node-llama-cpp 3.0.0-beta.14 → 3.0.0-beta.16

package/dist/evaluator/LlamaContext/LlamaContext.d.ts

@@ -1,6 +1,8 @@
 import { EventRelay } from "lifecycle-utils";
 import { Token } from "../../types.js";
 import { LlamaGrammarEvaluationState } from "../LlamaGrammarEvaluationState.js";
+import { TokenMeter } from "../TokenMeter.js";
+import { TokenBias } from "../TokenBias.js";
 import { ContextShiftOptions, ContextTokensDeleteRange, EvaluationPriority, LlamaContextSequenceRepeatPenalty } from "./types.js";
 import type { LlamaModel } from "../LlamaModel.js";
 export declare class LlamaContext {
@@ -13,6 +15,11 @@ export declare class LlamaContext {
     get model(): LlamaModel;
     get contextSize(): number;
     get batchSize(): number;
+    /**
+     * The actual size of the state in the memory in bytes.
+     * This value is provided by `llama.cpp` and doesn't include all the memory overhead of the context.
+     */
+    get stateSize(): number;
     getAllocatedContextSize(): number;
     get totalSequences(): number;
     get sequencesLeft(): number;
@@ -21,10 +28,15 @@ export declare class LlamaContext {
      * When there are no sequences left, this method will throw an error.
      * @param [options]
      */
-    getSequence({ contextShift: { size: contextShiftSize, strategy: contextShiftStrategy } }?: {
+    getSequence({ contextShift: { size: contextShiftSize, strategy: contextShiftStrategy }, _tokenMeter }?: {
         contextShift?: ContextShiftOptions;
     }): LlamaContextSequence;
     dispatchPendingBatch(): void;
+    /**
+     * Print the timings of token evaluation since that last print for this context.
+     * > **Note:** it prints on the `LlamaLogLevel.info` level, so if you set the level of your `Llama` instance higher than that,
+     * it won't print anything.
+     */
     printTimings(): Promise<void>;
 }
 export declare class LlamaContextSequence {
@@ -38,6 +50,7 @@ export declare class LlamaContextSequence {
     get model(): LlamaModel;
     get nextTokenIndex(): number;
     get contextTokens(): Token[];
+    get tokenMeter(): TokenMeter;
     get isLoadedToMemory(): boolean;
     compareContextTokens(tokens: Token[]): {
         firstDifferentIndex: number;
@@ -49,7 +62,7 @@ export declare class LlamaContextSequence {
     clearHistory(): Promise<void>;
     /**
      * Erase context tokens in the provided ranges to free up space for new tokens to be generated.
-     * the start and end of each range are exclusive.
+     * The start of each range is inclusive, and the end of each range is exclusive.
      * For example, the range `{start: 0, end: 1}` will remove the token at the `0` index only.
      */
     eraseContextTokenRanges(ranges: ContextTokensDeleteRange[]): Promise<void>;
@@ -57,13 +70,19 @@ export declare class LlamaContextSequence {
      * @param tokens
      * @param [options]
      */
-    evaluate(tokens: Token[], { temperature, minP, topK, topP, grammarEvaluationState, repeatPenalty, evaluationPriority, contextShift: { size: contextShiftSize, strategy: contextShiftStrategy }, yieldEosToken }?: {
+    evaluate(tokens: Token[], { temperature, minP, topK, topP, grammarEvaluationState, repeatPenalty, tokenBias, evaluationPriority, contextShift: { size: contextShiftSize, strategy: contextShiftStrategy }, yieldEosToken }?: {
         temperature?: number;
         minP?: number;
         topK?: number;
         topP?: number;
         grammarEvaluationState?: LlamaGrammarEvaluationState | (() => LlamaGrammarEvaluationState | undefined);
         repeatPenalty?: LlamaContextSequenceRepeatPenalty;
+        /**
+         * Adjust the probability of tokens being generated.
+         * Can be used to bias the model to generate tokens that you want it to lean towards,
+         * or to avoid generating tokens that you want it to avoid.
