跳转至

NVIDIA H100 GPUs 上的 Qwen3-235B-A22B 吞吐量优化

结论

qwen3-235b-a22b-h100

Qwen3-235B-A22B-Instruct-2507 在 H100 GPUs 上的吞吐量优化推荐配置：

推理服务命令 
1
vllm serve Qwen/Qwen3-235B-A22B-Instruct-2507-FP8 -tp 4 --max-num-batched-tokens 16384

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: 8424.86
Mean TPOT(ms): 147.56		 Total TPS: 7787.17x2 (+84.9%)
Mean TPOT(ms): 186.26
短提示词 Total TPS: 14668.27
Mean TPOT(ms): 212.98		 Total TPS: 21211.83x2 (+189.2%)
Mean TPOT(ms): 125.18
中等长度提示词 Total TPS: 23858.69
Mean TPOT(ms): 101.08		 Total TPS: 18950.04x2 (+58.9%)
Mean TPOT(ms): 125.93
长提示词 Total TPS: 21113.53
Mean TPOT(ms): 68.94		 Total TPS: 16478.79x2 (+56.1%)
Mean TPOT(ms): 89.52
超长提示词 Total TPS: 22657.51
Mean TPOT(ms): 132.10		 Total TPS: 15366.32x2 (+35.6%)
Mean TPOT(ms): 194.60

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。
  4. 量化可能影响准确率，但对于大多数模型，FP8 量化的准确率下降可控制在 1% 以内。详情参见评估结果。因此，强烈建议在高吞吐量推理服务场景中使用 FP8 量化。

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

Qwen3-235B-A22B-Instruct-2507

硬件

单节点配备 8 × NVIDIA H100 SXM GPU。

引擎版本

  • vLLM: v0.11.0
  • SGLang: v0.5.5.post1

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/Qwen3-235B-A22B-Instruct-2507 --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/Qwen3-235B-A22B-Instruct-2507 --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/Qwen3-235B-A22B-Instruct-2507 -tp 8
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:                     984
Benchmark duration (s):                  48.28
Total input tokens:                      212894
Total generated tokens:                  193865
Request throughput (req/s):              20.38
Output token throughput (tok/s):         4015.36
Peak output token throughput (tok/s):    9638.00
Peak concurrent requests:                984.00
Total Token throughput (tok/s):          8424.86
---------------Time to First Token----------------
Mean TTFT (ms):                          8730.69
Median TTFT (ms):                        8707.89
P99 TTFT (ms):                           12940.37
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          147.56
Median TPOT (ms):                        114.67
P99 TPOT (ms):                           342.35
---------------Inter-token Latency----------------
Mean ITL (ms):                           81.31
Median ITL (ms):                         51.14
P99 ITL (ms):                            677.61
==================================================

SGLang

推理服务脚本 
1
python3 -m sglang.launch_server --model-path Qwen/Qwen3-235B-A22B-Instruct-2507 --tp-size 8
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:                     984
Benchmark duration (s):                  53.18
Total input tokens:                      214137
Total generated tokens:                  193829
Request throughput (req/s):              18.50
Output token throughput (tok/s):         3644.72
Peak output token throughput (tok/s):    8791.00
Peak concurrent requests:                984.00
Total Token throughput (tok/s):          7671.31
---------------Time to First Token----------------
Mean TTFT (ms):                          11715.28
Median TTFT (ms):                        6274.12
P99 TTFT (ms):                           32860.14
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          178.35
Median TPOT (ms):                        102.70
P99 TPOT (ms):                           841.80
---------------Inter-token Latency----------------
Mean ITL (ms):                           88.35
Median ITL (ms):                         44.39
P99 ITL (ms):                            732.90
==================================================

结果：vLLM (8424.86 tok/s) > SGLang (7671.31 tok/s)

2. vLLM 中的 量化

推理服务脚本 
1
2
# TP8 only is not compatible with the model weight. Use EP as well.
vllm serve Qwen/Qwen3-235B-A22B-Instruct-2507-FP8 -tp 8 --enable-expert-parallel
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:                     984
Benchmark duration (s):                  49.88
Total input tokens:                      211761
Total generated tokens:                  192350
Request throughput (req/s):              19.73
Output token throughput (tok/s):         3856.59
Peak output token throughput (tok/s):    9788.00
Peak concurrent requests:                984.00
Total Token throughput (tok/s):          8102.36
---------------Time to First Token----------------
Mean TTFT (ms):                          9755.64
Median TTFT (ms):                        9742.53
P99 TTFT (ms):                           14508.64
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          169.52
Median TPOT (ms):                        129.08
P99 TPOT (ms):                           353.11
---------------Inter-token Latency----------------
Mean ITL (ms):                           88.39
Median ITL (ms):                         49.03
P99 ITL (ms):                            343.14
==================================================

