@qvac/sdk 0.14.1 → 0.16.0

package/dist/schemas/unload-model.js

@@ -1,18 +1,18 @@
-import { z } from "zod";
+import { z } from 'zod';
 export const unloadModelParamsSchema = z.object({
     modelId: z.string(),
     clearStorage: z.boolean().default(false),
-    autoClose: z.boolean().optional(),
+    autoClose: z.boolean().optional()
 });
 export const unloadModelRequestSchema = z.object({
-    type: z.literal("unloadModel"),
+    type: z.literal('unloadModel'),
     modelId: z.string(),
-    clearStorage: z.boolean().default(false),
+    clearStorage: z.boolean().default(false)
 });
 export const unloadModelResponseSchema = z.object({
-    type: z.literal("unloadModel"),
+    type: z.literal('unloadModel'),
     success: z.boolean(),
     error: z.string().optional(),
     hasActiveModels: z.boolean().optional(),
-    hasActiveProviders: z.boolean().optional(),
+    hasActiveProviders: z.boolean().optional()
 });

package/dist/schemas/vla.d.ts

@@ -1,4 +1,4 @@
-import { z } from "zod";
+import { z } from 'zod';
 export declare const vlaConfigSchema: z.ZodObject<{
     backend: z.ZodOptional<z.ZodEnum<{
         cpu: "cpu";
@@ -19,6 +19,11 @@ export declare const vlaHparamsSchema: z.ZodObject<{
         discrete: "discrete";
         continuous: "continuous";
     }>>;
+    imageInputMode: z.ZodOptional<z.ZodEnum<{
+        pixels: "pixels";
+        patches: "patches";
+    }>>;
+    imagePatchElems: z.ZodOptional<z.ZodNumber>;
 }, z.core.$strip>;
 export type VlaHparams = z.infer<typeof vlaHparamsSchema>;
 export declare const vlaStatsSchema: z.ZodObject<{
@@ -78,6 +83,11 @@ export declare const vlaHparamsResponseSchema: z.ZodObject<{
             discrete: "discrete";
             continuous: "continuous";
         }>>;
+        imageInputMode: z.ZodOptional<z.ZodEnum<{
+            pixels: "pixels";
+            patches: "patches";
+        }>>;
+        imagePatchElems: z.ZodOptional<z.ZodNumber>;
     }, z.core.$strip>;
     backendName: z.ZodNullable<z.ZodString>;
 }, z.core.$strip>;

