lib.rs 29.4 KB
Newer Older
1
2
3
// SPDX-FileCopyrightText: Copyright (c) 2024-2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
// SPDX-License-Identifier: Apache-2.0

4
5
6
use futures::StreamExt;
use once_cell::sync::OnceCell;
use pyo3::exceptions::PyStopAsyncIteration;
7
8
use pyo3::types::PyBytes;
use pyo3::types::{PyDict, PyList, PyString};
9
10
11
use pyo3::IntoPyObjectExt;
use pyo3::{exceptions::PyException, prelude::*};
use rs::pipeline::network::Ingress;
12
use std::path::PathBuf;
13
14
15
use std::{fmt::Display, sync::Arc};
use tokio::sync::Mutex;

Neelay Shah's avatar
Neelay Shah committed
16
use dynamo_runtime::{
Ryan Olson's avatar
Ryan Olson committed
17
    self as rs, logging,
18
19
20
    pipeline::{
        network::egress::push_router::RouterMode as RsRouterMode, EngineStream, ManyOut, SingleIn,
    },
21
    protocols::annotated::Annotated as RsAnnotated,
22
    traits::DistributedRuntimeProvider,
23
24
};

Neelay Shah's avatar
Neelay Shah committed
25
use dynamo_llm::{self as llm_rs};
26
27
use dynamo_llm::{entrypoint::RouterConfig, kv_router::KvRouterConfig};

28
29
use crate::llm::local_model::ModelRuntimeConfig;

30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
#[pyclass(eq, eq_int)]
#[derive(Clone, Debug, PartialEq)]
pub enum RouterMode {
    RoundRobin,
    Random,
    KV,
}

impl From<RouterMode> for RsRouterMode {
    fn from(mode: RouterMode) -> Self {
        match mode {
            RouterMode::RoundRobin => Self::RoundRobin,
            RouterMode::Random => Self::Random,
            RouterMode::KV => Self::KV,
        }
    }
}
47

48
mod context;
49
mod engine;
50
mod http;
51
mod llm;
52
53
54
55
56
57
58
59
60
61
62
63
64

type JsonServerStreamingIngress =
    Ingress<SingleIn<serde_json::Value>, ManyOut<RsAnnotated<serde_json::Value>>>;

static INIT: OnceCell<()> = OnceCell::new();

const DEFAULT_ANNOTATED_SETTING: Option<bool> = Some(true);

/// A Python module implemented in Rust. The name of this function must match
/// the `lib.name` setting in the `Cargo.toml`, else Python will not be able to
/// import the module.
#[pymodule]
fn _core(m: &Bound<'_, PyModule>) -> PyResult<()> {
Ryan Olson's avatar
Ryan Olson committed
65
    logging::init();
Yan Ru Pei's avatar
Yan Ru Pei committed
66
    m.add_function(wrap_pyfunction!(llm::kv::compute_block_hash_for_seq_py, m)?)?;
Ryan Olson's avatar
Ryan Olson committed
67
    m.add_function(wrap_pyfunction!(log_message, m)?)?;
68
    m.add_function(wrap_pyfunction!(register_llm, m)?)?;
69
70
    m.add_function(wrap_pyfunction!(llm::entrypoint::make_engine, m)?)?;
    m.add_function(wrap_pyfunction!(llm::entrypoint::run_input, m)?)?;
71

