xc-llm-ascend/docs/source/tutorials/pd_colocated_mooncake_multi_instance.md

PD-Colocated with Mooncake Multi-Instance

Getting Started

vLLM-Ascend now supports PD-colocated deployment with Mooncake features. This guide provides step-by-step instructions to test these features with constrained resources.

Using the Qwen2.5-72B-Instruct model as an example, this guide demonstrates how to use vllm-ascend v0.11.0 (with vLLM v0.11.0) on two Atlas 800T A2 nodes to deploy two vLLM instances. Each instance occupies 4 NPU cards and uses PD-colocated deployment.

Verify Multi-Node Communication Environment

Physical Layer Requirements

• The two Atlas 800T A2 nodes must be physically interconnected via a RoCE network. Without RoCE interconnection, cross-node KV Cache access performance will be significantly degraded.
• All NPU cards must communicate properly. Intra-node communication uses HCCS, while inter-node communication uses the RoCE network.

Verification Process

The following process serves as a reference example. Please modify parameters such as IP addresses according to your actual environment.

1. Single Node Verification:

   Execute the following commands sequentially. The results must all be success and the status must be UP:

   # Check the remote switch ports
   for i in {0..7}; do hccn_tool -i $i -lldp -g | grep Ifname; done
   # Get the link status of the Ethernet ports (UP or DOWN)
   for i in {0..7}; do hccn_tool -i $i -link -g ; done
   # Check the network health status
   for i in {0..7}; do hccn_tool -i $i -net_health -g ; done
   # View the network detected IP configuration
   for i in {0..7}; do hccn_tool -i $i -netdetect -g ; done
   # View gateway configuration
   for i in {0..7}; do hccn_tool -i $i -gateway -g ; done
   

2. Check NPU Network Configuration:

   Ensure that the hccn.conf file exists in the environment. If using Docker, mount it into the container.

   cat /etc/hccn.conf
   

3. Get NPU IP Addresses:

   for i in {0..7}; do hccn_tool -i $i -ip -g; done
   

4. Cross-Node PING Test:

   # Execute the following command on each node, replacing x.x.x.x
   # with the target node's NPU card address.
   for i in {0..7}; do hccn_tool -i $i -ping -g address x.x.x.x; done
   

Run with Docker

Start a Docker container on each node.

# Update the vllm-ascend image
export IMAGE=quay.io/ascend/vllm-ascend:v0.11.0
export NAME=vllm-ascend

# Run the container using the defined variables
# This test uses four NPU cards to create the container.
# Mount the hccn.conf file from the host node into the container.
docker run --rm \
--name $NAME \
--net=host \
--shm-size=1g \
--device /dev/davinci0 \
--device /dev/davinci1 \
--device /dev/davinci2 \
--device /dev/davinci3 \
--device /dev/davinci_manager \
--device /dev/devmm_svm \
--device /dev/hisi_hdc \
-v /usr/local/dcmi:/usr/local/dcmi \
-v /usr/local/Ascend/driver/tools/hccn_tool:\
/usr/local/Ascend/driver/tools/hccn_tool \
-v /usr/local/bin/npu-smi:/usr/local/bin/npu-smi \
-v /usr/local/Ascend/driver/lib64/:/usr/local/Ascend/driver/lib64/ \
-v /usr/local/Ascend/driver/version.info:/usr/local/Ascend/driver/version.info \
-v /etc/ascend_install.info:/etc/ascend_install.info \
-v /etc/hccn.conf:/etc/hccn.conf \
-v /root/.cache:/root/.cache \
-it $IMAGE bash

(Optional) Install Mooncake

Mooncake is pre-installed and functional in the v0.11.0 image. The following installation steps are optional.

Mooncake is the serving platform for Kimi, a leading LLM service provided by Moonshot AI. Installation and compilation guide: https://github.com/kvcache-ai/Mooncake?tab=readme-ov-file#build-and-use-binaries.

First, obtain the Mooncake project using the following command:

git clone -b v0.3.7.post2 --depth 1 https://github.com/kvcache-ai/Mooncake.git
cd Mooncake
git submodule update --init --recursive

Install MPI:

apt-get install mpich libmpich-dev -y

Install the relevant dependencies (Go installation is not required):

bash dependencies.sh -y

Compile and install:

mkdir build
cd build
cmake .. -DUSE_ASCEND_DIRECT=ON
make -j
make install

After installation, verify that Mooncake is installed correctly:

python -c "import mooncake; print(mooncake.__file__)"
# Expected output path:
# /usr/local/Ascend/ascend-toolkit/latest/python/
# site-packages/mooncake/__init__.py

Start Mooncake Master Service

To start the Mooncake master service in one of the node containers, use the following command:

docker exec -it vllm-ascend bash
cd /vllm-workspace/Mooncake
mooncake_master --port 50088 \
  --eviction_high_watermark_ratio 0.95 \
  --eviction_ratio 0.05

Parameter	Value	Explanation
port	50088	Port for the master service
eviction_high_watermark_ratio	0.95	High watermark ratio (95% threshold)
eviction_ratio	0.05	Percentage to evict when full (5%)

