@qvac/sdk 0.15.0 → 0.16.0

package/dist/server/bare/plugins/sdcpp-generation/plugin.d.ts

@@ -1,5 +1,5 @@
 import { type CreateModelParams, type PluginModelResult, type ResolveContext, type ResolveResult, type SdcppConfig } from '../../../../schemas/index';
-type DiffusionArtifactKey = 'clipLModelPath' | 'clipGModelPath' | 'clipVisionModelPath' | 't5XxlModelPath' | 'llmModelPath' | 'vaeModelPath' | 'highNoiseDiffusionModelPath' | 'esrganModelPath';
+type DiffusionArtifactKey = 'clipLModelPath' | 'clipGModelPath' | 'clipVisionModelPath' | 't5XxlModelPath' | 'llmModelPath' | 'vaeModelPath' | 'highNoiseDiffusionModelPath' | 'uncondModelPath' | 'audioVaeModelPath' | 'embeddingsConnectorsModelPath' | 'esrganModelPath';
 /**
  * Stable-diffusion.cpp plugin for image diffusion, upscaling, and Wan video.
  *
@@ -23,6 +23,10 @@ export declare const diffusionPlugin: {
             gpu: "gpu";
             cpu: "cpu";
         }>>;
+        'main-gpu': import("zod").ZodOptional<import("zod").ZodUnion<readonly [import("zod").ZodNumber, import("zod").ZodEnum<{
+            integrated: "integrated";
+            dedicated: "dedicated";
+        }>]>>;
         prediction: import("zod").ZodOptional<import("zod").ZodEnum<{
             auto: "auto";
             eps: "eps";
@@ -291,6 +295,43 @@ export declare const diffusionPlugin: {
                 classification: "classification";
             }>, import("zod").ZodLiteral<"vad">]>>;
         }, import("zod/v4/core").$strip>]>>;
+        uncondModelSrc: import("zod").ZodOptional<import("zod").ZodUnion<readonly [import("zod").ZodString, import("zod").ZodObject<{
+            src: import("zod").ZodString;
+            name: import("zod").ZodOptional<import("zod").ZodString>;
+            modelId: import("zod").ZodOptional<import("zod").ZodString>;
+            registryPath: import("zod").ZodOptional<import("zod").ZodString>;
+            registrySource: import("zod").ZodOptional<import("zod").ZodString>;
+            blobCoreKey: import("zod").ZodOptional<import("zod").ZodString>;
+            blobIndex: import("zod").ZodOptional<import("zod").ZodNumber>;
+            engine: import("zod").ZodOptional<import("zod").ZodString>;
+            expectedSize: import("zod").ZodOptional<import("zod").ZodNumber>;
+            sha256Checksum: import("zod").ZodOptional<import("zod").ZodString>;
+            addon: import("zod").ZodOptional<import("zod").ZodUnion<readonly [import("zod").ZodEnum<{
+                embeddings: "embeddings";
+                diffusion: "diffusion";
+                tts: "tts";
+                ocr: "ocr";
+                vla: "vla";
+                "llamacpp-completion": "llamacpp-completion";
+                "whispercpp-transcription": "whispercpp-transcription";
+                "bci-whispercpp-transcription": "bci-whispercpp-transcription";
+                "llamacpp-embedding": "llamacpp-embedding";
+                "nmtcpp-translation": "nmtcpp-translation";
+                "onnx-tts": "onnx-tts";
+                "tts-ggml": "tts-ggml";
+                "parakeet-transcription": "parakeet-transcription";
+                "ggml-ocr": "ggml-ocr";
+                "sdcpp-generation": "sdcpp-generation";
+                "ggml-vla": "ggml-vla";
+                "ggml-classification": "ggml-classification";
+                llm: "llm";
+                whisper: "whisper";
+                bci: "bci";
+                nmt: "nmt";
+                parakeet: "parakeet";
+                classification: "classification";
+            }>, import("zod").ZodLiteral<"vad">]>>;
+        }, import("zod/v4/core").$strip>]>>;
         clipVisionModelSrc: import("zod").ZodOptional<import("zod").ZodUnion<readonly [import("zod").ZodString, import("zod").ZodObject<{
             src: import("zod").ZodString;
             name: import("zod").ZodOptional<import("zod").ZodString>;
@@ -328,6 +369,80 @@ export declare const diffusionPlugin: {
                 classification: "classification";
             }>, import("zod").ZodLiteral<"vad">]>>;
         }, import("zod/v4/core").$strip>]>>;
+        audioVaeModelSrc: import("zod").ZodOptional<import("zod").ZodUnion<readonly [import("zod").ZodString, import("zod").ZodObject<{
+            src: import("zod").ZodString;
+            name: import("zod").ZodOptional<import("zod").ZodString>;
+            modelId: import("zod").ZodOptional<import("zod").ZodString>;
+            registryPath: import("zod").ZodOptional<import("zod").ZodString>;
+            registrySource: import("zod").ZodOptional<import("zod").ZodString>;
+            blobCoreKey: import("zod").ZodOptional<import("zod").ZodString>;
+            blobIndex: import("zod").ZodOptional<import("zod").ZodNumber>;
+            engine: import("zod").ZodOptional<import("zod").ZodString>;
+            expectedSize: import("zod").ZodOptional<import("zod").ZodNumber>;
+            sha256Checksum: import("zod").ZodOptional<import("zod").ZodString>;
+            addon: import("zod").ZodOptional<import("zod").ZodUnion<readonly [import("zod").ZodEnum<{
+                embeddings: "embeddings";
+                diffusion: "diffusion";
+                tts: "tts";
+                ocr: "ocr";
+                vla: "vla";
+                "llamacpp-completion": "llamacpp-completion";
+                "whispercpp-transcription": "whispercpp-transcription";
+                "bci-whispercpp-transcription": "bci-whispercpp-transcription";
+                "llamacpp-embedding": "llamacpp-embedding";
+                "nmtcpp-translation": "nmtcpp-translation";
+                "onnx-tts": "onnx-tts";
+                "tts-ggml": "tts-ggml";
+                "parakeet-transcription": "parakeet-transcription";
+                "ggml-ocr": "ggml-ocr";
+                "sdcpp-generation": "sdcpp-generation";
+                "ggml-vla": "ggml-vla";
+                "ggml-classification": "ggml-classification";
