Example: Offload KV cache to CPU

In this example, we will show you how to offload KV cache to CPU memory.

Note

Besides CPU memory, LMCache also supports offloading KV cache to many different destinations. See Supported offloading destinations for more details.

Prerequisites

Before you begin, make sure you have:

• vLLM v1 with LMCache installed (see Installation)

• A GPU that can run a LLM

Use CPU offloading in offline inference

This section demonstrates how to use CPU memory offloading in offline inference scenarios using LMCache with vLLM. The example script we use here is available in vLLM examples. See the examples README to understand how to run the script for vLLM v1.

First, set up the necessary environment variables for LMCache:

import os

# Set token chunk size to 256
os.environ["LMCACHE_CHUNK_SIZE"] = "256"
# Enable CPU memory backend
os.environ["LMCACHE_LOCAL_CPU"] = "True"
# Set CPU memory limit to 5GB
os.environ["LMCACHE_MAX_LOCAL_CPU_SIZE"] = "5.0"

Next, configure vLLM with LMCache integration:

from vllm import LLM, SamplingParams
from vllm.config import KVTransferConfig

# Configure KV cache transfer to use LMCache
ktc = KVTransferConfig(
    kv_connector="LMCacheConnectorV1",
    kv_role="kv_both",
)

# Initialize LLM with LMCache configuration
# Adjust gpu_memory_utilization based on your GPU memory
llm = LLM(model="Qwen/Qwen3-8B",
          kv_transfer_config=ktc,
          max_model_len=8000,
          gpu_memory_utilization=0.8)

Now you can run inference with automatic KV cache offloading:

# Create example prompts with shared prefix
shared_prompt = "Hello, how are you?" * 1000
prompts = [
    shared_prompt + "Hello, my name is",
]

# Define sampling parameters
sampling_params = SamplingParams(temperature=0, top_p=0.95, max_tokens=10)

# Run inference
outputs = llm.generate(prompts, sampling_params)
for output in outputs:
    generated_text = output.outputs[0].text
    print(f"Generated text: {generated_text!r}")

When the inference is complete, clean up the LMCache backend:

from lmcache.v1.cache_engine import LMCacheEngineBuilder
from lmcache.integration.vllm.utils import ENGINE_NAME

LMCacheEngineBuilder.destroy(ENGINE_NAME)

During inference, LMCache will automatically handle storing and managing KV cache in CPU memory. You can monitor this through the logs, which will show messages like:

LMCache INFO: Storing KV cache for 6006 out of 6006 tokens for request 0

This indicates that the KV cache has been successfully offloaded to CPU memory.

Note

• Adjust gpu_memory_utilization based on your GPU’s available memory

• The CPU offloading buffer size can be adjusted through LMCACHE_MAX_LOCAL_CPU_SIZE

Use CPU offloading in online inference

This section demonstrates how to use CPU memory offloading in online serving scenarios.

First, create a configuration file named lmcache_config.yaml with the following content:

chunk_size: 256
local_cpu: true
max_local_cpu_size: 5

Note

LMCache supports extensive configuration through a lmcache_config.yaml file where you can customize chunk sizes, memory limits, storage backends, and more. We’ll cover advanced configuration options in later examples. For now, let’s run a minimal example with default configuration.

Launch the vLLM server with LMCache integration using environment variables. Here’s an example command:

LMCACHE_CONFIG_FILE=lmcache_config.yaml \
vllm serve \
    Qwen/Qwen3-8B \
    --kv-transfer-config \
    '{"kv_connector":"LMCacheConnectorV1",
      "kv_role":"kv_both"
    }'

Key parameters explained:

• LMCACHE_CONFIG_FILE: Path to the LMCache configuration file.

