@qvac/sdk 0.15.0 → 0.17.0

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

@@ -1,4 +1,4 @@
-import { type CreateModelParams, type PluginModelResult } from '../../../../schemas/index';
+import { type CreateModelParams, type PluginModelResult } from '../../../../schemas/index.js';
 export declare const vlaPlugin: {
     modelType: "ggml-vla";
     displayName: string;
@@ -12,7 +12,7 @@ export declare const vlaPlugin: {
     }, import("zod/v4/core").$strip>;
     createModel(params: CreateModelParams): PluginModelResult;
     handlers: {
-        vlaRun: import("../../../../schemas/index").PluginHandlerDefinition<import("zod").ZodObject<{
+        vlaRun: import("../../../../schemas/index.js").PluginHandlerDefinition<import("zod").ZodObject<{
             type: import("zod").ZodLiteral<"vlaRun">;
             modelId: import("zod").ZodString;
             images: import("zod").ZodArray<import("zod").ZodString>;
@@ -37,7 +37,7 @@ export declare const vlaPlugin: {
                 backendDevice: import("zod").ZodOptional<import("zod").ZodNumber>;
             }, import("zod/v4/core").$strip>>;
         }, import("zod/v4/core").$strip>>;
-        vlaHparams: import("../../../../schemas/index").PluginHandlerDefinition<import("zod").ZodObject<{
+        vlaHparams: import("../../../../schemas/index.js").PluginHandlerDefinition<import("zod").ZodObject<{
             type: import("zod").ZodLiteral<"vlaHparams">;
             modelId: import("zod").ZodString;
         }, import("zod/v4/core").$strip>, import("zod").ZodObject<{
@@ -53,6 +53,11 @@ export declare const vlaPlugin: {
                     discrete: "discrete";
                     continuous: "continuous";
                 }>>;
+                imageInputMode: import("zod").ZodOptional<import("zod").ZodEnum<{
+                    pixels: "pixels";
+                    patches: "patches";
+                }>>;
+                imagePatchElems: import("zod").ZodOptional<import("zod").ZodNumber>;
             }, import("zod/v4/core").$strip>;
             backendName: import("zod").ZodNullable<import("zod").ZodString>;
         }, import("zod/v4/core").$strip>>;

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

@@ -16,7 +16,7 @@ function wrapVlaModel(inner, loadOpts) {
 }
 export const vlaPlugin = definePlugin({
     modelType: ModelType.ggmlVla,
-    displayName: 'VLA (SmolVLA / π₀.₅ ggml)',
+    displayName: 'VLA (SmolVLA / π₀.₅ / GR00T ggml)',
     addonPackage: ADDON_VLA,
     loadConfigSchema: vlaConfigSchema,
     createModel(params) {

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

@@ -1,12 +1,13 @@
-export { llmPlugin } from './llamacpp-completion/plugin';
-export { embeddingsPlugin } from './llamacpp-embedding/plugin';
-export { whisperPlugin } from './whispercpp-transcription/plugin';
-export { bciPlugin } from './bci-whispercpp-transcription/plugin';
-export { parakeetPlugin } from './parakeet-transcription/plugin';
-export { nmtPlugin } from './nmtcpp-translation/plugin';
-export { ttsPlugin } from './tts-ggml/plugin';
-export { ocrPlugin } from './ggml-ocr/plugin';
-export { diffusionPlugin } from './sdcpp-generation/plugin';
-export { vlaPlugin } from './ggml-vla/plugin';
-export { classificationPlugin } from './ggml-classification/plugin';
+export { llmPlugin } from './llamacpp-completion/plugin.js';
+export { embeddingsPlugin } from './llamacpp-embedding/plugin.js';
+export { whisperPlugin } from './whispercpp-transcription/plugin.js';
+export { bciPlugin } from './bci-whispercpp-transcription/plugin.js';
+export { parakeetPlugin } from './parakeet-transcription/plugin.js';
+export { nmtPlugin } from './nmtcpp-translation/plugin.js';
+export { ttsPlugin } from './tts-ggml/plugin.js';
+export { ocrPlugin } from './ggml-ocr/plugin.js';
+export { diffusionPlugin } from './sdcpp-generation/plugin.js';
+export { vlaPlugin } from './ggml-vla/plugin.js';
+export { classificationPlugin } from './ggml-classification/plugin.js';
+export { audioGenPlugin } from '../../../server/bare/plugins/audiogen-ggml/plugin.js';
 //# sourceMappingURL=index.d.ts.map

