@qvac/sdk 0.14.1 → 0.15.0

package/dist/server/bare/plugins/llamacpp-completion/ops/kv-cache-session.js

@@ -1,8 +1,8 @@
-import { promises as fsPromises } from "bare-fs";
-import path from "bare-path";
-import { findMatchingCache, generateConfigHash, getCacheFilePath, getCurrentCacheInfo, renameCacheFile, deleteCache as deleteCacheUtil, } from "../../../ops/kv-cache-utils.js";
-import { logCacheSaveError, logCacheStatus, } from "./cache-logger.js";
-import { getServerLogger } from "../../../../../logging/index.js";
+import { promises as fsPromises } from 'bare-fs';
+import path from 'bare-path';
+import { findMatchingCache, generateConfigHash, getCacheFilePath, getCurrentCacheInfo, renameCacheFile, deleteCache as deleteCacheUtil } from '../../../ops/kv-cache-utils.js';
+import { logCacheSaveError, logCacheStatus } from './cache-logger.js';
+import { getServerLogger } from '../../../../../logging/index.js';
 // Used by cross-model paths that have no `RequestContext` (e.g.
 // `deleteKvCacheState`). Per-session call sites receive a logger from
 // the caller — typically `withRequestContext(...)`.
@@ -89,13 +89,13 @@ export function createKvCacheSession(modelId, options) {
     function makeHandle(cachePath, registryKey) {
         const handle = {
             cachePath,
-            savedCount: cachedMessageCounts.get(cachePath) ?? 0,
+            savedCount: cachedMessageCounts.get(cachePath) ?? 0
         };
         turnState.set(handle, {
             cachePath,
             registryKey,
             committed: false,
-            rolledBack: false,
+            rolledBack: false
         });
         return handle;
     }
@@ -133,15 +133,11 @@ export function createKvCacheSession(modelId, options) {
         // `targetCachePath`.
         const existingCache = await findMatchingCache(modelId, input.configHash, input.history);
         const preResponseCacheInfo = await getCurrentCacheInfo(modelId, input.configHash, input.history);
-        const cachePath = existingCache !== null
-            ? existingCache.cachePath
-            : preResponseCacheInfo.cachePath;
-        const cacheKeyForRegistry = existingCache !== null
-            ? existingCache.cacheKey
-            : preResponseCacheInfo.cacheKey;
+        const cachePath = existingCache !== null ? existingCache.cachePath : preResponseCacheInfo.cachePath;
+        const cacheKeyForRegistry = existingCache !== null ? existingCache.cacheKey : preResponseCacheInfo.cacheKey;
         const registryKey = initRegistryKey(modelId, input.configHash, cacheKeyForRegistry);
         const cacheExists = existingCache !== null;
-        logCacheStatus("auto", cacheExists);
+        logCacheStatus('auto', cacheExists);
         if (!cacheExists) {
             await input.primeIfMissing(cachePath);
             await verifyPrimedFile(cachePath, logger);
@@ -150,7 +146,7 @@ export function createKvCacheSession(modelId, options) {
         return makeHandle(cachePath, registryKey);
     }
     async function beginTurn(input) {
-        if (input.kind === "custom")
+        if (input.kind === 'custom')
             return beginCustom(input);
         return beginAuto(input);
     }
@@ -165,7 +161,7 @@ export function createKvCacheSession(modelId, options) {
         }
         if (state.committed || state.rolledBack)
             return;
-        if (result.kind === "static") {
+        if (result.kind === 'static') {
             // Custom-key path: the addon wrote the new cache state inline
             // at the same path. Verify the file persisted (the addon
             // currently swallows save errors — see TODO in
@@ -238,7 +234,7 @@ export function createKvCacheSession(modelId, options) {
         beginTurn,
         commitTurn,
         rollback,
-        dropStaleSavedCount,
+        dropStaleSavedCount
     };
 }
 // ----- module-level administrative API -----
@@ -273,7 +269,7 @@ export function createKvCacheSession(modelId, options) {
  * no-op on absent keys).
