feat(examples): add gpu-k8s-arc-sglang GPU Kubernetes deployment example - #422
Draft
powderluv wants to merge 1 commit into
Draft
feat(examples): add gpu-k8s-arc-sglang GPU Kubernetes deployment example#422powderluv wants to merge 1 commit into
powderluv wants to merge 1 commit into
Conversation
A self-contained example for a native k0s GPU Kubernetes cluster on SPUR
nodes: ARC (actions-runner-controller) CI runners for a sglang fork, gated
auto-deploy of an sglang serving Deployment, and optional backend RoCE/RDMA
for distributed inference.
Genericized from a working AMD MI355X (gfx950) deployment: hostnames ->
gpu-node-N, provider mount path -> /mnt/scratch, RUNBOOK.md is a fill-in
template. Mesh/rail/pod subnets are RFC1918 example config.
Changes:
- examples/gpu-k8s-arc-sglang/{00-network..50-rdma} staged deploy tree
- k0s + Calico bird native routing over a WireGuard mesh; AMD GPU device plugin
- ARC GPU runner scale set (hand-authored dind spec) + non-GPU deploy runner
with scoped RBAC; gated fork-CI deploy job
- sglang serving Deployment/Service/PVC/Ingress (ROCm, ungated model)
- 50-rdma: RoCEv2 pod exposure + 2-node TP and prefill/decode disagg examples
- bootstrap.sh / teardown.sh / Makefile / README / RUNBOOK template
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
Codecov Report✅ All modified and coverable lines are covered by tests. Additional details and impacted files@@ Coverage Diff @@
## main #422 +/- ##
==========================================
- Coverage 70.86% 70.85% -0.01%
==========================================
Files 140 140
Lines 41849 41849
==========================================
- Hits 29656 29652 -4
- Misses 12193 12197 +4 🚀 New features to boost your workflow:
|
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Adds a self-contained example under
examples/gpu-k8s-arc-sglang/for standing up a native k0s GPU Kubernetes cluster on SPUR nodes: ARC (actions-runner-controller) CI runners for a sglang fork, gated auto-deploy of an sglang serving Deployment, and optional backend RoCE/RDMA for distributed inference.Genericized from a working AMD MI355X (gfx950) 4-node deployment.
What's included
10-cluster/) — k0s + Calico bird native routing over a WireGuard mesh (single overlay).20-gpu/) — AMD GPU device plugin + node labelling.30-arc/) — GPU runner scale set (hand-authored dind spec so/dev/kfd+/dev/drireach the dind sidecar) and a non-GPU deploy runner with least-privilege RBAC; GitHub App setup docs.40-serving/) — sglang Deployment/Service/PVC/Ingress (ROCm, ungated model) + a gated fork-CIdeployjob thatkubectl applys on green CI via an in-cluster ServiceAccount (scoped to thesglangnamespace).50-rdma/) — RoCEv2 pod exposure (hostNetwork + SR-IOV options), 2-node tensor-parallel, and prefill/decode disaggregation (MoRI) reference manifests.bootstrap.sh/teardown.sh/Makefile/README.md/RUNBOOK.mdtemplate.Genericization / safety
Redacted for public publication: hostnames →
gpu-node-N/head-node, provider mount path →/mnt/scratch, the internal operational RUNBOOK replaced with a fill-in template. Mesh/rail/pod subnets (10.44.0.0/16, etc.) are RFC1918 example config the example itself provisions. No secrets/tokens/keys —github-app-secret.example.yamlis a placeholder template.Status
Opened as a draft / review branch. Cluster bring-up, the GPU device plugin, ARC runners, and the gated serving deploy path were validated on real MI355X hardware. The
50-rdma/distributed-inference manifests are references — cross-node RDMA KV-transfer has an open NIC-QoS (nicctlPFC/DCQCN) dependency noted inline.🤖 Generated with Claude Code