diff --git a/config/manifests/bbr-example/httproute_bbr_lora.yaml b/config/manifests/bbr-example/httproute_bbr_lora.yaml new file mode 100644 index 000000000..8fc341a3f --- /dev/null +++ b/config/manifests/bbr-example/httproute_bbr_lora.yaml @@ -0,0 +1,71 @@ +apiVersion: gateway.networking.k8s.io/v1 +kind: HTTPRoute +metadata: + name: llm-llama-route +spec: + parentRefs: + - group: gateway.networking.k8s.io + kind: Gateway + name: inference-gateway + rules: + - backendRefs: + - group: inference.networking.k8s.io + kind: InferencePool + name: vllm-llama3-8b-instruct + matches: + - path: + type: PathPrefix + value: / + headers: + - type: Exact + name: X-Gateway-Model-Name + value: 'meta-llama/Llama-3.1-8B-Instruct' + - path: + type: PathPrefix + value: / + headers: + - type: Exact + name: X-Gateway-Model-Name + value: 'food-review-1' + timeouts: + request: 300s +--- +apiVersion: gateway.networking.k8s.io/v1 +kind: HTTPRoute +metadata: + name: llm-deepseek-route #give this HTTPRoute any name that helps you to group and track the matchers +spec: + parentRefs: + - group: gateway.networking.k8s.io + kind: Gateway + name: inference-gateway + rules: + - backendRefs: + - group: inference.networking.k8s.io + kind: InferencePool + name: vllm-deepseek-r1 + matches: + - path: + type: PathPrefix + value: / + headers: + - type: Exact + name: X-Gateway-Model-Name + value: 'deepseek/vllm-deepseek-r1' + - path: + type: PathPrefix + value: / + headers: + - type: Exact + name: X-Gateway-Model-Name + value: 'ski-resorts' + - path: + type: PathPrefix + value: / + headers: + - type: Exact + name: X-Gateway-Model-Name + value: 'movie-critique' + timeouts: + request: 300s +--- \ No newline at end of file diff --git a/config/manifests/vllm/sim-deployment-1.yaml b/config/manifests/vllm/sim-deployment-1.yaml new file mode 100644 index 000000000..d9c032a70 --- /dev/null +++ b/config/manifests/vllm/sim-deployment-1.yaml @@ -0,0 +1,44 @@ +apiVersion: apps/v1 +kind: Deployment +metadata: + name: vllm-deepseek-r1 +spec: + replicas: 1 + selector: + matchLabels: + app: vllm-deepseek-r1 + template: + metadata: + labels: + app: vllm-deepseek-r1 + spec: + containers: + - name: vllm-sim + image: ghcr.io/llm-d/llm-d-inference-sim:v0.4.0 + imagePullPolicy: Always + args: + - --model + - deepseek/vllm-deepseek-r1 + - --port + - "8000" + - --max-loras + - "2" + - --lora-modules + - '{"name": "ski-resorts"}' + - '{"name": "movie-critique"}' + env: + - name: POD_NAME + valueFrom: + fieldRef: + fieldPath: metadata.name + - name: NAMESPACE + valueFrom: + fieldRef: + fieldPath: metadata.namespace + ports: + - containerPort: 8000 + name: http + protocol: TCP + resources: + requests: + cpu: 10m diff --git a/site-src/guides/serve-multiple-genai-models.md b/site-src/guides/serve-multiple-genai-models.md index c94ff1a77..e3e5e4c08 100644 --- a/site-src/guides/serve-multiple-genai-models.md +++ b/site-src/guides/serve-multiple-genai-models.md @@ -1,44 +1,42 @@ -# Serve multiple generative AI models +# Serve multiple generative AI models and multiple LoRAs for the base AI models -A company wants to deploy multiple large language models (LLMs) to a cluster to serve different workloads. -For example, they might want to deploy a Gemma3 model for a chatbot interface and a DeepSeek model for a recommendation application (or as in the example in this guide, a combination of a Llama3 model and a smaller Phi4 model).. You may choose to locate these 2 models at 2 different L7 url paths and follow the steps described in the [`Getting started`](index.md) guide for each such model as already described. However you may also need to serve multiple models located at the same L7 url path and rely on parsing information such as -the Model name in the LLM prompt requests as defined in the OpenAI API format which is commonly used by most models. For such Model-aware routing, you can use the Body-Based Routing feature as described in this guide. +A company may need to