跳到主要内容
版本：最新版

ML-Based Model Selection

ML-based selection learns which model in a decision's candidate pool is most appropriate for a request. It is useful when historical evaluation data carries more information than a fixed priority order or a small set of hand-written rules.

The selector runs after a routing decision has matched. It can choose only from that decision's modelRefs; it does not discover, deploy, or authenticate provider models.

Available selectors​

SelectorHow it choosesConsider it when
KNNCombines recorded quality and speed across similar queriesSimilar requests tend to favor the same model and traceability matters
KMeansMaps a request to a learned clusterThe workload forms stable clusters and lookup cost matters
SVMUses a learned decision boundaryCandidate models separate cleanly in feature space
MLPScores candidates with a neural networkYou have enough data for a non-linear selector and can operate its runtime dependency

There is no universally best selector. Compare each candidate against simple baselines such as a fixed default, random choice, and the best single model on the same held-out dataset.

Before you train​

You need:

  • two or more OpenAI-compatible model endpoints
  • representative queries and ground-truth answers or another defensible scoring method
  • enough repeated coverage to evaluate each candidate model across important workload slices
  • an embedding model that is identical during training and online inference
  • a plan for secrets, rate limits, cost, and retention of model responses

Benchmarking sends every selected query to multiple provider endpoints and stores their responses, quality scores, and latency. Treat the output as sensitive when prompts or responses contain user data.

Dashboard workflow​

Open /ml-setup in the Dashboard to run the guided workflow:

  1. Upload a model-endpoint YAML file and a query JSONL file.
  2. Benchmark the candidate models.
  3. Train one or more selectors.
  4. Define decisions and download a configuration fragment.

The Benchmark and Train steps produce data and model artifacts under the Dashboard's ML data directory.

Current configuration export

The generated ml-model-selection-values.yaml is a migration fragment, not a standalone canonical Router configuration. Its config.model_selection, config.strategy, and config.decisions fields must be reviewed and mapped to global.router.model_selection, global.router.strategy, and routing.decisions in a complete config. Add the required listeners and providers, then run vllm-sr validate --config ... before deployment.

Command-line workflow​

1. Install the training dependencies​

cd src/training/model_selection/ml_model_selection
python3 -m venv .venv
source .venv/bin/activate
python -m pip install -r requirements.txt

2. Prepare evaluation queries​

Use one JSON object per line. ground_truth is required to score model output; category is optional but useful for sliced evaluation.

{"query":"What is the derivative of x^2?","ground_truth":"2x","category":"math","metric":"MATH"}
{"query":"Which city is the capital of France?","ground_truth":"B","category":"other","metric":"em_mc","choices":"A) London B) Paris C) Berlin D) Rome"}

Choose a metric that matches the task. The benchmark supports exact/containment matching, multiple-choice extraction, GSM8K and MATH answer extraction, text F1, and code evaluation. Inspect the benchmark output rather than assuming one metric is suitable for every domain.

3. Describe the candidate endpoints​

Keep credentials in environment variables. Do not write literal API keys into the YAML file.

models:
  - name: local-small
    endpoint: http://localhost:8000/v1
  - name: hosted-model
    endpoint: https://provider.example/v1
    api_key: ${PROVIDER_API_KEY}

4. Benchmark every candidate​

python benchmark.py \
  --queries queries.jsonl \
  --model-config models.yaml \
  --output benchmark-output.jsonl \
  --concurrency 4

Start with low concurrency. Increase it only after confirming that every endpoint can sustain the request rate and that provider rate limits are not distorting latency measurements.

5. Train selectors​

python train.py \
  --data-file benchmark-output.jsonl \
  --output-dir models

By default the script trains KNN, KMeans, SVM, and MLP artifacts. Use --algorithm knn|kmeans|svm|mlp to train one selector, or --skip-mlp when the PyTorch dependency is unavailable. Training accepts cpu, cuda, or mps. The current Router decision factory runs the loaded MLP artifact on CPU; its device field is accepted for compatibility but is not wired to selection.

The output directory contains JSON artifacts such as knn_model.json, kmeans_model.json, svm_model.json, and mlp_model.json. Their contents are specific to the benchmarked model names, embedding model, and feature layout.

6. Configure the Router​

Merge the selector settings into a complete canonical config. This example is a fragment; the full file still needs listeners, providers, and any signals used by the decision.

global:
  router:
    model_selection:
      ml:
        models_path: /models/selection
        embedding_dim: 1024
        knn:
          k: 5
          pretrained_path: /models/selection/knn_model.json

routing:
  decisions:
    - name: math
      description: Route math requests with the trained KNN selector.
      priority: 100
      rules:
        operator: AND
        conditions:
          - type: domain
            name: math
      algorithm:
        type: knn
      modelRefs:
        - model: local-small
        - model: hosted-model

The configured embedding_dim and online embedding model must match the training artifacts. Model names in modelRefs must match the names recorded in the benchmark data and the configured provider aliases.

vllm-sr validate --config config.yaml

Evaluate before rollout​

Use a held-out dataset that was not used to fit or tune the selector. Report:

  • answer-quality metric by workload slice
  • selected-model distribution
  • end-to-end latency and provider cost
  • regret relative to an oracle that picks the best evaluated response
  • comparison with fixed-default, best-single-model, and random baselines
  • failures, timeouts, and excluded samples

Publish the dataset revision, source commit, model revisions, embedding model, hardware, command, and raw report with any headline result. Percentages or QPS without that provenance do not describe expected production performance.

Common problems​

Artifact not found​

Check models_path and each pretrained_path from inside the Router runtime, not only on the host. Mount or package the files at the same paths used by the configuration.

Embedding dimension mismatch​

Use the same embedding model and feature layout for training and inference, and set embedding_dim to the exported artifact's dimension.

Poor held-out quality​

Confirm that model names and labels match, inspect class and domain coverage, look for train/test leakage, and compare against simple baselines. More complex selectors do not compensate for unrepresentative benchmark data.

References​