v0.9.8.9-rc.1

LoxiLB Inference Gateway v0.9.8.9-rc.1.

Container images (multi-arch, amd64+arm64):

• ghcr.io/loxilb-io/loxilb-inference-gateway:v0.9.8.9-rc.1 (Ubuntu 22.04 base)
• ghcr.io/loxilb-io/loxilb-inference-gateway:v0.9.8.9-rc.1-u24 (Ubuntu 24.04 base)

Attached artifacts: .deb (Ubuntu 20.04/22.04/24.04), .rpm
(RHEL/Rocky/Alma 9), binary tarballs, SHA256SUMS, and an SPDX
SBOM per image variant. See the README for install instructions
and the kernel baseline required by the eBPF datapath.

Images are scanned for fixable CRITICAL CVEs before they are
published, signed with Cosign (keyless), and carry SLSA build
provenance and SBOM attestations. Verify with:

cosign verify ghcr.io/loxilb-io/loxilb-inference-gateway:v0.9.8.9-rc.1 \
  --certificate-identity-regexp '^https://github\.com/loxilb-io/loxilb-inference-gateway/' \
  --certificate-oidc-issuer https://token.actions.githubusercontent.com
gh attestation verify oci://ghcr.io/loxilb-io/loxilb-inference-gateway:v0.9.8.9-rc.1 --owner loxilb-io

This is a fork of loxilb-io/loxilb
with inference-aware L4/L7 routing for LLM serving fleets.

What's Changed

• feat(api): identify the product on GET /version for client flavor detection by @NLX-SeokHwanKong in #64
• fix(cicd): mirror the hardened docker-inspect probe into the seagull common.sh by @NLX-SeokHwanKong in #58
• fix(docker): link the GHCR package to this repository by @NLX-SeokHwanKong in #61
• fix(docker): keep host build artifacts out of the image context by @NLX-SeokHwanKong in #67
• fix(api): reject security=3 instead of silently serving plaintext by @NLX-SeokHwanKong in #66
• fix(api): page and filter GET /logs over the whole file by @NLX-SeokHwanKong in #65
• feat(metrics): scope CPU utilization to the container it runs in by @NLX-SeokHwanKong in #68
• SGLang P/D disaggregation support + engine-dialect refactor by @NLX-SeokHwanKong in #69
• chore(hygiene): catch campaign test ids, work items and private-tree links by @NLX-SeokHwanKong in #70
• feat(gateway): TensorRT-LLM as a first-class engine — KV event plane, P/D dialect, admission guards by @NLX-SeokHwanKong in #71
• feat: CHWBL enhancements — origin-5xx demotion for plain-LB and user-prefix affinity fallback by @NLX-SeokHwanKong in #73
• feat: llama.cpp (llama-server) as the fourth supported engine — typed plain-LB with CHWBL affinity by @NLX-SeokHwanKong in #72
• feat: AI-QoS — per-tenant/per-model token quotas, rule policers, and the Tier-1 L7 byte shaper by @NLX-SeokHwanKong in #74
• fix: restore the AI-QoS series onto main (stranded by the #72/#74 stacked-merge race) by @NLX-SeokHwanKong in #76
• fix(api): declare model_name on keyed loadbalancer DELETE routes by @backguynn in #77
• AI-QoS backlog: shaper observability, per-tenant token burst, transit-egress policing by @NLX-SeokHwanKong in #78
• feat(cicd): k3d in-cluster inference-gateway scenarios (Service + Gateway API) by @backguynn in #79
• Authentication plane separation: per-service key policy, PostgreSQL credential stores, fail-closed outages by @NLX-SeokHwanKong in #80
• Multi-LLM operations guide and load-balancing doc refresh by @NLX-SeokHwanKong in #81

New Contributors

• @backguynn made their first contribution in #77

Full Changelog: v0.9.8.7...v0.9.8.9-rc.1

v0.9.8.7

LoxiLB Inference Gateway v0.9.8.7.

Container images (multi-arch, amd64+arm64):

• ghcr.io/loxilb-io/loxilb-inference-gateway:v0.9.8.7 (Ubuntu 22.04 base)
• ghcr.io/loxilb-io/loxilb-inference-gateway:v0.9.8.7-u24 (Ubuntu 24.04 base)

Attached artifacts: .deb (Ubuntu 20.04/22.04/24.04), .rpm
(RHEL/Rocky/Alma 9), binary tarballs, SHA256SUMS, and an SPDX
SBOM per image variant. See the README for install instructions
and the kernel baseline required by the eBPF datapath.

Images are scanned for fixable CRITICAL CVEs before they are
published, signed with Cosign (keyless), and carry SLSA build
provenance and SBOM attestations. Verify with:

cosign verify ghcr.io/loxilb-io/loxilb-inference-gateway:v0.9.8.7 \
  --certificate-identity-regexp '^https://github\.com/loxilb-io/loxilb-inference-gateway/' \
  --certificate-oidc-issuer https://token.actions.githubusercontent.com
gh attestation verify oci://ghcr.io/loxilb-io/loxilb-inference-gateway:v0.9.8.7 --owner loxilb-io

This is a fork of loxilb-io/loxilb
with inference-aware L4/L7 routing for LLM serving fleets.

What's Changed

• chore: post-public-flip cleanup — fork-safe CI, internal-ID scrub, issue templates by @NLX-SeokHwanKong in #54
• fix(docker): build gobgp from source to clear the release CVE gate by @NLX-SeokHwanKong in #55
• chore: bump loxilb-ebpf submodule to the internal-ID scrub tip by @NLX-SeokHwanKong in #56
• fix(cicd): restore the KV hash core and repair scrub-damaged shell arithmetic by @NLX-SeokHwanKong in #57
• fix(cicd): stop the datapath before deleting the pkg-smoke topology by @NLX-SeokHwanKong in #59
• fix(cicd): make tier3 deb smoke opt-in and drop eBPF refs before netdev delete by @NLX-SeokHwanKong in #60

Full Changelog: mcp/v0.1.0...v0.9.8.7

loxilb-mcp 0.1.0

Standalone loxilb-mcp 0.1.0. See mcp/README.md for install and configuration.

loxilb-mcp 0.1.0-rc.2

Standalone loxilb-mcp 0.1.0-rc.2. See mcp/README.md for install and configuration.

loxilb-mcp 0.1.0-rc.1

Standalone loxilb-mcp 0.1.0-rc.1. See mcp/README.md for install and configuration.

Nightly packages (rolling)

Rolling nightly .deb packages for loxilb-inference-gateway,
rebuilt from main each night. Not a versioned release — for
official, supported artifacts see the tagged releases produced by
release.yml.

Current build: v0.9.8.9-nightly.20260901