How to use this folder
Planning
This folder is the home for forward-looking planning for FlexInfer (feature inventory, roadmap slices, and implementation plans).
If you’re looking for “what exists today” docs, start here instead:
- Project overview:
README.md - Roadmap (high-level):
ROADMAP.md - Implementation status (detailed):
docs/IMPLEMENTATION_STATUS.md - Architecture overview:
docs/dev/architecture.md - CRD / contract docs:
docs/specs/README.md - v1alpha2 user docs:
docs/user/models-v1alpha2.md
How to use this folder
feature-inventory.md: what features exist, what’s stable, what’s experimental, and what’s missing.next-roadmap.md: the next series of features/enhancements to implement (prioritized).spec-capsule-template.md: reusable source-backed template, checklist, and examples for new feature slices.slice-readiness-gate.md: ready-for-implementation gate for target modules, agent delegation notes, validation, generated artifacts, and rollback notes.roadmap-reconciliation.md: post-merge checklist for keeping roadmap docs,.loomcontext, and GitLab issues aligned.context-curve-benchmark.md: spec capsule for the reporting-only long-context benchmark curve MVP.docker-major-dependency-rollout.md: rollout sequencing, matrix, and rollback plan for issue#21major Docker base updates.rocm-gfx1100-deploy-swap-tracing-slice.md: PR-2 readiness/proof capsule for gfx1100 deploy/swap controls, latency metrics, and opt-in tracing.context-aware-router-execution.md: execution checklist for issue#8(context-aware routing).phase-1-controller-api-hardening.md: PR-sized checklist for the next series of controller hardening work.../design/multi-tenancy.md: M1 design for namespace-isolated tenancy baseline.phase-5-multi-cluster.md: forward-looking checklist for multi-cluster federation (future).- Runtime/dev benchmark docs:
../dev/context-curve-benchmarking.mdand../dev/gemma4-benchmarking.md.
Current Focus (February 2026)
- Mixed-vendor homelab clusters (k3s) with AMD ROCm
gfx1100(RX 7900) and NVIDIA Maxwellsm_52(GTX 980 Ti). - Making scheduling decisions "obviously correct" by keeping node labels + telemetry stable:
flexinfer.ai/gpu.vendor,flexinfer.ai/gpu.arch,flexinfer.ai/gpu.vramflexinfer.ai/gpu-free-memory(scheduler headroom scoring)
- Backend guardrails and docs for the two problem-child GPU classes:
- ROCm gfx1100 guide:
docs/user/backends-rocm-gfx1100.md - Maxwell guide:
docs/user/backends-maxwell.mdandbuild/README-maxwell.md
- ROCm gfx1100 guide:
Reference homelab topology used while hardening these paths:
cblevins-7900xtx: AMD RX 7900 XTX (ROCmgfx1100)cblevins-5930k: AMD RX 7900 XTX (ROCmgfx1100)cblevins-gtx980ti: NVIDIA GTX 980 Ti (Maxwellsm_52)
Operational note (k3s homelab): if a node goes NotReady/unreachable, old FlexInfer pods can get stuck Terminating on that node and block Helm rollouts. Force-delete the stuck pods in flexinfer-system to unblock upgrades.
Planning checklist
Before implementation starts on a multi-file feature or operational workflow
change, copy spec-capsule-template.md, then use slice-readiness-gate.md to
mark it Ready only after filling in:
- Goal, non-goals, users/operators, and current evidence.
- Requirements and acceptance criteria.
- PR-sized implementation slices with target files/modules.
- Agent delegation notes for parallel work, including safe-to-edit modules, files/modules to avoid, local verification, and expected output signals.
- Validation commands and generated artifacts.
- Rollout/backout notes for Helm, CRD, Flux, and runtime-image changes.
- After merge, run
roadmap-reconciliation.mdbefore closing tracking issues.
Loom context pack
For deeper “planning artifacts” (product spec, implementation plan, decisions/worklog), see:
.loom/00-index.md