{
  "format": "damnit.inference-benchmark",
  "schema_version": 1,
  "record_id": "20261112T192745Z-inference",
  "generated_at": "2026-11-12T19:27:45.091Z",
  "kind": "inference-decode",
  "status": "PASS",
  "engine": {
    "name": "llama.cpp",
    "build_commit": "cd959bae9",
    "binary": "C:/Users/aaron/llama-bin/cpu/llama-bench.exe"
  },
  "hardware": {
    "platform": "win32",
    "arch": "x64",
    "release": "10.0.26300",
    "hostname": "LAPTOP-IRK4S2RH",
    "cpu_model": "11th Gen Intel(R) Core(TM) i3-1115G4 @ 3.00GHz",
    "logical_cpus": 4,
    "total_memory_bytes": 8379490304
  },
  "model": {
    "path": "C:/Users/aaron/models/qwen2.5-3b-instruct-q4_k_m.gguf",
    "bytes": 1929903264,
    "sha256": "9c9f56a391a3abbd5b89d0245bf6106081bcc3173119d4229235dd9d23253f94",
    "n_params": 3085938688
  },
  "run": {
    "threads": 4,
    "rounds": 3,
    "prompt_tokens": 512,
    "gen_tokens": 128,
    "command": "C:/Users/aaron/llama-bin/cpu/llama-bench.exe -m C:/Users/aaron/models/qwen2.5-3b-instruct-q4_k_m.gguf -t 4 -p 512 -n 128 -r 3 -o json"
  },
  "results": {
    "prefill_tokens_per_second": 34.933841,
    "prefill_stddev": 6.362619,
    "decode_tokens_per_second": 6.34733,
    "decode_stddev": 0.517095,
    "samples": 3
  },
  "bandwidth": {
    "bytes_per_token": 1929903264,
    "achieved_gbps": 12.25,
    "kernel_gbps": 23,
    "kernel_record": "20261111T195524Z-laptop-irk4s2rh",
    "utilization": 0.5326
  },
  "raw": {
    "exit_code": 0,
    "stdout": "[\r\n  {\r\n    \"build_commit\": \"cd959bae9\",\r\n    \"build_number\": 12110,\r\n    \"cpu_info\": \"11th Gen Intel(R) Core(TM) i3-1115G4 @ 3.00GHz\",\r\n    \"gpu_info\": \"\",\r\n    \"backends\": \"CPU\",\r\n    \"model_filename\": \"C:/Users/aaron/models/qwen2.5-3b-instruct-q4_k_m.gguf\",\r\n    \"model_type\": \"qwen2 3B Q4_K - Medium\",\r\n    \"model_size\": 1923946496,\r\n    \"model_n_params\": 3085938688,\r\n    \"n_batch\": 2048,\r\n    \"n_ubatch\": 512,\r\n    \"n_threads\": 4,\r\n    \"cpu_mask\": \"0x0\",\r\n    \"cpu_strict\": false,\r\n    \"poll\": 50,\r\n    \"type_k\": \"f16\",\r\n    \"type_v\": \"f16\",\r\n    \"n_gpu_layers\": -1,\r\n    \"n_cpu_moe\": 0,\r\n    \"split_mode\": \"layer\",\r\n    \"main_gpu\": 0,\r\n    \"no_kv_offload\": false,\r\n    \"flash_attn\": -1,\r\n    \"devices\": \"auto\",\r\n    \"tensor_split\": \"0.00\",\r\n    \"tensor_buft_overrides\": \"none\",\r\n    \"load_mode\": \"auto\",\r\n    \"lazy_mode\": \"auto\",\r\n    \"embeddings\": false,\r\n    \"no_op_offload\": 0,\r\n    \"no_host\": false,\r\n    \"repack\": true,\r\n    \"fit_target\": 0,\r\n    \"fit_min_ctx\": 0,\r\n    \"n_prompt\": 512,\r\n    \"n_gen\": 0,\r\n    \"n_depth\": 0,\r\n    \"test_time\": \"2026-11-12T19:25:36Z\",\r\n    \"avg_ns\": 15011598866,\r\n    \"stddev_ns\": 4224362326,\r\n    \"avg_ts\": 34.933841,\r\n    \"stddev_ts\": 6.362619,\r\n    \"samples_ns\": [ 18301169400, 14078086600, 12655540600 ],\r\n    \"samples_ts\": [ 27.9764, 36.3686, 40.4566 ]\r\n  },\r\n  {\r\n    \"build_commit\": \"cd959bae9\",\r\n    \"build_number\": 12110,\r\n    \"cpu_info\": \"11th Gen Intel(R) Core(TM) i3-1115G4 @ 3.00GHz\",\r\n    \"gpu_info\": \"\",\r\n    \"backends\": \"CPU\",\r\n    \"model_filename\": \"C:/Users/aaron/models/qwen2.5-3b-instruct-q4_k_m.gguf\",\r\n    \"model_type\": \"qwen2 3B Q4_K - Medium\",\r\n    \"model_size\": 1923946496,\r\n    \"model_n_params\": 3085938688,\r\n    \"n_batch\": 2048,\r\n    \"n_ubatch\": 512,\r\n    \"n_threads\": 4,\r\n    \"cpu_mask\": \"0x0\",\r\n    \"cpu_strict\": false,\r\n    \"poll\": 50,\r\n    \"type_k\": \"f16\",\r\n    \"type_v\": \"f16\",\r\n    \"n_gpu_layers\": -1,\r\n    \"n_cpu_moe\": 0,\r\n    \"split_mode\": \"layer\",\r\n    \"main_gpu\": 0,\r\n    \"no_kv_offload\": false,\r\n    \"flash_attn\": -1,\r\n    \"devices\": \"auto\",\r\n    \"tensor_split\": \"0.00\",\r\n    \"tensor_buft_overrides\": \"none\",\r\n    \"load_mode\": \"auto\",\r\n    \"lazy_mode\": \"auto\",\r\n    \"embeddings\": false,\r\n    \"no_op_offload\": 0,\r\n    \"no_host\": false,\r\n    \"repack\": true,\r\n    \"fit_target\": 0,\r\n    \"fit_min_ctx\": 0,\r\n    \"n_prompt\": 0,\r\n    \"n_gen\": 128,\r\n    \"n_depth\": 0,\r\n    \"test_time\": \"2026-11-12T19:26:41Z\",\r\n    \"avg_ns\": 20254888133,\r\n    \"stddev_ns\": 3451113735,\r\n    \"avg_ts\": 6.347330,\r\n    \"stddev_ts\": 0.517095,\r\n    \"samples_ns\": [ 20300810300, 18593259200, 21870594900 ],\r\n    \"samples_ts\": [ 6.30517, 6.88422, 5.85261 ]\r\n  }\r\n]\r\n",
    "stderr": "load_backend: loaded RPC backend from C:\\Users\\aaron\\llama-bin\\cpu\\ggml-rpc.dll\r\nload_backend: loaded CPU backend from C:\\Users\\aaron\\llama-bin\\cpu\\ggml-cpu-icelake.dll\r\n"
  },
  "limits": [
    "Decode is measured with llama.cpp, a standard engine, not the Damnit kernel.",
    "The ceiling comes from the published full-file kernel scan on the same machine, not from an inference run.",
    "A dense model reads every weight once per token; KV cache and attention reads are small at these contexts but are not zero."
  ],
  "problems": [],
  "elapsed_ms": 132667
}
