Перейти к содержанию

Оптимизация пропускной способности GLM-4.5-Air на NVIDIA A100 GPU

Вывод

glm-4.5-air-a100

Рекомендуемая конфигурация для оптимизации пропускной способности GLM-4.5-Air на A100 GPU:

Команда обслуживания модели 
1
2
python3 -m sglang.launch_server --model-path zai-org/GLM-4.5-Air \
--tp-size 4 --ep-size 4 --tool-call-parser glm --reasoning-parser glm45

Сравнение результатов бенчмарков до и после оптимизации:

Сценарий бенчмарка базовая конфигурация (vLLM без оптимизации) После оптимизации (в пересчёте на то же число GPU)
ShareGPT Общий TPS: 3903.44
Среднее TPOT(ms): 104.59		 Общий TPS: 6245.77 (+60.0%)
Среднее TPOT(ms): 309.31
Короткий запрос Общий TPS: 6271.24
Среднее TPOT(ms): 153.00		 Общий TPS: 13012.20 (+107.5%)
Среднее TPOT(ms): 1246.57
Запрос средней длины Общий TPS: 8643.75
Среднее TPOT(ms): 204.31		 Общий TPS: 10956.49 (+26.8%)
Среднее TPOT(ms): 165.34
Длинный запрос Общий TPS: 6975.10
Среднее TPOT(ms): 267.38		 Общий TPS: 9100.71 (+30.5%)
Среднее TPOT(ms): 124.68
Очень длинный запрос Общий TPS: 6435.97
Среднее TPOT(ms): 305.47		 Общий TPS: 7945.66 (+23.5%)
Среднее TPOT(ms): 276.41

Примечание

  1. Наши бенчмарки не охватывают все возможные сочетания оптимизаций. Например, отправной точкой дальнейшей настройки становится движок инференса с лучшей производительностью при настройках по умолчанию. Такое отсечение вариантов даёт локальный оптимум, который может не совпадать с глобальным.
  2. Есть и другие методы оптимизации, зависящие от сценария использования: максимальный размер батча, настройки планировщика, расширенный KV-кэш, CUDA graph и т. д. Выводы этого документа могут служить отправной точкой для более целенаправленной оптимизации.
  3. Тесты проводились на определённых аппаратных и программных конфигурациях. Развитие движков инференса может привести к новым выводам.

Если что-то упущено или требуется обновление с учётом новых изменений, сообщите нам.

Цель оптимизации

Высокая пропускная способность при высокой конкурентности запросов.

Условия эксперимента

Модель

zai-org/GLM-4.5-Air

Оборудование

4 × NVIDIA A100 SXM 80GB GPU на одном узле.

Версия движка

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

Набор данных бенчмарка

  1. ShareGPT
  2. Случайный набор данных с различной длиной последовательностей:
    • Очень длинный запрос: 32000 входных токенов, 100 выходных токенов
    • Длинный запрос: 4000 входных токенов, 200 выходных токенов
    • Запрос средней длины: 2000 входных токенов, 100 выходных токенов
    • Короткий запрос: 128 входных токенов, 4 выходных токенов

Скрипт бенчмарка

Для оценки производительности модели используется vLLM bench CLI. Бенчмарк запускается следующей командой:

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 zai-org/GLM-4.5-Air --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 zai-org/GLM-4.5-Air --backend openai-chat --endpoint /v1/chat/completions --dataset-name random --random-input-len 4000 --random-output-len 200 --num-prompts 500 --seed 42

Результаты эксперимента

1. Выбор движка инференса

vLLM

Скрипт обслуживания модели 
1
vllm serve zai-org/GLM-4.5-Air -tp 4
Результат бенчмарка 
 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):                  106.15
Total input tokens:                      214465
Total generated tokens:                  199904
Request throughput (req/s):              9.42
Output token throughput (tok/s):         1883.14
Peak output token throughput (tok/s):    3380.00
Peak concurrent requests:                1000.00
Total Token throughput (tok/s):          3903.44
---------------Time to First Token----------------
Mean TTFT (ms):                          38971.22
Median TTFT (ms):                        37502.20
P99 TTFT (ms):                           76635.16
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          104.59
Median TPOT (ms):                        101.18
P99 TPOT (ms):                           210.38
---------------Inter-token Latency----------------
Mean ITL (ms):                           91.67
Median ITL (ms):                         87.56
P99 ITL (ms):                            178.94
==================================================

