@qvac/sdk 0.10.2 → 0.12.0

package/dist/schemas/error.js

@@ -1,5 +1,21 @@
 import { z } from "zod";
 import { QvacErrorBase } from "@qvac/error";
+/**
+ * Wire shape for errors thrown across the RPC boundary. The fields are
+ * the union of (a) the legacy `QvacErrorBase` serialisation (`name`,
+ * `code`, `message`, `stack`, `cause`, `timestamp`) and (b) the new
+ * `typedFields` map (0.11.0+) carrying per-class structured data the
+ * client-side reconstructor uses to rebuild the original typed error.
+ *
+ * `typedFields` is opaque on the wire — `z.unknown()` — and the
+ * per-class reconstructor in `client/rpc/rpc-error.ts` casts each
+ * member at the boundary. The single-map shape keeps the schema
+ * compact regardless of how many typed-error classes the SDK
+ * eventually surfaces across RPC. New typed-error classes that need
+ * cross-RPC reconstruction add a `toErrorResponseFields()` method on
+ * the server side and a row to the reconstructor map on the client
+ * side; the schema itself doesn't change.
+ */
 export const errorResponseSchema = z.object({
     type: z.literal("error"),
     message: z.string(),
@@ -8,14 +24,22 @@ export const errorResponseSchema = z.object({
     name: z.string().optional(),
     code: z.number().optional(),
     cause: z.unknown().optional(),
+    typedFields: z.record(z.string(), z.unknown()).optional(),
 });
+function hasTypedFields(error) {
+    return (error !== null &&
+        typeof error === "object" &&
+        "toErrorResponseFields" in error &&
+        typeof error
+            .toErrorResponseFields === "function");
+}
 function isQvacError(error) {
     return error instanceof QvacErrorBase;
 }
 export function createErrorResponse(error) {
     if (isQvacError(error)) {
         const qvacData = error.toJSON();
-        return {
+        const response = {
             type: "error",
             name: qvacData.name,
             code: qvacData.code,
@@ -23,6 +47,10 @@ export function createErrorResponse(error) {
             stack: qvacData.stack,
             timestamp: new Date().toISOString(),
         };
+        if (hasTypedFields(error)) {
+            response.typedFields = error.toErrorResponseFields();
+        }
+        return response;
     }
     const message = error instanceof Error ? error.message : String(error);
     const stack = error instanceof Error ? error.stack : undefined;

package/dist/schemas/finetune.d.ts

@@ -336,6 +336,7 @@ export declare const finetuneRunRequestSchema: z.ZodObject<{
         resume: "resume";
     }>>;
     withProgress: z.ZodOptional<z.ZodBoolean>;
+    requestId: z.ZodOptional<z.ZodString>;
 }, z.core.$strict>;
 export declare const finetuneGetStateRequestSchema: z.ZodObject<{
     modelId: z.ZodString;
@@ -466,6 +467,7 @@ export declare const finetuneRequestSchema: z.ZodUnion<readonly [z.ZodObject<{
         resume: "resume";
     }>>;
     withProgress: z.ZodOptional<z.ZodBoolean>;
+    requestId: z.ZodOptional<z.ZodString>;
 }, z.core.$strict>, z.ZodObject<{
     modelId: z.ZodString;
     options: z.ZodObject<{

package/dist/schemas/finetune.js

@@ -96,6 +96,11 @@ export const finetuneRunRequestSchema = finetuneRunParamsSchemaBase
     type: z.literal("finetune"),
     operation: z.enum(["start", "resume"]).optional(),
     withProgress: z.boolean().optional(),
+    requestId: z
+        .string()
+        .min(1)
+        .optional()
+        .describe("Stable identifier for this in-flight finetune, generated by the client at call time. Surfaces on the registry so callers can target it with `cancel({ requestId })`. Optional on the wire — the server falls back to a server-generated id when the field is missing."),
 })
     .strict();
 export const finetuneGetStateRequestSchema = finetuneRunParamsSchemaBase

