@meetkai/mka1 0.51.2 → 0.51.3

package/src/models/components/deploymentcreate.ts

@@ -13,6 +13,17 @@ import {
   AcceleratorSpec$Outbound,
   AcceleratorSpec$outboundSchema,
 } from "./acceleratorspec.js";
+import {
+  DeploymentAPIFormat,
+  DeploymentAPIFormat$inboundSchema,
+  DeploymentAPIFormat$outboundSchema,
+} from "./deploymentapiformat.js";
+import {
+  DeploymentScalingCreate,
+  DeploymentScalingCreate$inboundSchema,
+  DeploymentScalingCreate$Outbound,
+  DeploymentScalingCreate$outboundSchema,
+} from "./deploymentscalingcreate.js";
 import {
   DeployStrategy,
   DeployStrategy$inboundSchema,
@@ -23,12 +34,6 @@ import {
   Engine$inboundSchema,
   Engine$outboundSchema,
 } from "./engine.js";
-import {
-  Scaling,
-  Scaling$inboundSchema,
-  Scaling$Outbound,
-  Scaling$outboundSchema,
-} from "./scaling.js";
 
 /**
  * Spin up an inference server.
@@ -42,6 +47,10 @@ export type DeploymentCreate = {
    * Request an accelerator by type and count, with optional fallback types.
    */
   accelerator: AcceleratorSpec;
+  /**
+   * Gateway inference surface exposed by the deployment runtime.
+   */
+  apiFormat?: DeploymentAPIFormat | undefined;
   /**
    * Require an endpoint key (separate from the model's own API key)
    */
@@ -55,9 +64,13 @@ export type DeploymentCreate = {
    */
   engineArgs?: Array<string> | undefined;
   /**
-   * Name/ID of a built image; defaults to the engine's stock image
+   * Digest-pinned OCI image for pure BYO mode; absent or null selects managed mode
    */
   image?: string | null | undefined;
+  /**
+   * Hard aggregate hourly spend ceiling for all replicas and rollout overlap
+   */
+  maxUsdHr: number;
   /**
    * ID of a registered model (see Serving Models)
    */
@@ -67,9 +80,17 @@ export type DeploymentCreate = {
    */
   name: string;
   /**
-   * Autoscaling configuration for a deployment.
+   * Container serving port required for pure BYO mode
+   */
+  port?: number | undefined;
+  /**
+   * Managed runtime recipe key; omitted to resolve uniquely from model metadata
+   */
+  runtime?: string | undefined;
+  /**
+   * Create-time scaling. Omitted min_containers defaults to one; explicit zero is preserved.
    */
-  scaling?: Scaling | undefined;
+  scaling?: DeploymentScalingCreate | undefined;
   secrets?: Array<string> | undefined;
   /**
    * How a config change or rollback is rolled out.
@@ -88,32 +109,42 @@ export const DeploymentCreate$inboundSchema: z.ZodType<
   unknown
 > = z.object({
   accelerator: AcceleratorSpec$inboundSchema,
+  api_format: DeploymentAPIFormat$inboundSchema.optional(),
   endpoint_auth: z.boolean().default(true),
   engine: Engine$inboundSchema.optional(),
   engine_args: z.array(z.string()).optional(),
   image: z.nullable(z.string()).optional(),
+  max_usd_hr: z.number(),
   model: z.string(),
   name: z.string(),
-  scaling: Scaling$inboundSchema.optional(),
+  port: z.number().int().optional(),
+  runtime: z.string().optional(),
+  scaling: DeploymentScalingCreate$inboundSchema.optional(),
   secrets: z.array(z.string()).optional(),
   strategy: DeployStrategy$inboundSchema.optional(),
   volume: z.nullable(z.string()).optional(),
 }).transform((v) => {
   return remap$(v, {
+    "api_format": "apiFormat",
     "endpoint_auth": "endpointAuth",
     "engine_args": "engineArgs",
+    "max_usd_hr": "maxUsdHr",
   });
 });
 /** @internal */
 export type DeploymentCreate$Outbound = {
   accelerator: AcceleratorSpec$Outbound;
+  api_format?: string | undefined;
   endpoint_auth: boolean;
   engine?: string | undefined;
   engine_args?: Array<string> | undefined;
   image?: string | null | undefined;
+  max_usd_hr: number;
   model: string;
   name: string;
-  scaling?: Scaling$Outbound | undefined;
+  port?: number | undefined;
+  runtime?: string | undefined;
+  scaling?: DeploymentScalingCreate$Outbound | undefined;
   secrets?: Array<string> | undefined;
   strategy?: string | undefined;
   volume?: string | null | undefined;
@@ -126,20 +157,26 @@ export const DeploymentCreate$outboundSchema: z.ZodType<
   DeploymentCreate
 > = z.object({
   accelerator: AcceleratorSpec$outboundSchema,
+  apiFormat: DeploymentAPIFormat$outboundSchema.optional(),
   endpointAuth: z.boolean().default(true),
   engine: Engine$outboundSchema.optional(),
   engineArgs: z.array(z.string()).optional(),
   image: z.nullable(z.string()).optional(),
+  maxUsdHr: z.number(),
   model: z.string(),
   name: z.string(),
-  scaling: Scaling$outboundSchema.optional(),
+  port: z.number().int().optional(),
+  runtime: z.string().optional(),
+  scaling: DeploymentScalingCreate$outboundSchema.optional(),
   secrets: z.array(z.string()).optional(),
   strategy: DeployStrategy$outboundSchema.optional(),
   volume: z.nullable(z.string()).optional(),
 }).transform((v) => {
   return remap$(v, {
+    apiFormat: "api_format",
     endpointAuth: "endpoint_auth",
     engineArgs: "engine_args",
+    maxUsdHr: "max_usd_hr",
   });
 });
 