+                llm: "llm";
+                whisper: "whisper";
+                bci: "bci";
+                nmt: "nmt";
+                parakeet: "parakeet";
+                classification: "classification";
+            }>, import("zod").ZodLiteral<"vad">]>>;
+        }, import("zod/v4/core").$strip>]>>;
+        embeddingsConnectorsModelSrc: import("zod").ZodOptional<import("zod").ZodUnion<readonly [import("zod").ZodString, import("zod").ZodObject<{
+            src: import("zod").ZodString;
+            name: import("zod").ZodOptional<import("zod").ZodString>;
+            modelId: import("zod").ZodOptional<import("zod").ZodString>;
+            registryPath: import("zod").ZodOptional<import("zod").ZodString>;
+            registrySource: import("zod").ZodOptional<import("zod").ZodString>;
+            blobCoreKey: import("zod").ZodOptional<import("zod").ZodString>;
+            blobIndex: import("zod").ZodOptional<import("zod").ZodNumber>;
+            engine: import("zod").ZodOptional<import("zod").ZodString>;
+            expectedSize: import("zod").ZodOptional<import("zod").ZodNumber>;
+            sha256Checksum: import("zod").ZodOptional<import("zod").ZodString>;
+            addon: import("zod").ZodOptional<import("zod").ZodUnion<readonly [import("zod").ZodEnum<{
+                embeddings: "embeddings";
+                diffusion: "diffusion";
+                tts: "tts";
+                ocr: "ocr";
+                vla: "vla";
+                "llamacpp-completion": "llamacpp-completion";
+                "whispercpp-transcription": "whispercpp-transcription";
+                "bci-whispercpp-transcription": "bci-whispercpp-transcription";
+                "llamacpp-embedding": "llamacpp-embedding";
+                "nmtcpp-translation": "nmtcpp-translation";
+                "onnx-tts": "onnx-tts";
+                "tts-ggml": "tts-ggml";
+                "parakeet-transcription": "parakeet-transcription";
+                "ggml-ocr": "ggml-ocr";
+                "sdcpp-generation": "sdcpp-generation";
+                "ggml-vla": "ggml-vla";
+                "ggml-classification": "ggml-classification";
+                llm: "llm";
+                whisper: "whisper";
+                bci: "bci";
+                nmt: "nmt";
+                parakeet: "parakeet";
+                classification: "classification";
+            }>, import("zod").ZodLiteral<"vad">]>>;
+        }, import("zod/v4/core").$strip>]>>;
         upscaler: import("zod").ZodOptional<import("zod").ZodObject<{
             type: import("zod").ZodOptional<import("zod").ZodLiteral<"esrgan">>;
             model_src: import("zod").ZodOptional<import("zod").ZodUnion<readonly [import("zod").ZodString, import("zod").ZodObject<{
@@ -439,6 +554,11 @@ export declare const diffusionPlugin: {
             stats: import("zod").ZodOptional<import("zod").ZodObject<{
                 modelLoadMs: import("zod").ZodOptional<import("zod").ZodNumber>;
                 generationMs: import("zod").ZodOptional<import("zod").ZodNumber>;
+                conditionerMs: import("zod").ZodOptional<import("zod").ZodNumber>;
+                denoiseMs: import("zod").ZodOptional<import("zod").ZodNumber>;
+                vaeMs: import("zod").ZodOptional<import("zod").ZodNumber>;
+                postProcessMs: import("zod").ZodOptional<import("zod").ZodNumber>;
+                stepsPerSecond: import("zod").ZodOptional<import("zod").ZodNumber>;
                 totalGenerationMs: import("zod").ZodOptional<import("zod").ZodNumber>;
                 totalWallMs: import("zod").ZodOptional<import("zod").ZodNumber>;
                 totalSteps: import("zod").ZodOptional<import("zod").ZodNumber>;
@@ -530,6 +650,7 @@ export declare const diffusionPlugin: {
             vae_tiling: import("zod").ZodOptional<import("zod").ZodBoolean>;
             vae_tile_size: import("zod").ZodOptional<import("zod").ZodUnion<readonly [import("zod").ZodNumber, import("zod").ZodString]>>;
             vae_tile_overlap: import("zod").ZodOptional<import("zod").ZodNumber>;
+            temporal_tiling: import("zod").ZodOptional<import("zod").ZodBoolean>;
             cache_mode: import("zod").ZodOptional<import("zod").ZodEnum<{
                 disabled: "disabled";
                 easycache: "easycache";
@@ -560,6 +681,11 @@ export declare const diffusionPlugin: {
                 height: import("zod").ZodOptional<import("zod").ZodNumber>;
                 modelLoadMs: import("zod").ZodOptional<import("zod").ZodNumber>;
                 generationMs: import("zod").ZodOptional<import("zod").ZodNumber>;
+                conditionerMs: import("zod").ZodOptional<import("zod").ZodNumber>;
+                denoiseMs: import("zod").ZodOptional<import("zod").ZodNumber>;
+                vaeMs: import("zod").ZodOptional<import("zod").ZodNumber>;
+                postProcessMs: import("zod").ZodOptional<import("zod").ZodNumber>;
+                stepsPerSecond: import("zod").ZodOptional<import("zod").ZodNumber>;
                 totalGenerationMs: import("zod").ZodOptional<import("zod").ZodNumber>;
                 totalWallMs: import("zod").ZodOptional<import("zod").ZodNumber>;
                 totalSteps: import("zod").ZodOptional<import("zod").ZodNumber>;
@@ -570,6 +696,8 @@ export declare const diffusionPlugin: {
                 totalVideoFrames: import("zod").ZodOptional<import("zod").ZodNumber>;
                 videoFrames: import("zod").ZodOptional<import("zod").ZodNumber>;
                 fps: import("zod").ZodOptional<import("zod").ZodNumber>;
+                hasAudio: import("zod").ZodOptional<import("zod").ZodBoolean>;
+                audioSampleRate: import("zod").ZodOptional<import("zod").ZodNumber>;
             }, import("zod/v4/core").$strip>>;
         }, import("zod/v4/core").$strip>>;
         upscaleStream: import("../../../../schemas/index").PluginHandlerDefinition<import("zod").ZodObject<{

