@qvac/sdk 0.8.3 → 0.9.0

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

@@ -121,6 +121,7 @@ export declare const nmtConfigBaseSchema: z.ZodDiscriminatedUnion<[z.ZodObject<{
             "onnx-tts": "onnx-tts";
             "parakeet-transcription": "parakeet-transcription";
             "onnx-ocr": "onnx-ocr";
+            "sdcpp-generation": "sdcpp-generation";
             llm: "llm";
             whisper: "whisper";
             embeddings: "embeddings";
@@ -128,6 +129,7 @@ export declare const nmtConfigBaseSchema: z.ZodDiscriminatedUnion<[z.ZodObject<{
             parakeet: "parakeet";
             tts: "tts";
             ocr: "ocr";
+            diffusion: "diffusion";
         }>, z.ZodLiteral<"vad">]>>;
     }, z.core.$strip>]>>;
     dstVocabSrc: z.ZodOptional<z.ZodUnion<readonly [z.ZodString, z.ZodObject<{
@@ -149,6 +151,7 @@ export declare const nmtConfigBaseSchema: z.ZodDiscriminatedUnion<[z.ZodObject<{
             "onnx-tts": "onnx-tts";
             "parakeet-transcription": "parakeet-transcription";
             "onnx-ocr": "onnx-ocr";
+            "sdcpp-generation": "sdcpp-generation";
             llm: "llm";
             whisper: "whisper";
             embeddings: "embeddings";
@@ -156,6 +159,7 @@ export declare const nmtConfigBaseSchema: z.ZodDiscriminatedUnion<[z.ZodObject<{
             parakeet: "parakeet";
             tts: "tts";
             ocr: "ocr";
+            diffusion: "diffusion";
         }>, z.ZodLiteral<"vad">]>>;
     }, z.core.$strip>]>>;
     normalize: z.ZodOptional<z.ZodNumber>;
@@ -190,6 +194,7 @@ export declare const nmtConfigBaseSchema: z.ZodDiscriminatedUnion<[z.ZodObject<{
                 "onnx-tts": "onnx-tts";
                 "parakeet-transcription": "parakeet-transcription";
                 "onnx-ocr": "onnx-ocr";
+                "sdcpp-generation": "sdcpp-generation";
                 llm: "llm";
                 whisper: "whisper";
                 embeddings: "embeddings";
@@ -197,6 +202,7 @@ export declare const nmtConfigBaseSchema: z.ZodDiscriminatedUnion<[z.ZodObject<{
                 parakeet: "parakeet";
                 tts: "tts";
                 ocr: "ocr";
+                diffusion: "diffusion";
             }>, z.ZodLiteral<"vad">]>>;
         }, z.core.$strip>]>;
         srcVocabSrc: z.ZodOptional<z.ZodUnion<readonly [z.ZodString, z.ZodObject<{
@@ -218,6 +224,7 @@ export declare const nmtConfigBaseSchema: z.ZodDiscriminatedUnion<[z.ZodObject<{
                 "onnx-tts": "onnx-tts";
                 "parakeet-transcription": "parakeet-transcription";
                 "onnx-ocr": "onnx-ocr";
+                "sdcpp-generation": "sdcpp-generation";
                 llm: "llm";
                 whisper: "whisper";
                 embeddings: "embeddings";
@@ -225,6 +232,7 @@ export declare const nmtConfigBaseSchema: z.ZodDiscriminatedUnion<[z.ZodObject<{
                 parakeet: "parakeet";
                 tts: "tts";
                 ocr: "ocr";
+                diffusion: "diffusion";
             }>, z.ZodLiteral<"vad">]>>;
         }, z.core.$strip>]>>;
         dstVocabSrc: z.ZodOptional<z.ZodUnion<readonly [z.ZodString, z.ZodObject<{
@@ -246,6 +254,7 @@ export declare const nmtConfigBaseSchema: z.ZodDiscriminatedUnion<[z.ZodObject<{
                 "onnx-tts": "onnx-tts";
                 "parakeet-transcription": "parakeet-transcription";
                 "onnx-ocr": "onnx-ocr";
+                "sdcpp-generation": "sdcpp-generation";
                 llm: "llm";
                 whisper: "whisper";
                 embeddings: "embeddings";
@@ -253,6 +262,7 @@ export declare const nmtConfigBaseSchema: z.ZodDiscriminatedUnion<[z.ZodObject<{
                 parakeet: "parakeet";
                 tts: "tts";
                 ocr: "ocr";
+                diffusion: "diffusion";
             }>, z.ZodLiteral<"vad">]>>;
         }, z.core.$strip>]>>;
         normalize: z.ZodOptional<z.ZodNumber>;
@@ -442,6 +452,7 @@ export declare const nmtConfigSchema: z.ZodPipe<z.ZodDiscriminatedUnion<[z.ZodOb
             "onnx-tts": "onnx-tts";
             "parakeet-transcription": "parakeet-transcription";