package/dist/schemas/get-model-info.d.ts

@@ -36,13 +36,15 @@ export declare const modelInfoSchema: z.ZodObject<{
     sha256Checksum: z.ZodString;
     addon: z.ZodEnum<{
         embeddings: "embeddings";
+        diffusion: "diffusion";
+        tts: "tts";
+        ocr: "ocr";
+        vla: "vla";
         llm: "llm";
         whisper: "whisper";
         nmt: "nmt";
         parakeet: "parakeet";
-        tts: "tts";
-        ocr: "ocr";
-        diffusion: "diffusion";
+        classification: "classification";
         vad: "vad";
         other: "other";
     }>;
@@ -83,13 +85,15 @@ export declare const getModelInfoResponseSchema: z.ZodObject<{
         sha256Checksum: z.ZodString;
         addon: z.ZodEnum<{
             embeddings: "embeddings";
+            diffusion: "diffusion";
+            tts: "tts";
+            ocr: "ocr";
+            vla: "vla";
             llm: "llm";
             whisper: "whisper";
             nmt: "nmt";
             parakeet: "parakeet";
-            tts: "tts";
-            ocr: "ocr";
-            diffusion: "diffusion";
+            classification: "classification";
             vad: "vad";
             other: "other";
         }>;

package/dist/schemas/get-model-info.js

@@ -99,6 +99,8 @@ export const modelInfoSchema = z.object({
         "tts",
         "ocr",
         "diffusion",
+        "vla",
+        "classification",
         "other",
     ])
         .describe("Inference addon / capability category this model belongs to."),

package/dist/schemas/index.d.ts

@@ -26,6 +26,8 @@ export * from "./error";
 export * from "./rag";
 export * from "./ocr";
 export * from "./sdcpp-config";
+export * from "./vla";
+export * from "./classification";
 export * from "./shard";
 export * from "./suspend";
 export * from "./resume";
@@ -33,6 +35,7 @@ export * from "./state";
 export { SDK_CLIENT_ERROR_CODES } from "./sdk-errors-client";
 export { SDK_SERVER_ERROR_CODES } from "./sdk-errors-server";
 export { REGISTRY_ERROR_CODES } from "./sdk-errors-registry";
+export { ERR_CODES as RAG_ERROR_CODES } from "@qvac/rag/errors";
 export { qvacConfigSchema, deviceMatchSchema, deviceConfigDefaultsSchema, devicePatternSchema, type QvacConfig, type DeviceMatch, type DeviceConfigDefaults, type DevicePattern, } from "./sdk-config";
 export { PROFILING_KEY, PROFILING_TRAILER_KEY, DELEGATION_BREAKDOWN_KEY, OPERATION_EVENT_KEY, MODEL_EXECUTION_KEY, profilerModeSchema, serverBreakdownSchema, delegationBreakdownSchema, operationEventSchema, profilingRequestMetaSchema, profilingResponseMetaSchema, perCallProfilingSchema, type ProfilerMode, type ServerBreakdown, type DelegationBreakdown, type OperationEvent, type ProfilingRequestMeta, type ProfilingResponseMeta, type PerCallProfiling, } from "./profiling";
 export { runtimeContextSchema, type RuntimeContext } from "./runtime-context";

package/dist/schemas/index.js

@@ -27,6 +27,8 @@ export * from "./error.js";
 export * from "./rag.js";
 export * from "./ocr.js";
 export * from "./sdcpp-config.js";
+export * from "./vla.js";
+export * from "./classification.js";
 export * from "./shard.js";
 export * from "./suspend.js";
 export * from "./resume.js";
@@ -34,6 +36,7 @@ export * from "./state.js";
 export { SDK_CLIENT_ERROR_CODES } from "./sdk-errors-client.js";
 export { SDK_SERVER_ERROR_CODES } from "./sdk-errors-server.js";
 export { REGISTRY_ERROR_CODES } from "./sdk-errors-registry.js";
+export { ERR_CODES as RAG_ERROR_CODES } from "@qvac/rag/errors.js";
 export { qvacConfigSchema, deviceMatchSchema, deviceConfigDefaultsSchema, devicePatternSchema, } from "./sdk-config.js";
 export { PROFILING_KEY, PROFILING_TRAILER_KEY, DELEGATION_BREAKDOWN_KEY, OPERATION_EVENT_KEY, MODEL_EXECUTION_KEY, profilerModeSchema, serverBreakdownSchema, delegationBreakdownSchema, operationEventSchema, profilingRequestMetaSchema, profilingResponseMetaSchema, perCallProfilingSchema, } from "./profiling.js";
 export { runtimeContextSchema } from "./runtime-context.js";

