fix: pass absolute --config-path to verl; hydra resolved ours under verl/trainer

Author: Devesh-Maheshwari

chtc/train_grpo.sh

@@ -58,7 +58,7 @@ pip install -e ".[dev,gpu]" --quiet
 # Pin to v0.7.0 (same as llm-starter, known to work with this image).
 echo "==> cloning + installing verl@v0.7.0"
 if [ ! -d "verl" ]; then
-    git clone https://github.com/volcengine/verl.git -b v0.7.0 --depth 1
+    git clone https://github.com/volcengine/verl.git -b v0.7.0 
 fi
 pip install -e verl --quiet
 
@@ -103,13 +103,17 @@ else
 fi
 
 # --- Launch verl ---
-echo "==> python -m verl.trainer.main_ppo --config-path=configs --config-name=${CONFIG_NAME}"
+# Hydra resolves --config-path RELATIVE TO the entry point's source dir
+# (verl/trainer/) by default. Pass an absolute path so it picks up our
+# project's configs/ instead of looking inside verl's own source tree.
+CONFIG_PATH_ABS="$(pwd)/configs"
+echo "==> python -m verl.trainer.main_ppo --config-path=${CONFIG_PATH_ABS} --config-name=${CONFIG_NAME}"
 
 # Don't bail on training failure — we still want the placeholder tarball
 # transferred so the job doesn't end up held.
 set +e
 python -m verl.trainer.main_ppo \
-    --config-path=configs \
+    --config-path="${CONFIG_PATH_ABS}" \
     --config-name="${CONFIG_NAME}" \
     ${MODEL_OVERRIDE}
 TRAIN_EXIT=$?