@qvac/sdk 0.14.0 → 0.15.0

package/dist/schemas/common.d.ts

@@ -1,4 +1,4 @@
-import { z } from "zod";
+import { z } from 'zod';
 export declare const requestSchema: z.ZodUnion<readonly [z.ZodObject<{
     type: z.ZodLiteral<"heartbeat">;
     delegate: z.ZodOptional<z.ZodObject<{
@@ -39,23 +39,24 @@ export declare const requestSchema: z.ZodUnion<readonly [z.ZodObject<{
         repeat_penalty: z.ZodOptional<z.ZodNumber>;
         stop_sequences: z.ZodOptional<z.ZodArray<z.ZodString>>;
         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<{
+        'cache-type-k': z.ZodOptional<z.ZodString>;
+        'cache-type-v': z.ZodOptional<z.ZodString>;
+        'main-gpu': z.ZodOptional<z.ZodUnion<readonly [z.ZodNumber, z.ZodEnum<{
             integrated: "integrated";
             dedicated: "dedicated";
         }>]>>;
-        "split-mode": z.ZodOptional<z.ZodEnum<{
+        'split-mode': z.ZodOptional<z.ZodEnum<{
             none: "none";
             layer: "layer";
             row: "row";
         }>>;
-        "tensor-split": z.ZodOptional<z.ZodString>;
+        'tensor-split': z.ZodOptional<z.ZodString>;
         openclCacheDir: z.ZodOptional<z.ZodString>;
         reasoning_budget: z.ZodOptional<z.ZodNumber>;
         projectionModelSrc: z.ZodOptional<z.ZodUnion<readonly [z.ZodString, z.ZodObject<{
@@ -100,6 +101,7 @@ export declare const requestSchema: z.ZodUnion<readonly [z.ZodObject<{
             batched: "batched";
             sequential: "sequential";
         }>>;
+        'mmproj-use-gpu': z.ZodOptional<z.ZodBoolean>;
     }, z.core.$strip>;
 }, z.core.$strict>, z.ZodObject<{
     type: z.ZodLiteral<"loadModel">;
@@ -335,8 +337,8 @@ export declare const requestSchema: z.ZodUnion<readonly [z.ZodObject<{
     modelConfig: z.ZodObject<{
         gpuLayers: z.ZodOptional<z.ZodNumber>;
         device: z.ZodOptional<z.ZodEnum<{
-            cpu: "cpu";
             gpu: "gpu";
+            cpu: "cpu";
         }>>;
         batchSize: z.ZodOptional<z.ZodNumber>;
         pooling: z.ZodOptional<z.ZodEnum<{
@@ -1103,8 +1105,8 @@ export declare const requestSchema: z.ZodUnion<readonly [z.ZodObject<{
         }>>;
         threads: z.ZodOptional<z.ZodNumber>;
         device: z.ZodOptional<z.ZodEnum<{
-            cpu: "cpu";
             gpu: "gpu";
+            cpu: "cpu";
         }>>;
         prediction: z.ZodOptional<z.ZodEnum<{
             auto: "auto";
@@ -1537,13 +1539,14 @@ export declare const requestSchema: z.ZodUnion<readonly [z.ZodObject<{
         repeat_penalty?: number | undefined;
         stop_sequences?: string[] | undefined;
         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;
-        "split-mode"?: "none" | "layer" | "row" | undefined;
-        "tensor-split"?: string | undefined;
+        'cache-type-k'?: string | undefined;
+        'cache-type-v'?: string | undefined;
+        'main-gpu'?: number | "integrated" | "dedicated" | undefined;
+        'split-mode'?: "none" | "layer" | "row" | undefined;
+        'tensor-split'?: string | undefined;
         openclCacheDir?: string | undefined;
         reasoning_budget?: number | undefined;
         projectionModelSrc?: string | {
@@ -1560,6 +1563,7 @@ export declare const requestSchema: z.ZodUnion<readonly [z.ZodObject<{
             addon?: "embeddings" | "diffusion" | "tts" | "ocr" | "vla" | "llamacpp-completion" | "whispercpp-transcription" | "bci-whispercpp-transcription" | "llamacpp-embedding" | "nmtcpp-translation" | "onnx-tts" | "tts-ggml" | "parakeet-transcription" | "ggml-ocr" | "sdcpp-generation" | "ggml-vla" | "ggml-classification" | "llm" | "whisper" | "bci" | "nmt" | "parakeet" | "classification" | "vad" | undefined;
