.. code:: bash

   docker run \
       -it \
       --shm-size 32g \
       --gpus all \
       -v {Huggingface-Cache-Path}:/root/.cache \
       --ipc=host \
       --network=host \
       --privileged \
       --name sglang_{your-name} \
       lmsysorg/sglang:dev \
       /bin/zsh

If you need to restart after exiting the container:

.. code:: bash

   docker start -i sglang_{your-name}

Update Python and Configure the Virtual Environment using uv

.. code:: bash

   cd ~
   git clone https://github.com/volcengine/verl.git
   cd verl

   # Install verl
   python3 -m uv pip install .
   python3 -m uv pip install -r ./requirements_sglang.txt

   # Manually install flash-attn
   python3 -m uv pip install wheel
   python3 -m uv pip install packaging
   python3 -m uv pip install flash-attn --no-build-isolation --no-deps

Set Up a Local Retrieval Engine

The local retrieval files are large—prepare sufficient disk space.
Downloading is about 60–70 GB, and uncompressed takes about 132 GB:

.. code:: bash

   conda activate retriever

   save_path=/the/path/to/save
   python examples/sglang_multiturn/search_r1_like/local_dense_retriever/download.py --save_path $save_path
   cat $save_path/part_* > $save_path/e5_Flat.index
   gzip -d $save_path/wiki-18.jsonl.gz

Start the Local flat e5 Retrieval Server

.. code:: bash

   export WANDB_API_KEY={YOUR_WANDB_API_KEY}

   # Define a timestamp function
   function now() {
       date '+%Y-%m-%d-%H-%M'
   }

**Preprocess the Dataset**

.. code:: bash

   # Ensure the now() function is defined
   # Create a logs directory
   mkdir -p logs

   # Set GPUs and run with a suitable log path
   export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7

   nohup bash examples/sglang_multiturn/search_r1_like/run_qwen2.5-3b_instruct_search_multiturn.sh \
     trainer.experiment_name=qwen2.5-3b-it_rm-searchR1-like-sgl-multiturn-$(now) \
     > logs/searchR1-like$(now).log 2>&1 &

Custom Search Configuration
---------------------------

To enable multi-turn reasoning, set the following fields in your config:

.. code:: yaml

   actor_rollout_ref:
     rollout:
       name: "sglang"
       multi_turn:
         enable: True

You must specify ``retrieval_service_url`` in ``examples/sglang_multiturn/config/tool_config/search_tool_config.yaml``, and properly configure concurrency. For more details on concurrency, refer to the Sandbox Fusion example:

.. code:: yaml

   tools:
     - class_name: verl.tools.search_tool.SearchTool
       config:
         retrieval_service_url: http://127.0.0.1:8000/retrieve
         num_workers: 120
         rate_limit: 120
         timeout: 30

The retriever input/output formats are as follows. If your service
parameters match, only modify ``retrieval_service_url``. You can also
customize in ``search_r1_like_utils.py``.

.. code:: python

   Input format:
   {
     "queries": ["What is Python?", "Tell me about neural networks."],
     "topk": 3,
     "return_scores": true
   }

   Output format (when return_scores=True, similarity scores are returned):
   {
       "result": [
           [   # Results for each query
               {
                   "document": doc, "score": score
               },
               # ... more documents
           ],
           # ... results for other queries
       ]
   }

Notes
-----

1. The total training time is about 27 hours; meanwhile, the validation
   dataset is very large (51 k), and each validation takes about 6000 s.
   (Therefore, ``val_before_train=False`` by default)