• --kv-transfer-config: Configures LMCache integration

  • kv_connector: Specifies the LMCache connector

  • kv_role: Set to “kv_both” for both storing and loading KV cache

Once the server is running, you can send requests to it using curl. Here’s an example of how to send a request to the vLLM server with LMCache integration:

curl http://localhost:8000/v1/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "Qwen/Qwen3-8B",
    "prompt": "<|im_start|>system\nYou are a helpful AI assistant.<|im_end|>\n<|im_start|>user\nWhat is the capital of France?<|im_end|>\n<|im_start|>assistant\n",
    "max_tokens": 100,
    "temperature": 0.7
  }'

You should see the following logs:

LMCache INFO: Storing KV cache for 31 out of 31 tokens for request cmpl-274bcaa80837444dbf9fbba4155d2620-0 (vllm_v1_adapter.py:497:lmcache.integration.vllm.vllm_v1_adapter)

Once you send the same curl request again, you should see the following logs:

LMCache INFO: Reqid: cmpl-4ddf8863a6ac4dc3b6a952f2a107e9b2-0, Total tokens 31, LMCache hit tokens: 30, need to load: 14 (vllm_v1_adapter.py:543:lmcache.integration.vllm.vllm_v1_adapter)

Example: CPU offloading benefits

This section demonstrates the performance benefits of using CPU offloading with LMCache. We’ll use a script that generates multiple prompts and compare the performance with and without LMCache.

Prerequisites (Setup)

• A CUDA GPU. The example picks a model that fits the GPU automatically:

  • Qwen/Qwen3-8B (bf16) when the GPU has ~36 GiB or more (e.g. A100-80G, H100).

  • Qwen/Qwen3-8B-FP8 with kv_cache_dtype="fp8" when the GPU has ~24 GiB and supports native FP8 (Ada Lovelace / Hopper, sm_89+; e.g. L4, L40, RTX 4090).

  • Qwen/Qwen3-1.7B as the fallback for smaller GPUs (~10 GiB and up), including Ampere 24 GiB cards (RTX A5000, RTX 3090) where FP8 is unsupported.

• Sufficient CPU memory. The example clamps the LMCache pinned host buffer to fit your system RAM and RLIMIT_MEMLOCK (ulimit -l), so it also works on smaller hosts without manual tuning.

Example script

Save the following script as cpu-offloading.py:

# SPDX-License-Identifier: Apache-2.0
"""
This file demonstrates the example usage of cpu offloading
with LMCache in vLLM v1.

Note that lmcache needs to be installed to run this example.
Learn more about LMCache in https://github.com/LMCache/LMCache.
"""
import os
import torch
import argparse
import time
from lmcache.v1.cache_engine import LMCacheEngineBuilder
from lmcache.integration.vllm.utils import ENGINE_NAME
from vllm import LLM, SamplingParams
from vllm.config import KVTransferConfig

def parse_arguments() -> argparse.Namespace:
    """Parse command line arguments."""
    parser = argparse.ArgumentParser(description="CPU offloading example with LMCache")
    parser.add_argument("--num-prompts", type=int, default=10,
                      help="Number of prompts to generate (default: 10)")
    parser.add_argument("--num-tokens", type=int, default=10000,
                      help="Number of tokens per prompt (default: 10000)")
    parser.add_argument("--enable-lmcache", action="store_true",
                      help="Enable LMCache for CPU offloading (default: True)")
    return parser.parse_args()

def pick_cpu_size_gb(workload_gb: float) -> float:
    """
    Clamp the LMCache pinned host buffer to fit system RAM and RLIMIT_MEMLOCK.

    cudaHostAlloc pins pages, so the buffer cannot exceed total RAM nor the
    per-process memlock limit (`ulimit -l`). On hosts where either is small,
    the original "1.5 GB per 10k tokens" formula fails with cudaErrorMemoryAllocation.