+         */
+        tokenBias?: TokenBias | (() => TokenBias);
         /**
          * When a lot of tokens are queued for the next batch, more than the configured `batchSize`, the tokens for each sequence will be
          * evaluated based on the strategy chosen for the context.
@@ -104,3 +123,11 @@ export declare class LlamaContextSequence {
         contextShift?: ContextShiftOptions;
     }): Promise<void>;
 }
+export declare function getDefaultContextBatchSize({ contextSize, sequences }: {
+    contextSize: number;
+    sequences: number;
+}): number;
+export declare function getDefaultContextSequences(): number;
+export declare function getDefaultModelContextSize({ trainContextSize }: {
+    trainContextSize?: number;
+}): number;

package/dist/evaluator/LlamaContext/LlamaContext.js

@@ -2,7 +2,8 @@ import { DisposeAggregator, EventRelay, withLock, DisposedError, AsyncDisposeAgg
 import { removeNullFields } from "../../utils/removeNullFields.js";
 import { compareTokens } from "../../utils/compareTokens.js";
 import { DisposeGuard } from "../../utils/DisposeGuard.js";
-import { resolveBatchItemsPrioritizingStrategy } from "./utils/resolveBatchItemsPrioritizingStrategy.js";
+import { TokenMeter } from "../TokenMeter.js";
+import { resolveBatchItemsPrioritizationStrategy } from "./utils/resolveBatchItemsPrioritizationStrategy.js";
 export class LlamaContext {
     /** @internal */ _llama;
     /** @internal */ _ctx;
@@ -25,7 +26,7 @@ export class LlamaContext {
     /** @internal */ _allocatedContextSize;
     /** @internal */ _disposed = false;
     onDispose = new EventRelay();
-    constructor({ _model }, { sequences = 1, seed = null, contextSize = _model.trainContextSize, batchSize = Math.min(contextSize * sequences, 512), threads = 6, batching: { dispatchSchedule: batchingDispatchSchedule = "nextTick", itemsPrioritizingStrategy: batchingItemsPrioritizingStrategy = "maximumParallelism" } = {}, _embeddings, _noSeed }) {
+    constructor({ _model }, { sequences, seed = null, contextSize, batchSize, threads = 6, batching: { dispatchSchedule: batchingDispatchSchedule = "nextTick", itemPrioritizationStrategy: batchingItemsPrioritizationStrategy = "maximumParallelism" } = {}, _embeddings, _noSeed }) {
         if (_model.disposed)
             throw new DisposedError();
         this._llama = _model._llama;
@@ -39,13 +40,14 @@ export class LlamaContext {
             seed: seed != null ? Math.max(-1, Math.floor(seed)) : undefined,
             contextSize: this._contextSize * this._totalSequences,
             batchSize: this._batchSize,
+            sequences: this._totalSequences,
             threads: Math.max(0, Math.floor(threads)),
             embeddings: _embeddings,
             noSeed: _noSeed
         }));
         this._batchingOptions = {
             dispatchSchedule: batchingDispatchSchedule,
-            itemsPrioritizingStrategy: batchingItemsPrioritizingStrategy
+            itemPrioritizationStrategy: batchingItemsPrioritizationStrategy
         };
         this._reclaimUnusedSequenceId = this._reclaimUnusedSequenceId.bind(this);
         this._disposeAggregator.add(() => {
@@ -82,6 +84,14 @@ export class LlamaContext {
     get batchSize() {
         return this._batchSize;
     }
+    /**
+     * The actual size of the state in the memory in bytes.
+     * This value is provided by `llama.cpp` and doesn't include all the memory overhead of the context.
+     */
+    get stateSize() {
+        this._ensureNotDisposed();
+        return this._ctx.getStateSize();
+    }
     getAllocatedContextSize() {
         this._ensureNotDisposed();
         if (this._allocatedContextSize == null)
@@ -99,7 +109,7 @@ export class LlamaContext {
      * When there are no sequences left, this method will throw an error.