package/src/models/components/deploymentscalingcreate.ts

@@ -0,0 +1,90 @@
+/*
+ * Code generated by Speakeasy (https://speakeasy.com). DO NOT EDIT.
+ */
+
+import * as z from "zod/v3";
+import { remap as remap$ } from "../../lib/primitives.js";
+import { safeParse } from "../../lib/schemas.js";
+import { Result as SafeParseResult } from "../../types/fp.js";
+import { SDKValidationError } from "../errors/sdkvalidationerror.js";
+
+/**
+ * Create-time scaling. Omitted min_containers defaults to one; explicit zero is preserved.
+ */
+export type DeploymentScalingCreate = {
+  bufferContainers?: number | undefined;
+  maxConcurrentInputs?: number | undefined;
+  maxContainers?: number | undefined;
+  /**
+   * Desired warm replicas; omitted defaults to one and explicit zero creates a stopped deployment
+   */
+  minContainers?: number | undefined;
+  scaledownWindowS?: number | undefined;
+};
+
+/** @internal */
+export const DeploymentScalingCreate$inboundSchema: z.ZodType<
+  DeploymentScalingCreate,
+  z.ZodTypeDef,
+  unknown
+> = z.object({
+  buffer_containers: z.number().int().default(0),
+  max_concurrent_inputs: z.number().int().default(100),
+  max_containers: z.number().int().default(1),
+  min_containers: z.number().int().default(1),
+  scaledown_window_s: z.number().int().default(300),
+}).transform((v) => {
+  return remap$(v, {
+    "buffer_containers": "bufferContainers",
+    "max_concurrent_inputs": "maxConcurrentInputs",
+    "max_containers": "maxContainers",
+    "min_containers": "minContainers",
+    "scaledown_window_s": "scaledownWindowS",
+  });
+});
+/** @internal */
+export type DeploymentScalingCreate$Outbound = {
+  buffer_containers: number;
+  max_concurrent_inputs: number;
+  max_containers: number;
+  min_containers: number;
+  scaledown_window_s: number;
+};
+
+/** @internal */
+export const DeploymentScalingCreate$outboundSchema: z.ZodType<
+  DeploymentScalingCreate$Outbound,
+  z.ZodTypeDef,
+  DeploymentScalingCreate
+> = z.object({
+  bufferContainers: z.number().int().default(0),
+  maxConcurrentInputs: z.number().int().default(100),
+  maxContainers: z.number().int().default(1),
+  minContainers: z.number().int().default(1),
+  scaledownWindowS: z.number().int().default(300),
+}).transform((v) => {
+  return remap$(v, {
+    bufferContainers: "buffer_containers",
+    maxConcurrentInputs: "max_concurrent_inputs",
+    maxContainers: "max_containers",
+    minContainers: "min_containers",
+    scaledownWindowS: "scaledown_window_s",
+  });
+});
+
+export function deploymentScalingCreateToJSON(
+  deploymentScalingCreate: DeploymentScalingCreate,
+): string {
+  return JSON.stringify(
+    DeploymentScalingCreate$outboundSchema.parse(deploymentScalingCreate),
+  );
+}
+export function deploymentScalingCreateFromJSON(
+  jsonString: string,
+): SafeParseResult<DeploymentScalingCreate, SDKValidationError> {
+  return safeParse(
+    jsonString,
+    (x) => DeploymentScalingCreate$inboundSchema.parse(JSON.parse(x)),
+    `Failed to parse 'DeploymentScalingCreate' from JSON`,
+  );
+}