72
73
74
75
76
77
    m.add_class::<DistributedRuntime>()?;
    m.add_class::<CancellationToken>()?;
    m.add_class::<Namespace>()?;
    m.add_class::<Component>()?;
    m.add_class::<Endpoint>()?;
    m.add_class::<Client>()?;
78
    m.add_class::<EtcdClient>()?;
79
    m.add_class::<AsyncResponseStream>()?;
80
    m.add_class::<llm::disagg_router::DisaggregatedRouter>()?;
81
82
83
    m.add_class::<llm::entrypoint::EntrypointArgs>()?;
    m.add_class::<llm::entrypoint::EngineConfig>()?;
    m.add_class::<llm::entrypoint::EngineType>()?;
84
85
    m.add_class::<llm::entrypoint::RouterConfig>()?;
    m.add_class::<llm::entrypoint::KvRouterConfig>()?;
86
    m.add_class::<llm::kv::WorkerMetricsPublisher>()?;
87
    m.add_class::<llm::model_card::ModelDeploymentCard>()?;
88
    m.add_class::<llm::local_model::ModelRuntimeConfig>()?;
89
90
    m.add_class::<llm::preprocessor::OAIChatPreprocessor>()?;
    m.add_class::<llm::backend::Backend>()?;
91
92
    m.add_class::<llm::kv::OverlapScores>()?;
    m.add_class::<llm::kv::KvIndexer>()?;
93
    m.add_class::<llm::kv::ApproxKvIndexer>()?;
94
95
96
    m.add_class::<llm::kv::EndpointKvMetrics>()?;
    m.add_class::<llm::kv::AggregatedMetrics>()?;
    m.add_class::<llm::kv::KvMetricsAggregator>()?;
97
    m.add_class::<llm::kv::KvEventPublisher>()?;
Yan Ru Pei's avatar
Yan Ru Pei committed
98
99
    m.add_class::<llm::kv::RadixTree>()?;
    m.add_class::<llm::kv::ZmqKvEventListener>()?;
100
101
    m.add_class::<llm::kv::ZmqKvEventPublisher>()?;
    m.add_class::<llm::kv::ZmqKvEventPublisherConfig>()?;
102
    m.add_class::<llm::kv::KvRecorder>()?;
103
    m.add_class::<llm::nats::NatsQueue>()?;
104
105
106
    m.add_class::<http::HttpService>()?;
    m.add_class::<http::HttpError>()?;
    m.add_class::<http::HttpAsyncEngine>()?;
107
    m.add_class::<context::PyContext>()?;
108
    m.add_class::<EtcdKvCache>()?;
109
    m.add_class::<ModelType>()?;
110
111
112
113
    m.add_class::<llm::kv::ForwardPassMetrics>()?;
    m.add_class::<llm::kv::WorkerStats>()?;
    m.add_class::<llm::kv::KvStats>()?;
    m.add_class::<llm::kv::SpecDecodeStats>()?;
114
    m.add_class::<RouterMode>()?;
115
116
117

    engine::add_to_module(m)?;

118
119
120
    #[cfg(feature = "block-manager")]
    llm::block_manager::add_to_module(m)?;

121
122
123
124
125
126
127
128
129
130
    Ok(())
}

pub fn to_pyerr<E>(err: E) -> PyErr
where
    E: Display,
{
    PyException::new_err(format!("{}", err))
}

Ryan Olson's avatar
Ryan Olson committed
131
132
133
134
135
136
137
/// Log a message from Python with file and line info
#[pyfunction]
#[pyo3(text_signature = "(level, message, module, file, line)")]
fn log_message(level: &str, message: &str, module: &str, file: &str, line: u32) {
    logging::log_message(level, message, module, file, line);
}

138
#[pyfunction]
139
#[pyo3(signature = (model_type, endpoint, model_path, model_name=None, context_length=None, kv_cache_block_size=None, router_mode=None, migration_limit=0, runtime_config=None, user_data=None))]
140
#[allow(clippy::too_many_arguments)]
141
142
143
fn register_llm<'p>(
    py: Python<'p>,
    model_type: ModelType,
144
145
146
    endpoint: Endpoint,
    model_path: &str,
    model_name: Option<&str>,
147
148
    context_length: Option<u32>,
    kv_cache_block_size: Option<u32>,
149
    router_mode: Option<RouterMode>,
150
    migration_limit: u32,
151
    runtime_config: Option<ModelRuntimeConfig>,
152
    user_data: Option<&Bound<'p, PyDict>>,
153
154
155
156
157
) -> PyResult<Bound<'p, PyAny>> {
    let model_type_obj = match model_type {
        ModelType::Chat => llm_rs::model_type::ModelType::Chat,
        ModelType::Completion => llm_rs::model_type::ModelType::Completion,
        ModelType::Backend => llm_rs::model_type::ModelType::Backend,
158
        ModelType::Embedding => llm_rs::model_type::ModelType::Embedding,
159
160
    };

161
162
    let inner_path = model_path.to_string();
    let model_name = model_name.map(|n| n.to_string());
163
164
165
    let router_mode = router_mode.unwrap_or(RouterMode::RoundRobin);
    let router_config = RouterConfig::new(router_mode.into(), KvRouterConfig::default());

166
167
168
169
170
171
172
    let user_data_json = user_data
        .map(|dict| pythonize::depythonize(dict))
        .transpose()
        .map_err(|err| {
            PyErr::new::<PyException, _>(format!("Failed to convert user_data: {}", err))
        })?;

