// Copyright 2025 Google LLC
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
//     http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.

syntax = "proto3";

package google.cloud.gkerecommender.v1;

import "google/api/annotations.proto";
import "google/api/client.proto";
import "google/api/field_behavior.proto";

option csharp_namespace = "Google.Cloud.GkeRecommender.V1";
option go_package = "cloud.google.com/go/gkerecommender/apiv1/gkerecommenderpb;gkerecommenderpb";
option java_multiple_files = true;
option java_outer_classname = "GkeRecommenderProto";
option java_package = "com.google.cloud.gkerecommender.v1";
option php_namespace = "Google\\Cloud\\GkeRecommender\\V1";
option ruby_package = "Google::Cloud::GkeRecommender::V1";

// GKE Inference Quickstart (GIQ) service provides profiles with performance
// metrics for popular models and model servers across multiple accelerators.
// These profiles help generate optimized best practices for running inference
// on GKE.
service GkeInferenceQuickstart {
  option (google.api.default_host) = "gkerecommender.googleapis.com";
  option (google.api.oauth_scopes) =
      "https://www.googleapis.com/auth/cloud-platform";

  // Fetches available models. Open-source models follow the Huggingface Hub
  // `owner/model_name` format.
  rpc FetchModels(FetchModelsRequest) returns (FetchModelsResponse) {
    option (google.api.http) = {
      get: "/v1/models:fetch"
    };
  }

  // Fetches available model servers. Open-source model servers use simplified,
  // lowercase names (e.g., `vllm`).
  rpc FetchModelServers(FetchModelServersRequest)
      returns (FetchModelServersResponse) {
    option (google.api.http) = {
      get: "/v1/modelServers:fetch"
    };
  }

  // Fetches available model server versions. Open-source servers use their own
  // versioning schemas (e.g., `vllm` uses semver like `v1.0.0`).
  //
  // Some model servers have different versioning schemas depending on the
  // accelerator. For example, `vllm` uses semver on GPUs, but returns nightly
  // build tags on TPUs. All available versions will be returned when different
  // schemas are present.
  rpc FetchModelServerVersions(FetchModelServerVersionsRequest)
      returns (FetchModelServerVersionsResponse) {
    option (google.api.http) = {
      get: "/v1/modelServerVersions:fetch"
    };
  }

  // Fetches available profiles. A profile contains performance metrics and
  // cost information for a specific model server setup. Profiles can be
  // filtered by parameters. If no filters are provided, all profiles are
  // returned.
  //
  // Profiles display a single value per performance metric based on the
  // provided performance requirements. If no requirements are given, the
  // metrics represent the inflection point. See [Run best practice inference
  // with GKE Inference Quickstart
  // recipes](https://cloud.google.com/kubernetes-engine/docs/how-to/machine-learning/inference/inference-quickstart#how)
  // for details.
  rpc FetchProfiles(FetchProfilesRequest) returns (FetchProfilesResponse) {
    option (google.api.http) = {
      post: "/v1/profiles:fetch"
      body: "*"
    };
  }

  // Generates an optimized deployment manifest for a given model and model
  // server, based on the specified accelerator, performance targets, and
  // configurations. See [Run best practice inference with GKE Inference
  // Quickstart
  // recipes](https://cloud.google.com/kubernetes-engine/docs/how-to/machine-learning/inference/inference-quickstart)
  // for deployment details.
  rpc GenerateOptimizedManifest(GenerateOptimizedManifestRequest)
      returns (GenerateOptimizedManifestResponse) {
    option (google.api.http) = {
      post: "/v1/optimizedManifest:generate"
      body: "*"
    };
  }

  // Fetches all of the benchmarking data available for a profile. Benchmarking
  // data returns all of the performance metrics available for a given model
  // server setup on a given instance type.
  rpc FetchBenchmarkingData(FetchBenchmarkingDataRequest)
      returns (FetchBenchmarkingDataResponse) {
    option (google.api.http) = {
      post: "/v1/benchmarkingData:fetch"
      body: "*"
    };
  }
}

