@atomic-chat/core 0.5.1

package/dist/models/embed.js

@@ -0,0 +1,83 @@
+/** Embeddings through a core-owned llama-server, preserving the extension's batching contract. */
+import { AtomicCoreError } from '../contracts/index.js';
+export class EmbedService {
+    deps;
+    constructor(deps) {
+        this.deps = deps;
+    }
+    async embed(provider, modelId, input, ubatchSize) {
+        if (!Array.isArray(input) || input.some((value) => typeof value !== 'string')) {
+            throw new AtomicCoreError('INVALID_ARGUMENT', 'input must be an array of strings');
+        }
+        if (!Number.isSafeInteger(ubatchSize) || ubatchSize < 2) {
+            throw new AtomicCoreError('INVALID_ARGUMENT', 'ubatch_size must be an integer of at least 2');
+        }
+        let session = this.deps.findSession(provider, modelId) ?? (await this.deps.load(provider, modelId));
+        let reloadedForEmbedding = false;
+        const result = {
+            model: session.model_id,
+            object: 'list',
+            usage: { prompt_tokens: 0, total_tokens: 0 },
+            data: [],
+        };
+        for (const { batch, offset } of buildEmbedBatches(input, ubatchSize)) {
+            let response = await this.request(session, batch);
+            if (response.status === 501 && !reloadedForEmbedding) {
+                reloadedForEmbedding = true;
+                await this.deps.unload(provider, modelId).catch(() => { });
+                session = await this.deps.load(provider, modelId);
+                response = await this.request(session, batch);
+            }
+            if (!response.ok) {
+                const body = await response.text().catch(() => '');
+                throw new AtomicCoreError('IO_ERROR', `Embedding request failed with HTTP ${response.status}`, body);
+            }
+            const body = (await response.json());
+            result.usage.prompt_tokens += body.usage?.prompt_tokens ?? 0;
+            result.usage.total_tokens += body.usage?.total_tokens ?? 0;
+            for (const item of body.data ?? [])
+                result.data.push({ ...item, index: item.index + offset });
+        }
+        result.model = session.model_id;
+        return result;
+    }
+    request(session, input) {
+        return (this.deps.fetch ?? fetch)(`http://127.0.0.1:${session.port}/v1/embeddings`, {
+            method: 'POST',
+            headers: { 'Content-Type': 'application/json', 'Authorization': `Bearer ${session.api_key}` },
+            body: JSON.stringify({ input, model: session.model_id, encoding_format: 'float' }),
+        });
+    }
+}
+/** Same 3-chars/token, 50% safety-margin packing the legacy extension uses. */
+export function buildEmbedBatches(input, ubatchSize) {
+    const limit = Math.floor(ubatchSize * 0.5);
+    const batches = [];
+    let batch = [];
+    let tokens = 0;
+    let offset = 0;
+    const flush = () => {
+        if (batch.length === 0)
+            return;
+        batches.push({ batch, offset });
+        offset += batch.length;
+        batch = [];
+        tokens = 0;
+    };
+    for (const text of input) {
+        const estimated = Math.max(1, Math.ceil(text.length / 3));
+        if (estimated > limit) {
+            flush();
+            batches.push({ batch: [text], offset });
+            offset++;
+            continue;
+        }
+        if (batch.length && tokens + estimated > limit)
+            flush();
+        batch.push(text);
+        tokens += estimated;
+    }
+    flush();
+    return batches;
+}
+//# sourceMappingURL=embed.js.map

