Оптимизация пропускной способности Qwen3-235B-A22B на NVIDIA A100 GPU

Вывод

Рекомендуемая конфигурация для оптимизации пропускной способности Qwen3-235B-A22B-Instruct-2507 на A100 GPU:

Команда обслуживания модели

vllm serve Qwen/Qwen3-235B-A22B-Instruct-2507-FP8 -tp 4

Сравнение результатов бенчмарков до и после оптимизации:

Сценарий бенчмарка	базовая конфигурация (vLLM без оптимизации)	После оптимизации
ShareGPT	Общий TPS: 3792.50 Среднее TPOT(ms): 123.36	Общий TPS: 3019.19x2 (+59.2%) Среднее TPOT(ms): 150.72
Короткий запрос	Общий TPS: 7449.43 Среднее TPOT(ms): 155.16	Общий TPS: 4961.66x2 (+33.2%) Среднее TPOT(ms): 265.87
Запрос средней длины	Общий TPS: 9342.67 Среднее TPOT(ms): 190.79	Общий TPS: 5333.91x2 (+14.2%) Среднее TPOT(ms): 342.39
Длинный запрос	Общий TPS: 8170.72 Среднее TPOT(ms): 205.64	Общий TPS: 4633.41x2 (+13.4%) Среднее TPOT(ms): 347.13
Очень длинный запрос	Общий TPS: 6657.49 Среднее TPOT(ms): 298.86	Общий TPS: 4132.51x2 (+24.1%) Среднее TPOT(ms): 481.71

Примечание

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.
Есть и другие методы оптимизации, зависящие от сценария использования: максимальный размер батча, настройки планировщика, расширенный KV-кэш, CUDA graph, Torch Compile и т. д. Выводы этого документа могут служить отправной точкой для более целенаправленной оптимизации.
Тесты проводились на определённых аппаратных и программных конфигурациях. Развитие движков инференса может привести к новым выводам.
Квантизация может влиять на точность, однако для большинства моделей FP8 обеспечивает снижение точности менее чем на 1%. Подробности см. в результатах оценки. Поэтому для обслуживания с высокой пропускной способностью настоятельно рекомендуется квантизация FP8.

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

Qwen3-235B-A22B-Instruct-2507

Оборудование

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

Версия движка

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

Набор данных бенчмарка

ShareGPT
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:

# 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

Результаты эксперимента

1. Выбор движка инференса

vLLM

Скрипт обслуживания модели

vllm serve Qwen/Qwen3-235B-A22B-Instruct-2507 -tp 8

Результат бенчмарка

============ Serving Benchmark Result ============
Successful requests:                     1000
Benchmark duration (s):                  109.00
Total input tokens:                      217393
Total generated tokens:                  196004
Request throughput (req/s):              9.17
Output token throughput (tok/s):         1798.14
Peak output token throughput (tok/s):    3297.00
Peak concurrent requests:                1000.00
Total Token throughput (tok/s):          3792.50
---------------Time to First Token----------------
Mean TTFT (ms):                          38065.91
Median TTFT (ms):                        37597.63
P99 TTFT (ms):                           77082.58
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          123.36
Median TPOT (ms):                        103.68
P99 TPOT (ms):                           448.36
---------------Inter-token Latency----------------
Mean ITL (ms):                           100.76
Median ITL (ms):                         86.83
P99 ITL (ms):                            185.41
==================================================

SGLang

Скрипт обслуживания модели

python3 -m sglang.launch_server --model-path Qwen/Qwen3-235B-A22B-Instruct-2507 --tp-size 8

Результат бенчмарка

============ Serving Benchmark Result ============
Successful requests:                     1000
Benchmark duration (s):                  89.35
Total input tokens:                      217393
Total generated tokens:                  196763
Request throughput (req/s):              11.19
Output token throughput (tok/s):         2202.21
Peak output token throughput (tok/s):    4926.00
Peak concurrent requests:                1000.00
Total Token throughput (tok/s):          4635.31
---------------Time to First Token----------------
Mean TTFT (ms):                          21130.67
Median TTFT (ms):                        15483.85
P99 TTFT (ms):                           51960.77
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          235.10
Median TPOT (ms):                        163.74
P99 TPOT (ms):                           1171.03
---------------Inter-token Latency----------------
Mean ITL (ms):                           136.35
Median ITL (ms):                         86.55
P99 ITL (ms):                            500.15
==================================================

