disagg_multimodal.sh

#!/bin/bash
# SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
# SPDX-License-Identifier: Apache-2.0

set -e
trap 'echo Cleaning up...; kill 0' EXIT

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

# Environment variables with defaults
export DYNAMO_HOME=${DYNAMO_HOME:-"/workspace"}
export MODEL_PATH=${MODEL_PATH:-"Qwen/Qwen3-VL-2B-Instruct"}
export SERVED_MODEL_NAME=${SERVED_MODEL_NAME:-"Qwen3-VL-2B-Instruct"}
export PREFILL_ENGINE_ARGS=${PREFILL_ENGINE_ARGS:-"$DYNAMO_HOME/examples/backends/trtllm/engine_configs/qwen3-vl-2b-instruct/prefill.yaml"}
export DECODE_ENGINE_ARGS=${DECODE_ENGINE_ARGS:-"$DYNAMO_HOME/examples/backends/trtllm/engine_configs/qwen3-vl-2b-instruct/decode.yaml"}
export PREFILL_CUDA_VISIBLE_DEVICES=${PREFILL_CUDA_VISIBLE_DEVICES:-"0"}
export DECODE_CUDA_VISIBLE_DEVICES=${DECODE_CUDA_VISIBLE_DEVICES:-"1"}
export MODALITY=${MODALITY:-"multimodal"}

HTTP_PORT="${DYN_HTTP_PORT:-8000}"
print_launch_banner --multimodal "Launching Disaggregated Multimodal Serving (2 GPUs)" "$MODEL_PATH" "$HTTP_PORT"

# run frontend
# dynamo.frontend accepts either --http-port flag or DYN_HTTP_PORT env var (defaults to 8000)
python3 -m dynamo.frontend &

# run prefill worker
CUDA_VISIBLE_DEVICES=$PREFILL_CUDA_VISIBLE_DEVICES python3 -m dynamo.trtllm \
  --model-path "$MODEL_PATH" \
  --served-model-name "$SERVED_MODEL_NAME" \
  --extra-engine-args  "$PREFILL_ENGINE_ARGS" \
  --modality "$MODALITY" \
  --disaggregation-mode prefill &

# run decode worker
CUDA_VISIBLE_DEVICES=$DECODE_CUDA_VISIBLE_DEVICES python3 -m dynamo.trtllm \
  --model-path "$MODEL_PATH" \
  --served-model-name "$SERVED_MODEL_NAME" \
  --extra-engine-args  "$DECODE_ENGINE_ARGS" \
  --modality "$MODALITY" \
  --disaggregation-mode decode &

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