Merge pull request #3713 from AI-Hypercomputer:hlo-identical-tests

PiperOrigin-RevId: 907074733

Author: Google-ML-Automation

.github/workflows/build_and_test_maxtext.yml

@@ -218,7 +218,7 @@ jobs:
       base_image: maxtext-unit-test-tpu:py312
       cloud_runner: linux-x86-ct6e-180-4tpu
       pytest_marker: 'not cpu_only and not gpu_only and integration_test and not post_training'
-      pytest_addopts: '--ignore=tests/post_training'
+      pytest_addopts: '--ignore=tests/post_training --ignore=tests/integration/hlo_diff_test.py'
       xla_python_client_mem_fraction: 0.75
       tf_force_gpu_allow_growth: false
       container_resource_option: "--privileged"

.github/workflows/run_tests_against_package.yml

@@ -70,6 +70,10 @@ on:
         description: 'If false, maxtext_sha must be provided for checkout'
         type: boolean
         default: false
+      is_update_hlo:
+        required: false
+        type: boolean
+        default: false
 
 permissions:
   contents: read
@@ -168,13 +172,19 @@ jobs:
           else
             SPLIT_ARGS=""
           fi
-          $PYTHON_EXE -m pytest ${INPUTS_PYTEST_ADDOPTS} \
-            -v \
-            -m "${FINAL_PYTEST_MARKER}" \
-            --durations=0 \
-            $PYTEST_COV_ARGS \
-            $SPLIT_ARGS \
-            ${INPUTS_PYTEST_EXTRA_ARGS}
+          
+          # Setup substitution: If manually updating HLO, skip tests execution and run only the update script instead!
+          if [ "${INPUTS_IS_UPDATE_HLO}" == "true" ]; then
+            python3 tests/utils/update_hlo_references.py
+          else
+            $PYTHON_EXE -m pytest ${INPUTS_PYTEST_ADDOPTS} \
+              -v \
+              -m "${FINAL_PYTEST_MARKER}" \
+              --durations=0 \
+              $PYTEST_COV_ARGS \
+              $SPLIT_ARGS \
+              ${INPUTS_PYTEST_EXTRA_ARGS}
+          fi
 
         env:
           INPUTS_IS_SCHEDULED_RUN: ${{ inputs.is_scheduled_run }}
@@ -185,6 +195,14 @@ jobs:
           INPUTS_WORKER_GROUP: ${{ inputs.worker_group }}
           INPUTS_PYTEST_EXTRA_ARGS: ${{ inputs.pytest_extra_args }}
           INPUTS_MAXTEXT_INSTALLED: ${{ inputs.maxtext_installed }}
+          INPUTS_IS_UPDATE_HLO: ${{ inputs.is_update_hlo }}
+      - name: Upload Reference HLO
+        if: ${{ inputs.is_update_hlo }}
+        uses: actions/upload-artifact@v4
+        with:
+          name: reference-hlo
+          path: tests/utils/reference_hlo_*.txt
+          if-no-files-found: ignore
       - name: Upload results to Codecov
         if: ${{ !inputs.maxtext_installed }} # Skip code coverage upload for maxtext image testing
         uses: codecov/codecov-action@v5

.github/workflows/run_tests_coordinator.yml

@@ -47,6 +47,10 @@ on:
         description: 'If false, maxtext_sha must be provided for checkout'
         type: boolean
         default: false
+      is_update_hlo:
+        required: false
+        type: boolean
+        default: false
 
 permissions:
   contents: read
@@ -150,3 +154,4 @@ jobs:
       worker_group: ${{ matrix.worker_group }}
       total_workers: ${{ contains(inputs.flavor, 'cpu-unit') && 2 || 1 }}
       maxtext_sha: ${{ inputs.maxtext_sha }}
+      is_update_hlo: ${{ inputs.is_update_hlo }}

.github/workflows/update_reference_hlo.yml

@@ -0,0 +1,49 @@
+name: "Update HLO References (for hlo_diff_test.py)"
+
+on:
+  workflow_dispatch:
+permissions:
+  contents: read
+
+jobs:
+  build-wheel:
+    uses: ./.github/workflows/build_package.yml
+    with:
+      device_type: tpu
+      device_name: v6e-4
+      cloud_runner: linux-x86-n2-16-buildkit
+
+  run-tests:
+    needs: build-wheel
+    uses: ./.github/workflows/run_tests_coordinator.yml
+    with:
+      flavor: tpu-integration
+      base_image: maxtext-unit-test-tpu:py312
+      is_scheduled_run: false
+      maxtext_sha: ${{ github.sha }}
+      is_update_hlo: true
+
+  commit-changes:
+    needs: run-tests # Wait for tests to finish
+    runs-on: ubuntu-latest
+    permissions:
+      contents: write
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v4
+        with:
+          ref: ${{ github.ref }}
+
+      - name: Download Reference HLO
+        uses: actions/download-artifact@v4
+        with:
+          name: reference-hlo
+          path: tests/utils/
+
+      - name: Commit and Push changes
+        run: |
+          git config --global user.name "github-actions[bot]"
+          git config --global user.email "github-actions[bot]@users.noreply.github.com"
+          git add tests/utils/reference_hlo_*.txt
+          git commit -m "Update reference HLO from CI artifact"
+          git push