  */
 export async function deleteKvCacheState(target) {
-    if ("all" in target) {
+    if ('all' in target) {
         const removed = await deleteCacheUtil({ all: true });
         cachedMessageCounts.clear();
         initializedCaches.clear();
@@ -284,7 +280,7 @@ export async function deleteKvCacheState(target) {
     }
     const removedPath = await deleteCacheUtil({
         kvCacheKey: target.kvCacheKey,
-        ...(target.modelId !== undefined && { modelId: target.modelId }),
+        ...(target.modelId !== undefined && { modelId: target.modelId })
     });
     // Prefix-cleanup the in-memory counts. The on-disk directory tree
     // is `{kvCacheRoot}/{kvCacheKey}[/{modelId}]/`, so every entry in
@@ -296,7 +292,7 @@ export async function deleteKvCacheState(target) {
     // facing kvCacheKey (and optionally narrow by modelId).
     clearInitializedCachesByScope({
         cacheKey: target.kvCacheKey,
-        ...(target.modelId !== undefined && { modelId: target.modelId }),
+        ...(target.modelId !== undefined && { modelId: target.modelId })
     });
 }
 // ----- private helpers -----
@@ -398,8 +394,8 @@ function clearInitializedCachesByScope(scope) {
         return;
     }
     for (const key of initializedCaches) {
-        const firstSep = key.indexOf(":");
-        const secondSep = key.indexOf(":", firstSep + 1);
+        const firstSep = key.indexOf(':');
+        const secondSep = key.indexOf(':', firstSep + 1);
         if (firstSep === -1 || secondSep === -1)
             continue;
         const entryModelId = key.slice(0, firstSep);
@@ -439,7 +435,7 @@ export const __kvCacheSessionTestHooks = {
     resetForTest() {
         cachedMessageCounts.clear();
         initializedCaches.clear();
-    },
+    }
 };
 // Re-export `generateConfigHash` from the path utilities so callers of
 // the session can compute the hash without separately importing

package/dist/server/bare/plugins/llamacpp-completion/ops/kv-cache-state.js

@@ -30,8 +30,8 @@
 export function decideCachedHistorySlice(savedCount, cacheExists, history) {
     if (!cacheExists || history.length === 0) {
         return {
-            messages: history.filter((msg) => msg.role !== "system"),
-            clearStaleCount: false,
+            messages: history.filter((msg) => msg.role !== 'system'),
+            clearStaleCount: false
         };
     }
     const canSlice = savedCount > 0 && savedCount <= history.length;
@@ -42,11 +42,9 @@ export function decideCachedHistorySlice(savedCount, cacheExists, history) {
     // mid-decode and still recorded `history.length + 1`. Treat it as a
     // bad state and resend the full (system-stripped) history.
     const useSlice = sliced !== null && sliced.length > 0;
-    const messages = useSlice
-        ? sliced
-        : history.filter((msg) => msg.role !== "system");
+    const messages = useSlice ? sliced : history.filter((msg) => msg.role !== 'system');
     return {
         messages,
-        clearStaleCount: !useSlice && savedCount > 0,
+        clearStaleCount: !useSlice && savedCount > 0
     };
 }

package/dist/server/bare/plugins/llamacpp-completion/plugin.d.ts

@@ -1,4 +1,4 @@
-import { type CreateModelParams, type PluginModelResult, type ResolveContext, type LlmConfigInput } from "../../../../schemas/index";
+import { type CreateModelParams, type PluginModelResult, type ResolveContext, type LlmConfigInput } from '../../../../schemas/index';
 export declare const llmPlugin: {
     modelType: "llamacpp-completion";
     displayName: string;
@@ -21,23 +21,24 @@ export declare const llmPlugin: {
         repeat_penalty: import("zod").ZodOptional<import("zod").ZodNumber>;
         stop_sequences: import("zod").ZodOptional<import("zod").ZodArray<import("zod").ZodString>>;
         n_discarded: import("zod").ZodOptional<import("zod").ZodNumber>;
+        parallel: import("zod").ZodOptional<import("zod").ZodNumber>;
         tools: import("zod").ZodOptional<import("zod").ZodBoolean>;
         toolsMode: import("zod").ZodOptional<import("zod").ZodEnum<{
             static: "static";
             dynamic: "dynamic";
         }>>;
-        "cache-type-k": import("zod").ZodOptional<import("zod").ZodString>;
-        "cache-type-v": import("zod").ZodOptional<import("zod").ZodString>;
-        "main-gpu": import("zod").ZodOptional<import("zod").ZodUnion<readonly [import("zod").ZodNumber, import("zod").ZodEnum<{
+        'cache-type-k': import("zod").ZodOptional<import("zod").ZodString>;
+        'cache-type-v': import("zod").ZodOptional<import("zod").ZodString>;
