@qvac/sdk 0.18.2 → 0.19.0

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

@@ -1,32 +0,0 @@
-import type { AbortSignal } from 'bare-abort-controller';
-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[];
-} | {
-    type: 'token';
-    id: string;
-    token: string;
-};
-type BatchModelResult = {
-    id: string;
-    output: string;
-};
-type BatchModelStreamResult = {
-    ids: string[];
-    results: BatchModelResult[];
-    modelExecutionMs: number;
-    stats?: CompletionStats;
-};
-export declare function batchCompletion(params: {
-    modelId: string;
-    prompts: BatchCompletionStreamPrompt[];
-}, opts: {
-    signal: AbortSignal;
-    scope: DisposableScope;
-    logger?: Logger;
-}): AsyncGenerator<BatchModelEvent, BatchModelStreamResult, unknown>;
-export {};
-//# sourceMappingURL=batch-completion-stream.d.ts.map

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

@@ -1,100 +0,0 @@
-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 '../../../../../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 { prependToolsToHistory } from '../../../../../server/utils/tool-integration.js';
-const logger = getServerLogger();
-function runBatchModel(model, prompts) {
-    const run = model.run.bind(model);
-    return run(prompts);
-}
-function mergeGenerationParams(generationParams, responseFormat) {
-    if (!responseFormat)
-        return generationParams;
-    const jsonSchema = getResponseFormatJsonSchema(responseFormat);
-    if (jsonSchema === undefined)
-        return generationParams;
-    return {
-        ...(generationParams ?? {}),
-        json_schema: jsonSchema
-    };
-}
-function renderPromptHistory(prompt, options) {
-    const tools = options.toolsEnabled && prompt.tools && prompt.tools.length > 0 ? prompt.tools : undefined;
-    let historyWithTools = prompt.history;
-    if (tools) {
-        historyWithTools = prependToolsToHistory(prompt.history, tools);
-    }
-    // Uses the same attachment expansion as single completion: each SDK
-    // attachment becomes an addon `type: "media"` message before the text turn.
-    return transformMessages(historyWithTools);
-}
-function buildBatchPrompt(prompt, options) {
-    const mergedGenerationParams = mergeGenerationParams(prompt.generationParams, prompt.responseFormat);
-    return {
-        ...(prompt.id !== undefined && { id: prompt.id }),
-        prompt: renderPromptHistory(prompt, options),
-        ...(mergedGenerationParams && {
-            runOptions: { generationParams: mergedGenerationParams }
-        })
-    };
-}
-function isBatchOutputChunk(output) {
-    return (output !== null &&
-        typeof output === 'object' &&
-        'id' in output &&
-        'chunk' in output &&
-        typeof output.id === 'string' &&
-        typeof output.chunk === 'string');
-}
-export async function* batchCompletion(params, opts) {
-    const { modelId, prompts } = params;
-    const { signal, scope } = opts;
-    const requestLogger = opts.logger ?? logger;
-    const model = getModel(modelId);
-    const modelConfig = getModelConfig(modelId);
-    const renderOptions = {
-        toolsEnabled: modelConfig.tools === true
-    };
-    // Cancel via the batch response, mirroring completion(): the addon routes
-    // response.cancel() to cancelJob for this batch group only, so a concurrent
-    // peer batch or completion on the same model keeps decoding.
-    let activeResponse = null;
-    const cancelActive = () => {
-        activeResponse?.cancel().catch((err) => {
-            requestLogger.warn(`[cancel] batch response.cancel() rejected during abort for modelId=${modelId}: ${err instanceof Error ? err.message : String(err)}`);
-        });
-    };
-    const onAbort = () => cancelActive();
-    signal.addEventListener('abort', onAbort, { once: true });
-    if (signal.aborted)
-        onAbort();
-    scope.defer(() => {
-        signal.removeEventListener('abort', onAbort);
-        activeResponse = null;
-    });
-    const addonPrompts = prompts.map((prompt) => buildBatchPrompt(prompt, renderOptions));
-    const modelStart = nowMs();
-    const response = await runBatchModel(model, addonPrompts);
-    activeResponse = response;
-    if (signal.aborted)
-        cancelActive();
-    const ids = response.ids;
-    yield { type: 'ids', ids };
-    for await (const output of response.iterate()) {
-        if (!isBatchOutputChunk(output))
-            continue;
-        yield { type: 'token', id: output.id, token: output.chunk };
-    }
-    const results = await response.await();
-    const modelExecutionMs = nowMs() - modelStart;
-    const stats = normalizeCompletionStats(response.stats);
-    return {
-        ...buildStreamResult(modelExecutionMs, stats),
-        ids,
-        results
-    };
-}

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

@@ -1,7 +0,0 @@
-export declare function logCacheStatus(cacheKey: string, isReusing: boolean): void;
-export declare function logCacheInit(cacheKey: string, systemPrompt: string): void;
-export declare function logMessagesToAddon(messages: unknown[], phase?: 'CACHE_INIT' | 'PROMPT_SEND' | 'NO_CACHE'): void;
-export declare function logCacheDisabled(): void;
-export declare function logCacheSave(sessionPath: string): void;
-export declare function logCacheSaveError(sessionPath: string, err: unknown): void;
-//# sourceMappingURL=cache-logger.d.ts.map