Результат: SGLang(4635.31 tok/s) > vLLM (3792.50 tok/s)

2. Quantization в vLLM

SGLang с квантизацией FP8 не работает на GPU A100. vLLM поддерживает выполнение с квантизацией w8a16 на A100 через ядро Marlin.

vLLM FP8

Скрипт обслуживания модели

# TP8 only is not compatible for the model weight. Use EP as well.
vllm serve Qwen/Qwen3-235B-A22B-Instruct-2507-FP8 -tp 8 --enable-expert-parallel

Результат бенчмарка

============ Serving Benchmark Result ============
Successful requests:                     1000
Benchmark duration (s):                  110.68
Total input tokens:                      217393
Total generated tokens:                  196648
Request throughput (req/s):              9.04
Output token throughput (tok/s):         1776.73
Peak output token throughput (tok/s):    3106.00
Peak concurrent requests:                1000.00
Total Token throughput (tok/s):          3740.89
---------------Time to First Token----------------
Mean TTFT (ms):                          37366.87
Median TTFT (ms):                        35525.98
P99 TTFT (ms):                           78616.18
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          115.74
Median TPOT (ms):                        109.21
P99 TPOT (ms):                           228.96
---------------Inter-token Latency----------------
Mean ITL (ms):                           101.28
Median ITL (ms):                         91.72
P99 ITL (ms):                            235.47
==================================================

3. Параллелизм в vLLM

TP4

Скрипт обслуживания модели

vllm serve Qwen/Qwen3-235B-A22B-Instruct-2507-FP8 -tp 4

Результат бенчмарка

============ Serving Benchmark Result ============
Successful requests:                     1000
Benchmark duration (s):                  136.64
Total input tokens:                      217393
Total generated tokens:                  195140
Request throughput (req/s):              7.32
Output token throughput (tok/s):         1428.16
Peak output token throughput (tok/s):    2852.00
Peak concurrent requests:                1000.00
Total Token throughput (tok/s):          3019.19
---------------Time to First Token----------------
Mean TTFT (ms):                          45879.29
Median TTFT (ms):                        43216.75
P99 TTFT (ms):                           99903.30
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          150.72
Median TPOT (ms):                        143.39
P99 TPOT (ms):                           297.17
---------------Inter-token Latency----------------
Mean ITL (ms):                           131.17
Median ITL (ms):                         119.69
P99 ITL (ms):                            311.46
==================================================

TP4 EP4

Скрипт обслуживания модели

vllm serve Qwen/Qwen3-235B-A22B-Instruct-2507-FP8 -tp 4 --enable-expert-parallel

Результат бенчмарка

============ Serving Benchmark Result ============
Successful requests:                     1000
Benchmark duration (s):                  146.74
Total input tokens:                      217393
Total generated tokens:                  195590
Request throughput (req/s):              6.81
Output token throughput (tok/s):         1332.94
Peak output token throughput (tok/s):    2501.00
Peak concurrent requests:                1000.00
Total Token throughput (tok/s):          2814.47
---------------Time to First Token----------------
Mean TTFT (ms):                          48930.06
Median TTFT (ms):                        45498.27
P99 TTFT (ms):                           106740.39
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          161.49
Median TPOT (ms):                        154.84
P99 TPOT (ms):                           329.32
---------------Inter-token Latency----------------
Mean ITL (ms):                           140.35
Median ITL (ms):                         128.38
P99 ITL (ms):                            340.60
==================================================

4. Максимальное число токенов в батче в vLLM

Скрипт обслуживания модели

vllm serve Qwen/Qwen3-235B-A22B-Instruct-2507-FP8 -tp 4 --max-num-batched-tokens 4096

Результат бенчмарка

# --max-num-batched-tokens 4096
============ Serving Benchmark Result ============
Successful requests:                     1000
Benchmark duration (s):                  136.64
Total input tokens:                      217393
Total generated tokens:                  195953
Request throughput (req/s):              7.32
Output token throughput (tok/s):         1434.07
Peak output token throughput (tok/s):    2784.00
Peak concurrent requests:                1000.00
Total Token throughput (tok/s):          3025.05
---------------Time to First Token----------------
Mean TTFT (ms):                          45389.84
Median TTFT (ms):                        41651.71
P99 TTFT (ms):                           98975.37
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          155.37
Median TPOT (ms):                        142.48
P99 TPOT (ms):                           509.56
---------------Inter-token Latency----------------
Mean ITL (ms):                           131.19
Median ITL (ms):                         114.03
P99 ITL (ms):                            395.26
==================================================

