@realtimex/node-llama-cpp 0.1.0

package/dist/evaluator/TokenMeter.js

@@ -0,0 +1,74 @@
+/**
+ * Tracks the usage of tokens.
+ */
+export class TokenMeter {
+    _inputTokens = 0;
+    _outputTokens = 0;
+    /**
+     * The number of input tokens used
+     */
+    get usedInputTokens() {
+        return this._inputTokens;
+    }
+    /**
+     * The number of tokens generated by a model
+     */
+    get usedOutputTokens() {
+        return this._outputTokens;
+    }
+    /**
+     * Get the current state of the token meter
+     */
+    getState() {
+        return {
+            usedInputTokens: this.usedInputTokens,
+            usedOutputTokens: this.usedOutputTokens
+        };
+    }
+    /**
+     * Log the usage of tokens
+     */
+    useTokens(tokens, type) {
+        if (tokens < 0)
+            throw new RangeError("Tokens cannot be negative");
+        else if (tokens === 0)
+            return;
+        if (type === "input")
+            this._inputTokens += tokens;
+        else if (type === "output")
+            this._outputTokens += tokens;
+        else {
+            void type;
+            throw new TypeError(`Unknown token type: ${type}`);
+        }
+    }
+    /**
+     * Get the difference between the current meter and another meter
+     */
+    diff(meter) {
+        return TokenMeter.diff(this, meter);
+    }
+    /**
+     * Log the usage of tokens on multiple meters
+     */
+    static useTokens(meters, tokens, type) {
+        if (meters == null)
+            return;
+        if (meters instanceof TokenMeter)
+            meters.useTokens(tokens, type);
+        else {
+            for (const meter of meters)
+                meter.useTokens(tokens, type);
+        }
+    }
+    /**
+     * Get the difference between two meters
+     */
+    static diff(meter1, meter2) {
+        return {
+            usedInputTokens: meter1.usedInputTokens - meter2.usedInputTokens,
+            usedOutputTokens: meter1.usedOutputTokens - meter2.usedOutputTokens
+        };
+    }
+}
+//# sourceMappingURL=TokenMeter.js.map

package/dist/evaluator/TokenMeter.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"TokenMeter.js","sourceRoot":"","sources":["../../src/evaluator/TokenMeter.ts"],"names":[],"mappings":"AAAA;;GAEG;AACH,MAAM,OAAO,UAAU;IACX,YAAY,GAAW,CAAC,CAAC;IACzB,aAAa,GAAW,CAAC,CAAC;IAElC;;OAEG;IACH,IAAW,eAAe;QACtB,OAAO,IAAI,CAAC,YAAY,CAAC;IAC7B,CAAC;IAED;;OAEG;IACH,IAAW,gBAAgB;QACvB,OAAO,IAAI,CAAC,aAAa,CAAC;IAC9B,CAAC;IAED;;OAEG;IACI,QAAQ;QACX,OAAO;YACH,eAAe,EAAE,IAAI,CAAC,eAAe;YACrC,gBAAgB,EAAE,IAAI,CAAC,gBAAgB;SAC1C,CAAC;IACN,CAAC;IAED;;OAEG;IACI,SAAS,CAAC,MAAc,EAAE,IAAwB;QACrD,IAAI,MAAM,GAAG,CAAC;YACV,MAAM,IAAI,UAAU,CAAC,2BAA2B,CAAC,CAAC;aACjD,IAAI,MAAM,KAAK,CAAC;YACjB,OAAO;QAEX,IAAI,IAAI,KAAK,OAAO;YAChB,IAAI,CAAC,YAAY,IAAI,MAAM,CAAC;aAC3B,IAAI,IAAI,KAAK,QAAQ;YACtB,IAAI,CAAC,aAAa,IAAI,MAAM,CAAC;aAC5B,CAAC;YACF,KAAM,IAAqB,CAAC;YAC5B,MAAM,IAAI,SAAS,CAAC,uBAAuB,IAAI,EAAE,CAAC,CAAC;QACvD,CAAC;IACL,CAAC;IAED;;OAEG;IACI,IAAI,CAAC,KAAmC;QAC3C,OAAO,UAAU,CAAC,IAAI,CAAC,IAAI,EAAE,KAAK,CAAC,CAAC;IACxC,CAAC;IAED;;OAEG;IACI,MAAM,CAAC,SAAS,CACnB,MAAuF,EACvF,MAAc,EACd,IAAwB;QAExB,IAAI,MAAM,IAAI,IAAI;YACd,OAAO;QAEX,IAAI,MAAM,YAAY,UAAU;YAC5B,MAAM,CAAC,SAAS,CAAC,MAAM,EAAE,IAAI,CAAC,CAAC;aAC9B,CAAC;YACF,KAAK,MAAM,KAAK,IAAI,MAAM;gBACtB,KAAK,CAAC,SAAS,CAAC,MAAM,EAAE,IAAI,CAAC,CAAC;QACtC,CAAC;IACL,CAAC;IAED;;OAEG;IACI,MAAM,CAAC,IAAI,CACd,MAAoC,EACpC,MAAoC;QAEpC,OAAO;YACH,eAAe,EAAE,MAAM,CAAC,eAAe,GAAG,MAAM,CAAC,eAAe;YAChE,gBAAgB,EAAE,MAAM,CAAC,gBAAgB,GAAG,MAAM,CAAC,gBAAgB;SACtE,CAAC;IACN,CAAC;CACJ"}

package/dist/evaluator/utils/chunkDocument.d.ts

@@ -0,0 +1,86 @@
+import { LlamaContextSequence } from "../LlamaContext/LlamaContext.js";
+import { Token, Tokenizer } from "../../types.js";
+import { LlamaText } from "../../utils/LlamaText.js";
+/**
+ * Chunk the given document using a given context sequence to use the chunks for RAG (Retrieval Augmented Generation) embeddings.
+ *
+ * This chunking method is fast and efficient, and utilizes as much parallelization as your hardware allows.
+ *
+ * Based on https://github.com/ZeroEntropy-AI/llama-chunk
+ * @experimental - this API is experimental and may change or be removed in subsequent releases
+ * @hidden
+ */
+export declare function experimentalChunkDocument(options: {
+    contextSequence: LlamaContextSequence;
+    document: string;
+    /**
+     * The tokens to use as separators for chunking the document.
+     * Passed to the `getSystemPrompt` function to generate the prompt.
+     */
+    separatorTokens?: Token[];
+    getSystemPrompt?(options: {
+        separatorTokens: Token[];
+        tokenizer: Tokenizer;
+        maxChunkSize?: number;
+    }): LlamaText | string;
+    /**
+     * Maximum number of tokens to allow in a chunk.
+     *
+     * As a chunk size approaches this limit, the higher the probability of a separator token being inserted.
+     *
+     * Set to `0` to disable this mechanism.
+     *
+     * Defaults to `500`.
+     */
+    maxChunkSize?: number;
+    /**
+     * The alignment curve for the maximum chunk size mechanism.
+     *
+     * Adjust the value based on the behavior of the model.
+     *
+     * Play around with values between `1` and `4` to see what works best for you.
+     *
+     * Set to `1` to disable this mechanism.
+     *
+     * Defaults to `4`.
+     */
+    maxChunkSizeAlignmentCurve?: number;
+    /**
+     * Append the next few tokens (up to `maxTokens`) to the current chunk if their trimmed content
+     * matches any of the texts in `trimmedTexts`
+     */
+    syntaxAlignment?: {
+        /**
+         * The maximum number of tokens to append to the current chunk if their trimmed content matches any of the texts in `trimmedTexts`.
+         *
+         * Default: `4`
+         */
+        maxTokens?: number;
+        /**
+         * The trimmed texts to match for, to append the token to the current chunk.
+         *
+         * Default: `["", ".", ";"]`
+         */
+        trimmedTexts?: string[];
+    };
+    /**
+     * The number of tokens to skip before starting to use the generated separator tokens to split the document.
+     */
+    skipFirstTokens?: number;
+    /**
+     * The number of recent probabilities to keep in the trail for normalization.
+     *
+     * Adjust the value based on the behavior of the model.
+     *
+     * Defaults to `200`.
+     */
+    normalizationTrailSize?: number;
+    /**
+     * Called when a chunk is generated with the tokens that make up the chunk and the separator token used to split the chunk.
+     */
+    onChunkTokens?(chunkTokens: Token[], usedSeparatorToken: Token): void;
+    /**
+     * Called when a chunk is generated with the text that makes up the chunk and the separator token used to split the chunk.
+     */
+    onChunkText?(chunkText: string, usedSeparatorToken: Token): void;
+}): Promise<string[]>;

