@eyepop.ai/eyepop 3.26.0-rc.2 → 3.27.0

package/dist/eyepop.index.d.mts

@@ -64,6 +64,13 @@ interface Prediction extends StreamTime {
     source_url?: string;
     source_type?: string;
     system_timestamp?: number;
+    /**
+     * Compute units the frame's remote calls (a VLM ability's calls to vlm-server) used, in whole units. Absent
+     * when they used none, as for a local model.
+     */
+    compute_units?: number;
+    /** The same in milli compute units; from prediction version 4. */
+    milli_compute_units?: number;
     objects?: Array<PredictedObject>;
     classes?: Array<PredictedClass>;
     texts?: Array<PredictedText>;
@@ -665,6 +672,16 @@ interface InferRequest {
     refresh?: boolean;
     transform_into?: TransformInto;
     response_schema?: ResponseSchema;
+    /**
+     * Auto-label task. The worker derives the prompt and output schema from the task and
+     * `prompts`, so `text_prompt`, `transform_into` and `response_schema` must be unset.
+     * Object detection returns `objects`, image classification `classes`, each labeled with a
+     * prompt's label. A task takes one image. Leave `config.max_new_tokens` unset to use the
+     * worker's default, which leaves room for a detection answer.
+     */
+    task?: AutoAnnotateTask;
+    /** Labels and what each one means, for `task`; required with a task. */
+    prompts?: AutoAnnotatePrompt[];
 }
 interface EvaluateFilter {
     partitions?: string[];
@@ -742,17 +759,45 @@ declare enum VlmAbilityStatus {
     draft = "draft",
     published = "published"
 }
+/**
+ * The model a user-defined VLM ability runs on. `qwen3` (the default when neither `model` nor
+ * `worker_release` is set) runs on EyePop's GPUs; `gpt_6_astra` and `gemini_4_argon` call a remote model.
+ */
+type VlmModel = 'qwen3' | 'gpt_6_astra' | 'gemini_4_argon';
+/** The VLM worker release each {@link VlmModel} runs on. */
+declare const VLM_MODEL_WORKER_RELEASES: Readonly<Record<VlmModel, string>>;
 interface AbilityAliasEntry {
     alias: string;
     tag: string;
 }
+/**
+ * A VLM ability runs in one of two modes:
+ * - free-form: `text_prompt`, optionally with `transform_into` and `response_schema`;
+ * - task: `task` with a non-empty `prompts`. The worker derives the prompt and output schema
+ *   from them, as for {@link InferRequest.task}, so `text_prompt`, `transform_into` and
+ *   `response_schema` must be unset.
+ *
+ * `DataEndpoint.createVlmAbility` refuses a create that mixes the two or has neither.
+ */
 interface VlmAbilityCreate {
     name: string;
     description: string;
-    worker_release: string;
-    text_prompt: string;
-    transform_into: TransformInto;
+    /**
+     * @deprecated Use `model`. Still accepted when it names the release of a {@link VlmModel}
+     * (`qwen3_instruct`, `gpt-6-astra` or `gemini-4-argon`, compared lowercase with `_` as `-`),
+     * and it must then agree with `model` if both are set.
+     */
+    worker_release?: string;
+    /** The model the ability runs on; the Dataset API defaults to `qwen3`. */
+    model?: VlmModel;
+    /** The prompt of a free-form ability; required unless `task` is set. */
+    text_prompt?: string;
+    transform_into?: TransformInto;
     response_schema?: ResponseSchema;
+    /** Auto-label task; see {@link InferRequest.task}. Excludes `text_prompt`, `transform_into` and `response_schema`. */
+    task?: AutoAnnotateTask;
+    /** Labels and what each one means, for `task`; required and non-empty with a task. */
+    prompts?: AutoAnnotatePrompt[];
     config: InferRuntimeConfig;
     is_public: boolean;
     video_chunk_length_ns?: number;
@@ -761,10 +806,19 @@ interface VlmAbilityCreate {
 interface VlmAbilityUpdate {
     name?: string;
     description?: string;
+    /**
+     * @deprecated Use `model`. Still accepted when it names the release of a {@link VlmModel}
+     * (`qwen3_instruct`, `gpt-6-astra` or `gemini-4-argon`, compared lowercase with `_` as `-`),
+     * and it must then agree with `model` if both are set.
+     */
     worker_release?: string;
+    /** The model the ability runs on; left out, the stored model is kept. */
+    model?: VlmModel;
     text_prompt?: string;
     transform_into?: TransformInto;
     response_schema?: ResponseSchema;
+    task?: AutoAnnotateTask;
+    prompts?: AutoAnnotatePrompt[];
     config?: InferRuntimeConfig;
     is_public?: boolean;
     video_chunk_length_ns?: number;
@@ -780,10 +834,19 @@ interface VlmAbility {
     name: string;
     description: string;
     vlm_ability_group_uuid?: string;
+    /** @deprecated Use `model`. The worker release the ability runs on, as stored. */
     worker_release?: string;
+    /**
+     * The model the ability runs on; `null` for a legacy worker release that is none of the
+     * {@link VlmModel}s. `DataEndpoint` derives it from `worker_release` when the Dataset API does not
+     * send it (`qwen3` when the release is unset), and reads a model this SDK does not know as `null`.
+     */
+    model?: VlmModel | null;
     text_prompt?: string;
     transform_into?: TransformInto;
     response_schema?: ResponseSchema;
+    task?: AutoAnnotateTask;
+    prompts?: AutoAnnotatePrompt[];
     config?: InferRuntimeConfig;
     video_chunk_length_ns?: number;
     video_chunk_overlap?: number;
@@ -990,6 +1053,12 @@ interface AssetUuidSource {
 type Source = FileSource | StreamSource | PathSource | UrlSource | AssetUuidSource | MediaStreamSource;
 interface ResultStream extends AsyncIterable<Prediction> {
     cancel(): void;
+    /**
+     * What the stream's remote calls (a VLM ability's calls to vlm-server) used so far, in milli compute units:
+     * a running total while it runs, its total once it ended. Undefined until reported, and where a backend does
+     * not report it; a local model reports none. Informational: the called API bills its calls itself.