             "onnx-ocr": "onnx-ocr";
+            "sdcpp-generation": "sdcpp-generation";
             llm: "llm";
             whisper: "whisper";
             embeddings: "embeddings";
@@ -449,6 +460,7 @@ export declare const nmtConfigSchema: z.ZodPipe<z.ZodDiscriminatedUnion<[z.ZodOb
             parakeet: "parakeet";
             tts: "tts";
             ocr: "ocr";
+            diffusion: "diffusion";
         }>, z.ZodLiteral<"vad">]>>;
     }, z.core.$strip>]>>;
     dstVocabSrc: z.ZodOptional<z.ZodUnion<readonly [z.ZodString, z.ZodObject<{
@@ -470,6 +482,7 @@ export declare const nmtConfigSchema: z.ZodPipe<z.ZodDiscriminatedUnion<[z.ZodOb
             "onnx-tts": "onnx-tts";
             "parakeet-transcription": "parakeet-transcription";
             "onnx-ocr": "onnx-ocr";
+            "sdcpp-generation": "sdcpp-generation";
             llm: "llm";
             whisper: "whisper";
             embeddings: "embeddings";
@@ -477,6 +490,7 @@ export declare const nmtConfigSchema: z.ZodPipe<z.ZodDiscriminatedUnion<[z.ZodOb
             parakeet: "parakeet";
             tts: "tts";
             ocr: "ocr";
+            diffusion: "diffusion";
         }>, z.ZodLiteral<"vad">]>>;
     }, z.core.$strip>]>>;
     normalize: z.ZodOptional<z.ZodNumber>;
@@ -511,6 +525,7 @@ export declare const nmtConfigSchema: z.ZodPipe<z.ZodDiscriminatedUnion<[z.ZodOb
                 "onnx-tts": "onnx-tts";
                 "parakeet-transcription": "parakeet-transcription";
                 "onnx-ocr": "onnx-ocr";
+                "sdcpp-generation": "sdcpp-generation";
                 llm: "llm";
                 whisper: "whisper";
                 embeddings: "embeddings";
@@ -518,6 +533,7 @@ export declare const nmtConfigSchema: z.ZodPipe<z.ZodDiscriminatedUnion<[z.ZodOb
                 parakeet: "parakeet";
                 tts: "tts";
                 ocr: "ocr";
+                diffusion: "diffusion";
             }>, z.ZodLiteral<"vad">]>>;
         }, z.core.$strip>]>;
         srcVocabSrc: z.ZodOptional<z.ZodUnion<readonly [z.ZodString, z.ZodObject<{
@@ -539,6 +555,7 @@ export declare const nmtConfigSchema: z.ZodPipe<z.ZodDiscriminatedUnion<[z.ZodOb
                 "onnx-tts": "onnx-tts";
                 "parakeet-transcription": "parakeet-transcription";
                 "onnx-ocr": "onnx-ocr";
+                "sdcpp-generation": "sdcpp-generation";
                 llm: "llm";
                 whisper: "whisper";
                 embeddings: "embeddings";
@@ -546,6 +563,7 @@ export declare const nmtConfigSchema: z.ZodPipe<z.ZodDiscriminatedUnion<[z.ZodOb
                 parakeet: "parakeet";
                 tts: "tts";
                 ocr: "ocr";
+                diffusion: "diffusion";
             }>, z.ZodLiteral<"vad">]>>;
         }, z.core.$strip>]>>;
         dstVocabSrc: z.ZodOptional<z.ZodUnion<readonly [z.ZodString, z.ZodObject<{
@@ -567,6 +585,7 @@ export declare const nmtConfigSchema: z.ZodPipe<z.ZodDiscriminatedUnion<[z.ZodOb
                 "onnx-tts": "onnx-tts";
                 "parakeet-transcription": "parakeet-transcription";
                 "onnx-ocr": "onnx-ocr";
+                "sdcpp-generation": "sdcpp-generation";
                 llm: "llm";
                 whisper: "whisper";
                 embeddings: "embeddings";
@@ -574,6 +593,7 @@ export declare const nmtConfigSchema: z.ZodPipe<z.ZodDiscriminatedUnion<[z.ZodOb
                 parakeet: "parakeet";
                 tts: "tts";
                 ocr: "ocr";
+                diffusion: "diffusion";
             }>, z.ZodLiteral<"vad">]>>;
         }, z.core.$strip>]>>;
         normalize: z.ZodOptional<z.ZodNumber>;
@@ -684,7 +704,7 @@ export declare const nmtConfigSchema: z.ZodPipe<z.ZodDiscriminatedUnion<[z.ZodOb
         engine?: string | undefined;
         expectedSize?: number | undefined;
         sha256Checksum?: string | undefined;
-        addon?: "llamacpp-completion" | "whispercpp-transcription" | "llamacpp-embedding" | "nmtcpp-translation" | "onnx-tts" | "parakeet-transcription" | "onnx-ocr" | "llm" | "whisper" | "embeddings" | "nmt" | "parakeet" | "tts" | "ocr" | "vad" | undefined;
