基于llama.cpp的推理性能测试

llama - batched-bench

1、下载编译llamap.cpp

git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
# rm -rf build; cmake -S . -B build -DLLAMA_CUBLAS=ON && cmake --build build --config Release

2、在如下地址下载模型

https://hf-mirror.com/TinyLlama/TinyLlama-1.1B-Chat-v0.2-GGUF/tree/main

3、运行测试

usage: ./batched-bench MODEL_PATH [N_KV_MAX] [IS_PP_SHARED] [NGL] <PP> <TG> <PL>
  <PP>, <TG> and PL are comma-separated lists of numbers without spaces

  example: ./batched-bench ggml-model-f16.gguf 2048 0 999 128,256,512 128,256 1,2,4,8,16,32

例子：./batched-bench ./models/llama-7b/ggml-model-q4_0.gguf 2048 0 99

main: n_kv_max = 2048, is_pp_shared = 0, n_gpu_layers = 99, n_threads = 4, n_threads_batch = 4

|    PP |     TG |    B |   N_KV |   T_PP s | S_PP t/s |   T_TG s | S_TG t/s |      T s |    S t/s |
|-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
|   128 |    128 |    1 |    256 |    0.056 |  2306.14 |    1.022 |   125.20 |    1.078 |   237.51 |
|   128 |    128 |    2 |    512 |    0.060 |  4232.94 |    1.018 |   251.49 |    1.078 |   474.77 |
|   128 |    128 |    4 |   1024 |    0.106 |  4838.77 |    1.211 |   422.71 |    1.317 |   777.49 |
|   128 |    128 |    8 |   2048 |    0.248 |  4135.55 |    1.977 |   517.87 |    2.225 |   920.47 |
|   128 |    256 |    1 |    384 |    0.043 |  2990.44 |    2.024 |   126.47 |    2.067 |   185.77 |
|   128 |    256 |    2 |    768 |    0.058 |  4380.94 |    2.010 |   254.69 |    2.069 |   371.24 |
|   128 |    256 |    4 |   1536 |    0.107 |  4790.19 |    2.517 |   406.85 |    2.624 |   585.42 |
|   256 |    128 |    1 |    384 |    0.060 |  4250.87 |    1.008 |   127.00 |    1.068 |   359.50 |
|   256 |    128 |    2 |    768 |    0.107 |  4806.20 |    1.005 |   254.61 |    1.112 |   690.64 |
|   256 |    128 |    4 |   1536 |    0.247 |  4149.39 |    1.320 |   387.95 |    1.567 |   980.50 |
|   256 |    256 |    1 |    512 |    0.060 |  4248.96 |    2.013 |   127.18 |    2.073 |   246.97 |
|   256 |    256 |    2 |   1024 |    0.108 |  4761.68 |    2.009 |   254.85 |    2.117 |   483.81 |
|   256 |    256 |    4 |   2048 |    0.250 |  4092.33 |    2.751 |   372.21 |    3.001 |   682.36 |
|   512 |    128 |    1 |    640 |    0.112 |  4574.08 |    1.008 |   127.01 |    1.120 |   571.55 |
|   512 |    128 |    2 |   1280 |    0.247 |  4142.86 |    1.032 |   248.11 |    1.279 |  1000.80 |
|   512 |    256 |    1 |    768 |    0.111 |  4622.02 |    2.016 |   127.00 |    2.127 |   361.15 |
|   512 |    256 |    2 |   1536 |    0.248 |  4134.78 |    2.096 |   244.30 |    2.343 |   655.44 |
|  1024 |    128 |    1 |   1152 |    0.254 |  4029.31 |    1.006 |   127.20 |    1.260 |   914.00 |
|  1024 |    256 |    1 |   1280 |    0.253 |  4041.34 |    2.008 |   127.48 |    2.262 |   565.97 |

llama_print_timings:        load time =     693.46 ms
llama_print_timings:      sample time =       0.00 ms /     1 runs   (    0.00 ms per token,      inf tokens per second)
llama_print_timings: prompt eval time =   21770.83 ms / 17936 tokens (    1.21 ms per token,   823.85 tokens per second)
llama_print_timings:        eval time =   12101.92 ms /  1536 runs   (    7.88 ms per token,   126.92 tokens per second)
llama_print_timings:       total time =   34486.78 ms / 19472 tokens

结果数据说明：

- `PP` - prompt tokens per batch
- `TG` - generated tokens per batch
- `B` - number of batches
- `N_KV` - required KV cache size
- `T_PP` - prompt processing time (i.e. time to first token)
- `S_PP` - prompt processing speed (`(B*PP)/T_PP` or `PP/T_PP`)
- `T_TG` - time to generate all batches
- `S_TG` - text generation speed (`(B*TG)/T_TG`)
- `T` - total time
- `S` - total speed (i.e. all tokens / total time)

问题记录

推理正常运行，但是没有offload到GPU

1、查看日志显示找不到GPU设备

2、查看GPU设备正常

(base) root@node06[/root]# nvidia-smi 
Tue Mar 19 09:27:34 2024   
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 460.91.03    Driver Version: 460.91.03    CUDA Version: 11.2     |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|                               |                      |               MIG M. |
|===============================+======================+======================|
|   0  Tesla T4            Off  | 00000000:04:00.0 Off |                    0 |
| N/A   30C    P8     9W /  70W |      0MiB / 15109MiB |      0%      Default |
|                               |                      |                  N/A |
+-------------------------------+----------------------+----------------------+
                                                                       