deploy multiple large language models (LLMs) in a cluster to support different workloads. For example, a Llama model could power a chatbot interface, while a DeepSeek model might serve a recommendation application. One approach is to expose these models on separate Layer 7 (L7) URL paths and follow the steps in the [`Getting Started (Latest/Main)`](getting-started-latest.md) guide for each model. + +However, one may also need to serve multiple models from the same L7 URL path. To achieve this, the system needs to extract information (such as the model name) from the request body (i.e., the LLM prompt). This pattern of serving multiple models behind a single endpoint is common among providers and is generally expected by clients. The OpenAI API format requires the model name to be specified in the request body. For such model-aware routing, use the Body-Based Routing (BBR) feature described in this guide. + +Additionally, each base AI model can have multiple Low-Rank Adaptations ([LoRAs](https://www.ibm.com/think/topics/lora)). LoRAs associated with the same base model are served by the same backend inference server that hosts the base model. A LoRA name is also provided as the model name in the request body. ## How -The following diagram illustrates how an Inference Gateway routes requests to different models based on the model name. -The model name is extracted by [Body-Based routing](https://github.com/kubernetes-sigs/gateway-api-inference-extension/blob/main/pkg/bbr/README.md) (BBR) - from the request body to the header. The header is then matched to dispatch - requests to different `InferencePool` (and their EPPs) instances. +[Body-Based Router (BBR)](https://github.com/kubernetes-sigs/gateway-api-inference-extension/blob/main/pkg/bbr/README.md) extracts the model name from the request body and adds it to the `X-Gateway-Model-Name` header. This header is then used for matching and routing the request to the appropriate `InferencePool` and its associated Endpoint Picker Extension (EPP) instances. ### Example Model-Aware Routing using Body-Based Routing (BBR) -This guide assumes you have already setup the cluster for basic model serving as described in the [`Getting started`](index.md) guide and this guide describes the additional steps needed from that point onwards in order to deploy and exercise an example of routing across multiple models. - +This guide assumes you have already setup the cluster for basic model serving as described in the [`Getting started (Latest/Main)`](getting-started-latest.md) guide. In what follows, this guide describes the additional steps required to deploy and test routing across multiple models and multiple LoRAs, where several LoRAs may be associated with a single base model. ### Deploy Body-Based Routing Extension -To enable body-based routing, you need to deploy the Body-Based Routing ExtProc server using Helm. This is a separate ExtProc server from the EndPoint Picker and when installed, is automatically inserted at the start of the gateway's ExtProc chain ahead of other EtxProc servers such as EPP. +To enable body-based routing, deploy the BBR `ext_proc` server using Helm. This server is independent of EPP. Once installed, it is automatically added as the first filter in the gateway’s filter chain, ahead of other `ext_proc` servers such as EPP. -First install this server. Depending on your Gateway provider, you can use one of the following commands: +Select an appropriate tab depending on your Gateway provider: === "GKE" - ```bash - helm install body-based-router \ - --set provider.name=gke \ - --version v1.0.0 \ - oci://registry.k8s.io/gateway-api-inference-extension/charts/body-based-routing - ``` + ```bash + helm install body-based-router \ + --set provider.name=gke \ + --version v0 \ + oci://us-central1-docker.pkg.dev/k8s-staging-images/gateway-api-inference-extension/charts/body-based-routing + ``` === "Istio" - ```bash - helm install body-based-router \ - --set provider.name=istio \ - --version v1.0.0 \ - oci://registry.k8s.io/gateway-api-inference-extension/charts/body-based-routing - ``` + ```bash + helm install body-based-router \ + --set provider.name=istio \ + --version v0 \ + oci://us-central1-docker.pkg.dev/k8s-staging-images/gateway-api-inference-extension/charts/body-based-routing + ``` === "Kgateway" @@ -66,65 +64,167 @@ First install this server. Depending on your Gateway provider, you can use one o === "Other" - ```bash - helm install body-based-router \ - --version v1.0.0 \ - oci://registry.k8s.io/gateway-api-inference-extension/charts/body-based-routing - ``` + ```bash + helm install body-based-router \ + --version v0 \ + oci://us-central1-docker.pkg.dev/k8s-staging-images/gateway-api-inference-extension/charts/body-based-routing + ``` -Once this is installed, verify that the BBR pod is running without errors using the command `kubectl get pods`. +After the installation, verify that the BBR pod is running without errors: + +```bash +kubectl get pods +``` ### Serving a Second Base Model -Next deploy the second base model that will be served from the same L7 path (which is `/`) as the `meta-llama/Llama-3.1-8B-Instruct` model already being served after following the steps from the [`Getting started`](index.md) guide. In this example the 2nd model is `microsoft/Phi-4-mini-instruct` a relatively small model ( about 3B parameters) from HuggingFace. Note that for this exercise, there need to be atleast 2 GPUs available on the system one each for the two models being served. Serve the second model via the following command. + +The example uses a vLLM simulator since this is the least common denominator configuration that can be run in every environment. The model, `deepseek/vllm-deepseek-r1`, will be served from the same `/` L7 path, as in the previous example from the [Getting Started (Latest/Main)](getting-started-latest.md) guide. + +Deploy the second base model: ```bash -kubectl apply -f https://raw.githubusercontent.com/kubernetes-sigs/gateway-api-inference-extension/refs/heads/main/config/manifests/bbr-example/vllm-phi4-mini.yaml +kubectl apply -f https://github.com/kubernetes-sigs/gateway-api-inference-extension/raw/main/config/manifests/vllm/sim-deployment-1.yaml +``` + +The overall setup is as follows. Two base models are deployed: `meta-llama/Llama-3.1-8B-Instruct` and `deepseek/vllm-deepseek-r1`. Additionally, the `food-review-1` LoRA is associated with `meta-llama/Llama-3.1-8B-Instruct`, while the `ski-resorts` and `movie-critique` LoRAs are associated with `deepseek/vllm-deepseek-r1`. + +⚠️ **Note**: LoRA names must be unique across the base AI models (i.e., across the backend inference server deployments) + +Review the YAML definition. + +```yaml + apiVersion: apps/v1 + kind: Deployment + metadata: + name: vllm-deepseek-r1 + spec: + replicas: 1 + selector: + matchLabels: + app: vllm-deepseek-r1 + template: + metadata: + labels: + app: vllm-deepseek-r1 + spec: + containers: + - name: vllm-sim + image: ghcr.io/llm-d/llm-d-inference-sim:v0.4.0 + imagePullPolicy: Always + args: + - --model + - deepseek/vllm-deepseek-r1 + - --port + - "8000" + - --max-loras + - "2" + - --lora-modules + - '{"name": "ski-resorts"}' + - '{"name": "movie-critique"}' + env: + - name: POD_NAME + valueFrom: + fieldRef: + fieldPath: metadata.name + - name: NAMESPACE + valueFrom: + fieldRef: + fieldPath: metadata.namespace + ports: + - containerPort: 8000 + name: http + protocol: TCP + resources: + requests: + scpu: 10m +``` + +Verify that the second base model pod is running without errors: + +```bash +kubectl get pods ``` -Once this is installed, and after allowing for model download and startup time which can last several minutes, verify that the pod with this 2nd LLM phi4-mini, is running without errors using the command `kubectl get pods`. ### Deploy the 2nd InferencePool and Endpoint Picker Extension -We also want to use an InferencePool and