+        addon?: "llamacpp-completion" | "whispercpp-transcription" | "llamacpp-embedding" | "nmtcpp-translation" | "onnx-tts" | "parakeet-transcription" | "onnx-ocr" | "sdcpp-generation" | "llm" | "whisper" | "embeddings" | "nmt" | "parakeet" | "tts" | "ocr" | "diffusion" | "vad" | undefined;
     } | undefined;
     dstVocabSrc?: string | {
         src: string;
@@ -697,7 +717,7 @@ export declare const nmtConfigSchema: z.ZodPipe<z.ZodDiscriminatedUnion<[z.ZodOb
         engine?: string | undefined;
         expectedSize?: number | undefined;
         sha256Checksum?: string | undefined;
-        addon?: "llamacpp-completion" | "whispercpp-transcription" | "llamacpp-embedding" | "nmtcpp-translation" | "onnx-tts" | "parakeet-transcription" | "onnx-ocr" | "llm" | "whisper" | "embeddings" | "nmt" | "parakeet" | "tts" | "ocr" | "vad" | undefined;
+        addon?: "llamacpp-completion" | "whispercpp-transcription" | "llamacpp-embedding" | "nmtcpp-translation" | "onnx-tts" | "parakeet-transcription" | "onnx-ocr" | "sdcpp-generation" | "llm" | "whisper" | "embeddings" | "nmt" | "parakeet" | "tts" | "ocr" | "diffusion" | "vad" | undefined;
     } | undefined;
     normalize?: number | undefined;
     pivotModel?: {
@@ -712,7 +732,7 @@ export declare const nmtConfigSchema: z.ZodPipe<z.ZodDiscriminatedUnion<[z.ZodOb
             engine?: string | undefined;
             expectedSize?: number | undefined;
             sha256Checksum?: string | undefined;
-            addon?: "llamacpp-completion" | "whispercpp-transcription" | "llamacpp-embedding" | "nmtcpp-translation" | "onnx-tts" | "parakeet-transcription" | "onnx-ocr" | "llm" | "whisper" | "embeddings" | "nmt" | "parakeet" | "tts" | "ocr" | "vad" | undefined;
+            addon?: "llamacpp-completion" | "whispercpp-transcription" | "llamacpp-embedding" | "nmtcpp-translation" | "onnx-tts" | "parakeet-transcription" | "onnx-ocr" | "sdcpp-generation" | "llm" | "whisper" | "embeddings" | "nmt" | "parakeet" | "tts" | "ocr" | "diffusion" | "vad" | undefined;
         };
         mode?: "full" | undefined;
         beamsize?: number | undefined;
@@ -734,7 +754,7 @@ export declare const nmtConfigSchema: z.ZodPipe<z.ZodDiscriminatedUnion<[z.ZodOb
             engine?: string | undefined;
             expectedSize?: number | undefined;
             sha256Checksum?: string | undefined;
-            addon?: "llamacpp-completion" | "whispercpp-transcription" | "llamacpp-embedding" | "nmtcpp-translation" | "onnx-tts" | "parakeet-transcription" | "onnx-ocr" | "llm" | "whisper" | "embeddings" | "nmt" | "parakeet" | "tts" | "ocr" | "vad" | undefined;
+            addon?: "llamacpp-completion" | "whispercpp-transcription" | "llamacpp-embedding" | "nmtcpp-translation" | "onnx-tts" | "parakeet-transcription" | "onnx-ocr" | "sdcpp-generation" | "llm" | "whisper" | "embeddings" | "nmt" | "parakeet" | "tts" | "ocr" | "diffusion" | "vad" | undefined;
         } | undefined;
         dstVocabSrc?: string | {
             src: string;
@@ -747,7 +767,7 @@ export declare const nmtConfigSchema: z.ZodPipe<z.ZodDiscriminatedUnion<[z.ZodOb
             engine?: string | undefined;
             expectedSize?: number | undefined;
             sha256Checksum?: string | undefined;
-            addon?: "llamacpp-completion" | "whispercpp-transcription" | "llamacpp-embedding" | "nmtcpp-translation" | "onnx-tts" | "parakeet-transcription" | "onnx-ocr" | "llm" | "whisper" | "embeddings" | "nmt" | "parakeet" | "tts" | "ocr" | "vad" | undefined;
+            addon?: "llamacpp-completion" | "whispercpp-transcription" | "llamacpp-embedding" | "nmtcpp-translation" | "onnx-tts" | "parakeet-transcription" | "onnx-ocr" | "sdcpp-generation" | "llm" | "whisper" | "embeddings" | "nmt" | "parakeet" | "tts" | "ocr" | "diffusion" | "vad" | undefined;
         } | undefined;
         normalize?: number | undefined;
     } | undefined;
@@ -801,7 +821,7 @@ export declare const nmtConfigSchema: z.ZodPipe<z.ZodDiscriminatedUnion<[z.ZodOb
         engine?: string | undefined;
         expectedSize?: number | undefined;
         sha256Checksum?: string | undefined;