         } | undefined;
         image_tile_mode?: "disabled" | "batched" | "sequential" | undefined;
+        'mmproj-use-gpu'?: boolean | undefined;
     };
     modelName?: string | undefined;
     withProgress?: boolean | undefined;
@@ -1738,7 +1742,7 @@ export declare const requestSchema: z.ZodUnion<readonly [z.ZodObject<{
     modelType: "llamacpp-embedding";
     modelConfig: {
         gpuLayers?: number | undefined;
-        device?: "cpu" | "gpu" | undefined;
+        device?: "gpu" | "cpu" | undefined;
         batchSize?: number | undefined;
         pooling?: "none" | "mean" | "cls" | "last" | "rank" | undefined;
         attention?: "causal" | "non-causal" | undefined;
@@ -1955,7 +1959,7 @@ export declare const requestSchema: z.ZodUnion<readonly [z.ZodObject<{
     modelConfig?: {
         mode: "diffusion" | "upscale" | "video";
         threads?: number | undefined;
-        device?: "cpu" | "gpu" | undefined;
+        device?: "gpu" | "cpu" | undefined;
         prediction?: "auto" | "eps" | "v" | "edm_v" | "flow" | "flux2_flow" | undefined;
         type?: "auto" | "f32" | "f16" | "bf16" | "q2_k" | "q3_k" | "q4_0" | "q4_1" | "q4_k" | "q5_0" | "q5_1" | "q5_k" | "q6_k" | "q8_0" | undefined;
         rng?: "cpu" | "cuda" | "std_default" | undefined;
@@ -2157,13 +2161,14 @@ export declare const requestSchema: z.ZodUnion<readonly [z.ZodObject<{
         repeat_penalty?: number | undefined;
         stop_sequences?: string[] | undefined;
         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;
-        "split-mode"?: "none" | "layer" | "row" | undefined;
-        "tensor-split"?: string | undefined;
+        'cache-type-k'?: string | undefined;
+        'cache-type-v'?: string | undefined;
+        'main-gpu'?: number | "integrated" | "dedicated" | undefined;
+        'split-mode'?: "none" | "layer" | "row" | undefined;
+        'tensor-split'?: string | undefined;
         openclCacheDir?: string | undefined;
         reasoning_budget?: number | undefined;
         projectionModelSrc?: string | {
@@ -2180,6 +2185,7 @@ export declare const requestSchema: z.ZodUnion<readonly [z.ZodObject<{
             addon?: "embeddings" | "diffusion" | "tts" | "ocr" | "vla" | "llamacpp-completion" | "whispercpp-transcription" | "bci-whispercpp-transcription" | "llamacpp-embedding" | "nmtcpp-translation" | "onnx-tts" | "tts-ggml" | "parakeet-transcription" | "ggml-ocr" | "sdcpp-generation" | "ggml-vla" | "ggml-classification" | "llm" | "whisper" | "bci" | "nmt" | "parakeet" | "classification" | "vad" | undefined;
         } | undefined;
         image_tile_mode?: "disabled" | "batched" | "sequential" | undefined;
+        'mmproj-use-gpu'?: boolean | undefined;
     };
     modelName?: string | undefined;
     withProgress?: boolean | undefined;
@@ -2358,7 +2364,7 @@ export declare const requestSchema: z.ZodUnion<readonly [z.ZodObject<{
     modelType: "llamacpp-embedding";
     modelConfig: {
         gpuLayers?: number | undefined;
-        device?: "cpu" | "gpu" | undefined;
+        device?: "gpu" | "cpu" | undefined;
         batchSize?: number | undefined;
         pooling?: "none" | "mean" | "cls" | "last" | "rank" | undefined;
         attention?: "causal" | "non-causal" | undefined;
@@ -2576,7 +2582,7 @@ export declare const requestSchema: z.ZodUnion<readonly [z.ZodObject<{
     modelConfig?: {
         mode: "diffusion" | "upscale" | "video";
         threads?: number | undefined;