package/dist/models/embed.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"embed.js","sourceRoot":"","sources":["../../src/models/embed.ts"],"names":[],"mappings":"AAAA,kGAAkG;AAClG,OAAO,EAAE,eAAe,EAAE,MAAM,uBAAuB,CAAA;AAsBvD,MAAM,OAAO,YAAY;IACM;IAA7B,YAA6B,IAAe;QAAf,SAAI,GAAJ,IAAI,CAAW;IAAG,CAAC;IAEhD,KAAK,CAAC,KAAK,CACT,QAAyB,EACzB,OAAe,EACf,KAAe,EACf,UAAkB;QAElB,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC,IAAI,KAAK,CAAC,IAAI,CAAC,CAAC,KAAK,EAAE,EAAE,CAAC,OAAO,KAAK,KAAK,QAAQ,CAAC,EAAE,CAAC;YAC9E,MAAM,IAAI,eAAe,CAAC,kBAAkB,EAAE,mCAAmC,CAAC,CAAA;QACpF,CAAC;QACD,IAAI,CAAC,MAAM,CAAC,aAAa,CAAC,UAAU,CAAC,IAAI,UAAU,GAAG,CAAC,EAAE,CAAC;YACxD,MAAM,IAAI,eAAe,CAAC,kBAAkB,EAAE,8CAA8C,CAAC,CAAA;QAC/F,CAAC;QACD,IAAI,OAAO,GAAG,IAAI,CAAC,IAAI,CAAC,WAAW,CAAC,QAAQ,EAAE,OAAO,CAAC,IAAI,CAAC,MAAM,IAAI,CAAC,IAAI,CAAC,IAAI,CAAC,QAAQ,EAAE,OAAO,CAAC,CAAC,CAAA;QACnG,IAAI,oBAAoB,GAAG,KAAK,CAAA;QAChC,MAAM,MAAM,GAAsB;YAChC,KAAK,EAAE,OAAO,CAAC,QAAQ;YACvB,MAAM,EAAE,MAAM;YACd,KAAK,EAAE,EAAE,aAAa,EAAE,CAAC,EAAE,YAAY,EAAE,CAAC,EAAE;YAC5C,IAAI,EAAE,EAAE;SACT,CAAA;QACD,KAAK,MAAM,EAAE,KAAK,EAAE,MAAM,EAAE,IAAI,iBAAiB,CAAC,KAAK,EAAE,UAAU,CAAC,EAAE,CAAC;YACrE,IAAI,QAAQ,GAAG,MAAM,IAAI,CAAC,OAAO,CAAC,OAAO,EAAE,KAAK,CAAC,CAAA;YACjD,IAAI,QAAQ,CAAC,MAAM,KAAK,GAAG,IAAI,CAAC,oBAAoB,EAAE,CAAC;gBACrD,oBAAoB,GAAG,IAAI,CAAA;gBAC3B,MAAM,IAAI,CAAC,IAAI,CAAC,MAAM,CAAC,QAAQ,EAAE,OAAO,CAAC,CAAC,KAAK,CAAC,GAAG,EAAE,GAAE,CAAC,CAAC,CAAA;gBACzD,OAAO,GAAG,MAAM,IAAI,CAAC,IAAI,CAAC,IAAI,CAAC,QAAQ,EAAE,OAAO,CAAC,CAAA;gBACjD,QAAQ,GAAG,MAAM,IAAI,CAAC,OAAO,CAAC,OAAO,EAAE,KAAK,CAAC,CAAA;YAC/C,CAAC;YACD,IAAI,CAAC,QAAQ,CAAC,EAAE,EAAE,CAAC;gBACjB,MAAM,IAAI,GAAG,MAAM,QAAQ,CAAC,IAAI,EAAE,CAAC,KAAK,CAAC,GAAG,EAAE,CAAC,EAAE,CAAC,CAAA;gBAClD,MAAM,IAAI,eAAe,CAAC,UAAU,EAAE,sCAAsC,QAAQ,CAAC,MAAM,EAAE,EAAE,IAAI,CAAC,CAAA;YACtG,CAAC;YACD,MAAM,IAAI,GAAG,CAAC,MAAM,QAAQ,CAAC,IAAI,EAAE,CAAkB,CAAA;YACrD,MAAM,CAAC,KAAK,CAAC,aAAa,IAAI,IAAI,CAAC,KAAK,EAAE,aAAa,IAAI,CAAC,CAAA;YAC5D,MAAM,CAAC,KAAK,CAAC,YAAY,IAAI,IAAI,CAAC,KAAK,EAAE,YAAY,IAAI,CAAC,CAAA;YAC1D,KAAK,MAAM,IAAI,IAAI,IAAI,CAAC,IAAI,IAAI,EAAE;gBAAE,MAAM,CAAC,IAAI,CAAC,IAAI,CAAC,EAAE,GAAG,IAAI,EAAE,KAAK,EAAE,IAAI,CAAC,KAAK,GAAG,MAAM,EAAE,CAAC,CAAA;QAC/F,CAAC;QACD,MAAM,CAAC,KAAK,GAAG,OAAO,CAAC,QAAQ,CAAA;QAC/B,OAAO,MAAM,CAAA;IACf,CAAC;IAEO,OAAO,CAAC,OAAoB,EAAE,KAAe;QACnD,OAAO,CAAC,IAAI,CAAC,IAAI,CAAC,KAAK,IAAI,KAAK,CAAC,CAAC,oBAAoB,OAAO,CAAC,IAAI,gBAAgB,EAAE;YAClF,MAAM,EAAE,MAAM;YACd,OAAO,EAAE,EAAE,cAAc,EAAE,kBAAkB,EAAE,eAAe,EAAE,UAAU,OAAO,CAAC,OAAO,EAAE,EAAE;YAC7F,IAAI,EAAE,IAAI,CAAC,SAAS,CAAC,EAAE,KAAK,EAAE,KAAK,EAAE,OAAO,CAAC,QAAQ,EAAE,eAAe,EAAE,OAAO,EAAE,CAAC;SACnF,CAAC,CAAA;IACJ,CAAC;CACF;AAED,+EAA+E;AAC/E,MAAM,UAAU,iBAAiB,CAC/B,KAAe,EACf,UAAkB;IAElB,MAAM,KAAK,GAAG,IAAI,CAAC,KAAK,CAAC,UAAU,GAAG,GAAG,CAAC,CAAA;IAC1C,MAAM,OAAO,GAA+C,EAAE,CAAA;IAC9D,IAAI,KAAK,GAAa,EAAE,CAAA;IACxB,IAAI,MAAM,GAAG,CAAC,CAAA;IACd,IAAI,MAAM,GAAG,CAAC,CAAA;IACd,MAAM,KAAK,GAAG,GAAG,EAAE;QACjB,IAAI,KAAK,CAAC,MAAM,KAAK,CAAC;YAAE,OAAM;QAC9B,OAAO,CAAC,IAAI,CAAC,EAAE,KAAK,EAAE,MAAM,EAAE,CAAC,CAAA;QAC/B,MAAM,IAAI,KAAK,CAAC,MAAM,CAAA;QACtB,KAAK,GAAG,EAAE,CAAA;QACV,MAAM,GAAG,CAAC,CAAA;IACZ,CAAC,CAAA;IACD,KAAK,MAAM,IAAI,IAAI,KAAK,EAAE,CAAC;QACzB,MAAM,SAAS,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,IAAI,CAAC,IAAI,CAAC,IAAI,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC,CAAA;QACzD,IAAI,SAAS,GAAG,KAAK,EAAE,CAAC;YACtB,KAAK,EAAE,CAAA;YACP,OAAO,CAAC,IAAI,CAAC,EAAE,KAAK,EAAE,CAAC,IAAI,CAAC,EAAE,MAAM,EAAE,CAAC,CAAA;YACvC,MAAM,EAAE,CAAA;YACR,SAAQ;QACV,CAAC;QACD,IAAI,KAAK,CAAC,MAAM,IAAI,MAAM,GAAG,SAAS,GAAG,KAAK;YAAE,KAAK,EAAE,CAAA;QACvD,KAAK,CAAC,IAAI,CAAC,IAAI,CAAC,CAAA;QAChB,MAAM,IAAI,SAAS,CAAA;IACrB,CAAC;IACD,KAAK,EAAE,CAAA;IACP,OAAO,OAAO,CAAA;AAChB,CAAC"}