+        'main-gpu': import("zod").ZodOptional<import("zod").ZodUnion<readonly [import("zod").ZodNumber, import("zod").ZodEnum<{
             integrated: "integrated";
             dedicated: "dedicated";
         }>]>>;
-        "split-mode": import("zod").ZodOptional<import("zod").ZodEnum<{
+        'split-mode': import("zod").ZodOptional<import("zod").ZodEnum<{
             none: "none";
             layer: "layer";
             row: "row";
         }>>;
-        "tensor-split": import("zod").ZodOptional<import("zod").ZodString>;
+        'tensor-split': import("zod").ZodOptional<import("zod").ZodString>;
         openclCacheDir: import("zod").ZodOptional<import("zod").ZodString>;
         reasoning_budget: import("zod").ZodOptional<import("zod").ZodNumber>;
         projectionModelSrc: import("zod").ZodOptional<import("zod").ZodUnion<readonly [import("zod").ZodString, import("zod").ZodObject<{
@@ -82,6 +83,7 @@ export declare const llmPlugin: {
             batched: "batched";
             sequential: "sequential";
         }>>;
+        'mmproj-use-gpu': import("zod").ZodOptional<import("zod").ZodBoolean>;
     }, import("zod/v4/core").$strip>;
     resolveConfig(cfg: LlmConfigInput, ctx: ResolveContext): Promise<{
         config: {
@@ -102,16 +104,18 @@ export declare const llmPlugin: {
             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;
             image_tile_mode?: "disabled" | "batched" | "sequential" | undefined;
+            'mmproj-use-gpu'?: boolean | undefined;
         };
         artifacts?: never;
     } | {
@@ -133,16 +137,18 @@ export declare const llmPlugin: {
             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;
             image_tile_mode?: "disabled" | "batched" | "sequential" | undefined;
+            'mmproj-use-gpu'?: boolean | undefined;
         };
         artifacts: {
             projectionModelPath: string;
@@ -150,6 +156,168 @@ export declare const llmPlugin: {
     }>;
     createModel(params: CreateModelParams): PluginModelResult;
     handlers: {
+        batchCompletionStream: import("../../../../schemas/index").PluginHandlerDefinition<import("zod").ZodObject<{
+            modelId: import("zod").ZodString;
+            stream: import("zod").ZodOptional<import("zod").ZodBoolean>;
+            captureThinking: import("zod").ZodOptional<import("zod").ZodBoolean>;
+            emitRawDeltas: import("zod").ZodOptional<import("zod").ZodBoolean>;
+            toolDialect: import("zod").ZodOptional<import("zod").ZodEnum<{
+                hermes: "hermes";
+                pythonic: "pythonic";
+                json: "json";
+                harmony: "harmony";
+                qwen35: "qwen35";
+                gemma4: "gemma4";
+            }>>;
+            requestId: import("zod").ZodOptional<import("zod").ZodString>;
+            prompts: import("zod").ZodArray<import("zod").ZodObject<{
+                id: import("zod").ZodOptional<import("zod").ZodString>;
+                history: import("zod").ZodArray<import("zod").ZodObject<{
+                    role: import("zod").ZodString;
+                    content: import("zod").ZodString;
+                    attachments: import("zod").ZodOptional<import("zod").ZodArray<import("zod").ZodObject<{
+                        path: import("zod").ZodString;
+                    }, import("zod/v4/core").$strip>>>;
+                }, import("zod/v4/core").$strip>>;
+                generationParams: import("zod").ZodOptional<import("zod").ZodObject<{
+                    temp: import("zod").ZodOptional<import("zod").ZodNumber>;
+                    top_p: import("zod").ZodOptional<import("zod").ZodNumber>;
+                    top_k: import("zod").ZodOptional<import("zod").ZodNumber>;
+                    predict: import("zod").ZodOptional<import("zod").ZodNumber>;
+                    seed: import("zod").ZodOptional<import("zod").ZodNumber>;
+                    frequency_penalty: import("zod").ZodOptional<import("zod").ZodNumber>;
+                    presence_penalty: import("zod").ZodOptional<import("zod").ZodNumber>;
+                    repeat_penalty: import("zod").ZodOptional<import("zod").ZodNumber>;
+                    reasoning_budget: import("zod").ZodOptional<import("zod").ZodNumber>;
+                    remove_thinking_from_context: import("zod").ZodOptional<import("zod").ZodBoolean>;
+                }, import("zod/v4/core").$strict>>;
+                responseFormat: import("zod").ZodOptional<import("zod").ZodDiscriminatedUnion<[import("zod").ZodObject<{
