v0.9.14

0.9.14 (2026-08-05)

Features

• api: add mxfp4 quantization and correct Blackwell runtime reality (#1381) (382af35)
• controller: opt-in Prometheus scrape annotations on inference pods (#1366) (2e8f8a8)
• dashboards: add a llama.cpp runtime dashboard (#1371) (5fedb88)
• dashboards: add an operator-health row to llmkube-inference (#1369) (ee5145b)
• dashboards: SGLang queue-wait, stage latency, token mix and HiCache L3 panels (#1370) (d1a57ca)
• observability: AMD GPU tier — amdgpu-exporter DaemonSet, metric contract, dashboard reconcile (#1354) (a9e560f)

Bug Fixes

• chart: correct stale Blackwell platform floors (R570 is EOL) (#1380) (445664e)
• chart: make GPU alerts exporter-agnostic instead of DCGM-only (#1372) (402a686)
• controller: correct the speculative-decoding draft flag and dedupe --metrics (#1392) (48d7d49)
• controller: skip model-cache PVC for pvc:// sources (#1363) (9ec0bf9)
• foreman: fail loudly when a revision's prior attempt cannot be restored (#1365) (b9b7685)
• foreman: run the coder gate's Go checks with GOTOOLCHAIN=auto (#1389) (ce828dd)
• router: apply external.model as modelNameOverride on backendRefs (#1398) (847b14d)
• router: compile external backends in the Gateway data plane (#1396) (5045dd0)
• router: rewrite body.model for external backends via bodyMutation (#1401) (2fc6992)
• router: surface a gateway reconcile failure as a Warning Event (#1400) (4145a67)
• router: use the non-deprecated events API for the gateway recorder (#1409) (435dd72)
• skip GitHub auth and repo clone for freeform tasks without a repo (#1350) (c1d0cc1)
• stop wasted-turn context ballooning on max_tokens truncation (#1352) (eaac3ef)
• test: make t.Fatal terminality explicit for staticcheck SA5011 (#1410) (7cf58c7)

Reverts

• remove the inert model-rewrite paths for external backends (#1402) (422b9ef)

Documentation

• add observability maintainer (#1407) (3fb54e2)
• fix autoscaling guide to use a metric llama.cpp actually exports (#1386) (bc3fca2)