Address review feedback on GRPO sandbox integration

Signed-off-by: Wei Du <wedu@nvidia.com>

Author: wedu-nvidia

dockerfiles/README.md

@@ -32,4 +32,17 @@ We directly use official `nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc8` image.
 
 ## Building sglang image
 
+<<<<<<< HEAD
 We directly use official `lmsysorg/sglang:v0.5.10.post1` image.
+=======
+We directly use official `lmsysorg/sglang:v0.5.8` image.
+
+## NeMo-RL image default
+
+The sample local cluster config currently defaults `containers.nemo-rl` to the pinned private image
+`nvcr.io/nvidian/nemo-rl:733c7d6-48669320`. This image is accessible to NVIDIA employees who already have access
+to the corresponding internal registry.
+
+If you do not have access to that registry, build the NeMo-RL image yourself from the matching commit in the
+upstream NeMo-RL repository and point your cluster config at that locally built image instead.
+>>>>>>> 1a725b3d (Address review feedback on GRPO sandbox integration)

nemo_skills/pipeline/nemo_rl/grpo.py

@@ -333,7 +333,9 @@ def grpo_nemo_rl(
         help="If specified, will reuse the code from this experiment. "
         "Can provide an experiment name or an experiment object if running from code.",
     ),
-    config: str = typer.Option(None, help="Override training config YAML; defaults to the upstream container config"),
+    config: str = typer.Option(
+        None, help="Override training config YAML; defaults to the upstream container config for nano-v3"
+    ),
     container: str = typer.Option(None, help="Override container image for NeMo-RL training/conversion jobs"),
     with_sandbox: bool = typer.Option(False, help="If True, will start a sandbox container alongside this job"),
     sandbox_container: str = typer.Option(None, help="Override container image for the sandbox sidecar"),