+     */
+    readonly milliComputeUnits?: number | undefined;
 }
 declare enum PopComponentType {
     FORWARD = "forward",
@@ -1315,9 +1384,15 @@ interface Options {
      */
     apiKey?: string | undefined;
     /**
-     * Bearer token accepted by the compute API.
+     * Bearer token accepted by the compute API: a login token, which the SDK exchanges at the
+     * compute API, or an already exchanged token, which it uses as is.
      */
     accessToken?: string | undefined;
+    /**
+     * The account a login token (`accessToken` or `oAuth2`) is exchanged for, sent to the compute API
+     * as `account_uuid`. Needed when the user has no single account to pick.
+     */
+    accountId?: string | undefined;
     /**
      * Temporary authentication token for client side execution.
      */
@@ -1415,6 +1490,7 @@ declare class Endpoint<T extends Endpoint<T>> {
     protected _options: Options;
     private _token;
     private _expire_token_time;
+    private _exchanging;
     protected _client: HttpClient | null;
     private _statusRetryHandlers;
     protected _limit: Semaphore;
@@ -1435,11 +1511,19 @@ declare class Endpoint<T extends Endpoint<T>> {
     connect(): Promise<T>;
     disconnect(wait?: boolean): Promise<void>;
     session(): Promise<Session>;
-    protected statusHandler401(_statusCode: number): void;
+    protected statusHandler401(_statusCode: number, failedToken?: string | null): void;
     protected fetchWithRetry(fetcher: () => Promise<Response>, retries?: number): Promise<Response>;
     protected reconnect(): Promise<T>;
     protected authorizationHeader(): Promise<string>;
+    private exchange;
+    private storeExchangedToken;
+    /**
+     * A token that already carries an account, or that the compute API refuses to exchange
+     * (see NOT_EXCHANGEABLE), is used as is. Any other refusal is raised.
+     */
+    private exchangeLoginToken;
     private currentAccessToken;
+    private authenticate;
 }
 
 /**
@@ -1625,9 +1709,15 @@ declare class InferAssetJob implements ResultStream {
     protected _inferRequest: InferRequest;
     protected _response: Promise<Prediction[]> | null;
     protected _runInfo: InferRunInfo | null;
+    protected _milliComputeUnits: number | undefined;
     protected _controller: AbortController;
     protected _requestLogger: Logger;
     [Symbol.asyncIterator](): AsyncIterator<Prediction>;
+    /**
+     * The milli compute units vlm-server billed for the inference, 0 on a cache hit; undefined before it answered
+     * or when it reported none.
+     */
+    get milliComputeUnits(): number | undefined;
     constructor(assetUrl: string, inferRequest: InferRequest, vlmFetch: VlmFetch, requestLogger: Logger);
     start(done: () => void): InferAssetJob;
     cancel(): void;
@@ -1651,6 +1741,60 @@ declare class EyePopDataWebSocketError extends Error {
     readonly sourceError: unknown;
     constructor(url: string, event: unknown);
 }
+/** Why `createVlmAbility` or `updateVlmAbility` refused an ability's mode fields. */
+declare enum VlmAbilityModeErrorReason {
+    /** Neither a non-empty `text_prompt` nor `task` is set (create only). */
+    missing_mode = "missing_mode",
+    /** `task` is not an `AutoAnnotateTask`. */
+    unknown_task = "unknown_task",
+    /** `task` or `prompts` is set together with `text_prompt`, `transform_into` or `response_schema`. */
+    task_excludes_free_form = "task_excludes_free_form",
+    /** `task` is set without `prompts` (create only). */
+    task_without_prompts = "task_without_prompts",
+    /** `prompts` is set without `task` (create only). */
+    prompts_without_task = "prompts_without_task",
+    /** `prompts` is empty or not a list. */
+    empty_prompts = "empty_prompts",
+    /** An entry of `prompts` lacks a non-empty `label` or `prompt`. */
+    invalid_prompt = "invalid_prompt"
+}
+/** Thrown, before any request, for a VLM ability that is neither free-form nor a task, or mixes them. */
+declare class VlmAbilityModeError extends Error {
+    readonly reason: VlmAbilityModeErrorReason;
+    /** The fields at fault, for example the free-form fields set next to a task, or `prompts[0]`. */
+    readonly fields: string[];
+    constructor(reason: VlmAbilityModeErrorReason, message: string, fields?: string[]);
+}
+/** Why `createVlmAbility` or `updateVlmAbility` refused an ability's `model` or `worker_release`. */
+declare enum VlmAbilityModelErrorReason {
+    /** `model` is not a {@link VlmModel}. */
+    unknown_model = "unknown_model",
+    /** `worker_release` is not the release of a {@link VlmModel}. */
+    unknown_worker_release = "unknown_worker_release",
+    /** `model` and `worker_release` are both set and name different models. */
+    model_conflict = "model_conflict"
+}
+/** Thrown, before any request, for a VLM ability whose `model` or deprecated `worker_release` the Dataset API would refuse. */
+declare class VlmAbilityModelError extends Error {
+    readonly reason: VlmAbilityModelErrorReason;
+    /** The fields at fault: `model`, `worker_release` or both. */
+    readonly fields: string[];
+    constructor(reason: VlmAbilityModelErrorReason, message: string, fields?: string[]);
+}
+/** The model of an ability that names neither `model` nor `worker_release`. */
+declare const DEFAULT_VLM_MODEL: VlmModel;
+/**
+ * The comparable form of a worker release name: lowercase, `_` replaced by `-`.
+ * `qwen3_instruct` and `Qwen3-Instruct` both normalize to `qwen3-instruct`.
