check-signal:
@@ -133,7 +135,7 @@ jobs:
           exit 0
 
       - name: Pull base image
-        run: docker pull ${{ env.BASE_IMAGE }}
+        run: bash .github/scripts/docker_pull_with_retry.sh "${{ env.BASE_IMAGE }}"
 
       - name: Generate Dockerfile
         run: |
@@ -341,7 +343,7 @@ jobs:
           exit 0
 
       - name: Pull base image
-        run: docker pull ${{ env.BASE_IMAGE }}
+        run: bash .github/scripts/docker_pull_with_retry.sh "${{ env.BASE_IMAGE }}"
 
       - name: Generate Dockerfile
         run: |

.github/workflows/vllm_benchmark.yaml

@@ -16,6 +16,8 @@ env:
   VLLM_BRANCH: "main"
   VLLM_REPOSITORY_URL: "https://github.com/vllm-project/vllm"
   BASE_IMAGE: rocm/vllm-dev:nightly
+  DOCKER_PULL_MAX_ATTEMPTS: "3"
+  DOCKER_PULL_RETRY_DELAY_SECONDS: "10"
   AITER_WHEEL_ARTIFACT_NAME: aiter-vllm-whl-${{ github.run_id }}
   GITHUB_REPO_URL: ${{ github.event.pull_request.head.repo.clone_url || 'https://github.com/ROCm/aiter.git' }}
   GITHUB_COMMIT_SHA: ${{ github.event.pull_request.head.sha || github.sha }}
@@ -81,7 +83,7 @@ jobs:
       - name: Build Aiter wheel in vLLM base image
         run: |
           set -euo pipefail
-          docker pull ${{ env.BASE_IMAGE }}
+          bash .github/scripts/docker_pull_with_retry.sh "${{ env.BASE_IMAGE }}"
           docker run --rm \
             --network=host \
             -v "${{ github.workspace }}:/workspace" \
@@ -144,6 +146,14 @@ jobs:
             extra_env_args: -e VLLM_USE_TRITON_FLASH_ATTN=0
 
     steps:
+      - name: Checkout Docker pull retry helper
+        uses: actions/checkout@v6
+        with:
+          repository: ${{ github.event.pull_request.head.repo.full_name || github.repository }}
+          ref: ${{ github.event.pull_request.head.sha || github.sha }}
+          sparse-checkout: |
+            .github/scripts
+
       - name: Download Aiter wheel artifact
         uses: actions/download-artifact@v4
         with:
@@ -170,7 +180,7 @@ jobs:
 
       - name: Download the vLLM base image
         run: |
-          docker pull ${{ env.BASE_IMAGE }}
+          bash .github/scripts/docker_pull_with_retry.sh "${{ env.BASE_IMAGE }}"
 
       - name: Run benchmarks
         run: |