package/src/models/components/deploymentscalingupdate.ts

@@ -0,0 +1,87 @@
+/*
+ * Code generated by Speakeasy (https://speakeasy.com). DO NOT EDIT.
+ */
+
+import * as z from "zod/v3";
+import { remap as remap$ } from "../../lib/primitives.js";
+import { safeParse } from "../../lib/schemas.js";
+import { Result as SafeParseResult } from "../../types/fp.js";
+import { SDKValidationError } from "../errors/sdkvalidationerror.js";
+
+/**
+ * Sparse scaling patch; omitted fields retain their effective values.
+ */
+export type DeploymentScalingUpdate = {
+  bufferContainers?: number | undefined;
+  maxConcurrentInputs?: number | undefined;
+  maxContainers?: number | undefined;
+  minContainers?: number | undefined;
+  scaledownWindowS?: number | undefined;
+};
+
+/** @internal */
+export const DeploymentScalingUpdate$inboundSchema: z.ZodType<
+  DeploymentScalingUpdate,
+  z.ZodTypeDef,
+  unknown
+> = z.object({
+  buffer_containers: z.number().int().optional(),
+  max_concurrent_inputs: z.number().int().optional(),
+  max_containers: z.number().int().optional(),
+  min_containers: z.number().int().optional(),
+  scaledown_window_s: z.number().int().optional(),
+}).transform((v) => {
+  return remap$(v, {
+    "buffer_containers": "bufferContainers",
+    "max_concurrent_inputs": "maxConcurrentInputs",
+    "max_containers": "maxContainers",
+    "min_containers": "minContainers",
+    "scaledown_window_s": "scaledownWindowS",
+  });
+});
+/** @internal */
+export type DeploymentScalingUpdate$Outbound = {
+  buffer_containers?: number | undefined;
+  max_concurrent_inputs?: number | undefined;
+  max_containers?: number | undefined;
+  min_containers?: number | undefined;
+  scaledown_window_s?: number | undefined;
+};
+
+/** @internal */
+export const DeploymentScalingUpdate$outboundSchema: z.ZodType<
+  DeploymentScalingUpdate$Outbound,
+  z.ZodTypeDef,
+  DeploymentScalingUpdate
+> = z.object({
+  bufferContainers: z.number().int().optional(),
+  maxConcurrentInputs: z.number().int().optional(),
+  maxContainers: z.number().int().optional(),
+  minContainers: z.number().int().optional(),
+  scaledownWindowS: z.number().int().optional(),
+}).transform((v) => {
+  return remap$(v, {
+    bufferContainers: "buffer_containers",
+    maxConcurrentInputs: "max_concurrent_inputs",
+    maxContainers: "max_containers",
+    minContainers: "min_containers",
+    scaledownWindowS: "scaledown_window_s",
+  });
+});
+
+export function deploymentScalingUpdateToJSON(
+  deploymentScalingUpdate: DeploymentScalingUpdate,
+): string {
+  return JSON.stringify(
+    DeploymentScalingUpdate$outboundSchema.parse(deploymentScalingUpdate),
+  );
+}
+export function deploymentScalingUpdateFromJSON(
+  jsonString: string,
+): SafeParseResult<DeploymentScalingUpdate, SDKValidationError> {
+  return safeParse(
+    jsonString,
+    (x) => DeploymentScalingUpdate$inboundSchema.parse(JSON.parse(x)),
+    `Failed to parse 'DeploymentScalingUpdate' from JSON`,
+  );
+}

