Development build. This page describes main, not a released package. The latest published Lodestar.Embeddings is 0.7.0 — read its documentation.

Home › Embeddings › Loading vocabularies

BpeFilesLoader.LoadAsync

Reads a vocab.json and merges.txt pair, asynchronously.

public static Task<BpeVocabulary> LoadAsync(Stream vocabJson, Stream merges, ArtifactLoadOptions options = null, bool byteLevel = true, CancellationToken cancellationToken = default)

Parameters — vocabJson and merges are the two streams, never disposed here; vocabJsonPath and mergesPath are the two files. options bounds what will be accepted and defaults to ArtifactLoadOptions's own defaults. byteLevel says whether the model is a byte-level BPE, which GPT-2 and its descendants are. cancellationToken cancels the read.

Returns — Task<BpeVocabulary>, completing with the loaded vocabulary.

Exceptions — ArgumentNullException for a null source or path. InvalidDataException when the content is not the format expected, declares a model this loader does not read, or exceeds a bound in options — the message names both the limit and the value. OperationCanceledException when cancellationToken is signalled.

Example — both streams, awaited together.

using Lodestar.Embeddings.Persistence;
using Lodestar.Embeddings.Tokenization;

static async Task<BpeVocabulary> LoadAsync() =>
    await BpeFilesLoader.LoadAsync(File.OpenRead("vocab.json"), File.OpenRead("merges.txt"));

BpeVocabulary vocab = LoadAsync().GetAwaiter().GetResult();

Remarks — Same pair, same ordering rules and same refusals as Load; only the I/O is asynchronous. The two streams are both read and neither is disposed, so the caller closes them.

The same-checkpoint requirement from Load applies unchanged, and is no easier to notice here.

Applies to — net10.0, netstandard2.0.

See also — BpeFilesLoader.Load, BpeFilesLoader, the persistence index.