package/dist/server/bare/plugins/sdcpp-generation/plugin.js

@@ -1,10 +1,12 @@
 import ImgStableDiffusion, { EsrganUpscaler, VideoStableDiffusion } from '@qvac/diffusion-cpp';
 import addonLogging from '@qvac/diffusion-cpp/addonLogging.js';
 import { definePlugin, defineHandler, sdcppConfigSchema, diffusionRequestSchema, diffusionStreamResponseSchema, videoRequestSchema, videoStreamResponseSchema, upscaleRequestSchema, upscaleStreamResponseSchema, ModelType, ADDON_DIFFUSION } from '../../../../schemas/index.js';
-import { createStreamLogger, registerAddonLogger } from '../../../../logging/index.js';
+import { createStreamLogger, registerAddonLogger, getServerLogger } from '../../../../logging/index.js';
 import { ModelLoadFailedError } from '../../../../utils/errors-server.js';
+import { isMobile } from '../../registry/runtime-context-registry.js';
+import { stripMultiGpuKeys } from '../../../utils/multi-gpu-mobile.js';
 import { diffusion } from './ops/diffusion.js';
-import { video } from './ops/video.js';
+import { markLtxVideoModel, video } from './ops/video.js';
 import { upscale } from './ops/upscale.js';
 // Single source of truth for `SdcppConfig.upscaler.*` → addon-config key
 // mapping. Used by both the diffusion-mode (post-generation upscaler) and the
@@ -52,13 +54,36 @@ export const diffusionPlugin = definePlugin({
     addonPackage: ADDON_DIFFUSION,
     loadConfigSchema: sdcppConfigSchema,
     async resolveConfig(cfg, ctx) {
+        if (cfg.uncondModelSrc && cfg.mode !== 'diffusion') {
+            throw new ModelLoadFailedError('modelConfig.uncondModelSrc is Ideogram 4 diffusion only. ' +
+                "Use mode: 'diffusion' or remove uncondModelSrc.");
+        }
+        if (cfg.uncondModelSrc && (!cfg.llmModelSrc || !cfg.vaeModelSrc)) {
+            throw new ModelLoadFailedError('modelConfig.uncondModelSrc selects the Ideogram 4 layout and requires ' +
+                'modelConfig.llmModelSrc (Qwen3-VL) and modelConfig.vaeModelSrc.');
+        }
         // Standalone-upscaler mode never references auxiliary models: the primary
         // modelSrc IS the ESRGAN file. Skip resolution to avoid downloading
         // unused encoders/VAEs and to keep load fast.
         if (cfg.mode === 'upscale') {
             return { config: cfg };
         }
-        const { clipLModelSrc, clipGModelSrc, clipVisionModelSrc, t5XxlModelSrc, llmModelSrc, vaeModelSrc, highNoiseDiffusionModelSrc, upscaler, ...rest } = cfg;
+        const { clipLModelSrc, clipGModelSrc, clipVisionModelSrc, t5XxlModelSrc, llmModelSrc, vaeModelSrc, highNoiseDiffusionModelSrc, uncondModelSrc, audioVaeModelSrc, embeddingsConnectorsModelSrc, upscaler, ...rest } = cfg;
+        // audioVae / embeddingsConnectors are LTX-2 video companions. Reject them
+        // up front — before any companion download — when the selected layout
+        // can't consume them, instead of silently resolving the file and then
+        // dropping it in createModel. Both fields are new, so this cannot break
+        // existing callers.
+        if (embeddingsConnectorsModelSrc && cfg.mode !== 'video') {
+            throw new ModelLoadFailedError('modelConfig.embeddingsConnectorsModelSrc selects the LTX-2 video ' +
+                "layout and is only valid with mode: 'video'.");
+        }
+        if (audioVaeModelSrc && !(cfg.mode === 'video' && embeddingsConnectorsModelSrc)) {
+            throw new ModelLoadFailedError('modelConfig.audioVaeModelSrc is LTX-2 video only. It requires ' +
+                "mode: 'video' together with modelConfig.embeddingsConnectorsModelSrc " +
+                '(which selects the LTX-2 layout). Add embeddingsConnectorsModelSrc ' +
+                'or remove audioVaeModelSrc.');
+        }
         // Video jobs do not apply ESRGAN so we drop the whole `upscaler` object.
         const effectiveUpscaler = cfg.mode === 'video' ? undefined : upscaler;
         const { model_src: esrganModelSrc, ...upscalerRuntime } = effectiveUpscaler ?? {};
@@ -74,6 +99,9 @@ export const diffusionPlugin = definePlugin({
             llmModelSrc,
             vaeModelSrc,
             highNoiseDiffusionModelSrc,
