Disable caching for long strings. (#1676)

tokenizers/src/models/bpe/model.rs

@@ -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)
     }

tokenizers/src/models/unigram/model.rs

@@ -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)
         }
     }

tokenizers/src/utils/cache.rs

@@ -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.