node-llama-cpp 3.0.0-beta.3 → 3.0.0-beta.30

package/dist/evaluator/LlamaModel/LlamaModel.js

@@ -0,0 +1,679 @@
+import process from "process";
+import path from "path";
+import { AsyncDisposeAggregator, DisposedError, EventRelay, withLock } from "lifecycle-utils";
+import { removeNullFields } from "../../utils/removeNullFields.js";
+import { DisposeGuard } from "../../utils/DisposeGuard.js";
+import { LlamaLocks, LlamaLogLevel, LlamaVocabularyType, LlamaVocabularyTypeValues } from "../../bindings/types.js";
+import { readGgufFileInfo } from "../../gguf/readGgufFileInfo.js";
+import { GgufInsights } from "../../gguf/insights/GgufInsights.js";
+import { getConsoleLogPrefix } from "../../utils/getConsoleLogPrefix.js";
+import { getReadablePath } from "../../cli/utils/getReadablePath.js";
+import { LlamaContext } from "../LlamaContext/LlamaContext.js";
+import { LlamaEmbeddingContext } from "../LlamaEmbeddingContext.js";
+import { TokenAttribute, TokenAttributes } from "./utils/TokenAttributes.js";
+const defaultLoraThreads = 6;
+const defaultLoraScale = 1;
+const defaultUseMmap = true;
+export class LlamaModel {
+    /** @internal */ _llama;
+    /** @internal */ _model;
+    /** @internal */ _backendModelDisposeGuard;
+    /** @internal */ _tokens;
+    /** @internal */ _modelPath;
+    /** @internal */ _fileInfo;
+    /** @internal */ _fileInsights;
+    /** @internal */ _gpuLayers;
+    /** @internal */ _filename;
+    /** @internal */ _disposedState = { disposed: false };
+    /** @internal */ _disposeAggregator = new AsyncDisposeAggregator();
+    /** @internal */ _llamaPreventDisposalHandle;
+    /** @internal */ _typeDescription;
+    /** @internal */ _trainContextSize;
+    /** @internal */ _embeddingVectorSize;
+    /** @internal */ _vocabularyType;
+    tokenizer;
+    onDispose = new EventRelay();
+    constructor({ modelPath, gpuLayers, vocabOnly, useMmap, useMlock, checkTensors, onLoadProgress, loadSignal }, { _llama, _fileInfo, _fileInsights }) {
+        this._llama = _llama;
+        this._fileInfo = _fileInfo;
+        this._modelPath = path.resolve(process.cwd(), modelPath);
+        this._fileInsights = _fileInsights;
+        this._gpuLayers = gpuLayers;
+        this._backendModelDisposeGuard = new DisposeGuard([this._llama._backendDisposeGuard]);
+        this._llamaPreventDisposalHandle = this._llama._backendDisposeGuard.createPreventDisposalHandle();
+        this._model = new this._llama._bindings.AddonModel(this._modelPath, removeNullFields({
+            addonExports: this._llama._bindings,
+            gpuLayers,
+            vocabOnly,
+            useMmap,
+            useMlock: _llama.supportsMlock
+                ? useMlock
+                : undefined,
+            checkTensors: checkTensors ?? false,
+            onLoadProgress: onLoadProgress == null
+                ? undefined
+                : (loadPercentage) => {
+                    try {
+                        onLoadProgress(loadPercentage);
+                    }
+                    catch (err) {
+                        // the native addon code calls this function, so there's no use to throw an error here
+                        console.error(err);
+                    }
+                },
+            hasLoadAbortSignal: loadSignal != null
+        }));
+        this._tokens = LlamaModelTokens._create(this._model, this._disposedState);
+        this._filename = path.basename(modelPath);
+        this._disposeAggregator.add(() => {
+            this._disposedState.disposed = true;
+        });
+        this._disposeAggregator.add(this.onDispose.dispatchEvent);
+        this._disposeAggregator.add(this._llama.onDispose.createListener(disposeModelIfReferenced.bind(null, new WeakRef(this))));
+        this._disposeAggregator.add(async () => {
+            await this._backendModelDisposeGuard.acquireDisposeLock();
+            await this._model.dispose();
+            this._llamaPreventDisposalHandle.dispose();
+        });
+        this.tokenize = this.tokenize.bind(this);
+        this.detokenize = this.detokenize.bind(this);
+        this.isSpecialToken = this.isSpecialToken.bind(this);
+        this.tokenize.detokenize = this.detokenize;
