vLLM con ray a mano
- Alejandro Rivero
- hace 7 días
- 1 Min. de lectura
#necesarioexport SSL_CERT_FILE=/fs/agustina/arivero/supercomplex/.local/lib/python3.11/site-packages/certifi/cacert.pem export RAY_NODE_MANAGER_HEARTBEAT_TIMEOUT_MILLISECONDS=20000 # 20 seconds export RAY_GCS_HEARTBEAT_TIMEOUT_SECONDS=60 1064 export RAY_DISABLE_PLACEMENT_GROUP_BUNDLING=1#opcionales por lios de debug, igual ni existen. 1116 export RAY_NODE_IP_ADDRESS=172.22.5.1 1134 export NCCL_DEBUG=INFO 1135 export NCCL_DEBUG_SUBSYS=ALL 1136 export NCCL_P2P_DISABLE=0 # or 1 if P2P is not supported 1137 export RAY_DEDUP_LOGS=0 1154 export NCCL_SOCKET_IFNAME=ib1
module load python-math/3.11.4ray start --head --node-ip-address=172.22.5.1 --num-gpus=4ray list placement-groups --detailvllm serve deepseek-ai/DeepSeek-R1 --distributed-executor-backend=ray --trust-remote-code --pipeline-parallel-size 2 --tensor-parallel-size 4



Comentarios