-        device?: "cpu" | "gpu" | undefined;
+        device?: "gpu" | "cpu" | undefined;
         prediction?: "auto" | "eps" | "v" | "edm_v" | "flow" | "flux2_flow" | undefined;
         type?: "auto" | "f32" | "f16" | "bf16" | "q2_k" | "q3_k" | "q4_0" | "q4_1" | "q4_k" | "q5_0" | "q5_1" | "q5_k" | "q6_k" | "q8_0" | undefined;
         rng?: "cpu" | "cuda" | "std_default" | undefined;
@@ -2953,6 +2959,77 @@ export declare const requestSchema: z.ZodUnion<readonly [z.ZodObject<{
     requestId: z.ZodOptional<z.ZodString>;
     type: z.ZodLiteral<"completionStream">;
 }, z.core.$strip>, z.ZodObject<{
+    modelId: z.ZodString;
+    stream: z.ZodOptional<z.ZodBoolean>;
+    captureThinking: z.ZodOptional<z.ZodBoolean>;
+    emitRawDeltas: z.ZodOptional<z.ZodBoolean>;
+    toolDialect: z.ZodOptional<z.ZodEnum<{
+        hermes: "hermes";
+        pythonic: "pythonic";
+        json: "json";
+        harmony: "harmony";
+        qwen35: "qwen35";
+        gemma4: "gemma4";
+    }>>;
+    requestId: z.ZodOptional<z.ZodString>;
+    prompts: z.ZodArray<z.ZodObject<{
+        id: z.ZodOptional<z.ZodString>;
+        history: z.ZodArray<z.ZodObject<{
+            role: z.ZodString;
+            content: z.ZodString;
+            attachments: z.ZodOptional<z.ZodArray<z.ZodObject<{
+                path: z.ZodString;
+            }, z.core.$strip>>>;
+        }, z.core.$strip>>;
+        generationParams: z.ZodOptional<z.ZodObject<{
+            temp: z.ZodOptional<z.ZodNumber>;
+            top_p: z.ZodOptional<z.ZodNumber>;
+            top_k: z.ZodOptional<z.ZodNumber>;
+            predict: z.ZodOptional<z.ZodNumber>;
+            seed: z.ZodOptional<z.ZodNumber>;
+            frequency_penalty: z.ZodOptional<z.ZodNumber>;
+            presence_penalty: z.ZodOptional<z.ZodNumber>;
+            repeat_penalty: z.ZodOptional<z.ZodNumber>;
+            reasoning_budget: z.ZodOptional<z.ZodNumber>;
+            remove_thinking_from_context: z.ZodOptional<z.ZodBoolean>;
+        }, z.core.$strict>>;
+        responseFormat: z.ZodOptional<z.ZodDiscriminatedUnion<[z.ZodObject<{
+            type: z.ZodLiteral<"text">;
+        }, z.core.$strict>, z.ZodObject<{
+            type: z.ZodLiteral<"json_object">;
+        }, z.core.$strict>, z.ZodObject<{
+            type: z.ZodLiteral<"json_schema">;
+            json_schema: z.ZodObject<{
+                name: z.ZodString;
+                description: z.ZodOptional<z.ZodString>;
+                schema: z.ZodRecord<z.ZodString, z.ZodUnknown>;
+                strict: z.ZodOptional<z.ZodBoolean>;
+            }, z.core.$strict>;
+        }, z.core.$strict>], "type">>;
+        tools: z.ZodOptional<z.ZodArray<z.ZodObject<{
+            type: z.ZodLiteral<"function">;
+            name: z.ZodString;
+            description: z.ZodString;
+            parameters: z.ZodObject<{
+                type: z.ZodLiteral<"object">;
+                properties: z.ZodRecord<z.ZodString, z.ZodObject<{
+                    type: z.ZodEnum<{
+                        string: "string";
+                        number: "number";
+                        boolean: "boolean";
+                        object: "object";
+                        array: "array";
+                        integer: "integer";
+                    }>;
+                    description: z.ZodOptional<z.ZodString>;
+                    enum: z.ZodOptional<z.ZodArray<z.ZodString>>;
+                }, z.core.$strip>>;
+                required: z.ZodOptional<z.ZodArray<z.ZodString>>;
+            }, z.core.$strip>;