package/dist/schemas/llamacpp-config.d.ts

@@ -30,7 +30,18 @@ export declare const llmConfigBaseSchema: z.ZodObject<{
     }>>;
     "cache-type-k": z.ZodOptional<z.ZodString>;
     "cache-type-v": z.ZodOptional<z.ZodString>;
+    "main-gpu": z.ZodOptional<z.ZodUnion<readonly [z.ZodNumber, z.ZodEnum<{
+        integrated: "integrated";
+        dedicated: "dedicated";
+    }>]>>;
+    "split-mode": z.ZodOptional<z.ZodEnum<{
+        none: "none";
+        layer: "layer";
+        row: "row";
+    }>>;
+    "tensor-split": z.ZodOptional<z.ZodString>;
     openclCacheDir: z.ZodOptional<z.ZodString>;
+    reasoning_budget: z.ZodOptional<z.ZodUnion<readonly [z.ZodLiteral<-1>, z.ZodLiteral<0>]>>;
     projectionModelSrc: z.ZodOptional<z.ZodUnion<readonly [z.ZodString, z.ZodObject<{
         src: z.ZodString;
         name: z.ZodOptional<z.ZodString>;
@@ -44,21 +55,26 @@ export declare const llmConfigBaseSchema: z.ZodObject<{
         sha256Checksum: z.ZodOptional<z.ZodString>;
         addon: z.ZodOptional<z.ZodUnion<readonly [z.ZodEnum<{
             embeddings: "embeddings";
+            diffusion: "diffusion";
+            tts: "tts";
+            ocr: "ocr";
+            vla: "vla";
             "llamacpp-completion": "llamacpp-completion";
             "whispercpp-transcription": "whispercpp-transcription";
             "llamacpp-embedding": "llamacpp-embedding";
             "nmtcpp-translation": "nmtcpp-translation";
             "onnx-tts": "onnx-tts";
+            "tts-ggml": "tts-ggml";
             "parakeet-transcription": "parakeet-transcription";
             "onnx-ocr": "onnx-ocr";
             "sdcpp-generation": "sdcpp-generation";
+            "ggml-vla": "ggml-vla";
+            "ggml-classification": "ggml-classification";
             llm: "llm";
             whisper: "whisper";
             nmt: "nmt";
             parakeet: "parakeet";
-            tts: "tts";
-            ocr: "ocr";
-            diffusion: "diffusion";
+            classification: "classification";
         }>, z.ZodLiteral<"vad">]>>;
     }, z.core.$strip>]>>;
 }, z.core.$strip>;
@@ -94,7 +110,18 @@ export declare const llmConfigSchema: z.ZodPipe<z.ZodObject<{
     }>>;
     "cache-type-k": z.ZodOptional<z.ZodString>;
     "cache-type-v": z.ZodOptional<z.ZodString>;
+    "main-gpu": z.ZodOptional<z.ZodUnion<readonly [z.ZodNumber, z.ZodEnum<{
+        integrated: "integrated";
+        dedicated: "dedicated";
+    }>]>>;
+    "split-mode": z.ZodOptional<z.ZodEnum<{
+        none: "none";
+        layer: "layer";
+        row: "row";
+    }>>;
+    "tensor-split": z.ZodOptional<z.ZodString>;
     openclCacheDir: z.ZodOptional<z.ZodString>;
+    reasoning_budget: z.ZodOptional<z.ZodUnion<readonly [z.ZodLiteral<-1>, z.ZodLiteral<0>]>>;
     projectionModelSrc: z.ZodOptional<z.ZodUnion<readonly [z.ZodString, z.ZodObject<{
         src: z.ZodString;
         name: z.ZodOptional<z.ZodString>;
@@ -108,21 +135,26 @@ export declare const llmConfigSchema: z.ZodPipe<z.ZodObject<{
         sha256Checksum: z.ZodOptional<z.ZodString>;
         addon: z.ZodOptional<z.ZodUnion<readonly [z.ZodEnum<{
             embeddings: "embeddings";
+            diffusion: "diffusion";
+            tts: "tts";
+            ocr: "ocr";
+            vla: "vla";
             "llamacpp-completion": "llamacpp-completion";
             "whispercpp-transcription": "whispercpp-transcription";
             "llamacpp-embedding": "llamacpp-embedding";
             "nmtcpp-translation": "nmtcpp-translation";
             "onnx-tts": "onnx-tts";
+            "tts-ggml": "tts-ggml";
             "parakeet-transcription": "parakeet-transcription";
             "onnx-ocr": "onnx-ocr";
             "sdcpp-generation": "sdcpp-generation";
+            "ggml-vla": "ggml-vla";
+            "ggml-classification": "ggml-classification";
             llm: "llm";
             whisper: "whisper";
             nmt: "nmt";
             parakeet: "parakeet";
-            tts: "tts";
-            ocr: "ocr";
-            diffusion: "diffusion";
+            classification: "classification";
         }>, z.ZodLiteral<"vad">]>>;
     }, z.core.$strip>]>>;
 }, z.core.$strip>, z.ZodTransform<{
@@ -147,7 +179,11 @@ export declare const llmConfigSchema: z.ZodPipe<z.ZodObject<{
     toolsMode?: "static" | "dynamic" | undefined;
     "cache-type-k"?: string | undefined;
     "cache-type-v"?: string | undefined;
+    "main-gpu"?: number | "integrated" | "dedicated" | undefined;
+    "split-mode"?: "none" | "layer" | "row" | undefined;
+    "tensor-split"?: string | undefined;
