[sttts]

Incorporated upstream; closing as superseded. This selected-shape correction landed through ai-dynamo/dynamo#17308, including both DGDR consumers. The source PR #17262 is also superseded. The separate remaining asyncio change is tracked in #17640. Branches and review history are preserved.

Summary

This is an on-top fix for ai-dynamo/dynamo#17262 and the heterogeneous-node review finding in discussion_r4403133517.

The source PR correctly discovers total same-GPU-type inventory, but used that total directly as the executable profiling budget. A cluster with 8-GPU and 4-GPU nodes of the same type therefore produced a budget of 12 even though one deployment shape cannot schedule 12 GPUs across those unequal nodes.

This change separates the two concepts without adding a Kubernetes API field:

• GPUInfo.TotalGPUs remains diagnostic same-type inventory.
• GPUInfo.SelectedShapeGPUs counts only nodes matching the selected GPU type, GPUs per node, and VRAM per GPU.
• v1beta1 automatic spec.hardware.totalGpus uses selected-shape capacity, capped at 32 as before.
• Explicit totalGpus remains authoritative.
• Unknown GPU types are grouped by normalized model identity rather than treating every unknown model as the same empty SKU.

Relates to ai-dynamo#17192.

API behavior

All hardware fields remain optional. No Kubernetes schema field is added.

spec:
  hardware:
    gpuSku: h100_sxm        # OPTIONAL; discovered when omitted
    numGpusPerNode: 8       # OPTIONAL; selected shape
    vramMb: 81920           # OPTIONAL; selected shape
    # totalGpus: omitted    # OPTIONAL; defaults to selected-shape capacity, capped at 32

With same-type 8-GPU and 4-GPU nodes, discovery records inventory 12 but defaults totalGpus to 8.

A caller may choose another shape by supplying its budget explicitly:

spec:
  hardware:
    gpuSku: h100_sxm
    numGpusPerNode: 4
    totalGpus: 4

A shape override without an explicit budget is rejected because discovery cannot derive executable capacity for that different shape:

spec:
  hardware:
    gpuSku: h100_sxm
    numGpusPerNode: 4
    # INVALID when discovery selected 8 GPUs per node: totalGpus is required

Discovered nodes	Inventory	Selected-shape capacity	Automatic budget
8 + 4, same type/VRAM	12	8	8
8 + 8, same type/VRAM	16	16	16
8 + 8, same type/different VRAM	16	8	8
8 + 4, explicit totalGpus: 64	12	8	64

Generated artifacts

The v1beta1 CRD, Pydantic model, raw/full API references, and curated DGDR reference are regenerated or aligned with the Go API comments.

Validation

• go test ./internal/gpu -count=1
• discovery table covers both 8+4 and 4+8, proving the maximum selected shape is independent of node observation order
• KUBEBUILDER_ASSETS=... go test ./internal/controller -count=1 (73.511s)
• golangci-lint v2.15.0 run --timeout=5m ./... (0 issues)
• make manifests generate-pydantic generate-api-docs
• python3 docs/fern/scripts/gen_kubernetes_api.py --check
• python3 docs/fern/scripts/docs_lint.py (0 errors; existing advisory warnings only)
• git diff --check

[github-actions]

👋 Hi sttts! Thank you for contributing to Pranjulcr7/dynamo.

Just a reminder: The NVIDIA Test Github Validation CI runs an essential subset of the testing framework to quickly catch errors.Your PR reviewers may elect to test the changes comprehensively before approving your changes.

🚀