CI workflow for Flash Attn (#41857)

ci for flash attn

Co-authored-by: ydshieh <[email protected]>

Author: ydshieh

.github/workflows/model_jobs.yml

@@ -28,6 +28,9 @@ on:
       report_repo_id:
         required: false
         type: string
+      pytest_marker:
+        required: false
+        type: string
 
 env:
   HF_HOME: /mnt/cache
@@ -137,7 +140,7 @@ jobs:
       - name: Run all tests on GPU
         working-directory: /transformers
         run: |
-          script -q -c "PATCH_TESTING_METHODS_TO_COLLECT_OUTPUTS=yes _PATCHED_TESTING_METHODS_OUTPUT_DIR=/transformers/reports/${{ env.machine_type }}_${{ inputs.report_name_prefix }}_${{ env.matrix_folders }}_test_reports python3 -m pytest -rsfE -v --make-reports=${{ env.machine_type }}_${{ inputs.report_name_prefix }}_${{ env.matrix_folders }}_test_reports tests/${{ matrix.folders }}" test_outputs.txt
+          script -q -c "PATCH_TESTING_METHODS_TO_COLLECT_OUTPUTS=yes _PATCHED_TESTING_METHODS_OUTPUT_DIR=/transformers/reports/${{ env.machine_type }}_${{ inputs.report_name_prefix }}_${{ env.matrix_folders }}_test_reports python3 -m pytest -rsfE -v -m '${{ inputs.pytest_marker }}' --make-reports=${{ env.machine_type }}_${{ inputs.report_name_prefix }}_${{ env.matrix_folders }}_test_reports tests/${{ matrix.folders }}" test_outputs.txt
           ls -la
           # Extract the exit code from the output file
           EXIT_CODE=$(tail -1 test_outputs.txt | grep -o 'COMMAND_EXIT_CODE="[0-9]*"' | cut -d'"' -f2)

.github/workflows/self-scheduled-flash-attn-caller.yml

@@ -0,0 +1,60 @@
+name: Nvidia CI - Flash Attn
+
+on:
+  repository_dispatch:
+  schedule:
+    - cron: "17 2 * * *"
+  push:
+    branches:
+      - run_nvidia_ci_flash_attn*
+  workflow_dispatch:
+    inputs:
+      prev_workflow_run_id:
+        description: 'previous workflow run id to compare'
+        type: string
+        required: false
+        default: ""
+      other_workflow_run_id:
+        description: 'other workflow run id to compare'
+        type: string
+        required: false
+        default: ""
+
+
+# Used for `push` to easily modify the target workflow runs to compare against
+env:
+    prev_workflow_run_id: ""
+    other_workflow_run_id: ""
+
+
+jobs:
+  setup:
+    name: Setup
+    runs-on: ubuntu-22.04
+    steps:
+      - name: Setup
+        run: |
+          mkdir "setup_values"
+          echo "${{ inputs.prev_workflow_run_id || env.prev_workflow_run_id }}" > "setup_values/prev_workflow_run_id.txt"
+          echo "${{ inputs.other_workflow_run_id || env.other_workflow_run_id }}" > "setup_values/other_workflow_run_id.txt"
+
+      - name: Upload artifacts
+        uses: actions/upload-artifact@v4
+        with:
+          name: setup_values
+          path: setup_values
+
+
+  model-ci:
+    name: Model CI
+    uses: ./.github/workflows/self-scheduled.yml
+    with:
+      job: run_models_gpu
+      slack_report_channel: "#transformers-ci-flash-attn"
+      docker: huggingface/transformers-all-latest-gpu
+      ci_event: Daily CI
+      runner_type: "a10"
+      report_repo_id: hf-internal-testing/transformers_flash_attn_ci
+      commit_sha: ${{ github.sha }}
+      pytest_marker: "flash_attn_test or flash_attn_3_test"
+    secrets: inherit

.github/workflows/self-scheduled.yml

@@ -38,6 +38,10 @@ on:
         default: ""
         required: false
         type: string
+      pytest_marker:
+        required: false
+        type: string
+
 
 env:
   HF_HOME: /mnt/cache
@@ -127,6 +131,7 @@ jobs:
       commit_sha: ${{ inputs.commit_sha || github.sha }}
       runner_type: ${{ inputs.runner_type }}
       report_repo_id: ${{ inputs.report_repo_id }}
+      pytest_marker: ${{ inputs.pytest_marker }}
     secrets: inherit
 
   run_trainer_and_fsdp_gpu:

utils/notification_service.py

@@ -1407,7 +1407,10 @@ def pop_default(l: list[Any], i: int, default: Any) -> Any:
     if not os.path.isdir(os.path.join(os.getcwd(), f"ci_results_{job_name}")):
         os.makedirs(os.path.join(os.getcwd(), f"ci_results_{job_name}"))
 
-    nvidia_daily_ci_workflow = "huggingface/transformers/.github/workflows/self-scheduled-caller.yml"
+    nvidia_daily_ci_workflow = (
+        "huggingface/transformers/.github/workflows/self-scheduled-caller.yml",
+        "huggingface/transformers/.github/workflows/self-scheduled-flash-attn-caller.yml",
+    )
     amd_daily_ci_workflows = (
         "huggingface/transformers/.github/workflows/self-scheduled-amd-mi325-caller.yml",
         "huggingface/transformers/.github/workflows/self-scheduled-amd-mi355-caller.yml",