package/dist/evaluator/utils/chunkDocument.js

@@ -0,0 +1,212 @@
+import { LlamaText, SpecialTokensText } from "../../utils/LlamaText.js";
+import { resolveChatWrapper } from "../../chatWrappers/utils/resolveChatWrapper.js";
+import { safeEventCallback } from "../../utils/safeEventCallback.js";
+import { maxRecentDetokenizerTokens } from "../../consts.js";
+/**
+ * Chunk the given document using a given context sequence to use the chunks for RAG (Retrieval Augmented Generation) embeddings.
+ *
+ * This chunking method is fast and efficient, and utilizes as much parallelization as your hardware allows.
+ *
+ * Based on https://github.com/ZeroEntropy-AI/llama-chunk
+ * @experimental - this API is experimental and may change or be removed in subsequent releases
+ * @hidden
+ */
+export async function experimentalChunkDocument(options) {
+    const { contextSequence, document, separatorTokens = findAppropriateSeparatorTokens(contextSequence.model), getSystemPrompt = getDefaultPrompt, maxChunkSize = 500, maxChunkSizeAlignmentCurve = 4, syntaxAlignment: { maxTokens: maxSyntaxAlignment = 4, trimmedTexts: syntaxAlignmentTrimmedTexts = ["", ".", ";"] } = {}, skipFirstTokens = 3, normalizationTrailSize = 100 } = options;
+    const onChunkTokens = safeEventCallback(options.onChunkTokens);
+    const onChunkText = safeEventCallback(options.onChunkText);
+    if (separatorTokens.length === 0)
+        throw new Error("Separator tokens must be provided");
+    const chatHistory = [{
+            type: "system",
+            text: LlamaText(getSystemPrompt({
+                separatorTokens,
+                tokenizer: contextSequence.model.tokenizer,
+                maxChunkSize: maxChunkSize <= 0
+                    ? undefined
+                    : maxChunkSize
+            })).toJSON()
+        }, {
+            type: "user",
+            text: document
+        }, {
+            type: "model",
+            response: [""]
+        }];
+    const chatWrapper = resolveChatWrapper(contextSequence.model);
+    const { contextText } = chatWrapper.generateContextState({ chatHistory });
+    const initialContextTokens = contextText.tokenize(contextSequence.model.tokenizer, "trimLeadingSpace");
+    const documentTokens = contextSequence.model.tokenize(document, false, "trimLeadingSpace");
+    const syntaxAlignmentTrimmedTextsSet = new Set(syntaxAlignmentTrimmedTexts);
+    if (initialContextTokens.length + documentTokens.length > contextSequence.context.contextSize)
+        throw new Error("The context size is too small to chunk the given document");
+    const evaluateInput = initialContextTokens.slice();
+    for (let i = 0; i < documentTokens.length - 1; i++) {
+        const token = documentTokens[i];
+        evaluateInput.push([token, {
+                generateNext: {
+                    probabilities: true
+                }
+            }]);
+    }
+    let weight = 1;
+    const recentProbabilitiesTrail = [];
+    let chunkStartIndex = 0;
+    let lastPushedSeparatorIndex = 0;
+    const chunks = [];
+    const res = [];
+    function pushSeparatorIndex(separateIndex, separatorToken) {
+        lastPushedSeparatorIndex = separateIndex;
+        if (separateIndex <= chunkStartIndex)
+            return;
+        let endIndex = separateIndex;
+        for (let i = 0; i < maxSyntaxAlignment && documentTokens[endIndex + i] != null; i++) {
+            const text = contextSequence.model.detokenize([documentTokens[endIndex + i]]);
+            if (!syntaxAlignmentTrimmedTextsSet.has(text.trim()))
+                break;
+            endIndex++;
+        }
+        const chunk = documentTokens.slice(chunkStartIndex, endIndex);
+        const text = contextSequence.model.detokenize(chunk, false, documentTokens.slice(chunkStartIndex - maxRecentDetokenizerTokens, chunkStartIndex));
+        chunks.push(chunk);
+        chunkStartIndex = endIndex;
+        onChunkTokens?.(chunk, separatorToken);
+        onChunkText?.(text, separatorToken);
+        res.push(text);
+    }
+    await contextSequence.controlledEvaluate(evaluateInput, {
+        onTokenResult(inputTokenIndex, result) {
+            const i = inputTokenIndex - initialContextTokens.length;
+            const nextProbabilities = result?.next?.probabilities;
+            const nextDocumentToken = documentTokens[i + 1];
+            if (nextProbabilities == null)
+                throw new Error("received no result for token " + i);
+            const topProbabilityScore = nextProbabilities.entries()
+                .next().value?.[1];
+            const [usedSeparatorToken, separatorProbability] = separatorTokens
+                .filter((token) => token !== nextDocumentToken) // avoid splitting on document tokens
+                .map((token) => [token, nextProbabilities.get(token)])
+                .filter((pair) => pair[1] != null)
+                .reduce(([tokenA, probabilityA], [tokenB, probabilityB]) => {
+                if (probabilityA >= probabilityB)
+                    return [tokenA, probabilityA];
+                return [tokenB, probabilityB];
+            }, [separatorTokens[0], 0]);
+            if (topProbabilityScore == null || separatorProbability == null || separatorProbability === 0)
+                return;
+            // console.log(
+            //     i, contextSequence.model.detokenize([documentTokens[i]!]),
+            //     Array.from(nextProbabilities.entries()).slice(0, 5)
+            //         .map(([token, probability]) => [contextSequence.model.detokenize([token], true), probability])
+            // );
+            if (separatorProbability >= topProbabilityScore)
+                pushSeparatorIndex(i + 1, usedSeparatorToken);
+            else if (i > skipFirstTokens) {
+                const adjustedProbability = separatorProbability + (weight * (1 - separatorProbability));
+                let maxChunkSizeAlignment = 0;