package/dist/schemas/vla.js

@@ -1,20 +1,20 @@
-import { z } from "zod";
+import { z } from 'zod';
 const BASE64_PATTERN = /^(?:[A-Za-z0-9+/]{4})*(?:[A-Za-z0-9+/]{2}==|[A-Za-z0-9+/]{3}=)?$/;
 // ============================================
 // Load-time config
 // ============================================
 export const vlaConfigSchema = z.object({
     backend: z
-        .enum(["auto", "cpu"])
+        .enum(['auto', 'cpu'])
         .optional()
-        .describe("Backend selection passed to `VlaModel.load({ backend })`. " +
+        .describe('Backend selection passed to `VlaModel.load({ backend })`. ' +
         "`'auto'` (default) prefers an accepted GPU (Vulkan / Metal / OpenCL) and falls back to CPU. " +
         "`'cpu'` forces CPU regardless of available accelerators."),
     verbosity: z
         .number()
         .int()
         .optional()
-        .describe("Native log verbosity forwarded to the addon (0=ERROR, 1=WARN, 2=INFO, 3=DEBUG)."),
+        .describe('Native log verbosity forwarded to the addon (0=ERROR, 1=WARN, 2=INFO, 3=DEBUG).')
 });
 // ============================================
 // Hparams (returned by the addon after load)
@@ -31,16 +31,33 @@ export const vlaHparamsSchema = z.object({
         .int()
         .positive()
         .optional()
-        .describe("Number of camera views the model expects (2 for SmolVLA, 3 for π₀.₅). " +
-        "Pass exactly this many preprocessed frames in `images`. Optional for " +
-        "back-compat — older addon builds may omit it."),
+        .describe('Number of camera views the model expects (2 for SmolVLA and GR00T, 3 ' +
+        'for π₀.₅). Pass exactly this many preprocessed frames in `images`. ' +
+        'Optional for back-compat — older addon builds may omit it.'),
     stateInputMode: z
-        .enum(["continuous", "discrete"])
+        .enum(['continuous', 'discrete'])
         .optional()
-        .describe("How the robot state is consumed. `'continuous'` (SmolVLA): the `state` " +
-        "Float32Array is projected by an in-model linear layer. `'discrete'` " +
-        "(π₀.₅): the state is tokenised into the language prompt and the `state` " +
-        "buffer is ignored — pass an empty `Float32Array(0)`. Optional for back-compat."),
+        .describe("How the robot state is consumed. `'continuous'` (SmolVLA, GR00T): the " +
+        '`state` Float32Array is projected by an in-model linear layer. ' +
+        "`'discrete'` (π₀.₅): the state is tokenised into the language prompt " +
+        'and the `state` buffer is ignored — pass an empty `Float32Array(0)`. ' +
+        'Optional for back-compat.'),
+    imageInputMode: z
+        .enum(['pixels', 'patches'])
+        .optional()
+        .describe("How images are supplied. `'pixels'` (SmolVLA, π₀.₅): each `images` " +
+        'entry is a `3 · imgWidth · imgHeight` CHW plane from ' +
+        "`vlaPreprocessImage`. `'patches'` (GR00T): each entry is a " +
+        'pre-patchified buffer of length `imagePatchElems`. Optional for ' +
+        'back-compat.'),
+    imagePatchElems: z
+        .number()
+        .int()
+        .nonnegative()
+        .optional()
+        .describe("For patch-input models (`imageInputMode === 'patches'`), the exact " +
+        'per-camera patch buffer length each `images` entry must have. Absent ' +
+        'for pixel-input models.')
 });
 // ============================================
 // Stats
@@ -59,7 +76,7 @@ export const vlaStatsSchema = z.object({
     backendDevice: z
         .number()
         .optional()
-        .describe("0 = CPU backend, 1 = GPU backend (Vulkan / Metal / OpenCL)."),
+        .describe('0 = CPU backend, 1 = GPU backend (Vulkan / Metal / OpenCL).')
 });
 // ============================================
 // Run request / response (wire format)
@@ -70,62 +87,65 @@ export const vlaStatsSchema = z.object({
 // and handle the encoding internally.
 // ============================================
 export const vlaRunRequestSchema = z.object({
-    type: z.literal("vlaRun"),
+    type: z.literal('vlaRun'),
     modelId: z.string(),
     images: z
         .array(z.string().min(1).regex(BASE64_PATTERN))
         .min(1)
-        .describe("Base64-encoded preprocessed images. Each entry is the underlying " +
-        "ArrayBuffer of a `Float32Array` produced by `vlaPreprocessImage(...)`. " +
-        "Length per image must equal `3 * imgWidth * imgHeight`."),
+        .describe('Base64-encoded preprocessed images, one per camera. Each entry is the ' +
+        'underlying ArrayBuffer of a `Float32Array`. For pixel-input models ' +
+        '(SmolVLA, π₀.₅) it comes from `vlaPreprocessImage(...)` and its length ' +
+        'must equal `3 * imgWidth * imgHeight`. For patch-input models (GR00T, ' +
+        "`hparams.imageInputMode === 'patches'`) it is a pre-patchified buffer " +
+        'of length `hparams.imagePatchElems`.'),
     imgWidth: z.number().int().positive(),
     imgHeight: z.number().int().positive(),
     state: z
         .string()
         .regex(BASE64_PATTERN)
-        .describe("Base64-encoded `Float32Array`. For continuous-state models (SmolVLA) " +
-        "this is length `hparams.maxStateDim` — use " +
-        "`vlaPadState(state, hparams.maxStateDim)` to zero-pad. For " +
+        .describe('Base64-encoded `Float32Array`. For continuous-state models (SmolVLA, GR00T) ' +
+        'this is length `hparams.maxStateDim` — use ' +
+        '`vlaPadState(state, hparams.maxStateDim)` to zero-pad. For ' +
         "discrete-state models (π₀.₅, `stateInputMode: 'discrete'`) the state " +
-        "is tokenised into the prompt and this buffer is ignored — encode an " +
-        "empty `Float32Array(0)` (which is the empty string)."),
+        'is tokenised into the prompt and this buffer is ignored — encode an ' +
+        'empty `Float32Array(0)` (which is the empty string).'),
     tokens: z
         .string()
         .min(1)
         .regex(BASE64_PATTERN)
-        .describe("Base64-encoded `Int32Array` of length `hparams.tokenizerMaxLength`."),
+        .describe('Base64-encoded `Int32Array` of length `hparams.tokenizerMaxLength`.'),
     mask: z
         .string()
         .min(1)
         .regex(BASE64_PATTERN)
-        .describe("Base64-encoded `Uint8Array` of length `hparams.tokenizerMaxLength`."),
+        .describe('Base64-encoded `Uint8Array` of length `hparams.tokenizerMaxLength`.'),
     noise: z
         .string()
         .min(1)
         .regex(BASE64_PATTERN)
         .optional()
-        .describe("Optional base64-encoded `Float32Array` of length " +
-        "`hparams.chunkSize * hparams.maxActionDim`. When omitted the addon " +
-        "samples its own prior."),
+        .describe('Optional base64-encoded `Float32Array` of length ' +
+        '`hparams.chunkSize * hparams.maxActionDim`. When omitted the addon ' +
+        'samples its own prior.')
 });
 export const vlaRunResponseSchema = z.object({
     actions: z
         .string()
         .min(1)
         .regex(BASE64_PATTERN)
-        .describe("Base64-encoded `Float32Array` of length `hparams.chunkSize * hparams.actionDim`."),
+        .describe('Base64-encoded `Float32Array` of length `hparams.chunkSize * hparams.actionDim`.'),
     actionDim: z.number().int().positive(),
     chunkSize: z.number().int().positive(),
-    stats: vlaStatsSchema.optional(),
+    stats: vlaStatsSchema.optional()
 });
 // ============================================
 // Hparams request / response (plugin handler)
 // ============================================
 export const vlaHparamsRequestSchema = z.object({
-    type: z.literal("vlaHparams"),
-    modelId: z.string(),
+    type: z.literal('vlaHparams'),
+    modelId: z.string()
 });
 export const vlaHparamsResponseSchema = z.object({
     hparams: vlaHparamsSchema,
-    backendName: z.string().nullable(),
+    backendName: z.string().nullable()
 });

