Distillative-AI
diff --git a/‎.github/workflows/unit-tests-mbridge-recipes.yaml‎
Lines changed: 80 additions & 5 deletions b/‎.github/workflows/unit-tests-mbridge-recipes.yaml‎
Lines changed: 80 additions & 5 deletions
diff --git a/‎bionemo-recipes/recipes/evo2_megatron/README.md‎
Lines changed: 9 additions & 4 deletions b/‎bionemo-recipes/recipes/evo2_megatron/README.md‎
Lines changed: 9 additions & 4 deletions
diff --git a/‎bionemo-recipes/recipes/evo2_megatron/examples/conftest.py‎
Lines changed: 29 additions & 0 deletions b/‎bionemo-recipes/recipes/evo2_megatron/examples/conftest.py‎
Lines changed: 29 additions & 0 deletions
@@ -25,6 +25,7 @@ jobs:
       any_changed: ${{ steps.changed-files.outputs.any_changed }}
       all_changed_files: ${{ steps.changed-files.outputs.all_changed_files }}
       dirs: ${{ steps.set-dirs.outputs.dirs }}
+      labels: ${{ steps.set-dirs.outputs.labels }}
 
     steps:
       - id: get-pr-info
@@ -118,6 +119,14 @@ jobs:
           ')
           echo "dirs=$DIRS_WITH_IMAGES" >> $GITHUB_OUTPUT
 
+          # Emit PR labels as a JSON array so downstream jobs can gate on ciflow:* labels.
+          if [[ "$PR_INFO" != "null" && "$PR_INFO" != "" ]]; then
+            LABELS=$(echo "$PR_INFO" | jq -c '[.labels[]?.name]' 2>/dev/null || echo "[]")
+          else
+            LABELS="[]"
+          fi
+          echo "labels=$LABELS" >> $GITHUB_OUTPUT
+
       - name: Show output
         run: |
           echo "=== Changed Files Analysis ==="
@@ -195,17 +204,83 @@ jobs:
           fi
           pytest -v .
 
+  run-tests-notebooks:
+    needs: changed-dirs
+    runs-on: linux-amd64-gpu-l4-latest-1
+    # Mirrors the framework workflow's notebook-trigger pattern (label-only on PRs,
+    # auto on merge_group + nightly schedule). Currently scoped to evo2_megatron --
+    # the only megatron recipe with example notebooks.
+    if: |
+        contains(needs.changed-dirs.outputs.dirs, 'bionemo-recipes/recipes/evo2_megatron') &&
+        (
+          (github.event_name == 'schedule') ||
+          (github.event_name == 'merge_group') ||
+          contains(fromJSON(needs.changed-dirs.outputs.labels || '[]'), 'ciflow:all-recipes') ||
+          contains(fromJSON(needs.changed-dirs.outputs.labels || '[]'), 'ciflow:notebooks')
+        )
+    name: "mbridge-notebook-tests (evo2_megatron)"
+    container:
+      image: svcbionemo023/bionemo-framework:pytorch26.04-py3-squashed
+      options: --shm-size=16G
+      env:
+        CI: true
+        HF_TOKEN: ${{ secrets.HF_TOKEN }}
+        HF_HOME: /cache/huggingface
+        BIONEMO_DATA_SOURCE: ngc
+
+    steps:
+      - name: Show GPU info
+        run: nvidia-smi
+      - name: Setup proxy cache
+        uses: nv-gha-runners/setup-proxy-cache@main
+
+      - name: Checkout repository
+        uses: actions/checkout@v4
+        with:
+          sparse-checkout: |
+            bionemo-recipes/recipes/evo2_megatron
+            sub-packages/bionemo-recipeutils
+            sub-packages/bionemo-core
+          sparse-checkout-cone-mode: false
+
+      - name: Cache Hugging Face models
+        uses: actions/cache@v4
+        with:
+          path: /cache/huggingface
+          key: ${{ runner.os }}-huggingface-evo2_megatron-notebooks-${{ github.sha }}
+          restore-keys: |
+            ${{ runner.os }}-huggingface-evo2_megatron-notebooks-
+            ${{ runner.os }}-huggingface-evo2_megatron-
+            ${{ runner.os }}-huggingface-
+
+      - name: Install dependencies
+        working-directory: bionemo-recipes/recipes/evo2_megatron
+        run: |
+          bash .ci_build.sh
+          source .ci_test_env.sh
+          pip install nbval
+
+      - name: Run notebook tests
+        working-directory: bionemo-recipes/recipes/evo2_megatron
+        run: |
+          source .ci_test_env.sh
+          FAST_CI_MODE=1 pytest -v -s --nbval-lax -x -p no:python \
+            examples/lora-fine-tuning-tutorial.ipynb
+
   verify-mbridge-recipe-tests:
-    needs: unit-tests
+    needs:
+      - changed-dirs
+      - unit-tests
+      - run-tests-notebooks
     runs-on: ubuntu-latest
     if: always()
     steps:
-      - name: Check unit-tests matrix status
+      - name: Check test job statuses
         run: |
-          if [[ "${{ needs.unit-tests.result }}" == "failure" || "${{ needs.unit-tests.result }}" == "cancelled" ]]; then
-            echo "Some mbridge unit-tests matrix jobs have failed or been cancelled!"
+          if [[ "${{ contains(needs.*.result, 'failure') || contains(needs.*.result, 'cancelled') }}" == "true" ]]; then
+            echo "Some mbridge test jobs have failed or been cancelled!"
             exit 1
           else
-            echo "All mbridge unit-tests matrix jobs have completed successfully or were skipped!"
+            echo "All mbridge test jobs have completed successfully or were skipped!"
             exit 0
           fi
@@ -276,6 +276,10 @@ freezes the entire base model and attaches low-rank adapter matrices to the
 modules you specify, with an optional escape hatch to keep selected modules
 fully trainable.
 
