lib.rs 39.7 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
use pyo3::IntoPyObjectExt;
use pyo3::{exceptions::PyException, prelude::*};
11
use rand::seq::IteratorRandom as _;
12
use rs::pipeline::network::Ingress;
13
use std::net::{IpAddr, Ipv4Addr, SocketAddr, SocketAddrV4};
14
use std::path::PathBuf;
15
use std::time::Duration;
16
17
18
use std::{fmt::Display, sync::Arc};
use tokio::sync::Mutex;

Neelay Shah's avatar
Neelay Shah committed
19
use dynamo_runtime::{
Ryan Olson's avatar
Ryan Olson committed
20
    self as rs, logging,
21
    pipeline::{
22
23
        context::Context as RsContext, network::egress::push_router::RouterMode as RsRouterMode,
        EngineStream, ManyOut, SingleIn,
24
    },
25
    protocols::annotated::Annotated as RsAnnotated,
26
    traits::DistributedRuntimeProvider,
27
28
};

Neelay Shah's avatar
Neelay Shah committed
29
use dynamo_llm::{self as llm_rs};
30
31
use dynamo_llm::{entrypoint::RouterConfig, kv_router::KvRouterConfig};

32
33
use crate::llm::local_model::ModelRuntimeConfig;

34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
#[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,
        }
    }
}
51

52
mod context;
53
mod engine;
54
mod http;
55
mod llm;
56
mod parsers;
57
58
59
60
61
62
63
64
65
66
67
68
69

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
70
    logging::init();
Yan Ru Pei's avatar
Yan Ru Pei committed
71
    m.add_function(wrap_pyfunction!(llm::kv::compute_block_hash_for_seq_py, m)?)?;
Ryan Olson's avatar
Ryan Olson committed
72
    m.add_function(wrap_pyfunction!(log_message, m)?)?;
73
    m.add_function(wrap_pyfunction!(register_llm, m)?)?;
74
75
    m.add_function(wrap_pyfunction!(llm::entrypoint::make_engine, m)?)?;
    m.add_function(wrap_pyfunction!(llm::entrypoint::run_input, m)?)?;
76

77
78
79
80
81
82
    m.add_class::<DistributedRuntime>()?;
    m.add_class::<CancellationToken>()?;
    m.add_class::<Namespace>()?;
    m.add_class::<Component>()?;
    m.add_class::<Endpoint>()?;
    m.add_class::<Client>()?;
83
    m.add_class::<EtcdClient>()?;
84
    m.add_class::<AsyncResponseStream>()?;
85
    m.add_class::<llm::disagg_router::DisaggregatedRouter>()?;
86
87
88
    m.add_class::<llm::entrypoint::EntrypointArgs>()?;
    m.add_class::<llm::entrypoint::EngineConfig>()?;
    m.add_class::<llm::entrypoint::EngineType>()?;
89
90
    m.add_class::<llm::entrypoint::RouterConfig>()?;
    m.add_class::<llm::entrypoint::KvRouterConfig>()?;
91
    m.add_class::<llm::kv::WorkerMetricsPublisher>()?;
92
    m.add_class::<llm::model_card::ModelDeploymentCard>()?;
93
    m.add_class::<llm::local_model::ModelRuntimeConfig>()?;
94
95
    m.add_class::<llm::preprocessor::OAIChatPreprocessor>()?;
    m.add_class::<llm::backend::Backend>()?;
96
97
    m.add_class::<llm::kv::OverlapScores>()?;
    m.add_class::<llm::kv::KvIndexer>()?;
98
    m.add_class::<llm::kv::ApproxKvIndexer>()?;
99
100
101
    m.add_class::<llm::kv::EndpointKvMetrics>()?;
    m.add_class::<llm::kv::AggregatedMetrics>()?;
    m.add_class::<llm::kv::KvMetricsAggregator>()?;
102
    m.add_class::<llm::kv::KvEventPublisher>()?;
Yan Ru Pei's avatar
Yan Ru Pei committed
103
104
    m.add_class::<llm::kv::RadixTree>()?;
    m.add_class::<llm::kv::ZmqKvEventListener>()?;
105
106
    m.add_class::<llm::kv::ZmqKvEventPublisher>()?;
    m.add_class::<llm::kv::ZmqKvEventPublisherConfig>()?;
107
    m.add_class::<llm::kv::KvRecorder>()?;
108
109
110
    m.add_class::<http::HttpService>()?;
    m.add_class::<http::HttpError>()?;
    m.add_class::<http::HttpAsyncEngine>()?;
111
    m.add_class::<context::Context>()?;
112
    m.add_class::<EtcdKvCache>()?;
113
    m.add_class::<ModelType>()?;
114
    m.add_class::<ModelInput>()?;
115
116
117
118
    m.add_class::<llm::kv::ForwardPassMetrics>()?;
    m.add_class::<llm::kv::WorkerStats>()?;
    m.add_class::<llm::kv::KvStats>()?;
    m.add_class::<llm::kv::SpecDecodeStats>()?;
119
120
    m.add_class::<llm::kv::KvPushRouter>()?;
    m.add_class::<llm::kv::KvPushRouterStream>()?;
121
    m.add_class::<RouterMode>()?;
122
123

