fix race conditions and add offline tokenizer loading api

Cargo.lock

@@ -572,6 +572,16 @@ dependencies = [
  "percent-encoding",
 ]
 
+[[package]]
+name = "fs2"
+version = "0.4.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "9564fc758e15025b46aa6643b1b77d047d1a56a1aea6e01002ac0c7026876213"
+dependencies = [
+ "libc",
+ "winapi",
+]
+
 [[package]]
 name = "futures"
 version = "0.3.31"
@@ -1324,6 +1334,7 @@ dependencies = [
  "bstr",
  "clap",
  "fancy-regex",
+ "fs2",
  "futures",
  "image",
  "pretty_assertions",
@@ -2556,6 +2567,28 @@ version = "0.1.8"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "53a85b86a771b1c87058196170769dd264f66c0782acf1ae6cc51bfd64b39082"
 
+[[package]]
+name = "winapi"
+version = "0.3.9"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "5c839a674fcd7a98952e593242ea400abe93992746761e38641405d28b00f419"
+dependencies = [
+ "winapi-i686-pc-windows-gnu",
+ "winapi-x86_64-pc-windows-gnu",
+]
+
+[[package]]
+name = "winapi-i686-pc-windows-gnu"
+version = "0.4.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ac3b87c63620426dd9b991e5ce0329eff545bccbbb34f3be09ff6fb6ab51b7b6"
+
+[[package]]
+name = "winapi-x86_64-pc-windows-gnu"
+version = "0.4.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "712e227841d057c1ee1cd2fb22fa7e5a5461ae8e48fa2ca79ec42cfc1931183f"
+
 [[package]]
 name = "windows-core"
 version = "0.61.0"

Cargo.toml

@@ -18,6 +18,7 @@ wasm-binding = ["wasm-bindgen", "serde-wasm-bindgen", "wasm-bindgen-futures"]
 [dependencies]
 anyhow = "1.0.98"
 base64 = "0.22.1"
+fs2 = "0.4.3"
 image = "0.25.6"
 serde = { version = "1.0.219", features = ["derive"] }
 serde_json = { version = "1.0.140", features = ["preserve_order"] }

python/openai_harmony/__init__.py

@@ -36,13 +36,10 @@ from pydantic import BaseModel, Field
 try:
     from .openai_harmony import (
         HarmonyError as HarmonyError,  # expose the actual Rust error directly
-    )
+        PyHarmonyEncoding as _PyHarmonyEncoding,  # type: ignore
-    from .openai_harmony import PyHarmonyEncoding as _PyHarmonyEncoding  # type: ignore
-    from .openai_harmony import (
         PyStreamableParser as _PyStreamableParser,  # type: ignore
-    )
-    from .openai_harmony import (
         load_harmony_encoding as _load_harmony_encoding,  # type: ignore
+        load_harmony_encoding_from_file as _load_harmony_encoding_from_file,  # type: ignore
     )
 
 except ModuleNotFoundError:  # pragma: no cover – raised during type-checking
@@ -690,6 +687,32 @@ def load_harmony_encoding(name: str | "HarmonyEncodingName") -> HarmonyEncoding:
     return HarmonyEncoding(inner)
 
 
+def load_harmony_encoding_from_file(
+    name: str,
+    vocab_file: str,
+    special_tokens: list[tuple[str, int]],
+    pattern: str,
+    n_ctx: int,
+    max_message_tokens: int,
+    max_action_length: int,
+    expected_hash: str | None = None,
+) -> HarmonyEncoding:
+    """Load a HarmonyEncoding from a local vocab file (offline usage).
+    Use this when network access is restricted or for reproducible builds where you want to avoid remote downloads.
+    """
+    inner: _PyHarmonyEncoding = _load_harmony_encoding_from_file(
+        name,
+        vocab_file,
+        special_tokens,
+        pattern,
+        n_ctx,
+        max_message_tokens,
+        max_action_length,
+        expected_hash,
+    )
+    return HarmonyEncoding(inner)
+
+
 # For *mypy* we expose a minimal stub of the `HarmonyEncodingName` enum.  At
 # **runtime** the user is expected to pass the *string* names because the Rust
 # side only operates on strings anyway.
@@ -718,4 +741,5 @@ __all__ = [
     "StreamableParser",
     "StreamState",
     "HarmonyError",
+    "load_harmony_encoding_from_file",
 ]