173
    pyo3_async_runtimes::tokio::future_into_py(py, async move {
174
175
176
177
178
        let mut builder = dynamo_llm::local_model::LocalModelBuilder::default();
        builder
            .model_path(Some(PathBuf::from(inner_path)))
            .model_name(model_name)
            .context_length(context_length)
179
            .kv_cache_block_size(kv_cache_block_size)
180
            .router_config(Some(router_config))
181
            .migration_limit(Some(migration_limit))
182
            .runtime_config(runtime_config.unwrap_or_default().inner)
183
            .user_data(user_data_json);
184
        // Download from HF, load the ModelDeploymentCard
185
        let mut local_model = builder.build().await.map_err(to_pyerr)?;
186
187
188
189
190
191
192
193
194
195
        // Advertise ourself on etcd so ingress can find us
        local_model
            .attach(&endpoint.inner, model_type_obj)
            .await
            .map_err(to_pyerr)?;

        Ok(())
    })
}

196
197
198
199
200
201
#[pyclass]
#[derive(Clone)]
struct EtcdKvCache {
    inner: Arc<rs::transports::etcd::KvCache>,
}

202
203
#[pyclass]
#[derive(Clone)]
Ryan Olson's avatar
Ryan Olson committed
204
pub struct DistributedRuntime {
205
206
207
208
    inner: rs::DistributedRuntime,
    event_loop: PyObject,
}

Ryan Olson's avatar
Ryan Olson committed
209
210
211
212
213
214
215
impl DistributedRuntime {
    #[allow(dead_code)]
    fn inner(&self) -> &rs::DistributedRuntime {
        &self.inner
    }
}

216
217
218
219
220
221
#[pyclass]
#[derive(Clone)]
struct EtcdClient {
    inner: rs::transports::etcd::Client,
}

222
#[pyclass]
223
#[derive(Clone)]
224
225
226
227
228
struct CancellationToken {
    inner: rs::CancellationToken,
}

#[pyclass]
229
#[derive(Clone)]
230
231
232
233
234
235
struct Namespace {
    inner: rs::component::Namespace,
    event_loop: PyObject,
}

#[pyclass]
236
#[derive(Clone)]
237
238
239
240
241
242
struct Component {
    inner: rs::component::Component,
    event_loop: PyObject,
}

#[pyclass]
243
#[derive(Clone)]
244
245
246
247
248
249
struct Endpoint {
    inner: rs::component::Endpoint,
    event_loop: PyObject,
}

#[pyclass]
250
#[derive(Clone)]
251
struct Client {
252
    router: rs::pipeline::PushRouter<serde_json::Value, RsAnnotated<serde_json::Value>>,
253
254
}

255
256
257
258
259
260
261
#[pyclass(eq, eq_int)]
#[derive(Clone, PartialEq)]
#[repr(i32)]
enum ModelType {
    Chat = 1,
    Completion = 2,
    Backend = 3,
262
    Embedding = 4,
263
264
}

265
266
267
#[pymethods]
impl DistributedRuntime {
    #[new]
268
    fn new(event_loop: PyObject, is_static: bool) -> PyResult<Self> {
269
        let worker = rs::Worker::from_settings().map_err(to_pyerr)?;
270
        INIT.get_or_try_init(|| {
271
            let primary = worker.tokio_runtime()?;
272
273
274
275
276
277
            pyo3_async_runtimes::tokio::init_with_runtime(primary)
                .map_err(|e| rs::error!("failed to initialize pyo3 static runtime: {:?}", e))?;
            rs::OK(())
        })
        .map_err(to_pyerr)?;

278
        let runtime = worker.runtime().clone();
279

280
281
282
283
284
285
286
287
288
289
290
        let inner =
            if is_static {
                runtime.secondary().block_on(
                    rs::DistributedRuntime::from_settings_without_discovery(runtime),
                )
            } else {
                runtime
                    .secondary()
                    .block_on(rs::DistributedRuntime::from_settings(runtime))
            };
        let inner = inner.map_err(to_pyerr)?;
291
292
293
294

        Ok(DistributedRuntime { inner, event_loop })
    }

Ryan Olson's avatar
Ryan Olson committed
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
    #[staticmethod]
    fn detached(py: Python) -> PyResult<Self> {
        let rt = rs::Worker::runtime_from_existing().map_err(to_pyerr)?;
        let handle = rt.primary();

        let inner = handle
            .block_on(rs::DistributedRuntime::from_settings(rt))
            .map_err(to_pyerr)?;