+            uncondModelSrc,
+            audioVaeModelSrc,
+            embeddingsConnectorsModelSrc,
             esrganModelSrc
         };
         const hasSources = Object.values(sources).some(Boolean);
@@ -81,7 +109,7 @@ export const diffusionPlugin = definePlugin({
             return { config: runtimeConfig };
         }
         const resolve = ctx.resolveModelPath;
-        const [clipLModelPath, clipGModelPath, clipVisionModelPath, t5XxlModelPath, llmModelPath, vaeModelPath, highNoiseDiffusionModelPath, esrganModelPath] = await Promise.all([
+        const [clipLModelPath, clipGModelPath, clipVisionModelPath, t5XxlModelPath, llmModelPath, vaeModelPath, highNoiseDiffusionModelPath, uncondModelPath, audioVaeModelPath, embeddingsConnectorsModelPath, esrganModelPath] = await Promise.all([
             clipLModelSrc ? resolve(clipLModelSrc) : undefined,
             clipGModelSrc ? resolve(clipGModelSrc) : undefined,
             clipVisionModelSrc ? resolve(clipVisionModelSrc) : undefined,
@@ -89,6 +117,9 @@ export const diffusionPlugin = definePlugin({
             llmModelSrc ? resolve(llmModelSrc) : undefined,
             vaeModelSrc ? resolve(vaeModelSrc) : undefined,
             highNoiseDiffusionModelSrc ? resolve(highNoiseDiffusionModelSrc) : undefined,
+            uncondModelSrc ? resolve(uncondModelSrc) : undefined,
+            audioVaeModelSrc ? resolve(audioVaeModelSrc) : undefined,
+            embeddingsConnectorsModelSrc ? resolve(embeddingsConnectorsModelSrc) : undefined,
             esrganModelSrc ? resolve(esrganModelSrc) : undefined
         ]);
         return {
@@ -101,6 +132,9 @@ export const diffusionPlugin = definePlugin({
                 ...(llmModelPath && { llmModelPath }),
                 ...(vaeModelPath && { vaeModelPath }),
                 ...(highNoiseDiffusionModelPath && { highNoiseDiffusionModelPath }),
+                ...(uncondModelPath && { uncondModelPath }),
+                ...(audioVaeModelPath && { audioVaeModelPath }),
+                ...(embeddingsConnectorsModelPath && { embeddingsConnectorsModelPath }),
                 ...(esrganModelPath && { esrganModelPath })
             }
         };
@@ -108,6 +142,13 @@ export const diffusionPlugin = definePlugin({
     createModel(params) {
         const { modelId, modelPath, modelConfig, artifacts } = params;
         const config = (modelConfig ?? {});
+        // no multi-gpu on mobile
+        if (isMobile()) {
+            const stripped = stripMultiGpuKeys(config);
+            if (stripped.length > 0) {
+                getServerLogger().warn(`[${ModelType.sdcppGeneration}:${modelId}] Multi-GPU parameters (${stripped.join(', ')}) are not supported on mobile (single-GPU device) — removing from config; model will load with single-GPU defaults`);
+            }
+        }
         // In diffusion mode the ESRGAN file (when post-generation upscale is
         // wanted) must come from upscaler.model_src — the primary modelPath is
         // the main diffusion checkpoint. Reject early with a clear error
@@ -133,28 +174,54 @@ export const diffusionPlugin = definePlugin({
             return { model };
         }
         if (config.mode === 'video') {
-            if (!artifacts?.['t5XxlModelPath']) {
-                throw new ModelLoadFailedError('modelConfig.t5XxlModelSrc is required in video mode. ' +
-                    'Provide the Wan text encoder model before loading the video pipeline.');
-            }
+            // Layout is selected the same way the addon self-detects it: the
+            // presence of the LTX-2 text-embedding connectors switches from the
+            // Wan layout (UMT5 via t5Xxl) to the LTX-2 layout (Gemma via llm +
+            // video VAE + connectors, optional audio VAE). Mirrors
+            // `SdModel::isLtxModel_ = !embeddingsConnectorsPath.empty()`.
+            const embeddingsConnectorsModelPath = artifacts?.['embeddingsConnectorsModelPath'];
             if (!artifacts?.['vaeModelPath']) {
                 throw new ModelLoadFailedError('modelConfig.vaeModelSrc is required in video mode. ' +
-                    'Provide the Wan VAE model before loading the video pipeline.');