      * @param [options]
      */
-    getSequence({ contextShift: { size: contextShiftSize = Math.min(100, Math.ceil(this.contextSize / 2)), strategy: contextShiftStrategy = "eraseBeginning" } = {} } = {}) {
+    getSequence({ contextShift: { size: contextShiftSize = Math.min(100, Math.ceil(this.contextSize / 2)), strategy: contextShiftStrategy = "eraseBeginning" } = {}, _tokenMeter } = {}) {
         this._ensureNotDisposed();
         const nextSequenceId = this._popSequenceId();
         if (nextSequenceId == null)
@@ -107,6 +117,7 @@ export class LlamaContext {
         return LlamaContextSequence._create({
             sequenceId: nextSequenceId,
             context: this,
+            tokenMeter: _tokenMeter,
             contextShift: {
                 size: contextShiftSize,
                 strategy: contextShiftStrategy
@@ -123,17 +134,18 @@ export class LlamaContext {
             this._currentDispatchBatchHandle = {};
             this._dispatchDecodeScheduled = false;
             this._batchDispatchPending = false;
-            let prioritizeStrategy;
-            try {
-                this._ensureNotDisposed();
-                prioritizeStrategy = resolveBatchItemsPrioritizingStrategy(this._batchingOptions.itemsPrioritizingStrategy);
-            }
-            catch (err) {
-                this._dispatchErrorForQueuedDecodesAndDequeue(new Set(this._queuedDecodes), err);
-                return;
-            }
-            let shouldHaveAnotherBatch = this._queuedDecodes.length > 0;
-            while (shouldHaveAnotherBatch) {
+            let shouldHaveAnotherLoop = this._queuedDecodes.length > 0;
+            const resolvePrioritizationStrategy = () => {
+                try {
+                    this._ensureNotDisposed();
+                    return resolveBatchItemsPrioritizationStrategy(this._batchingOptions.itemPrioritizationStrategy);
+                }
+                catch (err) {
+                    this._dispatchErrorForQueuedDecodesAndDequeue(new Set(this._queuedDecodes), err);
+                }
+                return null;
+            };
+            const getOrderedQueuedDecodes = (prioritizationStrategy) => {
                 const batchItemToQueuedDecodeMap = new Map();
                 const batchItemsList = [];
                 for (const queuedDecode of this._queuedDecodes) {
@@ -146,101 +158,132 @@ export class LlamaContext {
                 }
                 let prioritizedItems;
                 try {
-                    prioritizedItems = prioritizeStrategy({
+                    prioritizedItems = prioritizationStrategy({
                         items: batchItemsList,
                         size: this._batchSize
                     });
                 }
                 catch (err) {
                     this._dispatchErrorForQueuedDecodesAndDequeue(new Set(this._queuedDecodes), err);
-                    return;
+                    return null;
                 }
-                let batchTokenSlotsLeft = this._batchSize;
-                const afterDecodeActions = [];
-                const queuedDecodesToDelete = new Set();
-                const currentQueuedDecodeItems = new Set();
-                const currentBatchItems = [];
-                let currentBatchSize = 0;
-                for (const prioritizedItem of prioritizedItems) {
+                return prioritizedItems.map((prioritizedItem) => {
                     const queuedDecode = batchItemToQueuedDecodeMap.get(prioritizedItem.item);
                     if (queuedDecode == null)
                         throw new Error("Received invalid batch item. Make sure you keep the original object reference " +
                             "of the batch item on `item` on `PrioritizedBatchItem` in your custom prioritization strategy");
-                    const processAmount = Math.min(queuedDecode.tokens.length, prioritizedItem.processAmount, batchTokenSlotsLeft);
-                    if (processAmount <= 0)
+                    return {
+                        queuedDecode,
+                        processAmount: prioritizedItem.processAmount
+                    };
+                });
+            };
+            const fitQueuedDecodesToABatch = (queuedDecodes, batchSize) => {
+                const currentBatchItems = [];
+                let currentBatchSize = 0;
+                let batchTokenSlotsLeft = batchSize;
+                for (const { queuedDecode, processAmount } of queuedDecodes) {
+                    const resolvedProcessAmount = Math.min(processAmount <= 0 ? 1 : processAmount, queuedDecode.tokens.length, batchTokenSlotsLeft);