+        this.tokenize.isSpecialToken = this.isSpecialToken;
+        this.tokenizer = this.tokenize;
+    }
+    async dispose() {
+        if (this._disposedState.disposed)
+            return;
+        this._disposedState.disposed = true;
+        await this._disposeAggregator.dispose();
+    }
+    /** @hidden */
+    async [Symbol.asyncDispose]() {
+        await this.dispose();
+    }
+    get disposed() {
+        return this._disposedState.disposed;
+    }
+    get tokens() {
+        return this._tokens;
+    }
+    get filename() {
+        return this._filename;
+    }
+    get fileInfo() {
+        return this._fileInfo;
+    }
+    get fileInsights() {
+        return this._fileInsights;
+    }
+    /**
+     * Number of layers offloaded to the GPU.
+     * If GPU support is disabled, this will always be `0`.
+     */
+    get gpuLayers() {
+        return this._gpuLayers;
+    }
+    /**
+     * Total model size in memory in bytes
+     */
+    get size() {
+        this._ensureNotDisposed();
+        return this._model.getModelSize();
+    }
+    tokenize(text, specialTokens = false, options) {
+        this._ensureNotDisposed();
+        if (text === "")
+            return [];
+        if (specialTokens === "builtin") {
+            const builtinToken = text;
+            switch (builtinToken) {
+                case "BOS": return this.tokens.bos == null ? [] : [this.tokens.bos];
+                case "EOS": return this.tokens.eos == null ? [] : [this.tokens.eos];
+                case "NL": return this.tokens.nl == null ? [] : [this.tokens.nl];
+                case "EOT": return this.tokens.infill.eot == null ? [] : [this.tokens.infill.eot];
+            }
+            void (builtinToken);
+            throw new Error(`Unknown builtin special token: ${builtinToken}`);
+        }
+        if (options === "trimLeadingSpace") {
+            if (specialTokens) {
+                const countLeadingSpaces = (text) => {
+                    let count = 0;
+                    for (; count < text.length; count++) {
+                        if (text[count] !== " ")
+                            break;
+                    }
+                    return count;
+                };
+                const textLeadingSpaces = countLeadingSpaces(text);
+                const [workaroundToken, workaroundTokenString] = (this.tokens.bos != null && this.tokens.bosString != null)
+                    ? [this.tokens.bos, this.tokens.bosString]
+                    : (this.tokens.eos != null && this.tokens.eosString != null)
+                        ? [this.tokens.eos, this.tokens.eosString]
+                        : (this.tokens.nl != null && this.tokens.nlString != null)
+                            ? [this.tokens.nl, this.tokens.nlString]
+                            : (this.tokens.infill.eot != null && this.tokens.infill.eotString != null)
+                                ? [this.tokens.infill.eot, this.tokens.infill.eotString]
+                                : [null, null];
+                if (workaroundToken != null && workaroundTokenString != null) {
+                    const tokens = Array.from(this._model.tokenize(workaroundTokenString + text, true));
+                    const workaroundTokenIndex = tokens.indexOf(workaroundToken);
+                    // only use the tokenized output if it can be corrected, otherwise fallback to the default tokenization
+                    if (workaroundTokenIndex >= 0 && workaroundTokenIndex <= 1) {
+                        tokens.splice(0, workaroundTokenIndex + 1);
+                        if (countLeadingSpaces(this.detokenize(tokens, true)) === textLeadingSpaces)
+                            return tokens;
+                    }
+                }
+                const workaroundTokensString = "\n";
+                const workaroundTokens = Array.from(this._model.tokenize(workaroundTokensString, true));
+                if (text.startsWith(workaroundTokensString)) {
+                    const tokens = Array.from(this._model.tokenize(text, true));
+                    if (this.detokenize(tokens, true).startsWith(workaroundTokensString))
+                        return tokens;
+                }
