docs(training-guides): add GPU slicing with Dynamic Resource Allocation (DRA)

[typhoonzero]

What

Adds a new training-guides solution document showing how to request a slice of a GPU with Kubernetes Dynamic Resource Allocation (DRA) — resource.k8s.io/v1, GA in Kubernetes 1.34 — using the NVIDIA DRA driver, and run a LoRA supervised fine-tune inside that slice with Kubeflow Trainer v2.

The guide teaches the DRA model (DeviceClass → ResourceClaimTemplate → Pod resourceClaims + container resources.claims) and contrasts it with the whole-GPU device-plugin approach, so a small fine-tune no longer has to occupy an entire card.

Contents

• docs/en/training_guides/gpu-slicing-with-dra.mdx — device plugin vs. DRA, the three-object slice model, confirming the driver advertises slices, a DRA smoke test, a Trainer v2 fine-tune confined to a MIG slice, MIG profile sizing, the non-MIG time-slicing / MPS fallback, combining slices with Kueue quota, wiring the same claim into KFP / Volcano / plain Jobs, admin enablement, and troubleshooting.
• docs/en/training_guides/assets/dra/ — MIG and shared-GPU ResourceClaimTemplates, a DRA smoke Pod, a self-contained (no model/dataset download) LoRA-SFT TrainingRuntime, and a TrainJob.
• e2e/cases/c15_dra_gpu_slice.sh — end-to-end case that applies the template + runtime + TrainJob and asserts the fine-tune finishes inside the slice. It self-skips when no gpu.nvidia.com ResourceSlices are advertised, so it stays green on device-plugin-only clusters.
• index.mdx — one new row in the "Pick a path" table.

Validation

• doom lint clean (0 errors, 0 warnings).
• All three built-in DRA objects (both ResourceClaimTemplates + the smoke Pod) pass server-side dry-run against a live Kubernetes 1.34 cluster running the NVIDIA DRA driver v25.8.1 — the resource.k8s.io/v1 exactly request shape, CEL selectors, and the opaque GpuConfig are all accepted by the real API.
• Trainer v2 TrainingRuntime / TrainJob are modeled on the repo's published LlamaFactory runtime and parse cleanly.

Note on the live e2e

The full GPU e2e is deferred, not yet run live: the available dev cluster's only GPU is a HAMi-served A30 (actively serving an inference workload) and Kubeflow Trainer v2 is not installed there — activating the DRA kubelet-plugin on that card would conflict with HAMi and disrupt the running service. The MIG-profile attribute string in the selector follows the documented NVIDIA-driver convention; the guide explicitly directs readers to confirm it against their own kubectl get resourceslices -o yaml, and c15 will pin it down once a HAMi-free DRA GPU node is available.

🤖 Generated with Claude Code

[cloudflare-workers-and-pages]

Deploying alauda-ai with    Cloudflare Pages

Latest commit:	deb0db5
Status:	✅  Deploy successful!
Preview URL:	https://7e67fe79.alauda-ai.pages.dev
Branch Preview URL:	https://docs-gpu-slicing-with-dra.alauda-ai.pages.dev

View logs