+                if (maxChunkSize !== 0 && adjustedProbability < topProbabilityScore) {
+                    const leftProbability = 1 - adjustedProbability;
+                    const currentChunkSize = Math.max(0, 1 + i - chunkStartIndex);
+                    maxChunkSizeAlignment = currentChunkSize === 0
+                        ? 0
+                        : adjustExponential(leftProbability * Math.min(1, currentChunkSize / maxChunkSize), maxChunkSizeAlignmentCurve <= 0
+                            ? 1
+                            : maxChunkSizeAlignmentCurve, 0.8);
+                    if (currentChunkSize === maxChunkSize)
+                        maxChunkSizeAlignment = 1;
+                }
+                if (adjustedProbability + maxChunkSizeAlignment >= topProbabilityScore && adjustedProbability > 0) {
+                    pushSeparatorIndex(i + 1, usedSeparatorToken);
+                    // update the weight of the current token with the adjusted probability in the trail
+                    if (recentProbabilitiesTrail.length > 1) {
+                        weight /= recentProbabilitiesTrail.pop();
+                        recentProbabilitiesTrail.push(adjustedProbability);
+                        weight *= adjustedProbability;
+                    }
+                }
+            }
+            const nextDocumentTokenProbability = nextDocumentToken == null
+                ? undefined
+                : nextProbabilities.get(nextDocumentToken);
+            if (nextDocumentTokenProbability != null && nextDocumentTokenProbability > 0) {
+                recentProbabilitiesTrail.push(nextDocumentTokenProbability);
+                weight *= nextDocumentTokenProbability;
+                if (recentProbabilitiesTrail.length > normalizationTrailSize)
+                    weight /= recentProbabilitiesTrail.shift();
+            }
+        }
+    });
+    if (lastPushedSeparatorIndex !== documentTokens.length)
+        pushSeparatorIndex(documentTokens.length, separatorTokens[0]);
+    return res;
+}
+const idealTokenTexts = [
+    "\u6bb5", // means "section" in Chinese (according to https://github.com/ZeroEntropy-AI/llama-chunk)
+    "\u987f", // means "pause" in Chinese (according to Llama 3.1 8B and Qwen 2.5 3B)
+    "\u00a1", // inverted exclamation mark
+    "|",
+    "_"
+];
+function findAppropriateSeparatorTokens(model, maxTokens = 2) {
+    const idealTextsSet = new Set(idealTokenTexts);
+    const foundTokens = [];
+    for (const token of model.iterateAllTokens()) {
+        if (model.isSpecialToken(token))
+            continue;
+        const text = model.detokenize([token]);
+        const trimmedText = text.trim();
+        if (idealTextsSet.has(trimmedText)) {
+            const textIndex = idealTokenTexts.findIndex((idealText) => idealText === trimmedText);
+            if (foundTokens[textIndex] == null || text === trimmedText)
+                foundTokens[textIndex] = token;
+        }
+    }
+    const res = [];
+    for (let i = 0; i < idealTokenTexts.length; i++) {
+        const token = foundTokens[i];
+        if (token != null)
+            res.push(token);
+    }
+    return res.slice(0, maxTokens);
+}
+function getDefaultPrompt({ separatorTokens, tokenizer, maxChunkSize = 500 }) {
+    if (separatorTokens.length === 0)
+        throw new Error("No separator tokens provided");
+    else if (separatorTokens.length > 2)
+        throw new Error("Maximum of 2 separator tokens are supported");
+    return LlamaText.joinValues("\n", [
+        'Your job is to act as a "Chunker", for usage in RAG pipelines. The user will provide a long document.',
+        "",
+        "You should repeat the exact same message verbatim. EXCEPT, you should insert split tokens throughout the document.",
+        "",
+        "# Instructions",
+        LlamaText([
+            "- For splits, use `",
+            new SpecialTokensText(tokenizer.detokenize([separatorTokens[0]])),
+            '` as the "big split token" separator.'
+        ]),
+        separatorTokens.length > 1 && (LlamaText([
+            "- For small splits, use `",
+            new SpecialTokensText(tokenizer.detokenize([separatorTokens[1]])),
+            '` as the "big split token" separator.'
+        ])),
+        "- For example, in text document, small splits will be per-sentence, and big splits will be per-section. Do a big split BEFORE the header that defines a section.",
+        LlamaText([
+            "- You may get a user message that is unstructured or not structured cleanly. " +
+                "Still try to split that input as best as you can, even if it means doing a small split every ", Math.ceil(maxChunkSize / 5),
+            " characters, and a big split every ", Math.floor(maxChunkSize), " characters."
+        ]),
+        "- You should prefer to wait until the end of a newline or period to break, instead of breaking one or two tokens before that. If there are no newlines or periods, pick some other reasonable breakpoints instead.",
+        "- Your input could be anything - code, HTML, markdown, etc. You MUST try to output SOME split regardless of the input. Pick something reasonable! E.g. for nodejs, do a small split after every line or code block, and a big split after every function or class definitions.",
+        '- For HTML, add a small split token after every closing tag and sentence. Add a big split token after every closing tag of an "important" tag.',
+        "- Please note that you will sometimes not see your own splits in your previous output, that's OK, you MUST continue to try to output split tokens"
+    ].filter((x) => x !== false));
+}
+function adjustExponential(value, exponent, weight) {
+    if (value < 0)
+        return 0;
+    else if (value > 1)
+        return 1;
+    return (value * (1 - weight)) + (weight * Math.pow(value, exponent));
+}
+//# sourceMappingURL=chunkDocument.js.map

