llama-bench：测量 llama.cpp 的提示处理与生成性能

llama-bench 是 llama.cpp 的性能测试工具。

目录

• 命令语法
• 示例：不同模型、批次大小、线程数、GPU 卸载层数和预填充上下文
• 输出格式：Markdown、CSV、JSON、JSONL、SQL

命令语法

以下帮助输出来自指定的 v0.5.0 README，保留参数、默认值及占位符原样。它不是本机执行命令后得到的结果。

usage: llama-bench [options]

options:
  -h, --help
  --numa <distribute|isolate|numactl>       numa mode (default: disabled)
  -r, --repetitions <n>                     number of times to repeat each test (default: 5)
  --prio <-1|0|1|2|3>                       process/thread priority (default: 0)
  --delay <0...N> (seconds)                 delay between each test (default: 0)
  -o, --output <csv|json|jsonl|md|sql>      output format printed to stdout (default: md)
  -oe, --output-err <csv|json|jsonl|md|sql> output format printed to stderr (default: none)
  --list-devices                            list available devices and exit
  -v, --verbose                             verbose output
  --progress                                print test progress indicators
  --no-warmup                               skip warmup runs before benchmarking
  -fitt, --fit-target <MiB>                 fit model to device memory with this margin per device in MiB (default: off)
  -fitc, --fit-ctx <n>                      minimum ctx size for --fit-target (default: 4096)
  -rpc, --rpc <rpc_servers>                 register RPC devices (comma separated)

test parameters:
  -m, --model <filename>                    (default: models/7B/ggml-model-q4_0.gguf)
  -hf, -hfr, --hf-repo <user>/<model>[:quant] Hugging Face model repository; quant is optional, case-insensitive
                                            default to Q4_K_M, or falls back to the first file in the repo if Q4_K_M doesn't exist.
                                            example: ggml-org/GLM-4.7-Flash-GGUF:Q4_K_M
                                            (default: unused)
  -hff, --hf-file <file>                    Hugging Face model file. If specified, it will override the quant in --hf-repo
                                            (default: unused)
  -hft, --hf-token <token>                  Hugging Face access token
                                            (default: value from HF_TOKEN environment variable)
  -p, --n-prompt <n>                        (default: 512)
  -n, --n-gen <n>                           (default: 128)
  -pg <pp,tg>                               (default: )
  -d, --n-depth <n>                         (default: 0)
  -b, --batch-size <n>                      (default: 2048)
  -ub, --ubatch-size <n>                    (default: 512)
  -ctk, --cache-type-k <t>                  (default: f16)
  -ctv, --cache-type-v <t>                  (default: f16)
  -t, --threads <n>                         (default: system dependent)
  -C, --cpu-mask <hex,hex>                  (default: 0x0)
  --cpu-strict <0|1>                        (default: 0)
  --poll <0...100>                          (default: 50)
  -ngl, --n-gpu-layers <n>                  (default: -1)
  -ncmoe, --n-cpu-moe <n>                   (default: 0)
  -sm, --split-mode <none|layer|row|tensor> (default: layer)
  -mg, --main-gpu <i>                       (default: 0)
  -nkvo, --no-kv-offload <0|1>              (default: 0)
  -fa, --flash-attn <on|off|auto>           (default: auto)
  -dev, --device <dev0/dev1/...>            (default: auto)
  -lzm, --lazy-mode <on|auto|off>           (default: auto)
  -mmp, --mmap <0|1>                        (DEPRECATED IN FAVOUR OF --load-mode)
  -dio, --direct-io <0|1>                   (DEPRECATED IN FAVOUR OF --load-mode)
  -embd, --embeddings <0|1>                 (default: 0)
  -ts, --tensor-split <ts0/ts1/..>          (default: 0)
  -ot --override-tensor <tensor name pattern>=<buffer type>;...
                                            (default: disabled)
  -nopo, --no-op-offload <0|1>              (default: 0)
  --no-host <0|1>                           (default: 0)

Multiple values can be given for each parameter by separating them with ','
or by specifying the parameter multiple times. Ranges can be given as
'first-last' or 'first-last+step' or 'first-last*mult'.

llama-bench 可以执行三类测试：

• 提示处理（pp）：按批次处理输入提示，对应 -p。
• 文本生成（tg）：生成一串 token，对应 -n。
• 提示处理＋文本生成（pg）：先处理提示，再生成 token，对应 -pg。

