跳转至

NVIDIA H100 GPUs 上的 GPT-OSS-20B 吞吐量优化

结论

gpt-oss-20b-h100

GPT-OSS-20B 在 H100 GPUs 上的吞吐量优化推荐配置：

推理服务命令 
1
vllm serve openai/gpt-oss-20b --async-scheduling

优化前后的基准测试结果对比：

基准测试用例 基准配置（未做任何优化的 vLLM） 优化后
ShareGPT Total TPS: 16701.31
Mean TPOT(ms): 91.33		 Total TPS: 19128.79 (+14.5%)
Mean TPOT(ms): 86.81
短提示词 Total TPS: 28032.23
Mean TPOT(ms): 178.91		 Total TPS: 32611.91 (+16.3%)
Mean TPOT(ms): 75.14
中等长度提示词 Total TPS: 32113.65
Mean TPOT(ms): 149.76		 Total TPS: 33149.54 (+3.2%)
Mean TPOT(ms): 146.63
长提示词 Total TPS: 33236.17
Mean TPOT(ms): 184.26		 Total TPS: 33823.34 (+1.8%)
Mean TPOT(ms): 182.69
超长提示词 Total TPS: 29730.30
Mean TPOT(ms): 239.44		 Total TPS: 29718.21 (-0.0%)
Mean TPOT(ms): 239.49

说明

  1. 基准测试未覆盖所有可能的优化组合。例如，我们会选择默认配置下表现最好的推理引擎作为后续调优的起点。这种剪枝方式得到的是局部最优解，未必是全局最优解。
  2. 其他优化方法取决于具体使用场景，包括最大批大小、调度配置、扩展 KV cache、CUDA graph、Torch Compile 等。本文结论可作为进一步针对性优化的起点。
  3. 测试基于特定的硬件和软件环境。推理引擎的改进可能带来新的结论。

如有遗漏或需要根据最新变化更新的内容，请告诉我们。

优化目标

在高并发请求场景下实现高吞吐量。

实验环境

模型

GPT-OSS-20B

硬件

单个 NVIDIA H100 SXM GPU。

引擎版本

  • vLLM: v0.11.0
  • SGLang: v0.5.4.post2
  • TensorRT-LLM: v1.2.0rc1

基准测试数据集

  1. ShareGPT
  2. 不同序列长度的随机数据集：
    • 超长提示词：32000 输入 tokens，100 输出 tokens
    • 长提示词：4000 输入 tokens，200 输出 tokens
    • 中等长度提示词：2000 输入 tokens，100 输出 tokens
    • 短提示词：128 输入 tokens，4 输出 tokens

基准测试脚本

使用 vLLM bench CLI 工具对模型性能进行基准测试。运行基准测试的命令如下：

1
2
3
4
5
6
7
8
# Prepare the ShareGPT dataset
wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json

# Benchmark on ShareGPT dataset
vllm bench serve --model openai/gpt-oss-20b --backend openai-chat --endpoint /v1/chat/completions --dataset-name sharegpt --dataset-path ShareGPT_V3_unfiltered_cleaned_split.json --num-prompts 1000

# Benchmark on random dataset (fixed seed for reproducibility)
vllm bench serve --model openai/gpt-oss-20b --backend openai-chat --endpoint /v1/chat/completions --dataset-name random --random-input-len 4000 --random-output-len 200 --num-prompts 500 --seed 42

实验结果

1. 选择推理引擎

vLLM

推理服务脚本 
1
vllm serve openai/gpt-oss-20b
基准测试结果 
 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
============ Serving Benchmark Result ============
Successful requests:                     1000
Benchmark duration (s):                  24.81
Total input tokens:                      215312
Total generated tokens:                  193979
Request throughput (req/s):              40.30
Output token throughput (tok/s):         7817.03
Peak output token throughput (tok/s):    15953.00
Peak concurrent requests:                1000.00
Total Token throughput (tok/s):          16493.75
---------------Time to First Token----------------
Mean TTFT (ms):                          4881.62
Median TTFT (ms):                        4791.97
P99 TTFT (ms):                           8166.01
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          91.00
Median TPOT (ms):                        55.52
P99 TPOT (ms):                           237.48
---------------Inter-token Latency----------------
Mean ITL (ms):                           44.98
Median ITL (ms):                         31.92
P99 ITL (ms):                            248.25
==================================================

SGLang

推理服务脚本 
1
python3 -m sglang.launch_server --model-path openai/gpt-oss-20b
基准测试结果 
 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
============ Serving Benchmark Result ============
Successful requests:                     1000
Benchmark duration (s):                  73.05
Total input tokens:                      215312
Total generated tokens:                  195734
Request throughput (req/s):              13.69
Output token throughput (tok/s):         2679.49
Peak output token throughput (tok/s):    13804.00
Peak concurrent requests:                1000.00
Total Token throughput (tok/s):          5626.99
---------------Time to First Token----------------
Mean TTFT (ms):                          10412.75
Median TTFT (ms):                        10326.68
P99 TTFT (ms):                           11920.92
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          314.53
Median TPOT (ms):                        215.95
P99 TPOT (ms):                           1052.78
---------------Inter-token Latency----------------
Mean ITL (ms):                           161.29
Median ITL (ms):                         47.78
P99 ITL (ms):                            1929.13
==================================================