package/dist/server/bare/plugins/index.js

@@ -9,3 +9,4 @@ export { ocrPlugin } from './ggml-ocr/plugin.js';
 export { diffusionPlugin } from './sdcpp-generation/plugin.js';
 export { vlaPlugin } from './ggml-vla/plugin.js';
 export { classificationPlugin } from './ggml-classification/plugin.js';
+export { audioGenPlugin } from '../../../server/bare/plugins/audiogen-ggml/plugin.js';

package/dist/server/bare/plugins/llamacpp-completion/ops/batch-completion-stream.d.ts

@@ -1,7 +1,7 @@
 import type { AbortSignal } from 'bare-abort-controller';
-import type { BatchCompletionStreamPrompt, CompletionStats } from '../../../../../schemas/index';
-import type { DisposableScope } from '../../../runtime/disposable-scope';
-import type { Logger } from '../../../../../logging/types';
+import type { BatchCompletionStreamPrompt, CompletionStats } from '../../../../../schemas/index.js';
+import type { DisposableScope } from '../../../../../server/bare/runtime/disposable-scope.js';
+import type { Logger } from '../../../../../logging/types.js';
 type BatchModelEvent = {
     type: 'ids';
     ids: string[];

package/dist/server/bare/plugins/llamacpp-completion/ops/batch-completion-stream.js

@@ -1,12 +1,12 @@
 import { TOOLS_MODE } from '../../../../../schemas/tools.js';
-import { getModel, getModelConfig } from '../../../registry/model-registry.js';
+import { getModel, getModelConfig } from '../../../../../server/bare/registry/model-registry.js';
 import { getServerLogger } from '../../../../../logging/index.js';
 import { nowMs } from '../../../../../profiling/index.js';
 import { buildStreamResult } from '../../../../../profiling/model-execution.js';
-import { getResponseFormatJsonSchema } from '../../../../utils/response-format.js';
-import { transformMessages } from './completion-stream.js';
-import { normalizeCompletionStats } from './completion-stats.js';
-import { appendToolsToHistory, prependToolsToHistory } from '../../../../utils/tool-integration.js';
+import { getResponseFormatJsonSchema } from '../../../../../server/utils/response-format.js';
+import { transformMessages } from '../../../../../server/bare/plugins/llamacpp-completion/ops/completion-stream.js';
+import { normalizeCompletionStats } from '../../../../../server/bare/plugins/llamacpp-completion/ops/completion-stats.js';
+import { appendToolsToHistory, prependToolsToHistory } from '../../../../../server/utils/tool-integration.js';
 const logger = getServerLogger();
 function runBatchModel(model, prompts) {
     const run = model.run.bind(model);

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

@@ -1,11 +1,21 @@
-import type { LlmStats } from '../../../types/addon-responses';
+import type { CompletionStats } from '../../../../../schemas/index.js';
+import type { LlmStats } from '../../../../../server/bare/types/addon-responses.js';
 export declare function normalizeCompletionStats(stats: LlmStats | undefined): {
     timeToFirstToken?: number | undefined;
     tokensPerSecond?: number | undefined;
     cacheTokens?: number | undefined;
     promptTokens?: number | undefined;
     generatedTokens?: number | undefined;
+    emittedTokens?: number | undefined;
     avgConcurrentSeq?: number | undefined;
     backendDevice?: "gpu" | "cpu" | undefined;
 } | undefined;
+/**
+ * Attach the count of non-empty pieces the addon streamed without
+ * overwriting `generatedTokens` (`llama_perf` `n_eval`). Length / KV-cache
+ * decisions keep the decode count; usage reporting prefers `emittedTokens`.
+ * Zero is attached explicitly so inflated `n_eval` cannot be used as usage
+ * when nothing was streamed.
+ */
+export declare function withEmittedTokens(stats: CompletionStats | undefined, emittedPieces: number): CompletionStats | undefined;
 //# sourceMappingURL=completion-stats.d.ts.map

package/dist/server/bare/plugins/llamacpp-completion/ops/completion-stats.js

@@ -30,3 +30,19 @@ export function normalizeCompletionStats(stats) {
     }
     return normalized;
 }
+/**
+ * Attach the count of non-empty pieces the addon streamed without
+ * overwriting `generatedTokens` (`llama_perf` `n_eval`). Length / KV-cache
+ * decisions keep the decode count; usage reporting prefers `emittedTokens`.
+ * Zero is attached explicitly so inflated `n_eval` cannot be used as usage
+ * when nothing was streamed.
+ */
+export function withEmittedTokens(stats, emittedPieces) {
+    const emittedTokens = emittedPieces;
+    if (!stats && emittedTokens === 0)
+        return undefined;
+    return {
+        ...(stats ?? {}),
+        emittedTokens
+    };
+}