+                const tokens = Array.from(this._model.tokenize(workaroundTokensString + text, true));
+                // only use the tokenized output if it can be corrected, otherwise fallback to the default tokenization
+                if (workaroundTokens.length > 0 && workaroundTokens.every((token, index) => tokens[index] === token)) {
+                    tokens.splice(0, workaroundTokens.length);
+                    if (countLeadingSpaces(this.detokenize(tokens, true)) === textLeadingSpaces)
+                        return tokens;
+                }
+            }
+            else {
+                const workaroundTokensString = "\n";
+                const workaroundTokens = Array.from(this._model.tokenize(workaroundTokensString, false));
+                if (text.startsWith(workaroundTokensString)) {
+                    const tokens = Array.from(this._model.tokenize(text, false));
+                    if (this.detokenize(tokens, false).startsWith(workaroundTokensString))
+                        return tokens;
+                }
+                const tokens = Array.from(this._model.tokenize(workaroundTokensString + text, false));
+                // only use the tokenized output if it can be corrected, otherwise fallback to the default tokenization
+                if (workaroundTokens.length > 0 && workaroundTokens.every((token, index) => tokens[index] === token)) {
+                    tokens.splice(0, workaroundTokens.length);
+                    return tokens;
+                }
+            }
+        }
+        return Array.from(this._model.tokenize(text, specialTokens));
+    }
+    /**
+     * Transform tokens into text
+     * @param tokens - the tokens to detokenize.
+     * @param [specialTokens] - if set to `true`, special tokens will be detokenized to their corresponding token text representation.
+     * Recommended for debugging purposes only.
+     * Defaults to `false`.
+     */
+    detokenize(tokens, specialTokens = false) {
+        this._ensureNotDisposed();
+        if (tokens.length === 0)
+            return "";
+        return this._model.detokenize(Uint32Array.from(tokens), Boolean(specialTokens));
+    }
+    getTokenAttributes(token) {
+        if (this.vocabularyType === LlamaVocabularyType.none)
+            return TokenAttributes._create(token, TokenAttribute.undefined);
+        return TokenAttributes._create(token, this._model.getTokenAttributes(token));
+    }
+    /** Check whether the given token is a special token (a control-type token or a token with no normal text representation) */
+    isSpecialToken(token) {
+        if (token == null)
+            return false;
+        if (this.getTokenAttributes(token).control)
+            return true;
+        const normalText = this.detokenize([token], false);
+        if (normalText === "")
+            return this.detokenize([token], true) !== "";
+        return false;
+    }
+    /** Check whether the given token is an EOG (End Of Generation) token, like EOS or EOT. */
+    isEogToken(token) {
+        if (token == null)
+            return false;
+        return token === this.tokens.eos || token === this.tokens.infill.eot || this._model.isEogToken(token);
+    }
+    async createContext(options = {}) {
+        return await withLock(this._llama._memoryLock, LlamaLocks.loadToMemory, options.createSignal, async () => {
+            const preventDisposalHandle = this._backendModelDisposeGuard.createPreventDisposalHandle();
+            try {
+                return await LlamaContext._create(options, { _model: this });
+            }
+            finally {
+                preventDisposalHandle.dispose();
+            }
+        });
+    }
+    async createEmbeddingContext(options = {}) {
+        return await withLock(this._llama._memoryLock, LlamaLocks.loadToMemory, options.createSignal, async () => {
+            const preventDisposalHandle = this._backendModelDisposeGuard.createPreventDisposalHandle();
+            try {
+                return await LlamaEmbeddingContext._create({ _model: this }, options);
+            }
+            finally {
+                preventDisposalHandle.dispose();
+            }
+        });
+    }
+    /**
+     * Get warnings about the model file that would affect its usage.
