litellm/litellm-rust/crates/python-bridge/src/execution.rs
devin-ai-integration[bot] 46a185d3cd
feat(rust_bridge): count budget-check input tokens in Rust on all LLM routes (#40381)
* feat(rust_bridge): count budget-check input tokens in Rust on all LLM routes

Rust counts input tokens from the raw JSON body with the GIL released inside the existing budget reservation, covering every LLM route the auth dependency guards. It only fires for models on the Anthropic tokenizer when a budget is set, and Python counts whenever Rust is off, missing, or declines a body shape.

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

* perf(rust): count byte-level BPE tokens without the GPT-2 split regex (#40594)

The oniguruma run of the ByteLevel pre-tokenizer regex is about 90% of
encode_fast on a 100k token body (100 ms of the ~110 ms Rust admission
count in the gateway pod). A hand-written scanner that yields the same
pieces, then feeds the model directly, counts the same text in 10 ms.
It only engages for tokenizers with the Anthropic shape (optional NFKC,
ByteLevel without prefix space, no post-processor) and falls back to the
full encoder when the text contains an added token. Parity with
encode_fast is tested on random texts, the pieces are compared with the
real pre-tokenizer, and the \p{L}/\p{N}/\s tables are checked against
oniguruma for every code point.

NFKC runs through unicode-normalization-alignments, the crate and
Unicode tables NormalizedString::nfkc already uses, so the fast path
normalizes exactly what the full encoder would. Using the newer
unicode-normalization crate changed the count for 171 code points that
gained compatibility decompositions after Unicode 9 (U+32FF, U+A7F1..).
The fast normalizer is compared with the tokenizer's for every scalar
value and on random texts.

The scanner is built without mutable state: byte_char and mapped_len replace the const table builders and the reusable mapped buffer, and iter::successors replaces the stateful piece iterator. byte_chars_match_the_byte_level_alphabet checks the byte mapping against ByteLevel for every scalar value.

Co-authored-by: yassin <yassin@berri.ai>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

* fix(rust_bridge): bound concurrent token-count encodes and share the Anthropic tokenizer predicate

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

---------

Co-authored-by: yassin <yassin@berri.ai>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
2026-09-10 13:56:30 -07:00

