disagg_e_pd.sh

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

# 1 Encode + 1 PD worker for Qwen3-VL-2B-Instruct
# GPU 0: Encode (vision encoder)
# GPU 0: PD worker (prefill + decode, TP=1)

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=${SERVED_MODEL_NAME:-"Qwen/Qwen3-VL-2B-Instruct"}
export ENCODE_ENGINE_ARGS=${ENCODE_ENGINE_ARGS:-"$DYNAMO_HOME/examples/backends/trtllm/engine_configs/qwen3-vl-2b-instruct/encode.yaml"}
export PD_ENGINE_ARGS=${PD_ENGINE_ARGS:-"$DYNAMO_HOME/examples/backends/trtllm/engine_configs/qwen3-vl-2b-instruct/agg.yaml"}
export ENCODE_CUDA_VISIBLE_DEVICES=${ENCODE_CUDA_VISIBLE_DEVICES:-"0"}
export ENCODE_ENDPOINT=${ENCODE_ENDPOINT:-"dyn://dynamo.tensorrt_llm_encode.generate"}
export MODALITY=${MODALITY:-"multimodal"}
export ALLOWED_LOCAL_MEDIA_PATH=${ALLOWED_LOCAL_MEDIA_PATH:-"/tmp"}
export MAX_FILE_SIZE_MB=${MAX_FILE_SIZE_MB:-50}

# Extra arguments forwarded to the PD worker (e.g. --multimodal-embedding-cache-capacity-gb 10)
EXTRA_PD_ARGS=("$@")

HTTP_PORT="${DYN_HTTP_PORT:-8000}"
print_launch_banner --multimodal "Launching Multimodal E/PD" "$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 encode worker (vision encoder on GPU 0)
CUDA_VISIBLE_DEVICES=$ENCODE_CUDA_VISIBLE_DEVICES python3 -m dynamo.trtllm \
  --model-path "$MODEL_PATH" \
  --served-model-name "$SERVED_MODEL_NAME" \
  --extra-engine-args "$ENCODE_ENGINE_ARGS" \
  --modality "$MODALITY" \
  --allowed-local-media-path "$ALLOWED_LOCAL_MEDIA_PATH" \
  --max-file-size-mb "$MAX_FILE_SIZE_MB" \
  --disaggregation-mode encode &

# run PD worker 1 (GPU 0)
CUDA_VISIBLE_DEVICES=0 python3 -m dynamo.trtllm \
  --model-path "$MODEL_PATH" \
  --served-model-name "$SERVED_MODEL_NAME" \
  --extra-engine-args "$PD_ENGINE_ARGS" \
  --modality "$MODALITY" \
  --encode-endpoint "$ENCODE_ENDPOINT" \
  --disaggregation-mode prefill_and_decode \
  "${EXTRA_PD_ARGS[@]}" &

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