3. vLLM 中的 并行策略

TP4

推理服务脚本 
1
vllm serve Qwen/Qwen3-235B-A22B-Instruct-2507-FP8 -tp 4
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:                     984
Benchmark duration (s):                  54.12
Total input tokens:                      212992
Total generated tokens:                  194446
Request throughput (req/s):              18.18
Output token throughput (tok/s):         3592.57
Peak output token throughput (tok/s):    8605.00
Peak concurrent requests:                984.00
Total Token throughput (tok/s):          7527.80
---------------Time to First Token----------------
Mean TTFT (ms):                          9924.69
Median TTFT (ms):                        9896.14
P99 TTFT (ms):                           15574.61
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          169.20
Median TPOT (ms):                        121.96
P99 TPOT (ms):                           372.71
---------------Inter-token Latency----------------
Mean ITL (ms):                           89.89
Median ITL (ms):                         57.15
P99 ITL (ms):                            351.02
==================================================

TP4+EP

推理服务脚本 
1
vllm serve Qwen/Qwen3-235B-A22B-Instruct-2507-FP8 -tp 4 --enable-expert-parallel
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:                     984
Benchmark duration (s):                  57.08
Total input tokens:                      214311
Total generated tokens:                  192752
Request throughput (req/s):              17.24
Output token throughput (tok/s):         3376.99
Peak output token throughput (tok/s):    8005.00
Peak concurrent requests:                984.00
Total Token throughput (tok/s):          7131.69
---------------Time to First Token----------------
Mean TTFT (ms):                          10706.10
Median TTFT (ms):                        10676.53
P99 TTFT (ms):                           17226.88
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          190.24
Median TPOT (ms):                        131.62
P99 TPOT (ms):                           432.31
---------------Inter-token Latency----------------
Mean ITL (ms):                           96.98
Median ITL (ms):                         60.65
P99 ITL (ms):                            408.09
==================================================

4. vLLM 中的 注意力后端

默认使用 FlashAttention。

FlashInfer

推理服务脚本 
1
VLLM_ATTENTION_BACKEND=FLASHINFER vllm serve Qwen/Qwen3-235B-A22B-Instruct-2507-FP8 -tp 4
基准测试结果 
1
# Crash

XFormers

推理服务脚本 
1
VLLM_ATTENTION_BACKEND=XFORMERS vllm serve Qwen/Qwen3-235B-A22B-Instruct-2507-FP8 -tp 4
基准测试结果 
1
# Crash on inference

5. vLLM 中的 最大批处理 Token 数

推理服务脚本 
1
vllm serve Qwen/Qwen3-235B-A22B-Instruct-2507-FP8 -tp 4 --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
24
25
26
27
# --max-num-batched-tokens 16384
============ Serving Benchmark Result ============
Successful requests:                     984
Benchmark duration (s):                  52.46
Total input tokens:                      214381
Total generated tokens:                  194148
Request throughput (req/s):              18.76
Output token throughput (tok/s):         3700.75
Peak output token throughput (tok/s):    8574.00
Peak concurrent requests:                984.00
Total Token throughput (tok/s):          7787.17
---------------Time to First Token----------------
Mean TTFT (ms):                          9600.73
Median TTFT (ms):                        9525.78
P99 TTFT (ms):                           13887.13
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          186.26
Median TPOT (ms):                        108.31
P99 TPOT (ms):                           634.48
---------------Inter-token Latency----------------
Mean ITL (ms):                           86.62
Median ITL (ms):                         56.97
P99 ITL (ms):                            632.37
==================================================

# --max-num-batched-tokens 32768
# Crash on inference

优化选项汇总

优化选项 吞吐量提升
引擎选择 -
量化 -
并行策略 +78.7%
注意力后端 -
最大批处理 Token 数 +3.4%

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/Qwen3-235B-A22B-Instruct-2507 -tp 8
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):                  141.60
Total input tokens:                      3200000
Total generated tokens:                  8191
Request throughput (req/s):              0.71
Output token throughput (tok/s):         57.85
Peak output token throughput (tok/s):    346.00
Peak concurrent requests:                100.00
Total Token throughput (tok/s):          22657.51
---------------Time to First Token----------------
Mean TTFT (ms):                          70409.55
Median TTFT (ms):                        70737.29
P99 TTFT (ms):                           138084.17
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          132.10
Median TPOT (ms):                        134.49
P99 TPOT (ms):                           182.76
---------------Inter-token Latency----------------
Mean ITL (ms):                           129.44
Median ITL (ms):                         23.29
P99 ITL (ms):                            376.34
==================================================

# random 4K input
============ Serving Benchmark Result ============
Successful requests:                     500
Benchmark duration (s):                  98.99
Total input tokens:                      1997942
Total generated tokens:                  92185
Request throughput (req/s):              5.05
Output token throughput (tok/s):         931.21
Peak output token throughput (tok/s):    2450.00
Peak concurrent requests:                500.00
Total Token throughput (tok/s):          21113.53
---------------Time to First Token----------------
Mean TTFT (ms):                          47811.58
Median TTFT (ms):                        48562.54
P99 TTFT (ms):                           93161.32
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          68.94
Median TPOT (ms):                        72.58
P99 TPOT (ms):                           92.94
---------------Inter-token Latency----------------
Mean ITL (ms):                           68.75
Median ITL (ms):                         30.18
P99 ITL (ms):                            245.68
==================================================

