跳转至

NVIDIA A100 GPUs 上的 Qwen3-14B 吞吐量优化

结论

qwen3-14b-a100

Qwen3-14B 在 A100 GPUs 上的吞吐量优化推荐配置：

推理服务命令 
1
trtllm-serve Qwen/Qwen3-14B --max_seq_len=40960 --enable_chunked_prefill

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: 3922.41
Mean TPOT(ms): 114.34		 Total TPS: 5770.24 (+47.1%)
Mean TPOT(ms): 260.46
短提示词 Total TPS: 4916.63
Mean TPOT(ms): 233.35		 Total TPS: 8762.17 (+78.2%)
Mean TPOT(ms): 715.19
中等长度提示词 Total TPS: 5967.13
Mean TPOT(ms): 301.73		 Total TPS: 7474.00 (+25.3%)
Mean TPOT(ms): 313.69
长提示词 Total TPS: 5153.42
Mean TPOT(ms): 254.61		 Total TPS: 6418.65 (+24.6%)
Mean TPOT(ms): 196.10
超长提示词 Total TPS: 4430.14
Mean TPOT(ms): 445.58		 Total TPS: 4413.82 (-0.4%)
Mean TPOT(ms): 493.18

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

Qwen3-14B

硬件

单个 NVIDIA A100 SXM GPU。

引擎版本

  • vLLM: v0.11.0
  • SGLang: v0.5.5.post1
  • 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 Qwen/Qwen3-14B --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 Qwen/Qwen3-14B --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 Qwen/Qwen3-14B
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):                  106.82
Total input tokens:                      217393
Total generated tokens:                  201605
Request throughput (req/s):              9.36
Output token throughput (tok/s):         1887.31
Peak output token throughput (tok/s):    4616.00
Peak concurrent requests:                1000.00
Total Token throughput (tok/s):          3922.41
---------------Time to First Token----------------
Mean TTFT (ms):                          41875.09
Median TTFT (ms):                        39151.17
P99 TTFT (ms):                           83576.82
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          114.34
Median TPOT (ms):                        106.15
P99 TPOT (ms):                           262.52
---------------Inter-token Latency----------------
Mean ITL (ms):                           95.04
Median ITL (ms):                         85.55
P99 ITL (ms):                            271.69
==================================================

SGLang

推理服务脚本 
1
python3 -m sglang.launch_server --model-path Qwen/Qwen3-14B
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):                  79.35
Total input tokens:                      217393
Total generated tokens:                  201706
Request throughput (req/s):              12.60
Output token throughput (tok/s):         2542.06
Peak output token throughput (tok/s):    6721.00
Peak concurrent requests:                1000.00
Total Token throughput (tok/s):          5281.81
---------------Time to First Token----------------
Mean TTFT (ms):                          20365.34
Median TTFT (ms):                        19226.41
P99 TTFT (ms):                           37314.85
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          364.93
Median TPOT (ms):                        185.18
P99 TPOT (ms):                           2181.56
---------------Inter-token Latency----------------
Mean ITL (ms):                           141.30
Median ITL (ms):                         84.71
P99 ITL (ms):                            566.50
==================================================

TensorRT-LLM

推理服务脚本 
1
trtllm-serve Qwen/Qwen3-14B
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):                  72.71
Total input tokens:                      217393
Total generated tokens:                  201471
Request throughput (req/s):              13.75
Output token throughput (tok/s):         2770.99
Peak output token throughput (tok/s):    7336.00
Peak concurrent requests:                1000.00
Total Token throughput (tok/s):          5760.97
---------------Time to First Token----------------
Mean TTFT (ms):                          15258.26
Median TTFT (ms):                        14210.29
P99 TTFT (ms):                           37520.64
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          261.77
Median TPOT (ms):                        166.55
P99 TPOT (ms):                           915.21
---------------Inter-token Latency----------------
Mean ITL (ms):                           130.55
Median ITL (ms):                         85.40
P99 ITL (ms):                            931.59
==================================================

结果：TensorRT-LLM (5760.97 tok/s) > SGLang (5281.81 tok/s) > vLLM (3922.41 tok/s)

2. TensorRT-LLM/vLLM 中的 量化

TensorRT-LLM FP8

推理服务脚本 
1
trtllm-serve Qwen/Qwen3-14B-FP8
基准测试结果 
1
# RuntimeError: Unsupported SM version for FP8 block scaling GEMM