    engine::add_to_module(m)?;
124
    parsers::add_to_module(m)?;
125

126
127
128
    #[cfg(feature = "block-manager")]
    llm::block_manager::add_to_module(m)?;

129
130
131
132
133
134
135
136
137
138
    Ok(())
}

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

Ryan Olson's avatar
Ryan Olson committed
139
140
141
142
143
144
145
/// 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);
}

146
#[pyfunction]
147
#[pyo3(signature = (model_input, 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, custom_template_path=None))]
148
#[allow(clippy::too_many_arguments)]
149
150
fn register_llm<'p>(
    py: Python<'p>,
151
    model_input: ModelInput,
152
    model_type: ModelType,
153
154
155
    endpoint: Endpoint,
    model_path: &str,
    model_name: Option<&str>,
156
157
    context_length: Option<u32>,
    kv_cache_block_size: Option<u32>,
158
    router_mode: Option<RouterMode>,
159
    migration_limit: u32,
160
    runtime_config: Option<ModelRuntimeConfig>,
161
    user_data: Option<&Bound<'p, PyDict>>,
162
    custom_template_path: Option<&str>,
163
) -> PyResult<Bound<'p, PyAny>> {
164
165
166
    let model_input = match model_input {
        ModelInput::Text => llm_rs::model_type::ModelInput::Text,
        ModelInput::Tokens => llm_rs::model_type::ModelInput::Tokens,
167
168
    };

169
170
    let model_type_obj = model_type.inner;

171
172
    let inner_path = model_path.to_string();
    let model_name = model_name.map(|n| n.to_string());
173
174
175
    let router_mode = router_mode.unwrap_or(RouterMode::RoundRobin);
    let router_config = RouterConfig::new(router_mode.into(), KvRouterConfig::default());

176
177
178
179
180
181
182
183
184
185
186
187
188
    // Early validation of custom template path
    let custom_template_path_owned = custom_template_path
        .map(|s| {
            let path = PathBuf::from(s);
            if !path.exists() {
                return Err(PyErr::new::<pyo3::exceptions::PyFileNotFoundError, _>(
                    format!("Custom template file does not exist: {}", path.display()),
                ));
            }
            Ok(path)
        })
        .transpose()?;

189
190
191
192
193
194
195
    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))
        })?;

196
    pyo3_async_runtimes::tokio::future_into_py(py, async move {
197
198
199
200
201
        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)
202
            .kv_cache_block_size(kv_cache_block_size)
203
            .router_config(Some(router_config))
204
            .migration_limit(Some(migration_limit))
205
            .runtime_config(runtime_config.unwrap_or_default().inner)
206
207
            .user_data(user_data_json)
            .custom_template_path(custom_template_path_owned);
208
        // Download from HF, load the ModelDeploymentCard
209
        let mut local_model = builder.build().await.map_err(to_pyerr)?;
210
211
        // Advertise ourself on etcd so ingress can find us
        local_model
212
            .attach(&endpoint.inner, model_type_obj, model_input)
213
214
215
216
217
218
219
            .await
            .map_err(to_pyerr)?;

