Merge pull request opendatahub-io#379 from kserve/master

[pull] master from kserve:master

charts/kserve-resources/README.md

@@ -72,7 +72,7 @@ $ helm install kserve oci://ghcr.io/kserve/charts/kserve --version v0.13.0
 | kserve.servingruntime.lgbserver.tag | string | `"v0.13.0"` |  |
 | kserve.servingruntime.mlserver.image | string | `"docker.io/seldonio/mlserver"` |  |
 | kserve.servingruntime.mlserver.modelClassPlaceholder | string | `"{{.Labels.modelClass}}"` |  |
-| kserve.servingruntime.mlserver.tag | string | `"1.3.2"` |  |
+| kserve.servingruntime.mlserver.tag | string | `"1.5.0"` |  |
 | kserve.servingruntime.modelNamePlaceholder | string | `"{{.Name}}"` |  |
 | kserve.servingruntime.paddleserver.image | string | `"kserve/paddleserver"` |  |
 | kserve.servingruntime.paddleserver.tag | string | `"v0.13.0"` |  |

charts/kserve-resources/templates/clusterservingruntimes.yaml

@@ -53,14 +53,26 @@ spec:
       version: "1"
       autoSelect: true
       priority: 2
+    - name: xgboost
+      version: "2"
+      autoSelect: true
+      priority: 2
     - name: lightgbm
       version: "3"
       autoSelect: true
       priority: 2
+    - name: lightgbm
+      version: "4"
+      autoSelect: true
+      priority: 2
     - name: mlflow
       version: "1"
       autoSelect: true
       priority: 1
+    - name: mlflow
+      version: "2"
+      autoSelect: true
+      priority: 1
   protocolVersions:
     - v2
   containers:

charts/kserve-resources/values.yaml

@@ -95,7 +95,7 @@ kserve:
       tag: 2.6.2
     mlserver:
       image: docker.io/seldonio/mlserver
-      tag: 1.3.2
+      tag: 1.5.0
       modelClassPlaceholder: "{{.Labels.modelClass}}"
     sklearnserver:
       image: kserve/sklearnserver

config/runtimes/kserve-mlserver.yaml

@@ -20,14 +20,26 @@ spec:
       version: "1"
       autoSelect: true
       priority: 2
+    - name: xgboost
+      version: "2"
+      autoSelect: true
+      priority: 2
     - name: lightgbm
       version: "3"
       autoSelect: true
       priority: 2
+    - name: lightgbm
+      version: "4"
+      autoSelect: true
+      priority: 2
     - name: mlflow
       version: "1"
       autoSelect: true
       priority: 1
+    - name: mlflow
+      version: "2"
+      autoSelect: true
+      priority: 1
   protocolVersions:
     - v2
   containers:

config/runtimes/kustomization.yaml

@@ -25,7 +25,7 @@ images:
 
 - name: mlserver
   newName: docker.io/seldonio/mlserver
-  newTag: 1.3.2
+  newTag: 1.5.0
 
 - name: kserve-xgbserver
   newName: kserve/xgbserver

pkg/controller/v1beta1/inferenceservice/controller.go

@@ -168,7 +168,7 @@ func (r *InferenceServiceReconciler) Reconcile(ctx context.Context, req ctrl.Req
 	// Abort early if the resolved deployment mode is Serverless, but Knative Services are not available
 	if deploymentMode == constants.Serverless {
 		ksvcAvailable, checkKsvcErr := utils.IsCrdAvailable(r.ClientConfig, knservingv1.SchemeGroupVersion.String(), constants.KnativeServiceKind)
-		if err != nil {
+		if checkKsvcErr != nil {
 			return reconcile.Result{}, checkKsvcErr
 		}
 

python/huggingface_server.Dockerfile

@@ -1,4 +1,4 @@
-ARG BASE_IMAGE=nvidia/cuda:12.1.0-devel-ubuntu22.04
+ARG BASE_IMAGE=nvidia/cuda:12.4.1-devel-ubuntu22.04
 ARG VENV_PATH=/prod_venv
 
 FROM ${BASE_IMAGE} as builder
@@ -9,7 +9,7 @@ ARG POETRY_HOME=/opt/poetry
 ARG POETRY_VERSION=1.7.1
 
 # Install vllm
-ARG VLLM_VERSION=0.4.2
+ARG VLLM_VERSION=0.4.3
 
 RUN apt-get update -y && apt-get install gcc python3.10-venv python3-dev -y && apt-get clean && \
     rm -rf /var/lib/apt/lists/*
@@ -34,7 +34,7 @@ RUN cd huggingfaceserver && poetry install --no-interaction --no-cache
 
 RUN pip3 install vllm==${VLLM_VERSION}
 
-FROM nvidia/cuda:12.1.0-base-ubuntu22.04 as prod
+FROM nvidia/cuda:12.4.1-runtime-ubuntu22.04 as prod
 
