Updated examples with agnostic platforms

Author: Timothy Le

…lows/osmo/test_isaac_lab_serverless.yaml …orkflows/osmo/test_serverless_isaac.yamlapplications/osmo/workflows/osmo/test_isaac_lab_serverless.yaml renamed to applications/osmo/workflows/osmo/test_serverless_isaac.yaml applications/osmo/workflows/osmo/test_isaac_lab_serverless.yaml renamed to applications/osmo/workflows/osmo/test_serverless_isaac.yaml

@@ -1,3 +1,5 @@
+# Launches an Isaac Lab job on Nebius AI serverless from an OSMO CPU-only helper pod.
+# Waits for the remote job to finish and reports the final serverless status in the task logs.
 workflow:
   name: osmo-isaac-rl-launcher
   pool: default
@@ -6,7 +8,6 @@ workflow:
     nebius-cli:
       cpu: 1
       memory: 2Gi
-      platform: L40S
 
   tasks:
     - name: trigger-isaac-l40s-job
@@ -18,7 +19,6 @@ workflow:
         PARENT_ID: "project-XXX"
         NEBIUS_IAM_TOKEN: "<nebius iam get-access-token result>"
         NGC_API_KEY: "nvapi-XXX"
-
       command:
         - bash
         - -lc
@@ -40,8 +40,10 @@ workflow:
             "$PARENT_ID" > /root/.nebius/config.yaml
 
           echo "--- 3. Launch Isaac Lab L40S job ---"
+          # The OSMO task is just a CPU launcher pod. The actual serverless
+          # target is selected by the Nebius AI job platform below.
           JOB_OUTPUT=$(nebius ai job create \
-            --name "isaac-reachy-$(date +%s)" \
+            --name "isaac-serverless-$(date +%s)" \
             --image "nvcr.io/nvidia/isaac-lab:2.3.2" \
             --registry-username "\$oauthtoken" \
             --registry-password "$NGC_API_KEY" \

…mo/workflows/osmo/test_isaac_reachy.yaml …s/osmo/test_serverless_isaac_reachy.yamlapplications/osmo/workflows/osmo/test_isaac_reachy.yaml renamed to applications/osmo/workflows/osmo/test_serverless_isaac_reachy.yaml applications/osmo/workflows/osmo/test_isaac_reachy.yaml renamed to applications/osmo/workflows/osmo/test_serverless_isaac_reachy.yaml

@@ -1,3 +1,5 @@
+# Launches an Isaac Lab Reach/Franka training job on Nebius AI serverless from OSMO.
+# Mounts shared storage for logs/checkpoints, then waits for completion and surfaces failures.
 workflow:
   name: osmo-isaac-reachy-rl
   pool: default
@@ -6,7 +8,6 @@ workflow:
     nebius-cli:
       cpu: 1
       memory: 2Gi
-      platform: L40S
 
   tasks:
     - name: trigger-reachy-training
@@ -16,9 +17,9 @@ workflow:
       environment:
         SUBNET_ID: "vpcsubnet-XXX"
         PARENT_ID: "project-XXX"
-        FILESYSTEM_ID: "computefilesystem-XXX"
         NEBIUS_IAM_TOKEN: "NEBIUS_IAM_TOKEN"
         NGC_API_KEY: "nvapi-XXX"
+        FILESYSTEM_ID: "computefilesystem-XXX"
         # RL training parameters
         TASK: "Isaac-Reach-Franka-v0"  # swap for Isaac-Reach-Reachy-v0 when extension is installed
         NUM_ENVS: "4096"
@@ -43,24 +44,58 @@ workflow:
           mkdir -p /root/.nebius
           printf 'current-profile: default\nprofiles:\n  default:\n    endpoint: api.eu-north1.nebius.cloud:443\n    parent-id: %s\n' \
             "$PARENT_ID" > /root/.nebius/config.yaml
+          mkdir -p {{output}}
 