+                    if (resolvedProcessAmount <= 0) {
+                        if (batchTokenSlotsLeft === 0)
+                            break;
                         continue;
-                    batchTokenSlotsLeft -= processAmount;
+                    }
+                    batchTokenSlotsLeft -= resolvedProcessAmount;
+                    currentBatchSize += resolvedProcessAmount;
                     currentBatchItems.push({
                         queuedDecode,
-                        processAmount
+                        processAmount: resolvedProcessAmount
                     });
-                    currentBatchSize += processAmount;
                 }
-                let preventDisposalHandle;
+                return {
+                    currentBatchItems,
+                    currentBatchSize
+                };
+            };
+            const decodeTokenBatchItems = async (batchItems, currentBatchSize) => {
+                const afterDecodeActions = [];
+                const queuedDecodesToDelete = new Set();
+                const currentQueuedDecodeItems = new Set();
+                if (currentBatchSize !== 0)
+                    this._ctx.initBatch(currentBatchSize);
+                for (const { queuedDecode, processAmount } of batchItems) {
+                    let batchLogitIndex;
+                    try {
+                        const shouldGenerateLogitAtTheEnd = queuedDecode.generateLogitAtTheEnd &&
+                            processAmount === queuedDecode.tokens.length;
+                        const tokensToProcess = queuedDecode.tokens.slice(0, processAmount);
+                        const numberOfOutputTokens = shouldGenerateLogitAtTheEnd ? 1 : 0;
+                        TokenMeter.useTokens(queuedDecode.tokenMeter, Math.max(0, tokensToProcess.length - numberOfOutputTokens), "input");
+                        TokenMeter.useTokens(queuedDecode.tokenMeter, numberOfOutputTokens, "output");
+                        batchLogitIndex = this._ctx.addToBatch(queuedDecode.sequenceId, queuedDecode.firstTokenSequenceIndex, Uint32Array.from(tokensToProcess), shouldGenerateLogitAtTheEnd);
+                    }
+                    catch (err) {
+                        this._dispatchErrorForQueuedDecodesAndDequeue(new Set([queuedDecode]), err);
+                        continue;
+                    }
+                    currentQueuedDecodeItems.add(queuedDecode);
+                    if (queuedDecode.tokens.length === processAmount) {
+                        queuedDecodesToDelete.add(queuedDecode);
+                        afterDecodeActions.push({
+                            batchLogitIndex,
+                            response: queuedDecode.response,
+                            onDone: queuedDecode.onDone
+                        });
+                    }
+                    else {
+                        queuedDecode.tokens = queuedDecode.tokens.slice(processAmount);
+                        queuedDecode.firstTokenSequenceIndex += processAmount;
+                    }
+                }
+                for (let i = 0; i < this._queuedDecodes.length; i++) {
+                    const queuedDecode = this._queuedDecodes[i];
+                    if (queuedDecodesToDelete.has(queuedDecode)) {
+                        this._queuedDecodes.splice(i, 1);
+                        this._queuedDecodeSequenceIds.delete(queuedDecode.sequenceId);
+                        i--;
+                    }
+                }
                 try {
-                    preventDisposalHandle = this._backendContextDisposeGuard.createPreventDisposalHandle();
+                    if (currentBatchSize !== 0)
+                        await this._ctx.decodeBatch();
                 }
                 catch (err) {
-                    this._dispatchErrorForQueuedDecodesAndDequeue(new Set(this._queuedDecodes), err);
+                    this._dispatchErrorForQueuedDecodesAndDequeue(currentQueuedDecodeItems, err);
                     return;
                 }
-                try {
-                    if (currentBatchSize !== 0)
-                        this._ctx.initBatch(currentBatchSize);
-                    for (const { queuedDecode, processAmount } of currentBatchItems) {
-                        let batchLogitIndex;
+                for (const action of afterDecodeActions) {
+                    const [accept, reject] = action.response;
+                    if (action.onDone != null && action.batchLogitIndex != null) {
                         try {