package/dist/evaluator/utils/chunkDocument.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"chunkDocument.js","sourceRoot":"","sources":["../../../src/evaluator/utils/chunkDocument.ts"],"names":[],"mappings":"AAGA,OAAO,EAAC,SAAS,EAAE,iBAAiB,EAAC,MAAM,0BAA0B,CAAC;AACtE,OAAO,EAAC,kBAAkB,EAAC,MAAM,gDAAgD,CAAC;AAElF,OAAO,EAAC,iBAAiB,EAAC,MAAM,kCAAkC,CAAC;AACnE,OAAO,EAAC,0BAA0B,EAAC,MAAM,iBAAiB,CAAC;AAE3D;;;;;;;;GAQG;AACH,MAAM,CAAC,KAAK,UAAU,yBAAyB,CAAC,OA8E/C;IACG,MAAM,EACF,eAAe,EACf,QAAQ,EACR,eAAe,GAAG,8BAA8B,CAAC,eAAe,CAAC,KAAK,CAAC,EACvE,eAAe,GAAG,gBAAgB,EAClC,YAAY,GAAG,GAAG,EAClB,0BAA0B,GAAG,CAAC,EAC9B,eAAe,EAAE,EACb,SAAS,EAAE,kBAAkB,GAAG,CAAC,EACjC,YAAY,EAAE,2BAA2B,GAAG,CAAC,EAAE,EAAE,GAAG,EAAE,GAAG,CAAC,EAC7D,GAAG,EAAE,EACN,eAAe,GAAG,CAAC,EACnB,sBAAsB,GAAG,GAAG,EAC/B,GAAG,OAAO,CAAC;IAEZ,MAAM,aAAa,GAAG,iBAAiB,CAAC,OAAO,CAAC,aAAa,CAAC,CAAC;IAC/D,MAAM,WAAW,GAAG,iBAAiB,CAAC,OAAO,CAAC,WAAW,CAAC,CAAC;IAE3D,IAAI,eAAe,CAAC,MAAM,KAAK,CAAC;QAC5B,MAAM,IAAI,KAAK,CAAC,mCAAmC,CAAC,CAAC;IAEzD,MAAM,WAAW,GAAsB,CAAC;YACpC,IAAI,EAAE,QAAQ;YACd,IAAI,EAAE,SAAS,CAAC,eAAe,CAAC;gBAC5B,eAAe;gBACf,SAAS,EAAE,eAAe,CAAC,KAAK,CAAC,SAAS;gBAC1C,YAAY,EAAE,YAAY,IAAI,CAAC;oBAC3B,CAAC,CAAC,SAAS;oBACX,CAAC,CAAC,YAAY;aACrB,CAAC,CAAC,CAAC,MAAM,EAAE;SACf,EAAE;YACC,IAAI,EAAE,MAAM;YACZ,IAAI,EAAE,QAAQ;SACjB,EAAE;YACC,IAAI,EAAE,OAAO;YACb,QAAQ,EAAE,CAAC,EAAE,CAAC;SACjB,CAAC,CAAC;IACH,MAAM,WAAW,GAAG,kBAAkB,CAAC,eAAe,CAAC,KAAK,CAAC,CAAC;IAC9D,MAAM,EAAC,WAAW,EAAC,GAAG,WAAW,CAAC,oBAAoB,CAAC,EAAC,WAAW,EAAC,CAAC,CAAC;IACtE,MAAM,oBAAoB,GAAG,WAAW,CAAC,QAAQ,CAAC,eAAe,CAAC,KAAK,CAAC,SAAS,EAAE,kBAAkB,CAAC,CAAC;IACvG,MAAM,cAAc,GAAG,eAAe,CAAC,KAAK,CAAC,QAAQ,CAAC,QAAQ,EAAE,KAAK,EAAE,kBAAkB,CAAC,CAAC;IAC3F,MAAM,8BAA8B,GAAG,IAAI,GAAG,CAAC,2BAA2B,CAAC,CAAC;IAE5E,IAAI,oBAAoB,CAAC,MAAM,GAAG,cAAc,CAAC,MAAM,GAAG,eAAe,CAAC,OAAO,CAAC,WAAW;QACzF,MAAM,IAAI,KAAK,CAAC,2DAA2D,CAAC,CAAC;IAEjF,MAAM,aAAa,GAAkC,oBAAoB,CAAC,KAAK,EAAE,CAAC;IAClF,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,cAAc,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;QACjD,MAAM,KAAK,GAAG,cAAc,CAAC,CAAC,CAAE,CAAC;QACjC,aAAa,CAAC,IAAI,CAAC,CAAC,KAAK,EAAE;gBACvB,YAAY,EAAE;oBACV,aAAa,EAAE,IAAI;iBACtB;aACJ,CAAC,CAAC,CAAC;IACR,CAAC;IAED,IAAI,MAAM,GAAG,CAAC,CAAC;IACf,MAAM,wBAAwB,GAAa,EAAE,CAAC;IAE9C,IAAI,eAAe,GAAG,CAAC,CAAC;IACxB,IAAI,wBAAwB,GAAG,CAAC,CAAC;IACjC,MAAM,MAAM,GAAc,EAAE,CAAC;IAC7B,MAAM,GAAG,GAAa,EAAE,CAAC;IACzB,SAAS,kBAAkB,CAAC,aAAqB,EAAE,cAAqB;QACpE,wBAAwB,GAAG,aAAa,CAAC;QAEzC,IAAI,aAAa,IAAI,eAAe;YAChC,OAAO;QAEX,IAAI,QAAQ,GAAG,aAAa,CAAC;QAC7B,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,kBAAkB,IAAI,cAAc,CAAC,QAAQ,GAAG,CAAC,CAAC,IAAI,IAAI,EAAE,CAAC,EAAE,EAAE,CAAC;YAClF,MAAM,IAAI,GAAG,eAAe,CAAC,KAAK,CAAC,UAAU,CAAC,CAAC,cAAc,CAAC,QAAQ,GAAG,CAAC,CAAE,CAAC,CAAC,CAAC;YAC/E,IAAI,CAAC,8BAA8B,CAAC,GAAG,CAAC,IAAI,CAAC,IAAI,EAAE,CAAC;gBAChD,MAAM;YAEV,QAAQ,EAAE,CAAC;QACf,CAAC;QAED,MAAM,KAAK,GAAG,cAAc,CAAC,KAAK,CAAC,eAAe,EAAE,QAAQ,CAAC,CAAC;QAC9D,MAAM,IAAI,GAAG,eAAe,CAAC,KAAK,CAAC,UAAU,CACzC,KAAK,EACL,KAAK,EACL,cAAc,CAAC,KAAK,CAAC,eAAe,GAAG,0BAA0B,EAAE,eAAe,CAAC,CACtF,CAAC;QACF,MAAM,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC;QACnB,eAAe,GAAG,QAAQ,CAAC;QAE3B,aAAa,EAAE,CAAC,KAAK,EAAE,cAAc,CAAC,CAAC;QACvC,WAAW,EAAE,CAAC,IAAI,EAAE,cAAc,CAAC,CAAC;QACpC,GAAG,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;IACnB,CAAC;IAED,MAAM,eAAe,CAAC,kBAAkB,CAAC,aAAa,EAAE;QACpD,aAAa,CAAC,eAAe,EAAE,MAAM;YACjC,MAAM,CAAC,GAAG,eAAe,GAAG,oBAAoB,CAAC,MAAM,CAAC;YACxD,MAAM,iBAAiB,GAAG,MAAM,EAAE,IAAI,EAAE,aAAa,CAAC;YACtD,MAAM,iBAAiB,GAAG,cAAc,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC;YAEhD,IAAI,iBAAiB,IAAI,IAAI;gBACzB,MAAM,IAAI,KAAK,CAAC,+BAA+B,GAAG,CAAC,CAAC,CAAC;YAEzD,MAAM,mBAAmB,GAAG,iBAAiB,CAAC,OAAO,EAAE;iBAClD,IAAI,EAAE,CAAC,KAAK,EAAE,CAAC,CAAC,CAAC,CAAC;YACvB,MAAM,CAAC,kBAAkB,EAAE,oBAAoB,CAAC,GAAG,eAAe;iBAC7D,MAAM,CAAC,CAAC,KAAK,EAAE,EAAE,CAAC,KAAK,KAAK,iBAAiB,CAAC,CAAC,qCAAqC;iBACpF,GAAG,CAAC,CAAC,KAAK,EAAE,EAAE,CAAC,CAAC,KAAK,EAAE,iBAAiB,CAAC,GAAG,CAAC,KAAK,CAAC,CAAoD,CAAC;iBACxG,MAAM,CAAC,CAAC,IAAI,EAA+C,EAAE,CAAC,IAAI,CAAC,CAAC,CAAC,IAAI,IAAI,CAAC;iBAC9E,MAAM,CAAC,CAAC,CAAC,MAAM,EAAE,YAAY,CAAC,EAAE,CAAC,MAAM,EAAE,YAAY,CAAC,EAAE,EAAE;gBACvD,IAAI,YAAY,IAAI,YAAY;oBAC5B,OAAO,CAAC,MAAM,EAAE,YAAY,CAAC,CAAC;gBAElC,OAAO,CAAC,MAAM,EAAE