+                    'Provide the Wan or LTX-2 video VAE model before loading the video pipeline.');
+            }
+            const vaeModelPath = artifacts['vaeModelPath'];
+            let files;
+            if (embeddingsConnectorsModelPath) {
+                if (!artifacts['llmModelPath']) {
+                    throw new ModelLoadFailedError('modelConfig.llmModelSrc is required for LTX-2 video. ' +
+                        'Provide the Gemma text encoder model before loading the LTX-2 pipeline.');
+                }
+                files = {
+                    model: modelPath,
+                    vae: vaeModelPath,
+                    llm: artifacts['llmModelPath'],
+                    embeddingsConnectors: embeddingsConnectorsModelPath,
+                    ...(artifacts['audioVaeModelPath'] && {
+                        audioVae: artifacts['audioVaeModelPath']
+                    }),
+                    ...(artifacts['esrganModelPath'] && { esrgan: artifacts['esrganModelPath'] })
+                };
+            }
+            else {
+                if (!artifacts['t5XxlModelPath']) {
+                    throw new ModelLoadFailedError('modelConfig.t5XxlModelSrc is required in video mode. ' +
+                        'Provide the Wan text encoder model before loading the video pipeline.');
+                }
+                files = {
+                    model: modelPath,
+                    vae: vaeModelPath,
+                    t5Xxl: artifacts['t5XxlModelPath'],
+                    ...(artifacts['highNoiseDiffusionModelPath'] && {
+                        highNoiseDiffusionModel: artifacts['highNoiseDiffusionModelPath']
+                    }),
+                    ...(artifacts['clipVisionModelPath'] && {
+                        clipVision: artifacts['clipVisionModelPath']
+                    }),
+                    ...(artifacts['esrganModelPath'] && { esrgan: artifacts['esrganModelPath'] })
+                };
             }
-            const files = {
-                model: modelPath,
-                t5Xxl: artifacts['t5XxlModelPath'],
-                vae: artifacts['vaeModelPath'],
-                ...(artifacts?.['highNoiseDiffusionModelPath'] && {
-                    highNoiseDiffusionModel: artifacts['highNoiseDiffusionModelPath']
-                }),
-                ...(artifacts?.['clipVisionModelPath'] && {
-                    clipVision: artifacts['clipVisionModelPath']
-                }),
-                ...(artifacts?.['esrganModelPath'] && { esrgan: artifacts['esrganModelPath'] })
-            };
             /* eslint-disable @typescript-eslint/no-unused-vars */
-            const { clipLModelSrc, clipGModelSrc, clipVisionModelSrc, t5XxlModelSrc, llmModelSrc, vaeModelSrc, highNoiseDiffusionModelSrc, upscaler, mode, ...rest } = config;
+            const { clipLModelSrc, clipGModelSrc, clipVisionModelSrc, t5XxlModelSrc, llmModelSrc, vaeModelSrc, highNoiseDiffusionModelSrc, uncondModelSrc, audioVaeModelSrc, embeddingsConnectorsModelSrc, upscaler, mode, ...rest } = config;
             /* eslint-enable @typescript-eslint/no-unused-vars */
             const model = new VideoStableDiffusion({
                 files,
@@ -162,6 +229,8 @@ export const diffusionPlugin = definePlugin({
                 logger,
                 opts: { stats: true }
             });
+            if (embeddingsConnectorsModelPath)
+                markLtxVideoModel(model);
             return { model };
         }
         const files = {
@@ -171,6 +240,9 @@ export const diffusionPlugin = definePlugin({
             ...(artifacts?.['t5XxlModelPath'] && { t5Xxl: artifacts['t5XxlModelPath'] }),
             ...(artifacts?.['llmModelPath'] && { llm: artifacts['llmModelPath'] }),
             ...(artifacts?.['vaeModelPath'] && { vae: artifacts['vaeModelPath'] }),
+            ...(artifacts?.['uncondModelPath'] && {
+                uncondModel: artifacts['uncondModelPath']
+            }),
             ...(artifacts?.['esrganModelPath'] && { esrgan: artifacts['esrganModelPath'] })
         };
         // `mode` is consumed by this plugin to select the model class above; the