原文说明，除 -r、-o 和 -v 外，可重复给出选项来运行多组测试。每个 pp 或 tg 测试都会遍历指定选项的所有组合。多个值既可以用逗号分隔，例如 -n 16,32，也可以重复指定，例如 -n 16 -n 32。

每组测试重复 -r 指定的次数，并取平均值。结果给出平均 token 每秒（t/s）以及标准差。某些输出格式，例如 JSON，还包括各次重复的独立结果。

使用 -d <n> 可以指定测试开始时的上下文深度，先在 KV cache 中填充 <n> 个 token。

其他选项说明见同版本的 completion 示例。

计时范围：llama-bench 的测量不包括分词和采样耗时。

示例

以下命令、表格和结构化输出全部保留自原文。性能数字属于原文示例所用的模型、硬件与构建环境，本稿没有运行 benchmark，不能把这些数字当作当前 Laptop 的测量结果。表头保留输出字段名，便于与工具结果对照。

比较不同模型的文本生成

$ ./llama-bench -m models/7B/ggml-model-q4_0.gguf -m models/13B/ggml-model-q4_0.gguf -p 0 -n 128,256,512

比较不同批次大小的提示处理

$ ./llama-bench -n 0 -p 1024 -b 128,256,512,1024

比较不同线程数

$ ./llama-bench -n 0 -n 16 -p 64 -t 1,2,4,8,16,32

比较卸载到 GPU 的层数

$ ./llama-bench -ngl 10,20,30,31,32,33,34,35

比较预填充上下文

$ ./llama-bench -d 0,512

输出格式

默认使用 Markdown 输出。通过 -o 可以改用其他格式。

Markdown

$ ./llama-bench -o md

CSV

$ ./llama-bench -o csv

build_commit,build_number,cpu_info,gpu_info,backends,model_filename,model_type,model_size,model_n_params,n_batch,n_ubatch,n_threads,cpu_mask,cpu_strict,poll,type_k,type_v,n_gpu_layers,n_cpu_moe,split_mode,main_gpu,no_kv_offload,flash_attn,devices,tensor_split,tensor_buft_overrides,use_mmap,use_direct_io,embeddings,no_op_offload,no_host,fit_target,fit_min_ctx,n_prompt,n_gen,n_depth,test_time,avg_ns,stddev_ns,avg_ts,stddev_ts
"8cf427ff","5163","AMD Ryzen 7 7800X3D 8-Core Processor","NVIDIA GeForce RTX 4080","CUDA","models/Qwen2.5-7B-Instruct-Q4_K_M.gguf","qwen2 7B Q4_K - Medium","4677120000","7615616512","2048","512","8","0x0","0","50","f16","f16","-1","0","layer","0","0","-1","auto","0.00","none","1","0","0","0","0","0","0","512","0","0","2025-04-24T11:57:09Z","70285660","982040","7285.676949","100.064434"
"8cf427ff","5163","AMD Ryzen 7 7800X3D 8-Core Processor","NVIDIA GeForce RTX 4080","CUDA","models/Qwen2.5-7B-Instruct-Q4_K_M.gguf","qwen2 7B Q4_K - Medium","4677120000","7615616512","2048","512","8","0x0","0","50","f16","f16","-1","0","layer","0","0","-1","auto","0.00","none","1","0","0","0","0","0","0","0","128","0","2025-04-24T11:57:10Z","1067431600","3834831","119.915244","0.430617"