-                            batchLogitIndex = this._ctx.addToBatch(queuedDecode.sequenceId, queuedDecode.firstTokenSequenceIndex, Uint32Array.from(queuedDecode.tokens.slice(0, processAmount)), queuedDecode.generateLogitAtTheEnd && processAmount === queuedDecode.tokens.length);
+                            accept(action.onDone(action.batchLogitIndex ?? null));
                         }
                         catch (err) {
-                            this._dispatchErrorForQueuedDecodesAndDequeue(new Set([queuedDecode]), err);
-                            continue;
-                        }
-                        currentQueuedDecodeItems.add(queuedDecode);
-                        if (queuedDecode.tokens.length === processAmount) {
-                            queuedDecodesToDelete.add(queuedDecode);
-                            afterDecodeActions.push({
-                                batchLogitIndex,
-                                response: queuedDecode.response,
-                                onDone: queuedDecode.onDone
-                            });
-                        }
-                        else {
-                            queuedDecode.tokens = queuedDecode.tokens.slice(processAmount);
-                            queuedDecode.firstTokenSequenceIndex += processAmount;
-                        }
-                        if (batchTokenSlotsLeft === 0)
-                            break;
-                    }
-                    for (let i = 0; i < this._queuedDecodes.length; i++) {
-                        const queuedDecode = this._queuedDecodes[i];
-                        if (queuedDecodesToDelete.has(queuedDecode)) {
-                            this._queuedDecodes.splice(i, 1);
-                            this._queuedDecodeSequenceIds.delete(queuedDecode.sequenceId);
-                            i--;
+                            reject(err);
                         }
                     }
-                    shouldHaveAnotherBatch = this._queuedDecodes.length > 0;
-                    try {
-                        if (currentBatchSize !== 0)
-                            await this._ctx.decodeBatch();
-                    }
-                    catch (err) {
-                        this._dispatchErrorForQueuedDecodesAndDequeue(currentQueuedDecodeItems, err);
-                        return;
-                    }
-                    for (const action of afterDecodeActions) {
-                        const [accept, reject] = action.response;
-                        if (action.onDone != null && action.batchLogitIndex != null) {
-                            try {
-                                accept(action.onDone(action.batchLogitIndex ?? null));
-                            }
-                            catch (err) {
-                                reject(err);
-                            }
-                        }
-                        accept(undefined);
-                    }
+                    accept(undefined);
+                }
+            };
+            const prioritizationStrategy = resolvePrioritizationStrategy();
+            if (prioritizationStrategy == null)
+                return; // all queued items are rejected and dequeued when we get here
+            while (shouldHaveAnotherLoop) {
+                const orderedQueuedDecodes = getOrderedQueuedDecodes(prioritizationStrategy);
+                if (orderedQueuedDecodes == null)
+                    return; // all queued items are rejected and dequeued when we get here
+                const { currentBatchItems, currentBatchSize } = fitQueuedDecodesToABatch(orderedQueuedDecodes, this._batchSize);
+                let preventDisposalHandle;
+                try {
+                    preventDisposalHandle = this._backendContextDisposeGuard.createPreventDisposalHandle();
+                }
+                catch (err) {
+                    this._dispatchErrorForQueuedDecodesAndDequeue(new Set(this._queuedDecodes), err);
+                    return;
+                }
+                try {
+                    await decodeTokenBatchItems(currentBatchItems, currentBatchSize);
+                    shouldHaveAnotherLoop = this._queuedDecodes.length > 0;
                 }
                 finally {
                     preventDisposalHandle.dispose();
@@ -248,13 +291,18 @@ export class LlamaContext {
             }
         });
     }
+    /**
+     * Print the timings of token evaluation since that last print for this context.
+     * > **Note:** it prints on the `LlamaLogLevel.info` level, so if you set the level of your `Llama` instance higher than that,
+     * it won't print anything.