SGLang

Скрипт обслуживания модели 
1
python3 -m sglang.launch_server --model-path zai-org/GLM-4.5-Air --tp-size 4
Результат бенчмарка 
 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):                  67.25
Total input tokens:                      214465
Total generated tokens:                  199873
Request throughput (req/s):              14.87
Output token throughput (tok/s):         2971.88
Peak output token throughput (tok/s):    11709.00
Peak concurrent requests:                1000.00
Total Token throughput (tok/s):          6160.72
---------------Time to First Token----------------
Mean TTFT (ms):                          14772.52
Median TTFT (ms):                        14625.90
P99 TTFT (ms):                           22760.64
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          328.21
Median TPOT (ms):                        129.31
P99 TPOT (ms):                           2061.59
---------------Inter-token Latency----------------
Mean ITL (ms):                           109.79
Median ITL (ms):                         67.14
P99 ITL (ms):                            411.08
==================================================

Результат: SGLang(6160.72 tok/s) > vLLM (3903.44 tok/s) > TensorRT-LLM (Не поддерживается)

2. Quantization в SGLang

FP8

Скрипт обслуживания модели 
1
python3 -m sglang.launch_server --model zai-org/GLM-4.5-Air-FP8 --tp-size 4
Результат бенчмарка 
1
# Crash. FP8 not supported on A100

3. Параллелизм в SGLang

PP4

Скрипт обслуживания модели 
1
python3 -m sglang.launch_server --model zai-org/GLM-4.5-Air --pp-size 8
Результат бенчмарка 
1
# Crash. PP not supported for this setup.

TP4+EP4

Скрипт обслуживания модели 
1
python3 -m sglang.launch_server --model-path zai-org/GLM-4.5-Air  --tp-size 4 --ep-size 4
Результат бенчмарка 
 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):                  66.34
Total input tokens:                      214465
Total generated tokens:                  199904
Request throughput (req/s):              15.07
Output token throughput (tok/s):         3013.15
Peak output token throughput (tok/s):    9933.00
Peak concurrent requests:                1000.00
Total Token throughput (tok/s):          6245.77
---------------Time to First Token----------------
Mean TTFT (ms):                          8445.43
Median TTFT (ms):                        8321.65
P99 TTFT (ms):                           15245.84
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          309.31
Median TPOT (ms):                        130.15
P99 TPOT (ms):                           1969.56
---------------Inter-token Latency----------------
Mean ITL (ms):                           112.44
Median ITL (ms):                         75.24
P99 ITL (ms):                            484.22
==================================================

4. Бэкенд внимания в SGLang

В этой конфигурации SGLang по умолчанию использует бэкенд внимания flashinfer.

triton

Скрипт обслуживания модели 
1
python3 -m sglang.launch_server --model-path zai-org/GLM-4.5-Air  --tp-size 4 --ep-size 4 --attention-backend triton
Результат бенчмарка 
 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):                  68.84
Total input tokens:                      214465
Total generated tokens:                  199904
Request throughput (req/s):              14.53
Output token throughput (tok/s):         2904.08
Peak output token throughput (tok/s):    9443.00
Peak concurrent requests:                1000.00
Total Token throughput (tok/s):          6019.69
---------------Time to First Token----------------
Mean TTFT (ms):                          9348.81
Median TTFT (ms):                        9050.60
P99 TTFT (ms):                           17294.26
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          373.94
Median TPOT (ms):                        141.96
P99 TPOT (ms):                           2701.08
---------------Inter-token Latency----------------
Mean ITL (ms):                           120.58
Median ITL (ms):                         74.19
P99 ITL (ms):                            413.30
==================================================

fa3

Скрипт обслуживания модели 
1
python3 -m sglang.launch_server --model-path zai-org/GLM-4.5-Air  --tp-size 4 --ep-size 4 --attention-backend fa3
Результат бенчмарка 
 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):                  67.66
Total input tokens:                      214465
Total generated tokens:                  199904
Request throughput (req/s):              14.78
Output token throughput (tok/s):         2954.39
Peak output token throughput (tok/s):    8239.00
Peak concurrent requests:                1000.00
Total Token throughput (tok/s):          6123.98
---------------Time to First Token----------------
Mean TTFT (ms):                          8260.21
Median TTFT (ms):                        8222.20
P99 TTFT (ms):                           15115.68
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          310.40
Median TPOT (ms):                        133.38
P99 TPOT (ms):                           2082.45
---------------Inter-token Latency----------------
Mean ITL (ms):                           114.41
Median ITL (ms):                         78.46
P99 ITL (ms):                            397.40
==================================================