+ */
+declare function normalizeWorkerRelease(workerRelease: string): string;
+/** The model a worker release runs: `qwen3` for `null` or `undefined`, `null` for a release that is no model's. */
+declare function modelForWorkerRelease(workerRelease: string | null | undefined): VlmModel | null;
+/**
+ * The worker release to store for an ability: the canonical release of the model the fields name,
+ * `qwen3_instruct` when neither is set. Throws a {@link VlmAbilityModelError} like `createVlmAbility` does.
+ */
+declare function resolveWorkerRelease(model: VlmModel | null | undefined, workerRelease: string | null | undefined): string;
 declare class DataEndpoint extends Endpoint<DataEndpoint> {
     private datasetApiUrl;
     private vlmApiUrl;
@@ -2101,8 +2245,13 @@ declare class PipelineJob implements ResultStream {
     readonly whenCancelled: Promise<void>;
     openSource: OpenSource | null;
     stopOnPipeline: (() => void) | null;
+    private _milliComputeUnits;
     constructor(source: Message, media: Media[], live: boolean, params: ProcessParams, mediaStream?: MediaStream | undefined);
     get cancelled(): boolean;
+    /** The source's milli compute units: the latest total of its `usage` messages and its `end`. */
+    get milliComputeUnits(): number | undefined;
+    /** Keep a usage's or an end's milliComputeUnits: each carries the source's total so far. */
+    noteMilliComputeUnits(message: Message): void;
     command(id: string): Message;
     predicted(prediction: Prediction): void;
     finish(error?: Error | null): void;
@@ -2315,4 +2464,4 @@ declare namespace EyePop {
     function dataEndpoint(opts?: DataOptions): DataEndpoint;
 }
 
-export { type AbilityAliasEntry, AbstractJob, type AccessTokenAuth, type AliasResolution, type Annotation, AnnotationInclusionMode, AnnotationType, type ApiKeyAuth, type Area, AreaType, ArtifactType, type Asset, type AssetImport, AssetInclusionMode, AssetStatus, AssetUrlType, type AssetUuidSource, type Auth0Options, type Authentication, type AutoAnnotate, type AutoAnnotateParams, type AutoAnnotatePrompt, AutoAnnotateStatus, AutoAnnotateTask, type AutoPromptConfig, type AutoTask, type BaseArea, type BaseComponent, type BaseTask, type Box, COMBINED_UPLOAD_MAX_BYTES, type Camera, type CameraDistortion, type CameraExtrinsics, type CameraIntrinsics, type ChangeEvent, ChangeType, type ClassificationTask, type CloudBounds, type CloudBox, type ComponentFinderComponent, type ComponentParams, ComputeAccountMismatchError, ComputeApiError, type Contour, type ContourArea, type ContourFinderComponent, ContourType, type CreateWorkflow, type CreateWorkflowConfig, DEFAULT_PREDICTION_VERSION, DataApiType, DataEndpoint, type DataOptions, type DataSession, type Dataset, type DatasetAutoAnnotate, type DatasetAutoAnnotateCreate, type DatasetAutoAnnotateUpdate, type DatasetCreate, type DatasetHeroAssetUpdate, type DatasetParent, type DatasetUpdate, type DatasetVersion, type DatasetVersionAssetStats, type Depth, DepthMap, type DepthSemantic, type DownloadResponse, Endpoint, EndpointState, type EvaluateConfig, type EvaluateFilter, type EvaluateRequest, type EvaluateResponse, type EvaluateRunInfo, EvaluationStatus, ExportedBy, EyePop, EyePopDataWebSocketError, EyepopLineDecoder, type FileSource, type ForwardComponent, ForwardOperatorType, type HttpClient, type InferRequest, type InferRunInfo, type InferRuntimeConfig, type InferenceComponent, InferenceType, type ListWorkFlowItem, type ListWorkFlowItemMetadata, type ListWorkflowItemMetadataLabels, LoadFromAssetUuidJob, LoadFromGroupJob, LoadFromJob, LoadMediaStreamJob, type LocalAuth, MAX_OPEN_SOURCES, type Mask, type MediaStreamSource, type Model, type ModelCreate, type ModelExport, ModelExportStatus, ModelFormat, type ModelMetrics, ModelPrecisionType, type ModelSample, ModelStatus, ModelTask, type ModelTrainingProgress, ModelTrainingStage, ModelType, type ModelUpdate, type MotionDetectConfig, MotionModel, type MultipartContent, type MultipartPart, type OAuth2Auth, type OnChangeEvent, type Options, PIPELINE_PATH, PIPELINE_SOCKETS, type PathSource, PipelineBackend, PipelineCommandError, PipelineConfigurationError, type PipelineConfigurationProblem, PipelineJob, PipelineSocketClosedError, PipelineSourceError, type PlatformSupport, type Point2d, type Point3d, PointCloud, type Pop, type PopComponent, PopComponentType, type PopCrop, type PopDepthMap, type PopForward, type PopForwardOperator, type PopSelect, type PredictedClass, type PredictedEmbedding, type PredictedKeyPoint, type PredictedKeyPoints, type PredictedMesh, type PredictedMotion, type PredictedObject, type PredictedText, type Prediction, PredictionVersion, type ProcessParams, type ProcessRequest, QUATERNION_TOLERANCE, type QcAiHubExportParams, type Quaternion, type RectangleArea, type ResponseSchema, type ResultStream, type Roi, SUBPROTOCOL, SelectMode, Semaphore, type Session, type SessionAuth, SessionBackend, type Source, type SourceDefaults, Stream, type StreamEvent, type StreamSource, type StreamTime, TaskMode, TaskType, type TimeSpan, type TrackingComponent, TranscodeMode, type TransformInto, TransientPopId, UploadGroupJob, UploadJob, type UploadSource, type UrlSource, UserReview, type Vector3, type Vector3d, VideoMode, type VlmAbility, type VlmAbilityCreate, type VlmAbilityGroup, type VlmAbilityGroupCreate, type VlmAbilityGroupUpdate, VlmAbilityStatus, type VlmAbilityUpdate, WebrtcBase, WebrtcWhep, WebrtcWhip, type WorkerBackend, type WorkerContext, WorkerEndpoint, type WorkerOptions, type WorkerSession, type WorkerWork, type Workflow, WorkflowPhase, accessTokenCredential, apiKeyCredential, cloudOfDepth, cloudOfObject, cloudsOfPrediction, decodeDepthMap, decodePointCloud, EyePop as default, encodeMultipartFormData, getBaseUrl, isSelectOperator, popSelects, randomMultipartBoundary, readableStreamFromString, resolveAuth, validateCamera, validatePop };
