@eyepop.ai/eyepop 3.25.3 → 3.25.5

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[];
@@ -990,6 +1007,12 @@ 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 +1338,15 @@ 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 +1444,7 @@ 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 +1465,19 @@ 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 as an exchanged
+     * token ("API key or user login required"), is used as is. Any other refusal is raised.
+     */
+    private exchangeLoginToken;
     private currentAccessToken;
+    private authenticate;
 }
 
 /**
@@ -1477,6 +1515,11 @@ interface WorkerOptions extends Options {
      * Defaults to process.env['EYEPOP_PIPELINE_MODE'].
      */
     isPipelineMode?: boolean;
+    /**
+     * In pipeline mode, the most sockets the endpoint opens. Each runs its own pipeline, which plays its
+     * sources one after another; a new one opens only while every open one has a source. Defaults to 4.
+     */
+    pipelineSockets?: number;
 }
 
 /**
@@ -1620,9 +1663,15 @@ 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;
@@ -2014,10 +2063,12 @@ declare const PIPELINE_PATH = "/api/v1/pipeline";
 declare const SUBPROTOCOL = "eyepop.v1";
 /** A socket may have this many open sources; one more is answered `queue_full`. */
 declare const MAX_OPEN_SOURCES = 256;
+/** The most sockets an endpoint opens unless `pipelineSockets` says otherwise. */
+declare const PIPELINE_SOCKETS = 4;
 /** Uploads up to this size go as one combined message, larger ones in chunks. */
 declare const COMBINED_UPLOAD_MAX_BYTES: number;
 /** Why pipeline mode refused its configuration. */
-type PipelineConfigurationProblem = 'session_options' | 'job_queue_length' | 'no_pipeline_api_url' | 'insecure_url';
+type PipelineConfigurationProblem = 'session_options' | 'job_queue_length' | 'pipeline_sockets' | 'no_pipeline_api_url' | 'insecure_url';
 /** Pipeline mode cannot run as configured; `options` names the refused options for `session_options`. */
 declare class PipelineConfigurationError extends Error {
     readonly problem: PipelineConfigurationProblem;
@@ -2094,26 +2145,36 @@ 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;
     cancel(): void;
     [Symbol.asyncIterator](): AsyncIterator<Prediction>;
 }
-/** Runs jobs on one socket to the pipeline API, opened when the endpoint connects. */
+/**
+ * Runs jobs on up to `pipelineSockets` sockets to the pipeline API; connecting opens the first. A
+ * socket's pipeline plays its sources one after another, so a batch on one socket leaves the GPU
+ * mostly idle. While every open socket has a source and fewer than `pipelineSockets` are open, the next
+ * job opens another; each job goes on the socket that has room and the fewest open sources.
+ */
 declare class PipelineBackend implements WorkerBackend {
     private readonly ctx;
     private _pipelineApiUrl;
-    private _socket;
+    private _sockets;
     private _connecting;
     private _sendChain;
     private _disconnecting;
     private _nextId;
     private _currentPop;
     private readonly _warned;
-    private _heartbeat;
+    private _noExtraSocketUntil;
     private _reconnects;
     constructor(ctx: WorkerContext);
     reconnect(): Promise<void>;
@@ -2122,8 +2183,17 @@ declare class PipelineBackend implements WorkerBackend {
     popName(): string | null;
     authorizationHeader(): Promise<string>;
     private inOrder;
-    /** The open socket, or a new one with the last pop set on it. */
+    private openSockets;
+    /** An open socket, or a new one with the last pop set on it. */
     private ensureSocket;
+    /**
+     * The socket for the next job; the caller runs it in submission order. A live job goes on a socket
+     * of its own, an idle one or a new one, which `pipelineSockets` does not count. Any other job skips
+     * the sockets with a live source: a new one while every other socket has a source and fewer than
+     * `pipelineSockets` are open, else the one with room and the fewest sources.
+     */
+    private socketForJob;
+    /** A new socket with the last pop set on it; `patient` waits through 503 with Retry-After. */
     private openSocket;
     private startHeartbeat;
     private socketUrl;
@@ -2137,7 +2207,11 @@ declare class PipelineBackend implements WorkerBackend {
     /** Send a command that is answered by `result` or `error` under its id. */
     private command;
     pop(): Pop | null;
-    /** Set the pop on the socket's pipeline; a pop it refuses rejects, and the last pop it took stays. */
+    /**
+     * Set the pop on every open socket's pipeline; a pop the first refuses rejects, and the last pop it
+     * took stays. shortcut: a refusal from a later socket, which the first took, leaves the earlier ones
+     * on the new pop; a refusal is the pop's own fault, so every socket answers it alike.
+     */
     changePop(pop: Pop): Promise<void>;
     private setPopOn;
     submit(work: WorkerWork): Promise<ResultStream>;
@@ -2290,4 +2364,4 @@ 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, 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, 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 };