        Ok(DistributedRuntime {
            inner,
            event_loop: py.None(),
        })
    }

310
311
312
313
314
315
316
    fn namespace(&self, name: String) -> PyResult<Namespace> {
        Ok(Namespace {
            inner: self.inner.namespace(name).map_err(to_pyerr)?,
            event_loop: self.event_loop.clone(),
        })
    }

317
318
319
320
321
    fn etcd_client(&self) -> PyResult<Option<EtcdClient>> {
        match self.inner.etcd_client().clone() {
            Some(etcd_client) => Ok(Some(EtcdClient { inner: etcd_client })),
            None => Ok(None),
        }
322
323
    }

324
325
326
327
328
329
330
331
332
    fn shutdown(&self) {
        self.inner.runtime().shutdown();
    }

    fn event_loop(&self) -> PyObject {
        self.event_loop.clone()
    }
}

333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
#[pymethods]
impl EtcdKvCache {
    #[new]
    fn py_new(
        _etcd_client: &EtcdClient,
        _prefix: String,
        _initial_values: &Bound<'_, PyDict>,
    ) -> PyResult<Self> {
        // We can't create the KvCache here because it's async, so we'll return an error
        Err(PyErr::new::<pyo3::exceptions::PyRuntimeError, _>(
            "EtcdKvCache must be created using the 'new' class method",
        ))
    }

    #[staticmethod]
    #[allow(clippy::new_ret_no_self)]
    fn create<'p>(
        py: Python<'p>,
        etcd_client: &EtcdClient,
        prefix: String,
        initial_values: &Bound<'p, PyDict>,
    ) -> PyResult<Bound<'p, PyAny>> {
        let client = etcd_client.inner.clone();

        // Convert Python dict to Rust HashMap
        let mut rust_initial_values = std::collections::HashMap::new();
        for (key, value) in initial_values.iter() {
            let key_str = key.extract::<String>()?;

            // Handle both string and bytes values
            let value_bytes = if let Ok(bytes) = value.extract::<Vec<u8>>() {
                bytes
            } else if let Ok(string) = value.extract::<String>() {
                string.into_bytes()
            } else {
                return Err(PyErr::new::<pyo3::exceptions::PyTypeError, _>(
                    "Values must be either strings or bytes",
                ));
            };

            rust_initial_values.insert(key_str, value_bytes);
        }

        pyo3_async_runtimes::tokio::future_into_py(py, async move {
            let kv_cache = rs::transports::etcd::KvCache::new(client, prefix, rust_initial_values)
                .await
                .map_err(to_pyerr)?;

            Ok(EtcdKvCache {
                inner: Arc::new(kv_cache),
            })
        })
    }

    fn get<'p>(&self, py: Python<'p>, key: String) -> PyResult<Bound<'p, PyAny>> {
        let inner = self.inner.clone();

        pyo3_async_runtimes::tokio::future_into_py(py, async move {
            if let Some(value) = inner.get(&key).await {
                match Python::with_gil(|py| {
                    let py_obj = PyBytes::new(py, &value).into_pyobject(py)?;
                    Ok(py_obj.unbind().into_any())
                }) {
                    Ok(result) => Ok(result),
                    Err(e) => Err(e),
                }
            } else {
                Ok(Python::with_gil(|py| py.None()))
            }
        })
    }

    fn get_all<'p>(&self, py: Python<'p>) -> PyResult<Bound<'p, PyAny>> {
        let inner = self.inner.clone();

        pyo3_async_runtimes::tokio::future_into_py(py, async move {
            let all_values = inner.get_all().await;

            Python::with_gil(|py| {
                let dict = PyDict::new(py);
                for (key, value) in all_values {
                    // Strip the prefix from the key
                    let stripped_key = if let Some(stripped) = key.strip_prefix(&inner.prefix) {
                        stripped.to_string()
                    } else {
                        key
                    };
                    dict.set_item(stripped_key, PyBytes::new(py, &value))?;
                }
                let py_obj = dict.into_pyobject(py)?;
                Ok(py_obj.unbind().into_any())
            })
        })
    }

