first add

docker-compose_mega.yml

@@ -0,0 +1,38 @@
+
+services:
+  megatron-test:
+    image: nvcr.io/nvidia/pytorch:24.02-py3
+    shm_size: '560gb'
+    deploy:
+      resources:
+        reservations:
+          devices:
+            - driver: nvidia
+              count: all
+              capabilities: [gpu]
+    #runtime: nvidia
+    environment:
+      - NVIDIA_VISIBLE_DEVICES=all
+      - NVIDIA_DRIVER_CAPABILITIES=compute,utility
+        #- CUTLASS_PATH="/opt/cutlass"
+        #- CUDA_HOME="/usr/local/cuda"
+        #- PATH="${CUDA_HOME}/bin:${PATH}"
+        #- LD_LIBRARY_PATH="${CUDA_HOME}/lib64:${LD_LIBRARY_PATH}"
+    stdin_open: true
+    tty: true
+    privileged: true
+    cap_add:
+      - IPC_LOCK
+    volumes:
+      - /root/workspace:/mnt
+      - /dev/infiniband:/dev/infiniband
+    #  - /mnt/local-nvme:/root/
+    ports:
+      - "5000:5000"
+    # networks:
+    #   - ldh_overlay_network
+    network_mode: host
+
+# networks:
+#   ldh_overlay_network:
+#     external: true