package/dist/models/gguf/classify.d.ts

@@ -0,0 +1,29 @@
+/**
+ * Metadata-driven classification of a GGUF: embedding vs text, projector modality, embedded MTP
+ * head, effective context. Verbatim port of the relevant helpers in
+ * `extensions/llamacpp-upstream-extension/src/util.ts`.
+ */
+type Meta = Record<string, unknown> | undefined | null;
+/** True iff the load-error text is an MTP rejection (llama.cpp has no structured code for it). */
+export declare function matchesMtpLoadFailure(text: string): boolean;
+/** A combined Qwen GGUF whose MTP head is embedded (`{arch}.nextn_predict_layers` > 0). */
+export declare function hasEmbeddedMtp(metadata: Meta): boolean;
+export declare function isMtpCapable(metadata: Meta, mtpDraftPath: string): boolean;
+/** Architectures that cannot generate text; started as a chat model they crash the server. */
+export declare const NON_TEXT_GGUF_ARCHITECTURES: Set<string>;
+/** Weights that produce embeddings rather than text (load in embedding mode instead). */
+export declare function isEmbeddingGguf(metadata: Meta): boolean;
+/**
+ * The context to request: clamp to the trained maximum when known (llama.cpp aborts past it),
+ * otherwise pass the request through.
+ */
+export declare function effectiveCtxSize(requested: number | undefined, maxCtxTrain: number | undefined): number | undefined;
+/**
+ * Which modality an mmproj carries. `general.architecture` is `clip` for every projector; the
+ * modality lives in the `clip.*` keys. Unknown metadata falls back to vision.
+ */
+export declare function classifyProjector(metadata: Meta): {
+    vision: boolean;
+    audio: boolean;
+};
+export {};

package/dist/models/gguf/classify.js

@@ -0,0 +1,89 @@
+/**
+ * Metadata-driven classification of a GGUF: embedding vs text, projector modality, embedded MTP
+ * head, effective context. Verbatim port of the relevant helpers in
+ * `extensions/llamacpp-upstream-extension/src/util.ts`.
+ */
+const EMBEDDED_MTP_ARCHITECTURES = new Set(['qwen35', 'qwen35moe']);
+/** True iff the load-error text is an MTP rejection (llama.cpp has no structured code for it). */
+export function matchesMtpLoadFailure(text) {
+    if (!text)
+        return false;
+    return (/failed to create MTP context/i.test(text) ||
+        /context type MTP requested/i.test(text) ||
+        /doesn'?t contain MTP layers/i.test(text));
+}
+/** A combined Qwen GGUF whose MTP head is embedded (`{arch}.nextn_predict_layers` > 0). */
+export function hasEmbeddedMtp(metadata) {
+    if (!metadata)
+        return false;
+    const architecture = metadata['general.architecture'];
+    if (typeof architecture !== 'string' || !EMBEDDED_MTP_ARCHITECTURES.has(architecture))
+        return false;
+    const blockCount = Number(metadata[`${architecture}.block_count`]);
+    const nextn = Number(metadata[`${architecture}.nextn_predict_layers`]);
+    return Number.isInteger(blockCount) && Number.isInteger(nextn) && nextn > 0 && blockCount > nextn;
+}
+export function isMtpCapable(metadata, mtpDraftPath) {
+    return mtpDraftPath.length > 0 || hasEmbeddedMtp(metadata);
+}
+/** Architectures that cannot generate text; started as a chat model they crash the server. */
+export const NON_TEXT_GGUF_ARCHITECTURES = new Set([
+    'bert',
+    'modern-bert',
+    'nomic-bert',
+    'nomic-bert-moe',
+    'neo-bert',
+    'jina-bert-v2',
+    'jina-bert-v3',
+    'eurobert',
+    'gemma-embedding',
+    'llama-embed',
+    't5encoder',
+]);
+/** Weights that produce embeddings rather than text (load in embedding mode instead). */
+export function isEmbeddingGguf(metadata) {
+    const raw = metadata?.['general.architecture'];
+    const arch = typeof raw === 'string' ? raw.trim().toLowerCase() : '';
+    if (!arch)
+        return false;
+    if (NON_TEXT_GGUF_ARCHITECTURES.has(arch))
+        return true;
+    // Embedding/reranker conversions of a generative arch keep the arch name; a pooling type other
+    // than 0 (NONE) or a classifier head gives them away.
+    const pooling = metadata?.[`${arch}.pooling_type`];
+    const poolingStr = pooling == null ? '' : String(pooling).trim();
+    if (poolingStr !== '' && poolingStr !== '0')
+        return true;
+    return metadata?.[`${arch}.classifier.output_labels`] !== undefined;
+}
+/**
+ * The context to request: clamp to the trained maximum when known (llama.cpp aborts past it),
+ * otherwise pass the request through.
+ */
+export function effectiveCtxSize(requested, maxCtxTrain) {
+    if (typeof requested !== 'number' || !Number.isFinite(requested))
+        return requested;
+    if (typeof maxCtxTrain !== 'number' || !Number.isFinite(maxCtxTrain))
+        return requested;
+    if (maxCtxTrain <= 0)
+        return requested;
+    return Math.min(requested, maxCtxTrain);
+}
+/**
+ * Which modality an mmproj carries. `general.architecture` is `clip` for every projector; the
+ * modality lives in the `clip.*` keys. Unknown metadata falls back to vision.
+ */
+export function classifyProjector(metadata) {
+    if (!metadata)
+        return { vision: true, audio: false };
+    const truthy = (v) => String(v ?? '')
+        .trim()
+        .toLowerCase() === 'true';
+    const present = (v) => v !== undefined && v !== null && String(v).trim() !== '';
+    const vision = truthy(metadata['clip.has_vision_encoder']) || present(metadata['clip.vision.projector_type']);
+    const audio = truthy(metadata['clip.has_audio_encoder']) || present(metadata['clip.audio.projector_type']);
+    if (!vision && !audio)
+        return { vision: true, audio: false };
+    return { vision, audio };
+}
+//# sourceMappingURL=classify.js.map