src/encoding.rs

@@ -154,6 +154,31 @@ impl HarmonyEncoding {
             })
             .collect()
     }
+
+    pub fn from_local_file(
+        name: String,
+        vocab_file: &std::path::Path,
+        expected_hash: Option<&str>,
+        special_tokens: impl IntoIterator<Item = (String, u32)>,
+        pattern: &str,
+        n_ctx: usize,
+        max_message_tokens: usize,
+        max_action_length: usize,
+    ) -> anyhow::Result<Self> {
+        use crate::tiktoken_ext::public_encodings::load_encoding_from_file;
+        let bpe = load_encoding_from_file(vocab_file, expected_hash, special_tokens, pattern)?;
+        Ok(HarmonyEncoding {
+            name,
+            n_ctx,
+            max_message_tokens,
+            max_action_length,
+            tokenizer_name: vocab_file.display().to_string(),
+            tokenizer: std::sync::Arc::new(bpe),
+            format_token_mapping: Default::default(),
+            stop_formatting_tokens: Default::default(),
+            stop_formatting_tokens_for_assistant_actions: Default::default(),
+        })
+    }
 }
 
 // Methods for rendering conversations

src/py_module.rs

@@ -396,6 +396,38 @@ fn openai_harmony(_py: Python<'_>, m: &Bound<'_, PyModule>) -> PyResult<()> {
     }
     m.add_function(pyo3::wrap_pyfunction!(load_harmony_encoding_py, m)?)?;
 
+    // Convenience function to load a HarmonyEncoding from a local vocab file for offline
+    // scenarios or reproducible builds where remote download is not possible.
+    #[pyfunction(name = "load_harmony_encoding_from_file")]
+    fn load_harmony_encoding_from_file_py(
+        py: Python<'_>,
+        name: &str,
+        vocab_file: &str,
+        special_tokens: Vec<(String, u32)>,
+        pattern: &str,
+        n_ctx: usize,
+        max_message_tokens: usize,
+        max_action_length: usize,
+        expected_hash: Option<&str>,
+    ) -> PyResult<Py<PyHarmonyEncoding>> {
+        let encoding = HarmonyEncoding::from_local_file(
+            name.to_string(),
+            std::path::Path::new(vocab_file),
+            expected_hash,
+            special_tokens,
+            pattern,
+            n_ctx,
+            max_message_tokens,
+            max_action_length,
+        )
+        .map_err(|e| PyErr::new::<HarmonyError, _>(e.to_string()))?;
+        Py::new(py, PyHarmonyEncoding { inner: encoding })
+    }
+    m.add_function(pyo3::wrap_pyfunction!(
+        load_harmony_encoding_from_file_py,
+        m
+    )?)?;
+
     // Convenience functions to get the tool configs for the browser and python tools.
     #[pyfunction]
     fn get_tool_namespace_config(py: Python<'_>, tool: &str) -> PyResult<PyObject> {

src/tiktoken_ext/mod.rs

@@ -1,2 +1,2 @@
-mod public_encodings;
+pub mod public_encodings;
 pub use public_encodings::{set_tiktoken_base_url, Encoding};

src/tiktoken_ext/public_encodings.rs

@@ -9,6 +9,7 @@ use std::{
 use base64::{prelude::BASE64_STANDARD, Engine as _};
 
 use crate::tiktoken::{CoreBPE, Rank};
+use fs2::FileExt;
 use sha1::Sha1;
 use sha2::{Digest as _, Sha256};
 
@@ -420,24 +421,35 @@ fn download_or_find_cached_file(
 ) -> Result<PathBuf, RemoteVocabFileError> {
     let cache_dir = resolve_cache_dir()?;
     let cache_path = resolve_cache_path(&cache_dir, url);
-    if cache_path.exists() {
+    let lock_path = cache_path.with_extension("lock");
-        if verify_file_hash(&cache_path, expected_hash)? {
+    let lock_file = File::create(&lock_path).map_err(|e| {
-            return Ok(cache_path);
+        RemoteVocabFileError::IOError(format!("creating lock file {lock_path:?}"), e)
-        }
+    })?;
-        let _ = std::fs::remove_file(&cache_path);
+    lock_file
-    }
+        .lock_exclusive()
-    let hash = load_remote_file(url, &cache_path)?;
+        .map_err(|e| RemoteVocabFileError::IOError(format!("locking file {lock_path:?}"), e))?;
-    if let Some(expected_hash) = expected_hash {
+    let result = (|| {
-        if hash != expected_hash {
+        if cache_path.exists() {
+            if verify_file_hash(&cache_path, expected_hash)? {
+                return Ok(cache_path);
+            }
             let _ = std::fs::remove_file(&cache_path);
-            return Err(RemoteVocabFileError::HashMismatch {
-                file_url: url.to_string(),
-                expected_hash: expected_hash.to_string(),
-                computed_hash: hash,
-            });
         }
-    }
+        let hash = load_remote_file(url, &cache_path)?;
-    Ok(cache_path)
+        if let Some(expected_hash) = expected_hash {
+            if hash != expected_hash {
+                let _ = std::fs::remove_file(&cache_path);
+                return Err(RemoteVocabFileError::HashMismatch {
+                    file_url: url.to_string(),
+                    expected_hash: expected_hash.to_string(),
+                    computed_hash: hash,
+                });
+            }
+        }
+        Ok(cache_path)
+    })();
+    let _ = fs2::FileExt::unlock(&lock_file);
+    result
 }
 