+export { type AbilityAliasEntry, AbstractJob, type AccessTokenAuth, type AliasResolution, type Annotation, AnnotationInclusionMode, AnnotationType, type ApiKeyAuth, type Area, AreaType, ArtifactType, type Asset, type AssetImport, AssetInclusionMode, AssetStatus, AssetUrlType, type AssetUuidSource, type Auth0Options, type Authentication, type AutoAnnotate, type AutoAnnotateParams, type AutoAnnotatePrompt, AutoAnnotateStatus, AutoAnnotateTask, type AutoPromptConfig, type AutoTask, type BaseArea, type BaseComponent, type BaseTask, type Box, COMBINED_UPLOAD_MAX_BYTES, type Camera, type CameraDistortion, type CameraExtrinsics, type CameraIntrinsics, type ChangeEvent, ChangeType, type ClassificationTask, type CloudBounds, type CloudBox, type ComponentFinderComponent, type ComponentParams, ComputeAccountMismatchError, ComputeApiError, type Contour, type ContourArea, type ContourFinderComponent, ContourType, type CreateWorkflow, type CreateWorkflowConfig, DEFAULT_PREDICTION_VERSION, DEFAULT_VLM_MODEL, DataApiType, DataEndpoint, type DataOptions, type DataSession, type Dataset, type DatasetAutoAnnotate, type DatasetAutoAnnotateCreate, type DatasetAutoAnnotateUpdate, type DatasetCreate, type DatasetHeroAssetUpdate, type DatasetParent, type DatasetUpdate, type DatasetVersion, type DatasetVersionAssetStats, type Depth, DepthMap, type DepthSemantic, type DownloadResponse, Endpoint, EndpointState, type EvaluateConfig, type EvaluateFilter, type EvaluateRequest, type EvaluateResponse, type EvaluateRunInfo, EvaluationStatus, ExportedBy, EyePop, EyePopDataWebSocketError, EyepopLineDecoder, type FileSource, type ForwardComponent, ForwardOperatorType, type HttpClient, type InferRequest, type InferRunInfo, type InferRuntimeConfig, type InferenceComponent, InferenceType, type ListWorkFlowItem, type ListWorkFlowItemMetadata, type ListWorkflowItemMetadataLabels, LoadFromAssetUuidJob, LoadFromGroupJob, LoadFromJob, LoadMediaStreamJob, type LocalAuth, MAX_OPEN_SOURCES, type Mask, type MediaStreamSource, type Model, type ModelCreate, type ModelExport, ModelExportStatus, ModelFormat, type ModelMetrics, ModelPrecisionType, type ModelSample, ModelStatus, ModelTask, type ModelTrainingProgress, ModelTrainingStage, ModelType, type ModelUpdate, type MotionDetectConfig, MotionModel, type MultipartContent, type MultipartPart, type OAuth2Auth, type OnChangeEvent, type Options, PIPELINE_PATH, PIPELINE_SOCKETS, type PathSource, PipelineBackend, PipelineCommandError, PipelineConfigurationError, type PipelineConfigurationProblem, PipelineJob, PipelineSocketClosedError, PipelineSourceError, type PlatformSupport, type Point2d, type Point3d, PointCloud, type Pop, type PopComponent, PopComponentType, type PopCrop, type PopDepthMap, type PopForward, type PopForwardOperator, type PopSelect, type PredictedClass, type PredictedEmbedding, type PredictedKeyPoint, type PredictedKeyPoints, type PredictedMesh, type PredictedMotion, type PredictedObject, type PredictedText, type Prediction, PredictionVersion, type ProcessParams, type ProcessRequest, QUATERNION_TOLERANCE, type QcAiHubExportParams, type Quaternion, type RectangleArea, type ResponseSchema, type ResultStream, type Roi, SUBPROTOCOL, SelectMode, Semaphore, type Session, type SessionAuth, SessionBackend, type Source, type SourceDefaults, Stream, type StreamEvent, type StreamSource, type StreamTime, TaskMode, TaskType, type TimeSpan, type TrackingComponent, TranscodeMode, type TransformInto, TransientPopId, UploadGroupJob, UploadJob, type UploadSource, type UrlSource, UserReview, VLM_MODEL_WORKER_RELEASES, type Vector3, type Vector3d, VideoMode, type VlmAbility, type VlmAbilityCreate, type VlmAbilityGroup, type VlmAbilityGroupCreate, type VlmAbilityGroupUpdate, VlmAbilityModeError, VlmAbilityModeErrorReason, VlmAbilityModelError, VlmAbilityModelErrorReason, VlmAbilityStatus, type VlmAbilityUpdate, type VlmModel, WebrtcBase, WebrtcWhep, WebrtcWhip, type WorkerBackend, type WorkerContext, WorkerEndpoint, type WorkerOptions, type WorkerSession, type WorkerWork, type Workflow, WorkflowPhase, accessTokenCredential, apiKeyCredential, cloudOfDepth, cloudOfObject, cloudsOfPrediction, decodeDepthMap, decodePointCloud, EyePop as default, encodeMultipartFormData, getBaseUrl, isSelectOperator, modelForWorkerRelease, normalizeWorkerRelease, popSelects, randomMultipartBoundary, readableStreamFromString, resolveAuth, resolveWorkerRelease, validateCamera, validatePop };

package/dist/eyepop.index.d.ts

@@ -64,6 +64,13 @@ interface Prediction extends StreamTime {
     source_url?: string;
     source_type?: string;
     system_timestamp?: number;
+    /**
+     * Compute units the frame's remote calls (a VLM ability's calls to vlm-server) used, in whole units. Absent
+     * when they used none, as for a local model.