426 lines
13 KiB
Rust

use std::future::Future;
use std::panic::AssertUnwindSafe;
use std::time::Duration;
use futures_util::FutureExt;
use litellm_python_interop::{Pythonized, panic_to_pyerr, release_gil};
use pyo3::exceptions::PyRuntimeError;
use pyo3::prelude::*;
use serde::Serialize;
use tokio::runtime::{Handle, Runtime};
use tokio::time::{self, MissedTickBehavior};
pub(crate) fn run_sync<T, E, F>(
py: Python<'_>,
future: F,
map_error: fn(E) -> PyErr,
) -> PyResult<Py<PyAny>>
where
T: Serialize + Send + 'static,
E: Send + 'static,
F: Future<Output = Result<T, E>> + Send + 'static,
{
run_sync_on(
py,
pyo3_async_runtimes::tokio::get_runtime(),
future,
map_error,
)
}
fn run_sync_on<T, E, F>(
py: Python<'_>,
runtime: &Runtime,
future: F,
map_error: fn(E) -> PyErr,
) -> PyResult<Py<PyAny>>
where
T: Serialize + Send + 'static,
E: Send + 'static,
F: Future<Output = Result<T, E>> + Send + 'static,
{
if Handle::try_current().is_ok() {
return Err(PyRuntimeError::new_err(
"synchronous native routes cannot run from a Tokio context; use the async route",
));
}
let result = release_gil(py, move || runtime.block_on(wait_for_sync_result(future)))?;
let result = map_core_result(result, map_error)?;
Pythonized(result).into_pyobject(py).map(Bound::unbind)
}
pub(crate) fn run_async<T, E, F>(
py: Python<'_>,
future: F,
map_error: fn(E) -> PyErr,
) -> PyResult<Bound<'_, PyAny>>
where
T: Serialize + Send + 'static,
E: Send + 'static,
F: Future<Output = Result<T, E>> + Send + 'static,
{
pyo3_async_runtimes::tokio::future_into_py(py, async move {
let result = catch_future_panic(future).await?;
let result = map_core_result(result, map_error)?;
Ok(Pythonized(result))
})
}
fn map_core_result<T, E>(result: Result<T, E>, map_error: fn(E) -> PyErr) -> PyResult<T> {
match result {
Ok(value) => Ok(value),
Err(error) => Err(
std::panic::catch_unwind(AssertUnwindSafe(|| map_error(error)))
.map_err(panic_to_pyerr)?,
),
}
}
async fn catch_future_panic<T, E, F>(future: F) -> PyResult<Result<T, E>>
where
F: Future<Output = Result<T, E>>,
{
AssertUnwindSafe(future)
.catch_unwind()
.await
.map_err(panic_to_pyerr)
}
async fn wait_for_sync_result<T, E, F>(future: F) -> PyResult<Result<T, E>>
where
F: Future<Output = Result<T, E>>,
{
let future = catch_future_panic(future);
tokio::pin!(future);
let signal_interval = Duration::from_millis(50);
let mut signal_checks =
time::interval_at(time::Instant::now() + signal_interval, signal_interval);
signal_checks.set_missed_tick_behavior(MissedTickBehavior::Delay);
loop {
tokio::select! {
result = &mut future => return result,
_ = signal_checks.tick() => Python::attach(|py| py.check_signals())?,
}
}
}
#[cfg(test)]
mod tests {
use std::ffi::CString;
use std::future::poll_fn;
use std::sync::atomic::{AtomicUsize, Ordering};
use std::sync::{Arc, mpsc};
use std::task::Poll;
use std::thread;
use std::time::Instant;
use litellm_core::error::Error;
use pyo3::panic::PanicException;
use pyo3::types::{PyDict, PyModule};
use serde::Serializer;
use tokio::runtime::Builder;
use super::*;
fn runtime_error(error: Error) -> PyErr {
PyRuntimeError::new_err(error.to_string())
}
fn panicking_error_mapper(_error: Error) -> PyErr {
panic!("error mapper panicked")
}
struct PanickingOutput;
static ASYNC_PROBE_COMPLETED: AtomicUsize = AtomicUsize::new(0);
impl Serialize for PanickingOutput {
fn serialize<S>(&self, _serializer: S) -> Result<S::Ok, S::Error>
where
S: Serializer,
{
panic!("serializer panicked")
}
}
#[pyfunction]
fn async_serialization_panic(py: Python<'_>) -> PyResult<Bound<'_, PyAny>> {
run_async(py, async { Ok(PanickingOutput) }, runtime_error)
}
#[pyfunction]
fn async_runtime_probe(py: Python<'_>) -> PyResult<Bound<'_, PyAny>> {
run_async(
py,
async {
ASYNC_PROBE_COMPLETED.fetch_add(1, Ordering::SeqCst);
Ok(true)
},
runtime_error,
)
}
#[pyfunction]
fn runtime_worker_count() -> usize {
pyo3_async_runtimes::tokio::get_runtime()
.metrics()
.num_workers()
}
#[pyfunction]
fn runtime_is_responsive(_py: Python<'_>, expected_completions: usize) -> bool {
let completion_deadline = Instant::now() + Duration::from_secs(2);
while ASYNC_PROBE_COMPLETED.load(Ordering::SeqCst) < expected_completions {
if Instant::now() >= completion_deadline {
return false;
}
thread::sleep(Duration::from_millis(1));
}
let (heartbeat_tx, heartbeat_rx) = mpsc::sync_channel(1);
pyo3_async_runtimes::tokio::get_runtime().spawn(async move {
let _ = heartbeat_tx.send(());
});
heartbeat_rx.recv_timeout(Duration::from_secs(2)).is_ok()
}
fn extract_bool(py: Python<'_>, result: PyResult<Py<PyAny>>) -> bool {
result
.expect("route should complete")
.bind(py)
.extract()
.expect("result should convert")
}
#[test]
fn sync_runner_polls_future_on_the_caller_thread() {
Python::initialize();
Python::attach(|py| {
let caller_thread = std::thread::current().id();
let result = run_sync(
py,
async move { Ok(std::thread::current().id() == caller_thread) },
runtime_error,
);
assert!(extract_bool(py, result));
});
}
#[test]
fn sync_runner_releases_gil_while_waiting() {
Python::initialize();
Python::attach(|py| {
let result = run_sync(