// Request message for
// [GkeInferenceQuickstart.FetchModels][google.cloud.gkerecommender.v1.GkeInferenceQuickstart.FetchModels].
message FetchModelsRequest {
  // Optional. The target number of results to return in a single response.
  // If not specified, a default value will be chosen by the service.
  // Note that the response may include a partial list and a caller should
  // only rely on the response's
  // [next_page_token][google.cloud.gkerecommender.v1.FetchModelsResponse.next_page_token]
  // to determine if there are more instances left to be queried.
  optional int32 page_size = 1 [(google.api.field_behavior) = OPTIONAL];

  // Optional. The value of
  // [next_page_token][google.cloud.gkerecommender.v1.FetchModelsResponse.next_page_token]
  // received from a previous `FetchModelsRequest` call.
  // Provide this to retrieve the subsequent page in a multi-page list of
  // results. When paginating, all other parameters provided to
  // `FetchModelsRequest` must match the call that provided the page token.
  optional string page_token = 2 [(google.api.field_behavior) = OPTIONAL];
}

// Response message for
// [GkeInferenceQuickstart.FetchModels][google.cloud.gkerecommender.v1.GkeInferenceQuickstart.FetchModels].
message FetchModelsResponse {
  // Output only. List of available models. Open-source models follow the
  // Huggingface Hub `owner/model_name` format.
  repeated string models = 1 [(google.api.field_behavior) = OUTPUT_ONLY];

  // Output only. A token which may be sent as
  // [page_token][FetchModelsResponse.page_token] in a subsequent
  // `FetchModelsResponse` call to retrieve the next page of results.
  // If this field is omitted or empty, then there are no more results to
  // return.
  string next_page_token = 2 [(google.api.field_behavior) = OUTPUT_ONLY];
}

// Request message for
// [GkeInferenceQuickstart.FetchModelServers][google.cloud.gkerecommender.v1.GkeInferenceQuickstart.FetchModelServers].
message FetchModelServersRequest {
  // Required. The model for which to list model servers. Open-source models
  // follow the Huggingface Hub `owner/model_name` format. Use
  // [GkeInferenceQuickstart.FetchModels][google.cloud.gkerecommender.v1.GkeInferenceQuickstart.FetchModels]
  // to find available models.
  string model = 1 [(google.api.field_behavior) = REQUIRED];

  // Optional. The target number of results to return in a single response.
  // If not specified, a default value will be chosen by the service.
  // Note that the response may include a partial list and a caller should
  // only rely on the response's
  // [next_page_token][google.cloud.gkerecommender.v1.FetchModelServersResponse.next_page_token]
  // to determine if there are more instances left to be queried.
  optional int32 page_size = 2 [(google.api.field_behavior) = OPTIONAL];

  // Optional. The value of
  // [next_page_token][google.cloud.gkerecommender.v1.FetchModelServersResponse.next_page_token]
  // received from a previous `FetchModelServersRequest` call.
  // Provide this to retrieve the subsequent page in a multi-page list of
  // results. When paginating, all other parameters provided to
  // `FetchModelServersRequest` must match the call that provided the page
  // token.
  optional string page_token = 3 [(google.api.field_behavior) = OPTIONAL];
}

// Response message for
// [GkeInferenceQuickstart.FetchModelServers][google.cloud.gkerecommender.v1.GkeInferenceQuickstart.FetchModelServers].
message FetchModelServersResponse {
  // Output only. List of available model servers. Open-source model servers use
  // simplified, lowercase names (e.g., `vllm`).
  repeated string model_servers = 1 [(google.api.field_behavior) = OUTPUT_ONLY];

  // Output only. A token which may be sent as
  // [page_token][FetchModelServersResponse.page_token] in a subsequent
  // `FetchModelServersResponse` call to retrieve the next page of results.
  // If this field is omitted or empty, then there are no more results to
  // return.
  string next_page_token = 2 [(google.api.field_behavior) = OUTPUT_ONLY];
}

