node-llama-cpp 3.0.0-beta.13 → 3.0.0-beta.15

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

@@ -1,21 +1,26 @@
 import { EventRelay } from "lifecycle-utils";
 import { Token } from "../../types.js";
-import { LlamaModel } from "../LlamaModel.js";
 import { LlamaGrammarEvaluationState } from "../LlamaGrammarEvaluationState.js";
+import { GgufInsights } from "../../gguf/GgufInsights.js";
+import { TokenMeter } from "../TokenMeter.js";
+import { BuildGpu } from "../../bindings/types.js";
 import { ContextShiftOptions, ContextTokensDeleteRange, EvaluationPriority, LlamaContextOptions, LlamaContextSequenceRepeatPenalty } from "./types.js";
+import type { LlamaModel } from "../LlamaModel.js";
 export declare class LlamaContext {
     readonly onDispose: EventRelay<void>;
-    /**
-     * @param options
-     */
-    constructor({ model, sequences, seed, contextSize, batchSize, threads, batching: { dispatchSchedule: batchingDispatchSchedule, itemsPrioritizingStrategy: batchingItemsPrioritizingStrategy }, _embedding, _noSeed }: LlamaContextOptions);
-    dispose(): void;
+    private constructor();
+    dispose(): Promise<void>;
     /** @hidden */
-    [Symbol.dispose](): void;
+    [Symbol.asyncDispose](): Promise<void>;
     get disposed(): boolean;
     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;
@@ -24,10 +29,15 @@ 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 {
@@ -41,6 +51,7 @@ export declare class LlamaContextSequence {
     get model(): LlamaModel;
     get nextTokenIndex(): number;
     get contextTokens(): Token[];
+    get tokenMeter(): TokenMeter;
     get isLoadedToMemory(): boolean;
     compareContextTokens(tokens: Token[]): {
         firstDifferentIndex: number;
@@ -52,7 +63,7 @@ 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>;
@@ -107,3 +118,26 @@ export declare class LlamaContextSequence {
         contextShift?: ContextShiftOptions;
     }): Promise<void>;
 }
+export declare function resolveContextContextSizeOption({ contextSize, batchSize, sequences, modelFileInsights, modelGpuLayers, modelTrainContextSize, getVramState, llamaGpu, ignoreMemorySafetyChecks, isEmbeddingContext }: {
+    contextSize?: LlamaContextOptions["contextSize"];
+    batchSize?: LlamaContextOptions["batchSize"];
+    sequences: number;
+    modelFileInsights: GgufInsights;
+    modelGpuLayers: number;
+    modelTrainContextSize: number;
+    getVramState(): {
+        total: number;
+        free: number;
+    };
+    llamaGpu: BuildGpu;
+    ignoreMemorySafetyChecks?: boolean;
+    isEmbeddingContext?: boolean;
+}): number;
+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

@@ -1,11 +1,15 @@
-import { DisposeAggregator, EventRelay, withLock, DisposedError } from "lifecycle-utils";
+import { DisposeAggregator, EventRelay, withLock, DisposedError, AsyncDisposeAggregator } from "lifecycle-utils";
 import { removeNullFields } from "../../utils/removeNullFields.js";
 import { compareTokens } from "../../utils/compareTokens.js";
-import { resolveBatchItemsPrioritizingStrategy } from "./utils/resolveBatchItemsPrioritizingStrategy.js";
+import { DisposeGuard } from "../../utils/DisposeGuard.js";
+import { minAllowedContextSizeInCalculations } from "../../config.js";
+import { TokenMeter } from "../TokenMeter.js";
+import { resolveBatchItemsPrioritizationStrategy } from "./utils/resolveBatchItemsPrioritizationStrategy.js";
 export class LlamaContext {
     /** @internal */ _llama;
     /** @internal */ _ctx;
     /** @internal */ _onReclaimUnusedSequenceId = new EventRelay();
+    /** @internal */ _backendContextDisposeGuard;
     /** @internal */ _model;
     /** @internal */ _contextSize;
     /** @internal */ _batchSize;
@@ -14,7 +18,8 @@ export class LlamaContext {
     /** @internal */ _batchingOptions;
     /** @internal */ _queuedDecodeSequenceIds = new Set();
     /** @internal */ _queuedDecodes = [];
-    /** @internal */ _disposeAggregator = new DisposeAggregator();
+    /** @internal */ _disposeAggregator = new AsyncDisposeAggregator();
+    /** @internal */ _modelPreventDisposalHandle;
     /** @internal */ _nextGeneratedSequenceId = 0;
     /** @internal */ _dispatchDecodeScheduled = false;
     /** @internal */ _batchDispatchPending = false;
@@ -22,14 +27,13 @@ export class LlamaContext {
