Federation scheduler + KServe renderer (managed-kai)#63

Closed

dennis-upbound wants to merge 56 commits into

mainfrom

dennis/scheduler-1pager

Commits on May 6, 2026

design: scheduler & capability model 1-pager + deliverables (preview)

authored and
Dennis Ramdass
committed
Address Bassam's review: substrate / runtime split + node selector

authored and
Dennis Ramdass
committed

Commits on May 7, 2026

Address Bassam's review (full): #56 claims, replica==placement, ModelService

authored and
Dennis Ramdass
committed
Coherence pass + revert vocab to cluster-scoped

authored and
Dennis Ramdass
committed
Collapse node+device claims, delegate to in-cluster scheduler, v1/v2 cleanup

authored and
Dennis Ramdass
committed
Opinionated Kueue default + dual-path matching (label or DRA) + open questions

authored and
Dennis Ramdass
committed
Make pluggable substrate concrete: BYO scheduler + backend contract

authored and
Dennis Ramdass
committed
Clarify IR == existing ModelPlacement CRD; spell out 'intermediate representation'

authored and
Dennis Ramdass
committed
Refactor to broader API design framing + address Bassam's gpt-oss-20b notes

authored and
Dennis Ramdass
committed
Drop redundant modelplane-api/README; fold unique content into design appendix

authored and
Dennis Ramdass
committed
Once-over: tighten adapter substrate prose, drop stale open Qs, add diagram

authored and
Dennis Ramdass
committed
Rename modelplane-api/ -> proposed-modelplane-api/; tighten diagram

authored and
Dennis Ramdass
committed
Use Bassam's whiteboard verbatim for the diagram

authored and
Dennis Ramdass
committed
Reframe: federation planner vs DRA runtime allocator; instance-type macro

authored and
Dennis Ramdass
committed
Address Bassam's PR comments: drop requiredEngineFeatures, consumer-index discipline

authored and
Dennis Ramdass
committed
Align with Bassam's GPU-hardware taxonomy + reference clusters from cloud SKUs

authored and
Dennis Ramdass
committed
Fine-tooth pass on examples: fix layer mismatches, stale keys, factual errors

authored and
Dennis Ramdass
committed
Tighten doc + add EKS-no-DRA, multi-region, L40S-4x examples

authored and
Dennis Ramdass
committed
Reframe consumer-index discipline without comparing to a non-existent baseline

authored and
Dennis Ramdass
committed
Drop accidentally committed design/.DS_Store

authored and
Dennis Ramdass
committed
Second fine-tooth pass: layering, region attrs, vendor pin, macro split

authored and
Dennis Ramdass
committed
Add structural linter for the design examples

authored and
Dennis Ramdass
committed
Drop Problem + Who owns what; compress Open questions to a decision table

authored and
Dennis Ramdass
committed
Ignore design/.DS_Store

authored and
Dennis Ramdass
committed
Ignore personal notes (*.local.md, nic-1on1.md)

authored and
Dennis Ramdass
committed
Replace CapabilityVocabulary with InferenceClass; engine catalog goes to code

authored and
Dennis Ramdass
committed
Rename ModelService → InferenceProvider; drop supportedModels

authored and
Dennis Ramdass
committed
Make stack layering, BYO matrix, and break-glass scenarios explicit

authored and
Dennis Ramdass
committed
Restructure examples by resource type; add managed/BYO/break-glass examples; inline links in doc

authored and
Dennis Ramdass
committed
Final coherence pass: stale-ref + table sanity + missing rows

authored and
Dennis Ramdass
committed
Drop 'Nic-owned' / 'sketch' framing on InferenceProvider

authored and
Dennis Ramdass
committed
Frame KServeBackend.spec.engine as a proposed extension; drop examples/README

authored and
Dennis Ramdass
committed
Add KServeBackend extension XRD; drop sub-READMEs and Excalidraw

authored and
Dennis Ramdass
committed
Move modelplane-api.md under proposed-modelplane-api/

authored and
Dennis Ramdass
committed
Sweep cleanup: stale 'deleted' framing, missing intro, LINT.md vocab ref

authored and
Dennis Ramdass
committed
ModelEndpoint label selectors on InferenceProvider routes; reshape multi-region example

authored and
Dennis Ramdass
committed
WIP: Nic's API sketch
negz
committed

Commits on May 8, 2026

Iterate on simplified Modelplane API sketch
negz
committed
Drop serving profiles from ModelDeployment
negz
committed
Rename parallelism to topology, add instances for P:D ratio
negz
committed
Both KAI and Kueue first-class; add placement walkthroughs + multi-tenancy
Dennis Ramdass
committed
Refocus on scheduling; rename ModelPlacement→ModelReplica; defer API shape to #64
Dennis Ramdass
committed
Drop XRDs (defer to #64); add IR + lifecycle + Quickstart sections
Dennis Ramdass
committed
Split design doc into quickstart / advanced / scheduling / design
Dennis Ramdass
committed
Sketch composition functions for new API; compress docs to design pointer
Dennis Ramdass
committed
KAI / Kueue scheduler integration sketch
Dennis Ramdass
committed
Modularize composition functions; add 69 unit tests
Dennis Ramdass
committed
gitignore Python caches + .venv-test
Dennis Ramdass
committed
Align with K8s SIG-Scheduling vocab; document delta from main's scheduler
Dennis Ramdass
committed
Drop migration framing; describe scheduler properties directly
Dennis Ramdass
committed
Realign composition functions to Nic's style
Dennis Ramdass
committed
Narrow MR to managed-kai only; defer plugin system to follow-up
Dennis Ramdass
committed
Nest topology under workers, add workers.count for P:D ratio
negz
committed

Commits on May 9, 2026

Add cloud-specific provisioning to InferenceClass
negz
committed
Merge remote-tracking branch 'origin/pages' into dennis/scheduler-1pager
Dennis Ramdass
committed
Update to Nic's latest API shape: workers.{count, topology}
Dennis Ramdass
committed