// Request message for
// [GkeInferenceQuickstart.FetchModelServerVersions][google.cloud.gkerecommender.v1.GkeInferenceQuickstart.FetchModelServerVersions].
message FetchModelServerVersionsRequest {
  // Required. The model for which to list model server versions. Open-source
  // models follow the Huggingface Hub `owner/model_name` format. Use
  // [GkeInferenceQuickstart.FetchModels][google.cloud.gkerecommender.v1.GkeInferenceQuickstart.FetchModels]
  // to find available models.
  string model = 1 [(google.api.field_behavior) = REQUIRED];

  // Required. The model server for which to list versions. Open-source model
  // servers use simplified, lowercase names (e.g., `vllm`). Use
  // [GkeInferenceQuickstart.FetchModelServers][google.cloud.gkerecommender.v1.GkeInferenceQuickstart.FetchModelServers]
  // to find available model servers.
  string model_server = 2 [(google.api.field_behavior) = REQUIRED];

  // Optional. The target number of results to return in a single response.
  // If not specified, a default value will be chosen by the service.
  // Note that the response may include a partial list and a caller should
  // only rely on the response's
  // [next_page_token][google.cloud.gkerecommender.v1.FetchModelServerVersionsResponse.next_page_token]
  // to determine if there are more instances left to be queried.
  optional int32 page_size = 3 [(google.api.field_behavior) = OPTIONAL];

  // Optional. The value of
  // [next_page_token][google.cloud.gkerecommender.v1.FetchModelServerVersionsResponse.next_page_token]
  // received from a previous `FetchModelServerVersionsRequest` call.
  // Provide this to retrieve the subsequent page in a multi-page list of
  // results. When paginating, all other parameters provided to
  // `FetchModelServerVersionsRequest` must match the call that provided the
  // page token.
  optional string page_token = 4 [(google.api.field_behavior) = OPTIONAL];
}

// Response message for
// [GkeInferenceQuickstart.FetchModelServerVersions][google.cloud.gkerecommender.v1.GkeInferenceQuickstart.FetchModelServerVersions].
message FetchModelServerVersionsResponse {
  // Output only. A list of available model server versions.
  repeated string model_server_versions = 1
      [(google.api.field_behavior) = OUTPUT_ONLY];

  // Output only. A token which may be sent as
  // [page_token][FetchModelServerVersionsResponse.page_token] in a subsequent
  // `FetchModelServerVersionsResponse` call to retrieve the next page of
  // results. If this field is omitted or empty, then there are no more results
  // to return.
  string next_page_token = 2 [(google.api.field_behavior) = OUTPUT_ONLY];
}

// Request message for
// [GkeInferenceQuickstart.FetchBenchmarkingData][google.cloud.gkerecommender.v1.GkeInferenceQuickstart.FetchBenchmarkingData].
message FetchBenchmarkingDataRequest {
  // Required. The model server configuration to get benchmarking data for. Use
  // [GkeInferenceQuickstart.FetchProfiles][google.cloud.gkerecommender.v1.GkeInferenceQuickstart.FetchProfiles]
  // to find valid configurations.
  ModelServerInfo model_server_info = 1
      [(google.api.field_behavior) = REQUIRED];

  // Optional. The instance type to filter benchmarking data. Instance types are
  // in the format `a2-highgpu-1g`. If not provided, all instance types for the
  // given profile's `model_server_info` will be returned. Use
  // [GkeInferenceQuickstart.FetchProfiles][google.cloud.gkerecommender.v1.GkeInferenceQuickstart.FetchProfiles]
  // to find available instance types.
  string instance_type = 3 [(google.api.field_behavior) = OPTIONAL];

  // Optional. The pricing model to use for the benchmarking data. Defaults to
  // `spot`.
  string pricing_model = 4 [(google.api.field_behavior) = OPTIONAL];
}

// Response message for
// [GkeInferenceQuickstart.FetchBenchmarkingData][google.cloud.gkerecommender.v1.GkeInferenceQuickstart.FetchBenchmarkingData].
message FetchBenchmarkingDataResponse {
  // Output only. List of profiles containing their respective benchmarking
  // data.
  repeated Profile profile = 1 [(google.api.field_behavior) = OUTPUT_ONLY];
}