,YAAY,CAAC,CAAC;YAClC,CAAC,EAAE,CAAC,eAAe,CAAC,CAAC,CAAE,EAAE,CAAC,CAAC,CAAC,CAAC;YAEjC,IAAI,mBAAmB,IAAI,IAAI,IAAI,oBAAoB,IAAI,IAAI,IAAI,oBAAoB,KAAK,CAAC;gBACzF,OAAO;YAEX,eAAe;YACf,iEAAiE;YACjE,0DAA0D;YAC1D,yGAAyG;YACzG,KAAK;YAEL,IAAI,oBAAoB,IAAI,mBAAmB;gBAC3C,kBAAkB,CAAC,CAAC,GAAG,CAAC,EAAE,kBAAkB,CAAC,CAAC;iBAC7C,IAAI,CAAC,GAAG,eAAe,EAAE,CAAC;gBAC3B,MAAM,mBAAmB,GAAG,oBAAoB,GAAG,CAAC,MAAM,GAAG,CAAC,CAAC,GAAG,oBAAoB,CAAC,CAAC,CAAC;gBACzF,IAAI,qBAAqB,GAAG,CAAC,CAAC;gBAC9B,IAAI,YAAY,KAAK,CAAC,IAAI,mBAAmB,GAAG,mBAAmB,EAAE,CAAC;oBAClE,MAAM,eAAe,GAAG,CAAC,GAAG,mBAAmB,CAAC;oBAChD,MAAM,gBAAgB,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,CAAC,GAAG,CAAC,GAAG,eAAe,CAAC,CAAC;oBAC9D,qBAAqB,GAAG,gBAAgB,KAAK,CAAC;wBAC1C,CAAC,CAAC,CAAC;wBACH,CAAC,CAAC,iBAAiB,CACf,eAAe,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,gBAAgB,GAAG,YAAY,CAAC,EAC9D,0BAA0B,IAAI,CAAC;4BAC3B,CAAC,CAAC,CAAC;4BACH,CAAC,CAAC,0BAA0B,EAChC,GAAG,CACN,CAAC;oBAEN,IAAI,gBAAgB,KAAK,YAAY;wBACjC,qBAAqB,GAAG,CAAC,CAAC;gBAClC,CAAC;gBAED,IAAI,mBAAmB,GAAG,qBAAqB,IAAI,mBAAmB,IAAI,mBAAmB,GAAG,CAAC,EAAE,CAAC;oBAChG,kBAAkB,CAAC,CAAC,GAAG,CAAC,EAAE,kBAAkB,CAAC,CAAC;oBAE9C,oFAAoF;oBACpF,IAAI,wBAAwB,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;wBACtC,MAAM,IAAI,wBAAwB,CAAC,GAAG,EAAG,CAAC;wBAC1C,wBAAwB,CAAC,IAAI,CAAC,mBAAmB,CAAC,CAAC;wBACnD,MAAM,IAAI,mBAAmB,CAAC;oBAClC,CAAC;gBACL,CAAC;YACL,CAAC;YAED,MAAM,4BAA4B,GAAG,iBAAiB,IAAI,IAAI;gBAC1D,CAAC,CAAC,SAAS;gBACX,CAAC,CAAC,iBAAiB,CAAC,GAAG,CAAC,iBAAiB,CAAC,CAAC;YAC/C,IAAI,4BAA4B,IAAI,IAAI,IAAI,4BAA4B,GAAG,CAAC,EAAE,CAAC;gBAC3E,wBAAwB,CAAC,IAAI,CAAC,4BAA4B,CAAC,CAAC;gBAC5D,MAAM,IAAI,4BAA4B,CAAC;gBAEvC,IAAI,wBAAwB,CAAC,MAAM,GAAG,sBAAsB;oBACxD,MAAM,IAAI,wBAAwB,CAAC,KAAK,EAAG,CAAC;YACpD,CAAC;QACL,CAAC;KACJ,CAAC,CAAC;IAEH,IAAI,wBAAwB,KAAK,cAAc,CAAC,MAAM;QAClD,kBAAkB,CAAC,cAAc,CAAC,MAAM,EAAE,eAAe,CAAC,CAAC,CAAE,CAAC,CAAC;IAEnE,OAAO,GAAG,CAAC;AACf,CAAC;AAED,MAAM,eAAe,GAAG;IACpB,QAAQ,EAAE,0FAA0F;IACpG,QAAQ,EAAE,uEAAuE;IACjF,QAAQ,EAAE,4BAA4B;IACtC,GAAG;IACH,GAAG;CACN,CAAC;AACF,SAAS,8BAA8B,CAAC,KAAiB,EAAE,YAAoB,CAAC;IAC5E,MAAM,aAAa,GAAG,IAAI,GAAG,CAAC,eAAe,CAAC,CAAC;IAC/C,MAAM,WAAW,GAAY,EAAE,CAAC;IAEhC,KAAK,MAAM,KAAK,IAAI,KAAK,CAAC,gBAAgB,EAAE,EAAE,CAAC;QAC3C,IAAI,KAAK,CAAC,cAAc,CAAC,KAAK,CAAC;YAC3B,SAAS;QAEb,MAAM,IAAI,GAAG,KAAK,CAAC,UAAU,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC;QACvC,MAAM,WAAW,GAAG,IAAI,CAAC,IAAI,EAAE,CAAC;QAChC,IAAI,aAAa,CAAC,GAAG,CAAC,WAAW,CAAC,EAAE,CAAC;YACjC,MAAM,SAAS,GAAG,eAAe,CAAC,SAAS,CAAC,CAAC,SAAS,EAAE,EAAE,CAAC,SAAS,KAAK,WAAW,CAAC,CAAC;YAEtF,IAAI,WAAW,CAAC,SAAS,CAAC,IAAI,IAAI,IAAI,IAAI,KAAK,WAAW;gBACtD,WAAW,CAAC,SAAS,CAAC,GAAG,KAAK,CAAC;QACvC,CAAC;IACL,CAAC;IAED,MAAM,GAAG,GAAY,EAAE,CAAC;IACxB,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,eAAe,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;QAC9C,MAAM,KAAK,GAAG,WAAW,CAAC,CAAC,CAAC,CAAC;QAE7B,IAAI,KAAK,IAAI,IAAI;YACb,GAAG,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC;IACxB,CAAC;IAED,OAAO,GAAG,CAAC,KAAK,CAAC,CAAC,EAAE,SAAS,CAAC,CAAC;AACnC,CAAC;AAED,SAAS,gBAAgB,CAAC,EACtB,eAAe,EAAE,SAAS,EAAE,YAAY,GAAG,GAAG,EAGjD;IACG,IAAI,eAAe,CAAC,MAAM,KAAK,CAAC;QAC5B,MAAM,IAAI,KAAK,CAAC,8BAA8B,CAAC,CAAC;SAC/C,IAAI,eAAe,CAAC,MAAM,GAAG,CAAC;QAC/B,MAAM,IAAI,KAAK,CAAC,6CAA6C,CAAC,CAAC;IAEnE,OAAO,SAAS,CAAC,UAAU,CAAC,IAAI,EAAE;QAC9B,uGAAuG;QACvG,EAAE;QACF,oHAAoH;QACpH,EAAE;QACF,gBAAgB;QAChB,SAAS,CAAC;YACN,qBAAqB;YACrB,IAAI,iBAAiB,CAAC,SAAS,CAAC,UAAU,CAAC,CAAC,eAAe,CAAC,CAAC,CAAE,CAAC,CAAC,CAAC;YAClE,uCAAuC;SAC1C,CAAC;QACF,eAAe,CAAC,MAAM,GAAG,CAAC,IAAI,CAC1B,SAAS,CAAC;YACN,2BAA2B;YAC3B,IAAI,iBAAiB,CAAC,SAAS,CAAC,UAAU,CAAC,CAAC,eAAe,CAAC,CAAC,CAAE,CAAC,CAAC,CAAC;YAClE,uCAAuC;SAC1C,CAAC,CACL;QACD,kKAAkK;QAClK,SAAS,CAAC;YACN,+EAA+E;gBAC/E,+FAA+F,EAAE,IAAI,CAAC,IAAI,CAAC,YAAY,GAAG,CAAC,CAAC;YAC5H,qCAAqC,EAAE,IAAI,CAAC,KAAK,CAAC,YAAY,CAAC,EAAE,cAAc;SAClF,CAAC;QACF,oNAAoN;QACpN,gRAAgR;QAChR,gJAAgJ;QAChJ,mJAAmJ;KACtJ,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,KAAK,KAAK,CAAC,CAAC,CAAC;AAClC,CAAC;AAED,SAAS,iBAAiB,CAAC,KAAa,EAAE,QAAgB,EAAE,MAAc;IACtE,IAAI,KAAK,GAAG,CAAC;QACT,OAAO,CAAC,CAAC;SACR,IAAI,KAAK,GAAG,CAAC;QACd,OAAO,CAAC,CAAC;IAEb,OAAO,CAAC,KAAK,GAAG,CAAC,CAAC,GAAG,MAAM,CAAC,CAAC,GAAG,CAAC,MAAM,GAAG,IAAI,CAAC,GAAG,CAAC,KAAK,EAAE,QAAQ,CAAC,CAAC,CAAC;AACzE,CAAC"}

