".github/actions/docker-remote-build/action.yml" did not exist on "8969240a3546a83b25bdbe47fc80888f69e5aa7f"
agg.sh 2.3 KB
Newer Older
1
#!/bin/bash
2
# SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
3
# SPDX-License-Identifier: Apache-2.0
4
5
6
#
# Aggregated serving: single worker handles both prefill and decode.
# GPUs: 1
7

8
9
10
11
12
set -e
trap 'echo Cleaning up...; kill 0' EXIT

SCRIPT_DIR="$(dirname "$(readlink -f "$0")")"
source "$SCRIPT_DIR/../../../common/launch_utils.sh"
13

14
15
# Default values
MODEL="Qwen/Qwen3-0.6B"
16
ENABLE_OTEL=false
17
18
19

# Parse command line arguments
EXTRA_ARGS=()
20
21
while [[ $# -gt 0 ]]; do
    case $1 in
22
23
24
25
        --model-path)
            MODEL="$2"
            shift 2
            ;;
26
27
28
29
30
31
32
        --enable-otel)
            ENABLE_OTEL=true
            shift
            ;;
        -h|--help)
            echo "Usage: $0 [OPTIONS]"
            echo "Options:"
33
            echo "  --model-path <name>  Specify model (default: $MODEL)"
34
35
36
            echo "  --enable-otel        Enable OpenTelemetry tracing"
            echo "  -h, --help           Show this help message"
            echo ""
37
            echo "Additional SGLang/Dynamo flags can be passed and will be forwarded"
38
39
40
41
            echo "Note: System metrics are enabled by default on port 8081 (worker)"
            exit 0
            ;;
        *)
42
43
            EXTRA_ARGS+=("$1")
            shift
44
45
46
47
48
            ;;
    esac
done

# Enable tracing if requested
49
TRACE_ARGS=()
50
51
52
53
if [ "$ENABLE_OTEL" = true ]; then
    export DYN_LOGGING_JSONL=true
    export OTEL_EXPORT_ENABLED=1
    export OTEL_EXPORTER_OTLP_TRACES_ENDPOINT=${OTEL_EXPORTER_OTLP_TRACES_ENDPOINT:-http://localhost:4317}
54
    TRACE_ARGS+=(--enable-trace --otlp-traces-endpoint localhost:4317)
55
fi
56

57
HTTP_PORT="${DYN_HTTP_PORT:-8000}"
58
print_launch_banner "Launching Aggregated LLM Worker" "$MODEL" "$HTTP_PORT"
59

60
# run ingress
61
# dynamo.frontend accepts either --http-port flag or DYN_HTTP_PORT env var (defaults to 8000)
62
OTEL_SERVICE_NAME=dynamo-frontend \
63
python3 -m dynamo.frontend &
64

65
# run worker with metrics enabled
66
OTEL_SERVICE_NAME=dynamo-worker DYN_SYSTEM_PORT=${DYN_SYSTEM_PORT:-8081} \
67
python3 -m dynamo.sglang \
68
69
  --model-path "$MODEL" \
  --served-model-name "$MODEL" \
70
71
  --page-size 16 \
  --tp 1 \
72
73
  --trust-remote-code \
  --skip-tokenizer-init \
74
  --enable-metrics \
75
  "${TRACE_ARGS[@]}" \
76
77
78
79
  "${EXTRA_ARGS[@]}" &

# Exit on first worker failure; kill 0 in the EXIT trap tears down the rest
wait_any_exit