Blog running llm d without kubernetes (llm-d#333)

* running-llm-d-without-kubernetes

Signed-off-by: Ezra Silvera <ezra@il.ibm.com>

* Update blog author image

Signed-off-by: Pete Cheslock <pete.cheslock@redhat.com>

* blog(running-llm-d-without-kubernetes): FlowControl works in file-discovery mode

Update the parity section: FlowControl now works in file-discovery mode
with priority bands, fairness, ordering, and usage-limit policies
configured statically in EndpointPickerConfig.flowControl. Per-request
priority falls back to the configured default when no InferenceObjective
CRD is available; the x-flow-fairness-id header drives fairness within
a band.

Model-name rewriting (InferenceModelRewrite) is now the only CRD-driven
feature that remains unavailable outside Kubernetes.

Signed-off-by: Ezra Silvera <ezra@il.ibm.com>

* updated blog

Signed-off-by: Ezra Silvera <ezra@il.ibm.com>

* blog(running-llm-d-without-kubernetes): reframe around the hard problem

Shift the post from a deployment recipe toward the deep-technical llm-d
style, per review feedback. Lead with why routing intelligence was
entangled with Kubernetes and the EndpointDiscovery abstraction that
separates discovery from routing, rather than step-by-step deploy
commands (now covered by the upstream guide).

- Add a 'hard problem' section: RL rollout cost, inference workload
  diversity and naive framework routers, research-lab/HPC and
  benchmarking motivation.
- Slim the setup to a minimal example; trim Ray/Slurm to concept and
  link the guide for the full scripts.
- Fold discovery-mechanism extensions into the dynamic-substrates
  section; focus 'Research directions' on RL rollout gaps.
- Preserve inbound anchors (llm-d#1-the-endpoints-file, #troubleshooting,
  #parity-with-the-kubernetes-native-llm-d-deployment).

Signed-off-by: Ezra Silvera <ezra@il.ibm.com>

* blog(running-llm-d-without-kubernetes): drop endpoints-file schema table

Replace the per-field table with a one-line summary of the essential
fields and a pointer to the guide for the full reference. Keeps the
section lean; the llm-d#1-the-endpoints-file anchor is preserved.

Signed-off-by: Ezra Silvera <ezra@il.ibm.com>

* blog(running-llm-d-without-kubernetes): trim inline configs to match house style

Other technical posts (kvcache-wins, intelligent-scheduling,
native-kv-cache-offloading) keep deploy config in the guides and carry
only the idea inline. Drop the endpoints.yaml block and the
EndpointPickerConfig block, keeping the memorable one-liner
(dataLayer.discovery.pluginRef: file-discovery) in prose. Also soften
the closing line: drop weight-sync (llm-d does not own it) and replace
'for the fifth time' with 'yet again'.

Signed-off-by: Ezra Silvera <ezra@il.ibm.com>

* blog(running-llm-d-without-kubernetes): drop Troubleshooting section

The section opened with 'The full failure-mode list is in the guide' and
duplicated operational failure modes that belong in the guide. No other
technical post carries a troubleshooting reel; the blog stays conceptual
and defers how-to-debug to the guide.

Signed-off-by: Ezra Silvera <ezra@il.ibm.com>

* blog(running-llm-d-without-kubernetes): reframe parity section around substrate invariance

Replace the feature checklist with the architectural claim: because the
discovery seam is below the scheduler, file-discovery mode gets the whole
router unchanged, including capabilities not yet built - one routing
stack, substrate-invariant by construction. Keep the honest exceptions
(InferenceObjective priority, InferenceModelRewrite, endpoint-lifecycle
ownership) framed as inputs being migrated behind the same plugin pattern.
Retitle to 'One routing stack, every substrate'.

Signed-off-by: Ezra Silvera <ezra@il.ibm.com>

* blog(running-llm-d-without-kubernetes): add platform diagram, reframe substrate as platform

Add Figure 3 (llm-d-platform.drawio.svg): a draw.io diagram showing llm-d's
EndpointDiscovery module (Kube Discovery, File Discovery) discovering endpoints
on and serving requests from Kubernetes, Slurm, Ray, and bare metal, with the
RL frameworks (veRL, SkyRL, prime-rl) on top.

Replace the term 'substrate' with 'platform' throughout the post and caption.

Signed-off-by: Ezra Silvera <ezra@il.ibm.com>

* blog(running-llm-d-without-kubernetes): move platform diagram to the top as Figure 1

Move the platform overview diagram above the truncate so it serves as an
orienting hero (and shows on the blog listing), with a big-picture caption.
Renumber the file-discovery and Ray diagrams to Figures 2 and 3, and leave a
back-reference to Figure 1 in the 'One routing stack, every platform' section.

Signed-off-by: Ezra Silvera <ezra@il.ibm.com>

* blog(running-llm-d-without-kubernetes): drop broken well-lit-path link

The docs/well-lit-paths/no-kubernetes-deployment.md target does not exist
upstream (404), failing the CI link check. There is no no-kubernetes well-lit
path; the deploy-ready reference is the guide, already linked in the same
sentence - matching how other posts cite these guides.

Signed-off-by: Ezra Silvera <ezra@il.ibm.com>

* blog(running-llm-d-without-kubernetes): cite the no-Kubernetes infra doc

Restore the second upstream reference (the one the broken link intended) with
the correct path: docs/infrastructure/no-kubernetes-deployment.md, the
architecture/design overview - distinct from the step-by-step guide. It is filed
under docs/infrastructure upstream, not docs/well-lit-paths, so cite it as the
architecture doc rather than a well-lit path.

Signed-off-by: Ezra Silvera <ezra@il.ibm.com>

* blog(running-llm-d-without-kubernetes): replace "seam" metaphor with "interface"

Signed-off-by: Ezra Silvera <ezra@il.ibm.com>

* blog(running-llm-d-without-kubernetes): reframe benchmarking motivation around simplicity, cite InferenceX

Signed-off-by: Ezra Silvera <ezra@il.ibm.com>

* blog(running-llm-d-without-kubernetes): temper RL Kubernetes framing, note dual-managed worker nodes

Signed-off-by: Ezra Silvera <ezra@il.ibm.com>

* blog(running-llm-d-without-kubernetes): acknowledge Kubernetes as the default for production inference at scale

Signed-off-by: Ezra Silvera <ezra@il.ibm.com>

---------

Signed-off-by: Ezra Silvera <ezra@il.ibm.com>
Signed-off-by: Pete Cheslock <pete.cheslock@redhat.com>
Co-authored-by: Pete Cheslock <pete.cheslock@redhat.com>

Author: ezrasilvera

blog/2026-05-26_running-llm-d-without-kubernetes.md

@@ -422,3 +422,12 @@ rachelbrill:
   image_url: /img/blogs/rachelbrill.webp
   socials:
     linkedin: rachel-tzoref-brill-6645a11
+
+ezrasilvera:
+  name: Ezra Silvera
+  title: Senior Technical Staff Member, IBM
+  image_url: /img/blogs/ezrasilvera.webp
+  email: ezra@il.ibm.com
+  socials:
+    github: https://github.com/ezrasilvera
+    linkedin: ezra-silvera-9746682