     openclCacheDir?: string | undefined;
+    reasoning_budget?: 0 | -1 | undefined;
     projectionModelSrc?: string | {
         src: string;
         name?: string | undefined;
@@ -159,7 +195,7 @@ export declare const llmConfigSchema: z.ZodPipe<z.ZodObject<{
         engine?: string | undefined;
         expectedSize?: number | undefined;
         sha256Checksum?: string | undefined;
-        addon?: "embeddings" | "llamacpp-completion" | "whispercpp-transcription" | "llamacpp-embedding" | "nmtcpp-translation" | "onnx-tts" | "parakeet-transcription" | "onnx-ocr" | "sdcpp-generation" | "llm" | "whisper" | "nmt" | "parakeet" | "tts" | "ocr" | "diffusion" | "vad" | undefined;
+        addon?: "embeddings" | "diffusion" | "tts" | "ocr" | "vla" | "llamacpp-completion" | "whispercpp-transcription" | "llamacpp-embedding" | "nmtcpp-translation" | "onnx-tts" | "tts-ggml" | "parakeet-transcription" | "onnx-ocr" | "sdcpp-generation" | "ggml-vla" | "ggml-classification" | "llm" | "whisper" | "nmt" | "parakeet" | "classification" | "vad" | undefined;
     } | undefined;
 }, {
     ctx_size?: number | undefined;
@@ -183,7 +219,11 @@ export declare const llmConfigSchema: z.ZodPipe<z.ZodObject<{
     toolsMode?: "static" | "dynamic" | undefined;
     "cache-type-k"?: string | undefined;
     "cache-type-v"?: string | undefined;
+    "main-gpu"?: number | "integrated" | "dedicated" | undefined;
+    "split-mode"?: "none" | "layer" | "row" | undefined;
+    "tensor-split"?: string | undefined;
     openclCacheDir?: string | undefined;
+    reasoning_budget?: 0 | -1 | undefined;
     projectionModelSrc?: string | {
         src: string;
         name?: string | undefined;
@@ -195,7 +235,7 @@ export declare const llmConfigSchema: z.ZodPipe<z.ZodObject<{
         engine?: string | undefined;
         expectedSize?: number | undefined;
         sha256Checksum?: string | undefined;
-        addon?: "embeddings" | "llamacpp-completion" | "whispercpp-transcription" | "llamacpp-embedding" | "nmtcpp-translation" | "onnx-tts" | "parakeet-transcription" | "onnx-ocr" | "sdcpp-generation" | "llm" | "whisper" | "nmt" | "parakeet" | "tts" | "ocr" | "diffusion" | "vad" | undefined;
+        addon?: "embeddings" | "diffusion" | "tts" | "ocr" | "vla" | "llamacpp-completion" | "whispercpp-transcription" | "llamacpp-embedding" | "nmtcpp-translation" | "onnx-tts" | "tts-ggml" | "parakeet-transcription" | "onnx-ocr" | "sdcpp-generation" | "ggml-vla" | "ggml-classification" | "llm" | "whisper" | "nmt" | "parakeet" | "classification" | "vad" | undefined;
     } | undefined;
 }>>;
 export type LlmConfig = z.infer<typeof llmConfigSchema>;
@@ -227,6 +267,12 @@ export declare const embedConfigBaseSchema: z.ZodObject<{
         integrated: "integrated";
         dedicated: "dedicated";
     }>]>>;
+    splitMode: z.ZodOptional<z.ZodEnum<{
+        none: "none";
+        layer: "layer";
+        row: "row";
+    }>>;
+    tensorSplit: z.ZodOptional<z.ZodString>;
     verbosity: z.ZodOptional<z.ZodUnion<readonly [z.ZodLiteral<0>, z.ZodLiteral<1>, z.ZodLiteral<2>, z.ZodLiteral<3>]>>;
     openclCacheDir: z.ZodOptional<z.ZodString>;
 }, z.core.$strip>;
@@ -264,6 +310,12 @@ export declare const embedConfigSchema: z.ZodPipe<z.ZodObject<{
         integrated: "integrated";
         dedicated: "dedicated";
     }>]>>;
+    splitMode: z.ZodOptional<z.ZodEnum<{
+        none: "none";
+        layer: "layer";
+        row: "row";
+    }>>;
+    tensorSplit: z.ZodOptional<z.ZodString>;
     verbosity: z.ZodOptional<z.ZodUnion<readonly [z.ZodLiteral<0>, z.ZodLiteral<1>, z.ZodLiteral<2>, z.ZodLiteral<3>]>>;
     openclCacheDir: z.ZodOptional<z.ZodString>;
 }, z.core.$strip>, z.ZodTransform<{
@@ -275,6 +327,8 @@ export declare const embedConfigSchema: z.ZodPipe<z.ZodObject<{
     embdNormalize?: number | undefined;
     flashAttention?: "on" | "off" | "auto" | undefined;
     mainGpu?: number | "integrated" | "dedicated" | undefined;
+    splitMode?: "none" | "layer" | "row" | undefined;
+    tensorSplit?: string | undefined;
     verbosity?: 0 | 3 | 2 | 1 | undefined;
     openclCacheDir?: string | undefined;
 }, {
@@ -286,6 +340,8 @@ export declare const embedConfigSchema: z.ZodPipe<z.ZodObject<{
     embdNormalize?: number | undefined;
     flashAttention?: "on" | "off" | "auto" | undefined;
     mainGpu?: number | "integrated" | "dedicated" | undefined;
+    splitMode?: "none" | "layer" | "row" | undefined;
+    tensorSplit?: string | undefined;
     verbosity?: 0 | 3 | 2 | 1 | undefined;
     openclCacheDir?: string | undefined;
 }>>;

