Sequence Length Distributions for Advanced Benchmarking

April 17, 2026 · View on GitHub

Sequence length distributions allow you to specify complex patterns of input sequence length (ISL) and output sequence length (OSL) pairs with different probabilities and optional variance. This enables benchmarking of multiple use cases such as summarization and Q&A on one endpoint.

Overview

The sequence distribution feature provides benchmarking of mixed workloads with different ISL and OSL pairings.

Basic Usage

Example command

Add variance to make workloads more realistic:

aiperf profile \
    --model Qwen/Qwen3-0.6B \
    --endpoint-type chat \
    --endpoint /v1/chat/completions \
    --streaming \
    --url localhost:8000 \
    --sequence-distribution "64|10,32|8:70;256|40,128|20:20;1024|100,512|50:10" \

This creates:

  • 70% of requests with ISL ~ Normal(64, 10), OSL ~ Normal(32, 8)
  • 20% of requests with ISL ~ Normal(256, 40), OSL ~ Normal(128, 20)
  • 10% of requests with ISL ~ Normal(1024, 100), OSL ~ Normal(512, 50)

Values are automatically clamped to be at least 1.

Supported Formats

Basic:

"ISL1,OSL1:PROB1;ISL2,OSL2:PROB2;..."

With standard deviations:

"ISL1|STDDEV1,OSL1|STDDEV1:PROB1;ISL2|STDDEV2,OSL2|STDDEV2:PROB2"

2. Bracket Format

Basic:

"[(ISL1,OSL1):PROB1,(ISL2,OSL2):PROB2]"

With standard deviations:

"[(256|10,128|5):60,(512|20,256|15):40]"

3. JSON Format

Basic:

{"pairs": [{"isl": 256, "osl": 128, "prob": 60}, {"isl": 512, "osl": 256, "prob": 40}]}

With standard deviations:

{"pairs": [
  {"isl": 256, "isl_stddev": 10, "osl": 128, "osl_stddev": 5, "prob": 60},
  {"isl": 512, "isl_stddev": 20, "osl": 256, "osl_stddev": 15, "prob": 40}
]}

Examples

Example Case: Chatbot Workload Simulation

# Simulate typical chatbot traffic:
# - 70% short queries (quick questions)
# - 20% medium queries (explanations)
# - 10% long queries (complex tasks)

aiperf profile \
    --model Qwen/Qwen3-0.6B \
    --endpoint-type chat \
    --endpoint /v1/chat/completions \
    --streaming \
    --url localhost:8000 \
    --sequence-distribution "64|10,32|8:70;256|40,128|20:20;1024|100,512|50:10"

Sample Output (Successful Run):

INFO     Starting AIPerf System
INFO     Using sequence distribution: 70% (ISL~N(64,10), OSL~N(32,8)), 20% (ISL~N(256,40), OSL~N(128,20)), 10% (ISL~N(1024,100), OSL~N(512,50))
INFO     AIPerf System is PROFILING

Profiling: 100/100 |████████████████████████| 100% [02:15<00:00]

INFO     Benchmark completed successfully
INFO     Results saved to: artifacts/Qwen_Qwen3-0.6B-chat-concurrency1/

            NVIDIA AIPerf | LLM Metrics
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━┳━━━━━━━━┳━━━━━━━━━┳━━━━━━━━━┳━━━━━━━━┓
┃                         Metric ┃    avg ┃    min ┃     max ┃     p99 ┃    p50 ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━╇━━━━━━━━╇━━━━━━━━━╇━━━━━━━━━╇━━━━━━━━┩
│           Request Latency (ms) │ 567.34 │ 123.45 │ 4567.89 │ 4234.56 │ 345.67 │
│       Time to First Token (ms) │  78.90 │  23.45 │  234.56 │  212.34 │  67.89 │
│       Inter Token Latency (ms) │  14.23 │  11.34 │   19.45 │   18.90 │  13.89 │
│ Input Sequence Length (tokens) │ 189.45 │  48.00 │ 1234.00 │ 1156.78 │  67.00 │
│    Output Token Count (tokens) │  89.34 │  24.00 │  634.00 │  589.23 │  34.00 │
│     Request Throughput (req/s) │   7.45 │      - │       - │       - │      - │
└────────────────────────────────┴────────┴────────┴─────────┴─────────┴────────┘

JSON Export: artifacts/Qwen_Qwen3-0.6B-chat-concurrency1/profile_export_aiperf.json