package/dist/server/bare/plugins/tts-ggml/ops/text-to-speech.js

@@ -68,14 +68,7 @@ export async function* textToSpeech(params) {
             completeBuffer = completeBuffer.concat(Array.from(data.outputArray));
         }
         const modelExecutionMs = nowMs() - modelStart;
-        const stats = {
-            ...(response.stats?.audioDurationMs !== undefined && {
-                audioDuration: response.stats.audioDurationMs
-            }),
-            ...(response.stats?.totalSamples !== undefined && {
-                totalSamples: response.stats.totalSamples
-            })
-        };
+        const stats = collectTtsStats(response);
         yield { buffer: completeBuffer };
         return buildStreamResult(modelExecutionMs, hasDefinedValues(stats) ? stats : undefined);
     }
@@ -83,13 +76,6 @@ export async function* textToSpeech(params) {
         yield { buffer: Array.from(data.outputArray) };
     }
     const modelExecutionMs = nowMs() - modelStart;
-    const stats = {
-        ...(response.stats?.audioDurationMs !== undefined && {
-            audioDuration: response.stats.audioDurationMs
-        }),
-        ...(response.stats?.totalSamples !== undefined && {
-            totalSamples: response.stats.totalSamples
-        })
-    };
+    const stats = collectTtsStats(response);
     return buildStreamResult(modelExecutionMs, hasDefinedValues(stats) ? stats : undefined);
 }