package/dist/models/gguf/classify.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"classify.js","sourceRoot":"","sources":["../../../src/models/gguf/classify.ts"],"names":[],"mappings":"AAAA;;;;GAIG;AAIH,MAAM,0BAA0B,GAAG,IAAI,GAAG,CAAC,CAAC,QAAQ,EAAE,WAAW,CAAC,CAAC,CAAA;AAEnE,kGAAkG;AAClG,MAAM,UAAU,qBAAqB,CAAC,IAAY;IAChD,IAAI,CAAC,IAAI;QAAE,OAAO,KAAK,CAAA;IACvB,OAAO,CACL,+BAA+B,CAAC,IAAI,CAAC,IAAI,CAAC;QAC1C,6BAA6B,CAAC,IAAI,CAAC,IAAI,CAAC;QACxC,8BAA8B,CAAC,IAAI,CAAC,IAAI,CAAC,CAC1C,CAAA;AACH,CAAC;AAED,2FAA2F;AAC3F,MAAM,UAAU,cAAc,CAAC,QAAc;IAC3C,IAAI,CAAC,QAAQ;QAAE,OAAO,KAAK,CAAA;IAC3B,MAAM,YAAY,GAAG,QAAQ,CAAC,sBAAsB,CAAC,CAAA;IACrD,IAAI,OAAO,YAAY,KAAK,QAAQ,IAAI,CAAC,0BAA0B,CAAC,GAAG,CAAC,YAAY,CAAC;QAAE,OAAO,KAAK,CAAA;IACnG,MAAM,UAAU,GAAG,MAAM,CAAC,QAAQ,CAAC,GAAG,YAAY,cAAc,CAAC,CAAC,CAAA;IAClE,MAAM,KAAK,GAAG,MAAM,CAAC,QAAQ,CAAC,GAAG,YAAY,uBAAuB,CAAC,CAAC,CAAA;IACtE,OAAO,MAAM,CAAC,SAAS,CAAC,UAAU,CAAC,IAAI,MAAM,CAAC,SAAS,CAAC,KAAK,CAAC,IAAI,KAAK,GAAG,CAAC,IAAI,UAAU,GAAG,KAAK,CAAA;AACnG,CAAC;AAED,MAAM,UAAU,YAAY,CAAC,QAAc,EAAE,YAAoB;IAC/D,OAAO,YAAY,CAAC,MAAM,GAAG,CAAC,IAAI,cAAc,CAAC,QAAQ,CAAC,CAAA;AAC5D,CAAC;AAED,8FAA8F;AAC9F,MAAM,CAAC,MAAM,2BAA2B,GAAG,IAAI,GAAG,CAAC;IACjD,MAAM;IACN,aAAa;IACb,YAAY;IACZ,gBAAgB;IAChB,UAAU;IACV,cAAc;IACd,cAAc;IACd,UAAU;IACV,iBAAiB;IACjB,aAAa;IACb,WAAW;CACZ,CAAC,CAAA;AAEF,yFAAyF;AACzF,MAAM,UAAU,eAAe,CAAC,QAAc;IAC5C,MAAM,GAAG,GAAG,QAAQ,EAAE,CAAC,sBAAsB,CAAC,CAAA;IAC9C,MAAM,IAAI,GAAG,OAAO,GAAG,KAAK,QAAQ,CAAC,CAAC,CAAC,GAAG,CAAC,IAAI,EAAE,CAAC,WAAW,EAAE,CAAC,CAAC,CAAC,EAAE,CAAA;IACpE,IAAI,CAAC,IAAI;QAAE,OAAO,KAAK,CAAA;IACvB,IAAI,2BAA2B,CAAC,GAAG,CAAC,IAAI,CAAC;QAAE,OAAO,IAAI,CAAA;IACtD,+FAA+F;IAC/F,sDAAsD;IACtD,MAAM,OAAO,GAAG,QAAQ,EAAE,CAAC,GAAG,IAAI,eAAe,CAAC,CAAA;IAClD,MAAM,UAAU,GAAG,OAAO,IAAI,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,MAAM,CAAC,OAAO,CAAC,CAAC,IAAI,EAAE,CAAA;IAChE,IAAI,UAAU,KAAK,EAAE,IAAI,UAAU,KAAK,GAAG;QAAE,OAAO,IAAI,CAAA;IACxD,OAAO,QAAQ,EAAE,CAAC,GAAG,IAAI,2BAA2B,CAAC,KAAK,SAAS,CAAA;AACrE,CAAC;AAED;;;GAGG;AACH,MAAM,UAAU,gBAAgB,CAC9B,SAA6B,EAC7B,WAA+B;IAE/B,IAAI,OAAO,SAAS,KAAK,QAAQ,IAAI,CAAC,MAAM,CAAC,QAAQ,CAAC,SAAS,CAAC;QAAE,OAAO,SAAS,CAAA;IAClF,IAAI,OAAO,WAAW,KAAK,QAAQ,IAAI,CAAC,MAAM,CAAC,QAAQ,CAAC,WAAW,CAAC;QAAE,OAAO,SAAS,CAAA;IACtF,IAAI,WAAW,IAAI,CAAC;QAAE,OAAO,SAAS,CAAA;IACtC,OAAO,IAAI,CAAC,GAAG,CAAC,SAAS,EAAE,WAAW,CAAC,CAAA;AACzC,CAAC;AAED;;;GAGG;AACH,MAAM,UAAU,iBAAiB,CAAC,QAAc;IAC9C,IAAI,CAAC,QAAQ;QAAE,OAAO,EAAE,MAAM,EAAE,IAAI,EAAE,KAAK,EAAE,KAAK,EAAE,CAAA;IACpD,MAAM,MAAM,GAAG,CAAC,CAAU,EAAE,EAAE,CAC5B,MAAM,CAAC,CAAC,IAAI,EAAE,CAAC;SACZ,IAAI,EAAE;SACN,WAAW,EAAE,KAAK,MAAM,CAAA;IAC7B,MAAM,OAAO,GAAG,CAAC,CAAU,EAAE,EAAE,CAAC,CAAC,KAAK,SAAS,IAAI,CAAC,KAAK,IAAI,IAAI,MAAM,CAAC,CAAC,CAAC,CAAC,IAAI,EAAE,KAAK,EAAE,CAAA;IACxF,MAAM,MAAM,GACV,MAAM,CAAC,QAAQ,CAAC,yBAAyB,CAAC,CAAC,IAAI,OAAO,CAAC,QAAQ,CAAC,4BAA4B,CAAC,CAAC,CAAA;IAChG,MAAM,KAAK,GAAG,MAAM,CAAC,QAAQ,CAAC,wBAAwB,CAAC,CAAC,IAAI,OAAO,CAAC,QAAQ,CAAC,2BAA2B,CAAC,CAAC,CAAA;IAC1G,IAAI,CAAC,MAAM,IAAI,CAAC,KAAK;QAAE,OAAO,EAAE,MAAM,EAAE,IAAI,EAAE,KAAK,EAAE,KAAK,EAAE,CAAA;IAC5D,OAAO,EAAE,MAAM,EAAE,KAAK,EAAE,CAAA;AAC1B,CAAC"}