        Ok(())
    })
}

220
221
222
223
224
225
#[pyclass]
#[derive(Clone)]
struct EtcdKvCache {
    inner: Arc<rs::transports::etcd::KvCache>,
}

226
227
#[pyclass]
#[derive(Clone)]
Ryan Olson's avatar
Ryan Olson committed
228
pub struct DistributedRuntime {
229
230
231
232
    inner: rs::DistributedRuntime,
    event_loop: PyObject,
}

Ryan Olson's avatar
Ryan Olson committed
233
234
235
236
237
238
239
impl DistributedRuntime {
    #[allow(dead_code)]
    fn inner(&self) -> &rs::DistributedRuntime {
        &self.inner
    }
}

240
241
242
243
244
245
#[pyclass]
#[derive(Clone)]
struct EtcdClient {
    inner: rs::transports::etcd::Client,
}

246
#[pyclass]
247
#[derive(Clone)]
248
249
250
251
252
struct CancellationToken {
    inner: rs::CancellationToken,
}

#[pyclass]
253
#[derive(Clone)]
254
255
256
257
258
259
struct Namespace {
    inner: rs::component::Namespace,
    event_loop: PyObject,
}

#[pyclass]
260
#[derive(Clone)]
261
262
263
264
265
266
struct Component {
    inner: rs::component::Component,
    event_loop: PyObject,
}

#[pyclass]
267
#[derive(Clone)]
268
269
270
271
272
273
struct Endpoint {
    inner: rs::component::Endpoint,
    event_loop: PyObject,
}

#[pyclass]
274
#[derive(Clone)]
275
struct Client {
276
    router: rs::pipeline::PushRouter<serde_json::Value, RsAnnotated<serde_json::Value>>,
277
278
}

279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
#[pyclass]
#[derive(Clone, PartialEq)]
struct ModelType {
    inner: llm_rs::model_type::ModelType,
}

#[pymethods]
#[allow(non_upper_case_globals)]
impl ModelType {
    #[classattr]
    const Chat: Self = ModelType {
        inner: llm_rs::model_type::ModelType::Chat,
    };
    #[classattr]
    const Completions: Self = ModelType {
        inner: llm_rs::model_type::ModelType::Completions,
    };
    #[classattr]
    const Embedding: Self = ModelType {
        inner: llm_rs::model_type::ModelType::Embedding,
    };

    fn __or__(&self, other: &Self) -> Self {
        ModelType {
            inner: self.inner | other.inner,
        }
    }

    fn __str__(&self) -> String {
        self.inner.to_string()
    }
}

312
313
#[pyclass(eq, eq_int)]
#[derive(Clone, PartialEq)]
314
315
316
enum ModelInput {
    Text = 1,
    Tokens = 2,
317
318
}

319
320
321
#[pymethods]
impl DistributedRuntime {
    #[new]
322
    fn new(event_loop: PyObject, is_static: bool) -> PyResult<Self> {
323
        let worker = rs::Worker::from_settings().map_err(to_pyerr)?;
324
        INIT.get_or_try_init(|| {
325
            let primary = worker.tokio_runtime()?;
326
327
328
329
330
331
            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)?;

332
        let runtime = worker.runtime().clone();
333

334
335
336
337
338
339
340
341
342
343
344
        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)?;
345
346
347
348

        Ok(DistributedRuntime { inner, event_loop })
    }

Ryan Olson's avatar
Ryan Olson committed
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
    #[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(),
        })
    }

364
365
366
367
368
369
370
371
372
    /// Remove everything in an etcd namespace.
    /// Will be removed once we can clear the MDC automatically.
    fn temp_clear_namespace<'p>(&self, py: Python<'p>, name: String) -> PyResult<Bound<'p, PyAny>> {
        let inner = self.inner.clone();
        pyo3_async_runtimes::tokio::future_into_py(py, async move {
            inner.temp_clear_namespace(&name).await.map_err(to_pyerr)
        })
    }

