Example: Disaggregated prefill#

With LMCache as a KV cache transfer library, we can run disaggregated prefill with vLLM. Right now, LMCache uses NIXL as a transport layer to enable fast KV cache transfer via NVLink, RDMA, or TCP.

This guide demonstrates how to run LMCache with disaggregated prefill using a single prefiller and decoder setup (1P1D) on a single machine. The architecture splits the LLM inference into two stages: prefill and decode, running on separate GPUs for better resource utilization.

Prerequisites#

Before you begin, ensure you have:

  • At least 2 GPUs

  • Python packages installed:

    • lmcache (0.2.1 or above)

    • nixl (Install instructions here)

    • vllm (latest main branch)

    • httpx, fastapi, and uvicorn

  • A valid Hugging Face token (HF_TOKEN) with access to Llama 3.1 8B models

  • (Recommended) A machine with NVLink or RDMA enabled GPUs

Note

You can use ucx_perftest to check the GPU-GPU memory transfer and verify the NVLink or RDMA connection. Please refer to this link: UCX Performance Test.

Architecture Overview#

The disaggregated prefill setup consists of three main components:

  1. Prefiller Server (Port 8100): Handles the prefill phase of LLM inference

  2. Decoder Server (Port 8200): Manages the decoding/generation phase

  3. Proxy Server (Port 9000): Coordinates between prefiller and decoder

Configuration#

  1. Prefiller Server Configuration (lmcache-prefiller-config.yaml):

    # Disable CPU offloading since we're using NIXL for transfer
local_cpu: False
max_local_cpu_size: 0
max_local_disk_size: 0
remote_serde: NULL

# NIXL configuration for KV cache transfer
enable_nixl: True
nixl_role: "sender"          # Prefiller acts as KV cache sender
nixl_receiver_host: "localhost"  # Host where decoder is running
nixl_receiver_port: 55555        # Port where decoder is listening
nixl_buffer_size: 1073741824  # 1GB buffer for KV cache transfer
nixl_buffer_device: "cuda"   # Use GPU memory for buffer
nixl_enable_gc: True         # Enable garbage collection

  2. Decoder Server Configuration (lmcache-decoder-config.yaml):

    # Disable CPU offloading since we're using NIXL for transfer
local_cpu: False
max_local_cpu_size: 0
max_local_disk_size: 0
remote_serde: NULL

# NIXL configuration for KV cache transfer
enable_nixl: True
nixl_role: "receiver"        # Decoder acts as KV cache receiver
nixl_receiver_host: "localhost"  # Host where decoder is listening
nixl_receiver_port: 55555        # Port where decoder is listening
nixl_buffer_size: 1073741824  # 1GB buffer for KV cache transfer
nixl_buffer_device: "cuda"   # Use GPU memory for buffer
nixl_enable_gc: True         # Enable garbage collection

Step-by-Step Setup#

  1. Environment Setup

    Set your Hugging Face token before running the vLLM servers.

    export HF_TOKEN=your_hugging_face_token

  2. Launch the vLLM + LMCache Inference Servers

    You can launch the components individually:

    1. Launch Decoder (on GPU 1):

      UCX_TLS=cuda_ipc,cuda_copy,tcp \
    LMCACHE_CONFIG_FILE=lmcache-decoder-config.yaml \
    CUDA_VISIBLE_DEVICES=1 \
    vllm serve meta-llama/Llama-3.1-8B-Instruct \
    --port 8200 \
    --disable-log-requests \
    --kv-transfer-config \
    '{"kv_connector":"LMCacheConnectorV1","kv_role":"kv_consumer","kv_connector_extra_config": {"discard_partial_chunks": false, "lmcache_rpc_port": "consumer1"}}'

    2. Launch Prefiller (on GPU 0):

      UCX_TLS=cuda_ipc,cuda_copy,tcp \
    LMCACHE_CONFIG_FILE=lmcache-prefiller-config.yaml \
    CUDA_VISIBLE_DEVICES=0 \
    vllm serve meta-llama/Llama-3.1-8B-Instruct \
    --port 8100 \
    --disable-log-requests \
    --kv-transfer-config \
    '{"kv_connector":"LMCacheConnectorV1","kv_role":"kv_producer","kv_connector_extra_config": {"discard_partial_chunks": false, "lmcache_rpc_port": "producer1"}}'

    3. Launch a proxy server to coordinate between prefiller and decoder:

      The code for the proxy server is available in vLLM repo.

      wget https://raw.githubusercontent.com/vllm-project/vllm/main/examples/lmcache/disagg_prefill_lmcache_v1/disagg_proxy_server.py

python3 disagg_proxy_server.py \
    --host localhost \
    --port 9000 \
    --prefiller-host localhost \
    --prefiller-port 8100 \
    --decoder-host localhost \
    --decoder-port 8200

Note

The UCX_TLS environment variable is used to specify the transport layer for UCX (the example uses NVLink) The CUDA_VISIBLE_DEVICES environment variable is used to specify the GPUs to use for the servers.

  1. Verify Setup

    The servers are ready when you can access:

    • Prefiller: http://localhost:8100/v1/completions

    • Decoder: http://localhost:8200/v1/completions

    • Proxy: http://localhost:9000/v1/completions

Usage#

Send requests to the proxy server (port 9000) using either the completions or chat completions endpoint:

curl http://localhost:9000/v1/completions \
    -H "Content-Type: application/json" \
    -d '{
        "model": "meta-llama/Llama-3.1-8B-Instruct",
        "prompt": "Tell me a story",
        "max_tokens": 100
    }'

You can also test the setup with the following command, which runs the benchmark_serving.py from vLLM.

git clone https://github.com/vllm-project/vllm.git
cd vllm/benchmarks
python benchmark_serving.py --port 9000 --seed $(date +%s) \
    --model meta-llama/Llama-3.1-8B-Instruct \
    --dataset-name random --random-input-len 5000 --random-output-len 200 \
    --num-prompts 50 --burstiness 100 --request-rate 1

Monitoring#

The prefiller instance will log the throughput of KV cache transfer:

LMCache INFO: Store 5271 tokens takes: 6.5000 ms, throughput: 98.9889 GB/s; offload_time: 2.6594 ms, put_time: 3.4539 ms (cache_engine.py:190:lmcache.v1.cache_engine)

The decoder instance will log how many tokens are fetched from the LMCache:

LMCache INFO: Reqid: cmpl-b8bf01cbe47e4d108732ceeb4158d310-0, Total tokens 5170, LMCache hit tokens: 5169, need to load: 5169 (vllm_v1_adapter.py:543:lmcache.integration.vllm.vllm_v1_adapter)

The proxy server will log the TTFT of the prefiller node:

===============================
Num requests: 49
Prefill node TTFT stats:
- Average (ms): 0.1530598815606565
- Median (ms): 0.15739011764526367
- 99th Percentile (ms): 0.1643616008758545
===============================

Troubleshooting#

Common issues and solutions:

  1. GPU Requirements: Ensure you have at least 2 GPUs available

  2. Port Conflicts: Check if ports 8100, 8200, and 9000 are available

  3. HF Token: Verify your token starts with hf_ and has necessary model access

  4. CUDA Errors: Ensure CUDA_VISIBLE_DEVICES is set correctly for each server