+     */
+    compute_units?: number;
+    /** The same in milli compute units; from prediction version 4. */
+    milli_compute_units?: number;
     objects?: Array<PredictedObject>;
     classes?: Array<PredictedClass>;
     texts?: Array<PredictedText>;
@@ -665,6 +672,16 @@ interface InferRequest {
     refresh?: boolean;
     transform_into?: TransformInto;
     response_schema?: ResponseSchema;
+    /**
+     * Auto-label task. The worker derives the prompt and output schema from the task and
+     * `prompts`, so `text_prompt`, `transform_into` and `response_schema` must be unset.
+     * Object detection returns `objects`, image classification `classes`, each labeled with a
+     * prompt's label. A task takes one image. Leave `config.max_new_tokens` unset to use the
+     * worker's default, which leaves room for a detection answer.
+     */
+    task?: AutoAnnotateTask;
+    /** Labels and what each one means, for `task`; required with a task. */
+    prompts?: AutoAnnotatePrompt[];
 }
 interface EvaluateFilter {
     partitions?: string[];
@@ -742,17 +759,45 @@ declare enum VlmAbilityStatus {
     draft = "draft",
     published = "published"
 }
+/**
+ * The model a user-defined VLM ability runs on. `qwen3` (the default when neither `model` nor
+ * `worker_release` is set) runs on EyePop's GPUs; `gpt_6_astra` and `gemini_4_argon` call a remote model.
+ */
+type VlmModel = 'qwen3' | 'gpt_6_astra' | 'gemini_4_argon';
+/** The VLM worker release each {@link VlmModel} runs on. */
+declare const VLM_MODEL_WORKER_RELEASES: Readonly<Record<VlmModel, string>>;
 interface AbilityAliasEntry {
     alias: string;
     tag: string;
 }
+/**
+ * A VLM ability runs in one of two modes:
+ * - free-form: `text_prompt`, optionally with `transform_into` and `response_schema`;
+ * - task: `task` with a non-empty `prompts`. The worker derives the prompt and output schema
+ *   from them, as for {@link InferRequest.task}, so `text_prompt`, `transform_into` and
+ *   `response_schema` must be unset.
+ *
+ * `DataEndpoint.createVlmAbility` refuses a create that mixes the two or has neither.
+ */
 interface VlmAbilityCreate {
     name: string;
     description: string;
-    worker_release: string;
-    text_prompt: string;
-    transform_into: TransformInto;
+    /**
+     * @deprecated Use `model`. Still accepted when it names the release of a {@link VlmModel}
+     * (`qwen3_instruct`, `gpt-6-astra` or `gemini-4-argon`, compared lowercase with `_` as `-`),
+     * and it must then agree with `model` if both are set.
+     */
+    worker_release?: string;
+    /** The model the ability runs on; the Dataset API defaults to `qwen3`. */
+    model?: VlmModel;
+    /** The prompt of a free-form ability; required unless `task` is set. */
+    text_prompt?: string;
+    transform_into?: TransformInto;
     response_schema?: ResponseSchema;
+    /** Auto-label task; see {@link InferRequest.task}. Excludes `text_prompt`, `transform_into` and `response_schema`. */
+    task?: AutoAnnotateTask;
+    /** Labels and what each one means, for `task`; required and non-empty with a task. */
+    prompts?: AutoAnnotatePrompt[];
     config: InferRuntimeConfig;
     is_public: boolean;
     video_chunk_length_ns?: number;
@@ -761,10 +806,19 @@ interface VlmAbilityCreate {
 interface VlmAbilityUpdate {
     name?: string;
     description?: string;
+    /**
+     * @deprecated Use `model`. Still accepted when it names the release of a {@link VlmModel}
+     * (`qwen3_instruct`, `gpt-6-astra` or `gemini-4-argon`, compared lowercase with `_` as `-`),
+     * and it must then agree with `model` if both are set.
+     */
     worker_release?: string;
+    /** The model the ability runs on; left out, the stored model is kept. */
+    model?: VlmModel;
     text_prompt?: string;
     transform_into?: TransformInto;
     response_schema?: ResponseSchema;
+    task?: AutoAnnotateTask;
+    prompts?: AutoAnnotatePrompt[];
     config?: InferRuntimeConfig;
     is_public?: boolean;
     video_chunk_length_ns?: number;
@@ -780,10 +834,19 @@ interface VlmAbility {
     name: string;
     description: string;
     vlm_ability_group_uuid?: string;
+    /** @deprecated Use `model`. The worker release the ability runs on, as stored. */
     worker_release?: string;
+    /**
+     * The model the ability runs on; `null` for a legacy worker release that is none of the
+     * {@link VlmModel}s. `DataEndpoint` derives it from `worker_release` when the Dataset API does not
+     * send it (`qwen3` when the release is unset), and reads a model this SDK does not know as `null`.
+     */
+    model?: VlmModel | null;
     text_prompt?: string;
     transform_into?: TransformInto;
     response_schema?: ResponseSchema;
+    task?: AutoAnnotateTask;
+    prompts?: AutoAnnotatePrompt[];
     config?: InferRuntimeConfig;
     video_chunk_length_ns?: number;
     video_chunk_overlap?: number;
@@ -990,6 +1053,12 @@ interface AssetUuidSource {
 type Source = FileSource | StreamSource | PathSource | UrlSource | AssetUuidSource | MediaStreamSource;
 interface ResultStream extends AsyncIterable<Prediction> {
     cancel(): void;
+    /**
+     * What the stream's remote calls (a VLM ability's calls to vlm-server) used so far, in milli compute units:
+     * a running total while it runs, its total once it ended. Undefined until reported, and where a backend does
+     * not report it; a local model reports none. Informational: the called API bills its calls itself.