package/dist/schemas/llamacpp-config.js

@@ -30,6 +30,7 @@ export const llmConfigBaseSchema = z.object({
         z.number().int().min(1), // positive integer: fixed token count
     ])
         .optional(),
+    /** JS-side only: seeds conversation history. Never forwarded to the C++ addon. */
     system_prompt: z.string().optional(),
     no_mmap: z.boolean().optional(),
     verbosity: verbositySchema.optional(),
@@ -45,11 +46,20 @@ export const llmConfigBaseSchema = z.object({
         .optional(),
     "cache-type-k": z.string().optional(),
     "cache-type-v": z.string().optional(),
+    "main-gpu": z
+        .union([z.number().int().min(0), z.enum(["integrated", "dedicated"])])
+        .optional(),
+    "split-mode": z.enum(["none", "layer", "row"]).optional(),
+    "tensor-split": z.string().optional(),
     /**
      * Writable directory for OpenCL kernel binary cache. Required on Android
      * for fast GPU startup.
      */
     openclCacheDir: z.string().optional(),
+    /**
+     * Reasoning channel token budget. `-1` = unrestricted, `0` = disabled.
+     */
+    reasoning_budget: z.union([z.literal(-1), z.literal(0)]).optional(),
     projectionModelSrc: modelSrcInputSchema.optional(),
 });
 // Default values - typed as partial of the config
@@ -76,6 +86,8 @@ export const embedConfigBaseSchema = z.object({
     mainGpu: z
         .union([z.number().int().min(0), z.enum(["integrated", "dedicated"])])
         .optional(),
+    splitMode: z.enum(["none", "layer", "row"]).optional(),
+    tensorSplit: z.string().optional(),
     verbosity: verbositySchema.optional(),
     /**
      * Writable directory for OpenCL kernel binary cache. Required on Android