first add

docker-compose_nccl.yml

@@ -0,0 +1,28 @@
+version: '3.8'
+# https://github.com/mayooot/build-nccl-tests-with-pytorch
+services:
+  nccl-test-container:
+    image: mayooot/nccl-tests-with-pytorch:v0.0.2
+    container_name: nccl-test-container
+    network_mode: host
+    environment:
+      - PORT=1998
+      - PASS=P@88w0rd
+      - NVIDIA_VISIBLE_DEVICES=all
+      - NVIDIA_DRIVER_CAPABILITIES=compute,utility
+    volumes:
+      - ./id_rsa_finetune:/root/.ssh/id_rsa
+      - ./id_rsa.pub:/root/.ssh/id_rsa.pub
+    deploy:
+      resources:
+        reservations:
+          devices:
+            - driver: nvidia
+              count: all
+              capabilities: [gpu]
+    cap_add:
+      - IPC_LOCK
+    devices:
+      - /dev/infiniband:/dev/infiniband
+    shm_size: '32gb'
+    restart: unless-stopped