+     */
+    readonly milliComputeUnits?: number | undefined;
 }
 declare enum PopComponentType {
     FORWARD = "forward",
@@ -1315,9 +1384,15 @@ interface Options {
      */
     apiKey?: string | undefined;
     /**
-     * Bearer token accepted by the compute API.
+     * Bearer token accepted by the compute API: a login token, which the SDK exchanges at the
+     * compute API, or an already exchanged token, which it uses as is.
      */
     accessToken?: string | undefined;
+    /**
+     * The account a login token (`accessToken` or `oAuth2`) is exchanged for, sent to the compute API
+     * as `account_uuid`. Needed when the user has no single account to pick.
+     */
+    accountId?: string | undefined;
     /**
      * Temporary authentication token for client side execution.
      */
@@ -1415,6 +1490,7 @@ declare class Endpoint<T extends Endpoint<T>> {
     protected _options: Options;
     private _token;
     private _expire_token_time;
+    private _exchanging;
     protected _client: HttpClient | null;
     private _statusRetryHandlers;
     protected _limit: Semaphore;
@@ -1435,11 +1511,19 @@ declare class Endpoint<T extends Endpoint<T>> {
     connect(): Promise<T>;
     disconnect(wait?: boolean): Promise<void>;
     session(): Promise<Session>;
-    protected statusHandler401(_statusCode: number): void;
+    protected statusHandler401(_statusCode: number, failedToken?: string | null): void;
     protected fetchWithRetry(fetcher: () => Promise<Response>, retries?: number): Promise<Response>;
     protected reconnect(): Promise<T>;
     protected authorizationHeader(): Promise<string>;
+    private exchange;
+    private storeExchangedToken;
+    /**
+     * A token that already carries an account, or that the compute API refuses to exchange
+     * (see NOT_EXCHANGEABLE), is used as is. Any other refusal is raised.
+     */
+    private exchangeLoginToken;
     private currentAccessToken;
+    private authenticate;
 }
 
 /**
@@ -1625,9 +1709,15 @@ declare class InferAssetJob implements ResultStream {
     protected _inferRequest: InferRequest;
     protected _response: Promise<Prediction[]> | null;
     protected _runInfo: InferRunInfo | null;
+    protected _milliComputeUnits: number | undefined;
     protected _controller: AbortController;
     protected _requestLogger: Logger;
     [Symbol.asyncIterator](): AsyncIterator<Prediction>;
+    /**
+     * The milli compute units vlm-server billed for the inference, 0 on a cache hit; undefined before it answered
+     * or when it reported none.
+     */
+    get milliComputeUnits(): number | undefined;
     constructor(assetUrl: string, inferRequest: InferRequest, vlmFetch: VlmFetch, requestLogger: Logger);
     start(done: () => void): InferAssetJob;
     cancel(): void;
@@ -1651,6 +1741,60 @@ declare class EyePopDataWebSocketError extends Error {
     readonly sourceError: unknown;
     constructor(url: string, event: unknown);
 }
+/** Why `createVlmAbility` or `updateVlmAbility` refused an ability's mode fields. */
+declare enum VlmAbilityModeErrorReason {
+    /** Neither a non-empty `text_prompt` nor `task` is set (create only). */
+    missing_mode = "missing_mode",
+    /** `task` is not an `AutoAnnotateTask`. */
+    unknown_task = "unknown_task",
+    /** `task` or `prompts` is set together with `text_prompt`, `transform_into` or `response_schema`. */
+    task_excludes_free_form = "task_excludes_free_form",
+    /** `task` is set without `prompts` (create only). */
+    task_without_prompts = "task_without_prompts",
+    /** `prompts` is set without `task` (create only). */
+    prompts_without_task = "prompts_without_task",
+    /** `prompts` is empty or not a list. */
+    empty_prompts = "empty_prompts",
+    /** An entry of `prompts` lacks a non-empty `label` or `prompt`. */
+    invalid_prompt = "invalid_prompt"
+}
+/** Thrown, before any request, for a VLM ability that is neither free-form nor a task, or mixes them. */
+declare class VlmAbilityModeError extends Error {
+    readonly reason: VlmAbilityModeErrorReason;
+    /** The fields at fault, for example the free-form fields set next to a task, or `prompts[0]`. */
+    readonly fields: string[];
+    constructor(reason: VlmAbilityModeErrorReason, message: string, fields?: string[]);
+}
+/** Why `createVlmAbility` or `updateVlmAbility` refused an ability's `model` or `worker_release`. */
+declare enum VlmAbilityModelErrorReason {
+    /** `model` is not a {@link VlmModel}. */
+    unknown_model = "unknown_model",
+    /** `worker_release` is not the release of a {@link VlmModel}. */
+    unknown_worker_release = "unknown_worker_release",
+    /** `model` and `worker_release` are both set and name different models. */
+    model_conflict = "model_conflict"
+}
+/** Thrown, before any request, for a VLM ability whose `model` or deprecated `worker_release` the Dataset API would refuse. */
+declare class VlmAbilityModelError extends Error {
+    readonly reason: VlmAbilityModelErrorReason;
+    /** The fields at fault: `model`, `worker_release` or both. */
+    readonly fields: string[];
+    constructor(reason: VlmAbilityModelErrorReason, message: string, fields?: string[]);
+}
+/** The model of an ability that names neither `model` nor `worker_release`. */
+declare const DEFAULT_VLM_MODEL: VlmModel;
+/**
+ * The comparable form of a worker release name: lowercase, `_` replaced by `-`.
+ * `qwen3_instruct` and `Qwen3-Instruct` both normalize to `qwen3-instruct`.