docs/development/hlo_diff_testing.md

@@ -0,0 +1,90 @@
+# HLO Graph Diff Verification Testing
+
+This document provides context for the HLO Graph Diff tests, what HLO is, and how to manage reference baselines.
+
+## Related Files
+
+- **Test Logic**: `tests/integration/hlo_diff_test.py`
+- **Reference Checkpoints baselines**: `tests/utils/reference_hlo_*.txt`
+- **Update Helper script**: `tests/utils/update_hlo_references.py`
+- **GitHub Action Trigger Workflow**: `.github/workflows/update_reference_hlo.yml`
+
+## What is HLO?
+
+**HLO (High-Level Optimizer)** is the intermediate representation used by XLA (Accelerated Linear Algebra) to capture the lowering compiler graph structures.
+
+An HLO module records:
+
+- The sequences of low-level math operations (dot products, convolutions, additions).
+- Array tensor shapes and numerical precisions.
+- Multipod TPU cluster partitioning array sharding mappings.
+
+## Purpose of HloDiffTest
+
+The primary purpose of the `TestHloDiff` validation checks is to ensure that **refactoring PRs are purely refactoring code** and not unintentionally impacting graph compiler lowering or performance.
+
+- **For pure refactors:** The HLO graph layout should remain *strictly identical*. Any detected deviation flags that execution boundaries or operation pipelines might have changed under the hood.
+- **For dependency updates:** Changes to framework dependencies (like updating JAX or XLA versions) *are expected* to slightly alter compiled HLO output layouts, which makes baseline updates appropriate in those scenarios.
+
+______________________________________________________________________
+
+## How the Test Works
+
+This test runs automatically as part of the [`tpu-integration`](https://github.com/AI-Hypercomputer/maxtext/actions/workflows/build_and_test_maxtext.yml) CI test suite on every Pull Request.
+
+When the test method executes, it performs the following sequence of actions:
+
+1. **Triggers Compilation**: It runs the model training lifecycle compilation-only phase (invoking `train_compile.main()`) without actually allocating hardware compute nodes or running optimization passes.
+2. **Dumps HLO modules**: Instructs the XLA compiler back-end to capture optimizer operations lowering structure graphs and dump them to text files.
+3. **Strict comparison matches**: Compares the structural lines of the generated representation graph directly against baseline `.txt` copies stored under `tests/utils/`.
+
+______________________________________________________________________
+
+## Updating HLO reference files
+
+When intended architectures transformations alter graph lowering, reference file baselines require updates.
+
+> [!IMPORTANT]\
+> While running the update script locally is not the end of the world, **relying on local execution can cause remote CI tests to fail.**
+> The PR verification pipelines run the tests in a strictly locked GitHub Actions environment. The smallest discrepancies in local library installations will introduce slight backend lowering graph deviations. If your local execution leads to a remote CI check failure, rely on the GitHub Action trigger described below to generate environment-matching baselines.
+
+### Method 1: Run the manual GitHub Action Workflow (Highly Recommended)
+
+Triggering the CI workflow guarantees execution runs within the correct environment isolation scope.
+
+#### Option A: Using the GitHub UI
+
+1. Go to the Actions tab in the repository browser.
+2. Find the manual workflow: `Update HLO References (for hlo_diff_test.py)`.
+3. Run it targeting your PR workspace branch. It compiles the graph layout and commits the baseline update files back to the branch automatically.
+
+#### Option B: Using the GitHub CLI (`gh`)
+
+Alternatively, you can trigger the remote workflow via terminal CLI execution:
+
+```bash
+gh workflow run update_reference_hlo.yml --ref <branch>
+```
+
+> [!NOTE]
+> A successful run of the manual update workflow will add a new commit to your Pull Request branch. Once complete, you must:
+>
+> 1. Pull the new commit from remote.
+> 2. Squash the commits in your branch once again to keep your PR history clean.
+> 3. Push the squashed commit to remote.
+> 4. Retry the `tpu-integration` workflow to verify tests pass on your PR.
+
+### Method 2: Local Execution
+
+If you need to test or update baselines manually during development:
+
+```bash
+source .venv/bin/activate
+pytest tests/integration/hlo_diff_test.py -v
+```
+
+Or to force update the local baselines:
+
+```bash
+python3 tests/utils/update_hlo_references.py
+```