diff --git a/Cargo.toml b/Cargo.toml
index 84e8226..c00e7ad 100644
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -25,7 +25,7 @@ either = "1.13.0"
 macro_rules_attribute = "0.2.0"
 once_cell = "1.19.0"
 paste = "1.0.14"
-pyo3 = { version = "^0.27", features = ["extension-module"] }
+pyo3 = { version = "^0.28", features = ["extension-module"] }
 regex = "1.10.3"
 serde = "1.0.197"
 serde_json = "1.0.114"
diff --git a/src/tokenizer.rs b/src/tokenizer.rs
index 41c61f8..a6c53c5 100644
--- a/src/tokenizer.rs
+++ b/src/tokenizer.rs
@@ -131,7 +131,7 @@ impl SmirkTokenizer {
             .map(|x| EncodeInput::from(x.to_string()))
             .collect();
         // Release the GIL while tokenizing batch
-        let out = py.allow_threads(|| {
+        let out = py.detach(|| {
             self.tokenizer
                 .encode_batch_char_offsets(inputs, add_special_tokens)
                 .unwrap()
@@ -149,7 +149,7 @@ impl SmirkTokenizer {
         ids: Vec<Vec<u32>>,
         skip_special_tokens: bool,
     ) -> PyResult<Vec<String>> {
-        py.allow_threads(|| {
+        py.detach(|| {
             let sequences = ids.iter().map(|x| &x[..]).collect::<Vec<&[u32]>>();
             Ok(self
                 .tokenizer
@@ -434,7 +434,7 @@ impl SmirkTokenizer {
 
         // Train tokenizer
         let mut trainer: TrainerWrapper = builder.build().unwrap().into();
-        let _ = py.allow_threads(|| tokenizer.train_from_files(&mut trainer, files).unwrap());
+        let _ = py.detach(|| tokenizer.train_from_files(&mut trainer, files).unwrap());
         Ok(SmirkTokenizer::new(tokenizer))
     }
 }