+        }, z.core.$strip>>>;
+    }, z.core.$strict>>;
+    type: z.ZodLiteral<"batchCompletionStream">;
+}, z.core.$strict>, z.ZodObject<{
     type: z.ZodLiteral<"unloadModel">;
     modelId: z.ZodString;
     clearStorage: z.ZodDefault<z.ZodBoolean>;
@@ -3080,6 +3157,7 @@ export declare const requestSchema: z.ZodUnion<readonly [z.ZodObject<{
     modelId: z.ZodString;
     kind: z.ZodOptional<z.ZodEnum<{
         completion: "completion";
+        batchCompletion: "batchCompletion";
         embeddings: "embeddings";
         transcribe: "transcribe";
         translate: "translate";
@@ -3622,9 +3700,10 @@ export declare const responseSchema: z.ZodDiscriminatedUnion<[z.ZodObject<{
             cacheTokens: z.ZodOptional<z.ZodNumber>;
             promptTokens: z.ZodOptional<z.ZodNumber>;
             generatedTokens: z.ZodOptional<z.ZodNumber>;
+            avgConcurrentSeq: z.ZodOptional<z.ZodNumber>;
             backendDevice: z.ZodOptional<z.ZodEnum<{
-                cpu: "cpu";
                 gpu: "gpu";
+                cpu: "cpu";
             }>>;
         }, z.core.$strip>;
     }, z.core.$strip>, z.ZodObject<{
@@ -3650,6 +3729,96 @@ export declare const responseSchema: z.ZodDiscriminatedUnion<[z.ZodObject<{
             fullText: z.ZodString;
         }, z.core.$strip>>;
     }, z.core.$strict>]>>;
+}, z.core.$strict>, z.ZodObject<{
+    type: z.ZodLiteral<"batchCompletionStream">;
+    done: z.ZodOptional<z.ZodBoolean>;
+    ids: z.ZodOptional<z.ZodArray<z.ZodString>>;
+    events: z.ZodArray<z.ZodObject<{
+        id: z.ZodString;
+        event: z.ZodUnion<readonly [z.ZodObject<{
+            type: z.ZodLiteral<"contentDelta">;
+            seq: z.ZodNumber;
+            text: z.ZodString;
+        }, z.core.$strip>, z.ZodObject<{
+            type: z.ZodLiteral<"rawDelta">;
+            seq: z.ZodNumber;
+            text: z.ZodString;
+        }, z.core.$strip>, z.ZodObject<{
+            type: z.ZodLiteral<"thinkingDelta">;
+            seq: z.ZodNumber;
+            text: z.ZodString;
+        }, z.core.$strip>, z.ZodObject<{
+            type: z.ZodLiteral<"toolCall">;
+            seq: z.ZodNumber;
+            call: z.ZodObject<{
+                id: z.ZodString;
+                name: z.ZodString;
+                arguments: z.ZodRecord<z.ZodString, z.ZodUnknown>;
+                raw: z.ZodOptional<z.ZodString>;
+            }, z.core.$strip>;
+        }, z.core.$strip>, z.ZodObject<{
+            type: z.ZodLiteral<"toolError">;
+            seq: z.ZodNumber;
+            error: z.ZodObject<{
+                code: z.ZodEnum<{
+                    PARSE_ERROR: "PARSE_ERROR";
+                    VALIDATION_ERROR: "VALIDATION_ERROR";
+                    UNKNOWN_TOOL: "UNKNOWN_TOOL";
+                }>;
+                message: z.ZodString;
+                raw: z.ZodOptional<z.ZodString>;
+            }, z.core.$strip>;
+        }, z.core.$strip>, z.ZodObject<{
+            type: z.ZodLiteral<"completionStats">;
+            seq: z.ZodNumber;
+            stats: z.ZodObject<{
+                timeToFirstToken: z.ZodOptional<z.ZodNumber>;
+                tokensPerSecond: z.ZodOptional<z.ZodNumber>;
+                cacheTokens: z.ZodOptional<z.ZodNumber>;
+                promptTokens: z.ZodOptional<z.ZodNumber>;
+                generatedTokens: z.ZodOptional<z.ZodNumber>;
+                avgConcurrentSeq: z.ZodOptional<z.ZodNumber>;
+                backendDevice: z.ZodOptional<z.ZodEnum<{
+                    gpu: "gpu";
+                    cpu: "cpu";
