| 1 | # LMCache Examples |
| 2 | |
| 3 | This folder demonstrates how to use LMCache with vLLM v1 for KV cache |
| 4 | offloading, disaggregated prefilling, and KV cache sharing. |
| 5 | |
| 6 | ## Integration modes |
| 7 | |
| 8 | LMCache integrates with vLLM v1 in two ways: |
| 9 | |
| 10 | - **In-process mode** (`LMCacheConnectorV1`): LMCache runs inside the vLLM |
| 11 | process and is configured through environment variables or a YAML config |
| 12 | file (`LMCACHE_CONFIG_FILE`). This is the simplest way to add single-node |
| 13 | CPU/disk offloading. |
| 14 | - **Multi-process (MP) mode** (`LMCacheMPConnector`): LMCache runs as a |
| 15 | standalone server (`lmcache server`) that owns the KV cache storage; one or |
| 16 | more vLLM instances connect to it. This is the recommended mode for |
| 17 | distributed KV storage and for sharing KV cache across instances. See the |
| 18 | [LMCache docs](https://docs.lmcache.ai) for the full MP setup. |
| 19 | |
| 20 | ## 1. CPU offload (in-process) |
| 21 | |
| 22 | - `python cpu_offload_lmcache.py` - CPU offloading with `LMCacheConnectorV1` |
| 23 | for vLLM v1. |
| 24 | |
| 25 | ## 2. CPU offload (multi-process) |
| 26 | |
| 27 | - `bash cpu_offload_lmcache_mp.sh` - CPU offloading with `LMCacheMPConnector`, |
| 28 | using a standalone `lmcache server`. vLLM provides a built-in shortcut for |
| 29 | this setup via `--kv-offloading-backend lmcache` and |
| 30 | `--kv-offloading-size <GiB>`. |
| 31 | |
| 32 | ## 3. Disaggregated Prefill in vLLM v1 |
| 33 | |
| 34 | This example demonstrates how to run LMCache with disaggregated prefill using |
| 35 | NIXL on a single node. |
| 36 | |
| 37 | ### Prerequisites |
| 38 | |
| 39 | - Install [LMCache](https://github.com/LMCache/LMCache). You can simply run `pip install lmcache`. |
| 40 | - Install [NIXL](https://github.com/ai-dynamo/nixl). |
| 41 | - At least 2 GPUs |
| 42 | - Valid Hugging Face token (HF_TOKEN) for Llama 3.1 8B Instruct. |
| 43 | |
| 44 | ### Usage |
| 45 | |
| 46 | Run |
| 47 | `cd disagg_prefill_lmcache_v1` |
| 48 | to get into `disagg_prefill_lmcache_v1` folder, and then run |
| 49 | |
| 50 | ```bash |
| 51 | bash disagg_example_nixl.sh |
| 52 | ``` |
| 53 | |
| 54 | to run disaggregated prefill and benchmark the performance. |
| 55 | |
| 56 | ### Components |
| 57 | |
| 58 | #### Server Scripts |
| 59 | |
| 60 | - `disagg_prefill_lmcache_v1/disagg_vllm_launcher.sh` - Launches individual vLLM servers for prefill/decode, and also launches the proxy server. |
| 61 | - `disagg_prefill_lmcache_v1/disagg_proxy_server.py` - FastAPI proxy server that coordinates between prefiller and decoder |
| 62 | - `disagg_prefill_lmcache_v1/disagg_example_nixl.sh` - Main script to run the example |
| 63 | |
| 64 | #### Configuration |
| 65 | |
| 66 | - `disagg_prefill_lmcache_v1/configs/lmcache-prefiller-config.yaml` - Configuration for prefiller server |
| 67 | - `disagg_prefill_lmcache_v1/configs/lmcache-decoder-config.yaml` - Configuration for decoder server |
| 68 | |
| 69 | #### Log Files |
| 70 | |
| 71 | The main script generates several log files: |
| 72 | |
| 73 | - `prefiller.log` - Logs from the prefill server |
| 74 | - `decoder.log` - Logs from the decode server |
| 75 | - `proxy.log` - Logs from the proxy server |
| 76 | |
| 77 | ## 4. KV Cache Sharing |
| 78 | |
| 79 | The `kv_cache_sharing_lmcache_v1.py` example demonstrates how to share KV |
| 80 | caches between vLLM v1 instances through a centralized LMCache server. |