Merge pull request #64 from stackhpc/feat/flux-image-gen

Flux image gen chart improvements

Author: sd109

charts/flux-image-gen/templates/ui/configmap.yaml

@@ -13,4 +13,4 @@ data:
         address: {{ printf "http://%s.%s.svc:%v" ( printf "%s-%s-api" (include "flux-image-gen.fullname" $) . ) $.Release.Namespace $.Values.api.service.port }}
     {{- end }}
     example_prompt: |
-      {{- .Values.examplePrompt | nindent 6 -}}
+      {{- .Values.examplePrompt | trimSuffix "\n" | nindent 6 -}}

charts/flux-image-gen/values.yaml

@@ -85,13 +85,13 @@ api:
   deploymentStrategy:
     type: Recreate
 
-  # Downloading 100GB+ of model weights can take a long time so
-  # it's difficult to give these probes sensible default values...
+  # Downloading model weights can take a long time so it's
+  # difficult to give these probes sensible default values
+  # Is 30 minutes long enough...?
   startupProbe:
     # httpGet:
     #   path: /
     #   port: http
-    # Is 30 minutes long enough...?
     # failureThreshold: 180
     # periodSeconds: 10
   livenessProbe:

web-apps/chat/requirements.txt

@@ -1,4 +1,4 @@
-gradio<5
+gradio==4.21 # v4.41 breaks custom JS query param parsing
 gradio_client
 openai
 langchain

web-apps/flux-image-gen/api_server.py

@@ -31,6 +31,9 @@ class ImageGenInput(BaseModel):
     prompt: str
     add_sampling_metadata: bool
 
+@app.get("/")
+def health_check():
+    return "Server is running"
 
 @app.get("/model")
 async def get_model():

web-apps/flux-image-gen/gradio_ui.py

@@ -18,7 +18,9 @@ class Model(BaseModel):
 
 class AppSettings(BaseModel):
     models: List[Model]
-    example_prompt: str
+    example_prompt: str = "Yoda riding a skateboard."
+    title: str = "Flux Image Generation Demo"
+
 
 
 settings_path = pathlib.Path("/etc/gradio-app/gradio_config.yaml")
@@ -93,9 +95,8 @@ async def generate_image(
 
         return image, seed, filename, None
 
-
-with gr.Blocks() as demo:
-    gr.Markdown("# Flux Image Generation Demo")
+with gr.Blocks(title=settings.title) as demo:
+    gr.Markdown(f"# {settings.title}")
 
     with gr.Row():
         with gr.Column():

web-apps/image-analysis/requirements.txt

@@ -1,6 +1,6 @@
 pillow
 requests
-gradio<5
+gradio==4.21 # v4.41 breaks custom JS query param parsing
 gradio_client
 pydantic
 structlog

web-apps/kind-images.sh

@@ -7,14 +7,32 @@ if [[ -z $1 ]]; then
     exit 1
 fi
 
+# Work around storage limits in GH runners
+if [[ $CI == "true" ]]; then
+    DIR=/mnt/gimme-more-space
+    sudo mkdir -p $DIR
+    sudo chown -R $USER:$USER $DIR
+    TAR_PATH=$DIR/image.tar
+else
+    TAR_PATH="./image.tar"
+fi
+
 REMOTE_TAG=$1
 CLUSTER_NAME=${2:-kind}
 echo Kind cluster name: $CLUSTER_NAME
 KIND_TAG=local
 for image in $(find_images .); do
     full_name=ghcr.io/stackhpc/azimuth-llm-$image-ui
-    echo $full_name
+    echo $full_name:{$REMOTE_TAG,$KIND_TAG}
     docker pull $full_name:$REMOTE_TAG
     docker image tag $full_name:$REMOTE_TAG $full_name:$KIND_TAG
-    kind load docker-image -n $CLUSTER_NAME $full_name:$KIND_TAG
+    # NOTE(scott): The 'load docker-image' command saves the
+    # intermediate tar archive to /tmp which has limited space
+    # inside a GH runner so do each step manually here instead.
+    # kind load docker-image -n $CLUSTER_NAME $full_name:$KIND_TAG
+    # Apparently there's a separate 75G disk at /mnt so try using it.
+    docker image save -o $TAR_PATH $full_name:$KIND_TAG
+    docker image rm $full_name:{$REMOTE_TAG,$KIND_TAG}
+    kind load image-archive -n $CLUSTER_NAME $TAR_PATH
+    rm $TAR_PATH
 done