EndPoint Picker for this second model in addition to the Body Based Router in order to be able to schedule across multiple endpoints or LORA adapters within each base model. Hence we create these for our second model as follows. + +Set the Helm chart version (unless already set). + + ```bash + export IGW_CHART_VERSION=v0 + ``` + +Select a tab to follow the provider-specific instructions. === "GKE" - ```bash - export GATEWAY_PROVIDER=gke - helm install vllm-phi4-mini-instruct \ - --set inferencePool.modelServers.matchLabels.app=phi4-mini \ - --set provider.name=$GATEWAY_PROVIDER \ - --version v1.0.0 \ - oci://registry.k8s.io/gateway-api-inference-extension/charts/inferencepool - ``` + ```bash + export GATEWAY_PROVIDER=gke + helm install vllm-deepseek-r1 \ + --set inferencePool.modelServers.matchLabels.app=vllm-deepseek-r1 \ + --set provider.name=$GATEWAY_PROVIDER \ + --version $IGW_CHART_VERSION \ + oci://us-central1-docker.pkg.dev/k8s-staging-images/gateway-api-inference-extension/charts/inferencepool + ``` === "Istio" - ```bash - export GATEWAY_PROVIDER=istio - helm install vllm-phi4-mini-instruct \ - --set inferencePool.modelServers.matchLabels.app=phi4-mini \ - --set provider.name=$GATEWAY_PROVIDER \ - --version v1.0.0 \ - oci://registry.k8s.io/gateway-api-inference-extension/charts/inferencepool - ``` + ```bash + export GATEWAY_PROVIDER=istio + helm install vllm-deepseek-r1 \ + --set inferencePool.modelServers.matchLabels.app=vllm-deepseek-r1 \ + --set provider.name=$GATEWAY_PROVIDER \ + --version $IGW_CHART_VERSION \ + oci://us-central1-docker.pkg.dev/k8s-staging-images/gateway-api-inference-extension/charts/inferencepool + ``` +=== "Kgateway" + + ```bash + export GATEWAY_PROVIDER=none + helm install vllm-deepseek-r1 \ + --set inferencePool.modelServers.matchLabels.app=vllm-deepseek-r1 \ + --set provider.name=$GATEWAY_PROVIDER \ + --version $IGW_CHART_VERSION \ + oci://us-central1-docker.pkg.dev/k8s-staging-images/gateway-api-inference-extension/charts/inferencepool + ``` + +=== "NGINX Gateway Fabric" + + ```bash + export GATEWAY_PROVIDER=none + helm install vllm-deepseek-r1 \ + --set inferencePool.modelServers.matchLabels.app=vllm-deepseek-r1 \ + --set provider.name=$GATEWAY_PROVIDER \ + --version $IGW_CHART_VERSION \ + oci://us-central1-docker.pkg.dev/k8s-staging-images/gateway-api-inference-extension/charts/inferencepool + ``` -After executing this, verify that you see two InferencePools and two EPP pods, one per base model type, running without errors, using the CLIs `kubectl get inferencepools` and `kubectl get pods`. +After the installation, verify that you have two `InferencePools` and two EPP pods, one per base model type, running without errors -### Configure HTTPRoute +```bash +kubectl get inferencepools +``` + +```bash +kubectl get pods +``` + +### Configure HTTPRoutes -Before configuring the httproutes for the models, we need to delete the prior httproute created for the vllm-llama3-8b-instruct model because we will alter the routing to now also match on the model name as determined by the `X-Gateway-Model-Name` http header that will get inserted by the BBR extension after parsing the Model name from the body of the LLM request message. +Before configuring the HTTPRoutes for the models and their LoRAs, delete the existing HTTPRoute for the `meta-llama/Llama-3.1-8B-Instruct` model. The new routes will match the model name in the `X-Gateway-Model-Name` HTTP header, which is inserted by the BBR extension after parsing the model name from the LLM request body. ```bash kubectl delete httproute llm-route ``` -Now configure new HTTPRoutes, one per each model we want to serve via BBR using the following command which configures both routes. Also examine this manifest file, to see how the `X-Gateway-Model-Name` is used for a header match in the Gateway's rules to route requests to the correct Backend based on model name. For convenience the manifest is also listed below in order to view this