integrations/tgi/prometheus_metadata.yaml

platforms:
  - type: GKE
    launch_stage: GA
    detections:
      - characteristic_metric:
          metric_type: prometheus.googleapis.com/tgi_request_count/counter
    exporter_metadata:
      name: TGI
      doc_url: https://github.com/huggingface/text-generation-inference
      minimum_supported_version: "2.0.2"
    default_metrics:
      - name: prometheus.googleapis.com/tgi_batch_concat/counter
        prometheus_name: tgi_batch_concat
        kind: CUMULATIVE
        value_type: DOUBLE
      - name: prometheus.googleapis.com/tgi_batch_concat_duration/histogram
        prometheus_name: tgi_batch_concat_duration
        kind: CUMULATIVE
        value_type: DISTRIBUTION
      - name: prometheus.googleapis.com/tgi_batch_current_max_tokens/gauge
        prometheus_name: tgi_batch_current_max_tokens
        kind: GAUGE
        value_type: DOUBLE
      - name: prometheus.googleapis.com/tgi_batch_current_size/gauge
        prometheus_name: tgi_batch_current_size
        kind: GAUGE
        value_type: DOUBLE
      - name: prometheus.googleapis.com/tgi_batch_decode_duration/histogram
        prometheus_name: tgi_batch_decode_duration
        kind: CUMULATIVE
        value_type: DISTRIBUTION
      - name: prometheus.googleapis.com/tgi_batch_filter_duration/histogram
        prometheus_name: tgi_batch_filter_duration
        kind: CUMULATIVE
        value_type: DISTRIBUTION
      - name: prometheus.googleapis.com/tgi_batch_forward_duration/histogram
        prometheus_name: tgi_batch_forward_duration
        kind: CUMULATIVE
        value_type: DISTRIBUTION
      - name: prometheus.googleapis.com/tgi_batch_inference_count/counter
        prometheus_name: tgi_batch_inference_count
        kind: CUMULATIVE
        value_type: DOUBLE
      - name: prometheus.googleapis.com/tgi_batch_inference_duration/histogram
        prometheus_name: tgi_batch_inference_duration
        kind: CUMULATIVE
        value_type: DISTRIBUTION
      - name: prometheus.googleapis.com/tgi_batch_inference_success/counter
        prometheus_name: tgi_batch_inference_success
        kind: CUMULATIVE
        value_type: DOUBLE
      - name: prometheus.googleapis.com/tgi_queue_size/gauge
        prometheus_name: tgi_queue_size
        kind: GAUGE
        value_type: DOUBLE
      - name: prometheus.googleapis.com/tgi_request_count/counter
        prometheus_name: tgi_request_count
        kind: CUMULATIVE
        value_type: DOUBLE
      - name: prometheus.googleapis.com/tgi_request_duration/histogram
        prometheus_name: tgi_request_duration
        kind: CUMULATIVE
        value_type: DISTRIBUTION
      - name: prometheus.googleapis.com/tgi_request_generated_tokens/histogram
        prometheus_name: tgi_request_generated_tokens
        kind: CUMULATIVE
        value_type: DISTRIBUTION
      - name: prometheus.googleapis.com/tgi_request_inference_duration/histogram
        prometheus_name: tgi_request_inference_duration
        kind: CUMULATIVE
        value_type: DISTRIBUTION
      - name: prometheus.googleapis.com/tgi_request_input_length/histogram
        prometheus_name: tgi_request_input_length
        kind: CUMULATIVE
        value_type: DISTRIBUTION
      - name: prometheus.googleapis.com/tgi_request_max_new_tokens/histogram
        prometheus_name: tgi_request_max_new_tokens
        kind: CUMULATIVE
        value_type: DISTRIBUTION
      - name: prometheus.googleapis.com/tgi_request_mean_time_per_token_duration/histogram
        prometheus_name: tgi_request_mean_time_per_token_duration
        kind: CUMULATIVE
        value_type: DISTRIBUTION
      - name: prometheus.googleapis.com/tgi_request_queue_duration/histogram
        prometheus_name: tgi_request_queue_duration
        kind: CUMULATIVE
        value_type: DISTRIBUTION
      - name: prometheus.googleapis.com/tgi_request_skipped_tokens/histogram
        prometheus_name: tgi_request_skipped_tokens
        kind: CUMULATIVE
        value_type: DISTRIBUTION
      - name: prometheus.googleapis.com/tgi_request_success/counter
        prometheus_name: tgi_request_success
        kind: CUMULATIVE
        value_type: DOUBLE
      - name: prometheus.googleapis.com/tgi_request_validation_duration/histogram
        prometheus_name: tgi_request_validation_duration
        kind: CUMULATIVE
        value_type: DISTRIBUTION
    install_documentation_url: https://cloud.google.com/stackdriver/docs/managed-prometheus/exporters/tgi