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[];
@@ -990,6 +1007,12 @@ 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 +1338,15 @@ 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 +1444,7 @@ 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 +1465,19 @@ 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 as an exchanged
+     * token ("API key or user login required"), is used as is. Any other refusal is raised.
+     */
+    private exchangeLoginToken;
     private currentAccessToken;
+    private authenticate;
 }
 
 /**
@@ -1477,6 +1515,11 @@ interface WorkerOptions extends Options {
      * Defaults to process.env['EYEPOP_PIPELINE_MODE'].
      */
     isPipelineMode?: boolean;
+    /**
+     * In pipeline mode, the most sockets the endpoint opens. Each runs its own pipeline, which plays its
+     * sources one after another; a new one opens only while every open one has a source. Defaults to 4.
+     */
+    pipelineSockets?: number;
 }
 
 /**
@@ -1620,9 +1663,15 @@ 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;
@@ -2014,10 +2063,12 @@ declare const PIPELINE_PATH = "/api/v1/pipeline";
 declare const SUBPROTOCOL = "eyepop.v1";
 /** A socket may have this many open sources; one more is answered `queue_full`. */
 declare const MAX_OPEN_SOURCES = 256;
+/** The most sockets an endpoint opens unless `pipelineSockets` says otherwise. */
+declare const PIPELINE_SOCKETS = 4;
 /** Uploads up to this size go as one combined message, larger ones in chunks. */
 declare const COMBINED_UPLOAD_MAX_BYTES: number;
 /** Why pipeline mode refused its configuration. */
-type PipelineConfigurationProblem = 'session_options' | 'job_queue_length' | 'no_pipeline_api_url' | 'insecure_url';
+type PipelineConfigurationProblem = 'session_options' | 'job_queue_length' | 'pipeline_sockets' | 'no_pipeline_api_url' | 'insecure_url';
 /** Pipeline mode cannot run as configured; `options` names the refused options for `session_options`. */
 declare class PipelineConfigurationError extends Error {
     readonly problem: PipelineConfigurationProblem;
@@ -2094,26 +2145,36 @@ 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;
     cancel(): void;
     [Symbol.asyncIterator](): AsyncIterator<Prediction>;
 }
-/** Runs jobs on one socket to the pipeline API, opened when the endpoint connects. */
+/**
+ * Runs jobs on up to `pipelineSockets` sockets to the pipeline API; connecting opens the first. A
+ * socket's pipeline plays its sources one after another, so a batch on one socket leaves the GPU
+ * mostly idle. While every open socket has a source and fewer than `pipelineSockets` are open, the next
+ * job opens another; each job goes on the socket that has room and the fewest open sources.
+ */
 declare class PipelineBackend implements WorkerBackend {
     private readonly ctx;
     private _pipelineApiUrl;
-    private _socket;
+    private _sockets;
     private _connecting;
     private _sendChain;
     private _disconnecting;
     private _nextId;
     private _currentPop;
     private readonly _warned;
-    private _heartbeat;
+    private _noExtraSocketUntil;
     private _reconnects;
     constructor(ctx: WorkerContext);
     reconnect(): Promise<void>;
@@ -2122,8 +2183,17 @@ declare class PipelineBackend implements WorkerBackend {
     popName(): string | null;
     authorizationHeader(): Promise<string>;
     private inOrder;
-    /** The open socket, or a new one with the last pop set on it. */
+    private openSockets;
+    /** An open socket, or a new one with the last pop set on it. */
     private ensureSocket;
+    /**
+     * The socket for the next job; the caller runs it in submission order. A live job goes on a socket
+     * of its own, an idle one or a new one, which `pipelineSockets` does not count. Any other job skips
+     * the sockets with a live source: a new one while every other socket has a source and fewer than
+     * `pipelineSockets` are open, else the one with room and the fewest sources.
+     */
+    private socketForJob;
+    /** A new socket with the last pop set on it; `patient` waits through 503 with Retry-After. */
     private openSocket;
     private startHeartbeat;
     private socketUrl;
@@ -2137,7 +2207,11 @@ declare class PipelineBackend implements WorkerBackend {
     /** Send a command that is answered by `result` or `error` under its id. */
     private command;
     pop(): Pop | null;
-    /** Set the pop on the socket's pipeline; a pop it refuses rejects, and the last pop it took stays. */
+    /**
+     * Set the pop on every open socket's pipeline; a pop the first refuses rejects, and the last pop it
+     * took stays. shortcut: a refusal from a later socket, which the first took, leaves the earlier ones
+     * on the new pop; a refusal is the pop's own fault, so every socket answers it alike.
+     */
     changePop(pop: Pop): Promise<void>;
     private setPopOn;
     submit(work: WorkerWork): Promise<ResultStream>;
@@ -2290,4 +2364,4 @@ 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, 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, 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 };