routing configuration. +Now configure new HTTPRoutes for the two simulated models and their LoRAs that we want to serve via BBR using the following command which configures both routes. ```bash -kubectl apply -f https://github.com/kubernetes-sigs/gateway-api-inference-extension/raw/main/config/manifests/bbr-example/httproute_bbr.yaml +kubectl apply -f https://github.com/kubernetes-sigs/gateway-api-inference-extension/raw/main/config/manifests/bbr-example/httproute_bbr_lora.yaml ``` +Also examine the manifest file (see the yaml below), to see how the `X-Gateway-Model-Name` is used for a header match in the Gateway's rules to route requests to the correct Backend based on the model name. + ```yaml ---- apiVersion: gateway.networking.k8s.io/v1 kind: HTTPRoute metadata: @@ -145,16 +245,22 @@ spec: value: / headers: - type: Exact - #Body-Based routing(https://github.com/kubernetes-sigs/gateway-api-inference-extension/blob/main/pkg/bbr/README.md) is being used to copy the model name from the request body to the header. - name: X-Gateway-Model-Name # (1)! + name: X-Gateway-Model-Name value: 'meta-llama/Llama-3.1-8B-Instruct' + - path: + type: PathPrefix + value: / + headers: + - type: Exact + name: X-Gateway-Model-Name + value: 'food-review-1' timeouts: request: 300s --- apiVersion: gateway.networking.k8s.io/v1 kind: HTTPRoute metadata: - name: llm-phi4-route + name: llm-deepseek-route #give this HTTPRoute any name that helps you to group and track the matchers spec: parentRefs: - group: gateway.networking.k8s.io @@ -164,105 +270,197 @@ spec: - backendRefs: - group: inference.networking.k8s.io kind: InferencePool - name: vllm-phi4-mini-instruct + name: vllm-deepseek-r1 matches: - path: type: PathPrefix value: / headers: - type: Exact - #Body-Based routing(https://github.com/kubernetes-sigs/gateway-api-inference-extension/blob/main/pkg/bbr/README.md) is being used to copy the model name from the request body to the header. name: X-Gateway-Model-Name - value: 'microsoft/Phi-4-mini-instruct' + value: 'deepseek/vllm-deepseek-r1' + - path: + type: PathPrefix + value: / + headers: + - type: Exact + name: X-Gateway-Model-Name + value: 'ski-resorts' + - path: + type: PathPrefix + value: / + headers: + - type: Exact + name: X-Gateway-Model-Name + value: 'movie-critique' timeouts: request: 300s ---- +--- ``` -Before testing the setup, confirm that the HTTPRoute status conditions include `Accepted=True` and `ResolvedRefs=True` for both routes using the following commands. +⚠️ **Note** : +[Kubernetes API Gateway limits the total number of matchers per HTTPRoute to be less than 128](https://github.com/kubernetes-sigs/gateway-api/blob/df8c96c254e1ac6d5f5e0d70617f36143723d479/apis/v1/httproute_types.go#L128). + +Before testing the setup, confirm that the HTTPRoute status conditions include `Accepted=True` and `ResolvedRefs=True` for both routes using the following commands. ```bash kubectl get httproute llm-llama-route -o yaml ``` ```bash -kubectl get httproute llm-phi4-route -o yaml +kubectl get httproute llm-deepseek-route -o yaml ``` -## Try it out - -1. Get the gateway IP: - ```bash - IP=$(kubectl get gateway/inference-gateway -o jsonpath='{.status.addresses[0].value}'); PORT=80 - ``` +### Try the setup === "Chat Completions API" - 1. Send a few requests to Llama model as follows: - ```bash - curl -X POST -i ${IP}:${PORT}/v1/chat/completions \ - -H "Content-Type: application/json" \ - -d '{ - "model": "meta-llama/Llama-3.1-8B-Instruct", - "max_tokens": 100, - "temperature": 0, - "messages": [ - { - "role": "developer", - "content": "You are a helpful assistant." - }, - { - "role": "user", - "content": "Linux is said to be an open source kernel because " - } - ] - }' - ``` - - 2. Send a few requests to the Phi4 as follows: - ```bash - curl -X POST -i ${IP}:${PORT}/v1/chat/completions \ - -H "Content-Type: application/json" \ - -d '{ - "model": "microsoft/Phi-4-mini-instruct", - "max_tokens": 100, - "temperature": 