# Generated from run-config.json. Contains no credentials or cluster endpoints.
schema_version: 1
source: "run-config.json"
tested_configuration:
  schema: "stable-upstream-batch-interference-benchmark-v1"
  date: "2026-08-10"
  topology:
    endpoint_picker_replicas: 1
    model_replicas: 1
    gpus_per_model_replica: 1
  hardware:
    gpu_per_model_replica: "NVIDIA H100"
  model_service:
    model: "GPT-OSS 20B"
    runtime: "Red Hat AI Inference Server vLLM"
    tensor_parallel_size: 1
    max_model_len: 32768
    max_num_batched_tokens: 8192
    max_num_seqs: 128
    gpu_memory_utilization: 0.9
    prefix_cache: "disabled"
  endpoint_picker:
    version: "llm-d Endpoint Picker v0.9.0"
    flow_control_gate: "enabled"
    detector: "request-concurrency"
    max_concurrency: 128
    headroom: 0.1
    picker: "random"
    priority_bands:
      - 100
      - 50
      - 0
      - -10
    fairness: "round-robin within each priority band"
    reserved_capacity: "not configured"
    batch_eviction: "not configured"
  traffic:
    production_scenarios:
      - name: "realtime-only-reference"
        duration_s: 240
        analysis_windows:
          - name: "surge"
            start_s: 80
            end_s: 160
          - name: "recovery"
            start_s: 190
            end_s: 230
        tenants:
          - fairness_id: "realtime-chat"
            priority: 100
            input_tokens: 4096
            output_tokens: 128
            phases:
              - start_s: 60
                duration_s: 120
                rate_pattern: "noisy_sinusoidal"
                rate_center: 3.0
                rate_amplitude: 0.6
                period_s: 37
                phase_offset: 0.1
                rate_noise: 0.3
                rate_spike_probability: 0.02
      - name: "realtime-with-batch-already-running"
        duration_s: 240
        analysis_windows:
          - name: "baseline"
            start_s: 20
            end_s: 55
          - name: "surge"
            start_s: 80
            end_s: 160
          - name: "recovery"
            start_s: 190
            end_s: 230
        tenants:
          - fairness_id: "realtime-chat"
            priority: 100
            input_tokens: 4096
            output_tokens: 128
            phases:
              - start_s: 60
                duration_s: 120
                rate_pattern: "noisy_sinusoidal"
                rate_center: 3.0
                rate_amplitude: 0.6
                period_s: 37
                phase_offset: 0.1
                rate_noise: 0.3
                rate_spike_probability: 0.02
          - fairness_id: "batch-long-context"
            priority: -10
            input_tokens: 20000
            output_tokens: 128
            phases:
              - start_s: 0
                duration_s: 180
                rate_pattern: "noisy_sinusoidal"
                rate_center: 2.75
                rate_amplitude: 0.55
                period_s: 31
                phase_offset: 0.4
                rate_noise: 0.275
                rate_spike_probability: 0.02
  execution:
    arrival: "open-loop Poisson replay with noisy sinusoidal rates"
    production_repeats_per_arm: 3
    matched_realtime_trace: true
    cache_mode: "off"
  runner:
    scenario_source: "pipeline/benchmark.py"
    sha256: "3811ec26c46bf3a26fa643698ec54bf569bb4bc99c3ea22ca18f805cb077b8e0"
    traffic_driver: "GuideLLM"
    guidellm_version: "0.7.0"
    trace_compiler: "pipeline/guidellm_trace.py"
    launcher: "pipeline/run_guidellm_scenario.py"
    scenario_file: "scenarios.json"