# AIPerf 0.10.0+ Schema 2.0 — Long-Input Prefill Test Template
# Use case: Isolate memory-bandwidth prefill bottleneck for RAG / long-context workloads
# Concurrency is kept low (10) to avoid queueing artifacts

schema_version: "2.0"
benchmark:
  models:
    items:
      - name: "moonshotai/kimi-k2.6"
  endpoint:
    urls: ["https://ai.noris.de/v1"]
    type: "chat"
    streaming: true
  datasets:
    - name: "synthetic_long"
      type: "synthetic"
      prompt_template:
        content: "Analyze the following topic in depth and provide detailed insights."
      strategy:
        type: "fixed"
        max_total_tokens: 1152
        input_tokens: 1024   # Double the standard 512
        output_tokens: 128   # Short output to isolate prefill
  phases:
    - type: "concurrency"
      name: "profiling"
      duration: 300          # 5 minutes to amortize cold-start
      # Do NOT specify requests — duration drives the phase

# CLI usage:
#   aiperf profile --config this-file.yaml --concurrency 10 --ui none \
#       --api-key "$API_KEY" --artifact-dir results/longinput/

# YAML rules learned the hard way:
#   - datasets is a LIST (starts with "- name:", NOT "default:")
#   - prompt_template and strategy are siblings to name/type
#   - Never use ${VAR} in headers when passing --api-key via CLI