v0.6.0

Highlights

Scale on real serving-engine signals instead of proxying GPU utilization, plus MIG-aware collection and a native gRPC health check.

Added

• Inference-server scaling — vLLM queue depth & KV-cache usage (#28), Triton queue-wait & request-rate (#111), and a TGI tgi-inference profile (#64) — read directly from each engine's /metrics endpoint.
• MIG per-instance metrics — one entry per MIG compute instance via NVML (is_mig_instance, parent_index, mig_profile), with CUDA_VISIBLE_DEVICES MIG-UUID resolution for SLURM/Flux.
• gRPC health checking (#109) — implements the gRPC Health Checking Protocol; a background checker polls NVML every --health-check-interval (default 30s) and reports SERVING/NOT_SERVING.
• p95 / p99 multi-GPU aggregation (#69) — percentile reduction so one hot GPU doesn't dominate max.
• Flux shell plugin (#60) — deploy/flux/gpu-monitor.lua auto-starts/stops gpu-metrics alongside a Flux job.

Fixed

• gpuIndex values below -1 are now rejected with a clear error (#128).

Upgrade note

Helm liveness/readiness probes now default to the native gRPC probe type (probes.type: grpc, requires Kubernetes 1.24+). On older clusters set probes.type: http to keep the previous /healthz / /readyz probes.

Install

helm install keda-gpu-scaler oci://ghcr.io/pmady/charts/keda-gpu-scaler --version 0.6.0

Contributors

@venkata22a — inference-server scalers, MIG metrics, and gRPC health checking

Full Changelog: v0.5.0...v0.6.0