---
# Model Matrix for Embedding Performance Tests
# Defines which embedding models run which test scenarios

matrix:
  test_suite: "embedding-models"
  description: "Performance evaluation for vLLM embedding models on CPU"
  vllm_version_required: ">=0.11.0"  # Embedding benchmarking support

  # Test case mapping
  test_cases:
    # Test Case 1.1: Baseline Performance - English Model
    - id: "EMB-BASELINE-GRANITE-EN-R2-EMB512"
      description: "Baseline performance and scalability for English embedding model"
      model: "granite-embedding-english-r2"
      model_path: "ibm-granite/granite-embedding-english-r2"
      architecture: "MiniLM/BERT"
      focus: "Enterprise English Embedding Throughput"
      scenarios:
        - baseline-sweep
      quantization: false
      kv_cache: "1GiB"

    # Test Case 1.2: Baseline Performance - Multilingual Model
    - id: "EMB-BASELINE-GRANITE-ML-EMB512"
      description: "Baseline performance and scalability for Multilingual embedding model"
      model: "granite-embedding-278m-multilingual"
      model_path: "ibm-granite/granite-embedding-278m-multilingual"
      architecture: "XLM-RoBERTa"
      focus: "Multilingual Embedding Throughput"
      scenarios:
        - baseline-sweep
      quantization: false
      kv_cache: "1GiB"

    # Test Case 2.1: Latency and Stability - Multilingual Model
    - id: "EMB-LATENCY-GRANITE-ML-EMB512"
      description: "Latency scaling under concurrent load for Multilingual model"
      model: "granite-embedding-278m-multilingual"
      model_path: "ibm-granite/granite-embedding-278m-multilingual"
      architecture: "XLM-RoBERTa"
      focus: "P99 Latency Scaling for Multilingual Encoder"
      scenarios:
        - latency-concurrent
      quantization: false
      kv_cache: "1GiB"

  # Embedding model definitions
  embedding_models:
    # English Dense Encoder (MiniLM/BERT family)
    - name: "granite-embedding-english-r2"
      full_name: "ibm-granite/granite-embedding-english-r2"
      gated: false  # Public model (Apache 2.0), no token required
      architecture_family: "MiniLM/BERT"
      application_focus: "Encoder-Only (English Dense)"
      language: "English"
      parameters: "~110M"
      dimensions: 384
      max_sequence_length: 512
      default_scenarios:
        - baseline-sweep

    # Multilingual Dense Encoder (XLM-RoBERTa family)
    - name: "granite-embedding-278m-multilingual"
      full_name: "ibm-granite/granite-embedding-278m-multilingual"
      gated: false  # Public model (Apache 2.0), no token required
      architecture_family: "XLM-RoBERTa"
      application_focus: "Encoder-Only (Multilingual)"
      language: "Multilingual"
      parameters: "278M"
      dimensions: 768
      max_sequence_length: 512
      alternate_models:
        - "granite-embedding-107m-multilingual"
        - "Multilingual-e5-large"
      default_scenarios:
        - baseline-sweep
        - latency-concurrent

  # Common test parameters
  common_parameters:
    workload: "Embedding (512:1)"
    input_sequence_length: 512
    output_sequence_length: 1
    affinity: "auto"  # Handled by vLLM automatically
    dtype: "bfloat16"
    kv_cache_space: "1GiB"  # Minimal for embedding models (encoder-only)
    num_prompts: 1000

  # Platform configuration (vendor-specific)
  platform_config:
    cores: "VENDOR-DEFINED"
    numa_nodes: "VENDOR-DEFINED"
    notes: |
      - Core count and NUMA configuration should be optimized per platform
      - Intel Xeon platforms may benefit from specific core binding
      - AMD EPYC platforms may require different NUMA settings

  # Expected test duration (approximate)
  estimated_duration:
    baseline_sweep:
      per_model: "20-30 minutes"
      stages: 4  # inf, 25%, 50%, 75%
    latency_concurrent:
      per_model: "25-40 minutes"
      stages: 5  # 16, 32, 64, 128, 196 concurrency

  # Results storage structure
  results_structure: |
    results/embedding-models/
    ├── granite-embedding-english-r2/
    │   └── baseline/
    │       ├── sweep-inf.json
    │       ├── sweep-25pct.json
    │       ├── sweep-50pct.json
    │       └── sweep-75pct.json
    └── granite-embedding-278m-multilingual/
        ├── baseline/
        │   ├── sweep-inf.json
        │   ├── sweep-25pct.json
        │   ├── sweep-50pct.json
        │   └── sweep-75pct.json
        └── latency/
            ├── concurrent-16.json
            ├── concurrent-32.json
            ├── concurrent-64.json
            ├── concurrent-128.json
            └── concurrent-196.json