@qvac/sdk 0.17.1 → 0.18.0

package/dist/models/update-models/processing.js

@@ -1,65 +0,0 @@
-import { getAddonFromEngine, resolveCanonicalEngine } from '../../schemas/engine-addon-map.js';
-import { detectShardedModel } from './shards.js';
-export function toHexString(value) {
-    if (!value)
-        return '';
-    if (Buffer.isBuffer(value))
-        return value.toString('hex');
-    if (typeof value === 'string')
-        return value;
-    if (typeof value === 'object' && 'data' in value) {
-        return Buffer.from(value.data).toString('hex');
-    }
-    return '';
-}
-export function extractModelName(registryPath) {
-    const parts = registryPath.split('/');
-    if (parts.length >= 2) {
-        return parts[1] || parts[0] || '';
-    }
-    return (registryPath
-        .split('/')
-        .pop()
-        ?.replace(/\.\w+$/, '') || '');
-}
-export function processRegistryModel(model) {
-    const engine = resolveCanonicalEngine(model.engine);
-    if (!engine) {
-        console.warn(`⚠️  Skipping model with unknown engine "${model.engine}": ${model.path}`);
-        return null;
-    }
-    const filename = model.path.split('/').pop() || model.path;
-    const blobBinding = model.blobBinding;
-    const blobCoreKey = toHexString(blobBinding?.coreKey);
-    const blobBlockOffset = blobBinding?.blockOffset ?? 0;
-    const blobBlockLength = blobBinding?.blockLength ?? 0;
-    const blobByteOffset = blobBinding?.byteOffset ?? 0;
-    const expectedSize = blobBinding?.byteLength ?? 0;
-    // The sha256 lives on blobBinding at runtime (per the hyperschema),
-    // even though the TS types define it on QVACModelEntry. Try blobBinding first.
-    const sha256Checksum = blobBinding?.['sha256'] || model.sha256 || '';
-    const addon = getAddonFromEngine(engine);
-    const result = {
-        registryPath: model.path,
-        registrySource: model.source,
-        blobCoreKey,
-        blobBlockOffset,
-        blobBlockLength,
-        blobByteOffset,
-        modelId: filename,
-        addon,
-        expectedSize,
-        sha256Checksum,
-        engine,
-        modelName: extractModelName(model.path),
-        quantization: model.quantization || '',
-        params: model.params || '',
-        tags: model.tags || []
-    };
-    const shardDetection = detectShardedModel(filename);
-    if (shardDetection.isSharded) {
-        result.isShardPart = true;
-        result.shardInfo = shardDetection;
-    }
-    return result;
-}

package/dist/models/update-models/registry.d.ts

@@ -1,5 +0,0 @@
-import type { CollectOptions, ProcessedModel } from './types.js';
-export { processRegistryModel, extractModelName, toHexString } from './processing.js';
-export declare function collectModels(options?: CollectOptions): Promise<ProcessedModel[]>;
-export declare function deduplicateModels(models: ProcessedModel[], showDuplicates: boolean): ProcessedModel[];
-//# sourceMappingURL=registry.d.ts.map