5. Настройка CUDA graph в SGLang

Скрипт обслуживания модели 
1
python3 -m sglang.launch_server --model-path zai-org/GLM-4.5-Air  --tp-size 4 --ep-size 4 --cuda-graph-max-bs 512
Результат бенчмарка 
 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):                  66.38
Total input tokens:                      214465
Total generated tokens:                  199904
Request throughput (req/s):              15.06
Output token throughput (tok/s):         3011.43
Peak output token throughput (tok/s):    10146.00
Peak concurrent requests:                1000.00
Total Token throughput (tok/s):          6242.21
---------------Time to First Token----------------
Mean TTFT (ms):                          8458.72
Median TTFT (ms):                        8006.49
P99 TTFT (ms):                           15322.87
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          311.99
Median TPOT (ms):                        131.91
P99 TPOT (ms):                           1946.11
---------------Inter-token Latency----------------
Mean ITL (ms):                           112.67
Median ITL (ms):                         73.66
P99 ITL (ms):                            516.14
==================================================

Сводка методов оптимизации

Метод оптимизации Прирост пропускной способности
Выбор движка +57.8%
Quantization -
Параллелизм +1.4%
Бэкенд внимания -
Настройка CUDA graph -

Другие сценарии бенчмарка

Мы дополнительно проверили оптимизированную конфигурацию на разных нагрузках, чтобы оценить её обобщаемость.

Базовый скрипт обслуживания модели 
1
vllm serve zai-org/GLM-4.5-Air -tp 4
Результаты базового бенчмарка 
 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):                  498.76
Total input tokens:                      3200000
Total generated tokens:                  9987
Request throughput (req/s):              0.20
Output token throughput (tok/s):         20.02
Peak output token throughput (tok/s):    218.00
Peak concurrent requests:                100.00
Total Token throughput (tok/s):          6435.97
---------------Time to First Token----------------
Mean TTFT (ms):                          248013.29
Median TTFT (ms):                        248015.97
P99 TTFT (ms):                           491884.32
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          305.47
Median TPOT (ms):                        317.93
P99 TPOT (ms):                           319.14
---------------Inter-token Latency----------------
Mean ITL (ms):                           303.40
Median ITL (ms):                         309.84
P99 ITL (ms):                            406.26
==================================================

# random 4K input
============ Serving Benchmark Result ============
Successful requests:                     500
Benchmark duration (s):                  300.98
Total input tokens:                      1999787
Total generated tokens:                  99547
Request throughput (req/s):              1.66
Output token throughput (tok/s):         330.75
Peak output token throughput (tok/s):    1743.00
Peak concurrent requests:                500.00
Total Token throughput (tok/s):          6975.10
---------------Time to First Token----------------
Mean TTFT (ms):                          142573.31
Median TTFT (ms):                        141422.49
P99 TTFT (ms):                           289825.31
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          267.38
Median TPOT (ms):                        295.72
P99 TPOT (ms):                           300.72
---------------Inter-token Latency----------------
Mean ITL (ms):                           266.48
Median ITL (ms):                         295.37
P99 ITL (ms):                            307.18
==================================================

# random 2K input
============ Serving Benchmark Result ============
Successful requests:                     500
Benchmark duration (s):                  121.41
Total input tokens:                      999429
Total generated tokens:                  50000
Request throughput (req/s):              4.12
Output token throughput (tok/s):         411.83
Peak output token throughput (tok/s):    1796.00
Peak concurrent requests:                500.00
Total Token throughput (tok/s):          8643.75
---------------Time to First Token----------------
Mean TTFT (ms):                          60300.00
Median TTFT (ms):                        60039.46
P99 TTFT (ms):                           116422.28
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          204.31
Median TPOT (ms):                        224.99
P99 TPOT (ms):                           225.56
---------------Inter-token Latency----------------
Mean ITL (ms):                           202.57
Median ITL (ms):                         224.72
P99 ITL (ms):                            230.94
==================================================