package/dist/models/gguf/index.d.ts

@@ -0,0 +1,5 @@
+export * from './reader.js';
+export * from './kv-cache.js';
+export * from './support.js';
+export * from './classify.js';
+export * from './read-file.js';

package/dist/models/gguf/index.js

@@ -0,0 +1,6 @@
+export * from './reader.js';
+export * from './kv-cache.js';
+export * from './support.js';
+export * from './classify.js';
+export * from './read-file.js';
+//# sourceMappingURL=index.js.map

package/dist/models/gguf/index.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"index.js","sourceRoot":"","sources":["../../../src/models/gguf/index.ts"],"names":[],"mappings":"AAAA,cAAc,aAAa,CAAA;AAC3B,cAAc,eAAe,CAAA;AAC7B,cAAc,cAAc,CAAA;AAC5B,cAAc,eAAe,CAAA;AAC7B,cAAc,gBAAgB,CAAA"}

package/dist/models/gguf/kv-cache.d.ts

@@ -0,0 +1,27 @@
+/**
+ * KV-cache size estimate from GGUF metadata. Port of `gguf/utils.rs`
+ * (`kv_cache_bits_per_element`, `estimate_kv_cache_internal`). Error strings are the Rust
+ * `Display` texts, which the app receives verbatim.
+ */
+export interface KvCacheEstimate {
+    size: number;
+    per_token_size: number;
+}
+export declare class KvCacheError extends Error {
+    constructor(message: string);
+}
+export declare const KV_CACHE_ERRORS: {
+    readonly architectureNotFound: "Invalid metadata: architecture not found";
+    readonly blockCountInvalid: "Invalid metadata: block_count not found or invalid";
+    readonly headCountInvalid: "Invalid metadata: head_count not found or invalid";
+    readonly embeddingLengthInvalid: "Invalid metadata: embedding_length not found or invalid";
+    readonly contextLengthInvalid: "Invalid metadata: context_length not found or invalid";
+};
+/**
+ * Bits one KV element occupies under a llama.cpp `cache_type_*`. Block quants carry a scale per
+ * 32 elements (hence the half bits); unknown or absent reads as fp16.
+ */
+export declare function kvCacheBitsPerElement(cacheType: string | undefined | null): number;
+export declare function estimateKvCache(meta: Record<string, string>, ctxSize: number | undefined, cacheTypeK?: string | null, cacheTypeV?: string | null): KvCacheEstimate;
+/** `{arch}.context_length` when present and positive — the model's trained context. */
+export declare function ggufContextLength(meta: Record<string, string>): number | undefined;