 RUN apt-get update -y && apt-get install python3.10-venv -y && apt-get clean && \
     rm -rf /var/lib/apt/lists/*
@@ -58,6 +58,8 @@ ENV HF_HOME="/tmp/huggingface"
 ENV SAFETENSORS_FAST_GPU="1"
 # https://huggingface.co/docs/huggingface_hub/en/package_reference/environment_variables#hfhubdisabletelemetry
 ENV HF_HUB_DISABLE_TELEMETRY="1"
+# NCCL Lib path for vLLM. https://github.com/vllm-project/vllm/blob/ec784b2526219cd96159a52074ab8cd4e684410a/vllm/utils.py#L598-L602
+ENV VLLM_NCCL_SO_PATH="/lib/x86_64-linux-gnu/libnccl.so.2"
 
 USER 1000
 ENTRYPOINT ["python3", "-m", "huggingfaceserver"]

python/huggingfaceserver/huggingfaceserver/vllm/vllm_completions.py

@@ -144,10 +144,9 @@ async def create_completion(self, completion_request: CompletionRequest):
 
                 generators.append(
                     self.engine.generate(
-                        prompt,
+                        {"prompt": prompt, "prompt_token_ids": input_ids},
                         sampling_params,
                         f"{request_id}-{i}",
-                        prompt_token_ids=input_ids,
                     )
                 )
         except Exception as e:
@@ -175,7 +174,7 @@ async def create_completion(self, completion_request: CompletionRequest):
             )
 
         # Non-streaming response
-        final_res_batch: RequestOutput = [None] * len(prompts)
+        final_res_batch: List[RequestOutput] = [None] * len(prompts)
         try:
             async for i, res in result_generator:
                 final_res_batch[i] = res

python/huggingfaceserver/pyproject.toml

@@ -15,7 +15,7 @@ kserve = { path = "../kserve", extras = ["storage"], develop = true }
 transformers = "~4.40.2"
 accelerate = "~0.30.0"
 torch = "~2.3.0"
-vllm = { version = "^0.4.2", optional = true }
+vllm = { version = "^0.4.3", optional = true }
 
 [tool.poetry.extras]
 vllm = [

python/kserve/kserve/protocol/rest/openai/openai_proxy_model.py

@@ -243,8 +243,8 @@ async def create_completion(
         self, request: CompletionRequest
     ) -> Union[Completion, AsyncIterator[Completion]]:
         self.preprocess_completion_request(request)
-        req = self._build_request(self._completions_endpoint, request)
         if request.params.stream:
+            req = self._build_request(self._completions_endpoint, request)
             r = await self._http_client.send(req, stream=True)
             r.raise_for_status()
             it = AsyncMappingIterator(
@@ -254,23 +254,28 @@ async def create_completion(
             )
             return it
         else:
-            response = await self._http_client.send(req)
-            response.raise_for_status()
-            if self.skip_upstream_validation:
-                obj = response.json()
-                completion = Completion.model_construct(**obj)
-            else:
-                completion = Completion.model_validate_json(response.content)
+            completion = await self.generate_completion(request)
             self.postprocess_completion(completion, request)
             return completion
 
+    async def generate_completion(self, request: CompletionRequest) -> Completion:
+        req = self._build_request(self._completions_endpoint, request)
+        response = await self._http_client.send(req)
+        response.raise_for_status()
+        if self.skip_upstream_validation:
+            obj = response.json()
+            completion = Completion.model_construct(**obj)
+        else:
+            completion = Completion.model_validate_json(response.content)
+        return completion
+
     @error_handler
     async def create_chat_completion(
         self, request: ChatCompletionRequest
     ) -> Union[ChatCompletion, AsyncIterator[ChatCompletionChunk]]:
         self.preprocess_chat_completion_request(request)
-        req = self._build_request(self._chat_completions_endpoint, request)
         if request.params.stream:
+            req = self._build_request(self._chat_completions_endpoint, request)
             r = await self._http_client.send(req, stream=True)
             r.raise_for_status()
             it = AsyncMappingIterator(
@@ -280,12 +285,19 @@ async def create_chat_completion(
             )
             return it
         else:
-            response = await self._http_client.send(req)
-            response.raise_for_status()
-            if self.skip_upstream_validation:
-                obj = response.json()
-                chat_completion = ChatCompletion.model_construct(**obj)
-            else:
-                chat_completion = ChatCompletion.model_validate_json(response.content)
+            chat_completion = await self.generate_chat_completion(request)
             self.postprocess_chat_completion(chat_completion, request)
             return chat_completion
+
+    async def generate_chat_completion(
+        self, request: ChatCompletionRequest
+    ) -> ChatCompletion:
+        req = self._build_request(self._chat_completions_endpoint, request)
+        response = await self._http_client.send(req)
+        response.raise_for_status()
+        if self.skip_upstream_validation:
+            obj = response.json()
+            chat_completion = ChatCompletion.model_construct(**obj)
+        else:
+            chat_completion = ChatCompletion.model_validate_json(response.content)
+        return chat_completion