     /** @internal */ _allocatedContextSize;
     /** @internal */ _disposed = false;
     onDispose = new EventRelay();
-    /**
-     * @param options
-     */
-    constructor({ model, sequences = 1, seed = null, contextSize = model.trainContextSize, batchSize = contextSize, threads = 6, batching: { dispatchSchedule: batchingDispatchSchedule = "nextTick", itemsPrioritizingStrategy: batchingItemsPrioritizingStrategy = "maximumParallelism" } = {}, _embedding, _noSeed }) {
-        if (model.disposed)
+    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;
-        this._model = model;
+        this._llama = _model._llama;
+        this._model = _model;
+        this._backendContextDisposeGuard = new DisposeGuard([this._model._backendModelDisposeGuard]);
+        this._modelPreventDisposalHandle = this._model._backendModelDisposeGuard.createPreventDisposalHandle();
         this._totalSequences = Math.max(1, Math.floor(sequences));
         this._contextSize = Math.max(2, contextSize);
         this._batchSize = Math.max(batchSize, this._totalSequences);
@@ -37,30 +41,36 @@ 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)),
-            embedding: _embedding,
+            embeddings: _embeddings,
             noSeed: _noSeed
         }));
         this._batchingOptions = {
             dispatchSchedule: batchingDispatchSchedule,
-            itemsPrioritizingStrategy: batchingItemsPrioritizingStrategy
+            itemPrioritizationStrategy: batchingItemsPrioritizationStrategy
         };
         this._reclaimUnusedSequenceId = this._reclaimUnusedSequenceId.bind(this);
-        this._disposeAggregator.add(this._onReclaimUnusedSequenceId);
-        this._disposeAggregator.add(this.onDispose.dispatchEvent);
         this._disposeAggregator.add(() => {
-            this._ctx.dispose();
+            this._disposed = true;
         });
+        this._disposeAggregator.add(this._onReclaimUnusedSequenceId);
+        this._disposeAggregator.add(this.onDispose.dispatchEvent);
         this._disposeAggregator.add(this.model.onDispose.createListener(disposeContextIfReferenced.bind(null, new WeakRef(this))));
+        this._disposeAggregator.add(async () => {
+            await this._backendContextDisposeGuard.acquireDisposeLock();
+            await this._ctx.dispose();
+            this._modelPreventDisposalHandle.dispose();
+        });
     }
-    dispose() {
+    async dispose() {
         if (this._disposed)
             return;
         this._disposed = true;
-        this._disposeAggregator.dispose();
+        await this._disposeAggregator.dispose();
     }
     /** @hidden */
-    [Symbol.dispose]() {
+    [Symbol.asyncDispose]() {
         return this.dispose();
     }
     get disposed() {
@@ -75,6 +85,14 @@ 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)
@@ -92,7 +110,7 @@ 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)
@@ -100,6 +118,7 @@ export class LlamaContext {
         return LlamaContextSequence._create({
             sequenceId: nextSequenceId,
             context: this,
+            tokenMeter: _tokenMeter,
             contextShift: {
                 size: contextShiftSize,
                 strategy: contextShiftStrategy
@@ -116,17 +135,18 @@ 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) {
@@ -139,42 +159,65 @@ 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;
                 }
+                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 currentBatchItems) {
+                for (const { queuedDecode, processAmount } of batchItems) {
                     let batchLogitIndex;
                     try {
-                        batchLogitIndex = this._ctx.addToBatch(queuedDecode.sequenceId, queuedDecode.firstTokenSequenceIndex, Uint32Array.from(queuedDecode.tokens.slice(0, processAmount)), queuedDecode.generateLogitAtTheEnd && processAmount === queuedDecode.tokens.length);
+                        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);
@@ -193,8 +236,6 @@ export class LlamaContext {
                         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];
@@ -204,7 +245,6 @@ export class LlamaContext {
                         i--;
                     }
                 }