package/dist/models/gguf/kv-cache.js

@@ -0,0 +1,101 @@
+/**
+ * KV-cache size estimate from GGUF metadata. Port of `gguf/utils.rs`
+ * (`kv_cache_bits_per_element`, `estimate_kv_cache_internal`). Error strings are the Rust
+ * `Display` texts, which the app receives verbatim.
+ */
+import { parseRustU64 } from '../../util/index.js';
+export class KvCacheError extends Error {
+    constructor(message) {
+        super(message);
+        this.name = 'KvCacheError';
+    }
+}
+export const KV_CACHE_ERRORS = {
+    architectureNotFound: 'Invalid metadata: architecture not found',
+    blockCountInvalid: 'Invalid metadata: block_count not found or invalid',
+    headCountInvalid: 'Invalid metadata: head_count not found or invalid',
+    embeddingLengthInvalid: 'Invalid metadata: embedding_length not found or invalid',
+    contextLengthInvalid: 'Invalid metadata: context_length not found or invalid',
+};
+/**
+ * Bits one KV element occupies under a llama.cpp `cache_type_*`. Block quants carry a scale per
+ * 32 elements (hence the half bits); unknown or absent reads as fp16.
+ */
+export function kvCacheBitsPerElement(cacheType) {
+    switch (cacheType?.trim().toLowerCase()) {
+        case 'f32':
+            return 32;
+        case 'f16':
+        case 'bf16':
+            return 16;
+        case 'q8_0':
+            return 8.5;
+        case 'q5_1':
+            return 6;
+        case 'q5_0':
+            return 5.5;
+        case 'q4_1':
+            return 5;
+        case 'q4_0':
+        case 'iq4_nl':
+            return 4.5;
+        case 'turbo4':
+            return 4;
+        case 'turbo3':
+            return 3;
+        case 'turbo2':
+            return 2;
+        default:
+            return 16;
+    }
+}
+const u64 = (meta, key) => {
+    const raw = meta[key];
+    return raw === undefined ? undefined : parseRustU64(raw);
+};
+export function estimateKvCache(meta, ctxSize, cacheTypeK, cacheTypeV) {
+    const arch = meta['general.architecture'];
+    if (arch === undefined)
+        throw new KvCacheError(KV_CACHE_ERRORS.architectureNotFound);
+    const nLayer = u64(meta, `${arch}.block_count`);
+    if (nLayer === undefined || nLayer <= 0)
+        throw new KvCacheError(KV_CACHE_ERRORS.blockCountInvalid);
+    const headCount = u64(meta, `${arch}.attention.head_count`);
+    const headCountKv = u64(meta, `${arch}.attention.head_count_kv`);
+    const nHead = headCountKv !== undefined && headCountKv > 0 ? headCountKv : (headCount ?? 0);
+    if (nHead === 0)
+        throw new KvCacheError(KV_CACHE_ERRORS.headCountInvalid);
+    let keyLen = u64(meta, `${arch}.attention.key_length`) ?? 0;
+    let valLen = u64(meta, `${arch}.attention.value_length`) ?? 0;
+    if (keyLen === 0 || valLen === 0) {
+        const embLen = u64(meta, `${arch}.embedding_length`) ?? 0;
+        if (embLen > 0 && nHead > 0) {
+            const totalHeads = headCount ?? nHead;
+            const headDim = Math.floor(embLen / totalHeads);
+            keyLen = headDim;
+            valLen = headDim;
+        }
+    }
+    if (keyLen === 0 || valLen === 0)
+        throw new KvCacheError(KV_CACHE_ERRORS.embeddingLengthInvalid);
+    const maxCtx = u64(meta, `${arch}.context_length`);
+    if (maxCtx === undefined || maxCtx <= 0)
+        throw new KvCacheError(KV_CACHE_ERRORS.contextLengthInvalid);
+    const ctxLen = ctxSize === undefined ? maxCtx : Math.min(ctxSize, maxCtx);
+    const slidingWindowRaw = u64(meta, `${arch}.attention.sliding_window`);
+    const slidingWindow = slidingWindowRaw !== undefined && slidingWindowRaw > 0 ? slidingWindowRaw : undefined;
+    const bitsPerToken = nLayer * nHead * (keyLen * kvCacheBitsPerElement(cacheTypeK) + valLen * kvCacheBitsPerElement(cacheTypeV));
+    const kvPerToken = Math.ceil(bitsPerToken / 8);
+    const fullCost = ctxLen * kvPerToken;
+    const size = slidingWindow === undefined ? fullCost : Math.floor((fullCost + slidingWindow * kvPerToken) / 2);
+    return { size, per_token_size: kvPerToken };
+}
+/** `{arch}.context_length` when present and positive — the model's trained context. */
+export function ggufContextLength(meta) {
+    const arch = meta['general.architecture'];
+    if (arch === undefined)
+        return undefined;
+    const v = u64(meta, `${arch}.context_length`);
+    return v !== undefined && v > 0 ? v : undefined;
+}
+//# sourceMappingURL=kv-cache.js.map

