@qvac/sdk 0.17.0 → 0.18.0

package/dist/schemas/constants-registry.d.ts

@@ -39,10 +39,6 @@ export declare const constantsRegistry: {
         readonly ggmlVla: "ggml-vla";
         readonly ggmlClassification: "ggml-classification";
     }>;
-    readonly ToolsMode: z.ZodEnum<{
-        readonly static: "static";
-        readonly dynamic: "dynamic";
-    }>;
     readonly Verbosity: z.ZodEnum<{
         readonly ERROR: 0;
         readonly WARN: 1;
@@ -65,5 +61,23 @@ export declare const constantsRegistry: {
     readonly SupportedAudioFormat: z.ZodEnum<{
         [k: string]: string;
     }>;
+    readonly TtsPace: z.ZodEnum<{
+        [k: string]: string;
+    }>;
+    readonly TtsCosyvoice3Emotion: z.ZodEnum<{
+        [k: string]: string;
+    }>;
+    readonly TtsCosyvoice3InstructDialect: z.ZodEnum<{
+        [k: string]: string;
+    }>;
+    readonly TtsCosyvoice3InstructVolume: z.ZodEnum<{
+        [k: string]: string;
+    }>;
+    readonly TtsCosyvoice3InstructStyle: z.ZodEnum<{
+        [k: string]: string;
+    }>;
+    readonly AudioGenTaskType: z.ZodEnum<{
+        [k: string]: string;
+    }>;
 };
 //# sourceMappingURL=constants-registry.d.ts.map

package/dist/schemas/constants-registry.js

@@ -1,9 +1,17 @@
 import { z } from 'zod';
 import { ModelType } from './model-types.js';
-import { TOOLS_MODE } from './tools.js';
 import { VERBOSITY } from './llamacpp-config.js';
 import { PLUGIN_LLM, PLUGIN_EMBEDDING, PLUGIN_WHISPER, PLUGIN_BCI, PLUGIN_NMT, PLUGIN_TTS, PLUGIN_OCR, PLUGIN_DIFFUSION, PLUGIN_AUDIOGEN, PLUGIN_VLA, PLUGIN_CLASSIFICATION } from './plugin.js';
 import { SUPPORTED_AUDIO_FORMATS } from '../constants/audio.js';
+import { AUDIOGEN_TASK_TYPES } from './audio-gen.js';
+import { TTS_PACES, TTS_COSYVOICE3_EMOTIONS, TTS_COSYVOICE3_INSTRUCT_DIALECTS, TTS_COSYVOICE3_INSTRUCT_VOLUMES, TTS_COSYVOICE3_INSTRUCT_STYLES } from './text-to-speech.js';
+// Uppercases each vocabulary value into an identifier-safe varname (e.g.
+// 'slow' -> 'SLOW', 'cover-nofsq' -> 'COVER_NOFSQ'), mirroring the
+// SupportedAudioFormat derivation so names can't drift from the single
+// source-of-truth array.
+function enumFromVocabulary(values) {
+    return z.enum(Object.fromEntries(values.map((value) => [value.toUpperCase().replace(/-/g, '_'), value])));
+}
 /**
  * Every public constant `enum` from index.ts that downstream (non-JS)
  * client generators should get as a real, named type instead of a
@@ -30,7 +38,6 @@ import { SUPPORTED_AUDIO_FORMATS } from '../constants/audio.js';
  */
 export const constantsRegistry = {
     ModelType: z.enum(ModelType),
-    ToolsMode: z.enum(TOOLS_MODE),
     Verbosity: z.enum(VERBOSITY),
     PluginId: z.enum({
         LLM: PLUGIN_LLM,
@@ -48,5 +55,11 @@ export const constantsRegistry = {
     // Derived from SUPPORTED_AUDIO_FORMATS itself (e.g. '.mp3' -> 'MP3') rather
     // than a hand-written record, so the varnames can't drift from the one
     // source of truth every other consumer of SUPPORTED_AUDIO_FORMATS shares.
-    SupportedAudioFormat: z.enum(Object.fromEntries(SUPPORTED_AUDIO_FORMATS.map((ext) => [ext.slice(1).toUpperCase(), ext])))
+    SupportedAudioFormat: z.enum(Object.fromEntries(SUPPORTED_AUDIO_FORMATS.map((ext) => [ext.slice(1).toUpperCase(), ext]))),
+    TtsPace: enumFromVocabulary(TTS_PACES),
+    TtsCosyvoice3Emotion: enumFromVocabulary(TTS_COSYVOICE3_EMOTIONS),
+    TtsCosyvoice3InstructDialect: enumFromVocabulary(TTS_COSYVOICE3_INSTRUCT_DIALECTS),
+    TtsCosyvoice3InstructVolume: enumFromVocabulary(TTS_COSYVOICE3_INSTRUCT_VOLUMES),
+    TtsCosyvoice3InstructStyle: enumFromVocabulary(TTS_COSYVOICE3_INSTRUCT_STYLES),
+    AudioGenTaskType: enumFromVocabulary(AUDIOGEN_TASK_TYPES)
 };