TensorRT-LLM

推理服务脚本 
1
trtllm-serve openai/gpt-oss-20b
基准测试结果 
 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
============ Serving Benchmark Result ============
Successful requests:                     1000
Benchmark duration (s):                  53.03
Total input tokens:                      215312
Total generated tokens:                  191726
Request throughput (req/s):              18.86
Output token throughput (tok/s):         3615.45
Peak output token throughput (tok/s):    18641.00
Peak concurrent requests:                1000.00
Total Token throughput (tok/s):          7675.67
---------------Time to First Token----------------
Mean TTFT (ms):                          12248.87
Median TTFT (ms):                        12024.20
P99 TTFT (ms):                           22142.67
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          224.49
Median TPOT (ms):                        134.32
P99 TPOT (ms):                           552.52
---------------Inter-token Latency----------------
Mean ITL (ms):                           51.52
Median ITL (ms):                         0.07
P99 ITL (ms):                            963.17
==================================================

结果：vLLM (16493.75 tok/s) > TensorRT-LLM (7675.67 tok/s) > SGLang(5626.99 tok/s)

2. vLLM 中的 异步调度

推理服务脚本 
1
vllm serve openai/gpt-oss-20b --async-scheduling
基准测试结果 
 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
============ Serving Benchmark Result ============
Successful requests:                     1000
Benchmark duration (s):                  21.39
Total input tokens:                      215312
Total generated tokens:                  193919
Request throughput (req/s):              46.74
Output token throughput (tok/s):         9064.40
Peak output token throughput (tok/s):    22786.00
Peak concurrent requests:                1000.00
Total Token throughput (tok/s):          19128.79
---------------Time to First Token----------------
Mean TTFT (ms):                          4933.91
Median TTFT (ms):                        4761.74
P99 TTFT (ms):                           8029.82
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          86.81
Median TPOT (ms):                        52.23
P99 TPOT (ms):                           222.47
---------------Inter-token Latency----------------
Mean ITL (ms):                           38.92
Median ITL (ms):                         27.02
P99 ITL (ms):                            236.29
==================================================

5. vLLM 中的 批处理 Token 数调优

推理服务脚本 
1
vllm serve openai/gpt-oss-120b --async-scheduling --max-num-batched-tokens 16384
基准测试结果 
 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
============ Serving Benchmark Result ============
Successful requests:                     1000
Benchmark duration (s):                  21.56
Total input tokens:                      215312
Total generated tokens:                  194591
Request throughput (req/s):              46.39
Output token throughput (tok/s):         9027.14
Peak output token throughput (tok/s):    21433.00
Peak concurrent requests:                1000.00
Total Token throughput (tok/s):          19015.54
---------------Time to First Token----------------
Mean TTFT (ms):                          4957.11
Median TTFT (ms):                        4698.76
P99 TTFT (ms):                           7875.07
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          104.05
Median TPOT (ms):                        46.28
P99 TPOT (ms):                           443.08
---------------Inter-token Latency----------------
Mean ITL (ms):                           38.42
Median ITL (ms):                         24.01
P99 ITL (ms):                            448.17
==================================================

优化选项汇总

优化选项 吞吐量提升
引擎选择 -
异步调度 +16.0%
批处理 Token 数 -

其他基准测试用例

我们进一步对优化后的配置进行了基准测试，评估其在不同工作负载下的泛化表现。

基准配置的推理服务脚本 
1
vllm serve openai/gpt-oss-20b
基准配置的测试结果 
 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
# random 32K input
============ Serving Benchmark Result ============
Successful requests:                     100
Benchmark duration (s):                  107.95
Total input tokens:                      3200000
Total generated tokens:                  9350
Request throughput (req/s):              0.93
Output token throughput (tok/s):         86.62
Total Token throughput (tok/s):          29730.30
---------------Time to First Token----------------
Mean TTFT (ms):                          53687.69
Median TTFT (ms):                        53539.25
P99 TTFT (ms):                           105655.63
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          239.44
Median TPOT (ms):                        273.86
P99 TPOT (ms):                           275.66
---------------Inter-token Latency----------------
Mean ITL (ms):                           247.74
Median ITL (ms):                         266.37
P99 ITL (ms):                            513.87
==================================================

# random 4K input
============ Serving Benchmark Result ============
Successful requests:                     500
Benchmark duration (s):                  61.73
Total input tokens:                      1999249
Total generated tokens:                  52309
Request throughput (req/s):              8.10
Output token throughput (tok/s):         847.43
Total Token throughput (tok/s):          33236.17
---------------Time to First Token----------------
Mean TTFT (ms):                          31245.12
Median TTFT (ms):                        30820.97
P99 TTFT (ms):                           58702.17
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          184.26
Median TPOT (ms):                        225.55
P99 TPOT (ms):                           230.22
---------------Inter-token Latency----------------
Mean ITL (ms):                           192.92
Median ITL (ms):                         227.42
P99 ITL (ms):                            458.94
==================================================

