lib.rs 28.1 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
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
use dynamo_llm::{entrypoint::RouterConfig, kv_router::KvRouterConfig};

#[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,
        }
    }
}
45

46
mod engine;
47
mod http;
48
mod llm;
49
50
51
52
53
54
55
56
57
58
59
60
61

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
62
    logging::init();
Yan Ru Pei's avatar
Yan Ru Pei committed
63
    m.add_function(wrap_pyfunction!(llm::kv::compute_block_hash_for_seq_py, m)?)?;
Ryan Olson's avatar
Ryan Olson committed
64
    m.add_function(wrap_pyfunction!(log_message, m)?)?;
65
    m.add_function(wrap_pyfunction!(register_llm, m)?)?;
66
67
    m.add_function(wrap_pyfunction!(llm::entrypoint::make_engine, m)?)?;
    m.add_function(wrap_pyfunction!(llm::entrypoint::run_input, m)?)?;
68

69
70
71
72
73
74
    m.add_class::<DistributedRuntime>()?;
    m.add_class::<CancellationToken>()?;
    m.add_class::<Namespace>()?;
    m.add_class::<Component>()?;
    m.add_class::<Endpoint>()?;
    m.add_class::<Client>()?;
75
    m.add_class::<EtcdClient>()?;
76
    m.add_class::<AsyncResponseStream>()?;
77
    m.add_class::<llm::disagg_router::DisaggregatedRouter>()?;
78
79
80
    m.add_class::<llm::entrypoint::EntrypointArgs>()?;
    m.add_class::<llm::entrypoint::EngineConfig>()?;
    m.add_class::<llm::entrypoint::EngineType>()?;
81
82
    m.add_class::<llm::entrypoint::RouterConfig>()?;
    m.add_class::<llm::entrypoint::KvRouterConfig>()?;
83
    m.add_class::<llm::kv::WorkerMetricsPublisher>()?;
84
85
86
    m.add_class::<llm::model_card::ModelDeploymentCard>()?;
    m.add_class::<llm::preprocessor::OAIChatPreprocessor>()?;
    m.add_class::<llm::backend::Backend>()?;
87
88
    m.add_class::<llm::kv::OverlapScores>()?;
    m.add_class::<llm::kv::KvIndexer>()?;
89
    m.add_class::<llm::kv::ApproxKvIndexer>()?;
90
91
92
    m.add_class::<llm::kv::EndpointKvMetrics>()?;
    m.add_class::<llm::kv::AggregatedMetrics>()?;
    m.add_class::<llm::kv::KvMetricsAggregator>()?;
93
    m.add_class::<llm::kv::KvEventPublisher>()?;
Yan Ru Pei's avatar
Yan Ru Pei committed
94
95
    m.add_class::<llm::kv::RadixTree>()?;
    m.add_class::<llm::kv::ZmqKvEventListener>()?;
96
97
    m.add_class::<llm::kv::ZmqKvEventPublisher>()?;
    m.add_class::<llm::kv::ZmqKvEventPublisherConfig>()?;
98
    m.add_class::<llm::kv::KvRecorder>()?;
99
    m.add_class::<llm::nats::NatsQueue>()?;
100
101
102
    m.add_class::<http::HttpService>()?;
    m.add_class::<http::HttpError>()?;
    m.add_class::<http::HttpAsyncEngine>()?;
103
    m.add_class::<EtcdKvCache>()?;
104
    m.add_class::<ModelType>()?;
105
106
107
108
    m.add_class::<llm::kv::ForwardPassMetrics>()?;
    m.add_class::<llm::kv::WorkerStats>()?;
    m.add_class::<llm::kv::KvStats>()?;
    m.add_class::<llm::kv::SpecDecodeStats>()?;
109
    m.add_class::<RouterMode>()?;
110
111
112

    engine::add_to_module(m)?;

113
114
115
    #[cfg(feature = "block-manager")]
    llm::block_manager::add_to_module(m)?;

116
117
118
119
120
121
122
123
124
125
    Ok(())
}

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

Ryan Olson's avatar
Ryan Olson committed
126
127
128
129
130
131
132
/// 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);
}