package/dist/models/update-models/registry.js

@@ -1,87 +0,0 @@
-import { QVACRegistryClient } from '@qvac/registry-client';
-import { groupShardedModels } from './shards.js';
-import { groupCompanionSets } from './companions.js';
-import { processRegistryModel } from './processing.js';
-import { DEFAULT_REGISTRY_CORE_KEY } from '../../constants/index.js';
-// Re-export for backward compat
-export { processRegistryModel, extractModelName, toHexString } from './processing.js';
-export async function collectModels(options = {}) {
-    const { showDuplicates = false, noDedup = false } = options;
-    const models = [];
-    // eslint-disable-next-line @typescript-eslint/no-unsafe-assignment
-    const registryCoreKey = process.env['QVAC_REGISTRY_CORE_KEY'] ?? DEFAULT_REGISTRY_CORE_KEY;
-    const client = new QVACRegistryClient({ registryCoreKey });
-    try {
-        await client.ready();
-        const registryModels = await client.findModels({});
-        console.log(`📦 Found ${registryModels.length} entries in registry`);
-        for (const registryModel of registryModels) {
-            const processed = processRegistryModel(registryModel);
-            if (processed)
-                models.push(processed);
-        }
-    }
-    finally {
-        await client.close();
-    }
-    const groupedModels = groupShardedModels(models);
-    const withCompanions = groupCompanionSets(groupedModels);
-    if (noDedup) {
-        console.log(`\n⏭️  Skipping deduplication (--no-dedup flag set)`);
-        return withCompanions;
-    }
-    return deduplicateModels(withCompanions, showDuplicates);
-}
-export function deduplicateModels(models, showDuplicates) {
-    // Paths referenced as companion files in any companion set. Dropping one of
-    // these on sha256 collision would leave its companion set without a
-    // standalone RegistryItem that `getModelByPath()` can resolve, causing
-    // recursive registry resolutions (e.g. the Bergamot shared vocab) to fall
-    // into the metadata-less fallback where `expectedSize` is `0`.
-    const companionReferencedPaths = new Set();
-    for (const m of models) {
-        if (!m.companionSet)
-            continue;
-        for (const f of m.companionSet.files) {
-            companionReferencedPaths.add(`${f.registrySource}:${f.registryPath}`);
-        }
-    }
-    const seenChecksums = new Map();
-    const dedupedModels = [];
-    const skipped = [];
-    for (const model of models) {
-        if (!model.sha256Checksum || model.sha256Checksum === '') {
-            dedupedModels.push(model);
-            continue;
-        }
-        if (seenChecksums.has(model.sha256Checksum)) {
-            const pathKey = `${model.registrySource}:${model.registryPath}`;
-            if (companionReferencedPaths.has(pathKey)) {
-                dedupedModels.push(model);
-                continue;
-            }
-            skipped.push({
-                name: model.registryPath,
-                checksum: model.sha256Checksum,
-                reason: `Duplicate of ${seenChecksums.get(model.sha256Checksum)}`
-            });
-            continue;
-        }
-        seenChecksums.set(model.sha256Checksum, model.registryPath);
-        dedupedModels.push(model);
-    }
-    if (skipped.length > 0) {
-        console.log(`\n🧹 Removed ${skipped.length} duplicate model(s)`);
-        if (showDuplicates) {
-            skipped.forEach(({ name, checksum, reason }) => {
-                console.log(`  - ${name}`);
-                console.log(`    Checksum: ${checksum}`);
-                console.log(`    ${reason}`);
-            });
-        }
-        else {
-            console.log(`   Use --show-duplicates to see details`);
-        }
-    }
-    return dedupedModels;
-}

package/dist/models/update-models/shards.d.ts

@@ -1,4 +0,0 @@
-import type { ProcessedModel, ShardDetection } from './types.js';
-export declare function detectShardedModel(filename: string): ShardDetection;
-export declare function groupShardedModels(models: ProcessedModel[]): ProcessedModel[];
-//# sourceMappingURL=shards.d.ts.map

package/dist/models/update-models/shards.js

@@ -1,57 +0,0 @@
-export function detectShardedModel(filename) {
-    const shardPattern = /^(.+)-(\d{5})-of-(\d{5})(\.\w+)$/;
-    const match = filename.match(shardPattern);
-    if (match) {
-        return {
-            isSharded: true,
-            baseFilename: match[1],
-            currentShard: parseInt(match[2], 10),
-            totalShards: parseInt(match[3], 10),
-            extension: match[4]
-        };
-    }
-    return { isSharded: false };
-}
-export function groupShardedModels(models) {
-    const shardGroups = new Map();
-    const nonShardedModels = [];
-    for (const model of models) {
-        if (model.isShardPart && model.shardInfo) {
-            const baseKey = `${model.registrySource}:${model.shardInfo.baseFilename}`;
-            if (!shardGroups.has(baseKey)) {
-                shardGroups.set(baseKey, []);
-            }
-            shardGroups.get(baseKey).push(model);
-        }
-        else {
-            nonShardedModels.push(model);
-        }
-    }
-    const processedShards = [];
-    for (const [baseKey, shards] of shardGroups) {
-        shards.sort((a, b) => (a.shardInfo?.currentShard ?? 0) - (b.shardInfo?.currentShard ?? 0));
-        const firstShard = shards[0];
-        const totalExpectedShards = firstShard.shardInfo?.totalShards ?? 0;
-        if (shards.length !== totalExpectedShards) {
-            console.warn(`⚠️  Expected ${totalExpectedShards} shards but found ${shards.length} for ${baseKey}`);
-        }
-        const totalSize = shards.reduce((sum, s) => sum + s.expectedSize, 0);
-        const shardMetadata = shards.map((s) => ({
-            filename: s.modelId,
-            expectedSize: s.expectedSize,
-            sha256Checksum: s.sha256Checksum,
-            blobCoreKey: s.blobCoreKey,
-            blobBlockOffset: s.blobBlockOffset,
-            blobBlockLength: s.blobBlockLength,
-            blobByteOffset: s.blobByteOffset
-        }));
-        // eslint-disable-next-line @typescript-eslint/no-unused-vars
-        const { isShardPart: _shard, shardInfo: _info, ...rest } = firstShard;
-        processedShards.push({
-            ...rest,
-            expectedSize: totalSize,
-            shardMetadata
-        });
-    }
-    return [...nonShardedModels, ...processedShards];
-}