-        addon?: "llamacpp-completion" | "whispercpp-transcription" | "llamacpp-embedding" | "nmtcpp-translation" | "onnx-tts" | "parakeet-transcription" | "onnx-ocr" | "llm" | "whisper" | "embeddings" | "nmt" | "parakeet" | "tts" | "ocr" | "vad" | undefined;
+        addon?: "llamacpp-completion" | "whispercpp-transcription" | "llamacpp-embedding" | "nmtcpp-translation" | "onnx-tts" | "parakeet-transcription" | "onnx-ocr" | "sdcpp-generation" | "llm" | "whisper" | "embeddings" | "nmt" | "parakeet" | "tts" | "ocr" | "diffusion" | "vad" | undefined;
     } | undefined;
     dstVocabSrc?: string | {
         src: string;
@@ -814,7 +834,7 @@ export declare const nmtConfigSchema: z.ZodPipe<z.ZodDiscriminatedUnion<[z.ZodOb
         engine?: string | undefined;
         expectedSize?: number | undefined;
         sha256Checksum?: string | undefined;
-        addon?: "llamacpp-completion" | "whispercpp-transcription" | "llamacpp-embedding" | "nmtcpp-translation" | "onnx-tts" | "parakeet-transcription" | "onnx-ocr" | "llm" | "whisper" | "embeddings" | "nmt" | "parakeet" | "tts" | "ocr" | "vad" | undefined;
+        addon?: "llamacpp-completion" | "whispercpp-transcription" | "llamacpp-embedding" | "nmtcpp-translation" | "onnx-tts" | "parakeet-transcription" | "onnx-ocr" | "sdcpp-generation" | "llm" | "whisper" | "embeddings" | "nmt" | "parakeet" | "tts" | "ocr" | "diffusion" | "vad" | undefined;
     } | undefined;
     normalize?: number | undefined;
     pivotModel?: {
@@ -829,7 +849,7 @@ export declare const nmtConfigSchema: z.ZodPipe<z.ZodDiscriminatedUnion<[z.ZodOb
             engine?: string | undefined;
             expectedSize?: number | undefined;
             sha256Checksum?: string | undefined;
-            addon?: "llamacpp-completion" | "whispercpp-transcription" | "llamacpp-embedding" | "nmtcpp-translation" | "onnx-tts" | "parakeet-transcription" | "onnx-ocr" | "llm" | "whisper" | "embeddings" | "nmt" | "parakeet" | "tts" | "ocr" | "vad" | undefined;
+            addon?: "llamacpp-completion" | "whispercpp-transcription" | "llamacpp-embedding" | "nmtcpp-translation" | "onnx-tts" | "parakeet-transcription" | "onnx-ocr" | "sdcpp-generation" | "llm" | "whisper" | "embeddings" | "nmt" | "parakeet" | "tts" | "ocr" | "diffusion" | "vad" | undefined;
         };
         mode?: "full" | undefined;
         beamsize?: number | undefined;
@@ -851,7 +871,7 @@ export declare const nmtConfigSchema: z.ZodPipe<z.ZodDiscriminatedUnion<[z.ZodOb
             engine?: string | undefined;
             expectedSize?: number | undefined;
             sha256Checksum?: string | undefined;
-            addon?: "llamacpp-completion" | "whispercpp-transcription" | "llamacpp-embedding" | "nmtcpp-translation" | "onnx-tts" | "parakeet-transcription" | "onnx-ocr" | "llm" | "whisper" | "embeddings" | "nmt" | "parakeet" | "tts" | "ocr" | "vad" | undefined;
+            addon?: "llamacpp-completion" | "whispercpp-transcription" | "llamacpp-embedding" | "nmtcpp-translation" | "onnx-tts" | "parakeet-transcription" | "onnx-ocr" | "sdcpp-generation" | "llm" | "whisper" | "embeddings" | "nmt" | "parakeet" | "tts" | "ocr" | "diffusion" | "vad" | undefined;
         } | undefined;
         dstVocabSrc?: string | {
             src: string;
@@ -864,7 +884,7 @@ export declare const nmtConfigSchema: z.ZodPipe<z.ZodDiscriminatedUnion<[z.ZodOb
             engine?: string | undefined;
             expectedSize?: number | undefined;
             sha256Checksum?: string | undefined;
-            addon?: "llamacpp-completion" | "whispercpp-transcription" | "llamacpp-embedding" | "nmtcpp-translation" | "onnx-tts" | "parakeet-transcription" | "onnx-ocr" | "llm" | "whisper" | "embeddings" | "nmt" | "parakeet" | "tts" | "ocr" | "vad" | undefined;
+            addon?: "llamacpp-completion" | "whispercpp-transcription" | "llamacpp-embedding" | "nmtcpp-translation" | "onnx-tts" | "parakeet-transcription" | "onnx-ocr" | "sdcpp-generation" | "llm" | "whisper" | "embeddings" | "nmt" | "parakeet" | "tts" | "ocr" | "diffusion" | "vad" | undefined;
         } | undefined;
         normalize?: number | undefined;
     } | undefined;