# --max-num-batched-tokens 8192
============ Serving Benchmark Result ============
Successful requests:                     1000
Benchmark duration (s):                  137.59
Total input tokens:                      217393
Total generated tokens:                  195781
Request throughput (req/s):              7.27
Output token throughput (tok/s):         1422.97
Peak output token throughput (tok/s):    2724.00
Peak concurrent requests:                1000.00
Total Token throughput (tok/s):          3003.01
---------------Time to First Token----------------
Mean TTFT (ms):                          45405.05
Median TTFT (ms):                        41603.31
P99 TTFT (ms):                           99900.26
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          157.10
Median TPOT (ms):                        143.37
P99 TPOT (ms):                           561.98
---------------Inter-token Latency----------------
Mean ITL (ms):                           132.24
Median ITL (ms):                         120.38
P99 ITL (ms):                            306.32
==================================================

Сводка методов оптимизации

Метод оптимизации	Прирост пропускной способности
Выбор движка	+22.2%
Quantization	-
Параллелизм	+59.2%
Бэкенд внимания	-
Максимальное число токенов в батче	-

Другие сценарии бенчмарка

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

Базовый скрипт обслуживания модели

vllm serve Qwen/Qwen3-235B-A22B-Instruct-2507 -tp 8

Результаты базового бенчмарка

# random 32K input
============ Serving Benchmark Result ============
Successful requests:                     100
Benchmark duration (s):                  481.89
Total input tokens:                      3200000
Total generated tokens:                  8191
Request throughput (req/s):              0.21
Output token throughput (tok/s):         17.00
Peak output token throughput (tok/s):    140.00
Peak concurrent requests:                100.00
Total Token throughput (tok/s):          6657.49
---------------Time to First Token----------------
Mean TTFT (ms):                          240229.05
Median TTFT (ms):                        240137.10
P99 TTFT (ms):                           475251.39
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          298.86
Median TPOT (ms):                        307.94
P99 TPOT (ms):                           311.82
---------------Inter-token Latency----------------
Mean ITL (ms):                           295.13
Median ITL (ms):                         303.36
P99 ITL (ms):                            370.10
==================================================

# random 4K input
============ Serving Benchmark Result ============
Successful requests:                     500
Benchmark duration (s):                  255.70
Total input tokens:                      1997942
Total generated tokens:                  91352
Request throughput (req/s):              1.96
Output token throughput (tok/s):         357.26
Peak output token throughput (tok/s):    1403.00
Peak concurrent requests:                500.00
Total Token throughput (tok/s):          8170.72
---------------Time to First Token----------------
Mean TTFT (ms):                          122190.64
Median TTFT (ms):                        121615.01
P99 TTFT (ms):                           244882.42
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          205.64
Median TPOT (ms):                        220.33
P99 TPOT (ms):                           237.80
---------------Inter-token Latency----------------
Mean ITL (ms):                           204.69
Median ITL (ms):                         239.57
P99 ITL (ms):                            246.21
==================================================

# random 2K input
============ Serving Benchmark Result ============
Successful requests:                     500
Benchmark duration (s):                  112.15
Total input tokens:                      998175
Total generated tokens:                  49632
Request throughput (req/s):              4.46
Output token throughput (tok/s):         442.54
Peak output token throughput (tok/s):    1691.00
Peak concurrent requests:                500.00
Total Token throughput (tok/s):          9342.67
---------------Time to First Token----------------
Mean TTFT (ms):                          54722.50
Median TTFT (ms):                        54532.60
P99 TTFT (ms):                           106756.03
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          190.79
Median TPOT (ms):                        208.61
P99 TPOT (ms):                           209.09
---------------Inter-token Latency----------------
Mean ITL (ms):                           189.23
Median ITL (ms):                         208.52
P99 ITL (ms):                            211.58
==================================================

