model-name: openai/gpt-oss-20b
input-tokens: 73
output-tokens: 512
latency: 0.2
results:
    - concurrency: 1
      generation-speed: 56.75
      prompt-throughput: 102.85
      max-ttft: 0.71
      min-ttft: 0.71
      success-rate: 1
    - concurrency: 2
      generation-speed: 71.85
      prompt-throughput: 32.16
      max-ttft: 4.54
      min-ttft: 2.71
      success-rate: 1
    - concurrency: 4
      generation-speed: 138.9
      prompt-throughput: 74.88
      max-ttft: 3.9
      min-ttft: 1.6
      success-rate: 1
    - concurrency: 8
      generation-speed: 215.52
      prompt-throughput: 49.87
      max-ttft: 11.71
      min-ttft: 1.64
      success-rate: 1
    - concurrency: 16
      generation-speed: 353.45
      prompt-throughput: 134.57
      max-ttft: 8.68
      min-ttft: 1.25
      success-rate: 1
    - concurrency: 32
      generation-speed: 590.24
      prompt-throughput: 179.83
      max-ttft: 12.99
      min-ttft: 1.49
      success-rate: 1
    - concurrency: 64
      generation-speed: 960.35
      prompt-throughput: 311.89
      max-ttft: 14.98
      min-ttft: 0
      success-rate: 1
    - concurrency: 128
      generation-speed: 1630.4
      prompt-throughput: 440.14
      max-ttft: 21.23
      min-ttft: 0
      success-rate: 1
    - concurrency: 256
      generation-speed: 2275.96
      prompt-throughput: 420.71
      max-ttft: 44.42
      min-ttft: 2.19
      success-rate: 1
    - concurrency: 512
      generation-speed: 2289.36
      prompt-throughput: 415.2
      max-ttft: 90.02
      min-ttft: 2.34
      success-rate: 1
    - concurrency: 1024
      generation-speed: 2306.21
      prompt-throughput: 387.12
      max-ttft: 193.1
      min-ttft: 0
      success-rate: 1