bluejay/bluejay-infra
1
0
Fork 0
Code Issues Pull Requests 15 Actions Packages Projects Releases Wiki Activity

deploy(worldbuilder): enable live gpu backend

Browse Source
This commit is contained in:
Andrew Stoltz
2026-06-11 16:05:40 -05:00
parent 7103658342
commit c1a43c64b3
2 changed files with 29 additions and 33 deletions
Download Patch File Download Diff File Expand all files Collapse all files

40
apps/worldbuilder/README.md
View File

@@ -12,28 +12,27 @@ Source: `D:\git\FlowerCore\FlowerCore.WorldBuilder` (master)
in pfSense Unbound before this manifest is applied, or cert-manager
HTTP-01 silently exponential-backs-off ~2h.
Memory: `feedback_pfsense_dns_required_for_acme`.
2. **Image import to ALL RKE2 nodes** — pod can schedule to any of
`rke2-server` (10.0.56.11), `rke2-agent1` (10.0.56.12),
`rke2-agent2` (10.0.56.13). Build with:
2. **Image import to ALL Ready RKE2 nodes** — pod can currently schedule to
`rke2-server` (10.0.56.11) and `rke2-agent1` (10.0.56.12). Build with:
```bash
bash deploy/build.sh # in FlowerCore.WorldBuilder repo
podman save localhost/fc-worldbuilder:v<TAG> -o /tmp/fc-worldbuilder-v<TAG>.tar
for h in 10.0.56.11 10.0.56.12 10.0.56.13; do
scp /tmp/fc-worldbuilder-v<TAG>.tar fcadmin@$h:/tmp/
mkdir -p artifacts/deploy
podman save localhost/fc-worldbuilder:v<TAG> -o artifacts/deploy/fc-worldbuilder-v<TAG>.tar
for h in 10.0.56.11 10.0.56.12; do
ssh fcadmin@$h "mkdir -p /home/fcadmin/.fcv"
scp artifacts/deploy/fc-worldbuilder-v<TAG>.tar fcadmin@$h:/home/fcadmin/.fcv/
ssh fcadmin@$h \
"sudo /var/lib/rancher/rke2/bin/ctr -a /run/k3s/containerd/containerd.sock \
-n k8s.io images import /tmp/fc-worldbuilder-v<TAG>.tar"
-n k8s.io images import /home/fcadmin/.fcv/fc-worldbuilder-v<TAG>.tar"
done
```
Memory: `feedback_rke2_image_import_per_node_scp`.
3. **Bump image tag** in `worldbuilder.yaml` and git push.
ArgoCD ApplicationSet picks up within ~3 minutes.
4. **First production render** — open
`https://worldbuilder.iamworkin.lan/studio/c32e0000-0000-4000-8000-000000000004`
and confirm the Cyberpunk Blue Jay demo prompt loads with five seeded fake
generated images. This Sprint 32 visitor-safe profile uses
`ClientMode=fake`; switch the image-generation env vars back to ComfyUI only
for an operator-owned GPU render lane.
4. **First production render** — verify
`https://worldbuilder.iamworkin.lan/healthz`, open
`https://worldbuilder.iamworkin.lan/settings`, and confirm the image backend
reports ComfyUI before running an operator-owned render lane.
## Health probes
@@ -56,13 +55,8 @@ Source: `D:\git\FlowerCore\FlowerCore.WorldBuilder` (master)
## Image generation backend
Sprint 32 pins the Kubernetes profile to
`FlowerCore:WorldBuilder:ImageGeneration:ClientMode=fake` with
`BaseUrl=http://127.0.0.1:1`. That keeps the public/internal visitor demo
deterministic, avoids GPU exposure, and still exercises the studio/gallery
surface with persisted generated-image metadata.
The previous ComfyUI backend target was `http://10.0.56.20:8188` on
BLUEJAY-WS (R9700 / gfx1201 / ROCm 7.2.1). Re-enable it only in an
operator-owned follow-up that also verifies workstation reachability and image
import freshness.
The live internal profile now uses
`FlowerCore:WorldBuilder:ImageGeneration:ClientMode=comfyui` with
`BaseUrl=http://10.0.56.20:8188` on BLUEJAY-WS (R9700 / gfx1201 / ROCm 7.2).
Keep the public host pre-staging disabled unless the five safe-to-expose gates
are rechecked; the live GPU lane is operator-owned and internal-only.

22
apps/worldbuilder/worldbuilder.yaml
View File

@@ -5,10 +5,10 @@
#
# Image build (BLUEJAY-WS):
# bash deploy/build.sh # in FlowerCore.WorldBuilder repo
# podman save localhost/fc-worldbuilder:v<TAG> -o /tmp/fc-worldbuilder-v<TAG>.tar
# for h in 10.0.56.11 10.0.56.12 10.0.56.13; do
# scp /tmp/fc-worldbuilder-v<TAG>.tar fcadmin@$h:/tmp/
# ssh fcadmin@$h "sudo /var/lib/rancher/rke2/bin/ctr -a /run/k3s/containerd/containerd.sock -n k8s.io images import /tmp/fc-worldbuilder-v<TAG>.tar"
# podman save localhost/fc-worldbuilder:v<TAG> -o artifacts/deploy/fc-worldbuilder-v<TAG>.tar
# for h in 10.0.56.11 10.0.56.12; do
# scp artifacts/deploy/fc-worldbuilder-v<TAG>.tar fcadmin@$h:/home/fcadmin/.fcv/
# ssh fcadmin@$h "sudo /var/lib/rancher/rke2/bin/ctr -a /run/k3s/containerd/containerd.sock -n k8s.io images import /home/fcadmin/.fcv/fc-worldbuilder-v<TAG>.tar"
# done
---
apiVersion: v1
@@ -90,7 +90,7 @@ spec:
containers:
- name: web
# Bump tag for each rebuild. Initial deploy: v202605062048
image: localhost/fc-worldbuilder:v202605062048
image: localhost/fc-worldbuilder:v20260611-b4a0025-gpu
imagePullPolicy: Never
ports:
- containerPort: 8080
@@ -117,14 +117,16 @@ spec:
value: "/data/gallery"
- name: FlowerCore__WorldBuilder__Export__RootPath
value: "/data/exports"
# Visitor-safe Sprint 32 profile: fake backend keeps public demo
# rendering deterministic and avoids exposing BLUEJAY-WS GPU.
# Operator-approved live GPU lane. Internal-only host targets
# BLUEJAY-WS ComfyUI; keep public host pre-staging disabled below.
- name: FlowerCore__WorldBuilder__ImageGeneration__BaseUrl
value: "http://127.0.0.1:1"
value: "http://10.0.56.20:8188"
- name: FlowerCore__WorldBuilder__ImageGeneration__ClientMode
value: "fake"
value: "comfyui"
- name: FlowerCore__WorldBuilder__ImageGeneration__BackendId
value: "fake"
value: "comfyui"
- name: FlowerCore__WorldBuilder__ImageGeneration__VisitorSafe
value: "false"
resources:
# Cluster CPU-request budget runs hot (99% on all 3 nodes at deploy
# time) while actual CPU usage is well below capacity. Idle Blazor