fix: fix home dir for training image

Signed-off-by: anubhutiv <anubhutiv@nvidia.com>

Author: anubhutivyas

docker/rl/Dockerfile.nmp-rl-training

@@ -85,7 +85,13 @@ RUN bash /app/docker/scripts/remove-listed-files.sh /app/docker/rl/codec-file-re
 # Point runtime uv writes at the per-user cache instead, which is already owned by USER_UID above.
 # The prefetched venvs are unaffected: their symlinks are absolute /opt/uv_cache paths resolved at
 # build time, so changing UV_CACHE_DIR only redirects FUTURE uv operations.
+# HOME must follow the runtime user. The base sets HOME=/home/nmp-build for its build user
+# (uid 2000) and that value survives into the published image, but this stage runs as
+# ${USER_UID}, which cannot write there. Anything resolving `~` at import time then dies:
+# swanlab, which nemo_rl.utils.logger imports unconditionally, calls os.mkdir("~/.swanlab")
+# from its module body, so the GRPO driver fails before reading a config.
 ENV PATH="/opt/nemo_rl_venv/bin:${PATH}" \
+    HOME=/home/${USERNAME} \
     UV_CACHE_DIR=/home/${USERNAME}/.cache/uv
 ENTRYPOINT ["/opt/venv/bin/python"]
 CMD ["-m", "nmp.rl.tasks.training", "--help"]