Skip to content

Latest commit

 

History

History
40 lines (30 loc) · 2.03 KB

34.md

File metadata and controls

40 lines (30 loc) · 2.03 KB

LLM 速度测试 - 34B 单卡

使用模型 Yi-34B-Chat

机器配置 A40 24GB

测试方法

  • 零上下文时生成 1024 Tokens(禁用 EOS)
  • 生成三轮取平均值
  • 第一次生成前会进行 256 Tokens 的 Warm-up,记录热身前后显存变化

llama.cpp

llama_print_timings:        load time =      72.86 ms
llama_print_timings:      sample time =    4516.47 ms /  1023 runs   (    4.41 ms per token,   226.50 tokens per second)
llama_print_timings: prompt eval time =       0.00 ms /     1 tokens (    0.00 ms per token,      inf tokens per second)
llama_print_timings:        eval time =   39367.24 ms /  1023 runs   (   38.48 ms per token,    25.99 tokens per second)
llama_print_timings:       total time =   48351.40 ms /  1024 tokens
Output generated in 48.90 seconds (21.25 tokens/s, 1039 tokens, context 1, seed 1683353357)

llama_print_timings:        load time =      72.86 ms
llama_print_timings:      sample time =    4532.99 ms /  1023 runs   (    4.43 ms per token,   225.68 tokens per second)
llama_print_timings: prompt eval time =       0.00 ms /     1 tokens (    0.00 ms per token,      inf tokens per second)
llama_print_timings:        eval time =   39227.25 ms /  1023 runs   (   38.35 ms per token,    26.08 tokens per second)
llama_print_timings:       total time =   47776.10 ms /  1024 tokens
Output generated in 48.31 seconds (21.84 tokens/s, 1055 tokens, context 1, seed 1864893420)

llama_print_timings:        load time =      72.86 ms
llama_print_timings:      sample time =    4649.50 ms /  1023 runs   (    4.54 ms per token,   220.02 tokens per second)
llama_print_timings: prompt eval time =       0.00 ms /     1 tokens (    0.00 ms per token,      inf tokens per second)
llama_print_timings:        eval time =   39227.13 ms /  1023 runs   (   38.35 ms per token,    26.08 tokens per second)
llama_print_timings:       total time =   47838.45 ms /  1024 tokens
Output generated in 48.40 seconds (21.26 tokens/s, 1029 tokens, context 1, seed 314197904)

Average: 21.45 tokens/s

  • 加载后:21208 MiB