133
#[pyfunction]
134
#[pyo3(signature = (model_type, endpoint, model_path, model_name=None, context_length=None, kv_cache_block_size=None, router_mode=None, migration_limit=0))]
135
#[allow(clippy::too_many_arguments)]
136
137
138
fn register_llm<'p>(
    py: Python<'p>,
    model_type: ModelType,
139
140
141
    endpoint: Endpoint,
    model_path: &str,
    model_name: Option<&str>,
142
143
    context_length: Option<u32>,
    kv_cache_block_size: Option<u32>,
144
    router_mode: Option<RouterMode>,
145
    migration_limit: u32,
146
147
148
149
150
) -> 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,
151
        ModelType::Embedding => llm_rs::model_type::ModelType::Embedding,
152
153
    };

154
155
    let inner_path = model_path.to_string();
    let model_name = model_name.map(|n| n.to_string());
156
157
158
    let router_mode = router_mode.unwrap_or(RouterMode::RoundRobin);
    let router_config = RouterConfig::new(router_mode.into(), KvRouterConfig::default());

159
    pyo3_async_runtimes::tokio::future_into_py(py, async move {
160
161
162
163
164
        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)
165
            .kv_cache_block_size(kv_cache_block_size)
166
167
            .router_config(Some(router_config))
            .migration_limit(Some(migration_limit));
168
        // Download from HF, load the ModelDeploymentCard
169
        let mut local_model = builder.build().await.map_err(to_pyerr)?;
170
171
172
173
174
175
176
177
178
179
        // Advertise ourself on etcd so ingress can find us
        local_model
            .attach(&endpoint.inner, model_type_obj)
            .await
            .map_err(to_pyerr)?;

        Ok(())
    })
}

180
181
182
183
184
185
#[pyclass]
#[derive(Clone)]
struct EtcdKvCache {
    inner: Arc<rs::transports::etcd::KvCache>,
}

186
187
188
189
190
191
192
#[pyclass]
#[derive(Clone)]
struct DistributedRuntime {
    inner: rs::DistributedRuntime,
    event_loop: PyObject,
}

193
194
195
196
197
198
#[pyclass]
#[derive(Clone)]
struct EtcdClient {
    inner: rs::transports::etcd::Client,
}

199
#[pyclass]
200
#[derive(Clone)]
201
202
203
204
205
struct CancellationToken {
    inner: rs::CancellationToken,
}

#[pyclass]
206
#[derive(Clone)]
207
208
209
210
211
212
struct Namespace {
    inner: rs::component::Namespace,
    event_loop: PyObject,
}

#[pyclass]
213
#[derive(Clone)]
214
215
216
217
218
219
struct Component {
    inner: rs::component::Component,
    event_loop: PyObject,
}

#[pyclass]
220
#[derive(Clone)]
221
222
223
224
225
226
struct Endpoint {
    inner: rs::component::Endpoint,
    event_loop: PyObject,
}

#[pyclass]
227
#[derive(Clone)]
228
struct Client {
229
    router: rs::pipeline::PushRouter<serde_json::Value, RsAnnotated<serde_json::Value>>,
230
231
}

232
233
234
235
236
237
238
#[pyclass(eq, eq_int)]
#[derive(Clone, PartialEq)]
#[repr(i32)]
enum ModelType {
    Chat = 1,
    Completion = 2,
    Backend = 3,
239
    Embedding = 4,
240
241
}

242
243
244
#[pymethods]
impl DistributedRuntime {
    #[new]
245
    fn new(event_loop: PyObject, is_static: bool) -> PyResult<Self> {
246
        let worker = rs::Worker::from_settings().map_err(to_pyerr)?;
247
        INIT.get_or_try_init(|| {
248
            let primary = worker.tokio_runtime()?;
249
250
251
252
253
254
            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)?;

255
        let runtime = worker.runtime().clone();
256

257
258
259
260
261
262
263
264
265
266
267
        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)?;
268
269
270
271
272
273
274
275
276
277
278

        Ok(DistributedRuntime { inner, event_loop })
    }

    fn namespace(&self, name: String) -> PyResult<Namespace> {
        Ok(Namespace {
            inner: self.inner.namespace(name).map_err(to_pyerr)?,
            event_loop: self.event_loop.clone(),
        })
    }

279
280
281
282
283
    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),
        }
284
285
    }

286
287
288
289
290
291
292
293
294
    fn shutdown(&self) {
        self.inner.runtime().shutdown();
    }