+-----------------------------------------------------------------------------+
| Processes:                                                                  |
|  GPU   GI   CI        PID   Type   Process name                  GPU Memory |
|        ID   ID                                                   Usage      |
|=============================================================================|
|  No running processes found                                                 |
+-----------------------------------------------------------------------------+

3、分析代码文件，

4、报错是cuda和驱动版本不匹配

(base) root@node06[/root]# /usr/local/cuda-12.0/extras/demo_suite/deviceQuery
/usr/local/cuda-12.0/extras/demo_suite/deviceQuery Starting...

 CUDA Device Query (Runtime API) version (CUDART static linking)

cudaGetDeviceCount returned 35
-> CUDA driver version is insufficient for CUDA runtime version
Result = FAIL

官网查询对应关系

https://docs.nvidia.com/cuda/cuda-toolkit-release-notes/index.html

5、安装旧版本cuda 重新编译运行

wget https://developer.download.nvidia.com/compute/cuda/repos/rhel7/x86_64/cuda-rhel7.repo

yum -y install cuda-toolkit-11-2

(base) root@node06[/usr/local/cuda-11.2/samples/1_Utilities/deviceQuery]# ./deviceQuery 
./deviceQuery Starting...

 CUDA Device Query (Runtime API) version (CUDART static linking)

Detected 1 CUDA Capable device(s)

Device 0: "Tesla T4"
  CUDA Driver Version / Runtime Version          11.2 / 11.2
  CUDA Capability Major/Minor version number:    7.5
  Total amount of global memory:                 15110 MBytes (15843721216 bytes)
  (40) Multiprocessors, ( 64) CUDA Cores/MP:     2560 CUDA Cores
  GPU Max Clock rate:                            1590 MHz (1.59 GHz)
  Memory Clock rate:                             5001 Mhz
  Memory Bus Width:                              256-bit
  L2 Cache Size:                                 4194304 bytes
  Maximum Texture Dimension Size (x,y,z)         1D=(131072), 2D=(131072, 65536), 3D=(16384, 16384, 16384)
  Maximum Layered 1D Texture Size, (num) layers  1D=(32768), 2048 layers
  Maximum Layered 2D Texture Size, (num) layers  2D=(32768, 32768), 2048 layers
  Total amount of constant memory:               65536 bytes
  Total amount of shared memory per block:       49152 bytes
  Total shared memory per multiprocessor:        65536 bytes
  Total number of registers available per block: 65536
  Warp size:                                     32
  Maximum number of threads per multiprocessor:  1024
  Maximum number of threads per block:           1024
  Max dimension size of a thread block (x,y,z): (1024, 1024, 64)
  Max dimension size of a grid size    (x,y,z): (2147483647, 65535, 65535)
  Maximum memory pitch:                          2147483647 bytes
  Texture alignment:                             512 bytes
  Concurrent copy and kernel execution:          Yes with 3 copy engine(s)
  Run time limit on kernels:                     No
  Integrated GPU sharing Host Memory:            No
  Support host page-locked memory mapping:       Yes
  Alignment requirement for Surfaces:            Yes
  Device has ECC support:                        Enabled
  Device supports Unified Addressing (UVA):      Yes
  Device supports Managed Memory:                Yes
  Device supports Compute Preemption:            Yes
  Supports Cooperative Kernel Launch:            Yes
  Supports MultiDevice Co-op Kernel Launch:      Yes
  Device PCI Domain ID / Bus ID / location ID:   0 / 4 / 0
  Compute Mode:
     < Default (multiple host threads can use ::cudaSetDevice() with device simultaneously) >

deviceQuery, CUDA Driver = CUDART, CUDA Driver Version = 11.2, CUDA Runtime Version = 11.2, NumDevs = 1
Result = PASS

llama.cpp编译问题

低cuda版本编译报错：

添加 CUDA_DOCKER_ARCH参数，尝试改为=all，无法解决

其他参数值自行对应cuda尝试，如：compute_75（实测成功）

‘all’,‘all-major’,‘compute_35’,‘compute_37’, ‘compute_50’,‘compute_52’,‘compute_53’,‘compute_60’,‘compute_61’,‘compute_62’, ‘compute_70’,‘compute_72’,‘compute_75’,‘compute_80’,‘compute_86’,‘compute_87’, ‘lto_35’,‘lto_37’,‘lto_50’,‘lto_52’,‘lto_53’,‘lto_60’,‘lto_61’,‘lto_62’, ‘lto_70’,‘lto_72’,‘lto_75’,‘lto_80’,‘lto_86’,‘lto_87’,‘sm_35’,‘sm_37’,‘sm_50’, ‘sm_52’,‘sm_53’,‘sm_60’,‘sm_61’,‘sm_62’,‘sm_70’,‘sm_72’,‘sm_75’,‘sm_80’, ‘sm_86’,‘sm_87’.

最佳解决方案是部署高版本的驱动，驱动向前兼容，高版本驱动可以使用低版本的cuda