package/dist/scripts/contract/build-contract.d.ts

@@ -0,0 +1,38 @@
+import { z } from 'zod';
+export declare const contractDir: URL;
+export declare const callShapeByHandlerType: {
+    readonly reply: "request-reply";
+    readonly stream: "server-stream";
+    readonly duplex: "duplex";
+};
+type JsonSchema = Record<string, unknown>;
+export declare function toWireJsonSchema(schema: z.ZodType, io: 'input' | 'output', defName: string): JsonSchema;
+export declare function buildContract(): {
+    schemaDocument: {
+        $schema: string;
+        title: string;
+        description: string;
+        $defs: Record<string, JsonSchema>;
+    };
+    manifest: {
+        methods: {
+            progress?: {
+                condition: string;
+                responseSchema: string;
+            };
+            name: "cancel" | "transcribe" | "translate" | "finetune" | "loadModel" | "downloadAsset" | "rag" | "pluginInvoke" | "pluginInvokeStream" | "modelRegistryList" | "modelRegistrySearch" | "modelRegistryGetModel" | "completionStream" | "completionOrchestrate" | "batchCompletionStream" | "deleteCache" | "heartbeat" | "unloadModel" | "transcribeStream" | "bciTranscribe" | "bciTranscribeStream" | "loggingStream" | "embed" | "textToSpeech" | "textToSpeechStream" | "provide" | "stopProvide" | "getModelInfo" | "getLoadedModelInfo" | "ocrStream" | "diffusionStream" | "videoStream" | "upscaleStream" | "resume" | "suspend" | "state" | "classify";
+            callShape: "duplex" | "request-reply" | "server-stream";
+            requestSchema: string;
+            responseSchema: string;
+        }[];
+    };
+};
+export declare function renderContractFiles(): Promise<{
+    'schema.json': string;
+    'manifest.json': string;
+    'models.json': string;
+    'model-type-maps.json': string;
+    'error-codes.json': string;
+}>;
+export {};
+//# sourceMappingURL=build-contract.d.ts.map