{
  "schema_version": 1,
  "model": {
    "name": "llama-3.2-1b",
    "family": "llm",
    "dsp_arch": "v79",
    "tokenizer_pre": "llama3"
  },
  "chat": {
    "bos": "<|begin_of_text|>",
    "system": {
      "prefix": "<|start_header_id|>system<|end_header_id|>\n\n",
      "suffix": "<|eot_id|>"
    },
    "user": {
      "prefix": "<|start_header_id|>user<|end_header_id|>\n\n",
      "suffix": "<|eot_id|>"
    },
    "assistant": {
      "prefix": "<|start_header_id|>assistant<|end_header_id|>\n\n",
      "suffix": "<|eot_id|>"
    }
  },
  "params": {
    "hidden": 2048,
    "vocab": 128256,
    "n_layers": 16,
    "max_ctx": 512,
    "kv_dim": 512,
    "head_dim": 64,
    "eos_token_id": 128009
  },
  "artifacts": {
    "contexts": {
      "prefill": {
        "bin": "llama_full_wqo_o3.bin"
      },
      "decode": {
        "bin": "llama_dec_wqo_o3.bin"
      },
      "lmhead": {
        "bin": "llama_lmhead.bin"
      }
    },
    "embed": "llama_embed_f16.bin",
    "tokenizer": "tokenizer.json"
  },
  "plan": {
    "steps": [
      {
        "host": "llama_generate",
        "params": {
          "prefill": "llama_full_wqo",
          "decode": "llama_dec_wqo",
          "lmhead": "llama_lmhead_f16",
          "rope_factor": 32,
          "max_new": 64,
          "neg": -50000
        }
      }
    ]
  }
}