+                }>>;
+            }, z.core.$strip>;
+        }, z.core.$strip>, z.ZodObject<{
+            type: z.ZodLiteral<"completionDone">;
+            seq: z.ZodNumber;
+            stopReason: z.ZodLiteral<"error">;
+            error: z.ZodObject<{
+                message: z.ZodString;
+            }, z.core.$strip>;
+            raw: z.ZodOptional<z.ZodObject<{
+                fullText: z.ZodString;
+            }, z.core.$strip>>;
+        }, z.core.$strict>, z.ZodObject<{
+            type: z.ZodLiteral<"completionDone">;
+            seq: z.ZodNumber;
+            stopReason: z.ZodOptional<z.ZodEnum<{
+                cancelled: "cancelled";
+                eos: "eos";
+                length: "length";
+                stopSequence: "stopSequence";
+            }>>;
+            raw: z.ZodOptional<z.ZodObject<{
+                fullText: z.ZodString;
+            }, z.core.$strip>>;
+        }, z.core.$strict>]>;
+    }, z.core.$strict>>;
+    stats: z.ZodOptional<z.ZodObject<{
+        timeToFirstToken: z.ZodOptional<z.ZodNumber>;
+        tokensPerSecond: z.ZodOptional<z.ZodNumber>;
+        cacheTokens: z.ZodOptional<z.ZodNumber>;
+        promptTokens: z.ZodOptional<z.ZodNumber>;
+        generatedTokens: z.ZodOptional<z.ZodNumber>;
+        avgConcurrentSeq: z.ZodOptional<z.ZodNumber>;
+        backendDevice: z.ZodOptional<z.ZodEnum<{
+            gpu: "gpu";
+            cpu: "cpu";
+        }>>;
+    }, z.core.$strip>>;
 }, z.core.$strict>, z.ZodObject<{
     type: z.ZodLiteral<"unloadModel">;
     success: z.ZodBoolean;
@@ -3837,8 +4006,8 @@ export declare const responseSchema: z.ZodDiscriminatedUnion<[z.ZodObject<{
         tokensPerSecond: z.ZodOptional<z.ZodNumber>;
         totalTokens: z.ZodOptional<z.ZodNumber>;
         backendDevice: z.ZodOptional<z.ZodEnum<{
-            cpu: "cpu";
             gpu: "gpu";
+            cpu: "cpu";
         }>>;
         contextSize: z.ZodOptional<z.ZodNumber>;
     }, z.core.$strip>>;
@@ -4142,8 +4311,8 @@ export declare const responseSchema: z.ZodDiscriminatedUnion<[z.ZodObject<{
         height: z.ZodOptional<z.ZodNumber>;
         repeats: z.ZodOptional<z.ZodNumber>;
         backendDevice: z.ZodOptional<z.ZodEnum<{
-            cpu: "cpu";
             gpu: "gpu";
+            cpu: "cpu";
         }>>;
     }, z.core.$strip>>;
 }, z.core.$strip>, z.ZodObject<{

package/dist/schemas/common.js

@@ -1,38 +1,40 @@
-import { z } from "zod";
-import { perCallProfilingSchema } from "./profiling.js";
-import { heartbeatRequestSchema, heartbeatResponseSchema } from "./heartbeat.js";
-import { completionStreamRequestSchema, completionStreamResponseSchema, } from "./completion-stream.js";
-import { loadModelRequestSchema, loadModelResponseSchema, modelProgressUpdateSchema, } from "./load-model.js";
-import { downloadAssetRequestSchema, downloadAssetResponseSchema, } from "./download-asset.js";
-import { unloadModelRequestSchema, unloadModelResponseSchema, } from "./unload-model.js";
-import { transcribeRequestSchema, transcribeResponseSchema, transcribeStreamRequestSchema, transcribeStreamResponseSchema, } from "./transcription.js";
-import { bciTranscribeRequestSchema, bciTranscribeResponseSchema, bciTranscribeStreamRequestSchema, bciTranscribeStreamResponseSchema, } from "./bci.js";
-import { embedRequestSchema, embedResponseSchema } from "./embed.js";
-import { cancelRequestSchema, cancelResponseSchema } from "./cancel.js";
-import { provideRequestSchema, provideResponseSchema } from "./provide.js";