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

@@ -1,36 +0,0 @@
-import { getServerLogger } from '../../../../../logging/index.js';
-const logger = getServerLogger();
-function formatMessages(messages) {
-    return messages
-        .map((m) => {
-        const role = m.role || m.type || '?';
-        const content = m.content
-            ? m.content.length > 40
-                ? m.content.substring(0, 40) + '...'
-                : m.content
-            : m.name || '(no content)';
-        return `{${role}: "${content}"}`;
-    })
-        .join(', ');
-}
-export function logCacheStatus(cacheKey, isReusing) {
-    const status = isReusing ? 'REUSING' : 'CREATING';
-    logger.debug(`[kv-cache] [${cacheKey}] ${status} cache`);
-}
-export function logCacheInit(cacheKey, systemPrompt) {
-    const promptLen = systemPrompt.length;
-    logger.debug(`[kv-cache] [${cacheKey}] Initializing cache (prompt: ${promptLen} chars)`);
-}
-export function logMessagesToAddon(messages, phase = 'PROMPT_SEND') {
-    const typedMessages = messages;
-    logger.debug(`[kv-cache] [${phase}] Sending ${typedMessages.length} msg(s): [${formatMessages(typedMessages)}]`);
-}
-export function logCacheDisabled() {
-    logger.debug('[kv-cache] Cache disabled');
-}
-export function logCacheSave(sessionPath) {
-    logger.debug(`[kv-cache] Saving session: ...${sessionPath.slice(-20)}`);
-}
-export function logCacheSaveError(sessionPath, err) {
-    logger.warn(`[kv-cache] Failed to save session: ...${sessionPath.slice(-20)} — ${err instanceof Error ? err.message : String(err)}`);
-}

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

@@ -1,27 +0,0 @@
-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;
-export declare function stoppedByLength({ cancelled, effectivePredict, generatedTokens, stoppedAtContextBoundary }: {
-    cancelled: boolean;
-    effectivePredict: number | undefined;
-    generatedTokens: number | undefined;
-    stoppedAtContextBoundary: boolean;
-}): boolean;
-//# sourceMappingURL=completion-stats.d.ts.map

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

@@ -1,58 +0,0 @@
-function finiteNumber(value) {
-    return value !== undefined && Number.isFinite(value) ? value : undefined;
-}
-export function normalizeCompletionStats(stats) {
-    if (!stats)
-        return undefined;
-    const timeToFirstToken = finiteNumber(stats.TTFT);
-    const tokensPerSecond = finiteNumber(stats.TPS);
-    const cacheTokens = finiteNumber(stats.CacheTokens);
-    const promptTokens = finiteNumber(stats.promptTokens);
-    const generatedTokens = finiteNumber(stats.generatedTokens);
-    const avgConcurrentSeq = finiteNumber(stats.avgConcurrentSeq);
-    const normalized = {
-        ...(timeToFirstToken !== undefined && { timeToFirstToken }),
-        ...(tokensPerSecond !== undefined && { tokensPerSecond }),
-        ...(cacheTokens !== undefined && { cacheTokens }),
-        ...(promptTokens !== undefined && { promptTokens }),
-        ...(generatedTokens !== undefined && { generatedTokens }),
-        ...(avgConcurrentSeq !== undefined && { avgConcurrentSeq }),
-        ...(stats.backendDevice !== undefined && { backendDevice: stats.backendDevice })
-    };
-    if (timeToFirstToken === undefined &&
-        tokensPerSecond === undefined &&
-        cacheTokens === undefined &&
-        promptTokens === undefined &&
-        generatedTokens === undefined &&
-        avgConcurrentSeq === undefined &&
-        stats.backendDevice === undefined) {
-        return undefined;
-    }
-    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
-    };
-}
-export function stoppedByLength({ cancelled, effectivePredict, generatedTokens, stoppedAtContextBoundary }) {
-    if (cancelled)
-        return false;
-    if (stoppedAtContextBoundary)
-        return true;
-    return (effectivePredict !== undefined &&
-        effectivePredict > 0 &&
-        generatedTokens !== undefined &&
-        generatedTokens >= effectivePredict);
-}

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

@@ -1,47 +0,0 @@
-import type { AbortSignal } from 'bare-abort-controller';
-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;
-    toolCalls: ToolCall[];
-    stoppedAtContextBoundary: boolean;
-}
-interface ChatHistory {
-    role?: string;
-    content?: string;
-    type?: string;
-    name?: string;
-    description?: string;
-    parameters?: unknown;
-}
-export type CompletionGenerationParams = GenerationParams & {
-    json_schema?: string;
-};
-export declare function transformMessages(messages: Array<{
-    role: string;
-    content: string;
-    attachments?: {
-        path: string;
-    }[] | undefined;
-} | Tool>): ChatHistory[];
-export declare function completion(params: CompletionParams & {
-    tools?: Tool[];
-    generationParams?: GenerationParams;
-    toolDialect?: ToolDialect;
-    responseFormat?: ResponseFormat;
-}, opts: {
-    signal: AbortSignal;
-    scope: DisposableScope;
-    /**
-     * Request-scoped logger forwarded to `createKvCacheSession` so
-     * kv-cache lines share the request's lifecycle prefix. Falls
-     * back to the module-level server logger when omitted.
-     */
-    logger?: Logger;
-}): AsyncGenerator<{
-    token: string;
-}, CompletionResult, unknown>;
-export {};
-//# sourceMappingURL=completion-stream.d.ts.map