epd_disagg.sh

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

# Environment variables with defaults
export MODEL_PATH=${MODEL_PATH:-"Qwen/Qwen2-VL-7B-Instruct"}
export SERVED_MODEL_NAME=${SERVED_MODEL_NAME:-"Qwen/Qwen2-VL-7B-Instruct"}
export DISAGGREGATION_STRATEGY=${DISAGGREGATION_STRATEGY:-"decode_first"}
export PREFILL_ENGINE_ARGS=${PREFILL_ENGINE_ARGS:-"engine_configs/prefill.yaml"}
export DECODE_ENGINE_ARGS=${DECODE_ENGINE_ARGS:-"engine_configs/decode.yaml"}
export ENCODE_ENGINE_ARGS=${ENCODE_ENGINE_ARGS:-"engine_configs/encode.yaml"}
export PREFILL_CUDA_VISIBLE_DEVICES=${PREFILL_CUDA_VISIBLE_DEVICES:-"0"}
export DECODE_CUDA_VISIBLE_DEVICES=${DECODE_CUDA_VISIBLE_DEVICES:-"1"}
export ENCODE_CUDA_VISIBLE_DEVICES=${ENCODE_CUDA_VISIBLE_DEVICES:-"2"}
export ENCODE_ENDPOINT=${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}

# Setup cleanup trap
cleanup() {
    echo "Cleaning up background processes..."
    kill $DYNAMO_PID $PREFILL_PID $DECODE_PID $ENCODE_PID 2>/dev/null || true
    wait $DYNAMO_PID $PREFILL_PID $DECODE_PID $ENCODE_PID 2>/dev/null || true
    echo "Cleanup complete."
}
trap cleanup EXIT INT TERM

# run clear_namespace
python3 utils/clear_namespace.py --namespace dynamo

# run frontend
python3 -m dynamo.frontend --http-port 8000 &
DYNAMO_PID=$!

# run encode worker
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" \
  --disaggregation-strategy "$DISAGGREGATION_STRATEGY" \
  --modality "$MODALITY" \
  --allowed-local-media-path "$ALLOWED_LOCAL_MEDIA_PATH" \
  --max-file-size-mb "$MAX_FILE_SIZE_MB" \
  --disaggregation-mode encode &
ENCODE_PID=$!

# 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" \
  --disaggregation-strategy "$DISAGGREGATION_STRATEGY" \
  --modality "$MODALITY" \
  --disaggregation-mode prefill \
  --encode-endpoint "$ENCODE_ENDPOINT" &
PREFILL_PID=$!

# 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" \
  --disaggregation-strategy "$DISAGGREGATION_STRATEGY" \
  --modality "$MODALITY" \
  --disaggregation-mode decode &
DECODE_PID=$!

wait $DYNAMO_PID