coreml_config.json

{
  "model_name": "Decision-2.0-Kai-0.6B",
  "source": "vllm-sr/Decision-2.0-Kai-0.6B",
  "source_revision": "cd49ea3813fd8ba0928a9a23ef6c9a0f2f0cd764",
  "package": "Decision2KaiPacked.mlpackage",
  "functions": {"L256_N32": 14.3, "L512_N64": 27.8, "L1024_N128": 58.1, "L2048_N256": 142.3},
  "functions_note": "function name -> measured predict p50 ms on M5 Pro GPU, used to choose chunking",
  "pad_token_id": 151643,
  "inputs": {
    "input_ids": "[1, L] int32: shared prefix, then each question's suffix, then padding",
    "position_ids": "[1, L] int32: RoPE positions; a suffix continues from the prefix length",
    "mask": "[1, 1, L, L] float16 additive: 0 where allowed, -1e4 elsewhere",
    "cand_idx": "[N] int32: packed index of each option's last token",
    "query_idx": "[N] int32: packed index of that question's final token"
  },
  "outputs": {"logits": "[N] float32: one raw logit per option (before Score offsets and softmax)"}
}