 #[cfg(target_arch = "wasm32")]
@@ -572,4 +584,25 @@ mod tests {
             let _ = encoding.load().unwrap();
         }
     }
+
+    #[test]
+    fn test_parallel_load_encodings() {
+        use std::thread;
+
+        let encodings = Encoding::all();
+        for encoding in encodings {
+            let name = encoding.name();
+            let handles: Vec<_> = (0..8)
+                .map(|_| {
+                    let name = name.to_string();
+                    thread::spawn(move || {
+                        Encoding::from_name(&name).unwrap().load().unwrap();
+                    })
+                })
+                .collect();
+            for handle in handles {
+                handle.join().expect("Thread panicked");
+            }
+        }
+    }
 }

tests/test_harmony.py

@@ -35,6 +35,7 @@ from openai_harmony import (  # noqa: E402
     SystemContent,
     ToolDescription,
     load_harmony_encoding,
+    load_harmony_encoding_from_file,
 )
 from pydantic import ValidationError
 
@@ -949,3 +950,67 @@ def test_streamable_parser_tool_call_with_constrain_adjacent():
     ]
 
     assert parser.messages == expected
+
+
+def test_load_harmony_encoding_from_file(tmp_path):
+    import os
+    from openai_harmony import load_harmony_encoding_from_file
+
+    cache_dir = os.environ.get("TIKTOKEN_RS_CACHE_DIR")
+    if not cache_dir:
+        cache_dir = os.path.join(os.path.expanduser("~"), ".cache", "tiktoken-rs-cache")
+    import hashlib
+    url = "https://openaipublic.blob.core.windows.net/encodings/o200k_base.tiktoken"
+    cache_key = hashlib.sha1(url.encode()).hexdigest()
+    vocab_file = os.path.join(cache_dir, cache_key)
+    if not os.path.exists(vocab_file):
+        import pytest
+        pytest.skip("No local vocab file available for offline test")
+
+    special_tokens = [
+        ("<|startoftext|>", 199998),
+        ("<|endoftext|>", 199999),
+        ("<|reserved_200000|>", 200000),
+        ("<|reserved_200001|>", 200001),
+        ("<|return|>", 200002),
+        ("<|constrain|>", 200003),
+        ("<|reserved_200004|>", 200004),
+        ("<|channel|>", 200005),
+        ("<|start|>", 200006),
+        ("<|end|>", 200007),
+        ("<|message|>", 200008),
+        ("<|reserved_200009|>", 200009),
+        ("<|reserved_200010|>", 200010),
+        ("<|reserved_200011|>", 200011),
+        ("<|call|>", 200012),
+        ("<|reserved_200013|>", 200013),
+    ]
+    pattern = "|".join([
+        "[^\\r\\n\\p{L}\\p{N}]?[\\p{Lu}\\p{Lt}\\p{Lm}\\p{Lo}\\p{M}]*[\\p{Ll}\\p{Lm}\\p{Lo}\\p{M}]+(?i:'s|'t|'re|'ve|'m|'ll|'d)?",
+        "[^\\r\\n\\p{L}\\p{N}]?[\\p{Lu}\\p{Lt}\\p{Lm}\\p{Lo}\\p{M}]+[\\p{Ll}\\p{Lm}\\p{Lo}\\p{M}]*(?i:'s|'t|'re|'ve|'m|'ll|'d)?",
+        "\\p{N}{1,3}",
+        " ?[^\\s\\p{L}\\p{N}]+[\\r\\n/]*",
+        "\\s*[\\r\\n]+",
+        "\\s+(?!\\S)",
+        "\\s+",
+    ])
+    n_ctx = 8192
+    max_message_tokens = 4096
+    max_action_length = 256
+    expected_hash = "446a9538cb6c348e3516120d7c08b09f57c36495e2acfffe59a5bf8b0cfb1a2d"
+
+    encoding = load_harmony_encoding_from_file(
+        name="test_local",
+        vocab_file=vocab_file,
+        special_tokens=special_tokens,
+        pattern=pattern,
+        n_ctx=n_ctx,
+        max_message_tokens=max_message_tokens,
+        max_action_length=max_action_length,
+        expected_hash=expected_hash,
+    )
+
+    text = "Hello world!"
+    tokens = encoding.encode(text)
+    decoded = encoding.decode(tokens)
+    assert decoded.startswith("Hello world")