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

@@ -1,7 +1,7 @@
 import type { AbortSignal } from 'bare-abort-controller';
-import type { CompletionParams, CompletionStats, GenerationParams, ResponseFormat, Tool, ToolCall, ToolDialect } from '../../../../../schemas/index';
-import type { DisposableScope } from '../../../runtime/disposable-scope';
-import type { Logger } from '../../../../../logging/types';
+import type { CompletionParams, CompletionStats, GenerationParams, ResponseFormat, Tool, ToolCall, ToolDialect } from '../../../../../schemas/index.js';
+import type { DisposableScope } from '../../../../../server/bare/runtime/disposable-scope.js';
+import type { Logger } from '../../../../../logging/types.js';
 interface CompletionResult {
     modelExecutionMs: number;
     stats?: CompletionStats;

package/dist/server/bare/plugins/llamacpp-completion/ops/completion-stream.js

@@ -1,36 +1,20 @@
 import { TOOLS_MODE } from '../../../../../schemas/tools.js';
-import { logCacheDisabled, logCacheInit, logCacheSave, logMessagesToAddon } from './cache-logger.js';
-import { extractSystemPrompt, getCurrentCacheInfo } from '../../../ops/kv-cache-utils.js';
-import { getModel, getModelConfig } from '../../../registry/model-registry.js';
-import { decideCachedHistorySlice } from './kv-cache-state.js';
-import { createKvCacheSession, generateConfigHash } from './kv-cache-session.js';
-import { appendToolsToHistory, detectToolDialect, prependToolsToHistory } from '../../../../utils/tool-integration.js';
-import { parseToolCalls } from '../../../../utils/tools/index.js';
-import { getResponseFormatJsonSchema } from '../../../../utils/response-format.js';
-import { buildAutoCacheSaveHistory } from '../../../../utils/index.js';
+import { logCacheDisabled, logCacheInit, logCacheSave, logMessagesToAddon } from '../../../../../server/bare/plugins/llamacpp-completion/ops/cache-logger.js';
+import { extractSystemPrompt, getCurrentCacheInfo } from '../../../../../server/bare/ops/kv-cache-utils.js';
+import { getModel, getModelConfig } from '../../../../../server/bare/registry/model-registry.js';
+import { decideCachedHistorySlice, shouldCommitCachedTurn } from '../../../../../server/bare/plugins/llamacpp-completion/ops/kv-cache-state.js';
+import { createKvCacheSession, generateConfigHash } from '../../../../../server/bare/plugins/llamacpp-completion/ops/kv-cache-session.js';
+import { appendToolsToHistory, detectToolDialect, prependToolsToHistory } from '../../../../../server/utils/tool-integration.js';
+import { parseToolCalls } from '../../../../../server/utils/tools/index.js';
+import { getResponseFormatJsonSchema } from '../../../../../server/utils/response-format.js';
+import { buildAutoCacheSaveHistory } from '../../../../../server/utils/index.js';
 import { getServerLogger } from '../../../../../logging/index.js';
 import { AttachmentNotFoundError } from '../../../../../utils/errors-server.js';
 import { nowMs } from '../../../../../profiling/index.js';
 import { buildStreamResult } from '../../../../../profiling/model-execution.js';
-import { normalizeCompletionStats } from './completion-stats.js';
+import { normalizeCompletionStats, withEmittedTokens } from '../../../../../server/bare/plugins/llamacpp-completion/ops/completion-stats.js';
 import fs from 'bare-fs';
 const logger = getServerLogger();
-/**
- * Decide whether a completed turn earned the right to record its kv-cache
- * boundary. A `savedCount` is only safe to write when the turn ran to
- * completion AND produced at least one token — anything else (cancelled
- * mid-decode, zero-token reply, early EOS) leaves the on-disk cache file
- * in an unknown state relative to `history.length + 1`, and a stale entry
- * would slice the next turn's history down to an empty payload.