+     */
     async printTimings() {
         this._ensureNotDisposed();
         this._ctx.printTimings();
         await new Promise((accept) => setTimeout(accept, 0)); // wait for the logs to finish printing
     }
     /** @internal */
-    async _decodeTokens({ sequenceId, firstTokenSequenceIndex, tokens, generateLogitAtTheEnd = false, evaluationPriority = 5 }, onDone) {
+    async _decodeTokens({ sequenceId, firstTokenSequenceIndex, tokens, generateLogitAtTheEnd = false, evaluationPriority = 5, tokenMeter }, onDone) {
         return await new Promise((accept, reject) => {
             this._queuedDecodes.push({
                 sequenceId,
@@ -262,6 +310,7 @@ export class LlamaContext {
                 firstTokenSequenceIndex,
                 generateLogitAtTheEnd,
                 evaluationPriority,
+                tokenMeter,
                 response: [accept, reject],
                 onDone
             });
@@ -337,17 +386,44 @@ export class LlamaContext {
     }
     /** @internal */
     static async _create(options, { _model }) {
-        const context = new LlamaContext({ _model }, options);
+        const sequences = options.sequences ?? getDefaultContextSequences();
+        const contextSize = _model.fileInsights.configurationResolver.resolveContextContextSize(options.contextSize, {
+            batchSize: options.batchSize,
+            sequences: sequences,
+            modelGpuLayers: _model.gpuLayers,
+            modelTrainContextSize: _model.trainContextSize,
+            getVramState: () => _model._llama._vramOrchestrator.getMemoryState(),
+            llamaGpu: _model._llama.gpu,
+            ignoreMemorySafetyChecks: options.ignoreMemorySafetyChecks,
+            isEmbeddingContext: options._embeddings
+        });
+        const batchSize = options.batchSize ?? getDefaultContextBatchSize({ contextSize, sequences });
+        const vramRequiredEstimate = _model.fileInsights.estimateContextResourceRequirements({
+            contextSize,
+            sequences,
+            isEmbeddingContext: options._embeddings,
+            modelGpuLayers: _model.gpuLayers,
+            batchSize
+        }).gpuVram;
+        const context = new LlamaContext({ _model }, { ...options, contextSize, batchSize, sequences });
         const { createSignal } = options;
-        const contextLoaded = await context._ctx.init();
-        if (createSignal?.aborted) {
-            if (contextLoaded)
-                await context._ctx.dispose();
-            throw createSignal.reason;
+        const contextCreationMemoryReservation = options.ignoreMemorySafetyChecks
+            ? null
+            : _model._llama._vramOrchestrator.reserveMemory(vramRequiredEstimate);
+        try {
+            const contextLoaded = await context._ctx.init();
+            if (createSignal?.aborted) {
+                if (contextLoaded)
+                    await context._ctx.dispose();
+                throw createSignal.reason;
+            }
+            else if (!contextLoaded)
+                throw new Error("Failed to create context");
+            return context;
+        }
+        finally {
+            contextCreationMemoryReservation?.dispose?.();
         }
-        else if (!contextLoaded)
-            throw new Error("Failed to create context");
-        return context;
     }
 }
 export class LlamaContextSequence {
@@ -355,14 +431,16 @@ export class LlamaContextSequence {
     /** @internal */ _gcRegistry;
     /** @internal */ _context;
     /** @internal */ _contextShift;
+    /** @internal */ _tokenMeter;
     /** @internal */ _disposeAggregator = new DisposeAggregator();
     /** @internal */ _contextTokens = [];
     /** @internal */ _nextTokenIndex = 0;
     /** @internal */ _disposed = false;
     onDispose = new EventRelay();
-    constructor({ sequenceId, context, contextShift }) {
+    constructor({ sequenceId, context, tokenMeter, contextShift }) {
         this._sequenceId = sequenceId;
         this._context = context;
+        this._tokenMeter = tokenMeter ?? new TokenMeter();
         this._contextShift = contextShift;
         this._gcRegistry = new FinalizationRegistry(this._context._reclaimUnusedSequenceId);
         this._gcRegistry.register(this, sequenceId);
@@ -399,6 +477,9 @@ export class LlamaContextSequence {
     get contextTokens() {
         return this._contextTokens.slice();
     }
+    get tokenMeter() {
+        return this._tokenMeter;
+    }
     get isLoadedToMemory() {
         return !this._disposed;
     }
@@ -424,7 +505,7 @@ export class LlamaContextSequence {
     }
     /**
      * Erase context tokens in the provided ranges to free up space for new tokens to be generated.