package/dist/models/gguf/kv-cache.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"kv-cache.js","sourceRoot":"","sources":["../../../src/models/gguf/kv-cache.ts"],"names":[],"mappings":"AAAA;;;;GAIG;AAEH,OAAO,EAAE,YAAY,EAAE,MAAM,qBAAqB,CAAA;AAOlD,MAAM,OAAO,YAAa,SAAQ,KAAK;IACrC,YAAY,OAAe;QACzB,KAAK,CAAC,OAAO,CAAC,CAAA;QACd,IAAI,CAAC,IAAI,GAAG,cAAc,CAAA;IAC5B,CAAC;CACF;AAED,MAAM,CAAC,MAAM,eAAe,GAAG;IAC7B,oBAAoB,EAAE,0CAA0C;IAChE,iBAAiB,EAAE,oDAAoD;IACvE,gBAAgB,EAAE,mDAAmD;IACrE,sBAAsB,EAAE,yDAAyD;IACjF,oBAAoB,EAAE,uDAAuD;CACrE,CAAA;AAEV;;;GAGG;AACH,MAAM,UAAU,qBAAqB,CAAC,SAAoC;IACxE,QAAQ,SAAS,EAAE,IAAI,EAAE,CAAC,WAAW,EAAE,EAAE,CAAC;QACxC,KAAK,KAAK;YACR,OAAO,EAAE,CAAA;QACX,KAAK,KAAK,CAAC;QACX,KAAK,MAAM;YACT,OAAO,EAAE,CAAA;QACX,KAAK,MAAM;YACT,OAAO,GAAG,CAAA;QACZ,KAAK,MAAM;YACT,OAAO,CAAC,CAAA;QACV,KAAK,MAAM;YACT,OAAO,GAAG,CAAA;QACZ,KAAK,MAAM;YACT,OAAO,CAAC,CAAA;QACV,KAAK,MAAM,CAAC;QACZ,KAAK,QAAQ;YACX,OAAO,GAAG,CAAA;QACZ,KAAK,QAAQ;YACX,OAAO,CAAC,CAAA;QACV,KAAK,QAAQ;YACX,OAAO,CAAC,CAAA;QACV,KAAK,QAAQ;YACX,OAAO,CAAC,CAAA;QACV;YACE,OAAO,EAAE,CAAA;IACb,CAAC;AACH,CAAC;AAED,MAAM,GAAG,GAAG,CAAC,IAA4B,EAAE,GAAW,EAAsB,EAAE;IAC5E,MAAM,GAAG,GAAG,IAAI,CAAC,GAAG,CAAC,CAAA;IACrB,OAAO,GAAG,KAAK,SAAS,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC,CAAC,YAAY,CAAC,GAAG,CAAC,CAAA;AAC1D,CAAC,CAAA;AAED,MAAM,UAAU,eAAe,CAC7B,IAA4B,EAC5B,OAA2B,EAC3B,UAA0B,EAC1B,UAA0B;IAE1B,MAAM,IAAI,GAAG,IAAI,CAAC,sBAAsB,CAAC,CAAA;IACzC,IAAI,IAAI,KAAK,SAAS;QAAE,MAAM,IAAI,YAAY,CAAC,eAAe,CAAC,oBAAoB,CAAC,CAAA;IAEpF,MAAM,MAAM,GAAG,GAAG,CAAC,IAAI,EAAE,GAAG,IAAI,cAAc,CAAC,CAAA;IAC/C,IAAI,MAAM,KAAK,SAAS,IAAI,MAAM,IAAI,CAAC;QAAE,MAAM,IAAI,YAAY,CAAC,eAAe,CAAC,iBAAiB,CAAC,CAAA;IAElG,MAAM,SAAS,GAAG,GAAG,CAAC,IAAI,EAAE,GAAG,IAAI,uBAAuB,CAAC,CAAA;IAC3D,MAAM,WAAW,GAAG,GAAG,CAAC,IAAI,EAAE,GAAG,IAAI,0BAA0B,CAAC,CAAA;IAChE,MAAM,KAAK,GAAG,WAAW,KAAK,SAAS,IAAI,WAAW,GAAG,CAAC,CAAC,CAAC,CAAC,WAAW,CAAC,CAAC,CAAC,CAAC,SAAS,IAAI,CAAC,CAAC,CAAA;IAC3F,IAAI,KAAK,KAAK,CAAC;QAAE,MAAM,IAAI,YAAY,CAAC,eAAe,CAAC,gBAAgB,CAAC,CAAA;IAEzE,IAAI,MAAM,GAAG,GAAG,CAAC,IAAI,EAAE,GAAG,IAAI,uBAAuB,CAAC,IAAI,CAAC,CAAA;IAC3D,IAAI,MAAM,GAAG,GAAG,CAAC,IAAI,EAAE,GAAG,IAAI,yBAAyB,CAAC,IAAI,CAAC,CAAA;IAC7D,IAAI,MAAM,KAAK,CAAC,IAAI,MAAM,KAAK,CAAC,EAAE,CAAC;QACjC,MAAM,MAAM,GAAG,GAAG,CAAC,IAAI,EAAE,GAAG,IAAI,mBAAmB,CAAC,IAAI,CAAC,CAAA;QACzD,IAAI,MAAM,GAAG,CAAC,IAAI,KAAK,GAAG,CAAC,EAAE,CAAC;YAC5B,MAAM,UAAU,GAAG,SAAS,IAAI,KAAK,CAAA;YACrC,MAAM,OAAO,GAAG,IAAI,CAAC,KAAK,CAAC,MAAM,GAAG,UAAU,CAAC,CAAA;YAC/C,MAAM,GAAG,OAAO,CAAA;YAChB,MAAM,GAAG,OAAO,CAAA;QAClB,CAAC;IACH,CAAC;IACD,IAAI,MAAM,KAAK,CAAC,IAAI,MAAM,KAAK,CAAC;QAAE,MAAM,IAAI,YAAY,CAAC,eAAe,CAAC,sBAAsB,CAAC,CAAA;IAEhG,MAAM,MAAM,GAAG,GAAG,CAAC,IAAI,EAAE,GAAG,IAAI,iBAAiB,CAAC,CAAA;IAClD,IAAI,MAAM,KAAK,SAAS,IAAI,MAAM,IAAI,CAAC;QAAE,MAAM,IAAI,YAAY,CAAC,eAAe,CAAC,oBAAoB,CAAC,CAAA;IACrG,MAAM,MAAM,GAAG,OAAO,KAAK,SAAS,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,IAAI,CAAC,GAAG,CAAC,OAAO,EAAE,MAAM,CAAC,CAAA;IAEzE,MAAM,gBAAgB,GAAG,GAAG,CAAC,IAAI,EAAE,GAAG,IAAI,2BAA2B,CAAC,CAAA;IACtE,MAAM,aAAa,GAAG,gBAAgB,KAAK,SAAS,IAAI,gBAAgB,GAAG,CAAC,CAAC,CAAC,CAAC,gBAAgB,CAAC,CAAC,CAAC,SAAS,CAAA;IAE3G,MAAM,YAAY,GAChB,MAAM,GAAG,KAAK,GAAG,CAAC,MAAM,GAAG,qBAAqB,CAAC,UAAU,CAAC,GAAG,MAAM,GAAG,qBAAqB,CAAC,UAAU,CAAC,CAAC,CAAA;IAC5G,MAAM,UAAU,GAAG,IAAI,CAAC,IAAI,CAAC,YAAY,GAAG,CAAC,CAAC,CAAA;IAC9C,MAAM,QAAQ,GAAG,MAAM,GAAG,UAAU,CAAA;IACpC,MAAM,IAAI,GACR,aAAa,KAAK,SAAS,CAAC,CAAC,CAAC,QAAQ,CAAC,CAAC,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC,QAAQ,GAAG,aAAa,GAAG,UAAU,CAAC,GAAG,CAAC,CAAC,CAAA;IAClG,OAAO,EAAE,IAAI,EAAE,cAAc,EAAE,UAAU,EAAE,CAAA;AAC7C,CAAC;AAED,uFAAuF;AACvF,MAAM,UAAU,iBAAiB,CAAC,IAA4B;IAC5D,MAAM,IAAI,GAAG,IAAI,CAAC,sBAAsB,CAAC,CAAA;IACzC,IAAI,IAAI,KAAK,SAAS;QAAE,OAAO,SAAS,CAAA;IACxC,MAAM,CAAC,GAAG,GAAG,CAAC,IAAI,EAAE,GAAG,IAAI,iBAAiB,CAAC,CAAA;IAC7C,OAAO,CAAC,KAAK,SAAS,IAAI,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,SAAS,CAAA;AACjD,CAAC"}