+     *
+     * These warnings include all the warnings generated by `GgufInsights`, but are more comprehensive.
+     */
+    getWarnings() {
+        this._ensureNotDisposed();
+        const warnings = this._fileInsights.getWarnings(this._modelPath);
+        const modelFilePathText = `("${getReadablePath(this._modelPath)}")`;
+        try {
+            const specialTokenString = this.tokens.bosString || this.tokens.eosString || this.tokens.infill.eotString;
+            if (specialTokenString != null && specialTokenString !== "") {
+                const beforeTextNoSpecialTokens = "some test text here";
+                const afterTextNoSpecialTokens = this.detokenize(this.tokenize(beforeTextNoSpecialTokens, false, "trimLeadingSpace"));
+                if (beforeTextNoSpecialTokens !== afterTextNoSpecialTokens)
+                    warnings.push(`Using this model ${modelFilePathText} to tokenize text and then detokenize it resulted in a different text. ` +
+                        "There might be an issue with the model or the tokenizer implementation. " +
+                        "Using this model may not work as intended");
+                const beforeTextWithSpecialTokens = specialTokenString + beforeTextNoSpecialTokens;
+                const afterTextWithSpecialTokens = this.detokenize(this.tokenize(beforeTextWithSpecialTokens, true, "trimLeadingSpace"), true);
+                if (beforeTextWithSpecialTokens !== afterTextWithSpecialTokens)
+                    warnings.push(`Using this model ${modelFilePathText} to tokenize text with special tokens and then ` +
+                        "detokenize it resulted in a different text. " +
+                        "There might be an issue with the model or the tokenizer implementation. " +
+                        "Using this model may not work as intended");
+            }
+        }
+        catch (err) {
+            // do nothing
+        }
+        return warnings;
+    }
+    /** @hidden `ModelTypeDescription` type alias is too long in the documentation */
+    get typeDescription() {
+        this._ensureNotDisposed();
+        if (this._typeDescription == null)
+            this._typeDescription = this._model.getModelDescription();
+        return this._typeDescription;
+    }
+    /** The context size the model was trained on */
+    get trainContextSize() {
+        this._ensureNotDisposed();
+        if (this._trainContextSize == null)
+            this._trainContextSize = this._model.getTrainContextSize();
+        return this._trainContextSize;
+    }
+    /** The size of an embedding vector the model can produce */
+    get embeddingVectorSize() {
+        this._ensureNotDisposed();
+        if (this._embeddingVectorSize == null)
+            this._embeddingVectorSize = this._model.getEmbeddingVectorSize();
+        return this._embeddingVectorSize;
+    }
+    get vocabularyType() {
+        this._ensureNotDisposed();
+        if (this._vocabularyType == null) {
+            const vocabType = this._model.getVocabularyType();
+            this._vocabularyType = LlamaVocabularyTypeValues[vocabType];
+            if (this._vocabularyType == null) {
+                console.warn(getConsoleLogPrefix() + "Unknown vocabulary type:", vocabType);
+                this._vocabularyType = LlamaVocabularyType.none;
+            }
+        }
+        return this._vocabularyType;
+    }
+    /** @internal */
+    _ensureNotDisposed() {
+        if (this._disposedState.disposed)
+            throw new DisposedError();
+    }
+    /** @internal */
+    async _loadLora({ loraFilePath, baseModelPath, scale, threads }) {
+        await this._model.loadLora(path.resolve(process.cwd(), loraFilePath), scale ?? defaultLoraScale, Math.max(0, Math.floor(threads)), baseModelPath == null
+            ? undefined
+            : path.resolve(process.cwd(), baseModelPath));
+    }
+    /** @internal */
+    static async _create(modelOptions, { _llama }) {
+        const { loadSignal } = modelOptions;
+        let useMmap = modelOptions.useMmap ?? defaultUseMmap;
+        const loraOptions = typeof modelOptions.lora === "string"
+            ? { adapters: [{ loraFilePath: modelOptions.lora }] }
+            : modelOptions.lora;
+        if (loraOptions?.adapters != null && loraOptions.adapters.length > 0)
+            useMmap = false; // using LoRA with nmap crashes the process
+        const fileInfo = await readGgufFileInfo(modelOptions.modelPath, {
+            sourceType: "filesystem",
+            signal: loadSignal
+        });