- *
- * Replaces the pre-0.11.0 `shouldRecordSavedCount(wasCancelled, ...)` with
- * a signal-driven check that reads directly from the request's
- * `AbortSignal`. The local helper keeps the call sites in
- * `completion-stream.ts` honest without importing the registry every time.
- */
-function shouldCommitTurn(signal, producedTokens) {
-    return !signal.aborted && producedTokens;
-}
 function transformMessage(message) {
     const transformed = [];
     // Check if it's a tool definition (has type: "function")
@@ -63,8 +47,7 @@ function transformMessage(message) {
     return transformed;
 }
 function runModel(model, prompt, opts) {
-    const run = model.run.bind(model);
-    return run(prompt, opts);
+    return model.run(prompt, opts);
 }
 export function transformMessages(messages) {
     const transformed = [];
@@ -177,11 +160,14 @@ async function* processModelResponse(model, messagesToSend, tools, generationPar
     const response = await runModel(model, messagesToSend, hasRunOptions ? runOptions : undefined);
     let accumulatedText = '';
     let producedTokens = false;
+    let emittedPieces = 0;
     let toolCallsResult = [];
     for await (const token of response.iterate()) {
         const tokenStr = token;
-        if (tokenStr.length > 0)
+        if (tokenStr.length > 0) {
             producedTokens = true;
+            emittedPieces++;
+        }
         accumulatedText += tokenStr;
         yield { token: tokenStr };
     }
@@ -194,7 +180,7 @@ async function* processModelResponse(model, messagesToSend, tools, generationPar
         toolCallsResult = toolCalls;
     }
     const responseWithStats = response;
-    const stats = normalizeCompletionStats(responseWithStats.stats);
+    const stats = withEmittedTokens(normalizeCompletionStats(responseWithStats.stats), emittedPieces);
     return {
         ...buildStreamResult(modelExecutionMs, stats),
         toolCalls: toolCallsResult,
@@ -338,11 +324,17 @@ export async function* completion(params, opts) {
     /* cacheExists */ true, history, dynamicTools ? tools : undefined);
     logMessagesToAddon(messagesToSend, 'PROMPT_SEND');
     const result = yield* processModelResponse(model, messagesToSend, tools, mergedGenerationParams, { cacheKey: turn.cachePath, saveCacheToDisk: true }, dialect);
+    const shouldCommitTurn = shouldCommitCachedTurn({
+        aborted: signal.aborted,
+        producedTokens: result.producedTokens,
+        generatedTokens: result.stats?.generatedTokens,
+        predict: mergedGenerationParams?.predict ?? modelConfig.predict
+    });
     if (typeof kvCache === 'string') {
         // Custom-key path: the addon wrote the new cache state inline at
         // the same path. Either commit (records the boundary, suppresses
         // rollback) or fall through to the deferred rollback.
-        if (shouldCommitTurn(signal, result.producedTokens)) {
+        if (shouldCommitTurn) {
             await session.commitTurn(turn, {
                 kind: 'static',
                 messageCount: history.length + 1
@@ -362,10 +354,9 @@ export async function* completion(params, opts) {
         logger.warn(`[kv-cache] Auto cache tool-call turn; rolling back to avoid disk leak. path=${turn.cachePath}`);
         return result;
     }
-    if (!shouldCommitTurn(signal, result.producedTokens)) {
-        // Cancelled or zero-token turn — the addon wrote the file but its
-        // contents don't correspond to a clean turn boundary. Let the
-        // deferred rollback unlink it.
+    if (!shouldCommitTurn) {
+        // Cancelled, zero-token, or budget-exhausted turns do not establish
+        // a trustworthy message boundary.
         return result;
     }
     const savedHistory = buildAutoCacheSaveHistory(history.map((msg) => ({