-                shouldHaveAnotherBatch = this._queuedDecodes.length > 0;
                 try {
                     if (currentBatchSize !== 0)
                         await this._ctx.decodeBatch();
@@ -225,15 +265,45 @@ export class LlamaContext {
                     }
                     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();
+                }
             }
         });
     }
+    /**
+     * 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,
@@ -241,6 +311,7 @@ export class LlamaContext {
                 firstTokenSequenceIndex,
                 generateLogitAtTheEnd,
                 evaluationPriority,
+                tokenMeter,
                 response: [accept, reject],
                 onDone
             });
@@ -253,6 +324,8 @@ export class LlamaContext {
         if (this._disposed)
             return;
         void withLock(this, "context", async () => {
+            if (this._disposed)
+                return;
             this._ctx.disposeSequence(sequenceId);
             this._unusedSequenceIds.push(sequenceId);
             this._onReclaimUnusedSequenceId.dispatchEvent();
@@ -312,20 +385,65 @@ export class LlamaContext {
         if (this._disposed)
             throw new DisposedError();
     }
+    /** @internal */
+    static async _create(options, { _model }) {
+        const sequences = options.sequences ?? getDefaultContextSequences();
+        const contextSize = resolveContextContextSizeOption({
+            contextSize: options.contextSize,
+            batchSize: options.batchSize,
+            sequences: sequences,
+            modelFileInsights: _model.fileInsights,
+            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 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?.();
+        }
+    }
 }
 export class LlamaContextSequence {
     /** @internal */ _sequenceId;
     /** @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);
@@ -362,6 +480,9 @@ export class LlamaContextSequence {
     get contextTokens() {
         return this._contextTokens.slice();
     }
+    get tokenMeter() {
+        return this._tokenMeter;
+    }
     get isLoadedToMemory() {
         return !this._disposed;
     }
@@ -387,7 +508,7 @@ 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) {
@@ -396,6 +517,8 @@ export class LlamaContextSequence {
             this._ensureNotDisposed();
             if (ranges.length === 0)
                 return;
+            // if the deletion fails, we'll have to dispose the sequence and fill it up again
+            let deletionSuccessful = true;
             const resolvedRanges = ranges
                 .map(({ start, end }) => {
                 if (start === end)
@@ -425,15 +548,22 @@ export class LlamaContextSequence {
             let lastDeleteRangeEndPos = null;
             for (const range of resolvedRanges) {
                 this._contextTokens.splice(range.start - removedTokens, range.end - range.start);
-                this._context._ctx.removeTokenCellsFromSequence(this._sequenceId, range.start, range.end);
-                if (lastDeleteRangeEndPos != null && removedTokens > 0 && lastDeleteRangeEndPos !== range.start)
+                if (deletionSuccessful)
+                    deletionSuccessful &&= this._context._ctx.removeTokenCellsFromSequence(this._sequenceId, range.start, range.end);
+                if (deletionSuccessful && lastDeleteRangeEndPos != null && removedTokens > 0 && lastDeleteRangeEndPos !== range.start)
                     this._context._ctx.shiftSequenceTokenCells(this._sequenceId, lastDeleteRangeEndPos, range.start, -removedTokens);
                 removedTokens += range.end - range.start;
                 lastDeleteRangeEndPos = range.end;
             }
-            if (lastDeleteRangeEndPos != null && removedTokens > 0 && lastDeleteRangeEndPos !== this._nextTokenIndex)
+            if (deletionSuccessful && lastDeleteRangeEndPos != null && removedTokens > 0 && lastDeleteRangeEndPos !== this._nextTokenIndex)
                 this._context._ctx.shiftSequenceTokenCells(this._sequenceId, lastDeleteRangeEndPos, this._nextTokenIndex, -removedTokens);
             this._nextTokenIndex -= removedTokens;
+            if (deletionSuccessful)
+                return;
+            const newSequenceTokens = this._contextTokens.slice();
