input:
  shape: [1, "T"]
  dtype: int32
  name: tokens
  preprocessing:
    - apply_chat_template: deepseek_r1   # <|begin_of_sentence|>{system}<|User|>{prompt}<|Assistant|><think>\n
    - tokenize:
        tokenizer: embedded in the .gguf (tokenizer.ggml.*, BPE / gpt2 model, qwen2 pre-tokenizer)
        add_bos: true                    # bos id 151646
  notes: |
    `T` is the runtime sequence length. The GGUF carries its own vocab, merges
    and chat template, so no external tokenizer files are needed — llama.cpp
    tokenizes the prompt string internally. `[1, 4096]` in the run report is the
    pipeline's nominal export shape, not a fixed graph input: the trained
    context is 131072 tokens and the served window is whatever `n_ctx` the
    runtime is started with (4096 in the shipped recipe).

output:
  format: "Token ids generated one at a time; decoded to text by llama.cpp"
  postprocessing:
    decode:
      tokenizer: embedded in the .gguf
    reasoning_trace:
      opening_tag: "<think>"             # id 151648, emitted by the generation prompt
      closing_tag: "</think>"            # id 151649
      note: "final answer is the span after </think>"
    stop_on:
      - eos_token                        # id 151643 (<|end_of_sentence|>)
      - max_tokens

generation:
  # Sampling settings the reported MATH-500 pass@1 was measured under.
  temperature: 0.6
  top_p: 0.95
  max_tokens: 12288
  # n_ctx below is the shipped serving window, sized for the phone's KV-cache
  # budget. The MATH-500 evaluation ran at n_ctx 14336 instead — a 12288-token
  # generation budget does not fit in 4096.
  n_ctx: 4096
  n_threads: 4

model:
  format: gguf
  architecture: qwen2
  num_layers: 28
  hidden_size: 1536
  feed_forward_length: 8960
  num_attention_heads: 12
  num_kv_heads: 2
  rope_freq_base: 10000.0
  vocab_size: 151936
  trained_context_length: 131072

quantization:
  variant: Q4_K_M                        # general.file_type = 15 (MOSTLY_Q4_K_M)
  method: llama.cpp k-quant, imatrix-weighted
  default_tensor_type: Q4_K
  output_weight_type: Q6_K               # llama-quantize's own Q4_K_M default
  token_embd_type: Q4_K                  # ditto
  fp32_tensors: norms and attention biases
  imatrix_entries: 196
  imatrix_chunks: 256
  source_precision: f16

runtime:
  name: llama.cpp
  execution_backend: cpu
  target_hardware: vivo_x300
  build_flags: -march=armv8.6-a+dotprod+i8mm, GGML_CPU_KLEIDIAI=ON
  kleidiai_engaged: false                # no Q4_K tensor is eligible; ggml's own repack path runs instead