py,
async {
let gil_acquired = tokio::time::timeout(
Duration::from_secs(2),
tokio::task::spawn_blocking(|| Python::attach(|_| true)),
)
.await;
Ok(matches!(gil_acquired, Ok(Ok(true))))
},
runtime_error,
);
assert!(extract_bool(py, result));
});
}
#[test]
fn sync_runner_rejects_calls_from_a_tokio_context() {
Python::initialize();
let runtime = Builder::new_current_thread()
.enable_all()
.build()
.expect("runtime should build");
let error = runtime.block_on(async {
Python::attach(|py| {
run_sync::<bool, Error, _>(py, async { Ok(true) }, runtime_error)
.expect_err("sync route should reject a nested Tokio runtime")
})
});
assert_eq!(
error.to_string(),
"RuntimeError: synchronous native routes cannot run from a Tokio context; use the async route"
);
}
#[test]
fn sync_runner_can_drive_a_current_thread_runtime() {
Python::initialize();
let runtime = Builder::new_current_thread()
.enable_all()
.build()
.expect("runtime should build");
Python::attach(|py| {
let result = run_sync_on(
py,
&runtime,
async {
tokio::task::yield_now().await;
Ok(true)
},
runtime_error,
);
assert!(extract_bool(py, result));
});
}
#[test]
fn sync_runner_maps_a_panicked_future() {
Python::initialize();
Python::attach(|py| {
let error = run_sync::<bool, Error, _>(
py,
poll_fn(|_| -> Poll<Result<bool, Error>> { panic!("route future panicked") }),
runtime_error,
)
.expect_err("panicked route should become a Python exception");
assert!(error.is_instance_of::<PanicException>(py));
assert_eq!(error.to_string(), "PanicException: route future panicked");
});
}
#[test]
fn sync_runner_maps_a_panicked_error_mapper() {
Python::initialize();
Python::attach(|py| {
let error = run_sync::<bool, Error, _>(
py,
async { Err(Error::InvalidRequest("invalid".to_string())) },
panicking_error_mapper,
)
.expect_err("panicked mapper should become a Python exception");
assert!(error.is_instance_of::<PanicException>(py));
assert_eq!(error.to_string(), "PanicException: error mapper panicked");
});
}
#[test]
fn sync_runner_surfaces_serializer_panics() {
Python::initialize();
Python::attach(|py| {
let error = run_sync(py, async { Ok(PanickingOutput) }, runtime_error)
.expect_err("serializer panic should become a Python exception");
assert!(error.is_instance_of::<PanicException>(py));
assert_eq!(error.to_string(), "PanicException: serializer panicked");
});
}
#[test]
fn sync_runner_supports_concurrent_callers_on_the_shared_runtime() {
Python::initialize();
let barrier = Arc::new(tokio::sync::Barrier::new(2));
let callers: Vec<_> = (0..2)
.map(|_| {
let barrier = Arc::clone(&barrier);
thread::spawn(move || {
Python::attach(|py| {
extract_bool(
py,
run_sync(
py,
async move {
Ok(tokio::time::timeout(Duration::from_secs(2), barrier.wait())
.await
.is_ok())
},
runtime_error,
),
)
})
})
})
.collect();
let results: Vec<_> = callers
.into_iter()
.map(|caller| caller.join().expect("caller should not panic"))
.collect();
assert_eq!(results, vec![true, true]);
}
#[test]
fn async_runner_surfaces_serializer_panics() {
Python::initialize();
Python::attach(|py| {
let module = PyModule::new(py, "runtime").expect("module should be created");
module
.add_function(
wrap_pyfunction!(async_serialization_panic, &module)
.expect("function should wrap"),
)
.expect("function should register");
let locals = PyDict::new(py);
locals
.set_item("runtime", &module)
.expect("module should enter Python locals");
let code = CString::new(
r#"
import asyncio
async def exercise():
try:
await runtime.async_serialization_panic()
except BaseException as error:
assert type(error).__name__ == "PanicException"
assert str(error) == "serializer panicked"
else:
raise AssertionError("serializer panic was not raised")
asyncio.run(exercise())
"#,
)
.expect("Python source should not contain null bytes");
py.run(&code, Some(&locals), Some(&locals))
.expect("serializer panic should reach the Python awaiter");
});
}
#[test]
fn async_result_delivery_does_not_stall_tokio_workers() {
Python::initialize();
ASYNC_PROBE_COMPLETED.store(0, Ordering::SeqCst);
Python::attach(|py| {
let module = PyModule::new(py, "runtime").expect("module should be created");
for function in [
wrap_pyfunction!(async_runtime_probe, &module).expect("function should wrap"),
wrap_pyfunction!(runtime_worker_count, &module).expect("function should wrap"),
wrap_pyfunction!(runtime_is_responsive, &module).expect("function should wrap"),
] {
module
.add_function(function)
.expect("function should register");
}
let locals = PyDict::new(py);
locals
.set_item("runtime", &module)
.expect("module should enter Python locals");
let code = CString::new(
r#"
import asyncio
async def exercise():
worker_count = runtime.runtime_worker_count()
awaitables = [runtime.async_runtime_probe() for _ in range(worker_count)]
assert runtime.runtime_is_responsive(worker_count)
assert await asyncio.gather(*awaitables) == [True] * worker_count
asyncio.run(exercise())
"#,
)
.expect("Python source should not contain null bytes");
py.run(&code, Some(&locals), Some(&locals))
.expect("result delivery should leave Tokio workers responsive");
});
}
}