373
374
375
376
377
378
379
    fn namespace(&self, name: String) -> PyResult<Namespace> {
        Ok(Namespace {
            inner: self.inner.namespace(name).map_err(to_pyerr)?,
            event_loop: self.event_loop.clone(),
        })
    }

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
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
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
    /// Allocate a contiguous block of ports from the specified range and atomically reserve them.
    /// Returns a list of all allocated ports in order.
    #[pyo3(signature = (namespace, port_min, port_max, block_size, context=None))]
    fn allocate_port_block<'p>(
        &self,
        py: Python<'p>,
        namespace: &str,
        port_min: u16,
        port_max: u16,
        block_size: u16,
        context: Option<String>, // Optional info to store alongside the reservation
    ) -> PyResult<Bound<'p, PyAny>> {
        const MAX_ALLOCATE_ATTEMPTS: usize = 100;
        if block_size == 0 {
            return Err(PyErr::new::<pyo3::exceptions::PyValueError, _>(
                "Block size must be at least 1",
            ));
        }

        let Some(etcd_client) = self.inner.etcd_client() else {
            return Err(PyErr::new::<PyException, _>(
                "Static workers should not need to reserve ports",
            ));
        };

        let min = port_min;
        let max = port_max;

        // Compute maximum valid starting port (inclusive)
        let max_start_port = max.saturating_sub(block_size.saturating_sub(1));
        if max_start_port < min {
            return Err(PyErr::new::<PyException, _>(format!(
                "Port range {min}-{max} is too small for block size {block_size}",
            )));
        }

        // Randomize candidate starting ports to reduce contention/races
        let candidate_count =
            (max_start_port - port_min + 1).min(MAX_ALLOCATE_ATTEMPTS as u16) as usize;
        let mut rng = rand::rng();
        let candidate_ports: Vec<u16> =
            (port_min..=max_start_port).choose_multiple(&mut rng, candidate_count);

        let local_ip = match local_ip() {
            Ok(ip) => ip,
            Err(err) => {
                return Err(PyErr::new::<PyException, _>(format!(
                    "Failed fetching local IP address: {err}"
                )));
            }
        };

        let context_bytes = context.map(|s| s.as_bytes().to_vec()).unwrap_or_default();
        let namespace = namespace.to_owned();

        pyo3_async_runtimes::tokio::future_into_py(py, async move {
            for (attempt_idx, start_port) in candidate_ports.into_iter().enumerate() {
                let end_port_exclusive = start_port + block_size;
                let ports_to_reserve: Vec<u16> = (start_port..end_port_exclusive).collect();

                // Hold/bind all ports in the block
                let mut sockets = Vec::with_capacity(ports_to_reserve.len());
                let mut bind_failed = false;

                for &port in &ports_to_reserve {
                    match bind_tcp_port(port) {
                        Ok(sock) => sockets.push(sock),
                        Err(e) => {
                            tracing::error!(
                            "Failed to bind to port block starting at {start_port} (attempt {}): {e}",
                            attempt_idx + 1,
                        );
                            bind_failed = true;
                            break;
                        }
                    }
                }

                if bind_failed {
                    // Let previously bound sockets drop here
                    if attempt_idx < candidate_count - 1 {
                        tokio::time::sleep(Duration::from_millis(10)).await;
                    }
                    continue;
                }

                // With sockets held, reserve in ETCD
                let mut reserved_keys = Vec::with_capacity(ports_to_reserve.len());
                let mut reservation_failed = false;
                for port in &ports_to_reserve {
                    let key = make_port_key(&namespace, local_ip, *port).map_err(to_pyerr)?;

                    if let Err(e) = etcd_client
                        .kv_create(&key, context_bytes.clone(), None)
                        .await
                    {
                        tracing::error!(
                            "Failed to reserve port block starting at {start_port} (attempt {}): {e}",
                            attempt_idx + 1,
                        );
                        reservation_failed = true;
                        break;
                    }
                    reserved_keys.push(key);
                }

                if reservation_failed {
                    // Cleanup partial reservations
                    for key in reserved_keys {
                        if let Err(e) = etcd_client.kv_delete(key.as_str(), None).await {
                            tracing::warn!("Failed to cleanup reserved port {key}: {e}");
                        }
                    }