-import { stopProvideRequestSchema, stopProvideResponseSchema, } from "./stop-provide.js";
-import { translateRequestSchema, translateResponseSchema } from "./translate.js";
-import { loggingStreamRequestSchema, loggingStreamResponseSchema, } from "./logging-stream.js";
-import { ttsRequestSchema, ttsResponseSchema, textToSpeechStreamRequestSchema, textToSpeechStreamResponseSchema, } from "./text-to-speech.js";
-import { errorResponseSchema } from "./error.js";
-import { ragRequestSchema, ragResponseSchema, ragProgressUpdateSchema, } from "./rag.js";
-import { deleteCacheRequestSchema, deleteCacheResponseSchema, } from "./delete-cache.js";
-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, 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";
+import { z } from 'zod';
+import { perCallProfilingSchema } from './profiling.js';
+import { heartbeatRequestSchema, heartbeatResponseSchema } from './heartbeat.js';
+import { completionStreamRequestSchema, completionStreamResponseSchema } from './completion-stream.js';
+import { batchCompletionStreamRequestSchema, batchCompletionStreamResponseSchema } from './batch-completion-stream.js';
+import { loadModelRequestSchema, loadModelResponseSchema, modelProgressUpdateSchema } from './load-model.js';
+import { downloadAssetRequestSchema, downloadAssetResponseSchema } from './download-asset.js';
+import { unloadModelRequestSchema, unloadModelResponseSchema } from './unload-model.js';
+import { transcribeRequestSchema, transcribeResponseSchema, transcribeStreamRequestSchema, transcribeStreamResponseSchema } from './transcription.js';
+import { bciTranscribeRequestSchema, bciTranscribeResponseSchema, bciTranscribeStreamRequestSchema, bciTranscribeStreamResponseSchema } from './bci.js';
+import { embedRequestSchema, embedResponseSchema } from './embed.js';
+import { cancelRequestSchema, cancelResponseSchema } from './cancel.js';
+import { provideRequestSchema, provideResponseSchema } from './provide.js';
+import { stopProvideRequestSchema, stopProvideResponseSchema } from './stop-provide.js';
+import { translateRequestSchema, translateResponseSchema } from './translate.js';
+import { loggingStreamRequestSchema, loggingStreamResponseSchema } from './logging-stream.js';
+import { ttsRequestSchema, ttsResponseSchema, textToSpeechStreamRequestSchema, textToSpeechStreamResponseSchema } from './text-to-speech.js';
+import { errorResponseSchema } from './error.js';
+import { ragRequestSchema, ragResponseSchema, ragProgressUpdateSchema } from './rag.js';
+import { deleteCacheRequestSchema, deleteCacheResponseSchema } from './delete-cache.js';
+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, 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,
     downloadAssetRequestSchema,
     completionStreamRequestSchema,
+    batchCompletionStreamRequestSchema,
     unloadModelRequestSchema,
     transcribeRequestSchema,
     transcribeStreamRequestSchema,
@@ -63,13 +65,14 @@ export const requestSchema = z.union([
     suspendRequestSchema,
     resumeRequestSchema,
     stateRequestSchema,
-    classifyRequestSchema,
+    classifyRequestSchema
 ]);
-export const responseSchema = z.discriminatedUnion("type", [
+export const responseSchema = z.discriminatedUnion('type', [
     heartbeatResponseSchema,
     loadModelResponseSchema,
     downloadAssetResponseSchema,
     completionStreamResponseSchema,
+    batchCompletionStreamResponseSchema,
     unloadModelResponseSchema,
     modelProgressUpdateSchema,
     transcribeResponseSchema,
@@ -104,24 +107,24 @@ export const responseSchema = z.discriminatedUnion("type", [