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

@@ -1,4 +1,4 @@
-import type { FinetuneRunParams, FinetuneRunRequest, FinetuneProgress, FinetuneRequest, FinetuneResult, FinetuneStatus, FinetuneGetStateRequest } from '../../../../../schemas/index';
+import type { FinetuneRunParams, FinetuneRunRequest, FinetuneProgress, FinetuneRequest, FinetuneResult, FinetuneStatus, FinetuneGetStateRequest } from '../../../../../schemas/index.js';
 type FinetuneOptions = FinetuneRunParams['options'];
 export declare function clearFinetuneRuntimeState(modelId: string): void;
 export declare function getFinetuneStateFromCheckpoints(options: FinetuneOptions): FinetuneStatus;

package/dist/server/bare/plugins/llamacpp-completion/ops/finetune.js

@@ -51,10 +51,10 @@ var __disposeResources = (this && this.__disposeResources) || (function (Suppres
     return e.name = "SuppressedError", e.error = error, e.suppressed = suppressed, e;
 });
 import fs from 'bare-fs';
-import { getModel } from '../../../registry/model-registry.js';
+import { getModel } from '../../../../../server/bare/registry/model-registry.js';
 import { CompletionFailedError } from '../../../../../utils/errors-server.js';
-import { getRequestRegistry, withRequestContext } from '../../../runtime/index.js';
-import { generateServerRequestId } from '../../../runtime/request-id.js';
+import { getRequestRegistry, withRequestContext } from '../../../../../server/bare/runtime/index.js';
+import { generateServerRequestId } from '../../../../../server/bare/runtime/request-id.js';
 import { getServerLogger } from '../../../../../logging/index.js';
 const PAUSE_CHECKPOINT_PREFIX = 'pause_checkpoint_step_';
 const finetuneRuntimeState = new Set();

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

@@ -1,6 +1,6 @@
-import { generateConfigHash } from '../../../ops/kv-cache-utils';
-import type { CacheMessage } from '../../../../utils/index';
-import type { Logger } from '../../../../../logging/types';
+import { generateConfigHash } from '../../../../../server/bare/ops/kv-cache-utils.js';
+import { type CacheMessage } from '../../../../../server/utils/index.js';
+import type { Logger } from '../../../../../logging/types.js';
 export interface TurnHandle {
     /** Resolved on-disk cache file path the addon will read from / write to. */
     readonly cachePath: string;
@@ -59,7 +59,8 @@ export interface KvCacheSession {
      * primes the system-prompt cache if needed (delegated to
      * `input.primeIfMissing`), marks the cache initialized, and returns a
      * `TurnHandle` the handler attaches to `ctx.scope.defer(...)` for the
-     * rollback hook.
+     * rollback hook. Auto-cache path resolution is serialized with
+     * retention deletion before the handle is returned.
      */
     beginTurn(input: BeginTurnInput): Promise<TurnHandle>;
     /**
@@ -72,12 +73,13 @@ export interface KvCacheSession {
     commitTurn(turn: TurnHandle, result: CommitResult): Promise<void>;
     /**
      * Roll back an in-flight turn — atomically deletes the on-disk cache
-     * file, clears the in-memory `initializedCaches` entry, and forgets
-     * the `cachedMessageCounts` entry. **All three layers, always, in
-     * one place.** Idempotent: a turn that has already been committed
-     * or rolled back is a no-op on subsequent calls. Handlers register
-     * this via `ctx.scope.defer(...)` so it runs regardless of how the
-     * handler exits (success branch removes itself via `commitTurn`).
+     * file, clears the in-memory `initializedCaches` entry, forgets the
+     * `cachedMessageCounts` entry, releases the active-path ref, and
+     * removes orphaned marker metadata. Idempotent: a turn that has
+     * already been committed or rolled back is a no-op on subsequent
+     * calls. Handlers register this via `ctx.scope.defer(...)` so it
+     * runs regardless of how the handler exits (success branch removes
+     * itself via `commitTurn`).
      */
     rollback(turn: TurnHandle): Promise<void>;
     /**
@@ -148,7 +150,15 @@ export declare const __kvCacheSessionTestHooks: {
     setSavedCountForTest(cachePath: string, count: number): void;
     hasInitializedKey(modelId: string, configHash: string, cacheKey: string): boolean;
     markInitializedForTest(modelId: string, configHash: string, cacheKey: string): void;
+    getLastAutoCacheSweepMsForTest(): number;
+    setLastAutoCacheSweepMsForTest(value: number): void;
     resetForTest(): void;
+    waitForAutoCacheSweepForTest(): Promise<void>;
+    sweepAutoCachesForTest(options: {
+        maxBytes: number;
+        maxIdleMs: number;
+        nowMs: number;
+    }): Promise<void>;
 };
 export { generateConfigHash };
 //# sourceMappingURL=kv-cache-session.d.ts.map