# random 2K input
============ Serving Benchmark Result ============
Successful requests:                     500
Benchmark duration (s):                  32.58
Total input tokens:                      999120
Total generated tokens:                  47022
Request throughput (req/s):              15.35
Output token throughput (tok/s):         1443.44
Total Token throughput (tok/s):          32113.65
---------------Time to First Token----------------
Mean TTFT (ms):                          16515.86
Median TTFT (ms):                        16363.96
P99 TTFT (ms):                           30359.34
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          149.76
Median TPOT (ms):                        159.20
P99 TPOT (ms):                           225.79
---------------Inter-token Latency----------------
Mean ITL (ms):                           159.21
Median ITL (ms):                         220.69
P99 ITL (ms):                            450.04
==================================================

# random 128 input
============ Serving Benchmark Result ============
Successful requests:                     1000
Benchmark duration (s):                  4.70
Total input tokens:                      127755
Total generated tokens:                  4000
Request throughput (req/s):              212.76
Output token throughput (tok/s):         851.04
Total Token throughput (tok/s):          28032.23
---------------Time to First Token----------------
Mean TTFT (ms):                          3174.14
Median TTFT (ms):                        2892.24
P99 TTFT (ms):                           4415.25
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          178.91
Median TPOT (ms):                        208.03
P99 TPOT (ms):                           246.08
---------------Inter-token Latency----------------
Mean ITL (ms):                           274.82
Median ITL (ms):                         228.45
P99 ITL (ms):                            509.42
==================================================
优化后的推理服务脚本 
1
vllm serve openai/gpt-oss-20b --async-scheduling
优化后的测试结果 
 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
# random 32K input
============ Serving Benchmark Result ============
Successful requests:                     100
Benchmark duration (s):                  107.99
Total input tokens:                      3200000
Total generated tokens:                  9404
Request throughput (req/s):              0.93
Output token throughput (tok/s):         87.08
Peak output token throughput (tok/s):    1196.00
Peak concurrent requests:                100.00
Total Token throughput (tok/s):          29718.21
---------------Time to First Token----------------
Mean TTFT (ms):                          53565.41
Median TTFT (ms):                        53436.03
P99 TTFT (ms):                           105793.87
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          239.49
Median TPOT (ms):                        275.30
P99 TPOT (ms):                           276.86
---------------Inter-token Latency----------------
Mean ITL (ms):                           247.68
Median ITL (ms):                         269.91
P99 ITL (ms):                            515.55
==================================================

# random 4K input
============ Serving Benchmark Result ============
Successful requests:                     500
Benchmark duration (s):                  60.67
Total input tokens:                      1999249
Total generated tokens:                  52966
Request throughput (req/s):              8.24
Output token throughput (tok/s):         872.95
Peak output token throughput (tok/s):    7931.00
Peak concurrent requests:                500.00
Total Token throughput (tok/s):          33823.34
---------------Time to First Token----------------
Mean TTFT (ms):                          30513.46
Median TTFT (ms):                        30091.29
P99 TTFT (ms):                           58206.28
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          182.69
Median TPOT (ms):                        224.54
P99 TPOT (ms):                           229.32
---------------Inter-token Latency----------------
Mean ITL (ms):                           191.65
Median ITL (ms):                         226.01
P99 ITL (ms):                            458.93
==================================================

# random 2K input
============ Serving Benchmark Result ============
Successful requests:                     500
Benchmark duration (s):                  31.55
Total input tokens:                      999120
Total generated tokens:                  46814
Request throughput (req/s):              15.85
Output token throughput (tok/s):         1483.71
Peak output token throughput (tok/s):    12476.00
Peak concurrent requests:                500.00
Total Token throughput (tok/s):          33149.54
---------------Time to First Token----------------
Mean TTFT (ms):                          16044.42
Median TTFT (ms):                        16014.33
P99 TTFT (ms):                           29736.08
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          146.63
Median TPOT (ms):                        154.39
P99 TPOT (ms):                           223.69
---------------Inter-token Latency----------------
Mean ITL (ms):                           155.89
Median ITL (ms):                         216.15
P99 ITL (ms):                            447.46
==================================================

# random 128 input
============ Serving Benchmark Result ============
Successful requests:                     1000
Benchmark duration (s):                  4.04
Total input tokens:                      127755
Total generated tokens:                  4000
Request throughput (req/s):              247.52
Output token throughput (tok/s):         990.08
Peak output token throughput (tok/s):    1364.00
Peak concurrent requests:                1000.00
Total Token throughput (tok/s):          32611.91
---------------Time to First Token----------------
Mean TTFT (ms):                          2884.78
Median TTFT (ms):                        2641.45
P99 TTFT (ms):                           3879.45
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          75.14
Median TPOT (ms):                        16.98
P99 TPOT (ms):                           221.60
---------------Inter-token Latency----------------
Mean ITL (ms):                           156.10
Median ITL (ms):                         151.36
P99 ITL (ms):                            452.14
==================================================