    #[pyo3(signature = (key, value, lease_id=None))]
    fn put<'p>(
        &self,
        py: Python<'p>,
        key: String,
        value: Vec<u8>,
        lease_id: Option<i64>,
    ) -> PyResult<Bound<'p, PyAny>> {
        let inner = self.inner.clone();

        pyo3_async_runtimes::tokio::future_into_py(py, async move {
            inner.put(&key, value, lease_id).await.map_err(to_pyerr)?;
            Ok(())
        })
    }
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477

    fn delete<'p>(&self, py: Python<'p>, key: String) -> PyResult<Bound<'p, PyAny>> {
        let inner = self.inner.clone();

        pyo3_async_runtimes::tokio::future_into_py(py, async move {
            inner.delete(&key).await.map_err(to_pyerr)?;
            Ok(())
        })
    }

    fn clear_all<'p>(&self, py: Python<'p>) -> PyResult<Bound<'p, PyAny>> {
        let inner = self.inner.clone();

        pyo3_async_runtimes::tokio::future_into_py(py, async move {
            // Get all keys with the prefix
            let all_keys = inner
                .get_all()
                .await
                .keys()
                .cloned()
                .collect::<Vec<String>>();

            // Delete each key
            for key in all_keys {
                // Strip the prefix from the key before deleting
                if let Some(stripped_key) = key.strip_prefix(&inner.prefix) {
                    inner.delete(stripped_key).await.map_err(to_pyerr)?;
                } else {
                    inner.delete(&key).await.map_err(to_pyerr)?;
                }
            }

            Ok(())
        })
    }
478
479
}

480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
#[pymethods]
impl CancellationToken {
    fn cancel(&self) {
        self.inner.cancel();
    }

    fn cancelled<'p>(&self, py: Python<'p>) -> PyResult<Bound<'p, PyAny>> {
        let token = self.inner.clone();
        pyo3_async_runtimes::tokio::future_into_py(py, async move {
            token.cancelled().await;
            Ok(())
        })
    }
}

#[pymethods]
impl Component {
    fn endpoint(&self, name: String) -> PyResult<Endpoint> {
        let inner = self.inner.endpoint(name);
        Ok(Endpoint {
            inner,
            event_loop: self.event_loop.clone(),
        })
    }

    fn create_service<'p>(&self, py: Python<'p>) -> PyResult<Bound<'p, PyAny>> {
        let builder = self.inner.service_builder();
        pyo3_async_runtimes::tokio::future_into_py(py, async move {
            let _ = builder.create().await.map_err(to_pyerr)?;
            Ok(())
        })
    }
}

#[pymethods]
impl Endpoint {
516
    #[pyo3(signature = (generator, graceful_shutdown = true))]
517
518
519
520
    fn serve_endpoint<'p>(
        &self,
        py: Python<'p>,
        generator: PyObject,
521
        graceful_shutdown: Option<bool>,
522
523
524
525
526
527
    ) -> PyResult<Bound<'p, PyAny>> {
        let engine = Arc::new(engine::PythonAsyncEngine::new(
            generator,
            self.event_loop.clone(),
        )?);
        let ingress = JsonServerStreamingIngress::for_engine(engine).map_err(to_pyerr)?;
528
        let builder = self.inner.endpoint_builder().handler(ingress);
529
        let graceful_shutdown = graceful_shutdown.unwrap_or(true);
530
        pyo3_async_runtimes::tokio::future_into_py(py, async move {
531
532
533
534
535
            builder
                .graceful_shutdown(graceful_shutdown)
                .start()
                .await
                .map_err(to_pyerr)?;
536
537
538
539
            Ok(())
        })
    }

540
541
542
    fn client<'p>(&self, py: Python<'p>) -> PyResult<Bound<'p, PyAny>> {
        let inner = self.inner.clone();
        pyo3_async_runtimes::tokio::future_into_py(py, async move {
543
            let client = inner.client().await.map_err(to_pyerr)?;
544
545
546
547
548
549
            let push_router = rs::pipeline::PushRouter::<
                serde_json::Value,
                RsAnnotated<serde_json::Value>,
            >::from_client(client, Default::default())
            .await
            .map_err(to_pyerr)?;
550
551
552
            Ok(Client {
                router: push_router,
            })
553
554
        })
    }
555
556

    fn lease_id(&self) -> i64 {
557
558
559
560
561
        self.inner
            .drt()
            .primary_lease()
            .map(|l| l.id())
            .unwrap_or(0)
562
    }
563
564
565
566
567
568
569
570
571
572
573
574
575
}

#[pymethods]
impl Namespace {
    fn component(&self, name: String) -> PyResult<Component> {
        let inner = self.inner.component(name).map_err(to_pyerr)?;
        Ok(Component {
            inner,
            event_loop: self.event_loop.clone(),
        })
    }
}