// Request message for
// [GkeInferenceQuickstart.FetchProfiles][google.cloud.gkerecommender.v1.GkeInferenceQuickstart.FetchProfiles].
message FetchProfilesRequest {
  // Optional. The model to filter profiles by. Open-source models follow the
  // Huggingface Hub `owner/model_name` format. If not provided, all models are
  // returned. Use
  // [GkeInferenceQuickstart.FetchModels][google.cloud.gkerecommender.v1.GkeInferenceQuickstart.FetchModels]
  // to find available models.
  string model = 1 [(google.api.field_behavior) = OPTIONAL];

  // Optional. The model server to filter profiles by. If not provided, all
  // model servers are returned. Use
  // [GkeInferenceQuickstart.FetchModelServers][google.cloud.gkerecommender.v1.GkeInferenceQuickstart.FetchModelServers]
  // to find available model servers for a given model.
  string model_server = 2 [(google.api.field_behavior) = OPTIONAL];

  // Optional. The model server version to filter profiles by. If not provided,
  // all model server versions are returned. Use
  // [GkeInferenceQuickstart.FetchModelServerVersions][google.cloud.gkerecommender.v1.GkeInferenceQuickstart.FetchModelServerVersions]
  // to find available versions for a given model and server.
  string model_server_version = 3 [(google.api.field_behavior) = OPTIONAL];

  // Optional. The performance requirements to filter profiles. Profiles that do
  // not meet these requirements are filtered out. If not provided, all profiles
  // are returned.
  PerformanceRequirements performance_requirements = 4
      [(google.api.field_behavior) = OPTIONAL];

  // Optional. The target number of results to return in a single response. If
  // not specified, a default value will be chosen by the service. Note that the
  // response may include a partial list and a caller should only rely on the
  // response's
  // [next_page_token][google.cloud.gkerecommender.v1.FetchProfilesResponse.next_page_token]
  // to determine if there are more instances left to be queried.
  optional int32 page_size = 5 [(google.api.field_behavior) = OPTIONAL];

  // Optional. The value of
  // [next_page_token][google.cloud.gkerecommender.v1.FetchProfilesResponse.next_page_token]
  // received from a previous `FetchProfilesRequest` call.
  // Provide this to retrieve the subsequent page in a multi-page list of
  // results. When paginating, all other parameters provided to
  // `FetchProfilesRequest` must match the call that provided the page
  // token.
  optional string page_token = 6 [(google.api.field_behavior) = OPTIONAL];
}

// Performance requirements for a profile and or model deployment.
message PerformanceRequirements {
  // Optional. The target Normalized Time Per Output Token (NTPOT) in
  // milliseconds. NTPOT is calculated as `request_latency /
  // total_output_tokens`. If not provided, this target will not be enforced.
  optional int32 target_ntpot_milliseconds = 1
      [(google.api.field_behavior) = OPTIONAL];

  // Optional. The target Time To First Token (TTFT) in milliseconds. TTFT is
  // the time it takes to generate the first token for a request.  If not
  // provided, this target will not be enforced.
  optional int32 target_ttft_milliseconds = 2
      [(google.api.field_behavior) = OPTIONAL];

  // Optional. The target cost for running a profile's model server. If not
  // provided, this requirement will not be enforced.
  Cost target_cost = 3 [(google.api.field_behavior) = OPTIONAL];
}

// Represents an amount of money in a specific currency.
message Amount {
  // Output only. The whole units of the amount.
  // For example if `currencyCode` is `"USD"`, then 1 unit is one US dollar.
  int64 units = 1 [(google.api.field_behavior) = OUTPUT_ONLY];