JSON

$ ./llama-bench -o json

[
  {
    "build_commit": "8cf427ff",
    "build_number": 5163,
    "cpu_info": "AMD Ryzen 7 7800X3D 8-Core Processor",
    "gpu_info": "NVIDIA GeForce RTX 4080",
    "backends": "CUDA",
    "model_filename": "models/Qwen2.5-7B-Instruct-Q4_K_M.gguf",
    "model_type": "qwen2 7B Q4_K - Medium",
    "model_size": 4677120000,
    "model_n_params": 7615616512,
    "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": -1,
    "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",
    "use_mmap": true,
    "use_direct_io": false,
    "embeddings": false,
    "no_op_offload": 0,
    "no_host": false,
    "fit_target": 0,
    "fit_min_ctx": 0,
    "n_prompt": 512,
    "n_gen": 0,
    "n_depth": 0,
    "test_time": "2025-04-24T11:58:50Z",
    "avg_ns": 72135640,
    "stddev_ns": 1453752,
    "avg_ts": 7100.002165,
    "stddev_ts": 140.341520,
    "samples_ns": [ 74601900, 71632900, 71745200, 71952700, 70745500 ],
    "samples_ts": [ 6863.1, 7147.55, 7136.37, 7115.79, 7237.21 ]
  },
  {
    "build_commit": "8cf427ff",
    "build_number": 5163,
    "cpu_info": "AMD Ryzen 7 7800X3D 8-Core Processor",
    "gpu_info": "NVIDIA GeForce RTX 4080",
    "backends": "CUDA",
    "model_filename": "models/Qwen2.5-7B-Instruct-Q4_K_M.gguf",
    "model_type": "qwen2 7B Q4_K - Medium",
    "model_size": 4677120000,
    "model_n_params": 7615616512,
    "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": -1,
    "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",
    "use_mmap": true,
    "use_direct_io": false,
    "embeddings": false,
    "no_op_offload": 0,
    "no_host": false,
    "fit_target": 0,
    "fit_min_ctx": 0,
    "n_prompt": 0,
    "n_gen": 128,
    "n_depth": 0,
    "test_time": "2025-04-24T11:58:51Z",
    "avg_ns": 1076767880,
    "stddev_ns": 9449585,
    "avg_ts": 118.881588,
    "stddev_ts": 1.041811,
    "samples_ns": [ 1075361300, 1065089400, 1071761200, 1081934900, 1089692600 ],
    "samples_ts": [ 119.03, 120.178, 119.43, 118.307, 117.464 ]
  }
]

JSONL

$ ./llama-bench -o jsonl

{"build_commit": "8cf427ff", "build_number": 5163, "cpu_info": "AMD Ryzen 7 7800X3D 8-Core Processor", "gpu_info": "NVIDIA GeForce RTX 4080", "backends": "CUDA", "model_filename": "models/Qwen2.5-7B-Instruct-Q4_K_M.gguf", "model_type": "qwen2 7B Q4_K - Medium", "model_size": 4677120000, "model_n_params": 7615616512, "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": -1, "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", "use_mmap": true, "use_direct_io": false, "embeddings": false, "no_op_offload": 0, "no_host": false, "fit_target": 0, "fit_min_ctx": 0, "n_prompt": 512, "n_gen": 0, "n_depth": 0, "test_time": "2025-04-24T11:59:33Z", "avg_ns": 70497220, "stddev_ns": 883196, "avg_ts": 7263.609157, "stddev_ts": 90.940578, "samples_ns": [ 71551000, 71222800, 70364100, 69439100, 69909100 ],"samples_ts": [ 7155.74, 7188.71, 7276.44, 7373.37, 7323.8 ]}
{"build_commit": "8cf427ff", "build_number": 5163, "cpu_info": "AMD Ryzen 7 7800X3D 8-Core Processor", "gpu_info": "NVIDIA GeForce RTX 4080", "backends": "CUDA", "model_filename": "models/Qwen2.5-7B-Instruct-Q4_K_M.gguf", "model_type": "qwen2 7B Q4_K - Medium", "model_size": 4677120000, "model_n_params": 7615616512, "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": -1, "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", "use_mmap": true, "use_direct_io": false, "embeddings": false, "no_op_offload": 0, "no_host": false, "fit_target": 0, "fit_min_ctx": 0, "n_prompt": 0, "n_gen": 128, "n_depth": 0, "test_time": "2025-04-24T11:59:33Z", "avg_ns": 1068078400, "stddev_ns": 6279455, "avg_ts": 119.844681, "stddev_ts": 0.699739, "samples_ns": [ 1066331700, 1064864900, 1079042600, 1063328400, 1066824400 ],"samples_ts": [ 120.038, 120.203, 118.624, 120.377, 119.982 ]}