576
577
#[pymethods]
impl EtcdClient {
578
579
580
581
582
583
584
585
586
587
588
    #[pyo3(signature = (key, value, lease_id=None))]
    fn kv_create<'p>(
        &self,
        py: Python<'p>,
        key: String,
        value: Vec<u8>,
        lease_id: Option<i64>,
    ) -> PyResult<Bound<'p, PyAny>> {
        let client = self.inner.clone();
        pyo3_async_runtimes::tokio::future_into_py(py, async move {
            client
589
                .kv_create(&key, value, lease_id)
590
591
592
593
594
595
                .await
                .map_err(to_pyerr)?;
            Ok(())
        })
    }

596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
    #[pyo3(signature = (key, value, lease_id=None))]
    fn kv_create_or_validate<'p>(
        &self,
        py: Python<'p>,
        key: String,
        value: Vec<u8>,
        lease_id: Option<i64>,
    ) -> PyResult<Bound<'p, PyAny>> {
        let client = self.inner.clone();
        pyo3_async_runtimes::tokio::future_into_py(py, async move {
            client
                .kv_create_or_validate(key, value, lease_id)
                .await
                .map_err(to_pyerr)?;
            Ok(())
        })
    }

614
615
616
617
    fn primary_lease_id(&self) -> i64 {
        self.inner.lease_id()
    }

618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
    #[pyo3(signature = (key, value, lease_id=None))]
    fn kv_put<'p>(
        &self,
        py: Python<'p>,
        key: String,
        value: Vec<u8>,
        lease_id: Option<i64>,
    ) -> PyResult<Bound<'p, PyAny>> {
        let client = self.inner.clone();
        pyo3_async_runtimes::tokio::future_into_py(py, async move {
            client
                .kv_put(key, value, lease_id)
                .await
                .map_err(to_pyerr)?;
            Ok(())
        })
    }

    fn kv_get_prefix<'p>(&self, py: Python<'p>, prefix: String) -> PyResult<Bound<'p, PyAny>> {
        let client = self.inner.clone();
        pyo3_async_runtimes::tokio::future_into_py(py, async move {
            let result = client
                .kv_get_prefix(prefix)
                .await
                .map_err(|e| pyo3::exceptions::PyRuntimeError::new_err(e.to_string()))?;

            // Convert Vec<KeyValue> to a list of dictionaries
            let py_list = Python::with_gil(|py| {
                let list = PyList::empty(py);
                for kv in result {
                    let dict = PyDict::new(py);
                    dict.set_item("key", String::from_utf8_lossy(kv.key()).to_string())?;
                    dict.set_item("value", PyBytes::new(py, kv.value()))?;
                    dict.set_item("create_revision", kv.create_revision())?;
                    dict.set_item("mod_revision", kv.mod_revision())?;
                    dict.set_item("version", kv.version())?;
                    dict.set_item("lease", kv.lease())?;
                    list.append(dict)?;
                }
                Ok::<Py<PyList>, PyErr>(list.into())
            })?;

            Ok(py_list)
        })
    }
663
664
665
666
667
668
669
670

    fn revoke_lease<'p>(&self, py: Python<'p>, lease_id: i64) -> PyResult<Bound<'p, PyAny>> {
        let client = self.inner.clone();
        pyo3_async_runtimes::tokio::future_into_py(py, async move {
            client.revoke_lease(lease_id).await.map_err(to_pyerr)?;
            Ok(())
        })
    }
671
672
}

673
674
#[pymethods]
impl Client {
675
676
677
678
    /// Get list of current instances.
    /// Replaces endpoint_ids.
    fn instance_ids(&self) -> Vec<i64> {
        self.router.client.instance_ids()
679
680
    }

681
682
683
    /// Wait for an instance to be available for work.
    /// Replaces wait_for_endpoints.
    fn wait_for_instances<'p>(&self, py: Python<'p>) -> PyResult<Bound<'p, PyAny>> {
684
        let inner = self.router.client.clone();
685
        pyo3_async_runtimes::tokio::future_into_py(py, async move {
686
            inner
687
                .wait_for_instances()
688
689
690
                .await
                .map(|v| v.into_iter().map(|cei| cei.id()).collect::<Vec<i64>>())
                .map_err(to_pyerr)
691
692
693
694
695
696
697
698
699
700
701
        })
    }