package/dist/server/bare/delegate-rpc-client.js

@@ -5,6 +5,7 @@ import { DelegateConnectionFailedError } from "../../utils/errors-server.js";
 import { getServerLogger } from "../../logging/index.js";
 import { nowMs } from "../../profiling/index.js";
 import { cacheDelegationConnectionTime, clearPeerConnectionTracking, } from "../rpc/profiling/delegation-profiler.js";
+import { getNextCommandId } from "../rpc/rpc-utils.js";
 const logger = getServerLogger();
 // Map of active RPC instances by public key
 const activeRPCs = new Map();
@@ -13,7 +14,6 @@ const activeConnections = new Map();
 // Track whether the global connection handler has been registered
 let connectionHandlerRegistered = false;
 const HEALTH_CHECK_TIMEOUT_MS = 1500;
-import { getNextCommandId } from "../rpc/delegate-utils.js";
 function isHeartbeatResponse(payload) {
     return (typeof payload === "object" &&
         payload !== null &&
@@ -75,13 +75,13 @@ async function closeConnection(publicKey) {
         // Wait for the close event before returning so Hyperswarm's internal
         // _allConnections is cleaned up before we attempt to rejoin/flush.
         await new Promise((resolve) => {
-            const timeout = setTimeout(resolve, 5000);
             existingConnection.on("close", () => {
-                clearTimeout(timeout);
                 resolve();
             });
             existingConnection.destroy();
         });
+        if (activeConnections.get(publicKey) !== existingConnection)
+            return;
         activeConnections.delete(publicKey);
         activeRPCs.delete(publicKey);
         clearPeerConnectionTracking(publicKey);

package/dist/server/bare/hyperswarm.js

@@ -3,6 +3,7 @@ import Hyperswarm from "hyperswarm";
 import crypto from "bare-crypto";
 import { envSchema } from "../../schemas/provide.js";
 import { getSDKConfig } from "./registry/config-registry.js";
+import { registerSwarm, unregisterSwarm } from "./runtime-lifecycle.js";
 import { getServerLogger } from "../../logging/index.js";
 const logger = getServerLogger();
 function getHyperswarmSeedBuffer() {
@@ -59,13 +60,14 @@ function createSwarm(firewallConfig) {
     const swarmOptions = { seed, firewall, relayThrough: getRelays };
     return new Hyperswarm(swarmOptions);
 }
-let swarm;
+let swarm = null;
 const activeProviderTopics = new Set();
 export function getSwarm({ firewallConfig, } = {}) {
     if (swarm) {
         return swarm;
     }
     swarm = createSwarm(firewallConfig);
+    registerSwarm(swarm, { label: "shared-swarm", createdAt: Date.now() });
     return swarm;
 }
 export function registerProviderTopic(topic) {
@@ -78,8 +80,18 @@ export function hasActiveProviders() {
     return activeProviderTopics.size > 0;
 }
 export async function destroySwarm() {
-    if (swarm) {
+    if (!swarm)
+        return;
+    const ref = swarm;
+    swarm = null;
+    try {
+        await ref.destroy();
+        unregisterSwarm(ref);
         activeProviderTopics.clear();
-        return swarm.destroy();
+    }
+    catch (error) {
+        if (swarm === null)
+            swarm = ref;
+        throw error;
     }
 }

package/dist/server/bare/ops/embed.d.ts

@@ -1,12 +1,16 @@
-import { type EmbedParams } from "../../../schemas/index";
+import { type EmbedParams, type EmbedStats } from "../../../schemas/index";
+export interface EmbedResult {
+    embedding: number[] | number[][];
+    stats?: EmbedStats;
+}
 export declare function embed(params: {
     modelId: string;
     text: string;
-}): Promise<number[]>;
+}): Promise<EmbedResult>;
 export declare function embed(params: {
     modelId: string;
     text: string[];
-}): Promise<number[][]>;
-export declare function embed(params: EmbedParams): Promise<number[] | number[][]>;
+}): Promise<EmbedResult>;
+export declare function embed(params: EmbedParams): Promise<EmbedResult>;
 export declare function normalizeVector(vector: Float32Array): number[];
 //# sourceMappingURL=embed.d.ts.map