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

295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
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
#[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(())
        })
    }
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

    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(())
        })
    }
440
441
}

442
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
#[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 {
478
    #[pyo3(signature = (generator))]
479
480
481
482
483
484
485
486
487
488
    fn serve_endpoint<'p>(
        &self,
        py: Python<'p>,
        generator: PyObject,
    ) -> 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)?;
489
        let builder = self.inner.endpoint_builder().handler(ingress);
490
491
492
493
494
495
        pyo3_async_runtimes::tokio::future_into_py(py, async move {
            builder.start().await.map_err(to_pyerr)?;
            Ok(())
        })
    }

496
497
498
    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 {
499
            let client = inner.client().await.map_err(to_pyerr)?;
500
501
502
503
504
505
            let push_router = rs::pipeline::PushRouter::<
                serde_json::Value,
                RsAnnotated<serde_json::Value>,
            >::from_client(client, Default::default())
            .await
            .map_err(to_pyerr)?;
506
507
508
            Ok(Client {
                router: push_router,
            })
509
510
        })
    }
511
512

    fn lease_id(&self) -> i64 {
513
514
515
516
517
        self.inner
            .drt()
            .primary_lease()
            .map(|l| l.id())
            .unwrap_or(0)
518
    }
519
520
521
522
523
524
525
526
527
528
529
530
531
}

#[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(),
        })
    }
}

532
533
#[pymethods]
impl EtcdClient {
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
    #[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
                .kv_create(key, value, lease_id)
                .await
                .map_err(to_pyerr)?;
            Ok(())
        })
    }

552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
    #[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(())
        })
    }

570
571
572
573
    fn primary_lease_id(&self) -> i64 {
        self.inner.lease_id()
    }

574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
    #[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)
        })
    }
619
620
621
622
623
624
625
626

    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(())
        })
    }
627
628
}

629
630
#[pymethods]
impl Client {
631
632
633
634
    /// Get list of current instances.
    /// Replaces endpoint_ids.
    fn instance_ids(&self) -> Vec<i64> {
        self.router.client.instance_ids()
635
636
    }

637
638
639
    /// 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>> {
640
        let inner = self.router.client.clone();
641
        pyo3_async_runtimes::tokio::future_into_py(py, async move {
642
            inner
643
                .wait_for_instances()
644
645
646
                .await
                .map(|v| v.into_iter().map(|cei| cei.id()).collect::<Vec<i64>>())
                .map_err(to_pyerr)
647
648
649
650
651
652
653
654
655
656
657
        })
    }

    /// 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>> {
658
        if self.router.client.is_static() {
659
660
661
662
            self.r#static(py, request, annotated)
        } else {
            self.random(py, request, annotated)
        }
663
664
665
666
667
668
669
670
671
672
673
674
675
676
    }

    /// 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);
677
        let client = self.router.clone();
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700

        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);
701
        let client = self.router.clone();
702
703
704
705
706
707
708
709
710
711
712
713

        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.
714
    #[pyo3(signature = (request, instance_id, annotated=DEFAULT_ANNOTATED_SETTING))]
715
716
717
718
    fn direct<'p>(
        &self,
        py: Python<'p>,
        request: PyObject,
719
        instance_id: i64,
720
721
722
723
724
725
        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);
726
        let client = self.router.clone();
727
728
729

        pyo3_async_runtimes::tokio::future_into_py(py, async move {
            let stream = client
730
                .direct(request.into(), instance_id)
731
732
733
734
735
736
737
738
739
740
741
                .await
                .map_err(to_pyerr)?;

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

            Ok(AsyncResponseStream {
                rx: Arc::new(Mutex::new(rx)),
                annotated,
            })
        })
    }
742
743
744
745
746
747
748
749
750
751
752
753
754

    /// 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);
755
        let client = self.router.clone();
756
757
758
759
760
761
762
763
764
765
766
767

        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,
            })
        })
    }
768
769
770
}

async fn process_stream(
771
    stream: EngineStream<RsAnnotated<serde_json::Value>>,
772
773
774
775
776
    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
777
        let annotated: RsAnnotated<serde_json::Value> = response;
778
779
780
781
782
783
784
785
786
787
788
789
        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 {
790
            tracing::error!("Failed to send response: {:?}", e);
791
            break;
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
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
        }

        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")
        )
    }
}