+            this._nextTokenIndex = 0;
+            this._context._ctx.disposeSequence(this._sequenceId);
+            await this.evaluateWithoutGeneratingNewTokens(newSequenceTokens);
         });
     }
     /**
@@ -485,7 +615,7 @@ 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;
@@ -519,7 +649,7 @@ 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 () => {
@@ -539,7 +669,8 @@ export class LlamaContextSequence {
                     tokens: tokensToDecode,
                     firstTokenSequenceIndex: this._nextTokenIndex,
                     generateLogitAtTheEnd,
-                    evaluationPriority
+                    evaluationPriority,
+                    tokenMeter
                 }, !generateLogitAtTheEnd
                     ? undefined
                     : onDecodeDone);
@@ -586,10 +717,11 @@ 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
@@ -600,11 +732,70 @@ export class LlamaContextSequence {
 function disposeContextIfReferenced(contextRef) {
     const context = contextRef.deref();
     if (context != null)
-        context.dispose();
+        void context.dispose();
 }
 function disposeContextSequenceIfReferenced(contextRef) {
     const context = contextRef.deref();
     if (context != null)
         context.dispose();
 }
+export function resolveContextContextSizeOption({ contextSize, batchSize, sequences, modelFileInsights, modelGpuLayers, modelTrainContextSize, getVramState, llamaGpu, ignoreMemorySafetyChecks = false, isEmbeddingContext = false }) {
+    if (contextSize == null)
+        contextSize = "auto";
+    if (typeof contextSize === "number") {
+        const resolvedContextSize = Math.max(1, Math.floor(contextSize));
+        if (ignoreMemorySafetyChecks)
+            return resolvedContextSize;
+        const vramState = getVramState();
+        const contextVram = modelFileInsights.estimateContextResourceRequirements({
+            contextSize: resolvedContextSize,
+            batchSize: batchSize ?? getDefaultContextBatchSize({ contextSize: resolvedContextSize, sequences }),
+            modelGpuLayers: modelGpuLayers,
+            sequences,
+            isEmbeddingContext
+        }).gpuVram;
+        if (contextVram > vramState.free)
+            throw new Error(`The context size of ${resolvedContextSize}${sequences > 1 ? ` with ${sequences} sequences` : ""} is too large for the available VRAM`);
+        return resolvedContextSize;
+    }
+    else if (contextSize === "auto" || typeof contextSize === "object") {
+        if (llamaGpu === false)
+            return modelTrainContextSize;
+        const vramState = getVramState();
+        if (vramState.total === 0)
+            return modelTrainContextSize;
+        const freeVram = vramState.free;
+        const maxContextSize = contextSize === "auto"
+            ? getDefaultModelContextSize({ trainContextSize: modelTrainContextSize })
+            : Math.min(contextSize.max ?? getDefaultModelContextSize({ trainContextSize: modelTrainContextSize }), getDefaultModelContextSize({ trainContextSize: modelTrainContextSize }));
+        const minContextSize = contextSize === "auto"
+            ? minAllowedContextSizeInCalculations
+            : Math.max(contextSize.min ?? minAllowedContextSizeInCalculations, minAllowedContextSizeInCalculations);
+        for (let testContextSize = maxContextSize; testContextSize >= minContextSize; testContextSize--) {
+            const contextVram = modelFileInsights.estimateContextResourceRequirements({
+                contextSize: testContextSize,
+                batchSize: batchSize ?? getDefaultContextBatchSize({ contextSize: testContextSize, sequences }),
+                modelGpuLayers: modelGpuLayers,
+                sequences,
+                isEmbeddingContext
+            }).gpuVram;
+            if (contextVram <= freeVram)
+                return testContextSize;
+        }
+        if (ignoreMemorySafetyChecks)
+            return minContextSize;
+        throw new Error(`The available VRAM is too small to fit the context size of ${maxContextSize}${sequences > 1 ? ` with ${sequences} sequences` : ""}`);
+    }
+    throw new Error(`Invalid context size: "${contextSize}"`);
+}
+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