@qvac/sdk 0.11.0 → 0.12.1

package/dist/schemas/common.js

@@ -19,13 +19,14 @@ import { deleteCacheRequestSchema, deleteCacheResponseSchema, } from "./delete-c
 import { getModelInfoRequestSchema, getModelInfoResponseSchema, } from "./get-model-info.js";
 import { getLoadedModelInfoRequestSchema, getLoadedModelInfoResponseSchema, } from "./get-loaded-model-info.js";
 import { ocrStreamRequestSchema, ocrStreamResponseSchema } from "./ocr.js";
-import { diffusionStreamRequestSchema, diffusionStreamResponseSchema, upscaleStreamRequestSchema, upscaleStreamResponseSchema, } from "./sdcpp-config.js";
+import { diffusionStreamRequestSchema, diffusionStreamResponseSchema, videoStreamRequestSchema, videoStreamResponseSchema, upscaleStreamRequestSchema, upscaleStreamResponseSchema, } from "./sdcpp-config.js";
 import { finetuneRequestSchema, finetuneResponseSchema, finetuneProgressResponseSchema, } from "./finetune.js";
 import { pluginInvokeRequestSchema, pluginInvokeResponseSchema, pluginInvokeStreamRequestSchema, pluginInvokeStreamResponseSchema, } from "./plugin.js";
 import { modelRegistryListRequestSchema, modelRegistryListResponseSchema, modelRegistrySearchRequestSchema, modelRegistrySearchResponseSchema, modelRegistryGetModelRequestSchema, modelRegistryGetModelResponseSchema, } from "./registry.js";
 import { suspendRequestSchema, suspendResponseSchema } from "./suspend.js";
 import { resumeRequestSchema, resumeResponseSchema } from "./resume.js";
 import { stateRequestSchema, stateResponseSchema } from "./state.js";
+import { classifyRequestSchema, classifyResponseSchema, } from "./classification.js";
 export const requestSchema = z.union([
     heartbeatRequestSchema,
     loadModelRequestSchema,
@@ -48,6 +49,7 @@ export const requestSchema = z.union([
     getLoadedModelInfoRequestSchema,
     ocrStreamRequestSchema,
     diffusionStreamRequestSchema,
+    videoStreamRequestSchema,
     upscaleStreamRequestSchema,
     finetuneRequestSchema,
     pluginInvokeRequestSchema,
@@ -58,6 +60,7 @@ export const requestSchema = z.union([
     suspendRequestSchema,
     resumeRequestSchema,
     stateRequestSchema,
+    classifyRequestSchema,
 ]);
 export const responseSchema = z.discriminatedUnion("type", [
     heartbeatResponseSchema,
@@ -84,6 +87,7 @@ export const responseSchema = z.discriminatedUnion("type", [
     getLoadedModelInfoResponseSchema,
     ocrStreamResponseSchema,
     diffusionStreamResponseSchema,
+    videoStreamResponseSchema,
     upscaleStreamResponseSchema,
     finetuneResponseSchema,
     finetuneProgressResponseSchema,
@@ -95,6 +99,7 @@ export const responseSchema = z.discriminatedUnion("type", [
     suspendResponseSchema,
     resumeResponseSchema,
     stateResponseSchema,
+    classifyResponseSchema,
 ]);
 export const rpcOptionsSchema = z.object({
     timeout: z