SQL

SQL 输出适合导入 SQLite 数据库。可以把输出通过管道交给 sqlite3 命令行工具，将结果加入数据库。

$ ./llama-bench -o sql

CREATE TABLE IF NOT EXISTS llama_bench (
  build_commit TEXT,
  build_number INTEGER,
  cpu_info TEXT,
  gpu_info TEXT,
  backends TEXT,
  model_filename TEXT,
  model_type TEXT,
  model_size INTEGER,
  model_n_params INTEGER,
  n_batch INTEGER,
  n_ubatch INTEGER,
  n_threads INTEGER,
  cpu_mask TEXT,
  cpu_strict INTEGER,
  poll INTEGER,
  type_k TEXT,
  type_v TEXT,
  n_gpu_layers INTEGER,
  n_cpu_moe INTEGER,
  split_mode TEXT,
  main_gpu INTEGER,
  no_kv_offload INTEGER,
  flash_attn INTEGER,
  devices TEXT,
  tensor_split TEXT,
  tensor_buft_overrides TEXT,
  use_mmap INTEGER,
  use_direct_io INTEGER,
  embeddings INTEGER,
  no_op_offload INTEGER,
  no_host INTEGER,
  fit_target INTEGER,
  fit_min_ctx INTEGER,
  n_prompt INTEGER,
  n_gen INTEGER,
  n_depth INTEGER,
  test_time TEXT,
  avg_ns INTEGER,
  stddev_ns INTEGER,
  avg_ts REAL,
  stddev_ts REAL
);

INSERT INTO llama_bench (build_commit, build_number, cpu_info, gpu_info, backends, model_filename, model_type, model_size, model_n_params, n_batch, n_ubatch, n_threads, cpu_mask, cpu_strict, poll, type_k, type_v, n_gpu_layers, n_cpu_moe, split_mode, main_gpu, no_kv_offload, flash_attn, devices, tensor_split, tensor_buft_overrides, use_mmap, use_direct_io, embeddings, no_op_offload, no_host, fit_target, fit_min_ctx, n_prompt, n_gen, n_depth, test_time, avg_ns, stddev_ns, avg_ts, stddev_ts) VALUES ('8cf427ff', '5163', 'AMD Ryzen 7 7800X3D 8-Core Processor', 'NVIDIA GeForce RTX 4080', 'CUDA', 'models/Qwen2.5-7B-Instruct-Q4_K_M.gguf', 'qwen2 7B Q4_K - Medium', '4677120000', '7615616512', '2048', '512', '8', '0x0', '0', '50', 'f16', 'f16', '-1', '0', 'layer', '0', '0', '-1', 'auto', '0.00', 'none', '1', '0', '0', '0', '0', '0', '0', '512', '0', '0', '2025-04-24T12:00:08Z', '69905000', '519516', '7324.546977', '54.032613');
INSERT INTO llama_bench (build_commit, build_number, cpu_info, gpu_info, backends, model_filename, model_type, model_size, model_n_params, n_batch, n_ubatch, n_threads, cpu_mask, cpu_strict, poll, type_k, type_v, n_gpu_layers, n_cpu_moe, split_mode, main_gpu, no_kv_offload, flash_attn, devices, tensor_split, tensor_buft_overrides, use_mmap, use_direct_io, embeddings, no_op_offload, no_host, fit_target, fit_min_ctx, n_prompt, n_gen, n_depth, test_time, avg_ns, stddev_ns, avg_ts, stddev_ts) VALUES ('8cf427ff', '5163', 'AMD Ryzen 7 7800X3D 8-Core Processor', 'NVIDIA GeForce RTX 4080', 'CUDA', 'models/Qwen2.5-7B-Instruct-Q4_K_M.gguf', 'qwen2 7B Q4_K - Medium', '4677120000', '7615616512', '2048', '512', '8', '0x0', '0', '50', 'f16', 'f16', '-1', '0', 'layer', '0', '0', '-1', 'auto', '0.00', 'none', '1', '0', '0', '0', '0', '0', '0', '0', '128', '0', '2025-04-24T12:00:09Z', '1063608780', '4464130', '120.346696', '0.504647');

来源：The ggml authors，llama.cpp/tools/llama-bench README（v0.5.0）。本稿中文翻译并补充计时范围、原文输出和引用目标的边界说明；15 个代码/输出块及 6 张数据表保留完整，未使用 master 替换指定版本，未执行命令、下载模型或运行性能测试。

原文与本稿复用按 MIT 许可；完整告知如下：