package/dist/server/bare/ops/embed.js

@@ -1,29 +1,40 @@
 import { getModel } from "../registry/model-registry.js";
 import { embedParamsSchema } from "../../../schemas/index.js";
+import { buildUnaryResult } from "../../../profiling/model-execution.js";
 import { EmbedNoEmbeddingsError, EmbedFailedError, } from "../../../utils/errors-server.js";
+import { nowMs } from "../../../profiling/index.js";
 export async function embed(params) {
     const { modelId, text } = embedParamsSchema.parse(params);
     const model = getModel(modelId);
-    const response = await model.run(text);
-    const rawEmbeddings = (await response.await());
+    const modelStart = nowMs();
+    const response = (await model.run(text));
+    const rawEmbeddings = await response.await();
+    const modelExecutionMs = nowMs() - modelStart;
+    const stats = {
+        ...(response.stats?.total_time_ms !== undefined && { totalTime: response.stats.total_time_ms }),
+        ...(response.stats?.tokens_per_second !== undefined && { tokensPerSecond: response.stats.tokens_per_second }),
+        ...(response.stats?.total_tokens !== undefined && { totalTokens: response.stats.total_tokens }),
+        ...(response.stats?.backendDevice !== undefined && { backendDevice: response.stats.backendDevice }),
+    };
     const embeddingsArray = rawEmbeddings[0];
     if (Array.isArray(text)) {
         if (!embeddingsArray || embeddingsArray.length === 0) {
             throw new EmbedNoEmbeddingsError();
         }
-        return embeddingsArray.map((embeddingVector) => {
+        const embedding = embeddingsArray.map((embeddingVector) => {
             if (!embeddingVector || embeddingVector.length === 0) {
                 throw new EmbedNoEmbeddingsError();
             }
             return normalizeVector(embeddingVector);
         });
+        return buildUnaryResult({ embedding }, modelExecutionMs, stats);
     }
     else {
         const embeddingVector = embeddingsArray?.[0];
         if (!embeddingVector || embeddingVector.length === 0) {
             throw new EmbedNoEmbeddingsError();
         }
-        return normalizeVector(embeddingVector);
+        return buildUnaryResult({ embedding: normalizeVector(embeddingVector) }, modelExecutionMs, stats);
     }
 }
 export function normalizeVector(vector) {

package/dist/server/bare/ops/transcribe.d.ts

@@ -1,3 +1,7 @@
-import { type TranscribeParams } from "../../../schemas/index";
-export declare function transcribe(params: TranscribeParams): AsyncGenerator<string, void, void>;
+import { type TranscribeParams, type TranscribeStats } from "../../../schemas/index";
+export declare function transcribe(params: TranscribeParams): AsyncGenerator<string, {
+    modelExecutionMs: number;
+    stats?: TranscribeStats;
+}, void>;
+export declare function transcribeStream(modelId: string, audioInputStream: AsyncIterable<Buffer>, prompt?: string): AsyncGenerator<string, void, void>;
 //# sourceMappingURL=transcribe.d.ts.map