package/src/models/components/deploymentupdate.ts

@@ -7,6 +7,11 @@ import { remap as remap$ } from "../../lib/primitives.js";
 import { safeParse } from "../../lib/schemas.js";
 import { Result as SafeParseResult } from "../../types/fp.js";
 import { SDKValidationError } from "../errors/sdkvalidationerror.js";
+import {
+  DeploymentAPIFormat,
+  DeploymentAPIFormat$inboundSchema,
+  DeploymentAPIFormat$outboundSchema,
+} from "./deploymentapiformat.js";
 import {
   DeployStrategy,
   DeployStrategy$inboundSchema,
@@ -17,9 +22,12 @@ import {
  * Patch a deployment's config. Triggers a new revision (rolling by default).
  */
 export type DeploymentUpdate = {
+  apiFormat?: DeploymentAPIFormat | null | undefined;
   engineArgs?: Array<string> | null | undefined;
   image?: string | null | undefined;
   model?: string | null | undefined;
+  port?: number | null | undefined;
+  runtime?: string | null | undefined;
   secrets?: Array<string> | null | undefined;
   /**
    * How a config change or rollback is rolled out.
@@ -33,21 +41,28 @@ export const DeploymentUpdate$inboundSchema: z.ZodType<
   z.ZodTypeDef,
   unknown
 > = z.object({
+  api_format: z.nullable(DeploymentAPIFormat$inboundSchema).optional(),
   engine_args: z.nullable(z.array(z.string())).optional(),
   image: z.nullable(z.string()).optional(),
   model: z.nullable(z.string()).optional(),
+  port: z.nullable(z.number().int()).optional(),
+  runtime: z.nullable(z.string()).optional(),
   secrets: z.nullable(z.array(z.string())).optional(),
   strategy: DeployStrategy$inboundSchema.optional(),
 }).transform((v) => {
   return remap$(v, {
+    "api_format": "apiFormat",
     "engine_args": "engineArgs",
   });
 });
 /** @internal */
 export type DeploymentUpdate$Outbound = {
+  api_format?: string | null | undefined;
   engine_args?: Array<string> | null | undefined;
   image?: string | null | undefined;
   model?: string | null | undefined;
+  port?: number | null | undefined;
+  runtime?: string | null | undefined;
   secrets?: Array<string> | null | undefined;
   strategy?: string | undefined;
 };
@@ -58,13 +73,17 @@ export const DeploymentUpdate$outboundSchema: z.ZodType<
   z.ZodTypeDef,
   DeploymentUpdate
 > = z.object({
+  apiFormat: z.nullable(DeploymentAPIFormat$outboundSchema).optional(),
   engineArgs: z.nullable(z.array(z.string())).optional(),
   image: z.nullable(z.string()).optional(),
   model: z.nullable(z.string()).optional(),
+  port: z.nullable(z.number().int()).optional(),
+  runtime: z.nullable(z.string()).optional(),
   secrets: z.nullable(z.array(z.string())).optional(),
   strategy: DeployStrategy$outboundSchema.optional(),
 }).transform((v) => {
   return remap$(v, {
+    apiFormat: "api_format",
     engineArgs: "engine_args",
   });
 });

package/src/models/components/index.ts

@@ -135,9 +135,12 @@ export * from "./deleteorgautomodelresponse.js";
 export * from "./deletepromptresponse.js";
 export * from "./deleteresponseobject.js";
 export * from "./deployment.js";
+export * from "./deploymentapiformat.js";
 export * from "./deploymentcreate.js";
 export * from "./deploymentendpoint.js";
 export * from "./deploymentrollbackrequest.js";
+export * from "./deploymentscalingcreate.js";
+export * from "./deploymentscalingupdate.js";
 export * from "./deploymentupdate.js";
 export * from "./deploystrategy.js";
 export * from "./doubleclickaction.js";

package/src/models/operations/updatescaling.ts

@@ -15,7 +15,7 @@ export type UpdateScalingRequest = {
    * Optional external end-user identifier forwarded by the API gateway.
    */
   xOnBehalfOf?: string | undefined;
-  scaling: components.Scaling;
+  deploymentScalingUpdate: components.DeploymentScalingUpdate;
 };
 
 /** @internal */
@@ -26,19 +26,19 @@ export const UpdateScalingRequest$inboundSchema: z.ZodType<
 > = z.object({
   deployment_id: z.string(),
   "X-On-Behalf-Of": z.string().optional(),
-  Scaling: components.Scaling$inboundSchema,
+  DeploymentScalingUpdate: components.DeploymentScalingUpdate$inboundSchema,
 }).transform((v) => {
   return remap$(v, {
     "deployment_id": "deploymentId",
     "X-On-Behalf-Of": "xOnBehalfOf",
-    "Scaling": "scaling",
+    "DeploymentScalingUpdate": "deploymentScalingUpdate",
   });
 });
 /** @internal */
 export type UpdateScalingRequest$Outbound = {
   deployment_id: string;
   "X-On-Behalf-Of"?: string | undefined;
-  Scaling: components.Scaling$Outbound;
+  DeploymentScalingUpdate: components.DeploymentScalingUpdate$Outbound;
 };
 
 /** @internal */
@@ -49,12 +49,12 @@ export const UpdateScalingRequest$outboundSchema: z.ZodType<
 > = z.object({
   deploymentId: z.string(),
   xOnBehalfOf: z.string().optional(),
-  scaling: components.Scaling$outboundSchema,
+  deploymentScalingUpdate: components.DeploymentScalingUpdate$outboundSchema,
 }).transform((v) => {
   return remap$(v, {
     deploymentId: "deployment_id",
     xOnBehalfOf: "X-On-Behalf-Of",
-    scaling: "Scaling",
+    deploymentScalingUpdate: "DeploymentScalingUpdate",
   });
 });