                    // Sockets automatically released via RAII
                    if attempt_idx < candidate_count - 1 {
                        tokio::time::sleep(Duration::from_millis(10)).await;
                    }
                    continue;
                }

                // Success - sockets will be released automatically
                tracing::debug!("Reserved port block {ports_to_reserve:?}");
                return Ok(ports_to_reserve);
            }

            Err(PyErr::new::<PyException, _>(format!(
            "Failed to allocate and reserve a port block of size {block_size} from range {min}-{max} after {candidate_count} attempts")))
        })
    }

511
    fn do_not_use_etcd_client(&self) -> PyResult<Option<EtcdClient>> {
512
513
514
515
        match self.inner.etcd_client().clone() {
            Some(etcd_client) => Ok(Some(EtcdClient { inner: etcd_client })),
            None => Ok(None),
        }
516
517
    }

518
519
520
521
522
523
524
525
526
    fn shutdown(&self) {
        self.inner.runtime().shutdown();
    }

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

527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
// Bind a TCP port and return a socket held until dropped.
fn bind_tcp_port(port: u16) -> std::io::Result<socket2::Socket> {
    let sock = socket2::Socket::new(
        socket2::Domain::IPV4,
        socket2::Type::STREAM,
        Some(socket2::Protocol::TCP),
    )?;
    sock.set_reuse_address(true)?;
    let addr = SocketAddr::V4(SocketAddrV4::new(Ipv4Addr::UNSPECIFIED, port));
    sock.bind(&addr.into())?;
    Ok(sock)
}

fn make_port_key(namespace: &str, node_ip: IpAddr, port: u16) -> anyhow::Result<String> {
    Ok(format!("dyn://{namespace}/ports/{node_ip}/{port}"))
}

fn local_ip() -> Result<IpAddr, local_ip_address::Error> {
    local_ip_address::local_ip().or_else(|err| match err {
        local_ip_address::Error::LocalIpAddressNotFound => {
            // Fall back to IPv6 if no IPv4 addresses are found
            local_ip_address::local_ipv6()
        }
        _ => Err(err),
    })
}

554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
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
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
663
#[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(())
        })
    }
664
665
666
667
668
669
670
671
672

    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(())
        })
    }
673
674
}

675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
#[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 {
711
    #[pyo3(signature = (generator, graceful_shutdown = true, metrics_labels = None, health_check_payload = None))]
712
713
714
715
    fn serve_endpoint<'p>(
        &self,
        py: Python<'p>,
        generator: PyObject,
716
        graceful_shutdown: Option<bool>,
717
        metrics_labels: Option<Vec<(String, String)>>,
718
        health_check_payload: Option<&Bound<'p, PyDict>>,
719
720
721
722
723
724
    ) -> 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)?;
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746

        // Convert Python dict to serde_json::Value if provided and validate it's an object
        let health_payload_json = health_check_payload
            .map(|dict| pythonize::depythonize::<serde_json::Value>(dict))
            .transpose()
            .map_err(|err| {
                pyo3::exceptions::PyTypeError::new_err(format!(
                    "Failed to convert health_check_payload: {}",
                    err
                ))
            })?;

        // Require an object/dict
        if let Some(ref payload) = health_payload_json {
            if !payload.is_object() {
                return Err(pyo3::exceptions::PyTypeError::new_err(
                    "health_check_payload must be a JSON object (dict)",
                ));
            }
        }

        let mut builder = self
747
748
749
750
            .inner
            .endpoint_builder()
            .metrics_labels(metrics_labels)
            .handler(ingress);
751
752
753
754
755

        if let Some(payload) = health_payload_json {
            builder = builder.health_check_payload(payload);
        }