TensorRT-LLM AWQ

推理服务脚本 
1
trtllm-serve Qwen/Qwen3-14B-AWQ
基准测试结果 
1
# KeyError: 'weight'

vLLM FP8

推理服务脚本 
1
vllm serve Qwen/Qwen3-14B-FP8
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):                  125.66
Total input tokens:                      217393
Total generated tokens:                  201655
Request throughput (req/s):              7.96
Output token throughput (tok/s):         1604.75
Peak output token throughput (tok/s):    3823.00
Peak concurrent requests:                1000.00
Total Token throughput (tok/s):          3334.74
---------------Time to First Token----------------
Mean TTFT (ms):                          52287.47
Median TTFT (ms):                        48838.08
P99 TTFT (ms):                           104986.53
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          146.36
Median TPOT (ms):                        132.69
P99 TPOT (ms):                           372.46
---------------Inter-token Latency----------------
Mean ITL (ms):                           118.59
Median ITL (ms):                         96.50
P99 ITL (ms):                            380.19
==================================================

vLLM AWQ

推理服务脚本 
1
vllm serve Qwen/Qwen3-14B-AWQ
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):                  111.81
Total input tokens:                      217393
Total generated tokens:                  201461
Request throughput (req/s):              8.94
Output token throughput (tok/s):         1801.74
Peak output token throughput (tok/s):    4069.00
Peak concurrent requests:                1000.00
Total Token throughput (tok/s):          3745.97
---------------Time to First Token----------------
Mean TTFT (ms):                          47136.93
Median TTFT (ms):                        44343.62
P99 TTFT (ms):                           94568.79
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          129.77
Median TPOT (ms):                        120.87
P99 TPOT (ms):                           331.28
---------------Inter-token Latency----------------
Mean ITL (ms):                           106.86
Median ITL (ms):                         89.48
P99 ITL (ms):                            338.24
==================================================

3. TensorRT-LLM 中的 最大批处理 Token 数

推理服务脚本 
1
trtllm-serve Qwen/Qwen3-14B --max_num_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
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
# --max_num_tokens=16384
============ Serving Benchmark Result ============
Successful requests:                     1000
Benchmark duration (s):                  72.54
Total input tokens:                      217393
Total generated tokens:                  201631
Request throughput (req/s):              13.79
Output token throughput (tok/s):         2779.60
Peak output token throughput (tok/s):    5583.00
Peak concurrent requests:                1000.00
Total Token throughput (tok/s):          5776.50
---------------Time to First Token----------------
Mean TTFT (ms):                          15374.97
Median TTFT (ms):                        13585.74
P99 TTFT (ms):                           38913.73
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          268.46
Median TPOT (ms):                        159.86
P99 TPOT (ms):                           1618.81
---------------Inter-token Latency----------------
Mean ITL (ms):                           129.13
Median ITL (ms):                         87.94
P99 ITL (ms):                            1712.52
==================================================

# --max_num_tokens=32768
============ Serving Benchmark Result ============
Successful requests:                     1000
Benchmark duration (s):                  74.21
Total input tokens:                      217393
Total generated tokens:                  201567
Request throughput (req/s):              13.48
Output token throughput (tok/s):         2716.13
Peak output token throughput (tok/s):    5559.00
Peak concurrent requests:                1000.00
Total Token throughput (tok/s):          5645.51
---------------Time to First Token----------------
Mean TTFT (ms):                          16776.11
Median TTFT (ms):                        15050.03
P99 TTFT (ms):                           43000.86
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          234.79
Median TPOT (ms):                        150.33
P99 TPOT (ms):                           1486.15
---------------Inter-token Latency----------------
Mean ITL (ms):                           123.01
Median ITL (ms):                         87.61
P99 ITL (ms):                            525.09
==================================================

优化选项汇总

优化选项 吞吐量提升
引擎选择 +47.1%
量化 -
最大批处理 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 Qwen/Qwen3-14B
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
92
93
94
95
96
97
98
99
# random 32K input
============ Serving Benchmark Result ============
Successful requests:                     100
Benchmark duration (s):                  724.56
Total input tokens:                      3200000
Total generated tokens:                  9891
Request throughput (req/s):              0.14
Output token throughput (tok/s):         13.65
Peak output token throughput (tok/s):    130.00
Peak concurrent requests:                100.00
Total Token throughput (tok/s):          4430.14
---------------Time to First Token----------------
Mean TTFT (ms):                          359648.08
Median TTFT (ms):                        359424.54
P99 TTFT (ms):                           714338.15
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          445.58
Median TPOT (ms):                        462.54
P99 TPOT (ms):                           463.96
---------------Inter-token Latency----------------
Mean ITL (ms):                           441.08
Median ITL (ms):                         448.05
P99 ITL (ms):                            594.59
==================================================

# random 4K input
============ Serving Benchmark Result ============
Successful requests:                     500
Benchmark duration (s):                  407.05
Total input tokens:                      1997942
Total generated tokens:                  99783
Request throughput (req/s):              1.23
Output token throughput (tok/s):         245.13
Peak output token throughput (tok/s):    1304.00
Peak concurrent requests:                500.00
Total Token throughput (tok/s):          5153.42
---------------Time to First Token----------------
Mean TTFT (ms):                          197625.03
Median TTFT (ms):                        196621.85
P99 TTFT (ms):                           396056.32
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          254.61
Median TPOT (ms):                        272.93
P99 TPOT (ms):                           283.50
---------------Inter-token Latency----------------
Mean ITL (ms):                           253.82
Median ITL (ms):                         369.00
P99 ITL (ms):                            377.42
==================================================