    Args:
        workload_gb: Desired buffer size for the workload, in GiB.
    Returns:
        float: A buffer size in GiB that fits both caps, never below 1.0.
    """
    import psutil

    ram_gib = psutil.virtual_memory().total / (1024 ** 3)
    try:
        import resource
        memlock_soft, _ = resource.getrlimit(resource.RLIMIT_MEMLOCK)
        memlock_gib = (
            float("inf")
            if memlock_soft == resource.RLIM_INFINITY
            else memlock_soft / (1024 ** 3)
        )
    except ImportError:
        # `resource` is POSIX-only; on Windows treat memlock as unbounded.
        memlock_gib = float("inf")
    return max(min(workload_gb, ram_gib * 0.5, memlock_gib * 0.9), 1.0)

def setup_lmcache_environment(num_prompts: int, num_tokens: int) -> None:
    """
    Configure LMCache environment variables.
    Args:
        num_prompts: Number of prompts to process
        num_tokens: Number of tokens per prompt
    """
    workload_gb = num_prompts * num_tokens * 1.5 / 10000  # 1.5 GB per 10k tokens
    cpu_size = pick_cpu_size_gb(workload_gb)

    env_vars = {
        "LMCACHE_CHUNK_SIZE": "256",         # Set tokens per chunk
        "LMCACHE_LOCAL_CPU": "True",         # Enable local CPU backend
        "LMCACHE_MAX_LOCAL_CPU_SIZE": str(cpu_size)  # CPU memory limit (GB)
    }
    for key, value in env_vars.items():
        os.environ[key] = value

def pick_model_and_kwargs() -> tuple[str, dict]:
    """
    Pick a Qwen model that fits the current GPU's memory and compute capability.

    Tiers:
        - >= 36 GiB                    -> Qwen/Qwen3-8B (bf16)
        - >= 20 GiB and sm >= 89       -> Qwen/Qwen3-8B-FP8 (native FP8)
        - >= 10 GiB                    -> Qwen/Qwen3-1.7B
        - otherwise                    -> RuntimeError

    Returns:
        tuple[str, dict]: (model id, extra kwargs to pass to ``LLM``).
    Raises:
        RuntimeError: If no CUDA GPU is visible or it is too small.
    """
    if not torch.cuda.is_available():
        raise RuntimeError("No GPU available")

    total_gib = torch.cuda.get_device_properties(0).total_memory / (1024 ** 3)
    major, minor = torch.cuda.get_device_capability(0)
    sm = major * 10 + minor
    has_fp8 = sm >= 89  # Ada Lovelace / Hopper

    if total_gib >= 36:
        return "Qwen/Qwen3-8B", {}
    if total_gib >= 20 and has_fp8:
        print(f"[fallback] GPU {total_gib:.1f} GiB sm_{sm}: using Qwen3-8B-FP8")
        return "Qwen/Qwen3-8B-FP8", {"kv_cache_dtype": "fp8"}
    if total_gib >= 10:
        print(f"[fallback] GPU {total_gib:.1f} GiB sm_{sm}: using Qwen3-1.7B")
        return "Qwen/Qwen3-1.7B", {}
    raise RuntimeError(
        f"GPU has {total_gib:.1f} GiB; need at least 10 GiB for Qwen3-1.7B"
    )

def create_test_prompts(num_prompts: int = 10, num_tokens: int = 1000) -> list[str]:
    """
    Create test prompts with index prefix and dummy body.
    Args:
        num_prompts: Number of prompts to generate
        num_tokens: Approximate number of tokens per prompt (using 'Hi ' as token unit)
    Returns:
        list: List of prompts with format '[index] Hi Hi Hi...'
    """
    prompts = []
    dummy_text = "Hi " * num_tokens

    for i in range(num_prompts):
        prompt = f"[Prompt {i}] {dummy_text} how are you?"
        prompts.append(prompt)

    return prompts

def initialize_llm(max_len: int = 16384, enable_lmcache: bool = True) -> LLM:
    """
    Initialize the LLM with a model auto-selected for the current GPU.
    Args:
        max_len: Maximum sequence length
        enable_lmcache: Whether to wire up the LMCache KV connector
    Returns:
        LLM: Configured LLM instance
    """
    model_name, extra_kwargs = pick_model_and_kwargs()

    ktc = KVTransferConfig(
        kv_connector="LMCacheConnectorV1",
        kv_role="kv_both",
    ) if enable_lmcache else None