0, - "messages": [ - { - "role": "developer", - "content": "You are a helpful assistant." - }, - { - "role": "user", - "content": "2+2 is " - } - ] - }' - ``` + 1. Send a few requests to Llama model to test that it works as before, as follows: + + ```bash + curl -X POST -i ${IP}:${PORT}/v1/chat/completions \ + -H "Content-Type: application/json" \ + -d '{ + "model": "meta-llama/Llama-3.1-8B-Instruct", + "max_tokens": 100, + "temperature": 0, + "messages": [ + { + "role": "developer", + "content": "You are a helpful assistant." + }, + { + "role": "user", + "content": "Linux is said to be an open source kernel because " + } + ] + }' + ``` + + 1. Send a few requests to Deepseek model to test that it works, as follows: + + ```bash + curl -X POST -i ${IP}:${PORT}/v1/chat/completions \ + -H "Content-Type: application/json" \ + -d '{ + "model": "deepseek/vllm-deepseek-r1", + "max_tokens": 100, + "temperature": 0, + "messages": [ + { + "role": "developer", + "content": "You are a helpful assistant." + }, + { + "role": "user", + "content": "Linux is said to be an open source kernel because " + } + ] + }' + ``` + 1. Send a few requests to the LoRA of the Llama model as follows: + + ```bash + curl -X POST -i ${IP}:${PORT}/v1/chat/completions \ + -H "Content-Type: application/json" \ + -d '{ + "model": "food-review-1", + "max_tokens": 100, + "temperature": 0, + "messages": [ + { + "role": "reviewer", + "content": "You are a helpful assistant." + }, + { + "role": "user", + "content": "Write a review of the best restaurans in San-Francisco" + } + ] + }' + ``` + + 1. Send a few requests to one LoRA of the Deepseek model as follows: + + ```bash + curl -X POST -i ${IP}:${PORT}/v1/chat/completions \ + -H "Content-Type: application/json" \ + -d '{ + "model": "movie-critique", + "max_tokens": 100, + "temperature": 0, + "messages": [ + { + "role": "reviewer", + "content": "You are a helpful assistant." + }, + { + "role": "user", + "content": "What are the best movies of 2025?" + } + ] + }' + ``` + + 1. Send a few requests to another LoRA of the Deepseek model as follows: + + ```bash + curl -X POST -i ${IP}:${PORT}/v1/chat/completions \ + -H "Content-Type: application/json" \ + -d '{ + "model": "ski-resorts", + "max_tokens": 100, + "temperature": 0, + "messages": [ + { + "role": "reviewer", + "content": "You are a helpful assistant." + }, + { + "role": "user", + "content": "Tell mne about ski deals" + } + ] + }' + ``` === "Completions API" - 1. Send a few requests to Llama model as follows: - ```bash - curl -X POST -i ${IP}:${PORT}/v1/completions \ - -H "Content-Type: application/json" \ - -d '{ - "model": "meta-llama/Llama-3.1-8B-Instruct", - "prompt": "Linux is said to be an open source kernel because ", - "max_tokens": 100, - "temperature": 0 - }' - ``` - - 2. Send a few requests to the Phi4 as follows: - ```bash - curl -X POST -i ${IP}:${PORT}/v1/completions \ - -H "Content-Type: application/json" \ - -d '{ - "model": "microsoft/Phi-4-mini-instruct", - "prompt": "2+2 is ", - "max_tokens": 20, - "temperature": 0 - }' - ``` - + 1. Send a few requests to Llama model's LoRA as follows: + + ```bash + curl -X POST -i ${IP}:${PORT}/v1/completions \ + -H "Content-Type: application/json" \ + -d '{ + "model": "food-review-1", + "prompt": "Write as if you were a critic: San Francisco ", + "max_tokens": 100, + "temperature": 0 + }' + ``` + + 1. Send a few requests to the first Deepseek LoRA as follows: + + ```bash + curl -X POST -i ${IP}:${PORT}/v1/completions \ + -H "Content-Type: application/json" \ + -d '{ + "model": "ski-resorts", + "prompt": "What is the best ski resort in Austria?", + "max_tokens": 20, + "temperature": 0 + }' + ``` + + 1. Send a few requests to the second Deepseek LoRA as follows: + + ```bash + curl -X POST -i ${IP}:${PORT}/v1/completions \ + -H "Content-Type: application/json" \ + -d '{ + "model": "movie-critique", + "prompt": "Tell me about movies", + "max_tokens": 20, + "temperature": 0 + }' + ```