| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """This file demonstrates the example usage of remote KV cache sharing |
| 4 | with LMCache. |
| 5 | We will launch 2 vllm instances, and launch an additional LMCache server. |
| 6 | KV cache is transferred in the following manner: |
| 7 | (1) vLLM instance 1 -> LMCache server (KV cache store). |
| 8 | (2) LMCache server -> vLLM instance 2 (KV cache reuse/retrieve). |
| 9 | |
| 10 | Note that lmcache needs to be installed to run this example. |
| 11 | Learn more about LMCache in https://github.com/LMCache/LMCache. |
| 12 | """ |
| 13 | |
| 14 | import os |
| 15 | import subprocess |
| 16 | import time |
| 17 | from multiprocessing import Event, Process |
| 18 | |
| 19 | from lmcache.integration.vllm.utils import ENGINE_NAME |
| 20 | from lmcache.v1.cache_engine import LMCacheEngineBuilder |
| 21 | |
| 22 | from vllm import LLM, SamplingParams |
| 23 | from vllm.config import KVTransferConfig |
| 24 | |
| 25 | # LMCache-related environment variables |
| 26 | # The port to start LMCache server |
| 27 | port = 8100 |
| 28 | # LMCache is set to use 256 tokens per chunk |
| 29 | os.environ["LMCACHE_CHUNK_SIZE"] = "256" |
| 30 | # Disable local CPU backend in LMCache |
| 31 | os.environ["LMCACHE_LOCAL_CPU"] = "False" |
| 32 | # Set local CPU memory buffer limit to 5.0 GB |
| 33 | os.environ["LMCACHE_MAX_LOCAL_CPU_SIZE"] = "5.0" |
| 34 | # Set the remote URL for LMCache server |
| 35 | os.environ["LMCACHE_REMOTE_URL"] = f"lm://localhost:{port}" |
| 36 | # Set the serializer/deserializer between vllm and LMCache server |
| 37 | # `naive` indicates using raw bytes of the tensor without any compression |
| 38 | os.environ["LMCACHE_REMOTE_SERDE"] = "naive" |
| 39 | |
| 40 | prompts = [ |
| 41 | "Hello, how are you?" * 1000, |
| 42 | ] |
| 43 | |
| 44 | |
| 45 | def run_store(store_done, prompts): |
| 46 | # We use GPU 0 for KV cache store process. |
| 47 | os.environ["CUDA_VISIBLE_DEVICES"] = "0" |
| 48 | |
| 49 | sampling_params = SamplingParams(temperature=0, top_p=0.95, max_tokens=10) |
| 50 | |
| 51 | ktc = KVTransferConfig(kv_connector="LMCacheConnectorV1", kv_role="kv_both") |
| 52 | # Set GPU memory utilization to 0.8 for an A40 GPU with 40GB |
| 53 | # memory. Reduce the value if your GPU has less memory. |
| 54 | llm = LLM( |
| 55 | model="mistralai/Mistral-7B-Instruct-v0.2", |
| 56 | kv_transfer_config=ktc, |
| 57 | max_model_len=8000, |
| 58 | gpu_memory_utilization=0.8, |
| 59 | enforce_eager=True, |
| 60 | ) |
| 61 | |
| 62 | outputs = llm.generate(prompts, sampling_params) |
| 63 | for output in outputs: |
| 64 | generated_text = output.outputs[0].text |
| 65 | print(f"Generated text: {generated_text!r}") |
| 66 | print("KV cache store is finished.") |
| 67 | store_done.set() |
| 68 | |
| 69 | # Clean up lmcache backend |
| 70 | LMCacheEngineBuilder.destroy(ENGINE_NAME) |
| 71 | |
| 72 | |
| 73 | def run_retrieve(store_done, prompts, timeout=1): |
| 74 | # We use GPU 1 for KV cache retrieve process. |
| 75 | os.environ["CUDA_VISIBLE_DEVICES"] = "1" |
| 76 | |
| 77 | sampling_params = SamplingParams(temperature=0, top_p=0.95, max_tokens=10) |
| 78 | |
| 79 | ktc = KVTransferConfig(kv_connector="LMCacheConnectorV1", kv_role="kv_both") |
| 80 | # Set GPU memory utilization to 0.8 for an A40 GPU with 40GB |
| 81 | # of memory. Reduce the value if your GPU has less memory. |
| 82 | llm = LLM( |
| 83 | model="mistralai/Mistral-7B-Instruct-v0.2", |
| 84 | kv_transfer_config=ktc, |
| 85 | max_model_len=8000, |
| 86 | gpu_memory_utilization=0.8, |
| 87 | enforce_eager=True, |
| 88 | ) |
| 89 | |
| 90 | print("Waiting for KV cache store to finish...") |
| 91 | store_done.wait() |
| 92 | time.sleep(timeout) |
| 93 | |
| 94 | outputs = llm.generate(prompts, sampling_params) |
| 95 | for output in outputs: |
| 96 | generated_text = output.outputs[0].text |
| 97 | print(f"Generated text: {generated_text!r}") |
| 98 | |
| 99 | # Clean up lmcache backend |
| 100 | LMCacheEngineBuilder.destroy(ENGINE_NAME) |
| 101 | |
| 102 | |
| 103 | def run_lmcache_server(port): |
| 104 | server_proc = subprocess.Popen( |
| 105 | ["python", "-m", "lmcache.v1.server", "localhost", str(port)] |
| 106 | ) |
| 107 | return server_proc |
| 108 | |
| 109 | |
| 110 | def main(): |
| 111 | store_done = Event() |
| 112 | store_process = Process(target=run_store, args=(store_done, prompts)) |
| 113 | retrieve_process = Process(target=run_retrieve, args=(store_done, prompts)) |
| 114 | lmcache_server_process = run_lmcache_server(port) |
| 115 | |
| 116 | # Start KV cache store process |
| 117 | store_process.start() |
| 118 | |
| 119 | # Start KV cache retrieve process |
| 120 | retrieve_process.start() |
| 121 | |
| 122 | # Clean up the processes |
| 123 | store_process.join() |
| 124 | retrieve_process.terminate() |
| 125 | lmcache_server_process.terminate() |
| 126 | lmcache_server_process.wait() |
| 127 | |
| 128 | |
| 129 | if __name__ == "__main__": |
| 130 | main() |