NVIDIA A100 GPU üzerinde Qwen3-14B işlem hacmi optimizasyonu

Sonuç

A100 GPU üzerinde Qwen3-14B işlem hacmini optimize etmek için önerilen yapılandırma:

Model sunum komutu

trtllm-serve Qwen/Qwen3-14B --max_seq_len=40960 --enable_chunked_prefill

Optimizasyon öncesi ve sonrası benchmark sonuçlarının karşılaştırması:

Benchmark senaryosu	temel yapılandırma (optimizasyonsuz vLLM)	Optimize edilmiş
ShareGPT	Toplam TPS: 3922.41 Ortalama TPOT(ms): 114.34	Toplam TPS: 5770.24 (+47.1%) Ortalama TPOT(ms): 260.46
Kısa istem	Toplam TPS: 4916.63 Ortalama TPOT(ms): 233.35	Toplam TPS: 8762.17 (+78.2%) Ortalama TPOT(ms): 715.19
Orta uzunlukta istem	Toplam TPS: 5967.13 Ortalama TPOT(ms): 301.73	Toplam TPS: 7474.00 (+25.3%) Ortalama TPOT(ms): 313.69
Uzun istem	Toplam TPS: 5153.42 Ortalama TPOT(ms): 254.61	Toplam TPS: 6418.65 (+24.6%) Ortalama TPOT(ms): 196.10
Çok uzun istem	Toplam TPS: 4430.14 Ortalama TPOT(ms): 445.58	Toplam TPS: 4413.82 (-0.4%) Ortalama TPOT(ms): 493.18

Not

Benchmark testlerimiz olası tüm optimizasyon birleşimlerini kapsamaz. Örneğin, varsayılan yapılandırmasında en iyi performansı gösteren çıkarım motorunu sonraki ayarlamalar için başlangıç noktası olarak seçeriz. Bu eleme yaklaşımı yerel bir optimum verir; bu, genel optimum olmayabilir.
En büyük toplu iş boyutu, zamanlama yapılandırması, genişletilmiş KV önbelleği, CUDA graph, Torch Compile gibi belirli kullanım senaryolarına bağlı başka optimizasyonlar da vardır. Bu belgedeki sonuçlar, daha hedefli optimizasyonlar için başlangıç noktası olabilir.
Testler belirli donanım ve yazılım yapılandırmalarında yürütülmüştür. Çıkarım motorlarındaki gelişmeler sonuçları değiştirebilir.

Eksik noktalar veya yeni değişiklikleri yansıtacak güncellemeler varsa bize bildirin.

Optimizasyon hedefi

Yüksek eş zamanlı isteklerde yüksek işlem hacmi sağlar.

Deney kurulumu

Model

Qwen3-14B

Donanım

Tek NVIDIA A100 SXM GPU.

Motor sürümü

vLLM: v0.11.0
SGLang: v0.5.5.post1
TensorRT-LLM: v1.2.0rc1

Benchmark veri kümesi

ShareGPT
Farklı dizi uzunluklarına sahip rastgele veri kümesi:
- Çok uzun istem: 32000 giriş token'ı, 100 çıkış token'ı
- Uzun istem: 4000 giriş token'ı, 200 çıkış token'ı
- Orta uzunlukta istem: 2000 giriş token'ı, 100 çıkış token'ı
- Kısa istem: 128 giriş token'ı, 4 çıkış token'ı

Benchmark betiği

Model performansını ölçmek için vLLM bench CLI aracını kullanıyoruz. Benchmark şu komutla çalıştırılır:

# 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

Deney sonuçları

1. Çıkarım motoru seçimi

vLLM

Model sunum betiği

1	`vllm serve Qwen/Qwen3-14B`

Benchmark sonucu

============ 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

Model sunum betiği

python3 -m sglang.launch_server --model-path Qwen/Qwen3-14B

Benchmark sonucu

============ 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

Model sunum betiği

1	`trtllm-serve Qwen/Qwen3-14B`

Benchmark sonucu

============ 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
==================================================

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

2. TensorRT-LLM/vLLM için Quantization

TensorRT-LLM FP8

Model sunum betiği

1	`trtllm-serve Qwen/Qwen3-14B-FP8`

Benchmark sonucu

1	`# RuntimeError: Unsupported SM version for FP8 block scaling GEMM`

TensorRT-LLM AWQ

Model sunum betiği

1	`trtllm-serve Qwen/Qwen3-14B-AWQ`

Benchmark sonucu

1	`# KeyError: 'weight'`

vLLM FP8

Model sunum betiği

1	`vllm serve Qwen/Qwen3-14B-FP8`

Benchmark sonucu

============ 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

Model sunum betiği

1	`vllm serve Qwen/Qwen3-14B-AWQ`

Benchmark sonucu

============ 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 için Toplu işteki en fazla token sayısı

Model sunum betiği

trtllm-serve Qwen/Qwen3-14B --max_num_tokens=16384

Benchmark sonucu

# --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
==================================================

Optimizasyon seçeneklerinin özeti

Optimizasyon seçeneği	İşlem hacmi iyileşmesi
Motor seçimi	+47.1%
Quantization	-
Toplu işteki en fazla token sayısı	-

Diğer benchmark senaryoları

Optimize edilmiş yapılandırmanın farklı iş yüklerine genellenebilirliğini değerlendirmek için ek benchmark testleri yaptık.

Temel model sunum betiği

1	`vllm serve Qwen/Qwen3-14B`

Temel benchmark sonuçları

# 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
==================================================

Optimize edilmiş model sunum betiği

trtllm-serve Qwen/Qwen3-14B --max_seq_len=40960 --enable_chunked_prefill

Optimize edilmiş benchmark sonuçları

# 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
==================================================