Examples vLLM + LMCache w. CacheBlend

LMCache should be able to reduce the generation time of the second and following calls (even though the reused KV cache is not a prefix).

Some ad-hoc changes needed in vLLM

• In vllm/vllm/v1/worker/gpu_worker.py, comment out ensure_kv_transfer_initialized(vllm_config) in function def init_worker_distributed_environment.
• In the same file, add

from lmcache.v1.compute.models.utils import VLLMModelTracker
from lmcache.integration.vllm.utils import ENGINE_NAME
        
VLLMModelTracker.register_model(ENGINE_NAME, self.model_runner.model)
ensure_kv_transfer_initialized(self.vllm_config)

at the end of the function def load_model.

CPU offloading

• python blend.py - CacheBlend with CPU as backend

Disk offloading

• python blend.py --use-disk - CachBlend with local disk as backend