756
        let graceful_shutdown = graceful_shutdown.unwrap_or(true);
757
        pyo3_async_runtimes::tokio::future_into_py(py, async move {
758
759
760
761
762
            builder
                .graceful_shutdown(graceful_shutdown)
                .start()
                .await
                .map_err(to_pyerr)?;
763
764
765
766
            Ok(())
        })
    }

767
768
769
    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 {
770
            let client = inner.client().await.map_err(to_pyerr)?;
771
772
773
774
775
776
            let push_router = rs::pipeline::PushRouter::<
                serde_json::Value,
                RsAnnotated<serde_json::Value>,
            >::from_client(client, Default::default())
            .await
            .map_err(to_pyerr)?;
777
778
779
            Ok(Client {
                router: push_router,
            })
780
781
        })
    }
782
783

    fn lease_id(&self) -> i64 {
784
785
786
787
788
        self.inner
            .drt()
            .primary_lease()
            .map(|l| l.id())
            .unwrap_or(0)
789
    }
790
791
792
793
794
795
796
797
798
799
800
801
802
}

#[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(),
        })
    }
}

803
804
#[pymethods]
impl EtcdClient {
805
806
807
808
809
810
811
812
813
814
815
    #[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
816
                .kv_create(&key, value, lease_id)
817
818
819
820
821
822
                .await
                .map_err(to_pyerr)?;
            Ok(())
        })
    }

823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
    #[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(())
        })
    }

841
842
843
844
    fn primary_lease_id(&self) -> i64 {
        self.inner.lease_id()
    }

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
    #[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)
        })
    }
890
891
892
893
894
895
896
897

    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(())
        })
    }
898
899
}

900
901
#[pymethods]
impl Client {
902
903
904
905
    /// Get list of current instances.
    /// Replaces endpoint_ids.
    fn instance_ids(&self) -> Vec<i64> {
        self.router.client.instance_ids()
906
907
    }

908
909
910
    /// 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>> {
911
        let inner = self.router.client.clone();
912
        pyo3_async_runtimes::tokio::future_into_py(py, async move {
913
            inner
914
                .wait_for_instances()
915
916
917
                .await
                .map(|v| v.into_iter().map(|cei| cei.id()).collect::<Vec<i64>>())
                .map_err(to_pyerr)
918
919
920
921
        })
    }

    /// Issue a request to the endpoint using the default routing strategy.
922
    #[pyo3(signature = (request, annotated=DEFAULT_ANNOTATED_SETTING, context=None))]
923
924
925
926
927
    fn generate<'p>(
        &self,
        py: Python<'p>,
        request: PyObject,
        annotated: Option<bool>,
928
        context: Option<context::Context>,
929
    ) -> PyResult<Bound<'p, PyAny>> {
930
        if self.router.client.is_static() {
931
            self.r#static(py, request, annotated, context)
932
        } else {
933
            self.random(py, request, annotated, context)
934
        }
935
936
937
    }

    /// Send a request to the next endpoint in a round-robin fashion.
938
    #[pyo3(signature = (request, annotated=DEFAULT_ANNOTATED_SETTING, context=None))]
939
940
941
942
943
    fn round_robin<'p>(
        &self,
        py: Python<'p>,
        request: PyObject,
        annotated: Option<bool>,
944
        context: Option<context::Context>,
945
946
947
948
949
    ) -> 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);
950
        let client = self.router.clone();
951
952

        pyo3_async_runtimes::tokio::future_into_py(py, async move {
953
954
955
956
957
958
959
960
961
            let stream = match context {
                Some(context) => {
                    let request = RsContext::with_id(request, context.inner().id().to_string());
                    let stream = client.round_robin(request).await.map_err(to_pyerr)?;
                    context.inner().link_child(stream.context());
                    stream
                }
                _ => client.round_robin(request.into()).await.map_err(to_pyerr)?,
            };
962
963
964
965
966
967
968
969
970
            tokio::spawn(process_stream(stream, tx));
            Ok(AsyncResponseStream {
                rx: Arc::new(Mutex::new(rx)),
                annotated,
            })
        })
    }

