add workflow for syncing

.github/workflows/sync_branch.yaml

@@ -0,0 +1,24 @@
+name: sync main with dev
+on:
+  push:
+    branches:
+      - main
+
+jobs:
+  sync-branches:
+    runs-on: ubuntu-latest
+    name: syncing main with dev
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v2
+      - name: Set up Node
+        uses: actions/setup-node@v1
+        with:
+          node-version: 12
+      - name: Opening pull request
+        id: pull
+        uses: tretuna/[email protected]
+        with:
+          GITHUB_TOKEN: ${{secrets.GITHUB_TOKEN}}
+          FROM_BRANCH: "main"
+          TO_BRANCH: "dev"

CONTRIBUTING.md

@@ -4,14 +4,18 @@ Thanks for the interest in contributing to NeMo-Aligner. We do all of NeMo-Align
 
 # Pull Requests (PR) Guidelines
 
-**Send your PRs to the `main` branch**
+**Send your PRs to the `main` or `dev` branch**
 
 1) Make sure your PR does one thing. Have a clear answer to "What does this PR do?".
 2) Read General Principles and style guide below
 3) Make sure you sign your commits. E.g. use ``git commit -sS`` when committing.
 4) Make sure all unittests finish successfully before sending PR ``pytest`` or (if your dev box does not have GPU) ``pytest --cpu`` from the root folder
 5) Send your PR and request a review
 
+**NOTE**: The `main` branch uses a fixed NeMo version which we will update on every release. The `dev` branch is the branch that has all commits from `main` but uses NeMo's main branch: this branch is less stable but we run nightly tests on it to make sure everything works. We only provide the dockerfile that works with `main`, which is the branch most PRs should target unless they require the latest NeMo main (in which case they should target `dev`).
+
+Every release `dev` and `main` will sync to be the same.
+
 ## Unit tests
 Quick unit tests (locally, while developing)
 ```

nemo_aligner/models/nlp/gpt/megatron_gpt_reward_model.py

@@ -318,7 +318,7 @@ def on_load_checkpoint(self, checkpoint) -> None:
         """
         # mcore uses distributed checkpointing
         if "state_dict" in checkpoint and checkpoint["state_dict"]:
-            for index, module in enumerate(self.get_gpt_module_list()):
+            for index, module in enumerate(self.get_model_module_list()):
                 if parallel_state.get_virtual_pipeline_model_parallel_world_size() is not None:
                     checkpoint_state_dict = checkpoint["state_dict"][f"model_{index}"]
                 else:

nemo_aligner/utils/train_utils.py

@@ -38,7 +38,7 @@ def set_sync_funcs(ptl_model, forward_only):
         param_sync_func = ptl_model.sync_overlap_parameters
 
     # pipeline schedules will get these from ptl_model.model.config
-    for module in ptl_model.get_gpt_module_list():
+    for module in ptl_model.get_model_module_list():
         module.config.no_sync_func = no_sync_func
         module.config.grad_sync_func = grad_sync_func
         module.config.param_sync_func = param_sync_func