Include no sample example

quickstart/examples/no-sample-app/README.md

@@ -0,0 +1,122 @@
+# No Sample App Preset
+
+This preset is for when you want to provision the infrastructure around modelservice but not actually create a `modelservice` custom resource.
+This can be particularly helpful because of how helm works - it wants to make sure that it owns and deploys all resources in its chart via labels.
+The result of this, is that if you create a sample app through the helm charts, and you want to modify or delete your `modelservice` custom resource,
+you have to uninstall the whole stack, tweak an example values override file before re-installing through quickstart.
+
+This preset enables you to have more control over that. You can apply the "No Sample App" preset, use `helm template` to template manifests based on
+a different preset, extract the resourcess that make up the diff ([see below](./README.md#making-up-the-difference)), and then apply them by hand.
+The result should be you can make **some** modifications to your setup without having to use the quickstart `--uninstall` tack in the installer script,
+and install everything from the ground up again.
+
+It should be noted that you can only make **some** modifications because
+[the baseConfigMap](https://github.com/llm-d/llm-d-deployer/tree/dc90005ef2fc87f1e6dbac282dfc038187efde12/charts/llm-d/templates/modelservice/presets)
+presets for `modelservice` are immutable, any desired changes there will require uninstalling and reinstalling via the quickstart install script.
+
+## Deploying Infrastructure
+
+Assuming one is starting from the quickstart directory, a sample deploy might look like:
+
+```bash
+HF_TOKEN=$(HFTOKEN) ./llmd-installer.sh --namespace llm-d --values-file examples/no-sample-app/no-sample-app.yaml
+```
+
+## Validating the deployment
+
+You should expect to see only see infra related resources that can be plugged into any setup, specifically:
+
+- A gateway, which should spin a gateway deployment / pod.
+- A modelservice controller.
+- A redis master deployment. This is turned on by default in this example in case the desired modelservice should use the `baseConfigMapRef` of `base-gpu-with-nixl-and-redis-lookup-preset`. If you do not want to use this preset, you can optionally add the following to your `no-sample-app.yaml`:
+
+```yaml
+sampleApplication:
+  enabled: false
+redis
+  enabled: false
+```
+
+To sumarrize, in a successful deployment of this preset, expect to see:
+
+```log
+NAME                                       READY   STATUS    RESTARTS   AGE
+llm-d-inference-gateway-654ddbcb74-25vv4   1/1     Running   0          58m
+llm-d-modelservice-7844f89454-lcs2r        1/1     Running   0          58m
+llm-d-redis-master-8599696799-2wpv5        1/1     Running   0          58m
+```
+
+## Deploying the `ModelService`
+
+To see all resource templates that get created based on a "Sample App" which we will intentionally create manually in this example, you can
+[look here](https://github.com/llm-d/llm-d-deployer/tree/main/charts/llm-d/templates/sample-application).
+
+These templates will map into the following manifests that we will create later based on this `no-sample-app` vs another sample app scenario:
+
+1. A `modelservice` custom resource.
+    - This is the biggest component of a "Sample App" and will create many other children resources, such as prefill, decode, and `endpoint-picker` deployments, services, serviceAccounts, RBAC, an `inferencepool`, an `inferencemodel`, etc.
+2. An `httpRoute` that will be accepted by the gateway, with a backendRef of the `inferencepool` that will get created by the `modelservice` CR.
+3. (ISTIO ONLY) A `destinationRule`. We have support for Istio as a gateway control plane implementation but for now we are using kgateway as a default, so I don't believe this has ever been tested.
+4. A `ClusterRoleBinding` for the `endpoint-picker` that allow for scraping of `endpoint-picker` (epp) metrics with authentication.
+    - This is included in "sample app" because per "sample app" there is an `endpoint-picker` that gets deployed, with a different `serviceAccount`, named based off the `modelName` in the Sample App.
+
+## Making Up the Difference
+
+Lets walkthrough how one might use this "No Sample App" preset to provisoin the base infrastructure, and then create a sample app by hand to re-create one of the other presets.
+
+First lets pick a "Sample App" to re-create - in this case lets start with "no-features". We are going to use `helm template` with two values files, first
+the base values file, and then passing the "no features" preset as an override values file, and then dumping that to a temporary yaml file.
+
+```bash
+# navigate to the chart directory
+cd ../../../charts/llm-d/
+
+# set your namespace in your no-sample-app infra was already deployed
+: "${NAMESPACE:=llm-d}"
+
+TEMP=$(mktemp)
+
+helm template llm-d ./ -n ${NAMESPACE} --values ../../quickstart/examples/no-features/no-features.yaml > $TEMP
+```
+
+For those unfamiliar with `helm` or `helm template`, this above command will tell helm to spit out what manifests a `helm upgrade -i` or `helm install`
+would plan to create based on the merge of the two values files. It will treat the second values file with higher priority in this merge, and overwrite anything
+that appears in both with the value of the 2nd values file (in this case, our `no-features` preset we are trying to re-create).
+
+As mentioned above, that `$TEMP` file should include all the manifests from the `no-features` example, but if you remember we already applied our `no-sample-app` preset, so many of these manifests will already exist in our cluster.
+
+We want to grab just the `modelservice` CR, the `httpRoute`, and `clusterRoleBindigng` for EPP metrics. You can do this by hand, but based on the `no-features` preset were re-creating, you can automate it with the following (this yq):
+
+```bash
+yq -i eval-all '
+  select(
+    (.kind == "ModelService")
+    or (.kind == "HTTPRoute")
+    or (.kind == "ClusterRoleBinding" and .metadata.name == "llm-d-llm-d-modelservice-epp-metrics-scrape")
+  )
+  | del(.metadata.labels["helm.sh/chart"])
+  | del(.metadata.labels["app.kubernetes.io/managed-by"])
+' $TEMP
+
+kubectl apply -f $TEMP
+```
+
+> [!NOTE]
+> The `ClusterRoleBinding` name is based on the `modelservice` controller name, so it can vary.
+> It is suggested to copy the manifests by hand, but based on matching up the `CRB` name with the `no-feature` example, this example can be scripted.
+> This example also assumes you are using the default `kgateway` control plane and APIs rather than Istio.
+> This command depends on the `mikefarah` version of `yq` v4 or greater
+
+In addition to grabbing those resources, the above script will filter out the helm labels that would otherwise get templated into the manifests,
+allowing us to create and manage these resources by hand.
+
+Now our pods should reflect that of the `no-feature` example with our new `epp` and single `decode` pods:
+
+```log
+NAME                                                       READY   STATUS    RESTARTS   AGE
+llm-d-inference-gateway-654ddbcb74-25vv4                   1/1     Running   0          97m
+llm-d-modelservice-7844f89454-lcs2r                        1/1     Running   0          97m
+llm-d-redis-master-8599696799-2wpv5                        1/1     Running   0          97m
+meta-llama-llama-3-2-3b-instruct-decode-7c679bb8b4-kbxc6   2/2     Running   0          42s
+meta-llama-llama-3-2-3b-instruct-epp-8d4749d47-dnpf2       1/1     Running   0          42s
+```

quickstart/examples/no-sample-app/no-sample-app.yaml

@@ -0,0 +1,2 @@
+sampleApplication:
+  enabled: false