package/dist/server/bare/ops/transcribe.js

@@ -2,7 +2,13 @@ import { getModel, getModelConfig, getModelEntry, } from "../registry/model-regi
 import { ModelType, } from "../../../schemas/index.js";
 import { createAudioStream } from "../utils/audio-input.js";
 import { getServerLogger } from "../../../logging/index.js";
+import { TranscriptionFailedError } from "../../../utils/errors-server.js";
+import { nowMs } from "../../../profiling/index.js";
+import { buildStreamResult } from "../../../profiling/model-execution.js";
 const logger = getServerLogger();
+function hasRunStreaming(model) {
+    return "runStreaming" in model && typeof model.runStreaming === "function";
+}
 const SILENCE_MARKERS = {
     [ModelType.whispercppTranscription]: "[BLANK_AUDIO]",
     [ModelType.parakeetTranscription]: "[No speech detected]",
@@ -52,10 +58,13 @@ export async function* transcribe(params) {
     const silenceMarker = SILENCE_MARKERS[engineType] ?? "";
     const audioFormat = getAudioFormat(modelId, engineType);
     const originalConfig = await applyPrompt(modelId, params.prompt, engineType);
+    let modelExecutionMs = 0;
+    let response;
     try {
         const model = getModel(modelId);
         const audioStream = await createAudioStream(params.audioChunk, audioFormat);
-        const response = await model.run(audioStream);
+        const modelStart = nowMs();
+        response = (await model.run(audioStream));
         for await (const output of response.iterate()) {
             logger.debug("Streaming Transcription Update:", output);
             const text = output
@@ -66,6 +75,49 @@ export async function* transcribe(params) {
                 yield text;
             }
         }
+        modelExecutionMs = nowMs() - modelStart;
+    }
+    finally {
+        if (originalConfig) {
+            await restorePrompt(modelId, originalConfig);
+        }
+    }
+    const stats = {
+        ...(response?.stats?.audioDurationMs !== undefined && { audioDuration: response.stats.audioDurationMs }),
+        ...(response?.stats?.realTimeFactor !== undefined && { realTimeFactor: response.stats.realTimeFactor }),
+        ...(response?.stats?.tokensPerSecond !== undefined && { tokensPerSecond: response.stats.tokensPerSecond }),
+        ...(response?.stats?.totalTokens !== undefined && { totalTokens: response.stats.totalTokens }),
+        ...(response?.stats?.totalSegments !== undefined && { totalSegments: response.stats.totalSegments }),
+        ...(response?.stats?.whisperEncodeMs !== undefined && { whisperEncodeTime: response.stats.whisperEncodeMs }),
+        ...(response?.stats?.whisperDecodeMs !== undefined && { whisperDecodeTime: response.stats.whisperDecodeMs }),
+        ...(response?.stats?.encoderMs !== undefined && { encoderTime: response.stats.encoderMs }),
+        ...(response?.stats?.decoderMs !== undefined && { decoderTime: response.stats.decoderMs }),
+        ...(response?.stats?.melSpecMs !== undefined && { melSpecTime: response.stats.melSpecMs }),
+    };
+    return buildStreamResult(modelExecutionMs, stats);
+}
+export async function* transcribeStream(modelId, audioInputStream, prompt) {
+    const engineType = getEngineModelType(modelId);
+    const silenceMarker = SILENCE_MARKERS[engineType] ?? "";
+    const originalConfig = await applyPrompt(modelId, prompt, engineType);
+    try {
+        const model = getModel(modelId);
+        if (!hasRunStreaming(model)) {
+            throw new TranscriptionFailedError(`Model ${modelId} does not support streaming transcription`);
+        }
+        const response = await model.runStreaming(audioInputStream);
+        for await (const segments of response.iterate()) {
+            logger.debug("Live Transcription Update:", segments);
+            for (const segment of segments) {
+                if (!segment.text)
+                    continue;
+                if (silenceMarker && segment.text.includes(silenceMarker))
+                    continue;
+                if (segment.text.trim()) {
+                    yield segment.text;
+                }
+            }
+        }
     }
     finally {
         if (originalConfig) {

package/dist/server/bare/ops/translate.d.ts

@@ -1,5 +1,8 @@
 import { type TranslateParams, type TranslationStats } from "../../../schemas/index";
 export declare function getLanguage(code: string | undefined): string;
 export declare function isAfrican(code: string | undefined): boolean;
-export declare function translate(params: TranslateParams): AsyncGenerator<string, TranslationStats | undefined, unknown>;
+export declare function translate(params: TranslateParams): AsyncGenerator<string, {
+    modelExecutionMs: number;
+    stats?: TranslationStats;
+}, unknown>;
 //# sourceMappingURL=translate.d.ts.map