package/dist/models/update-models/types.d.ts

@@ -1,88 +0,0 @@
-import type { ModelRegistryEntryAddon, ModelRegistryEngine } from '../../schemas/registry.js';
-export interface ShardInfo {
-    isSharded: true;
-    baseFilename: string;
-    currentShard: number;
-    totalShards: number;
-    extension: string;
-}
-export interface NotSharded {
-    isSharded: false;
-}
-export type ShardDetection = ShardInfo | NotSharded;
-export interface BlobRef {
-    expectedSize: number;
-    sha256Checksum: string;
-    blobCoreKey: string;
-    blobBlockOffset: number;
-    blobBlockLength: number;
-    blobByteOffset: number;
-}
-export interface ShardMetadataEntry extends BlobRef {
-    filename: string;
-}
-/**
- * A single file within a companion set (e.g., an `.onnx` model or its `_data` file).
- *
- * @property key - Identifier for this file within the set
- *   Used by plugins to look up resolved paths from the download artifacts.
- * @property targetName - Filename used when writing to the cache directory.
- * @property primary - When true, this is the main file whose resolved path is returned to
- *   the addon. Exactly one file per set should be primary.
- */
-export interface CompanionSetMetadataEntry extends BlobRef {
-    key: string;
-    registryPath: string;
-    registrySource: string;
-    targetName: string;
-    primary?: boolean;
-}
-/**
- * Describes a group of files that must be downloaded and cached together
- *
- * @property setKey - Deterministic hash identifying this set. Used as the cache
- *   subdirectory: `<cacheDir>/sets/<setKey>/`.
- * @property primaryKey - The `key` value of the primary file in the `files` array.
- *   The resolver returns this file's path to the addon after download.
- * @property files - All files in the set, including the primary.
- */
-export interface CompanionSetMetadata {
-    setKey: string;
-    primaryKey: string;
-    files: readonly CompanionSetMetadataEntry[];
-}
-export interface ProcessedModel extends BlobRef {
-    registryPath: string;
-    registrySource: string;
-    modelId: string;
-    addon: ModelRegistryEntryAddon;
-    engine: ModelRegistryEngine;
-    modelName: string;
-    quantization: string;
-    params: string;
-    tags: string[];
-    isShardPart?: boolean;
-    shardInfo?: ShardInfo;
-    shardMetadata?: ShardMetadataEntry[];
-    companionSet?: CompanionSetMetadata;
-    isCompanionOnly?: boolean;
-    name?: string;
-}
-export interface CurrentModel {
-    name: string;
-    registryPath: string;
-}
-export interface CollectOptions {
-    showDuplicates?: boolean;
-    noDedup?: boolean;
-}
-export interface ExportNameInput {
-    path: string;
-    engine: ModelRegistryEngine;
-    name: string;
-    quantization: string;
-    params: string;
-    tags: string[];
-    usedNames: Set<string>;
-}
-//# sourceMappingURL=types.d.ts.map

package/dist/models/update-models/types.js

@@ -1 +0,0 @@
-export {};

package/dist/models/update-models/utils.d.ts

@@ -1,3 +0,0 @@
-export declare function formatSize(bytes: number): string;
-export declare function getCommitHash(short?: boolean): string;
-//# sourceMappingURL=utils.d.ts.map

package/dist/models/update-models/utils.js

@@ -1,21 +0,0 @@
-import { execSync } from 'child_process';
-export function formatSize(bytes) {
-    if (bytes < 1024)
-        return `${bytes}B`;
-    if (bytes < 1024 * 1024)
-        return `${(bytes / 1024).toFixed(1)}KB`;
-    if (bytes < 1024 * 1024 * 1024)
-        return `${(bytes / (1024 * 1024)).toFixed(1)}MB`;
-    return `${(bytes / (1024 * 1024 * 1024)).toFixed(1)}GB`;
-}
-export function getCommitHash(short = false) {
-    try {
-        const cmd = short ? 'git rev-parse --short HEAD' : 'git rev-parse HEAD';
-        return execSync(cmd, { encoding: 'utf-8' }).trim();
-    }
-    catch (error) {
-        throw new Error('Git is required to generate history file', {
-            cause: error
-        });
-    }
-}