Create a Mooncake Configuration File Named mooncake.json

The template for the mooncake.json file is as follows:

{
    "metadata_server": "P2PHANDSHAKE",
    "protocol": "ascend",
    "device_name": "",
    "use_ascend_direct": true,
    "master_server_address": "<your_server_ip>:50088",
    "global_segment_size": 107374182400
}

Parameter	Value	Explanation
metadata_server	P2PHANDSHAKE	Point-to-point handshake mode
protocol	ascend	Ascend proprietary protocol
use_ascend_direct	true	Enable direct hardware access
master_server_address	90.90.100.188:50088(for example)	Master server address
global_segment_size	107374182400	Size per segment (100 GB)

vLLM Instance Deployment

Create containers on both Node 1 and Node 2, and launch the Qwen2.5-72B-Instruct model service in each to test the reusability and performance of cross-node, cross-instance KV Cache. Instance 1 utilizes NPU cards [0-3] on the first Atlas 800T A2 server, while Instance 2 utilizes cards [0-3] on the second server.

Deploy Instance 1

Replace file paths, host, and port parameters based on your actual environment configuration.

export LD_LIBRARY_PATH=/usr/local/Ascend/ascend-toolkit/\
latest/python/site-packages:$LD_LIBRARY_PATH
export MOONCAKE_CONFIG_PATH="/vllm-workspace/mooncake.json"
# NPU buffer pool: quantity:size(MB)
# Allocates 4 buffers of 8MB each for KV transfer
export ASCEND_BUFFER_POOL=4:8

vllm serve <path_to_your_model>/Qwen2.5-72B-Instruct/ \
--served-model-name qwen \
--dtype bfloat16 \
--max-model-len 25600 \
--tensor-parallel-size 4 \
--host <your_server_ip> \
--port 8002 \
--max-num-batched-tokens 4096 \
--gpu-memory-utilization 0.9 \
--kv-transfer-config '{
      "kv_connector": "MooncakeConnectorStoreV1",
      "kv_role": "kv_both",
      "kv_connector_extra_config": {
          "use_layerwise": false,
          "mooncake_rpc_port": "0",
          "load_async": true,
          "register_buffer": true
      }
  }'

Deploy Instance 2

The deployment method for Instance 2 is identical to Instance 1. Simply modify the --host and --port parameters according to your Instance 2 configuration.

Configuration Parameters

Parameter	Value	Explanation
kv_connector	MooncakeConnectorStoreV1	Use StoreV1 version
kv_role	kv_both	Enable both produce and consume
use_layerwise	false	Transfer entire cache (see note)
mooncake_rpc_port	0	Automatic port assignment
load_async	true	Enable asynchronous loading
register_buffer	true	Required for PD-colocated mode

Note on use_layerwise:

• false: Transfer entire KV Cache (suitable for cross-node with sufficient bandwidth)
• true: Layer-by-layer transfer (suitable for single-node memory constraints)

Benchmark

We recommend using the aisbench tool to assess performance. The test uses Dataset A, consisting of fully random data, with the following configuration:

• Input/output tokens: 1024/10
• Total requests: 100
• Concurrency: 25

The test procedure consists of three steps:

Step 1: Baseline (No Cache)

Send Dataset A to Instance 1 on Node 1 and record the Time to First Token (TTFT) as TTFT1.

Preparation for Step 2

Before Step 2, send a fully random Dataset B to Instance 1. Due to the unified HBM/DRAM KV Cache with LRU (Least Recently Used) eviction policy, Dataset B's cache evicts Dataset A's cache from HBM, leaving Dataset A's cache only in Node 1's DRAM.

Step 2: Local DRAM Hit

Send Dataset A to Instance 1 again to measure the performance when hitting the KV Cache in local DRAM. Record the TTFT as TTFT2.

Step 3: Cross-Node DRAM Hit

Send Dataset A to Instance 2. With the Mooncake KV Cache pool, this results in a cross-node KV Cache hit from Node 1's DRAM. Record the TTFT as TTFT3.

Model Configuration:

from ais_bench.benchmark.models import VLLMCustomAPIChatStream
from ais_bench.benchmark.utils.model_postprocessors import extract_non_reasoning_content

models = [
    dict(
        attr="service",
        type=VLLMCustomAPIChatStream,
        abbr='vllm-api-stream-chat',
        path="<path_to_your_model>/Qwen2.5-72B-Instruct",
        model="qwen",
        request_rate = 0,
        retry = 2,
        host_ip = "<your_server_ip>",
        host_port = 8002,
        max_out_len = 10,
        batch_size= 25,
        trust_remote_code=False,
        generation_kwargs = dict(
            temperature = 0,
            ignore_eos = True,
        ),
    )
]

Performance Benchmarking Commands：

ais_bench --models vllm_api_stream_chat \
  --datasets gsm8k_gen_0_shot_cot_str_perf \
  --debug --summarizer default_perf --mode perf

Test Results

Requests	Concur	TTFT1 (ms)	TTFT2 (ms)	TTFT3 (ms)
100	25	2322	739	948