package/dist/schemas/completion-event.d.ts

@@ -5,6 +5,7 @@ export declare const completionStatsSchema: z.ZodObject<{
     timeToFirstToken: z.ZodOptional<z.ZodNumber>;
     tokensPerSecond: z.ZodOptional<z.ZodNumber>;
     cacheTokens: z.ZodOptional<z.ZodNumber>;
+    promptTokens: z.ZodOptional<z.ZodNumber>;
     generatedTokens: z.ZodOptional<z.ZodNumber>;
     backendDevice: z.ZodOptional<z.ZodEnum<{
         cpu: "cpu";
@@ -58,6 +59,7 @@ export declare const statsEventSchema: z.ZodObject<{
         timeToFirstToken: z.ZodOptional<z.ZodNumber>;
         tokensPerSecond: z.ZodOptional<z.ZodNumber>;
         cacheTokens: z.ZodOptional<z.ZodNumber>;
+        promptTokens: z.ZodOptional<z.ZodNumber>;
         generatedTokens: z.ZodOptional<z.ZodNumber>;
         backendDevice: z.ZodOptional<z.ZodEnum<{
             cpu: "cpu";
@@ -137,6 +139,7 @@ export declare const completionEventSchema: z.ZodUnion<readonly [z.ZodObject<{
         timeToFirstToken: z.ZodOptional<z.ZodNumber>;
         tokensPerSecond: z.ZodOptional<z.ZodNumber>;
         cacheTokens: z.ZodOptional<z.ZodNumber>;
+        promptTokens: z.ZodOptional<z.ZodNumber>;
         generatedTokens: z.ZodOptional<z.ZodNumber>;
         backendDevice: z.ZodOptional<z.ZodEnum<{
             cpu: "cpu";

package/dist/schemas/completion-event.js

@@ -4,6 +4,7 @@ export const completionStatsSchema = z.object({
     timeToFirstToken: z.number().optional(),
     tokensPerSecond: z.number().optional(),
     cacheTokens: z.number().optional(),
+    promptTokens: z.number().optional(),
     generatedTokens: z.number().optional(),
     backendDevice: z.enum(["cpu", "gpu"]).optional(),
 });

package/dist/schemas/completion-stream.d.ts

@@ -238,6 +238,7 @@ export declare const completionStreamResponseSchema: z.ZodObject<{
             timeToFirstToken: z.ZodOptional<z.ZodNumber>;
             tokensPerSecond: z.ZodOptional<z.ZodNumber>;
             cacheTokens: z.ZodOptional<z.ZodNumber>;
+            promptTokens: z.ZodOptional<z.ZodNumber>;
             generatedTokens: z.ZodOptional<z.ZodNumber>;
             backendDevice: z.ZodOptional<z.ZodEnum<{
                 cpu: "cpu";

package/dist/schemas/download-asset.d.ts

@@ -17,18 +17,23 @@ export declare const downloadAssetOptionsSchema: z.ZodPipe<z.ZodObject<{
             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";
+            classification: "classification";
         }>, z.ZodLiteral<"vad">]>>;
     }, z.core.$strip>]>;
     seed: z.ZodOptional<z.ZodBoolean>;
@@ -45,7 +50,7 @@ export declare const downloadAssetOptionsSchema: z.ZodPipe<z.ZodObject<{
         engine?: string | undefined;
         expectedSize?: number | undefined;
         sha256Checksum?: string | undefined;
-        addon?: "embeddings" | "diffusion" | "tts" | "ocr" | "llamacpp-completion" | "whispercpp-transcription" | "llamacpp-embedding" | "nmtcpp-translation" | "onnx-tts" | "parakeet-transcription" | "onnx-ocr" | "sdcpp-generation" | "llm" | "whisper" | "nmt" | "parakeet" | "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;
     };
 }, {
     assetSrc: string | {
@@ -59,7 +64,7 @@ export declare const downloadAssetOptionsSchema: z.ZodPipe<z.ZodObject<{
         engine?: string | undefined;
         expectedSize?: number | undefined;
         sha256Checksum?: string | undefined;
-        addon?: "embeddings" | "diffusion" | "tts" | "ocr" | "llamacpp-completion" | "whispercpp-transcription" | "llamacpp-embedding" | "nmtcpp-translation" | "onnx-tts" | "parakeet-transcription" | "onnx-ocr" | "sdcpp-generation" | "llm" | "whisper" | "nmt" | "parakeet" | "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;
     };
     seed?: boolean | undefined;
 }>>;
@@ -80,18 +85,23 @@ export declare const downloadAssetOptionsToRequestSchema: z.ZodPipe<z.ZodObject<
             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";
+            classification: "classification";
         }>, z.ZodLiteral<"vad">]>>;
     }, z.core.$strip>]>;
     seed: z.ZodOptional<z.ZodBoolean>;
@@ -116,7 +126,7 @@ export declare const downloadAssetOptionsToRequestSchema: z.ZodPipe<z.ZodObject<
         engine?: string | undefined;
         expectedSize?: number | undefined;
         sha256Checksum?: string | undefined;
-        addon?: "embeddings" | "diffusion" | "tts" | "ocr" | "llamacpp-completion" | "whispercpp-transcription" | "llamacpp-embedding" | "nmtcpp-translation" | "onnx-tts" | "parakeet-transcription" | "onnx-ocr" | "sdcpp-generation" | "llm" | "whisper" | "nmt" | "parakeet" | "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;
     };
     seed?: boolean | undefined;
     onProgress?: unknown;