+> **End-to-end example:** see [`examples/lora-fine-tuning-tutorial.ipynb`](examples/lora-fine-tuning-tutorial.ipynb)
+> for a runnable walkthrough that fine-tunes the 1B checkpoint for splice-site
+> classification, including a head-only baseline for comparison.
+
 ### Basic usage
 
 Add `--lora-finetune` to any `train_evo2` command alongside a checkpoint:
@@ -509,10 +513,11 @@ checkpoint.
 
 The `examples/` directory contains Jupyter notebooks demonstrating common workflows:
 
-| Notebook                     | Description                                            |
-| ---------------------------- | ------------------------------------------------------ |
-| `zeroshot_brca1.ipynb`       | Zero-shot BRCA1 variant effect prediction with Evo2 1B |
-| `fine-tuning-tutorial.ipynb` | Fine-tune the 1B checkpoint on human chromosomes       |
+| Notebook                          | Description                                                                                                            |
+| --------------------------------- | ---------------------------------------------------------------------------------------------------------------------- |
+| `zeroshot_brca1.ipynb`            | Zero-shot BRCA1 variant effect prediction with Evo2 1B                                                                 |
+| `fine-tuning-tutorial.ipynb`      | Fine-tune the 1B checkpoint on human chromosomes                                                                       |
+| `lora-fine-tuning-tutorial.ipynb` | LoRA fine-tune the 1B checkpoint for splice-site classification, with a head-only baseline for trainable-param savings |
 
 ## Docker build
 
 
@@ -0,0 +1,29 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: LicenseRef-Apache2
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# nbval hardcodes a 5s "no iopub message" timeout that fires when a cell
+# subprocess (torchrun) is silent during CUDA setup. Bump it for these notebooks.
+try:
+    import nbval.plugin
+except ImportError:
+    pass
+else:
+    _original_init = nbval.plugin.IPyNbCell.__init__
+
+    def _patched_init(self, *args, **kwargs):
+        _original_init(self, *args, **kwargs)
+        self.output_timeout = 300
+
+    nbval.plugin.IPyNbCell.__init__ = _patched_init