package/dist/gguf/consts.d.ts

@@ -0,0 +1,4 @@
+import retry from "async-retry";
+export declare const ggufDefaultFetchRetryOptions: retry.Options;
+export declare const defaultExtraAllocationSize: number;
+export declare const noDirectSubNestingGGufMetadataKeys: readonly string[];

package/dist/gguf/consts.js

@@ -0,0 +1,12 @@
+export const ggufDefaultFetchRetryOptions = {
+    retries: 10,
+    factor: 2,
+    minTimeout: 1000,
+    maxTimeout: 1000 * 16
+};
+export const defaultExtraAllocationSize = 1024 * 1024 * 4; // 4MB
+export const noDirectSubNestingGGufMetadataKeys = [
+    "general.license",
+    "tokenizer.chat_template"
+];
+//# sourceMappingURL=consts.js.map

package/dist/gguf/consts.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"consts.js","sourceRoot":"","sources":["../../src/gguf/consts.ts"],"names":[],"mappings":"AAEA,MAAM,CAAC,MAAM,4BAA4B,GAAkB;IACvD,OAAO,EAAE,EAAE;IACX,MAAM,EAAE,CAAC;IACT,UAAU,EAAE,IAAI;IAChB,UAAU,EAAE,IAAI,GAAG,EAAE;CACf,CAAC;AAEX,MAAM,CAAC,MAAM,0BAA0B,GAAG,IAAI,GAAG,IAAI,GAAG,CAAC,CAAC,CAAC,MAAM;AAEjE,MAAM,CAAC,MAAM,kCAAkC,GAAsB;IACjE,iBAAiB;IACjB,yBAAyB;CAC5B,CAAC"}

package/dist/gguf/errors/InvalidGgufMagicError.d.ts

@@ -0,0 +1,3 @@
+export declare class InvalidGgufMagicError extends Error {
+    constructor(expectedGgufMagic: string, actualGgufMagic: string);
+}

package/dist/gguf/errors/InvalidGgufMagicError.js

@@ -0,0 +1,6 @@
+export class InvalidGgufMagicError extends Error {
+    constructor(expectedGgufMagic, actualGgufMagic) {
+        super(`Invalid GGUF magic. Expected "${expectedGgufMagic}" but got "${actualGgufMagic}".`);
+    }
+}
+//# sourceMappingURL=InvalidGgufMagicError.js.map

package/dist/gguf/errors/InvalidGgufMagicError.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"InvalidGgufMagicError.js","sourceRoot":"","sources":["../../../src/gguf/errors/InvalidGgufMagicError.ts"],"names":[],"mappings":"AAAA,MAAM,OAAO,qBAAsB,SAAQ,KAAK;IAC5C,YAAmB,iBAAyB,EAAE,eAAuB;QACjE,KAAK,CAAC,iCAAiC,iBAAiB,cAAc,eAAe,IAAI,CAAC,CAAC;IAC/F,CAAC;CACJ"}

package/dist/gguf/errors/UnsupportedGgufValueTypeError.d.ts

@@ -0,0 +1,4 @@
+export declare class UnsupportedGgufValueTypeError extends Error {
+    readonly ggufValueType: number;
+    constructor(ggufValueType: number);
+}

package/dist/gguf/errors/UnsupportedGgufValueTypeError.js

@@ -0,0 +1,9 @@
+export class UnsupportedGgufValueTypeError extends Error {
+    ggufValueType;
+    constructor(ggufValueType) {
+        super(`Unsupported GGUF value type "${ggufValueType}"`);
+        Object.defineProperty(this, "ggufValueType", { enumerable: false });
+        this.ggufValueType = ggufValueType;
+    }
+}
+//# sourceMappingURL=UnsupportedGgufValueTypeError.js.map

package/dist/gguf/errors/UnsupportedGgufValueTypeError.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"UnsupportedGgufValueTypeError.js","sourceRoot":"","sources":["../../../src/gguf/errors/UnsupportedGgufValueTypeError.ts"],"names":[],"mappings":"AAAA,MAAM,OAAO,6BAA8B,SAAQ,KAAK;IACpC,aAAa,CAAS;IAEtC,YAAmB,aAAqB;QACpC,KAAK,CAAC,gCAAgC,aAAa,GAAG,CAAC,CAAC;QAExD,MAAM,CAAC,cAAc,CAAC,IAAI,EAAE,eAAoC,EAAE,EAAC,UAAU,EAAE,KAAK,EAAC,CAAC,CAAC;QAEvF,IAAI,CAAC,aAAa,GAAG,aAAa,CAAC;IACvC,CAAC;CACJ"}

package/dist/gguf/fileReaders/GgufFileReader.d.ts

@@ -0,0 +1,36 @@
+import { GgufReadOffset } from "../utils/GgufReadOffset.js";
+import { Promisable } from "../../utils/transformPromisable.js";
+export declare const valueTypeToBytesToRead: {
+    readonly uint8: 1;
+    readonly uint16: 2;
+    readonly uint32: 4;
+    readonly uint64: 8;
+    readonly int8: 1;
+    readonly int16: 2;
+    readonly int32: 4;
+    readonly int64: 8;
+    readonly float32: 4;
+    readonly float64: 8;
+    readonly bool: 1;
+};
+export declare abstract class GgufFileReader {
+    protected _buffer: Buffer<ArrayBuffer>;
+    abstract readByteRange(offset: number | GgufReadOffset, length: number): Promisable<Buffer>;
+    protected abstract ensureHasByteRange(offset: number | GgufReadOffset, length: number): Promisable<void>;
+    readUint8(offset: number | GgufReadOffset): Promisable<number>;
+    readUint16(offset: number | GgufReadOffset): Promisable<number>;
+    readUint32(offset: number | GgufReadOffset): Promisable<number>;
+    readUint64(offset: number | GgufReadOffset): Promisable<bigint>;
+    readInt8(offset: number | GgufReadOffset): Promisable<number>;
+    readInt16(offset: number | GgufReadOffset): Promisable<number>;
+    readInt32(offset: number | GgufReadOffset): Promisable<number>;
+    readInt64(offset: number | GgufReadOffset): Promisable<bigint>;
+    readFloat32(offset: number | GgufReadOffset): Promisable<number>;
+    readFloat64(offset: number | GgufReadOffset): Promisable<number>;
+    readBool(offset: number | GgufReadOffset): Promisable<boolean>;
+    readString(offset: number | GgufReadOffset): Promisable<string>;
+    readStringWithLength(offset: number | GgufReadOffset, length: number): Promisable<string>;
+    protected _addToBuffer(buffer: Buffer): void;
+    private _withBufferRead;
+    static castNumberIfSafe(value: bigint): number | bigint;
+}

