Skip to content

Releases: aitra-ai/aitra-meter

Aitra Meter v0.2.4

15 Jun 02:53
7ad2329

Choose a tag to compare

GitHub Actions dependency updates.

Changed

  • Bumped 10 GitHub Actions dependencies to latest versions

Aitra Meter v0.2.3

15 Jun 02:53
1dd169e

Choose a tag to compare

Opt-in OTLP metric export and pre-built Grafana dashboard.

Added

  • internal/export/otlp/ — OTLP emission of gen_ai.infrastructure.energy.* metrics via OTel Collector
  • Helm values: otel.enabled (default false), otel.endpoint, otel.protocol
  • Prometheus and OTLP pipelines run simultaneously when otel.enabled=true
  • helm/aitra-meter/files/grafana-dashboard.json — auto-provisions via Grafana sidecar with grafana.enabled=true
  • examples/alerting-rules.yaml — three reference PrometheusRule alerting rules: efficiency regression, GPU idle, measurement unstable

Aitra Meter v0.2.2

15 Jun 02:53

Choose a tag to compare

Zeus energy provider via Unix socket IPC, plus two derived efficiency metrics.

Added

  • internal/provider/energy/zeus/zeus.go — full Zeus implementation over /tmp/zeus.sock JSON-RPC (begin_window, end_window, idle_power, devices)
  • aitra_tokens_per_joule — inverse of J/token, efficiency in the output direction
  • aitra_gpu_utilization_efficiency — output tokens per watt of GPU power draw
  • Both metrics computed in ReportWindow from data already present in every WindowReport — no new provider calls

Aitra Meter v0.2.1

15 Jun 02:53
88fcb54

Choose a tag to compare

generic-prometheus inference provider — compatible with TGI, SGLang, Ollama, and Triton. Metric names and model label are configurable via environment variables. Implements the full InferenceMetricsProvider interface (ADR 0004).

Added

  • internal/provider/inference/genericprometheus/ — full implementation with configurable metric names
  • Table-driven tests covering metric scraping and model label resolution

Aitra Meter v0.2.0

15 Jun 02:52
d6a4e5b

Choose a tag to compare

SQLite backend via modernc.org/sqlite — pure Go, no CGO, no additional server required. Every measurement record is persisted. Storage backends remain pluggable via the Backend interface (ADR 0005).

Added

  • internal/storage/sqlite/ — full WriteBatch and QueryChargeback implementation
  • internal/storage/sqlite/sqlite_test.go — integration test: 52k rows, 30-day chargeback query under 10s
  • modernc.org/sqlite v1.34.5 added to go.mod

aitra-meter-0.1.0

22 May 22:42

Choose a tag to compare

Open-source Kubernetes-native AI inference efficiency measurement