# random 128 input
============ Serving Benchmark Result ============
Successful requests:                     1000
Benchmark duration (s):                  21.03
Total input tokens:                      127881
Total generated tokens:                  4000
Request throughput (req/s):              47.55
Output token throughput (tok/s):         190.21
Peak output token throughput (tok/s):    952.00
Peak concurrent requests:                1000.00
Total Token throughput (tok/s):          6271.24
---------------Time to First Token----------------
Mean TTFT (ms):                          16583.77
Median TTFT (ms):                        16009.09
P99 TTFT (ms):                           20726.61
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          153.00
Median TPOT (ms):                        153.83
P99 TPOT (ms):                           173.39
---------------Inter-token Latency----------------
Mean ITL (ms):                           114.74
Median ITL (ms):                         153.44
P99 ITL (ms):                            192.15
==================================================
Скрипт обслуживания после оптимизации 
1
2
python3 -m sglang.launch_server --model GLM-4.5-Air --tp-size 4 --ep-size 4 \
    --tool-call-parser glm --reasoning-parser glm45
Результаты бенчмарка после оптимизации 
 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):                  403.99
Total input tokens:                      3200000
Total generated tokens:                  9931
Request throughput (req/s):              0.25
Output token throughput (tok/s):         24.58
Peak output token throughput (tok/s):    332.00
Peak concurrent requests:                100.00
Total Token throughput (tok/s):          7945.66
---------------Time to First Token----------------
Mean TTFT (ms):                          200446.56
Median TTFT (ms):                        199894.57
P99 TTFT (ms):                           397275.29
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          276.41
Median TPOT (ms):                        295.02
P99 TPOT (ms):                           410.95
---------------Inter-token Latency----------------
Mean ITL (ms):                           271.13
Median ITL (ms):                         32.39
P99 ITL (ms):                            3555.22
==================================================

# random 4K input
============ Serving Benchmark Result ============
Successful requests:                     500
Benchmark duration (s):                  230.67
Total input tokens:                      1999787
Total generated tokens:                  99437
Request throughput (req/s):              2.17
Output token throughput (tok/s):         431.09
Peak output token throughput (tok/s):    1680.00
Peak concurrent requests:                500.00
Total Token throughput (tok/s):          9100.71
---------------Time to First Token----------------
Mean TTFT (ms):                          109316.41
Median TTFT (ms):                        113614.39
P99 TTFT (ms):                           222368.09
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          124.68
Median TPOT (ms):                        124.45
P99 TPOT (ms):                           179.33
---------------Inter-token Latency----------------
Mean ITL (ms):                           123.34
Median ITL (ms):                         59.70
P99 ITL (ms):                            294.40
==================================================

# random 2K input
============ Serving Benchmark Result ============
Successful requests:                     500
Benchmark duration (s):                  95.78
Total input tokens:                      999429
Total generated tokens:                  49964
Request throughput (req/s):              5.22
Output token throughput (tok/s):         521.66
Peak output token throughput (tok/s):    3160.00
Peak concurrent requests:                500.00
Total Token throughput (tok/s):          10956.49
---------------Time to First Token----------------
Mean TTFT (ms):                          45224.78
Median TTFT (ms):                        44492.00
P99 TTFT (ms):                           91909.36
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          165.34
Median TPOT (ms):                        163.74
P99 TPOT (ms):                           281.83
---------------Inter-token Latency----------------
Mean ITL (ms):                           162.06
Median ITL (ms):                         63.35
P99 ITL (ms):                            166.19
==================================================

# random 128 input
============ Serving Benchmark Result ============
Successful requests:                     1000
Benchmark duration (s):                  10.14
Total input tokens:                      127881
Total generated tokens:                  4000
Request throughput (req/s):              98.67
Output token throughput (tok/s):         394.66
Peak output token throughput (tok/s):    2166.00
Peak concurrent requests:                1000.00
Total Token throughput (tok/s):          13012.20
---------------Time to First Token----------------
Mean TTFT (ms):                          5882.75
Median TTFT (ms):                        5678.55
P99 TTFT (ms):                           9321.13
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          1246.57
Median TPOT (ms):                        1420.02
P99 TPOT (ms):                           2379.65
---------------Inter-token Latency----------------
Mean ITL (ms):                           747.94
Median ITL (ms):                         59.22
P99 ITL (ms):                            6782.02
==================================================