$ C:\quantdiff\llama.cpp\bin\llama-bench.exe -m C:\quantdiff\dl\campaign\22086870b009\Ornith-1.5-9B-Q8_0.gguf -ngl 999 -fa 1 -p 512 -n 128 -r 5 -t 8 -o json
# started 2026-10-22T05:52:00Z
ggml_cuda_init: found 1 CUDA devices (Total VRAM: 12199 MiB):
  Device 0: NVIDIA GeForce RTX 5070, compute capability 12.0, VMM: yes, VRAM: 12199 MiB
load_backend: loaded CUDA backend from C:\quantdiff\llama.cpp\bin\ggml-cuda.dll
load_backend: loaded RPC backend from C:\quantdiff\llama.cpp\bin\ggml-rpc.dll
load_backend: loaded CPU backend from C:\quantdiff\llama.cpp\bin\ggml-cpu-haswell.dll
[
  {
    "build_commit": "39da130fc",
    "build_number": 11770,
    "cpu_info": "AMD Ryzen 7 5800X 8-Core Processor             ",
    "gpu_info": "NVIDIA GeForce RTX 5070",
    "backends": "CUDA",
    "model_filename": "C:\\quantdiff\\dl\\campaign\\22086870b009\\Ornith-1.5-9B-Q8_0.gguf",
    "model_type": "qwen35 9B Q8_0",
    "model_size": 9775091712,
    "model_n_params": 9197093888,
    "n_batch": 2048,
    "n_ubatch": 512,
    "n_threads": 8,
    "cpu_mask": "0x0",
    "cpu_strict": false,
    "poll": 50,
    "type_k": "f16",
    "type_v": "f16",
    "n_gpu_layers": 999,
    "n_cpu_moe": 0,
    "split_mode": "layer",
    "main_gpu": 0,
    "no_kv_offload": false,
    "flash_attn": 1,
    "devices": "auto",
    "tensor_split": "0.00",
    "tensor_buft_overrides": "none",
    "load_mode": "auto",
    "lazy_mode": "auto",
    "embeddings": false,
    "no_op_offload": 0,
    "no_host": false,
    "repack": true,
    "fit_target": 0,
    "fit_min_ctx": 0,
    "n_prompt": 512,
    "n_gen": 0,
    "n_depth": 0,
    "test_time": "2026-10-22T05:52:03Z",
    "avg_ns": 106135680,
    "stddev_ns": 3453542,
    "avg_ts": 4827.935971,
    "stddev_ts": 150.670221,
    "samples_ns": [ 112286800, 104991600, 104834500, 104244900, 104320600 ],
    "samples_ts": [ 4559.75, 4876.58, 4883.89, 4911.51, 4907.95 ]
  },
  {
    "build_commit": "39da130fc",
    "build_number": 11770,
    "cpu_info": "AMD Ryzen 7 5800X 8-Core Processor             ",
    "gpu_info": "NVIDIA GeForce RTX 5070",
    "backends": "CUDA",
    "model_filename": "C:\\quantdiff\\dl\\campaign\\22086870b009\\Ornith-1.5-9B-Q8_0.gguf",
    "model_type": "qwen35 9B Q8_0",
    "model_size": 9775091712,
    "model_n_params": 9197093888,
    "n_batch": 2048,
    "n_ubatch": 512,
    "n_threads": 8,
    "cpu_mask": "0x0",
    "cpu_strict": false,
    "poll": 50,
    "type_k": "f16",
    "type_v": "f16",
    "n_gpu_layers": 999,
    "n_cpu_moe": 0,
    "split_mode": "layer",
    "main_gpu": 0,
    "no_kv_offload": false,
    "flash_attn": 1,
    "devices": "auto",
    "tensor_split": "0.00",
    "tensor_buft_overrides": "none",
    "load_mode": "auto",
    "lazy_mode": "auto",
    "embeddings": false,
    "no_op_offload": 0,
    "no_host": false,
    "repack": true,
    "fit_target": 0,
    "fit_min_ctx": 0,
    "n_prompt": 0,
    "n_gen": 128,
    "n_depth": 0,
    "test_time": "2026-10-22T05:52:04Z",
    "avg_ns": 1964898840,
    "stddev_ns": 30859102,
    "avg_ts": 65.156096,
    "stddev_ts": 1.018269,
    "samples_ns": [ 1949460000, 1997882600, 1940206400, 1998866600, 1938078600 ],
    "samples_ts": [ 65.6592, 64.0678, 65.9724, 64.0363, 66.0448 ]
  }
]

# exit 0 after 15s