    return LLM(
        model=model_name,
        kv_transfer_config=ktc,
        max_model_len=max_len,
        enable_prefix_caching=False,
        gpu_memory_utilization=0.9,
        **extra_kwargs,
    )

def generate_and_print_output(
    llm: LLM,
    prompts: list[str],
    sampling_params: SamplingParams,
) -> float:
    """
    Generate text and print the results.
    Args:
        llm: LLM instance
        prompts: List of input prompts
        sampling_params: Configured sampling parameters
    Returns:
        float: Time taken for generation in seconds
    """
    start_time = time.time()
    outputs = llm.generate(prompts, sampling_params)
    end_time = time.time()

    for output in outputs:
        generated_text = output.outputs[0].text
        print(f"Generated text: {generated_text!r}")

    return end_time - start_time

def main() -> None:
    """Main execution function."""
    # Parse command line arguments
    args = parse_arguments()

    # Setup environment if LMCache is enabled
    if args.enable_lmcache:
        setup_lmcache_environment(args.num_prompts, args.num_tokens)

    # Create prompts and sampling parameters
    prompts = create_test_prompts(num_prompts=args.num_prompts, num_tokens=args.num_tokens)
    sampling_params = SamplingParams(temperature=0, top_p=0.95, max_tokens=1)

    # Initialize model
    llm = initialize_llm(enable_lmcache=args.enable_lmcache)

    # First run
    print("\nFirst run:")
    first_run_time = generate_and_print_output(llm, prompts, sampling_params)
    print(f"First run time: {first_run_time:.2f} seconds")

    # Second run
    print("\nSecond run:")
    second_run_time = generate_and_print_output(llm, prompts, sampling_params)
    print(f"Second run time: {second_run_time:.2f} seconds")

    # Print speedup
    if first_run_time > 0:
        speedup = first_run_time / second_run_time
        print(f"\nSpeedup (first run / second run): {speedup:.2f}x")

    # Cleanup if LMCache was enabled
    if args.enable_lmcache:
        LMCacheEngineBuilder.destroy(ENGINE_NAME)

if __name__ == "__main__":
    main()

Running the Example

1. First, run the script without LMCache:

   python cpu-offloading.py
   

   You’ll see output like:

   Speedup (first run / second run): 1.00x
   

   Without LMCache, there’s no speedup between runs even if vLLM has prefix caching enabled. This is because the KV cache exceeds GPU memory and can’t be reused.

2. Now, run with LMCache enabled:

   python cpu-offloading.py --enable-lmcache
   

   You’ll see output like:

   Speedup (first run / second run): 7.43x
   

The significant speedup in the second case demonstrates how LMCache effectively manages KV cache offloading to CPU memory. When the total size of KV cache exceeds GPU memory, LMCache allows you to store and reuse the cache from CPU memory, resulting in much faster subsequent generations for prompts with shared prefixes.

Supported offloading destinations

LMCache now supports offloading KV cache to the following destinations:

• CPU memory

• Local file system

• Mooncake Storage

• InfiniStore

• Redis

• ValKey

Troubleshooting

If you encounter the following error:

(EngineCore_DP0 pid=55437) ERROR 10-04 14:44:47 [core.py:708] RuntimeError:
Cannot re-initialize CUDA in forked subprocess. To use CUDA with multiprocessing, you must use the 'spawn' start method

You can resolve this issue using one of the following methods:

• Set VLLM_WORKER_MULTIPROC_METHOD=spawn in the environment variables.

• Or update the Python code to guard usage of vllm behind a if __name__ == '__main__': block.

if __name__ == '__main__':
    from vllm import LLM, SamplingParams
    from vllm.config import KVTransferConfig
    from lmcache.v1.cache_engine import LMCacheEngineBuilder
    from lmcache.integration.vllm.utils import ENGINE_NAME
    main()

For details, please refer to the vLLM Troubleshooting Guide: Python multiprocessing.