package/dist/server/bare/plugins/tts-ggml/plugin.d.ts

@@ -16,6 +16,7 @@ export declare const ttsPlugin: {
         streamChunkTokens?: number | undefined;
         streamFirstChunkTokens?: number | undefined;
         cfmSteps?: number | undefined;
+        cfgRate?: number | undefined;
         threads?: number | undefined;
         nGpuLayers?: number | undefined;
         seed?: number | undefined;
@@ -27,6 +28,7 @@ export declare const ttsPlugin: {
         ttsNumInferenceSteps?: number | undefined;
         useGPU?: boolean | undefined;
         outputSampleRate?: number | undefined;
+        vulkanCacheDir?: string | undefined;
     }, string>>;
     createModel(params: CreateModelParams): PluginModelResult;
     handlers: {
@@ -46,6 +48,8 @@ export declare const ttsPlugin: {
             stats: import("zod").ZodOptional<import("zod").ZodObject<{
                 audioDuration: import("zod").ZodOptional<import("zod").ZodNumber>;
                 totalSamples: import("zod").ZodOptional<import("zod").ZodNumber>;
+                enhancerBackendDevice: import("zod").ZodOptional<import("zod").ZodNumber>;
+                enhancerBackendId: import("zod").ZodOptional<import("zod").ZodNumber>;
             }, import("zod/v4/core").$strip>>;
             chunkIndex: import("zod").ZodOptional<import("zod").ZodNumber>;
             sentenceChunk: import("zod").ZodOptional<import("zod").ZodString>;
@@ -69,6 +73,8 @@ export declare const ttsPlugin: {
             stats: import("zod").ZodOptional<import("zod").ZodObject<{
                 audioDuration: import("zod").ZodOptional<import("zod").ZodNumber>;
                 totalSamples: import("zod").ZodOptional<import("zod").ZodNumber>;
+                enhancerBackendDevice: import("zod").ZodOptional<import("zod").ZodNumber>;
+                enhancerBackendId: import("zod").ZodOptional<import("zod").ZodNumber>;
             }, import("zod/v4/core").$strip>>;
             chunkIndex: import("zod").ZodOptional<import("zod").ZodNumber>;
             sentenceChunk: import("zod").ZodOptional<import("zod").ZodString>;

package/dist/server/bare/plugins/tts-ggml/plugin.js

@@ -101,6 +101,7 @@ function createChatterboxModel(modelId, config, params, artifacts) {
             ? { streamFirstChunkTokens: config.streamFirstChunkTokens }
             : {}),
         ...(config.cfmSteps !== undefined ? { cfmSteps: config.cfmSteps } : {}),
+        ...(config.cfgRate !== undefined ? { cfgRate: config.cfgRate } : {}),
         ...(config.threads !== undefined ? { threads: config.threads } : {}),
         ...(config.nGpuLayers !== undefined ? { nGpuLayers: config.nGpuLayers } : {}),
         ...(config.seed !== undefined ? { seed: config.seed } : {}),
@@ -134,7 +135,8 @@ function createSupertonicModel(modelId, config, params, artifacts) {
             useGPU: config.useGPU ?? false,
             ...(config.outputSampleRate !== undefined
                 ? { outputSampleRate: config.outputSampleRate }
-                : {})
+                : {}),
+            ...(config.vulkanCacheDir !== undefined ? { vulkanCacheDir: config.vulkanCacheDir } : {})
         },
         logger,
         opts: { stats: true },

package/dist/server/bare/types/addon-responses.d.ts

@@ -26,6 +26,8 @@ export interface NmtResponse {
 export interface TtsStats {
     audioDurationMs?: number;
     totalSamples?: number;
+    enhancerBackendDevice?: number;
+    enhancerBackendId?: number;
 }
 export interface TtsResponse {
     stats?: TtsStats;

package/dist/server/bare/utils/tts-stats.d.ts

@@ -13,10 +13,14 @@ export type TtsOpYield = {
     chunkIndex?: number;
     sentenceChunk?: string;
 };
+type AddonTtsStats = {
+    audioDurationMs?: number;
+    totalSamples?: number;
+    enhancerBackendDevice?: number;
+    enhancerBackendId?: number;
+};
 export declare function collectTtsStats(response: {
-    stats?: {
-        audioDurationMs?: number;
-        totalSamples?: number;
-    };
+    stats?: AddonTtsStats;
 }): TtsStats;
+export {};
 //# sourceMappingURL=tts-stats.d.ts.map

package/dist/server/bare/utils/tts-stats.js

@@ -5,6 +5,12 @@ export function collectTtsStats(response) {
         }),
         ...(response.stats?.totalSamples !== undefined && {
             totalSamples: response.stats.totalSamples
+        }),
+        ...(response.stats?.enhancerBackendDevice !== undefined && {
+            enhancerBackendDevice: response.stats.enhancerBackendDevice
+        }),
+        ...(response.stats?.enhancerBackendId !== undefined && {
+            enhancerBackendId: response.stats.enhancerBackendId
         })
     };
 }

