$ C:\quantdiff\llama.cpp\bin\llama-bench.exe -m C:\quantdiff\dl\campaign\dbf58f4c413b\gemma-4-12b-it-Q3_K_M.gguf -ngl 999 -fa 1 -p 512 -n 128 -r 5 -t 8 -o json
# started 2026-10-22T07:19:17Z
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\\dbf58f4c413b\\gemma-4-12b-it-Q3_K_M.gguf",
    "model_type": "gemma4 ?B Q3_K - Medium",
    "model_size": 5678047424,
    "model_n_params": 11907350576,
    "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": "2026-10-22T07:19:20Z",
    "avg_ns": 174195700,
    "stddev_ns": 3537838,
    "avg_ts": 2940.167490,
    "stddev_ts": 58.133764,
    "samples_ns": [ 180521100, 172777800, 172644400, 172570100, 172465100 ],
    "samples_ts": [ 2836.23, 2963.34, 2965.63, 2966.91, 2968.72 ]
  },
  {
    "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\\dbf58f4c413b\\gemma-4-12b-it-Q3_K_M.gguf",
    "model_type": "gemma4 ?B Q3_K - Medium",
    "model_size": 5678047424,
    "model_n_params": 11907350576,
    "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": "2026-10-22T07:19:21Z",
    "avg_ns": 1646789080,
    "stddev_ns": 3269611,
    "avg_ts": 77.727260,
    "stddev_ts": 0.154177,
    "samples_ns": [ 1651333500, 1644046200, 1644347800, 1645056900, 1649161000 ],
    "samples_ts": [ 77.5131, 77.8567, 77.8424, 77.8089, 77.6152 ]
  }
]

# exit 0 after 13s