{
  "model": "/root/models/qwen",
  "context_tokens": 7680,
  "block_size": 256,
  "layers": 24,
  "blocks_per_layer": 30,
  "blocks_total": 720,
  "kv_blocks_paged_to_cpu": 720,
  "direct_reuse_max_logit_diff": 0.65625,
  "direct_next_token_same": true,
  "paged_vs_direct_max_logit_diff": 0.0,
  "max_logit_diff": 0.65625,
  "next_token_same": true,
  "full_token": " we",
  "paged_token": " we",
  "gpu_memory_allocated_mib": 3370.4,
  "gpu_memory_reserved_mib": 3436.0,
  "caveats": [
    "Reference KV paging on a 0.5B HF model, not vLLM connector integration.",
    "Single context/query; no concurrency.",
    "Blocks are paged CPU<->GPU manually; attention reconstruction is exact for this cache."
  ]
}