From d08bf5844a878c92c0153fb761f308e08258f45a Mon Sep 17 00:00:00 2001
From: sirius0xdev <sirius0xdev@gmail.com>
Date: Wed, 29 Apr 2026 04:12:36 +0000
Subject: [PATCH 1/2] feat: optimize RTX6000 vLLM KEDA scaling

- Rename deployment to rtx6000-brain-vllm to match scaler target ref
- Set deployment replicas to 0 (KEDA controlled)
- Add pollingInterval: 10s and cooldownPeriod: 30s for faster response

This enables scale-to-zero when idle and quick scale-up on first request.

See https://github.com/sirius0xdev/gcloud-lab/tree/master/infrastructure%2Fgpus%2Fbase
---
 infrastructure/gpus/base/keda-gpu-scaling/keda-vllm.yaml | 2 +-
 infrastructure/gpus/base/vllm-servers/rtx6000-vllm.yaml  | 4 ++--
 2 files changed, 3 insertions(+), 3 deletions(-)

diff --git a/infrastructure/gpus/base/keda-gpu-scaling/keda-vllm.yaml b/infrastructure/gpus/base/keda-gpu-scaling/keda-vllm.yaml
index c675bdd..eee9d2e 100644
--- a/infrastructure/gpus/base/keda-gpu-scaling/keda-vllm.yaml
+++ b/infrastructure/gpus/base/keda-gpu-scaling/keda-vllm.yaml
@@ -14,4 +14,4 @@ spec:
     min: 0
     max: 1
   targetPendingRequests: 1
-  scaledownPeriod: 900
+  scaledownPeriod: 900\n  pollingInterval: 10\n  cooldownPeriod: 30
diff --git a/infrastructure/gpus/base/vllm-servers/rtx6000-vllm.yaml b/infrastructure/gpus/base/vllm-servers/rtx6000-vllm.yaml
index 2e60c16..d5ccbbc 100644
--- a/infrastructure/gpus/base/vllm-servers/rtx6000-vllm.yaml
+++ b/infrastructure/gpus/base/vllm-servers/rtx6000-vllm.yaml
@@ -2,12 +2,12 @@
 apiVersion: apps/v1
 kind: Deployment
 metadata:
-  name: openclaw-brain-vllm
+  name: rtx6000-brain-vllm
   namespace: customer1
   labels:
     app: rtx6000-brain
 spec:
-  replicas: 1
+  replicas: 0
   selector:
     matchLabels:
       app: rtx6000-brain

From 7e7b7d414df117ac1bd7c580973d7cb501161264 Mon Sep 17 00:00:00 2001
From: sirius0xdev <sirius0xdev@gmail.com>
Date: Wed, 29 Apr 2026 04:14:46 +0000
Subject: [PATCH 2/2] fix: restrict KEDA scaling to /v1/(chat/)?completions
 paths only
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

- Switch to HTTPScaledObject (matches a100 pattern)
- Probes (/health, /v1/models) now ignored for scaling
- Keeps fast polling (10s), quick cooldown (30s), 15min idle scale-to-0

Health/readiness/startup probes hit pod IP directly (bypass service), so never triggered scaling anyway — but paths ensure only inference traffic scales.
---
 .../gpus/base/keda-gpu-scaling/keda-vllm.yaml   | 17 +++++++++++------
 1 file changed, 11 insertions(+), 6 deletions(-)

diff --git a/infrastructure/gpus/base/keda-gpu-scaling/keda-vllm.yaml b/infrastructure/gpus/base/keda-gpu-scaling/keda-vllm.yaml
index eee9d2e..1669e1c 100644
--- a/infrastructure/gpus/base/keda-gpu-scaling/keda-vllm.yaml
+++ b/infrastructure/gpus/base/keda-gpu-scaling/keda-vllm.yaml
@@ -1,17 +1,22 @@
-apiVersion: keda.sh/v1alpha1
-kind: ScaledObject
+apiVersion: http.keda.sh/v1alpha1
+kind: HTTPScaledObject
 metadata:
   name: rtx6000-scaling
   namespace: customer1
 spec:
+  hosts:
+  - rtx6000-brain-service.customer1.svc.cluster.local
   scaleTargetRef:
     name: rtx6000-brain-vllm
     kind: Deployment
     apiVersion: apps/v1
-  hosts:
-    - rtx6000-brain-service.customer1.svc.cluster.local
+    service: rtx6000-brain-service
+    port: 8000
+  paths:
+  - path: "^/v1/(chat/)?completions"
   replicas:
     min: 0
     max: 1
-  targetPendingRequests: 1
-  scaledownPeriod: 900\n  pollingInterval: 10\n  cooldownPeriod: 30
+  pollingInterval: 10
+  cooldownPeriod: 30
+  scaledownPeriod: 900