    /// Issue a request to the endpoint using the default routing strategy.
    #[pyo3(signature = (request, annotated=DEFAULT_ANNOTATED_SETTING))]
    fn generate<'p>(
        &self,
        py: Python<'p>,
        request: PyObject,
        annotated: Option<bool>,
    ) -> PyResult<Bound<'p, PyAny>> {
702
        if self.router.client.is_static() {
703
704
705
706
            self.r#static(py, request, annotated)
        } else {
            self.random(py, request, annotated)
        }
707
708
709
710
711
712
713
714
715
716
717
718
719
720
    }

    /// Send a request to the next endpoint in a round-robin fashion.
    #[pyo3(signature = (request, annotated=DEFAULT_ANNOTATED_SETTING))]
    fn round_robin<'p>(
        &self,
        py: Python<'p>,
        request: PyObject,
        annotated: Option<bool>,
    ) -> PyResult<Bound<'p, PyAny>> {
        let request: serde_json::Value = pythonize::depythonize(&request.into_bound(py))?;
        let annotated = annotated.unwrap_or(false);

        let (tx, rx) = tokio::sync::mpsc::channel(32);
721
        let client = self.router.clone();
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744

        pyo3_async_runtimes::tokio::future_into_py(py, async move {
            let stream = client.round_robin(request.into()).await.map_err(to_pyerr)?;
            tokio::spawn(process_stream(stream, tx));
            Ok(AsyncResponseStream {
                rx: Arc::new(Mutex::new(rx)),
                annotated,
            })
        })
    }

    /// Send a request to a random endpoint.
    #[pyo3(signature = (request, annotated=DEFAULT_ANNOTATED_SETTING))]
    fn random<'p>(
        &self,
        py: Python<'p>,
        request: PyObject,
        annotated: Option<bool>,
    ) -> PyResult<Bound<'p, PyAny>> {
        let request: serde_json::Value = pythonize::depythonize(&request.into_bound(py))?;
        let annotated = annotated.unwrap_or(false);

        let (tx, rx) = tokio::sync::mpsc::channel(32);
745
        let client = self.router.clone();
746
747
748
749
750
751
752
753
754
755
756
757

        pyo3_async_runtimes::tokio::future_into_py(py, async move {
            let stream = client.random(request.into()).await.map_err(to_pyerr)?;
            tokio::spawn(process_stream(stream, tx));
            Ok(AsyncResponseStream {
                rx: Arc::new(Mutex::new(rx)),
                annotated,
            })
        })
    }

    /// Directly send a request to a specific endpoint.
758
    #[pyo3(signature = (request, instance_id, annotated=DEFAULT_ANNOTATED_SETTING))]
759
760
761
762
    fn direct<'p>(
        &self,
        py: Python<'p>,
        request: PyObject,
763
        instance_id: i64,
764
765
766
767
768
769
        annotated: Option<bool>,
    ) -> PyResult<Bound<'p, PyAny>> {
        let request: serde_json::Value = pythonize::depythonize(&request.into_bound(py))?;
        let annotated = annotated.unwrap_or(false);

        let (tx, rx) = tokio::sync::mpsc::channel(32);
770
        let client = self.router.clone();
771
772
773

        pyo3_async_runtimes::tokio::future_into_py(py, async move {
            let stream = client
774
                .direct(request.into(), instance_id)
775
776
777
778
779
780
781
782
783
784
785
                .await
                .map_err(to_pyerr)?;

            tokio::spawn(process_stream(stream, tx));

            Ok(AsyncResponseStream {
                rx: Arc::new(Mutex::new(rx)),
                annotated,
            })
        })
    }
786
787
788
789
790
791
792
793
794
795
796
797
798

    /// Directly send a request to a pre-defined static worker
    #[pyo3(signature = (request, annotated=DEFAULT_ANNOTATED_SETTING))]
    fn r#static<'p>(
        &self,
        py: Python<'p>,
        request: PyObject,
        annotated: Option<bool>,
    ) -> PyResult<Bound<'p, PyAny>> {
        let request: serde_json::Value = pythonize::depythonize(&request.into_bound(py))?;
        let annotated = annotated.unwrap_or(false);

        let (tx, rx) = tokio::sync::mpsc::channel(32);
799
        let client = self.router.clone();
800
801
802
803
804
805
806
807
808
809
810
811

        pyo3_async_runtimes::tokio::future_into_py(py, async move {
            let stream = client.r#static(request.into()).await.map_err(to_pyerr)?;

            tokio::spawn(process_stream(stream, tx));