package/dist/schemas/engine-addon-map.d.ts

@@ -5,9 +5,12 @@ export declare const ENGINE_TO_ADDON: {
     readonly "llamacpp-embedding": "embeddings";
     readonly "nmtcpp-translation": "nmt";
     readonly "onnx-tts": "tts";
+    readonly "tts-ggml": "tts";
     readonly "onnx-ocr": "ocr";
     readonly "parakeet-transcription": "parakeet";
     readonly "sdcpp-generation": "diffusion";
+    readonly "ggml-vla": "vla";
+    readonly "ggml-classification": "classification";
     readonly "onnx-vad": "vad";
 };
 export declare function resolveCanonicalEngine(engine: string): ModelRegistryEngine | null;

package/dist/schemas/engine-addon-map.js

@@ -1,5 +1,5 @@
 import { ModelType } from "./model-types.js";
-import { ADDON_DIFFUSION, ADDON_EMBEDDING, ADDON_LLM, ADDON_NMT, ADDON_OCR, ADDON_PARAKEET, ADDON_TTS, ADDON_WHISPER, } from "./plugin.js";
+import { ADDON_CLASSIFICATION, ADDON_DIFFUSION, ADDON_EMBEDDING, ADDON_LLM, ADDON_NMT, ADDON_OCR, ADDON_PARAKEET, ADDON_TTS, ADDON_VLA, ADDON_WHISPER, } from "./plugin.js";
 import { modelRegistryEngineSchema, } from "./registry.js";
 // Canonical engine → addon mapping (exhaustive). `as const` preserves
 // per-key literals so the addon can be derived from the engine at the type level.
@@ -9,9 +9,12 @@ export const ENGINE_TO_ADDON = {
     [ModelType.llamacppEmbedding]: "embeddings",
     [ModelType.nmtcppTranslation]: "nmt",
     [ModelType.onnxTts]: "tts",
+    [ModelType.ttsGgml]: "tts",
     [ModelType.onnxOcr]: "ocr",
     [ModelType.parakeetTranscription]: "parakeet",
     [ModelType.sdcppGeneration]: "diffusion",
+    [ModelType.ggmlVla]: "vla",
+    [ModelType.ggmlClassification]: "classification",
     "onnx-vad": "vad",
 };
 // Legacy engine names → canonical engine.
@@ -21,29 +24,39 @@ const LEGACY_ENGINE_TO_CANONICAL = {
     [ADDON_WHISPER]: ModelType.whispercppTranscription,
     [ADDON_EMBEDDING]: ModelType.llamacppEmbedding,
     [ADDON_NMT]: ModelType.nmtcppTranslation,
-    [ADDON_TTS]: ModelType.onnxTts,
+    [ADDON_TTS]: ModelType.ttsGgml,
     [ADDON_OCR]: ModelType.onnxOcr,
     [ADDON_PARAKEET]: ModelType.parakeetTranscription,
     "@qvac/translation-llamacpp": ModelType.nmtcppTranslation,
     "@qvac/vad-silero": "onnx-vad",
-    "@qvac/tts": ModelType.onnxTts,
+    // Legacy package / tag names from the ONNX era — resolve to the GGML engine.
+    "@qvac/tts": ModelType.ttsGgml,
+    "@qvac/tts-onnx": ModelType.ttsGgml,
     // Tag-style names (used by some older registry entries)
     generation: ModelType.llamacppCompletion,
     transcription: ModelType.whispercppTranscription,
     embedding: ModelType.llamacppEmbedding,
     translation: ModelType.nmtcppTranslation,
     vad: "onnx-vad",
-    tts: ModelType.onnxTts,
+    tts: ModelType.ttsGgml,
     ocr: ModelType.onnxOcr,
     [ADDON_DIFFUSION]: ModelType.sdcppGeneration,
     diffusion: ModelType.sdcppGeneration,
+    [ADDON_VLA]: ModelType.ggmlVla,
+    vla: ModelType.ggmlVla,
+    [ADDON_CLASSIFICATION]: ModelType.ggmlClassification,
+    classification: ModelType.ggmlClassification,
 };
 // Resolves any engine string (legacy or canonical) to a validated canonical engine.
 // Returns null if the engine is not recognized.
 export function resolveCanonicalEngine(engine) {
     const direct = modelRegistryEngineSchema.safeParse(engine);
-    if (direct.success)
+    if (direct.success) {
+        // Registry rows and cached metadata may still say "onnx-tts"; route to GGML.
+        if (direct.data === ModelType.onnxTts)
+            return ModelType.ttsGgml;
         return direct.data;
+    }
     const canonical = LEGACY_ENGINE_TO_CANONICAL[engine];
     if (canonical)
         return canonical;