-     * the start and end of each range are exclusive.
+     * The start of each range is inclusive, and the end of each range is exclusive.
      * For example, the range `{start: 0, end: 1}` will remove the token at the `0` index only.
      */
     async eraseContextTokenRanges(ranges) {
@@ -486,7 +567,7 @@ export class LlamaContextSequence {
      * @param tokens
      * @param [options]
      */
-    evaluate(tokens, { temperature = 0, minP = 0, topK = 40, topP = 0.95, grammarEvaluationState, repeatPenalty, evaluationPriority = 5, contextShift: { size: contextShiftSize = this._contextShift.size, strategy: contextShiftStrategy = this._contextShift.strategy } = {}, yieldEosToken = false } = {}) {
+    evaluate(tokens, { temperature = 0, minP = 0, topK = 40, topP = 0.95, grammarEvaluationState, repeatPenalty, tokenBias, evaluationPriority = 5, contextShift: { size: contextShiftSize = this._contextShift.size, strategy: contextShiftStrategy = this._contextShift.strategy } = {}, yieldEosToken = false } = {}) {
         return this._evaluate(tokens, {
             temperature,
             minP,
@@ -494,6 +575,7 @@ export class LlamaContextSequence {
             topP,
             grammarEvaluationState,
             repeatPenalty,
+            tokenBias,
             evaluationPriority,
             contextShiftOptions: {
                 size: contextShiftSize,
@@ -522,7 +604,7 @@ export class LlamaContextSequence {
         }
     }
     /** @internal */
-    async *_evaluate(tokens, { temperature = 0, minP = 0, topK = 40, topP = 0.95, grammarEvaluationState, repeatPenalty, evaluationPriority = 5, generateNewTokens = true, contextShiftOptions, yieldEosToken = false }) {
+    async *_evaluate(tokens, { temperature = 0, minP = 0, topK = 40, topP = 0.95, grammarEvaluationState, repeatPenalty, tokenBias, evaluationPriority = 5, generateNewTokens = true, contextShiftOptions, yieldEosToken = false }) {
         this._ensureNotDisposed();
         let evalTokens = tokens;
         if (evalTokens.length === 0)
@@ -531,7 +613,7 @@ export class LlamaContextSequence {
         while (true) {
             this._ensureNotDisposed();
             // Evaluate to get the next token.