package/dist/server/rpc/create-server.js

@@ -7,9 +7,31 @@ export function createBareKitRPCServer() {
     const { IPC } = globalThis.BareKit;
     return new RPC(IPC, handleRequest);
 }
+/**
+ * Dial the client's IPC endpoint. Two encodings cross the handshake:
+ *
+ * - `tcp://host:port` — a loopback TCP endpoint. Used by clients whose
+ *   runtime has no portable server for filesystem sockets, notably the
+ *   Python `asyncio` transport (`start_unix_server` is Unix-only, and
+ *   asyncio has no cross-platform named-pipe server). One code path on
+ *   every OS, so Windows dials the same way Linux/macOS do.
+ * - anything else — a filesystem path: a Unix domain socket, or a
+ *   `\\.\pipe\...` named pipe on Windows, as the Node client (`node:net`)
+ *   creates.
+ */
+function connectToEndpoint(endpoint) {
+    if (endpoint.startsWith('tcp://')) {
+        const hostPort = endpoint.slice('tcp://'.length);
+        const sep = hostPort.lastIndexOf(':');
+        const host = hostPort.slice(0, sep);
+        const port = Number(hostPort.slice(sep + 1));
+        return connect(port, host);
+    }
+    return connect(endpoint);
+}
 export function createIPCClient(socketPath, options) {
     logger.info(`Connecting to IPC socket at ${socketPath}`);
-    const socket = connect(socketPath);
+    const socket = connectToEndpoint(socketPath);
     socket.on('connect', () => {
         logger.info('Connected to IPC server');
     });

package/dist/server/rpc/handler-registry.js

@@ -1,5 +1,6 @@
 import {} from '../../schemas/index.js';
 import { handleBatchCompletionStream } from './handlers/batch-completion-stream.js';
+import { handleCompletionOrchestrate } from './handlers/completion-orchestrate.js';
 import { handleCompletionStream } from './handlers/completion-stream.js';
 import { handleDownloadAsset } from './handlers/download-asset.js';
 import { handleLoadModel } from './handlers/load-model/index.js';
@@ -147,6 +148,10 @@ export const registry = {
         delegatedHandler: handleCompletionStreamDelegated,
         isDelegated: isModelDelegated
     },
+    // Deliberately no delegated handler: tool callbacks execute code on the
+    // client machine, so only the user's own local worker may orchestrate
+    // (see completionOrchestrateResponseSchema's contract note).
+    completionOrchestrate: { type: 'duplex', handler: handleCompletionOrchestrate },
     batchCompletionStream: {
         type: 'stream',
         handler: handleBatchCompletionStream

package/dist/server/rpc/handlers/completion-orchestrate.d.ts

@@ -0,0 +1,42 @@
+import type { CompletionOrchestrateRequest, CompletionOrchestrateResponse, CompletionStreamRequest, CompletionStreamResponse } from '../../../schemas/index';
+import { type ToolCallbackResult } from '../../../schemas/completion-stream';
+/**
+ * Reads newline-delimited JSON tool results off the duplex request stream
+ * and resolves per-callId waiters. Chunk boundaries don't align with lines,
+ * so a partial tail is buffered until its newline arrives.
+ *
+ * The pump is torn down by `close()` (called from a `finally` when
+ * orchestration ends, however it ends) so the upstream iterator is returned
+ * and the reader stops holding the request stream open.
+ */
+export declare class ToolResultReader {
+    private buffer;
+    private readonly results;
+    private readonly waiters;
+    private failed;
+    private closed;
+    private readonly iterator;
+    constructor(inputStream: AsyncIterable<Buffer>);
+    private pump;
+    private failWaiters;
+    private accept;
+    waitFor(callId: string, timeoutMs?: number): Promise<ToolCallbackResult>;
+    /**
+     * Stop the pump and release the upstream iterator. Idempotent. Pending
+     * waiters are failed so a caller blocked in `waitFor` can't hang after the
+     * stream is torn down.
+     */
+    close(): void;
+}
+type RunTurn = (request: CompletionStreamRequest) => AsyncGenerator<CompletionStreamResponse, void, unknown>;
+/**
+ * The orchestration core, pure over its collaborators so it unit-tests
+ * without the plugin registry: runs completion turns, forwards their events
+ * downstream, and when a turn requests tool calls, emits `toolCallback`
+ * frames and blocks on the client's results before starting the next turn
+ * with the extended history.
+ */
+export declare function orchestrateCompletion(request: CompletionOrchestrateRequest, runTurn: RunTurn, reader: Pick<ToolResultReader, 'waitFor'>): AsyncGenerator<CompletionOrchestrateResponse>;
+export declare function handleCompletionOrchestrate(request: CompletionOrchestrateRequest, inputStream: AsyncIterable<Buffer>): AsyncGenerator<CompletionOrchestrateResponse>;
+export {};
+//# sourceMappingURL=completion-orchestrate.d.ts.map