package/dist/server/bare/ops/translate.js

@@ -1,6 +1,8 @@
 import { getModel } from "../registry/model-registry.js";
 import { translateServerParamsSchema, normalizeModelType, ModelType, AFRICAN_LANGUAGES_MAP, } from "../../../schemas/index.js";
 import { getLangName } from "@qvac/langdetect-text";
+import { nowMs } from "../../../profiling/index.js";
+import { buildStreamResult } from "../../../profiling/model-execution.js";
 export function getLanguage(code) {
     if (!code)
         return "";
@@ -24,27 +26,22 @@ export async function* translate(params) {
     const model = getModel(modelId);
     const fromLanguage = getLanguage(from);
     const toLanguage = getLanguage(to);
-    const startTime = Date.now();
-    let processedTokens = 0;
     // Check if input is an array and model type is NMT
     if (Array.isArray(text) &&
         canonicalModelType === ModelType.nmtcppTranslation) {
         // Use runBatch for batch processing
+        const modelStart = nowMs();
         const translations = await model.runBatch(text);
+        const modelExecutionMs = nowMs() - modelStart;
         // Yield each translation with a newline separator
         for (let i = 0; i < translations.length; i++) {
             const translation = translations[i];
-            processedTokens++;
             yield translation;
             if (i < translations.length - 1) {
                 yield "\n";
             }
         }
-        const endTime = Date.now();
-        return {
-            processedTokens,
-            processingTime: endTime - startTime,
-        };
+        return { modelExecutionMs };
     }
     // Single text processing (for NMT or LLM)
     const singleText = Array.isArray(text) ? text[0] : text;
@@ -59,70 +56,36 @@ export async function* translate(params) {
                     : `${context ? `${context}. ` : ""}Translate the following text from ${fromLanguage} into ${toLanguage}. Only output the translation, nothing else.\n\n${fromLanguage}: ${singleText}\n${toLanguage}:`,
             },
         ];
+    const modelStart = nowMs();
     const response = await model.run(input);
     // Check if the response has an iterate method (like LLM models)
     if (canonicalModelType === ModelType.llamacppCompletion &&
         typeof response.iterate === "function") {
-        for await (const token of response.iterate()) {
-            processedTokens++;
+        const llmResponse = response;
+        for await (const token of llmResponse.iterate()) {
             yield token;
         }
-    }
-    else {
-        // For models that don't support iterate, create an async iterator using onUpdate
-        const tokenQueue = [];
-        let isComplete = false;
-        let resolveNext = null;
-        // Start the response processing
-        const responsePromise = response
-            .onUpdate((data) => {
-            processedTokens++;
-            if (resolveNext) {
-                // If there's a pending read, resolve it immediately
-                resolveNext({ value: data, done: false });
-                resolveNext = null;
-            }
-            else {
-                // Otherwise, queue the token
-                tokenQueue.push(data);
-            }
-        })
-            .await()
-            .then(() => {
-            isComplete = true;
-            if (resolveNext) {
-                resolveNext({ value: undefined, done: true });
-                resolveNext = null;
-            }
-        });
-        // Create an async iterator
-        const asyncIterator = {
-            async next() {
-                if (tokenQueue.length > 0) {
-                    return { value: tokenQueue.shift(), done: false };
-                }
-                if (isComplete) {
-                    return { value: undefined, done: true };
-                }
-                // Wait for the next token
-                return new Promise((resolve) => {
-                    resolveNext = resolve;
-                });
-            },
-            [Symbol.asyncIterator]() {
-                return this;
-            },
+        const modelExecutionMs = nowMs() - modelStart;
+        const stats = {
+            ...(llmResponse.stats?.TPS !== undefined && { tokensPerSecond: llmResponse.stats.TPS }),
+            ...(llmResponse.stats?.TTFT !== undefined && { timeToFirstToken: llmResponse.stats.TTFT }),
+            ...(llmResponse.stats?.CacheTokens !== undefined && { cacheTokens: llmResponse.stats.CacheTokens }),
+            ...(llmResponse.stats?.generatedTokens !== undefined && { totalTokens: llmResponse.stats.generatedTokens }),
         };
-        // Yield tokens as they come
-        for await (const token of asyncIterator) {
-            yield token;
-        }
-        // Ensure the response is fully processed
-        await responsePromise;
+        return buildStreamResult(modelExecutionMs, stats);
+    }
+    const nmtResponse = response;
+    for await (const token of nmtResponse.iterate()) {
+        yield token;
     }
-    const endTime = Date.now();
-    return {
-        processedTokens,
-        processingTime: endTime - startTime,
+    const modelExecutionMs = nowMs() - modelStart;
+    const stats = {
+        ...(nmtResponse.stats?.totalTime !== undefined && { totalTime: nmtResponse.stats.totalTime }),
+        ...(nmtResponse.stats?.totalTokens !== undefined && { totalTokens: nmtResponse.stats.totalTokens }),
+        ...(nmtResponse.stats?.decodeTime !== undefined && { decodeTime: nmtResponse.stats.decodeTime }),
+        ...(nmtResponse.stats?.encodeTime !== undefined && { encodeTime: nmtResponse.stats.encodeTime }),
+        ...(nmtResponse.stats?.TPS !== undefined && { tokensPerSecond: nmtResponse.stats.TPS }),
+        ...(nmtResponse.stats?.TTFT !== undefined && { timeToFirstToken: nmtResponse.stats.TTFT }),
     };
+    return buildStreamResult(modelExecutionMs, stats);
 }

package/dist/server/bare/ops/unload-model.js

@@ -2,6 +2,7 @@ import { promises as fsPromises } from "bare-fs";
 import path from "bare-path";
 import { unregisterModel } from "../registry/model-registry.js";
 import { unregisterAllLoggingStreams } from "../registry/logging-stream-registry.js";
+import { clearFinetuneRuntimeState } from "../plugins/llamacpp-completion/ops/finetune.js";
 import { unregisterAddonLogger, getServerLogger } from "../../../logging/index.js";
 import { unloadModelParamsSchema } from "../../../schemas/index.js";
 import { ModelNotLoadedError } from "../../../utils/errors-server.js";
@@ -13,6 +14,7 @@ export async function unloadModel(params) {
     if (!entry) {
         throw new ModelNotLoadedError(modelId);
     }
+    clearFinetuneRuntimeState(modelId);
     if (entry.local?.loader) {
         await entry.local.loader.close();
     }

package/dist/server/bare/plugins/index.d.ts

@@ -5,4 +5,5 @@ export { parakeetPlugin } from "./parakeet-transcription/plugin";
 export { nmtPlugin } from "./nmtcpp-translation/plugin";
 export { ttsPlugin } from "./onnx-tts/plugin";
 export { ocrPlugin } from "./onnx-ocr/plugin";
+export { diffusionPlugin } from "./sdcpp-generation/plugin";
 //# sourceMappingURL=index.d.ts.map