package/dist/gguf/fileReaders/GgufFileReader.js

@@ -0,0 +1,106 @@
+import { GgufReadOffset } from "../utils/GgufReadOffset.js";
+import { transformPromisable } from "../../utils/transformPromisable.js";
+export const valueTypeToBytesToRead = {
+    uint8: 1,
+    uint16: 2,
+    uint32: 4,
+    uint64: 8,
+    int8: 1,
+    int16: 2,
+    int32: 4,
+    int64: 8,
+    float32: 4,
+    float64: 8,
+    bool: 1
+};
+export class GgufFileReader {
+    _buffer = Buffer.alloc(0);
+    readUint8(offset) {
+        return this._withBufferRead(offset, valueTypeToBytesToRead.uint8, (resolvedOffset) => {
+            return this._buffer.readUInt8(resolvedOffset);
+        });
+    }
+    readUint16(offset) {
+        return this._withBufferRead(offset, valueTypeToBytesToRead.uint16, (resolvedOffset) => {
+            return this._buffer.readUInt16LE(resolvedOffset);
+        });
+    }
+    readUint32(offset) {
+        return this._withBufferRead(offset, valueTypeToBytesToRead.uint32, (resolvedOffset) => {
+            return this._buffer.readUInt32LE(resolvedOffset);
+        });
+    }
+    readUint64(offset) {
+        return this._withBufferRead(offset, valueTypeToBytesToRead.uint64, (resolvedOffset) => {
+            return this._buffer.readBigUInt64LE(resolvedOffset);
+        });
+    }
+    readInt8(offset) {
+        return this._withBufferRead(offset, valueTypeToBytesToRead.int8, (resolvedOffset) => {
+            return this._buffer.readInt8(resolvedOffset);
+        });
+    }
+    readInt16(offset) {
+        return this._withBufferRead(offset, valueTypeToBytesToRead.int16, (resolvedOffset) => {
+            return this._buffer.readInt16LE(resolvedOffset);
+        });
+    }
+    readInt32(offset) {
+        return this._withBufferRead(offset, valueTypeToBytesToRead.int32, (resolvedOffset) => {
+            return this._buffer.readInt32LE(resolvedOffset);
+        });
+    }
+    readInt64(offset) {
+        return this._withBufferRead(offset, valueTypeToBytesToRead.int64, (resolvedOffset) => {
+            return this._buffer.readBigInt64LE(resolvedOffset);
+        });
+    }
+    readFloat32(offset) {
+        return this._withBufferRead(offset, valueTypeToBytesToRead.float32, (resolvedOffset) => {
+            return this._buffer.readFloatLE(resolvedOffset);
+        });
+    }
+    readFloat64(offset) {
+        return this._withBufferRead(offset, valueTypeToBytesToRead.float64, (resolvedOffset) => {
+            return this._buffer.readDoubleLE(resolvedOffset);
+        });
+    }
+    readBool(offset) {
+        return this._withBufferRead(offset, valueTypeToBytesToRead.uint8, (resolvedOffset) => {
+            return this._buffer.readUInt8(resolvedOffset) === 1;
+        });
+    }
+    readString(offset) {
+        const readOffset = GgufReadOffset.resolveReadOffset(offset);
+        return transformPromisable(this.readUint64(readOffset), (length) => {
+            return this.readStringWithLength(readOffset, Number(length));
+        });
+    }
+    readStringWithLength(offset, length) {
+        const readLength = valueTypeToBytesToRead.uint8 * length;
+        return this._withBufferRead(offset, readLength, (resolvedOffset) => {
+            return this._buffer.toString("utf8", resolvedOffset, Math.min(resolvedOffset + readLength, this._buffer.length));
+        });
+    }
+    _addToBuffer(buffer) {
+        const newBuffer = Buffer.alloc(this._buffer.byteLength + buffer.byteLength);
+        this._buffer.copy(newBuffer);
+        buffer.copy(newBuffer, this._buffer.byteLength);
+        this._buffer = newBuffer;
+    }
+    _withBufferRead(offset, length, reader) {
+        return transformPromisable(this.ensureHasByteRange(offset, length), () => {
+            const resolvedOffset = GgufReadOffset.resolveReadOffset(offset);
+            return transformPromisable(reader(resolvedOffset.offset), (res) => {
+                resolvedOffset.moveBy(Math.min(length, this._buffer.length - resolvedOffset.offset));
+                return res;
+            });
+        });
+    }
+    static castNumberIfSafe(value) {
+        if (value > Number.MAX_SAFE_INTEGER)
+            return value;
+        return Number(value);
+    }
+}
+//# sourceMappingURL=GgufFileReader.js.map