  // Output only. Number of nano (10^-9) units of the amount.
  // The value must be between -999,999,999 and +999,999,999 inclusive.
  // If `units` is positive, `nanos` must be positive or zero.
  // If `units` is zero, `nanos` can be positive, zero, or negative.
  // If `units` is negative, `nanos` must be negative or zero.
  // For example $-1.75 is represented as `units`=-1 and `nanos`=-750,000,000.
  int32 nanos = 2 [(google.api.field_behavior) = OUTPUT_ONLY];
}

// Cost for running a model deployment on a given instance type. Currently, only
// USD currency code is supported.
message Cost {
  // Optional. The cost per million output tokens, calculated as:
  // $/output token = GPU $/s / (1/output-to-input-cost-ratio * input tokens/s +
  // output tokens/s)
  Amount cost_per_million_output_tokens = 1
      [(google.api.field_behavior) = OPTIONAL];

  // Optional. The cost per million input tokens. $/input token = ($/output
  // token) / output-to-input-cost-ratio.
  Amount cost_per_million_input_tokens = 2
      [(google.api.field_behavior) = OPTIONAL];

  // Optional. The pricing model used to calculate the cost. Can be one of:
  // `3-years-cud`, `1-year-cud`, `on-demand`, `spot`. If not provided, `spot`
  // will be used.
  string pricing_model = 3 [(google.api.field_behavior) = OPTIONAL];

  // Optional. The output-to-input cost ratio. This determines how the total GPU
  // cost is split between input and output tokens. If not provided, `4.0` is
  // used, assuming a 4:1 output:input cost ratio.
  optional float output_input_cost_ratio = 4
      [(google.api.field_behavior) = OPTIONAL];
}

// Represents a range of throughput values in tokens per second.
message TokensPerSecondRange {
  // Output only. The minimum value of the range.
  int32 min = 1 [(google.api.field_behavior) = OUTPUT_ONLY];

  // Output only. The maximum value of the range.
  int32 max = 2 [(google.api.field_behavior) = OUTPUT_ONLY];
}

// Represents a range of latency values in milliseconds.
message MillisecondRange {
  // Output only. The minimum value of the range.
  int32 min = 1 [(google.api.field_behavior) = OUTPUT_ONLY];

  // Output only. The maximum value of the range.
  int32 max = 2 [(google.api.field_behavior) = OUTPUT_ONLY];
}

// Performance range for a model deployment.
message PerformanceRange {
  // Output only. The range of throughput in output tokens per second. This is
  // measured as total_output_tokens_generated_by_server /
  // elapsed_time_in_seconds.
  TokensPerSecondRange throughput_output_range = 1
      [(google.api.field_behavior) = OUTPUT_ONLY];

  // Output only. The range of TTFT (Time To First Token) in milliseconds. TTFT
  // is the time it takes to generate the first token for a request.
  MillisecondRange ttft_range = 2 [(google.api.field_behavior) = OUTPUT_ONLY];

  // Output only. The range of NTPOT (Normalized Time Per Output Token) in
  // milliseconds. NTPOT is the request latency normalized by the number of
  // output tokens, measured as request_latency / total_output_tokens.
  MillisecondRange ntpot_range = 3 [(google.api.field_behavior) = OUTPUT_ONLY];
}

// Response message for
// [GkeInferenceQuickstart.FetchProfiles][google.cloud.gkerecommender.v1.GkeInferenceQuickstart.FetchProfiles].
message FetchProfilesResponse {
  // Output only. List of profiles that match the given model server info and
  // performance requirements (if provided).
  repeated Profile profile = 1 [(google.api.field_behavior) = OUTPUT_ONLY];

  // Output only. The combined range of performance values observed across all
  // profiles in this response.
  PerformanceRange performance_range = 2
      [(google.api.field_behavior) = OUTPUT_ONLY];

  // Output only. Additional comments related to the response.
  string comments = 3 [(google.api.field_behavior) = OUTPUT_ONLY];

  // Output only. A token which may be sent as
  // [page_token][FetchProfilesResponse.page_token] in a subsequent
  // `FetchProfilesResponse` call to retrieve the next page of results. If this
  // field is omitted or empty, then there are no more results to return.
  string next_page_token = 4 [(google.api.field_behavior) = OUTPUT_ONLY];
}