package/dist/schemas/index.d.ts

@@ -3,7 +3,7 @@ export * from './cancel.js';
 export * from './batch-completion-stream.js';
 export * from './completion-stream.js';
 export * from './completion-event.js';
-export { toolSchema, toolCallSchema, toolCallErrorSchema, TOOLS_MODE, type Tool, type ToolCall, type ToolCallError, type ToolCallWithCall, type ToolsMode } from './tools.js';
+export { toolSchema, toolCallSchema, toolCallErrorSchema, type Tool, type ToolCall, type ToolCallError, type ToolCallWithCall } from './tools.js';
 export * from './delegate.js';
 export * from './delete-cache.js';
 export * from './download-asset.js';

package/dist/schemas/index.js

@@ -4,7 +4,7 @@ export * from './cancel.js';
 export * from './batch-completion-stream.js';
 export * from './completion-stream.js';
 export * from './completion-event.js';
-export { toolSchema, toolCallSchema, toolCallErrorSchema, TOOLS_MODE } from './tools.js';
+export { toolSchema, toolCallSchema, toolCallErrorSchema } from './tools.js';
 export * from './delegate.js';
 export * from './delete-cache.js';
 export * from './download-asset.js';

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

@@ -36,10 +36,6 @@ export declare const llmConfigBaseSchema: z.ZodObject<{
     n_discarded: z.ZodOptional<z.ZodNumber>;
     parallel: z.ZodOptional<z.ZodNumber>;
     tools: z.ZodOptional<z.ZodBoolean>;
-    toolsMode: z.ZodOptional<z.ZodEnum<{
-        static: "static";
-        dynamic: "dynamic";
-    }>>;
     '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<{
@@ -98,6 +94,10 @@ export declare const llmConfigBaseSchema: z.ZodObject<{
         batched: "batched";
         sequential: "sequential";
     }>>;
+    image_no_upscale: z.ZodOptional<z.ZodEnum<{
+        on: "on";
+        off: "off";
+    }>>;
     'mmproj-use-gpu': z.ZodOptional<z.ZodBoolean>;
 }, z.core.$strip>;
 export type LlmConfigInput = z.infer<typeof llmConfigBaseSchema>;
@@ -133,10 +133,6 @@ export declare const llmConfigSchema: z.ZodPipe<z.ZodObject<{
     n_discarded: z.ZodOptional<z.ZodNumber>;
     parallel: z.ZodOptional<z.ZodNumber>;
     tools: z.ZodOptional<z.ZodBoolean>;
-    toolsMode: z.ZodOptional<z.ZodEnum<{
-        static: "static";
-        dynamic: "dynamic";
-    }>>;
     'cache-type-k': z.ZodOptional<z.ZodString>;
     'cache-type-v': z.ZodOptional<z.ZodString>;
     'main-gpu': z.ZodOptional<z.ZodUnion<readonly [z.ZodNumber, z.ZodEnum<{
@@ -195,6 +191,10 @@ export declare const llmConfigSchema: z.ZodPipe<z.ZodObject<{
         batched: "batched";
         sequential: "sequential";
     }>>;
+    image_no_upscale: z.ZodOptional<z.ZodEnum<{
+        on: "on";
+        off: "off";
+    }>>;
     'mmproj-use-gpu': z.ZodOptional<z.ZodBoolean>;
 }, z.core.$strip>, z.ZodTransform<{
     ctx_size: number | undefined;
@@ -216,7 +216,6 @@ export declare const llmConfigSchema: z.ZodPipe<z.ZodObject<{
     n_discarded?: number | undefined;
     parallel?: number | undefined;
     tools?: boolean | undefined;
-    toolsMode?: "static" | "dynamic" | undefined;
     'cache-type-k'?: string | undefined;
     'cache-type-v'?: string | undefined;
     'main-gpu'?: number | "integrated" | "dedicated" | undefined;
@@ -238,6 +237,7 @@ export declare const llmConfigSchema: z.ZodPipe<z.ZodObject<{
         addon?: "embeddings" | "diffusion" | "audiogen" | "tts" | "ocr" | "vla" | "llamacpp-completion" | "whispercpp-transcription" | "bci-whispercpp-transcription" | "llamacpp-embedding" | "nmtcpp-translation" | "onnx-tts" | "tts-ggml" | "parakeet-transcription" | "ggml-ocr" | "sdcpp-generation" | "audiogen-ggml" | "ggml-vla" | "ggml-classification" | "llm" | "whisper" | "bci" | "nmt" | "parakeet" | "classification" | "vad" | undefined;
     } | undefined;
     image_tile_mode: "disabled" | "batched" | "sequential" | undefined;
+    image_no_upscale?: "on" | "off" | undefined;
     'mmproj-use-gpu'?: boolean | undefined;
 }, {
     ctx_size?: number | undefined;
@@ -259,7 +259,6 @@ export declare const llmConfigSchema: z.ZodPipe<z.ZodObject<{
     n_discarded?: number | undefined;
     parallel?: number | undefined;
     tools?: boolean | undefined;
-    toolsMode?: "static" | "dynamic" | undefined;
     'cache-type-k'?: string | undefined;
     'cache-type-v'?: string | undefined;
     'main-gpu'?: number | "integrated" | "dedicated" | undefined;
@@ -281,6 +280,7 @@ export declare const llmConfigSchema: z.ZodPipe<z.ZodObject<{
         addon?: "embeddings" | "diffusion" | "audiogen" | "tts" | "ocr" | "vla" | "llamacpp-completion" | "whispercpp-transcription" | "bci-whispercpp-transcription" | "llamacpp-embedding" | "nmtcpp-translation" | "onnx-tts" | "tts-ggml" | "parakeet-transcription" | "ggml-ocr" | "sdcpp-generation" | "audiogen-ggml" | "ggml-vla" | "ggml-classification" | "llm" | "whisper" | "bci" | "nmt" | "parakeet" | "classification" | "vad" | undefined;
     } | undefined;
     image_tile_mode?: "disabled" | "batched" | "sequential" | undefined;
+    image_no_upscale?: "on" | "off" | undefined;
     'mmproj-use-gpu'?: boolean | undefined;
 }>>;
 export type LlmConfig = z.infer<typeof llmConfigSchema>;

