main
sh 71 lines 2.07 KB
Raw
1 #!/bin/bash
2
3 HOST="0.0.0.0"
4 PORT=8006
5 DATA_PARALLEL_SIZE=4
6 REDUNDANT_EXPERTS=0
7 LOCAL_MODEL_PATH="/models/models--deepseek-ai--DeepSeek-V2-Lite/snapshots/604d5664dddd88a0433dbae533b7fe9472482de0"
8 MODEL_NAME="deepseek-ai/DeepSeek-V2-Lite"
9
10 while [[ $# -gt 0 ]]; do
11 case $1 in
12 --dp)
13 DATA_PARALLEL_SIZE="$2"
14 shift 2
15 ;;
16 --re)
17 REDUNDANT_EXPERTS="$2"
18 shift 2
19 ;;
20 --host)
21 HOST="$2"
22 shift 2
23 ;;
24 --port)
25 PORT="$2"
26 shift 2
27 ;;
28 --model)
29 MODEL_NAME="$2"
30 shift 2
31 ;;
32 --local-model)
33 MODEL_NAME=$LOCAL_MODEL_PATH
34 shift
35 ;;
36 -h|--help)
37 echo "Usage: $0 [OPTIONS]"
38 echo "Options:"
39 echo " --dp SIZE Set data parallel size (default: 4)"
40 echo " --re SIZE Set redundant experts (default: 0)"
41 echo " --host HOST Set host address (default: 0.0.0.0)"
42 echo " --port PORT Set port number (default: 8006)"
43 echo " --model MODEL_NAME Set model name or path"
44 echo " -h, --help Show this help message"
45 exit 0
46 ;;
47 *)
48 echo "Unknown option: $1"
49 echo "Use -h or --help for usage information"
50 exit 1
51 ;;
52 esac
53 done
54
55 echo "Starting vLLM server for $MODEL_NAME with data parallel size: $DATA_PARALLEL_SIZE and redundant experts: $REDUNDANT_EXPERTS"
56
57 export RAY_DEDUP_LOGS=0
58 export VLLM_USE_DEEP_GEMM=1
59
60 vllm serve "$MODEL_NAME" \
61 --data-parallel-size "$DATA_PARALLEL_SIZE" \
62 --data-parallel-size-local "$DATA_PARALLEL_SIZE" \
63 --data-parallel-backend ray \
64 --enforce-eager \
65 --enable-expert-parallel \
66 --enable-eplb \
67 --all2all-backend allgather_reducescatter \
68 --num-redundant-experts "$REDUNDANT_EXPERTS" \
69 --trust-remote-code \
70 --host "$HOST" \
71 --port "$PORT"