+                    type: import("zod").ZodLiteral<"text">;
+                }, import("zod/v4/core").$strict>, import("zod").ZodObject<{
+                    type: import("zod").ZodLiteral<"json_object">;
+                }, import("zod/v4/core").$strict>, import("zod").ZodObject<{
+                    type: import("zod").ZodLiteral<"json_schema">;
+                    json_schema: import("zod").ZodObject<{
+                        name: import("zod").ZodString;
+                        description: import("zod").ZodOptional<import("zod").ZodString>;
+                        schema: import("zod").ZodRecord<import("zod").ZodString, import("zod").ZodUnknown>;
+                        strict: import("zod").ZodOptional<import("zod").ZodBoolean>;
+                    }, import("zod/v4/core").$strict>;
+                }, import("zod/v4/core").$strict>], "type">>;
+                tools: import("zod").ZodOptional<import("zod").ZodArray<import("zod").ZodObject<{
+                    type: import("zod").ZodLiteral<"function">;
+                    name: import("zod").ZodString;
+                    description: import("zod").ZodString;
+                    parameters: import("zod").ZodObject<{
+                        type: import("zod").ZodLiteral<"object">;
+                        properties: import("zod").ZodRecord<import("zod").ZodString, import("zod").ZodObject<{
+                            type: import("zod").ZodEnum<{
+                                string: "string";
+                                number: "number";
+                                boolean: "boolean";
+                                object: "object";
+                                array: "array";
+                                integer: "integer";
+                            }>;
+                            description: import("zod").ZodOptional<import("zod").ZodString>;
+                            enum: import("zod").ZodOptional<import("zod").ZodArray<import("zod").ZodString>>;
+                        }, import("zod/v4/core").$strip>>;
+                        required: import("zod").ZodOptional<import("zod").ZodArray<import("zod").ZodString>>;
+                    }, import("zod/v4/core").$strip>;
+                }, import("zod/v4/core").$strip>>>;
+            }, import("zod/v4/core").$strict>>;
+            type: import("zod").ZodLiteral<"batchCompletionStream">;
+        }, import("zod/v4/core").$strict>, import("zod").ZodObject<{
+            type: import("zod").ZodLiteral<"batchCompletionStream">;
+            done: import("zod").ZodOptional<import("zod").ZodBoolean>;
+            ids: import("zod").ZodOptional<import("zod").ZodArray<import("zod").ZodString>>;
+            events: import("zod").ZodArray<import("zod").ZodObject<{
+                id: import("zod").ZodString;
+                event: import("zod").ZodUnion<readonly [import("zod").ZodObject<{
+                    type: import("zod").ZodLiteral<"contentDelta">;
+                    seq: import("zod").ZodNumber;
+                    text: import("zod").ZodString;
+                }, import("zod/v4/core").$strip>, import("zod").ZodObject<{
+                    type: import("zod").ZodLiteral<"rawDelta">;
+                    seq: import("zod").ZodNumber;
+                    text: import("zod").ZodString;
+                }, import("zod/v4/core").$strip>, import("zod").ZodObject<{
+                    type: import("zod").ZodLiteral<"thinkingDelta">;
+                    seq: import("zod").ZodNumber;
+                    text: import("zod").ZodString;
+                }, import("zod/v4/core").$strip>, import("zod").ZodObject<{
+                    type: import("zod").ZodLiteral<"toolCall">;
+                    seq: import("zod").ZodNumber;
+                    call: import("zod").ZodObject<{
+                        id: import("zod").ZodString;
+                        name: import("zod").ZodString;
+                        arguments: import("zod").ZodRecord<import("zod").ZodString, import("zod").ZodUnknown>;
+                        raw: import("zod").ZodOptional<import("zod").ZodString>;
+                    }, import("zod/v4/core").$strip>;
+                }, import("zod/v4/core").$strip>, import("zod").ZodObject<{
+                    type: import("zod").ZodLiteral<"toolError">;
+                    seq: import("zod").ZodNumber;
+                    error: import("zod").ZodObject<{
+                        code: import("zod").ZodEnum<{
+                            PARSE_ERROR: "PARSE_ERROR";
+                            VALIDATION_ERROR: "VALIDATION_ERROR";