package/dist/gguf/fileReaders/GgufFileReader.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"GgufFileReader.js","sourceRoot":"","sources":["../../../src/gguf/fileReaders/GgufFileReader.ts"],"names":[],"mappings":"AAAA,OAAO,EAAC,cAAc,EAAC,MAAM,4BAA4B,CAAC;AAC1D,OAAO,EAAa,mBAAmB,EAAC,MAAM,oCAAoC,CAAC;AAEnF,MAAM,CAAC,MAAM,sBAAsB,GAAG;IAClC,KAAK,EAAE,CAAC;IACR,MAAM,EAAE,CAAC;IACT,MAAM,EAAE,CAAC;IACT,MAAM,EAAE,CAAC;IACT,IAAI,EAAE,CAAC;IACP,KAAK,EAAE,CAAC;IACR,KAAK,EAAE,CAAC;IACR,KAAK,EAAE,CAAC;IACR,OAAO,EAAE,CAAC;IACV,OAAO,EAAE,CAAC;IACV,IAAI,EAAE,CAAC;CACD,CAAC;AAEX,MAAM,OAAgB,cAAc;IACtB,OAAO,GAAG,MAAM,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC;IAK7B,SAAS,CAAC,MAA+B;QAC5C,OAAO,IAAI,CAAC,eAAe,CAAC,MAAM,EAAE,sBAAsB,CAAC,KAAK,EAAE,CAAC,cAAc,EAAE,EAAE;YACjF,OAAO,IAAI,CAAC,OAAO,CAAC,SAAS,CAAC,cAAc,CAAC,CAAC;QAClD,CAAC,CAAC,CAAC;IACP,CAAC;IAEM,UAAU,CAAC,MAA+B;QAC7C,OAAO,IAAI,CAAC,eAAe,CAAC,MAAM,EAAE,sBAAsB,CAAC,MAAM,EAAE,CAAC,cAAc,EAAE,EAAE;YAClF,OAAO,IAAI,CAAC,OAAO,CAAC,YAAY,CAAC,cAAc,CAAC,CAAC;QACrD,CAAC,CAAC,CAAC;IACP,CAAC;IAEM,UAAU,CAAC,MAA+B;QAC7C,OAAO,IAAI,CAAC,eAAe,CAAC,MAAM,EAAE,sBAAsB,CAAC,MAAM,EAAE,CAAC,cAAc,EAAE,EAAE;YAClF,OAAO,IAAI,CAAC,OAAO,CAAC,YAAY,CAAC,cAAc,CAAC,CAAC;QACrD,CAAC,CAAC,CAAC;IACP,CAAC;IAEM,UAAU,CAAC,MAA+B;QAC7C,OAAO,IAAI,CAAC,eAAe,CAAC,MAAM,EAAE,sBAAsB,CAAC,MAAM,EAAE,CAAC,cAAc,EAAE,EAAE;YAClF,OAAO,IAAI,CAAC,OAAO,CAAC,eAAe,CAAC,cAAc,CAAC,CAAC;QACxD,CAAC,CAAC,CAAC;IACP,CAAC;IAEM,QAAQ,CAAC,MAA+B;QAC3C,OAAO,IAAI,CAAC,eAAe,CAAC,MAAM,EAAE,sBAAsB,CAAC,IAAI,EAAE,CAAC,cAAc,EAAE,EAAE;YAChF,OAAO,IAAI,CAAC,OAAO,CAAC,QAAQ,CAAC,cAAc,CAAC,CAAC;QACjD,CAAC,CAAC,CAAC;IACP,CAAC;IAEM,SAAS,CAAC,MAA+B;QAC5C,OAAO,IAAI,CAAC,eAAe,CAAC,MAAM,EAAE,sBAAsB,CAAC,KAAK,EAAE,CAAC,cAAc,EAAE,EAAE;YACjF,OAAO,IAAI,CAAC,OAAO,CAAC,WAAW,CAAC,cAAc,CAAC,CAAC;QACpD,CAAC,CAAC,CAAC;IACP,CAAC;IAEM,SAAS,CAAC,MAA+B;QAC5C,OAAO,IAAI,CAAC,eAAe,CAAC,MAAM,EAAE,sBAAsB,CAAC,KAAK,EAAE,CAAC,cAAc,EAAE,EAAE;YACjF,OAAO,IAAI,CAAC,OAAO,CAAC,WAAW,CAAC,cAAc,CAAC,CAAC;QACpD,CAAC,CAAC,CAAC;IACP,CAAC;IAEM,SAAS,CAAC,MAA+B;QAC5C,OAAO,IAAI,CAAC,eAAe,CAAC,MAAM,EAAE,sBAAsB,CAAC,KAAK,EAAE,CAAC,cAAc,EAAE,EAAE;YACjF,OAAO,IAAI,CAAC,OAAO,CAAC,cAAc,CAAC,cAAc,CAAC,CAAC;QACvD,CAAC,CAAC,CAAC;IACP,CAAC;IAEM,WAAW,CAAC,MAA+B;QAC9C,OAAO,IAAI,CAAC,eAAe,CAAC,MAAM,EAAE,sBAAsB,CAAC,OAAO,EAAE,CAAC,cAAc,EAAE,EAAE;YACnF,OAAO,IAAI,CAAC,OAAO,CAAC,WAAW,CAAC,cAAc,CAAC,CAAC;QACpD,CAAC,CAAC,CAAC;IACP,CAAC;IAEM,WAAW,CAAC,MAA+B;QAC9C,OAAO,IAAI,CAAC,eAAe,CAAC,MAAM,EAAE,sBAAsB,CAAC,OAAO,EAAE,CAAC,cAAc,EAAE,EAAE;YACnF,OAAO,IAAI,CAAC,OAAO,CAAC,YAAY,CAAC,cAAc,CAAC,CAAC;QACrD,CAAC,CAAC,CAAC;IACP,CAAC;IAEM,QAAQ,CAAC,MAA+B;QAC3C,OAAO,IAAI,CAAC,eAAe,CAAC,MAAM,EAAE,sBAAsB,CAAC,KAAK,EAAE,CAAC,cAAc,EAAE,EAAE;YACjF,OAAO,IAAI,CAAC,OAAO,CAAC,SAAS,CAAC,cAAc,CAAC,KAAK,CAAC,CAAC;QACxD,CAAC,CAAC,CAAC;IACP,CAAC;IAEM,UAAU,CAAC,MAA+B;QAC7C,MAAM,UAAU,GAAG,cAAc,CAAC,iBAAiB,CAAC,MAAM,CAAC,CAAC;QAE5D,OAAO,mBAAmB,CAAC,IAAI,CAAC,UAAU,CAAC,UAAU,CAAC,EAAE,CAAC,MAAM,EAAE,EAAE;YAC/D,OAAO,IAAI,CAAC,oBAAoB,CAAC,UAAU,EAAE,MAAM,CAAC,MAAM,CAAC,CAAC,CAAC;QACjE,CAAC,CAAC,CAAC;IACP,CAAC;IAEM,oBAAoB,CAAC,MAA+B,EAAE,MAAc;QACvE,MAAM,UAAU,GAAG,sBAAsB,CAAC,KAAK,GAAG,MAAM,CAAC;QAEzD,OAAO,IAAI,CAAC,eAAe,CAAC,MAAM,EAAE,UAAU,EAAE,CAAC,cAAc,EAAE,EAAE;YAC/D,OAAO,IAAI,CAAC,OAAO,CAAC,QAAQ,CAAC,MAAM,EAAE,cAAc,EAAE,IAAI,CAAC,GAAG,CAAC,cAAc,GAAG,UAAU,EAAE,IAAI,CAAC,OAAO,CAAC,MAAM,CAAC,CAAC,CAAC;QACrH,CAAC,CAAC,CAAC;IACP,CAAC;IAES,YAAY,CAAC,MAAc;QACjC,MAAM,SAAS,GAAG,MAAM,CAAC,KAAK,CAAC,IAAI,CAAC,OAAO,CAAC,UAAU,GAAG,MAAM,CAAC,UAAU,CAAC,CAAC;QAC5E,IAAI,CAAC,OAAO,CAAC,IAAI,CAAC,SAAS,CAAC,CAAC;QAC7B,MAAM,CAAC,IAAI,CAAC,SAAS,EAAE,IAAI,CAAC,OAAO,CAAC,UAAU,CAAC,CAAC;QAEhD,IAAI,CAAC,OAAO,GAAG,SAAS,CAAC;IAC7B,CAAC;IAEO,eAAe,CAAI,MAA+B,EAAE,MAAc,EAAE,MAAqC;QAC7G,OAAO,mBAAmB,CAAC,IAAI,CAAC,kBAAkB,CAAC,MAAM,EAAE,MAAM,CAAC,EAAE,GAAG,EAAE;YACrE,MAAM,cAAc,GAAG,cAAc,CAAC,iBAAiB,CAAC,MAAM,CAAC,CAAC;YAEhE,OAAO,mBAAmB,CAAC,MAAM,CAAC,cAAc,CAAC,MAAM,CAAC,EAAE,CAAC,GAAG,EAAE,EAAE;gBAC9D,cAAc,CAAC,MAAM,CAAC,IAAI,CAAC,GAAG,CAAC,MAAM,EAAE,IAAI,CAAC,OAAO,CAAC,MAAM,GAAG,cAAc,CAAC,MAAM,CAAC,CAAC,CAAC;gBAErF,OAAO,GAAG,CAAC;YACf,CAAC,CAAC,CAAC;QACP,CAAC,CAAC,CAAC;IACP,CAAC;IAEM,MAAM,CAAC,gBAAgB,CAAC,KAAa;QACxC,IAAI,KAAK,GAAG,MAAM,CAAC,gBAAgB;YAC/B,OAAO,KAAK,CAAC;QAEjB,OAAO,MAAM,CAAC,KAAK,CAAC,CAAC;IACzB,CAAC;CACJ"}

package/dist/gguf/fileReaders/GgufFsFileReader.d.ts

@@ -0,0 +1,16 @@
+import { GgufReadOffset } from "../utils/GgufReadOffset.js";
+import { GgufFileReader } from "./GgufFileReader.js";
+type GgufFsFileReaderOptions = {
+    filePath: string;
+    signal?: AbortSignal;
+};
+export declare class GgufFsFileReader extends GgufFileReader {
+    readonly filePath: string;
+    private readonly _signal?;
+    constructor({ filePath, signal }: GgufFsFileReaderOptions);
+    readByteRange(offset: number | GgufReadOffset, length: number): Buffer<ArrayBuffer> | Promise<Buffer<ArrayBuffer>>;
+    protected ensureHasByteRange(offset: number | GgufReadOffset, length: number): Promise<void> | undefined;
+    private _readToExpandBufferUpToOffset;
+    private _readByteRange;
+}
+export {};