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

@@ -1,23 +0,0 @@
-/**
- * Helper for detecting the llama.cpp addon's `Cancelled` status error
- * (`LlmErrors.hpp::Cancelled = 27`) on the continuous-batching path.
- *
- * When a batch is cancelled while it still has prompts queued behind the
- * `parallel` slot limit, `ContinuousBatchScheduler::cancelPendingLocked`
- * fails the whole batch group with a `Cancelled` `StatusError`, which
- * `processBatch` rethrows — so `model.run(prompts)` rejects rather than
- * resolving with partial outputs (see
- * `packages/llm-llamacpp/docs/continuous-batching.md` "Cancellation
- * semantics"). In-flight slots are still cancelled gracefully, but the
- * shared group carries the error.
- *
- * The addon's JS-facing throw goes through `js_throw_error(env, code,
- * msg)` (see `JsUtils.hpp::JSCATCH`) where `code` comes from
- * `qvac_errors::StatusError::codeString()` and formats as
- * `"[ <addonId> :: Cancelled ]"`. This helper lets the batch handler
- * recognise that rejection so it can complete the run with cancelled
- * terminal events (matching the SDK's graceful cancel semantics) instead
- * of surfacing a generic `CompletionFailedError`.
- */
-export declare function isAddonCancelledError(err: unknown): boolean;
-//# sourceMappingURL=batch-cancelled.d.ts.map

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

@@ -1,30 +0,0 @@
-/**
- * Helper for detecting the llama.cpp addon's `Cancelled` status error
- * (`LlmErrors.hpp::Cancelled = 27`) on the continuous-batching path.
- *
- * When a batch is cancelled while it still has prompts queued behind the
- * `parallel` slot limit, `ContinuousBatchScheduler::cancelPendingLocked`
- * fails the whole batch group with a `Cancelled` `StatusError`, which
- * `processBatch` rethrows — so `model.run(prompts)` rejects rather than
- * resolving with partial outputs (see
- * `packages/llm-llamacpp/docs/continuous-batching.md` "Cancellation
- * semantics"). In-flight slots are still cancelled gracefully, but the
- * shared group carries the error.
- *
- * The addon's JS-facing throw goes through `js_throw_error(env, code,
- * msg)` (see `JsUtils.hpp::JSCATCH`) where `code` comes from
- * `qvac_errors::StatusError::codeString()` and formats as
- * `"[ <addonId> :: Cancelled ]"`. This helper lets the batch handler
- * recognise that rejection so it can complete the run with cancelled
- * terminal events (matching the SDK's graceful cancel semantics) instead
- * of surfacing a generic `CompletionFailedError`.
- */
-export function isAddonCancelledError(err) {
-    if (typeof err !== 'object' || err === null)
-        return false;
-    const code = err.code;
-    // Anchor to the codeString tail (`[ <addonId> :: Cancelled ]`) so a
-    // hypothetical sibling like `CancelledByPolicy` after an addon-side
-    // rename doesn't false-positive here.
-    return typeof code === 'string' && /::\s*Cancelled\s*\]/.test(code);
-}

package/dist/server/utils/expand-gguf-shards.d.ts

@@ -1,3 +0,0 @@
-/** Expand a sharded GGUF path into `.tensors.txt` plus all shards. */
-export declare function expandGGUFIntoShards(modelPath: string): string[];
-//# sourceMappingURL=expand-gguf-shards.d.ts.map

package/dist/server/utils/expand-gguf-shards.js

@@ -1,22 +0,0 @@
-const SHARD_PATTERN = /^(.+)-(\d{5})-of-(\d{5})\.gguf$/;
-/** Expand a sharded GGUF path into `.tensors.txt` plus all shards. */
-export function expandGGUFIntoShards(modelPath) {
-    const lastSep = Math.max(modelPath.lastIndexOf('/'), modelPath.lastIndexOf('\\'));
-    const dir = lastSep >= 0 ? modelPath.slice(0, lastSep) : '';
-    const sep = lastSep >= 0 ? modelPath.charAt(lastSep) : '/';
-    const filename = lastSep >= 0 ? modelPath.slice(lastSep + 1) : modelPath;
-    const match = filename.match(SHARD_PATTERN);
-    if (!match || !match[1] || !match[3])
-        return [modelPath];
-    const baseFilename = match[1];
-    const totalDigits = match[3];
-    const totalShards = Number.parseInt(totalDigits, 10);
-    if (totalShards <= 0)
-        return [modelPath];
-    const join = (name) => (dir ? `${dir}${sep}${name}` : name);
-    const shards = [join(`${baseFilename}.tensors.txt`)];
-    for (let i = 1; i <= totalShards; i++) {
-        shards.push(join(`${baseFilename}-${String(i).padStart(5, '0')}-of-${totalDigits}.gguf`));
-    }
-    return shards;
-}