// Model server information gives. Valid model server info combinations can
// be found using
// [GkeInferenceQuickstart.FetchProfiles][google.cloud.gkerecommender.v1.GkeInferenceQuickstart.FetchProfiles].
message ModelServerInfo {
  // Required. The model. Open-source models follow the Huggingface Hub
  // `owner/model_name` format. Use
  // [GkeInferenceQuickstart.FetchModels][google.cloud.gkerecommender.v1.GkeInferenceQuickstart.FetchModels]
  // to find available models.
  string model = 1 [(google.api.field_behavior) = REQUIRED];

  // Required. The model server. Open-source model servers use simplified,
  // lowercase names (e.g., `vllm`). Use
  // [GkeInferenceQuickstart.FetchModelServers][google.cloud.gkerecommender.v1.GkeInferenceQuickstart.FetchModelServers]
  // to find available servers.
  string model_server = 2 [(google.api.field_behavior) = REQUIRED];

  // Optional. The model server version. Use
  // [GkeInferenceQuickstart.FetchModelServerVersions][google.cloud.gkerecommender.v1.GkeInferenceQuickstart.FetchModelServerVersions]
  // to find available versions. If not provided, the latest available version
  // is used.
  string model_server_version = 3 [(google.api.field_behavior) = OPTIONAL];
}

// Resources used by a model deployment.
message ResourcesUsed {
  // Output only. The number of accelerators (e.g., GPUs or TPUs) used by the
  // model deployment on the Kubernetes node.
  int32 accelerator_count = 1 [(google.api.field_behavior) = OUTPUT_ONLY];
}

// Performance statistics for a model deployment.
message PerformanceStats {
  // Output only. The number of queries per second.
  // Note: This metric can vary widely based on context length and may not be a
  // reliable measure of LLM throughput.
  float queries_per_second = 1 [(google.api.field_behavior) = OUTPUT_ONLY];

  // Output only. The number of output tokens per second. This is the throughput
  // measured as total_output_tokens_generated_by_server /
  // elapsed_time_in_seconds.
  int32 output_tokens_per_second = 2
      [(google.api.field_behavior) = OUTPUT_ONLY];

  // Output only. The Normalized Time Per Output Token (NTPOT) in milliseconds.
  // This is the request latency normalized by the number of output tokens,
  // measured as request_latency / total_output_tokens.
  int32 ntpot_milliseconds = 3 [(google.api.field_behavior) = OUTPUT_ONLY];

  // Output only. The Time To First Token (TTFT) in milliseconds. This is the
  // time it takes to generate the first token for a request.
  int32 ttft_milliseconds = 4 [(google.api.field_behavior) = OUTPUT_ONLY];

  // Output only. The cost of running the model deployment.
  repeated Cost cost = 5 [(google.api.field_behavior) = OUTPUT_ONLY];
}

// A profile containing information about a model deployment.
message Profile {
  // Output only. The model server configuration. Use
  // [GkeInferenceQuickstart.FetchProfiles][google.cloud.gkerecommender.v1.GkeInferenceQuickstart.FetchProfiles]
  // to find valid configurations.
  ModelServerInfo model_server_info = 1
      [(google.api.field_behavior) = OUTPUT_ONLY];

  // Output only. The accelerator type. Expected format: `nvidia-h100-80gb`.
  string accelerator_type = 2 [(google.api.field_behavior) = OUTPUT_ONLY];

  // Output only. The TPU topology (if applicable).
  string tpu_topology = 3 [(google.api.field_behavior) = OUTPUT_ONLY];

  // Output only. The instance type. Expected format: `a2-highgpu-1g`.
  string instance_type = 4 [(google.api.field_behavior) = OUTPUT_ONLY];

  // Output only. The resources used by the model deployment.
  ResourcesUsed resources_used = 5 [(google.api.field_behavior) = OUTPUT_ONLY];

  // Output only. The performance statistics for this profile.
  repeated PerformanceStats performance_stats = 6
      [(google.api.field_behavior) = OUTPUT_ONLY];
}

