feat: add NeMo Gym rollouts pipeline

docs/basics/code-packaging.md

@@ -40,6 +40,21 @@ running commands from it.
     ```
     In all cases, uncommitted code will not be used.
 
+!!! note
+
+    You can override the default packaging behavior with the following environment variables:
+
+    - `NEMO_SKILLS_FORCE_PATTERN_PACKAGER=1` — Skip git-based packaging entirely and always use the installed
+      `nemo_skills` package tree (PatternPackager). Useful when you have an editable install and don't want
+      packaging tied to the git state of your current directory.
+    - `NEMO_SKILLS_FORCE_INSTALLED_PACKAGE=1` — When running from a git repo, use the installed `nemo_skills`
+      package instead of the repo's `nemo_skills/` directory. The git repo is still packaged, but `nemo_skills`
+      is picked up from the installed location. Useful when your repo checkout has extra files you don't want
+      uploaded.
+
+    Note that `NEMO_SKILLS_FORCE_INSTALLED_PACKAGE` has no effect when `NEMO_SKILLS_FORCE_PATTERN_PACKAGER`
+    is also set, since the latter bypasses the git repo branch entirely.
+
 
 Finally, it's important to keep in mind that whenever you submit a new experiment, NeMo-Run will create a copy of your
 code package both locally (inside `~/.nemo_run`) and on cluster (inside `ssh_tunnel/job_dir` path in your cluster config).

nemo_skills/evaluation/evaluator/livecodebench.py

@@ -211,7 +211,7 @@ async def eval_livecodebench_async(eval_config: LiveCodeBenchEvaluatorConfig):
     """Evaluation running within a sandbox."""
     async with sandbox_context(eval_config.sandbox) as sandbox:
         if not await _install_packages_in_sandbox(sandbox, eval_config):
-            return
+            raise RuntimeError("Failed to install livecodebench packages in sandbox.")
 
         jsonl_file = eval_config.input_file
         LOG.info(f"Processing file: {jsonl_file} in sandbox")
@@ -299,6 +299,10 @@ def eval_livecodebench(cfg):
         raise RuntimeError("The 'pypy3' interpreter requires a running sandbox, but the service was unreachable.")
 
     if sandbox_is_ready:
-        asyncio.run(eval_livecodebench_async(eval_config))
+        try:
+            asyncio.run(eval_livecodebench_async(eval_config))
+        except Exception as e:
+            LOG.warning(f"Sandbox evaluation failed: {e}. Falling back to local evaluation.")
+            eval_livecodebench_without_sandbox(eval_config)
     else:
         eval_livecodebench_without_sandbox(eval_config)

nemo_skills/pipeline/cli.py

@@ -26,6 +26,7 @@
 from nemo_skills.pipeline.generate import generate
 from nemo_skills.pipeline.megatron_lm.train import train_megatron_lm
 from nemo_skills.pipeline.nemo_evaluator import nemo_evaluator
+from nemo_skills.pipeline.nemo_gym_rollouts import nemo_gym_rollouts
 from nemo_skills.pipeline.nemo_rl.grpo import grpo_nemo_rl
 from nemo_skills.pipeline.nemo_rl.sft import sft_nemo_rl
 from nemo_skills.pipeline.prepare_data import prepare_data