+        const ggufInsights = await GgufInsights.from(fileInfo, _llama);
+        const gpuLayers = await ggufInsights.configurationResolver.resolveModelGpuLayers(modelOptions.gpuLayers, {
+            ignoreMemorySafetyChecks: modelOptions.ignoreMemorySafetyChecks
+        });
+        const vramRequiredEstimate = ggufInsights.estimateModelResourceRequirements({ gpuLayers: gpuLayers }).gpuVram;
+        const model = new LlamaModel({ ...modelOptions, gpuLayers, useMmap }, { _fileInfo: fileInfo, _fileInsights: ggufInsights, _llama });
+        const modelCreationMemoryReservation = modelOptions.ignoreMemorySafetyChecks
+            ? null
+            : _llama._vramOrchestrator.reserveMemory(vramRequiredEstimate);
+        const loggedWarnings = new Set();
+        function onAbort() {
+            model._model.abortActiveModelLoad();
+            loadSignal?.removeEventListener("abort", onAbort);
+        }
+        function logWarnings(warnings) {
+            for (const warning of warnings) {
+                if (loggedWarnings.has(warning))
+                    continue;
+                _llama._log(LlamaLogLevel.warn, warning);
+                loggedWarnings.add(warning);
+            }
+        }
+        if (loadSignal != null) {
+            if (loadSignal.aborted)
+                throw loadSignal.reason;
+            loadSignal.addEventListener("abort", onAbort);
+        }
+        logWarnings(ggufInsights.getWarnings(modelOptions.modelPath));
+        try {
+            const modelLoaded = await model._model.init();
+            if (loadSignal?.aborted) {
+                if (modelLoaded)
+                    await model._model.dispose();
+                throw loadSignal.reason;
+            }
+            else if (!modelLoaded)
+                throw new Error("Failed to load model");
+            loadSignal?.removeEventListener("abort", onAbort);
+            logWarnings(model.getWarnings());
+            if (loraOptions != null && loraOptions.adapters.length > 0) {
+                const loraThreads = loraOptions.threads ?? defaultLoraThreads;
+                let loadedAdapters = 0;
+                for (const adapter of loraOptions.adapters) {
+                    try {
+                        await model._loadLora({
+                            loraFilePath: adapter.loraFilePath,
+                            baseModelPath: adapter.baseModelPath,
+                            scale: adapter.scale,
+                            threads: loraThreads
+                        });
+                        loadedAdapters++;
+                        try {
+                            loraOptions.onLoadProgress?.(loadedAdapters / loraOptions.adapters.length);
+                        }
+                        catch (err) {
+                            console.error(err);
+                        }
+                    }
+                    catch (err) {
+                        await model._model.dispose();
+                        throw err;
+                    }
+                    if (loadSignal?.aborted) {
+                        await model._model.dispose();
+                        throw loadSignal.reason;
+                    }
+                }
+                logWarnings(model.getWarnings());
+            }
+            else if (loraOptions?.onLoadProgress != null) {
+                try {
+                    loraOptions.onLoadProgress(1);
+                }
+                catch (err) {
+                    console.error(err);
+                }
+            }
+            return model;
+        }
+        finally {
+            loadSignal?.removeEventListener("abort", onAbort);
+            modelCreationMemoryReservation?.dispose?.();
+        }
+    }
+}
+export class LlamaModelTokens {
+    /** @internal */ _model;
+    /** @internal */ _disposedState;
+    /** @internal */ _infillTokens;
+    /** @internal */ _bosToken;
+    /** @internal */ _eosToken;
+    /** @internal */ _nlToken;
+    /** @internal */ _bosString;
+    /** @internal */ _eosString;
+    /** @internal */ _nlString;
+    /** @internal */ _shouldPrependBosToken;
+    constructor(model, disposedState) {
+        this._model = model;
+        this._disposedState = disposedState;
+    }
+    /**
+     * @returns infill tokens
+     */
+    get infill() {
+        this._ensureNotDisposed();
+        if (this._infillTokens == null)
+            this._infillTokens = LlamaModelInfillTokens._create(this._model, this._disposedState);
+        return this._infillTokens;
+    }
+    /**
+     * @returns The BOS (Beginning Of Sequence) token.