# random 2K input
============ Serving Benchmark Result ============
Successful requests:                     500
Benchmark duration (s):                  43.91
Total input tokens:                      998175
Total generated tokens:                  49500
Request throughput (req/s):              11.39
Output token throughput (tok/s):         1127.26
Peak output token throughput (tok/s):    4021.00
Peak concurrent requests:                500.00
Total Token throughput (tok/s):          23858.69
---------------Time to First Token----------------
Mean TTFT (ms):                          21313.79
Median TTFT (ms):                        20242.32
P99 TTFT (ms):                           40939.00
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          101.08
Median TPOT (ms):                        108.64
P99 TPOT (ms):                           125.24
---------------Inter-token Latency----------------
Mean ITL (ms):                           100.16
Median ITL (ms):                         35.94
P99 ITL (ms):                            244.64
==================================================

# random 128 input
============ Serving Benchmark Result ============
Successful requests:                     984
Benchmark duration (s):                  8.83
Total input tokens:                      125650
Total generated tokens:                  3936
Request throughput (req/s):              111.38
Output token throughput (tok/s):         445.53
Peak output token throughput (tok/s):    1489.00
Peak concurrent requests:                984.00
Total Token throughput (tok/s):          14668.27
---------------Time to First Token----------------
Mean TTFT (ms):                          7042.17
Median TTFT (ms):                        6882.45
P99 TTFT (ms):                           8721.24
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          212.98
Median TPOT (ms):                        238.84
P99 TPOT (ms):                           242.28
---------------Inter-token Latency----------------
Mean ITL (ms):                           159.73
Median ITL (ms):                         235.21
P99 ITL (ms):                            250.35
==================================================
优化后的推理服务脚本 
1
vllm serve Qwen/Qwen3-235B-A22B-Instruct-2507-FP8 -tp 4 --max-num-batched-tokens 16384
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):                  208.79
Total input tokens:                      3200000
Total generated tokens:                  8281
Request throughput (req/s):              0.48
Output token throughput (tok/s):         39.66
Peak output token throughput (tok/s):    277.00
Peak concurrent requests:                100.00
Total Token throughput (tok/s):          15366.32
---------------Time to First Token----------------
Mean TTFT (ms):                          103754.50
Median TTFT (ms):                        103784.48
P99 TTFT (ms):                           205204.25
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          194.60
Median TPOT (ms):                        194.50
P99 TPOT (ms):                           277.94
---------------Inter-token Latency----------------
Mean ITL (ms):                           193.15
Median ITL (ms):                         27.60
P99 ITL (ms):                            1117.34
==================================================

# random 4K input
============ Serving Benchmark Result ============
Successful requests:                     500
Benchmark duration (s):                  126.76
Total input tokens:                      1997942
Total generated tokens:                  90966
Request throughput (req/s):              3.94
Output token throughput (tok/s):         717.60
Peak output token throughput (tok/s):    2139.00
Peak concurrent requests:                500.00
Total Token throughput (tok/s):          16478.79
---------------Time to First Token----------------
Mean TTFT (ms):                          60513.14
Median TTFT (ms):                        61048.35
P99 TTFT (ms):                           120001.31
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          89.52
Median TPOT (ms):                        94.78
P99 TPOT (ms):                           121.95
---------------Inter-token Latency----------------
Mean ITL (ms):                           89.17
Median ITL (ms):                         34.82
P99 ITL (ms):                            642.94
==================================================

# random 2K input
============ Serving Benchmark Result ============
Successful requests:                     500
Benchmark duration (s):                  55.28
Total input tokens:                      998175
Total generated tokens:                  49471
Request throughput (req/s):              9.04
Output token throughput (tok/s):         894.84
Peak output token throughput (tok/s):    3723.00
Peak concurrent requests:                500.00
Total Token throughput (tok/s):          18950.04
---------------Time to First Token----------------
Mean TTFT (ms):                          26860.95
Median TTFT (ms):                        26155.54
P99 TTFT (ms):                           51773.01
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          125.93
Median TPOT (ms):                        140.65
P99 TPOT (ms):                           152.49
---------------Inter-token Latency----------------
Mean ITL (ms):                           124.89
Median ITL (ms):                         39.21
P99 ITL (ms):                            632.84
==================================================

# random 128 input
============ Serving Benchmark Result ============
Successful requests:                     984
Benchmark duration (s):                  6.11
Total input tokens:                      125650
Total generated tokens:                  3936
Request throughput (req/s):              161.07
Output token throughput (tok/s):         644.28
Peak output token throughput (tok/s):    3676.00
Peak concurrent requests:                984.00
Total Token throughput (tok/s):          21211.83
---------------Time to First Token----------------
Mean TTFT (ms):                          5621.51
Median TTFT (ms):                        5525.90
P99 TTFT (ms):                           5803.57
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          125.18
Median TPOT (ms):                        152.34
P99 TPOT (ms):                           182.84
---------------Inter-token Latency----------------
Mean ITL (ms):                           93.89
Median ITL (ms):                         110.55
P99 ITL (ms):                            243.41
==================================================