+ */
+declare function normalizeWorkerRelease(workerRelease: string): string;
+/** The model a worker release runs: `qwen3` for `null` or `undefined`, `null` for a release that is no model's. */
+declare function modelForWorkerRelease(workerRelease: string | null | undefined): VlmModel | null;
+/**
+ * The worker release to store for an ability: the canonical release of the model the fields name,
+ * `qwen3_instruct` when neither is set. Throws a {@link VlmAbilityModelError} like `createVlmAbility` does.
+ */
+declare function resolveWorkerRelease(model: VlmModel | null | undefined, workerRelease: string | null | undefined): string;
 declare class DataEndpoint extends Endpoint<DataEndpoint> {
     private datasetApiUrl;
     private vlmApiUrl;
@@ -2101,8 +2245,13 @@ declare class PipelineJob implements ResultStream {
     readonly whenCancelled: Promise<void>;
     openSource: OpenSource | null;
     stopOnPipeline: (() => void) | null;
+    private _milliComputeUnits;
     constructor(source: Message, media: Media[], live: boolean, params: ProcessParams, mediaStream?: MediaStream | undefined);
     get cancelled(): boolean;
+    /** The source's milli compute units: the latest total of its `usage` messages and its `end`. */
+    get milliComputeUnits(): number | undefined;
+    /** Keep a usage's or an end's milliComputeUnits: each carries the source's total so far. */
+    noteMilliComputeUnits(message: Message): void;
     command(id: string): Message;
     predicted(prediction: Prediction): void;
     finish(error?: Error | null): void;
@@ -2315,4 +2464,4 @@ declare namespace EyePop {
     function dataEndpoint(opts?: DataOptions): DataEndpoint;
 }
 
-export { type AbilityAliasEntry, AbstractJob, type AccessTokenAuth, type AliasResolution, type Annotation, AnnotationInclusionMode, AnnotationType, type ApiKeyAuth, type Area, AreaType, ArtifactType, type Asset, type AssetImport, AssetInclusionMode, AssetStatus, AssetUrlType, type AssetUuidSource, type Auth0Options, type Authentication, type AutoAnnotate, type AutoAnnotateParams, type AutoAnnotatePrompt, AutoAnnotateStatus, AutoAnnotateTask, type AutoPromptConfig, type AutoTask, type BaseArea, type BaseComponent, type BaseTask, type Box, COMBINED_UPLOAD_MAX_BYTES, type Camera, type CameraDistortion, type CameraExtrinsics, type CameraIntrinsics, type ChangeEvent, ChangeType, type ClassificationTask, type CloudBounds, type CloudBox, type ComponentFinderComponent, type ComponentParams, ComputeAccountMismatchError, ComputeApiError, type Contour, type ContourArea, type ContourFinderComponent, ContourType, type CreateWorkflow, type CreateWorkflowConfig, DEFAULT_PREDICTION_VERSION, DataApiType, DataEndpoint, type DataOptions, type DataSession, type Dataset, type DatasetAutoAnnotate, type DatasetAutoAnnotateCreate, type DatasetAutoAnnotateUpdate, type DatasetCreate, type DatasetHeroAssetUpdate, type DatasetParent, type DatasetUpdate, type DatasetVersion, type DatasetVersionAssetStats, type Depth, DepthMap, type DepthSemantic, type DownloadResponse, Endpoint, EndpointState, type EvaluateConfig, type EvaluateFilter, type EvaluateRequest, type EvaluateResponse, type EvaluateRunInfo, EvaluationStatus, ExportedBy, EyePop, EyePopDataWebSocketError, EyepopLineDecoder, type FileSource, type ForwardComponent, ForwardOperatorType, type HttpClient, type InferRequest, type InferRunInfo, type InferRuntimeConfig, type InferenceComponent, InferenceType, type ListWorkFlowItem, type ListWorkFlowItemMetadata, type ListWorkflowItemMetadataLabels, LoadFromAssetUuidJob, LoadFromGroupJob, LoadFromJob, LoadMediaStreamJob, type LocalAuth, MAX_OPEN_SOURCES, type Mask, type MediaStreamSource, type Model, type ModelCreate, type ModelExport, ModelExportStatus, ModelFormat, type ModelMetrics, ModelPrecisionType, type ModelSample, ModelStatus, ModelTask, type ModelTrainingProgress, ModelTrainingStage, ModelType, type ModelUpdate, type MotionDetectConfig, MotionModel, type MultipartContent, type MultipartPart, type OAuth2Auth, type OnChangeEvent, type Options, PIPELINE_PATH, PIPELINE_SOCKETS, type PathSource, PipelineBackend, PipelineCommandError, PipelineConfigurationError, type PipelineConfigurationProblem, PipelineJob, PipelineSocketClosedError, PipelineSourceError, type PlatformSupport, type Point2d, type Point3d, PointCloud, type Pop, type PopComponent, PopComponentType, type PopCrop, type PopDepthMap, type PopForward, type PopForwardOperator, type PopSelect, type PredictedClass, type PredictedEmbedding, type PredictedKeyPoint, type PredictedKeyPoints, type PredictedMesh, type PredictedMotion, type PredictedObject, type PredictedText, type Prediction, PredictionVersion, type ProcessParams, type ProcessRequest, QUATERNION_TOLERANCE, type QcAiHubExportParams, type Quaternion, type RectangleArea, type ResponseSchema, type ResultStream, type Roi, SUBPROTOCOL, SelectMode, Semaphore, type Session, type SessionAuth, SessionBackend, type Source, type SourceDefaults, Stream, type StreamEvent, type StreamSource, type StreamTime, TaskMode, TaskType, type TimeSpan, type TrackingComponent, TranscodeMode, type TransformInto, TransientPopId, UploadGroupJob, UploadJob, type UploadSource, type UrlSource, UserReview, type Vector3, type Vector3d, VideoMode, type VlmAbility, type VlmAbilityCreate, type VlmAbilityGroup, type VlmAbilityGroupCreate, type VlmAbilityGroupUpdate, VlmAbilityStatus, type VlmAbilityUpdate, WebrtcBase, WebrtcWhep, WebrtcWhip, type WorkerBackend, type WorkerContext, WorkerEndpoint, type WorkerOptions, type WorkerSession, type WorkerWork, type Workflow, WorkflowPhase, accessTokenCredential, apiKeyCredential, cloudOfDepth, cloudOfObject, cloudsOfPrediction, decodeDepthMap, decodePointCloud, EyePop as default, encodeMultipartFormData, getBaseUrl, isSelectOperator, popSelects, randomMultipartBoundary, readableStreamFromString, resolveAuth, validateCamera, validatePop };
