gpt_oss_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 DYNAMO_HOME=${DYNAMO_HOME:-"/workspace"}
export MODEL_PATH=${MODEL_PATH:-"/model"}
export SERVED_MODEL_NAME=${SERVED_MODEL_NAME:-"openai/gpt-oss-120b"}
export PREFILL_ENGINE_ARGS=${PREFILL_ENGINE_ARGS:-"$DYNAMO_HOME/examples/backends/trtllm/engine_configs/gpt-oss-120b/prefill.yaml"}
export DECODE_ENGINE_ARGS=${DECODE_ENGINE_ARGS:-"$DYNAMO_HOME/examples/backends/trtllm/engine_configs/gpt-oss-120b/decode.yaml"}

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


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

# With tensor_parallel_size=4, each worker needs 4 GPUs
# run prefill worker
CUDA_VISIBLE_DEVICES=0,1,2,3 python3 -m dynamo.trtllm \
  --model-path "$MODEL_PATH" \
  --served-model-name "$SERVED_MODEL_NAME" \
  --extra-engine-args "$PREFILL_ENGINE_ARGS" \
  --dyn-reasoning-parser gpt_oss \
  --dyn-tool-call-parser harmony \
  --disaggregation-mode prefill \
  --max-num-tokens 20000 \
  --max-batch-size 32 \
  --free-gpu-memory-fraction 0.9 \
  --tensor-parallel-size 4 \
  --expert-parallel-size 4 &

# run decode worker
CUDA_VISIBLE_DEVICES=4,5,6,7 python3 -m dynamo.trtllm \
  --model-path "$MODEL_PATH" \
  --served-model-name "$SERVED_MODEL_NAME" \
  --extra-engine-args "$DECODE_ENGINE_ARGS" \
  --dyn-reasoning-parser gpt_oss \
  --dyn-tool-call-parser harmony \
  --disaggregation-mode decode \
  --max-num-tokens 16384 \
  --free-gpu-memory-fraction 0.9 \
  --tensor-parallel-size 4 \
  --expert-parallel-size 4