    /// Send a request to a random endpoint.
971
    #[pyo3(signature = (request, annotated=DEFAULT_ANNOTATED_SETTING, context=None))]
972
973
974
975
976
    fn random<'p>(
        &self,
        py: Python<'p>,
        request: PyObject,
        annotated: Option<bool>,
977
        context: Option<context::Context>,
978
979
980
981
982
    ) -> 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);
983
        let client = self.router.clone();
984
985

        pyo3_async_runtimes::tokio::future_into_py(py, async move {
986
987
988
989
990
991
992
993
994
            let stream = match context {
                Some(context) => {
                    let request = RsContext::with_id(request, context.inner().id().to_string());
                    let stream = client.random(request).await.map_err(to_pyerr)?;
                    context.inner().link_child(stream.context());
                    stream
                }
                _ => client.random(request.into()).await.map_err(to_pyerr)?,
            };
995
996
997
998
999
1000
1001
1002
1003
            tokio::spawn(process_stream(stream, tx));
            Ok(AsyncResponseStream {
                rx: Arc::new(Mutex::new(rx)),
                annotated,
            })
        })
    }

    /// Directly send a request to a specific endpoint.
1004
    #[pyo3(signature = (request, instance_id, annotated=DEFAULT_ANNOTATED_SETTING, context=None))]
1005
1006
1007
1008
    fn direct<'p>(
        &self,
        py: Python<'p>,
        request: PyObject,
1009
        instance_id: i64,
1010
        annotated: Option<bool>,
1011
        context: Option<context::Context>,
1012
1013
1014
1015
1016
    ) -> 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);
1017
        let client = self.router.clone();
1018
1019

        pyo3_async_runtimes::tokio::future_into_py(py, async move {
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
            let stream = match context {
                Some(context) => {
                    let request = RsContext::with_id(request, context.inner().id().to_string());
                    let stream = client
                        .direct(request, instance_id)
                        .await
                        .map_err(to_pyerr)?;
                    context.inner().link_child(stream.context());
                    stream
                }
                _ => client
                    .direct(request.into(), instance_id)
                    .await
                    .map_err(to_pyerr)?,
            };
1035
1036
1037
1038
1039
1040
1041
1042
1043

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

            Ok(AsyncResponseStream {
                rx: Arc::new(Mutex::new(rx)),
                annotated,
            })
        })
    }
1044
1045

    /// Directly send a request to a pre-defined static worker
1046
    #[pyo3(signature = (request, annotated=DEFAULT_ANNOTATED_SETTING, context=None))]
1047
1048
1049
1050
1051
    fn r#static<'p>(
        &self,
        py: Python<'p>,
        request: PyObject,
        annotated: Option<bool>,
1052
        context: Option<context::Context>,
1053
1054
1055
1056
1057
    ) -> 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);
1058
        let client = self.router.clone();
1059
1060

        pyo3_async_runtimes::tokio::future_into_py(py, async move {
1061
1062
1063
1064
1065
1066
1067
1068
1069
            let stream = match context {
                Some(context) => {
                    let request = RsContext::with_id(request, context.inner().id().to_string());
                    let stream = client.r#static(request).await.map_err(to_pyerr)?;
                    context.inner().link_child(stream.context());
                    stream
                }
                _ => client.r#static(request.into()).await.map_err(to_pyerr)?,
            };
1070
1071
1072
1073
1074
1075
1076
1077
1078

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

            Ok(AsyncResponseStream {
                rx: Arc::new(Mutex::new(rx)),
                annotated,
            })
        })
    }
1079
1080
1081
}

async fn process_stream(
1082
    stream: EngineStream<RsAnnotated<serde_json::Value>>,
1083
1084
1085
1086
1087
    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
1088
        let annotated: RsAnnotated<serde_json::Value> = response;
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
        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 {
1101
            tracing::error!("Failed to send response: {:?}", e);
1102
            break;
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
        }

        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")
        )
    }
}