+export { type AbilityAliasEntry, AbstractJob, type AccessTokenAuth, type AliasResolution, type Annotation, AnnotationInclusionMode, AnnotationType, type ApiKeyAuth, type Area, AreaType, ArtifactType, type Asset, type AssetImport, AssetInclusionMode, AssetStatus, AssetUrlType, type AssetUuidSource, type Auth0Options, type Authentication, type AutoAnnotate, type AutoAnnotateParams, type AutoAnnotatePrompt, AutoAnnotateStatus, AutoAnnotateTask, type AutoPromptConfig, type AutoTask, type BaseArea, type BaseComponent, type BaseTask, type Box, COMBINED_UPLOAD_MAX_BYTES, type Camera, type CameraDistortion, type CameraExtrinsics, type CameraIntrinsics, type ChangeEvent, ChangeType, type ClassificationTask, type CloudBounds, type CloudBox, type ComponentFinderComponent, type ComponentParams, ComputeAccountMismatchError, ComputeApiError, type Contour, type ContourArea, type ContourFinderComponent, ContourType, type CreateWorkflow, type CreateWorkflowConfig, DEFAULT_PREDICTION_VERSION, DEFAULT_VLM_MODEL, DataApiType, DataEndpoint, type DataOptions, type DataSession, type Dataset, type DatasetAutoAnnotate, type DatasetAutoAnnotateCreate, type DatasetAutoAnnotateUpdate, type DatasetCreate, type DatasetHeroAssetUpdate, type DatasetParent, type DatasetUpdate, type DatasetVersion, type DatasetVersionAssetStats, type Depth, DepthMap, type DepthSemantic, type DownloadResponse, Endpoint, EndpointState, type EvaluateConfig, type EvaluateFilter, type EvaluateRequest, type EvaluateResponse, type EvaluateRunInfo, EvaluationStatus, ExportedBy, EyePop, EyePopDataWebSocketError, EyepopLineDecoder, type FileSource, type ForwardComponent, ForwardOperatorType, type HttpClient, type InferRequest, type InferRunInfo, type InferRuntimeConfig, type InferenceComponent, InferenceType, type ListWorkFlowItem, type ListWorkFlowItemMetadata, type ListWorkflowItemMetadataLabels, LoadFromAssetUuidJob, LoadFromGroupJob, LoadFromJob, LoadMediaStreamJob, type LocalAuth, MAX_OPEN_SOURCES, type Mask, type MediaStreamSource, type Model, type ModelCreate, type ModelExport, ModelExportStatus, ModelFormat, type ModelMetrics, ModelPrecisionType, type ModelSample, ModelStatus, ModelTask, type ModelTrainingProgress, ModelTrainingStage, ModelType, type ModelUpdate, type MotionDetectConfig, MotionModel, type MultipartContent, type MultipartPart, type OAuth2Auth, type OnChangeEvent, type Options, PIPELINE_PATH, PIPELINE_SOCKETS, type PathSource, PipelineBackend, PipelineCommandError, PipelineConfigurationError, type PipelineConfigurationProblem, PipelineJob, PipelineSocketClosedError, PipelineSourceError, type PlatformSupport, type Point2d, type Point3d, PointCloud, type Pop, type PopComponent, PopComponentType, type PopCrop, type PopDepthMap, type PopForward, type PopForwardOperator, type PopSelect, type PredictedClass, type PredictedEmbedding, type PredictedKeyPoint, type PredictedKeyPoints, type PredictedMesh, type PredictedMotion, type PredictedObject, type PredictedText, type Prediction, PredictionVersion, type ProcessParams, type ProcessRequest, QUATERNION_TOLERANCE, type QcAiHubExportParams, type Quaternion, type RectangleArea, type ResponseSchema, type ResultStream, type Roi, SUBPROTOCOL, SelectMode, Semaphore, type Session, type SessionAuth, SessionBackend, type Source, type SourceDefaults, Stream, type StreamEvent, type StreamSource, type StreamTime, TaskMode, TaskType, type TimeSpan, type TrackingComponent, TranscodeMode, type TransformInto, TransientPopId, UploadGroupJob, UploadJob, type UploadSource, type UrlSource, UserReview, VLM_MODEL_WORKER_RELEASES, type Vector3, type Vector3d, VideoMode, type VlmAbility, type VlmAbilityCreate, type VlmAbilityGroup, type VlmAbilityGroupCreate, type VlmAbilityGroupUpdate, VlmAbilityModeError, VlmAbilityModeErrorReason, VlmAbilityModelError, VlmAbilityModelErrorReason, VlmAbilityStatus, type VlmAbilityUpdate, type VlmModel, WebrtcBase, WebrtcWhep, WebrtcWhip, type WorkerBackend, type WorkerContext, WorkerEndpoint, type WorkerOptions, type WorkerSession, type WorkerWork, type Workflow, WorkflowPhase, accessTokenCredential, apiKeyCredential, cloudOfDepth, cloudOfObject, cloudsOfPrediction, decodeDepthMap, decodePointCloud, EyePop as default, encodeMultipartFormData, getBaseUrl, isSelectOperator, modelForWorkerRelease, normalizeWorkerRelease, popSelects, randomMultipartBoundary, readableStreamFromString, resolveAuth, resolveWorkerRelease, validateCamera, validatePop };