vLLM における LMCache のプリフィル高速化の評価

# Standard vLLM serving
vllm serve Qwen/Qwen3-8B

# LMCache-enabled serving
##### lmcache_config.yaml
chunk_size: 256
local_cpu: true
max_local_cpu_size: 50
#####
LMCACHE_CONFIG_FILE=/root/lmcache_config.yaml vllm serve /root/Qwen3-8B \
--kv-transfer-config '{"kv_connector":"LMCacheConnectorV1", "kv_role":"kv_both"}'

# Multi-turn bench scripts
# Ref: https://github.com/vllm-project/vllm/tree/main/benchmarks/multi_turn

##### generate_multi_turn.json
{
    "filetype": "generate_conversations",
    "num_conversations": 24,
    "text_files": ["pg1184.txt"],
    "print_stats": false,
    "prompt_input": {
        "num_turns": {
            "distribution": "uniform",
            "min": 12,
            "max": 18
        },
        "common_prefix_num_tokens": {
            "distribution": "constant",
            "value": 500
        },
        "prefix_num_tokens": {
            "distribution": "lognormal",
            "average": 4000,
            "max": 20000
        },
        "num_tokens": {
            "distribution": "uniform",
            "min": 120,
            "max": 160
        }
    },
    "prompt_output": {
        "num_tokens": {
            "distribution": "uniform",
            "min": 80,
            "max": 120
        }
    }
}
#####

python benchmark_serving_multi_turn.py --model $MODEL_PATH --input-file generate_multi_turn.json --num-clients 10 --max-active-conversations 10