+                            UNKNOWN_TOOL: "UNKNOWN_TOOL";
+                        }>;
+                        message: import("zod").ZodString;
+                        raw: import("zod").ZodOptional<import("zod").ZodString>;
+                    }, import("zod/v4/core").$strip>;
+                }, import("zod/v4/core").$strip>, import("zod").ZodObject<{
+                    type: import("zod").ZodLiteral<"completionStats">;
+                    seq: import("zod").ZodNumber;
+                    stats: import("zod").ZodObject<{
+                        timeToFirstToken: import("zod").ZodOptional<import("zod").ZodNumber>;
+                        tokensPerSecond: import("zod").ZodOptional<import("zod").ZodNumber>;
+                        cacheTokens: import("zod").ZodOptional<import("zod").ZodNumber>;
+                        promptTokens: import("zod").ZodOptional<import("zod").ZodNumber>;
+                        generatedTokens: import("zod").ZodOptional<import("zod").ZodNumber>;
+                        avgConcurrentSeq: import("zod").ZodOptional<import("zod").ZodNumber>;
+                        backendDevice: import("zod").ZodOptional<import("zod").ZodEnum<{
+                            gpu: "gpu";
+                            cpu: "cpu";
+                        }>>;
+                    }, import("zod/v4/core").$strip>;
+                }, import("zod/v4/core").$strip>, import("zod").ZodObject<{
+                    type: import("zod").ZodLiteral<"completionDone">;
+                    seq: import("zod").ZodNumber;
+                    stopReason: import("zod").ZodLiteral<"error">;
+                    error: import("zod").ZodObject<{
+                        message: import("zod").ZodString;
+                    }, import("zod/v4/core").$strip>;
+                    raw: import("zod").ZodOptional<import("zod").ZodObject<{
+                        fullText: import("zod").ZodString;
+                    }, import("zod/v4/core").$strip>>;
+                }, import("zod/v4/core").$strict>, import("zod").ZodObject<{
+                    type: import("zod").ZodLiteral<"completionDone">;
+                    seq: import("zod").ZodNumber;
+                    stopReason: import("zod").ZodOptional<import("zod").ZodEnum<{
+                        cancelled: "cancelled";
+                        eos: "eos";
+                        length: "length";
+                        stopSequence: "stopSequence";
+                    }>>;
+                    raw: import("zod").ZodOptional<import("zod").ZodObject<{
+                        fullText: import("zod").ZodString;
+                    }, import("zod/v4/core").$strip>>;
+                }, import("zod/v4/core").$strict>]>;
+            }, import("zod/v4/core").$strict>>;
+            stats: import("zod").ZodOptional<import("zod").ZodObject<{
+                timeToFirstToken: import("zod").ZodOptional<import("zod").ZodNumber>;
+                tokensPerSecond: import("zod").ZodOptional<import("zod").ZodNumber>;
+                cacheTokens: import("zod").ZodOptional<import("zod").ZodNumber>;
+                promptTokens: import("zod").ZodOptional<import("zod").ZodNumber>;
+                generatedTokens: import("zod").ZodOptional<import("zod").ZodNumber>;
+                avgConcurrentSeq: import("zod").ZodOptional<import("zod").ZodNumber>;
+                backendDevice: import("zod").ZodOptional<import("zod").ZodEnum<{
+                    gpu: "gpu";
+                    cpu: "cpu";
+                }>>;
+            }, import("zod/v4/core").$strip>>;
+        }, import("zod/v4/core").$strict>>;
         completionStream: import("../../../../schemas/index").PluginHandlerDefinition<import("zod").ZodObject<{
             history: import("zod").ZodArray<import("zod").ZodObject<{
                 role: import("zod").ZodString;
@@ -264,9 +432,10 @@ export declare const llmPlugin: {
                     cacheTokens: import("zod").ZodOptional<import("zod").ZodNumber>;
                     promptTokens: import("zod").ZodOptional<import("zod").ZodNumber>;
                     generatedTokens: import("zod").ZodOptional<import("zod").ZodNumber>;
+                    avgConcurrentSeq: import("zod").ZodOptional<import("zod").ZodNumber>;
                     backendDevice: import("zod").ZodOptional<import("zod").ZodEnum<{
-                        cpu: "cpu";
                         gpu: "gpu";
+                        cpu: "cpu";
                     }>>;
                 }, import("zod/v4/core").$strip>;
             }, import("zod/v4/core").$strip>, import("zod").ZodObject<{