mii/tokenizers
1
0
Fork 0
mirror of https://github.com/mii443/tokenizers.git synced 2025-08-22 08:15:49 +00:00
Code Issues Packages Projects Releases Wiki Activity

Disable caching for long strings. (#1676)

Browse Source
This commit is contained in:
Nicolas Patry
2024-11-07 21:36:27 +08:00
committed by GitHub
parent c6b5c3eab7
commit 5aa9f6cff0
3 changed files with 11 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
tokenizers/src/models/bpe/model.rs
View File

@ -1,6 +1,6 @@
use super::{super::OrderedVocabIter, trainer::BpeTrainer, Error, Pair, Word};
use crate::tokenizer::{Model, Result, Token};
use crate::utils::cache::{Cache, DEFAULT_CACHE_CAPACITY};
use crate::utils::cache::{Cache, DEFAULT_CACHE_CAPACITY, MAX_LENGTH};
use crate::utils::iter::ResultShunt;
use serde_json::Value;
use std::borrow::Cow;
@ -482,7 +482,9 @@ impl BPE {
let word = self.merge_word(sequence)?;
let ret = self.word_to_tokens(&word).collect();
if let Some(ref cache) = self.cache {
cache.set(sequence.to_owned(), word);
if sequence.len() < MAX_LENGTH {
cache.set(sequence.to_owned(), word);
}
}
Ok(ret)
}

6
tokenizers/src/models/unigram/model.rs
View File

@ -4,7 +4,7 @@ use super::{
trie::{Trie, TrieBuilder},
};
use crate::tokenizer::{Model, Result, Token};
use crate::utils::cache::Cache;
use crate::utils::cache::{Cache, MAX_LENGTH};
use std::collections::HashMap;
use std::convert::TryInto;
@ -230,7 +230,9 @@ impl Unigram {
} else {
self.encode_unoptimized(sentence)?
};
self.cache.set(sentence.to_owned(), result.clone());
if sentence.len() < MAX_LENGTH {
self.cache.set(sentence.to_owned(), result.clone());
}
Ok(result)
}
}

3
tokenizers/src/utils/cache.rs
View File

@ -5,6 +5,9 @@ use std::sync::RwLock;
/// The default capacity for a `BPE`'s internal cache.
pub static DEFAULT_CACHE_CAPACITY: usize = 10_000;
/// The maximum length we should cache in a model
/// Strings that are too long have minimal chances to cache hit anyway
pub static MAX_LENGTH: usize = 256;
/// Provides a simple multithread cache to speed up BPE tokenization that will try to read values
/// concurrently but won't block if another thread is writing.