# random 128 input
============ Serving Benchmark Result ============
Successful requests:                     1000
Benchmark duration (s):                  17.68
Total input tokens:                      127698
Total generated tokens:                  4000
Request throughput (req/s):              56.56
Output token throughput (tok/s):         226.26
Peak output token throughput (tok/s):    948.00
Peak concurrent requests:                1000.00
Total Token throughput (tok/s):          7449.43
---------------Time to First Token----------------
Mean TTFT (ms):                          12767.63
Median TTFT (ms):                        12409.68
P99 TTFT (ms):                           17339.37
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          155.16
Median TPOT (ms):                        158.22
P99 TPOT (ms):                           168.04
---------------Inter-token Latency----------------
Mean ITL (ms):                           116.37
Median ITL (ms):                         157.64
P99 ITL (ms):                            179.04
==================================================

Скрипт обслуживания после оптимизации

vllm serve Qwen/Qwen3-235B-A22B-Instruct-2507-FP8 -tp 4

Результаты бенчмарка после оптимизации

# random 32K input
============ Serving Benchmark Result ============
Successful requests:                     100
Benchmark duration (s):                  776.36
Total input tokens:                      3200000
Total generated tokens:                  8300
Request throughput (req/s):              0.13
Output token throughput (tok/s):         10.69
Peak output token throughput (tok/s):    152.00
Peak concurrent requests:                100.00
Total Token throughput (tok/s):          4132.51
---------------Time to First Token----------------
Mean TTFT (ms):                          387908.55
Median TTFT (ms):                        387968.52
P99 TTFT (ms):                           766523.82
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          481.71
Median TPOT (ms):                        496.29
P99 TPOT (ms):                           513.27
---------------Inter-token Latency----------------
Mean ITL (ms):                           476.10
Median ITL (ms):                         490.85
P99 ITL (ms):                            612.03
==================================================

# random 4K input
============ Serving Benchmark Result ============
Successful requests:                     500
Benchmark duration (s):                  450.71
Total input tokens:                      1997942
Total generated tokens:                  90389
Request throughput (req/s):              1.11
Output token throughput (tok/s):         200.55
Peak output token throughput (tok/s):    1066.00
Peak concurrent requests:                500.00
Total Token throughput (tok/s):          4633.41
---------------Time to First Token----------------
Mean TTFT (ms):                          216485.01
Median TTFT (ms):                        215365.76
P99 TTFT (ms):                           436587.02
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          347.13
Median TPOT (ms):                        371.93
P99 TPOT (ms):                           412.29
---------------Inter-token Latency----------------
Mean ITL (ms):                           344.46
Median ITL (ms):                         425.81
P99 ITL (ms):                            436.76
==================================================

# random 2K input
============ Serving Benchmark Result ============
Successful requests:                     500
Benchmark duration (s):                  196.42
Total input tokens:                      998175
Total generated tokens:                  49495
Request throughput (req/s):              2.55
Output token throughput (tok/s):         251.99
Peak output token throughput (tok/s):    1116.00
Peak concurrent requests:                500.00
Total Token throughput (tok/s):          5333.91
---------------Time to First Token----------------
Mean TTFT (ms):                          95261.76
Median TTFT (ms):                        94982.78
P99 TTFT (ms):                           189238.63
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          342.39
Median TPOT (ms):                        376.64
P99 TPOT (ms):                           377.99
---------------Inter-token Latency----------------
Mean ITL (ms):                           339.59
Median ITL (ms):                         376.62
P99 ITL (ms):                            380.69
==================================================

# random 128 input
============ Serving Benchmark Result ============
Successful requests:                     1000
Benchmark duration (s):                  26.54
Total input tokens:                      127698
Total generated tokens:                  4000
Request throughput (req/s):              37.67
Output token throughput (tok/s):         150.70
Peak output token throughput (tok/s):    311.00
Peak concurrent requests:                1000.00
Total Token throughput (tok/s):          4961.66
---------------Time to First Token----------------
Mean TTFT (ms):                          17260.17
Median TTFT (ms):                        17240.71
P99 TTFT (ms):                           26154.32
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          265.87
Median TPOT (ms):                        272.95
P99 TPOT (ms):                           278.77
---------------Inter-token Latency----------------
Mean ITL (ms):                           199.40
Median ITL (ms):                         272.50
P99 ITL (ms):                            288.23
==================================================