// Request message for
// [GkeInferenceQuickstart.GenerateOptimizedManifest][google.cloud.gkerecommender.v1.GkeInferenceQuickstart.GenerateOptimizedManifest].
message GenerateOptimizedManifestRequest {
  // Required. The model server configuration to generate the manifest for. Use
  // [GkeInferenceQuickstart.FetchProfiles][google.cloud.gkerecommender.v1.GkeInferenceQuickstart.FetchProfiles]
  // to find valid configurations.
  ModelServerInfo model_server_info = 1
      [(google.api.field_behavior) = REQUIRED];

  // Required. The accelerator type. Use
  // [GkeInferenceQuickstart.FetchProfiles][google.cloud.gkerecommender.v1.GkeInferenceQuickstart.FetchProfiles]
  // to find valid accelerators for a given `model_server_info`.
  string accelerator_type = 2 [(google.api.field_behavior) = REQUIRED];

  // Optional. The kubernetes namespace to deploy the manifests in.
  string kubernetes_namespace = 3 [(google.api.field_behavior) = OPTIONAL];

  // Optional. The performance requirements to use for generating Horizontal Pod
  // Autoscaler (HPA) resources. If provided, the manifest includes HPA
  // resources to adjust the model server replica count to maintain the
  // specified targets (e.g., NTPOT, TTFT) at a P50 latency. Cost targets are
  // not currently supported for HPA generation. If the specified targets are
  // not achievable, the HPA manifest will not be generated.
  PerformanceRequirements performance_requirements = 4
      [(google.api.field_behavior) = OPTIONAL];

  // Optional. The storage configuration for the model. If not provided, the
  // model is loaded from Huggingface.
  StorageConfig storage_config = 5 [(google.api.field_behavior) = OPTIONAL];
}

// A Kubernetes manifest.
message KubernetesManifest {
  // Output only. Kubernetes resource kind.
  string kind = 1 [(google.api.field_behavior) = OUTPUT_ONLY];

  // Output only. Kubernetes API version.
  string api_version = 2 [(google.api.field_behavior) = OUTPUT_ONLY];

  // Output only. YAML content.
  string content = 3 [(google.api.field_behavior) = OUTPUT_ONLY];
}

// Response message for
// [GkeInferenceQuickstart.GenerateOptimizedManifest][google.cloud.gkerecommender.v1.GkeInferenceQuickstart.GenerateOptimizedManifest].
message GenerateOptimizedManifestResponse {
  // Output only. A list of generated Kubernetes manifests.
  repeated KubernetesManifest kubernetes_manifests = 1
      [(google.api.field_behavior) = OUTPUT_ONLY];

  // Output only. Comments related to deploying the generated manifests.
  repeated string comments = 2 [(google.api.field_behavior) = OUTPUT_ONLY];

  // Output only. Additional information about the versioned dependencies used
  // to generate the manifests. See [Run best practice inference with GKE
  // Inference Quickstart
  // recipes](https://cloud.google.com/kubernetes-engine/docs/how-to/machine-learning/inference/inference-quickstart)
  // for details.
  string manifest_version = 3 [(google.api.field_behavior) = OUTPUT_ONLY];
}

// Storage configuration for a model deployment.
message StorageConfig {
  // Optional. The Google Cloud Storage bucket URI to load the model from. This
  // URI must point to the directory containing the model's config file
  // (`config.json`) and model weights. A tuned GCSFuse setup can improve
  // LLM Pod startup time by more than 7x. Expected format:
  // `gs://<bucket-name>/<path-to-model>`.
  string model_bucket_uri = 1 [(google.api.field_behavior) = OPTIONAL];

  // Optional. The URI for the GCS bucket containing the XLA compilation cache.
  // If using TPUs, the XLA cache will be written to the same path as
  // `model_bucket_uri`. This can speed up vLLM model preparation for repeated
  // deployments.
  string xla_cache_bucket_uri = 2 [(google.api.field_behavior) = OPTIONAL];
}