           echo "--- 3. Launch Isaac Lab training job ---"
-          RUN_NAME="${TASK}-$(date +%s)"
-          JOB_OUTPUT=$(nebius ai job create \
-            --name "reachy-train-$(date +%s)" \
-            --image "nvcr.io/nvidia/isaac-lab:2.3.2" \
-            --registry-username "\$oauthtoken" \
-            --registry-password "$NGC_API_KEY" \
-            --platform "gpu-l40s-d" \
-            --preset "1gpu-16vcpu-96gb" \
-            --volume "$FILESYSTEM_ID:/mnt/storage" \
-            --container-command "bash" \
-            --args "-c '/workspace/isaaclab/isaaclab.sh -p /workspace/isaaclab/scripts/reinforcement_learning/rsl_rl/train.py --task $TASK --headless --num_envs $NUM_ENVS --max_iterations $MAX_ITERATIONS +runner.log_root_path=/mnt/storage/logs 2>&1'" \
-            --timeout "120m" \
-            --subnet-id "$SUBNET_ID" \
-            --parent-id "$PARENT_ID")
+          # The OSMO task is just a CPU launcher pod. The actual serverless
+          # target is selected by the Nebius AI job platform below.
+          if [[ -z "${FILESYSTEM_ID:-}" || "${FILESYSTEM_ID}" == "computefilesystem-XXX" ]]; then
+            echo "FILESYSTEM_ID must be set to a real filestore ID for persistent training outputs." >&2
+            exit 1
+          fi
+
+          LOG_ROOT_PATH="/mnt/storage/logs"
+          RESULTS_PATH="${LOG_ROOT_PATH}/${TASK}/"
+          echo "Using filestore for logs/checkpoints: ${FILESYSTEM_ID}"
+          JOB_CMD=(
+            nebius ai job create
+            --name "reachy-train-$(date +%s)"
+            --image "nvcr.io/nvidia/isaac-lab:2.3.2"
+            --registry-username "\$oauthtoken"
+            --registry-password "$NGC_API_KEY"
+            --platform "gpu-h100-sxm"
+            --preset "1gpu-16vcpu-200gb"
+            --volume "${FILESYSTEM_ID}:/mnt/storage"
+          )
+
+          JOB_CMD+=(
+            --container-command "bash"
+            --args "-c '/workspace/isaaclab/isaaclab.sh -p /workspace/isaaclab/scripts/reinforcement_learning/rsl_rl/train.py --task $TASK --headless --num_envs $NUM_ENVS --max_iterations $MAX_ITERATIONS +runner.log_root_path=${LOG_ROOT_PATH} 2>&1'"
+            --timeout "120m"
+            --subnet-id "$SUBNET_ID"
+            --parent-id "$PARENT_ID"
+          )
+
+          JOB_CREATE_RESPONSE_FILE="{{output}}/job-create-response.txt"
+          JOB_CREATE_ERROR_FILE="{{output}}/job-create-error.txt"
+          JOB_ID_FILE="{{output}}/job-id.txt"
+
+          if ! JOB_OUTPUT=$("${JOB_CMD[@]}" 2>&1); then
+            printf '%s\n' "$JOB_OUTPUT" | tee "$JOB_CREATE_ERROR_FILE"
+            echo "Job create failed before a job ID was returned." | tee -a "$JOB_CREATE_ERROR_FILE"
+            echo "If the plain serverless Isaac workflow succeeds, the likely culprit is the filestore mount: ${FILESYSTEM_ID}:/mnt/storage" | tee -a "$JOB_CREATE_ERROR_FILE"
+            exit 1
+          fi
+
+          printf '%s\n' "$JOB_OUTPUT" > "$JOB_CREATE_RESPONSE_FILE"
           echo "Job create response: $JOB_OUTPUT"
           JOB_ID=$(echo "$JOB_OUTPUT" | grep -oE 'aijob-[a-z0-9]+' | head -1)
+          if [[ -z "$JOB_ID" ]]; then
+            echo "Job create response did not include a job ID." | tee "$JOB_CREATE_ERROR_FILE"
+            printf '%s\n' "$JOB_OUTPUT" | tee -a "$JOB_CREATE_ERROR_FILE"
+            exit 1
+          fi
+          printf '%s\n' "$JOB_ID" > "$JOB_ID_FILE"
           echo "Submitted training job: $JOB_ID"
 
           echo "--- 4. Wait for training completion ---"
@@ -72,7 +107,7 @@ workflow:
             case "$STATUS" in
               SUCCEEDED|COMPLETED)
                 echo "Training completed successfully."
-                echo "Logs and checkpoints saved to: /mnt/storage/logs/$TASK/"
+                echo "Logs and checkpoints path: ${RESULTS_PATH}"
                 break ;;
               FAILED|ERROR|CANCELLED)
                 echo "Training failed with status: $STATUS"