-            const nextToken = await this._decodeTokens(evalTokens, generateNewTokens, evaluationPriority, contextShiftOptions, (batchLogitIndex) => {
+            const nextToken = await this._decodeTokens(evalTokens, generateNewTokens, evaluationPriority, this._tokenMeter, contextShiftOptions, (batchLogitIndex) => {
                 const repeatPenaltyTokens = repeatPenalty?.punishTokens instanceof Function
                     ? repeatPenalty.punishTokens()
                     : repeatPenalty?.punishTokens;
@@ -540,6 +622,7 @@ export class LlamaContextSequence {
                     : grammarEvaluationState;
                 if (resolvedGrammarEvaluationState != null && resolvedGrammarEvaluationState._llama !== this.model._llama)
                     throw new Error("The LlamaGrammar used by passed to this function was created with a different Llama instance than the one used by this sequence's model. Make sure you use the same Llama instance for both the model and the grammar.");
+                const { tokenBiasKeys, tokenBiasValues } = getTokenBiasesForAddon(tokenBias, this.model);
                 return this._context._ctx.sampleToken(batchLogitIndex, removeNullFields({
                     temperature,
                     minP,
@@ -551,6 +634,8 @@ export class LlamaContextSequence {
                         : undefined,
                     repeatPenaltyPresencePenalty: repeatPenalty?.presencePenalty,
                     repeatPenaltyFrequencyPenalty: repeatPenalty?.frequencyPenalty,
+                    tokenBiasKeys,
+                    tokenBiasValues,
                     grammarEvaluationState: resolvedGrammarEvaluationState?._state
                 }));
             });
@@ -565,7 +650,7 @@ export class LlamaContextSequence {
         }
     }
     /** @internal */
-    async _decodeTokens(tokens, generateLogit, evaluationPriority, contextShiftOptions, onDecodeDone) {
+    async _decodeTokens(tokens, generateLogit, evaluationPriority, tokenMeter, contextShiftOptions, onDecodeDone) {
         this._ensureNotDisposed();
         const tokensLeftToDecode = tokens.slice();
         return await withLock(this, "evaluate", async () => {
@@ -585,7 +670,8 @@ export class LlamaContextSequence {
                     tokens: tokensToDecode,
                     firstTokenSequenceIndex: this._nextTokenIndex,
                     generateLogitAtTheEnd,
-                    evaluationPriority
+                    evaluationPriority,
+                    tokenMeter
                 }, !generateLogitAtTheEnd
                     ? undefined
                     : onDecodeDone);
@@ -632,10 +718,11 @@ export class LlamaContextSequence {
      * We need this to make it impossible to manually create instances of this class outside the code of this library
      * @internal
      */
-    static _create({ sequenceId, context, contextShift: { size: contextShiftSize = Math.min(100, Math.ceil(context.contextSize / 2)), strategy: contextShiftStrategy = "eraseBeginning" } = {} }) {
+    static _create({ sequenceId, context, tokenMeter, contextShift: { size: contextShiftSize = Math.min(100, Math.ceil(context.contextSize / 2)), strategy: contextShiftStrategy = "eraseBeginning" } = {} }) {
         return new LlamaContextSequence({
             sequenceId,
             context,
+            tokenMeter,
             contextShift: {
                 size: contextShiftSize,
                 strategy: contextShiftStrategy
@@ -643,6 +730,34 @@ export class LlamaContextSequence {
         });
     }
 }
+function getTokenBiasesForAddon(tokenBias, currentModel) {
+    if (tokenBias == null)
+        return {
+            tokenBiasKeys: undefined,
+            tokenBiasValues: undefined
+        };
+    if (tokenBias instanceof Function)
+        tokenBias = tokenBias();
+    if (tokenBias._model !== currentModel)
+        throw new Error("This TokenBias instance was created with a different model than the one used by this context. " +
+            "Make sure you use the model instance of the context sequence for the TokenBias you use it with.");
+    const tokenBiasKeys = [];
+    const tokenBiasValues = [];
+    for (const [token, bias] of tokenBias._biases) {
+        tokenBiasKeys.push(token);
+        tokenBiasValues.push(bias);
+    }
+    if (tokenBiasKeys.length === 0 || tokenBiasValues.length === 0) {
+        return {
+            tokenBiasKeys: undefined,
+            tokenBiasValues: undefined
+        };
+    }
+    return {
+        tokenBiasKeys: Uint32Array.from(tokenBiasKeys),
+        tokenBiasValues: Float32Array.from(tokenBiasValues)
+    };
+}
 function disposeContextIfReferenced(contextRef) {
     const context = contextRef.deref();
     if (context != null)
@@ -653,4 +768,14 @@ function disposeContextSequenceIfReferenced(contextRef) {
     if (context != null)
         context.dispose();
 }
+export function getDefaultContextBatchSize({ contextSize, sequences }) {
+    return Math.min(contextSize * sequences, 512);
+}
+export function getDefaultContextSequences() {
+    return 1;
+}
+const defaultFallbackContextSize = 4096;
+export function getDefaultModelContextSize({ trainContextSize }) {
+    return trainContextSize ?? defaultFallbackContextSize;
+}
 //# sourceMappingURL=LlamaContext.js.map