# random 2K input
============ Serving Benchmark Result ============
Successful requests:                     500
Benchmark duration (s):                  175.66
Total input tokens:                      998175
Total generated tokens:                  50000
Request throughput (req/s):              2.85
Output token throughput (tok/s):         284.64
Peak output token throughput (tok/s):    1722.00
Peak concurrent requests:                500.00
Total Token throughput (tok/s):          5967.13
---------------Time to First Token----------------
Mean TTFT (ms):                          86908.30
Median TTFT (ms):                        86575.09
P99 TTFT (ms):                           170607.65
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          301.73
Median TPOT (ms):                        334.17
P99 TPOT (ms):                           334.55
---------------Inter-token Latency----------------
Mean ITL (ms):                           299.28
Median ITL (ms):                         333.99
P99 ITL (ms):                            338.95
==================================================

# random 128 input
============ Serving Benchmark Result ============
Successful requests:                     1000
Benchmark duration (s):                  26.79
Total input tokens:                      127698
Total generated tokens:                  4000
Request throughput (req/s):              37.33
Output token throughput (tok/s):         149.33
Peak output token throughput (tok/s):    618.00
Peak concurrent requests:                1000.00
Total Token throughput (tok/s):          4916.63
---------------Time to First Token----------------
Mean TTFT (ms):                          19451.46
Median TTFT (ms):                        19150.52
P99 TTFT (ms):                           26523.59
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          233.35
Median TPOT (ms):                        238.61
P99 TPOT (ms):                           251.05
---------------Inter-token Latency----------------
Mean ITL (ms):                           175.01
Median ITL (ms):                         237.82
P99 ITL (ms):                            253.18
==================================================
优化后的推理服务脚本 
1
trtllm-serve Qwen/Qwen3-14B --max_seq_len=40960 --enable_chunked_prefill
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):                  727.25
Total input tokens:                      3200000
Total generated tokens:                  9932
Request throughput (req/s):              0.14
Output token throughput (tok/s):         13.66
Peak output token throughput (tok/s):    169.00
Peak concurrent requests:                100.00
Total Token throughput (tok/s):          4413.82
---------------Time to First Token----------------
Mean TTFT (ms):                          361107.28
Median TTFT (ms):                        361543.94
P99 TTFT (ms):                           716851.40
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          493.18
Median TPOT (ms):                        500.49
P99 TPOT (ms):                           590.17
---------------Inter-token Latency----------------
Mean ITL (ms):                           488.09
Median ITL (ms):                         48.22
P99 ITL (ms):                            2215.91
==================================================

# random 4K input
============ Serving Benchmark Result ============
Successful requests:                     500
Benchmark duration (s):                  326.81
Total input tokens:                      1997942
Total generated tokens:                  99760
Request throughput (req/s):              1.53
Output token throughput (tok/s):         305.25
Peak output token throughput (tok/s):    1224.00
Peak concurrent requests:                500.00
Total Token throughput (tok/s):          6418.65
---------------Time to First Token----------------
Mean TTFT (ms):                          155664.65
Median TTFT (ms):                        152609.98
P99 TTFT (ms):                           316584.25
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          196.10
Median TPOT (ms):                        207.95
P99 TPOT (ms):                           218.23
---------------Inter-token Latency----------------
Mean ITL (ms):                           195.11
Median ITL (ms):                         55.84
P99 ITL (ms):                            1038.71
==================================================

# random 2K input
============ Serving Benchmark Result ============
Successful requests:                     500
Benchmark duration (s):                  140.24
Total input tokens:                      998175
Total generated tokens:                  50000
Request throughput (req/s):              3.57
Output token throughput (tok/s):         356.52
Peak output token throughput (tok/s):    2128.00
Peak concurrent requests:                500.00
Total Token throughput (tok/s):          7474.00
---------------Time to First Token----------------
Mean TTFT (ms):                          67624.26
Median TTFT (ms):                        65786.34
P99 TTFT (ms):                           134519.75
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          313.69
Median TPOT (ms):                        351.99
P99 TPOT (ms):                           365.92
---------------Inter-token Latency----------------
Mean ITL (ms):                           310.55
Median ITL (ms):                         63.96
P99 ITL (ms):                            1002.01
==================================================

# random 128 input
============ Serving Benchmark Result ============
Successful requests:                     1000
Benchmark duration (s):                  15.03
Total input tokens:                      127698
Total generated tokens:                  4000
Request throughput (req/s):              66.53
Output token throughput (tok/s):         266.13
Peak output token throughput (tok/s):    883.00
Peak concurrent requests:                1000.00
Total Token throughput (tok/s):          8762.17
---------------Time to First Token----------------
Mean TTFT (ms):                          8166.48
Median TTFT (ms):                        7882.94
P99 TTFT (ms):                           14856.34
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          715.19
Median TPOT (ms):                        853.12
P99 TPOT (ms):                           1028.67
---------------Inter-token Latency----------------
Mean ITL (ms):                           536.38
Median ITL (ms):                         638.54
P99 ITL (ms):                            1280.94
==================================================