neurocache_config.json

{
  "attention_layers": [
    22,
    23,
    24,
    25
  ],
  "auto_mapping": null,
  "base_model_name_or_path": "meta-llama/Llama-3.2-3B",
  "cache_dtype": "float16",
  "cache_size": 32768,
  "cache_type": "FIFO",
  "compression_factor": 6,
  "context_size": 1,
  "global_cache": false,
  "inference_mode": true,
  "neighborhood_size": 2,
  "neurocache_type": "ONDEVICE",
  "pooling_window": 8,
  "retrieval_map": {
    "22": 22
  },
  "retrieval_stride": 2,
  "similarity_fn": "l2",
  "task_type": "CAUSAL_LM",
  "topk": 4
}