Default Spectrum-X settings from GPU platform

[almaslennikov]

Summary

• derive multiplaneMode and numberOfPlanes from GPU platform plus the east-west NIC inventory
• default H100/H200/B200/GB200 to none/1 and CX8 B300/GB300 to the conservative GA swplb/2 path
• resolve --for preset hardware before profile defaults without persisting synthetic preset inventory
• complete explicit single-plane pairs and reject multiplane modes combined with one plane
• update user, reference, default-config, and bundled skill documentation

Platform decision

The current Network Operator documentation defines both SWPLB and HWPLB for B300 and GB300. Platform type therefore cannot identify which load-balancing mechanism is deployed. Launch Kit defaults the documented CX8 combinations to the GA SWPLB path; HWPLB remains an explicit site-topology choice. Quad-plane B300 also remains explicit because B300 supports both two and four planes.

Sources:

• https://docs.nvidia.com/networking/display/kubernetes2640/spectrum-x/spectrum-x.html
• https://docs.nvidia.com/networking/display/kubernetes2640/spectrum-x/quick-start.html

Validation

• go test -race -count=1 ./...
• CGO_ENABLED=0 go build ./...
• CGO_ENABLED=0 go vet ./...
• golangci-lint v2.11 run ./...
• mkdocs build --strict

[greptile-apps]

Greptile Summary

This PR derives Spectrum-X mode and plane-count defaults from GPU platform and east-west NIC inventory, while making preset hardware available before profile resolution without persisting synthetic inventory.

• Defaults H100, H200, B200, and GB200 to none/1 and B300 and GB300 to swplb/2.
• Completes explicit single-plane pairs and rejects load-balancing modes configured with one plane.
• Reorders --for preset synthesis so preset hardware drives defaults while preserving the source clusterConfig during write-back.
• Updates tests, default configuration comments, user documentation, and agent-facing skill references.

Confidence Score: 5/5

The PR appears safe to merge; no concrete changed-code defect remains after reviewing the default-resolution, validation, preset, and persistence paths.

The new platform mappings are covered for canonical discovery and preset identities, contradictory mode/plane pairs are rejected before rendering, and preset inventory drives resolution without being persisted over the source hardware configuration.

Important Files Changed

Filename	Overview
pkg/resolve/defaults.go	Adds platform-aware Spectrum-X defaults, single-plane companion completion, normalized NIC identifiers, and heterogeneous-group consistency checks.
pkg/resolve/validate.go	Rejects swplb and hwplb configurations that specify only one plane.
pkg/app/generate.go	Synthesizes --for preset inventory before profile resolution and uses a shallow write-back copy to preserve the source hardware inventory.
pkg/app/generate_profile_test.go	Verifies preset hardware drives defaults without replacing persisted source inventory.
pkg/resolve/defaults_test.go	Covers supported platform mappings, NIC fallback behavior, conflicting heterogeneous platforms, and inferred single-plane companions.
pkg/resolve/validate_test.go	Covers rejection of load-balancing modes paired with a single plane.
pkg/config/default-config.yaml	Updates embedded configuration guidance to describe platform-derived Spectrum-X defaults.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart LR
  A[Load source configuration] --> B{--for preset?}
  B -->|No| C[Use discovered clusterConfig]
  B -->|Yes| D[Synthesize transient preset clusterConfig]
  C --> E[Apply hardware defaults]
  D --> E
  E --> F[Apply explicit CLI overrides]
  F --> G[Validate resolved profile]
  G --> H[Render manifests]
  G --> I[Write resolved profile back]
  I --> J[Restore source clusterConfig for persistence]

Loading

_{Reviews (1): Last reviewed commit: "Default Spectrum-X settings from GPU pla..." | Re-trigger Greptile}