+     */
+    get bos() {
+        this._ensureNotDisposed();
+        if (this._bosToken == null)
+            this._bosToken = this._model.tokenBos();
+        if (this._bosToken === -1)
+            return null;
+        return this._bosToken;
+    }
+    /**
+     * @returns The EOS (End Of Sequence) token.
+     */
+    get eos() {
+        this._ensureNotDisposed();
+        if (this._eosToken == null)
+            this._eosToken = this._model.tokenEos();
+        if (this._eosToken === -1)
+            return null;
+        return this._eosToken;
+    }
+    /**
+     * @returns The NL (New Line) token.
+     */
+    get nl() {
+        this._ensureNotDisposed();
+        if (this._nlToken == null)
+            this._nlToken = this._model.tokenNl();
+        if (this._nlToken === -1)
+            return null;
+        return this._nlToken;
+    }
+    /**
+     * @returns The BOS (Beginning Of Sequence) token as a string.
+     */
+    get bosString() {
+        this._ensureNotDisposed();
+        const bosToken = this.bos;
+        if (bosToken == null)
+            return null;
+        if (this._bosString == null)
+            this._bosString = this._model.getTokenString(bosToken);
+        return this._bosString;
+    }
+    /**
+     * @returns The EOS (End Of Sequence) token as a string.
+     */
+    get eosString() {
+        this._ensureNotDisposed();
+        const eosToken = this.eos;
+        if (eosToken == null)
+            return null;
+        if (this._eosString == null)
+            this._eosString = this._model.getTokenString(eosToken);
+        return this._eosString;
+    }
+    /**
+     * @returns The NL (New Line) token as a string.
+     */
+    get nlString() {
+        this._ensureNotDisposed();
+        const nlToken = this.nl;
+        if (nlToken == null)
+            return null;
+        if (this._nlString == null)
+            this._nlString = this._model.getTokenString(nlToken);
+        return this._nlString;
+    }
+    /**
+     * @returns Whether we should prepend a BOS (Beginning Of Sequence) token for evaluations with this model.
+     */
+    get shouldPrependBosToken() {
+        this._ensureNotDisposed();
+        if (this._shouldPrependBosToken == null)
+            this._shouldPrependBosToken = this.bos != null && this._model.shouldPrependBosToken();
+        return this._shouldPrependBosToken;
+    }
+    /** @internal */
+    _ensureNotDisposed() {
+        if (this._disposedState.disposed)
+            throw new DisposedError();
+    }
+    /** @internal */
+    static _create(model, disposedState) {
+        return new LlamaModelTokens(model, disposedState);
+    }
+}
+export class LlamaModelInfillTokens {
+    /** @internal */ _model;
+    /** @internal */ _disposedState;
+    /** @internal */ _prefixToken;
+    /** @internal */ _middleToken;
+    /** @internal */ _suffixToken;
+    /** @internal */ _eotToken;
+    /** @internal */ _prefixString;
+    /** @internal */ _middleString;
+    /** @internal */ _suffixString;
+    /** @internal */ _eotString;
+    constructor(model, disposedState) {
+        this._model = model;
+        this._disposedState = disposedState;
+    }
+    /**
+     * @returns The beginning of infill prefix token.