package/dist/models/gguf/read-file.d.ts

@@ -0,0 +1,11 @@
+/**
+ * Read GGUF metadata from a file on disk. The parser in `reader.ts` stays pure (it takes bytes), so
+ * this is the only place that opens a model file: it hands the chunked reader growing prefixes of
+ * the file, which is how the app avoids mapping a 40 GB model to learn its context length.
+ */
+import type { GgufMetadata } from './reader.js';
+export interface ReadGgufFileOptions {
+    chunkSize?: number;
+    maxBytes?: number;
+}
+export declare function readGgufMetadataFromFile(path: string, options?: ReadGgufFileOptions): Promise<GgufMetadata>;

package/dist/models/gguf/read-file.js

@@ -0,0 +1,21 @@
+/**
+ * Read GGUF metadata from a file on disk. The parser in `reader.ts` stays pure (it takes bytes), so
+ * this is the only place that opens a model file: it hands the chunked reader growing prefixes of
+ * the file, which is how the app avoids mapping a 40 GB model to learn its context length.
+ */
+import { open } from 'node:fs/promises';
+import { readGgufMetadataChunked } from './reader.js';
+export async function readGgufMetadataFromFile(path, options = {}) {
+    const handle = await open(path, 'r');
+    try {
+        return await readGgufMetadataChunked(async (byteLength) => {
+            const buffer = Buffer.allocUnsafe(byteLength);
+            const { bytesRead } = await handle.read(buffer, 0, byteLength, 0);
+            return buffer.subarray(0, bytesRead);
+        }, options);
+    }
+    finally {
+        await handle.close().catch(() => { });
+    }
+}
+//# sourceMappingURL=read-file.js.map

package/dist/models/gguf/read-file.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"read-file.js","sourceRoot":"","sources":["../../../src/models/gguf/read-file.ts"],"names":[],"mappings":"AAAA;;;;GAIG;AAEH,OAAO,EAAE,IAAI,EAAE,MAAM,kBAAkB,CAAA;AACvC,OAAO,EAAE,uBAAuB,EAAE,MAAM,aAAa,CAAA;AAQrD,MAAM,CAAC,KAAK,UAAU,wBAAwB,CAC5C,IAAY,EACZ,UAA+B,EAAE;IAEjC,MAAM,MAAM,GAAG,MAAM,IAAI,CAAC,IAAI,EAAE,GAAG,CAAC,CAAA;IACpC,IAAI,CAAC;QACH,OAAO,MAAM,uBAAuB,CAAC,KAAK,EAAE,UAAkB,EAAE,EAAE;YAChE,MAAM,MAAM,GAAG,MAAM,CAAC,WAAW,CAAC,UAAU,CAAC,CAAA;YAC7C,MAAM,EAAE,SAAS,EAAE,GAAG,MAAM,MAAM,CAAC,IAAI,CAAC,MAAM,EAAE,CAAC,EAAE,UAAU,EAAE,CAAC,CAAC,CAAA;YACjE,OAAO,MAAM,CAAC,QAAQ,CAAC,CAAC,EAAE,SAAS,CAAC,CAAA;QACtC,CAAC,EAAE,OAAO,CAAC,CAAA;IACb,CAAC;YAAS,CAAC;QACT,MAAM,MAAM,CAAC,KAAK,EAAE,CAAC,KAAK,CAAC,GAAG,EAAE,GAAE,CAAC,CAAC,CAAA;IACtC,CAAC;AACH,CAAC"}

package/dist/models/gguf/reader.d.ts

@@ -0,0 +1,40 @@
+/**
+ * GGUF header reader. Port of `tauri-plugin-llamacpp-upstream/src/gguf/{helpers,types}.rs`.
+ *
+ * Reads magic, version, tensor count and the metadata KV block only — never the tensor table.
+ * Every value is stringified the way the Rust reader does (`to_string()`, `[a, b]` for arrays of
+ * up to 24 elements, a placeholder for longer ones), so downstream code sees identical strings.
+ *
+ * Pure over a byte buffer; callers feed it a file prefix and grow it while it asks for more.
+ */
+export interface GgufMetadata {
+    version: number;
+    tensor_count: number;
+    metadata: Record<string, string>;
+}
+export declare const GGUF_VALUE_TYPE_NAMES: readonly ["Uint8", "Int8", "Uint16", "Int16", "Uint32", "Int32", "Float32", "Bool", "String", "Array", "Uint64", "Int64", "Float64"];
+export declare const MAX_GGUF_STRING_BYTES: number;
+export declare const MAX_GGUF_ARRAY_LEN = 1000000;
+export declare const INLINE_ARRAY_MAX_LEN = 24;
+/** The buffer ended before the header did; feed more bytes and retry. */
+export declare class GgufNeedMoreData extends Error {
+    constructor();
+}
+/** The bytes are not a GGUF header the reader understands (same messages as the Rust reader). */
+export declare class GgufParseError extends Error {
+    constructor(message: string);
+}
+/**
+ * Parse the header from `bytes`. Throws `GgufNeedMoreData` when the buffer is too short and
+ * `GgufParseError` when the bytes are not a valid header.
+ */
+export declare function readGgufMetadata(bytes: Uint8Array): GgufMetadata;
+/**
+ * Drive `readGgufMetadata` over a source that yields the file in growing prefixes (a local file
+ * read in chunks, or HTTP range requests). `readPrefix(byteLength)` returns at least what it can;
+ * a prefix shorter than requested means EOF.
+ */
+export declare function readGgufMetadataChunked(readPrefix: (byteLength: number) => Promise<Uint8Array>, options?: {
+    chunkSize?: number;
+    maxBytes?: number;
+}): Promise<GgufMetadata>;