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

@@ -39,10 +39,12 @@ export declare const modelInfoSchema: z.ZodObject<{
         diffusion: "diffusion";
         tts: "tts";
         ocr: "ocr";
+        vla: "vla";
         llm: "llm";
         whisper: "whisper";
         nmt: "nmt";
         parakeet: "parakeet";
+        classification: "classification";
         vad: "vad";
         other: "other";
     }>;
@@ -86,10 +88,12 @@ export declare const getModelInfoResponseSchema: z.ZodObject<{
             diffusion: "diffusion";
             tts: "tts";
             ocr: "ocr";
+            vla: "vla";
             llm: "llm";
             whisper: "whisper";
             nmt: "nmt";
             parakeet: "parakeet";
+            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

@@ -58,18 +58,23 @@ export declare const llmConfigBaseSchema: z.ZodObject<{
             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";
+            classification: "classification";
         }>, z.ZodLiteral<"vad">]>>;
     }, z.core.$strip>]>>;
 }, z.core.$strip>;
@@ -133,18 +138,23 @@ export declare const llmConfigSchema: z.ZodPipe<z.ZodObject<{
             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";
+            classification: "classification";
         }>, z.ZodLiteral<"vad">]>>;
     }, z.core.$strip>]>>;
 }, z.core.$strip>, z.ZodTransform<{
@@ -185,7 +195,7 @@ export declare const llmConfigSchema: z.ZodPipe<z.ZodObject<{
         engine?: string | undefined;
         expectedSize?: number | undefined;
         sha256Checksum?: string | undefined;
-        addon?: "embeddings" | "diffusion" | "tts" | "ocr" | "llamacpp-completion" | "whispercpp-transcription" | "llamacpp-embedding" | "nmtcpp-translation" | "onnx-tts" | "parakeet-transcription" | "onnx-ocr" | "sdcpp-generation" | "llm" | "whisper" | "nmt" | "parakeet" | "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;
@@ -225,7 +235,7 @@ export declare const llmConfigSchema: z.ZodPipe<z.ZodObject<{
         engine?: string | undefined;
         expectedSize?: number | undefined;
         sha256Checksum?: string | undefined;
-        addon?: "embeddings" | "diffusion" | "tts" | "ocr" | "llamacpp-completion" | "whispercpp-transcription" | "llamacpp-embedding" | "nmtcpp-translation" | "onnx-tts" | "parakeet-transcription" | "onnx-ocr" | "sdcpp-generation" | "llm" | "whisper" | "nmt" | "parakeet" | "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>;