package/dist/schemas/llamacpp-config.js

@@ -1,6 +1,5 @@
 import { z } from 'zod';
 import { modelSrcInputSchema } from './model-src-utils.js';
-import { TOOLS_MODE } from './tools.js';
 /**
  * Upper bound for `reasoning_budget`. Mirrors the llm-llamacpp addon, which
  * stores the budget as a 32-bit `int` and rejects values above
@@ -47,10 +46,6 @@ export const llmConfigBaseSchema = z.object({
      */
     parallel: z.number().int().min(1).optional(),
     tools: z.boolean().optional(),
-    toolsMode: z
-        .enum([TOOLS_MODE.static, TOOLS_MODE.dynamic])
-        .describe('Controls tool placement in the prompt. "static" (default) prepends the tool set once and reuses it across the session. "dynamic" anchors tools after the last user message and trims them from the kv-cache after the chain resolves so each user prompt can carry its own tools.')
-        .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(),
@@ -76,6 +71,18 @@ export const llmConfigBaseSchema = z.object({
      * Ignored by text-only models. Default is `"sequential"`.
      */
     image_tile_mode: z.enum(['disabled', 'batched', 'sequential']).optional(),
+    /**
+     * idefics3-style image preprocessing rule (multimodal models only):
+     *   - `"on"`: round the image's long side up to a whole number of slices and
+     *     cap it, so an image smaller than the cap keeps its own resolution and
+     *     becomes far fewer slices.
+     *   - `"off"`: always stretch the long side to the cap.
+     * When unset, the model's own GGUF value is used. Ignored with a warning by
+     * models that do not use idefics3-style preprocessing. Changes the number of
+     * image tokens, and therefore both accuracy and encode time, so a checkpoint
+     * whose GGUF omits the key needs this set to preprocess correctly.
+     */
+    image_no_upscale: z.enum(['on', 'off']).optional(),
     /**
      * Run the multimodal projector (mmproj / vision encoder) on the GPU
      * (multimodal models only). `true` forces GPU, `false` forces CPU. When