+     */
+    get prefix() {
+        this._ensureNotDisposed();
+        if (this._prefixToken == null)
+            this._prefixToken = this._resolveSpecialToken(this._model.prefixToken(), ["<fim_prefix>"]);
+        if (this._prefixToken === -1)
+            return null;
+        return this._prefixToken;
+    }
+    /**
+     * @returns The beginning of infill middle token.
+     */
+    get middle() {
+        this._ensureNotDisposed();
+        if (this._middleToken == null)
+            this._middleToken = this._resolveSpecialToken(this._model.middleToken(), ["<fim_middle>"]);
+        if (this._middleToken === -1)
+            return null;
+        return this._middleToken;
+    }
+    /**
+     * @returns The beginning of infill suffix token.
+     */
+    get suffix() {
+        this._ensureNotDisposed();
+        if (this._suffixToken == null)
+            this._suffixToken = this._resolveSpecialToken(this._model.suffixToken(), ["<fim_suffix>"]);
+        if (this._suffixToken === -1)
+            return null;
+        return this._suffixToken;
+    }
+    /**
+     * @returns End of infill middle token (End Of Text).
+     */
+    get eot() {
+        this._ensureNotDisposed();
+        if (this._eotToken == null)
+            this._eotToken = this._model.eotToken();
+        if (this._eotToken === -1)
+            return null;
+        return this._eotToken;
+    }
+    /**
+     * @returns The beginning of infill prefix token as a string.
+     */
+    get prefixString() {
+        this._ensureNotDisposed();
+        const prefixToken = this.prefix;
+        if (prefixToken == null)
+            return null;
+        if (this._prefixString == null)
+            this._prefixString = this._model.getTokenString(prefixToken);
+        return this._prefixString;
+    }
+    /**
+     * @returns The beginning of infill middle token as a string.
+     */
+    get middleString() {
+        this._ensureNotDisposed();
+        const middleToken = this.middle;
+        if (middleToken == null)
+            return null;
+        if (this._middleString == null)
+            this._middleString = this._model.getTokenString(middleToken);
+        return this._middleString;
+    }
+    /**
+     * @returns The beginning of infill suffix token as a string.
+     */
+    get suffixString() {
+        this._ensureNotDisposed();
+        const suffixToken = this.suffix;
+        if (suffixToken == null)
+            return null;
+        if (this._suffixString == null)
+            this._suffixString = this._model.getTokenString(suffixToken);
+        return this._suffixString;
+    }
+    /**
+     * @returns End of infill middle token (End Of Text) as a string.
+     */
+    get eotString() {
+        this._ensureNotDisposed();
+        const eotToken = this.eot;
+        if (eotToken == null)
+            return null;
+        if (this._eotString == null)
+            this._eotString = this._model.getTokenString(eotToken);
+        return this._eotString;
+    }
+    /** @internal */
+    _ensureNotDisposed() {
+        if (this._disposedState.disposed)
+            throw new DisposedError();
+    }
+    /** @internal */
+    _resolveSpecialToken(token, fallbackTexts) {
+        if (token != null && token !== -1)
+            return token;
+        for (const text of fallbackTexts) {
+            const tokens = this._model.tokenize(text, true);
+            if (tokens.length !== 1)
+                continue;
+            return tokens[0];
+        }
+        return -1;
+    }
+    /** @internal */
+    static _create(model, disposedState) {
+        return new LlamaModelInfillTokens(model, disposedState);
+    }
+}
+function disposeModelIfReferenced(modelRef) {
+    const model = modelRef.deref();
+    if (model != null)
+        void model.dispose();
+}
+//# sourceMappingURL=LlamaModel.js.map