     suspendResponseSchema,
     resumeResponseSchema,
     stateResponseSchema,
-    classifyResponseSchema,
+    classifyResponseSchema
 ]);
 export const rpcOptionsSchema = z.object({
     timeout: z
         .number()
         .min(100)
         .optional()
-        .describe("Per-call RPC timeout in milliseconds; overrides the SDK-level default for this request only."),
+        .describe('Per-call RPC timeout in milliseconds; overrides the SDK-level default for this request only.'),
     healthCheckTimeout: z
         .number()
         .min(100)
         .optional()
-        .describe("Timeout in milliseconds for the health-check probe that precedes the RPC call."),
+        .describe('Timeout in milliseconds for the health-check probe that precedes the RPC call.'),
     forceNewConnection: z
         .boolean()
         .optional()
-        .describe("When `true`, skip any cached RPC connection and open a fresh one for this call."),
+        .describe('When `true`, skip any cached RPC connection and open a fresh one for this call.'),
     profiling: perCallProfilingSchema
         .optional()
-        .describe("Per-call profiler configuration; when present, overrides the SDK's global profiler settings for this request."),
+        .describe("Per-call profiler configuration; when present, overrides the SDK's global profiler settings for this request.")
 });

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

@@ -1,15 +1,16 @@
-import { z } from "zod";
-import { type Tool, type ToolCallWithCall } from "./tools";
-import type { ToolDialect } from "./completion-stream";
+import { z } from 'zod';
+import { type Tool, type ToolCallWithCall } from './tools';
+import type { ToolDialect } from './completion-stream';
 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>;
+    avgConcurrentSeq: z.ZodOptional<z.ZodNumber>;
     backendDevice: z.ZodOptional<z.ZodEnum<{
-        cpu: "cpu";
         gpu: "gpu";
+        cpu: "cpu";
     }>>;
 }, z.core.$strip>;
 export type CompletionStats = z.infer<typeof completionStatsSchema>;
@@ -61,9 +62,10 @@ export declare const statsEventSchema: z.ZodObject<{
         cacheTokens: z.ZodOptional<z.ZodNumber>;
         promptTokens: z.ZodOptional<z.ZodNumber>;
         generatedTokens: z.ZodOptional<z.ZodNumber>;
+        avgConcurrentSeq: z.ZodOptional<z.ZodNumber>;
         backendDevice: z.ZodOptional<z.ZodEnum<{
-            cpu: "cpu";
             gpu: "gpu";
+            cpu: "cpu";
         }>>;
     }, z.core.$strip>;
 }, z.core.$strip>;
@@ -141,9 +143,10 @@ export declare const completionEventSchema: z.ZodUnion<readonly [z.ZodObject<{
         cacheTokens: z.ZodOptional<z.ZodNumber>;
         promptTokens: z.ZodOptional<z.ZodNumber>;
         generatedTokens: z.ZodOptional<z.ZodNumber>;
+        avgConcurrentSeq: z.ZodOptional<z.ZodNumber>;
         backendDevice: z.ZodOptional<z.ZodEnum<{
-            cpu: "cpu";
             gpu: "gpu";
+            cpu: "cpu";
         }>>;
     }, z.core.$strip>;
 }, z.core.$strip>, z.ZodObject<{
@@ -225,8 +228,8 @@ export type CompletionRun = {
     toolCalls: Promise<ToolCallWithCall[]>;
     stats: Promise<CompletionStats | undefined>;
 };
-export type ToolCallingCapability = "textParse" | "none";
-export type ThinkingFramingCapability = "thinkTags" | "none";
+export type ToolCallingCapability = 'textParse' | 'none';
+export type ThinkingFramingCapability = 'thinkTags' | 'none';
 export type PluginCapabilities = {
     toolCalling: ToolCallingCapability;
     thinkingFraming: ThinkingFramingCapability;