dgx-spark-playbooks/nvidia/trt-llm/assets/docker-compose.yml
2025-10-12 18:14:37 +00:00

47 lines
1.2 KiB
YAML

version: '3.8'
services:
trtllm:
image: nvcr.io/nvidia/tensorrt-llm/release:1.0.0rc3
deploy:
replicas: 2
restart_policy:
condition: any
delay: 5s
max_attempts: 3
window: 120s
resources:
reservations:
generic_resources:
- discrete_resource_spec:
kind: 'NVIDIA_GPU'
value: 1
environment:
- UCX_NET_DEVICES=enp1s0f0np0,enp1s0f1np1
- NCCL_SOCKET_IFNAME=enp1s0f0np0,enp1s0f1np1
- OMPI_MCA_btl_tcp_if_include=enp1s0f0np0,enp1s0f1np1
- OMPI_MCA_orte_default_hostfile=/etc/openmpi-hostfile
- OMPI_MCA_rmaps_ppr_n_pernode=1
- OMPI_ALLOW_RUN_AS_ROOT=1
- OMPI_ALLOW_RUN_AS_ROOT_CONFIRM=1
entrypoint: /opt/trtllm-mn-entrypoint.sh
volumes:
- ~/.cache/huggingface/:/root/.cache/huggingface/
- ~/trtllm-mn-entrypoint.sh:/opt/trtllm-mn-entrypoint.sh
- ~/.ssh:/tmp/.ssh:ro
ulimits:
memlock: -1
stack: 67108864
networks:
- host
healthcheck:
test: ["CMD", "service", "ssh", "status"]
interval: 30s
timeout: 10s
retries: 10
networks:
host:
name: host
external: true