            Ok(AsyncResponseStream {
                rx: Arc::new(Mutex::new(rx)),
                annotated,
            })
        })
    }
812
813
814
}

async fn process_stream(
815
    stream: EngineStream<RsAnnotated<serde_json::Value>>,
816
817
818
819
820
    tx: tokio::sync::mpsc::Sender<RsAnnotated<PyObject>>,
) {
    let mut stream = stream;
    while let Some(response) = stream.next().await {
        // Convert the response to a PyObject using Python's GIL
821
        let annotated: RsAnnotated<serde_json::Value> = response;
822
823
824
825
826
827
828
829
830
831
832
833
        let annotated: RsAnnotated<PyObject> = annotated.map_data(|data| {
            let result = Python::with_gil(|py| match pythonize::pythonize(py, &data) {
                Ok(pyobj) => Ok(pyobj.into()),
                Err(e) => Err(e.to_string()),
            });
            result
        });

        let is_error = annotated.is_error();

        // Send the PyObject through the channel or log an error
        if let Err(e) = tx.send(annotated).await {
834
            tracing::error!("Failed to send response: {:?}", e);
835
            break;
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
        }

        if is_error {
            break;
        }
    }
}

#[pyclass]
struct AsyncResponseStream {
    rx: Arc<Mutex<tokio::sync::mpsc::Receiver<RsAnnotated<PyObject>>>>,
    annotated: bool,
}

#[pymethods]
impl AsyncResponseStream {
    /// This method is required to implement the `AsyncIterator` protocol.
    #[pyo3(name = "__aiter__")]
    fn aiter(slf: PyRef<Self>, py: Python) -> PyResult<Py<PyAny>> {
        slf.into_py_any(py)
    }
    /// This method is required to implement the `AsyncIterator` protocol.
    #[pyo3(name = "__anext__")]
    fn next<'p>(&self, py: Python<'p>) -> PyResult<Bound<'p, PyAny>> {
        let rx = self.rx.clone();
        let annotated = self.annotated;

        pyo3_async_runtimes::tokio::future_into_py(py, async move {
            loop {
                let value = rx.lock().await.recv().await;
                match value {
                    Some(pyobj) => {
                        let pyobj = match pyobj.ok() {
                            Ok(pyobj) => pyobj,
                            Err(e) => {
                                return Err(PyErr::new::<pyo3::exceptions::PyValueError, _>(e));
                            }
                        };

                        if annotated {
                            let object = Annotated { inner: pyobj };
                            #[allow(deprecated)]
                            let object = Python::with_gil(|py| object.into_py(py));
                            return Ok(object);
                        } else {
                            match pyobj.data {
                                Some(data) => return Ok(data),
                                None => continue,
                            }
                        }
                    }
                    None => return Err(PyStopAsyncIteration::new_err("Stream exhausted")),
                }
            }
        })
    }
}

#[pyclass]
struct Annotated {
    inner: RsAnnotated<PyObject>,
}

#[pymethods]
impl Annotated {
    #[new]
    fn new(data: PyObject) -> Self {
        Annotated {
            inner: RsAnnotated::from_data(data),
        }
    }

    fn is_error(&self) -> bool {
        self.inner.is_error()
    }

    fn data(&self) -> Option<PyObject> {
        self.inner.data.clone()
    }

    fn event(&self) -> Option<String> {
        self.inner.event.clone()
    }

    fn comments(&self) -> Option<Vec<String>> {
        self.inner.comment.clone()
    }

    fn id(&self) -> Option<String> {
        self.inner.id.clone()
    }

    #[pyo3(name = "__repr__")]
    fn _repr(&self, py: Python) -> String {
        let data = self.inner.data.clone().map(|obj| {
            obj.call_method0(py, "__repr__")
                .and_then(|repr_obj| repr_obj.extract::<Py<PyString>>(py))
                .map(|py_str| py_str.to_string_lossy(py).into_owned())
                .unwrap_or_else(|_| "<failed_repr>".to_string())
        });

        format!(
            "Annotated(data={}, event={}, comment={:?}, id={})",
            data.unwrap_or_else(|| "<no_data>".to_string()),
            self.inner.event.as_deref().unwrap_or("None"),
            self.inner.comment.as_deref().unwrap_or(&[]),
            self.inner.id.as_deref().unwrap_or("None")
        )
    }
}