lib.rs 35.9 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
use futures::StreamExt;
use once_cell::sync::OnceCell;
6
use pyo3::IntoPyObjectExt;
7
use pyo3::exceptions::PyStopAsyncIteration;
8
use pyo3::types::{PyDict, PyString};
9
use pyo3::{exceptions::PyException, prelude::*};
10
use rand::seq::IteratorRandom as _;
11
use rs::pipeline::network::Ingress;
12
use std::net::{IpAddr, Ipv4Addr, SocketAddr, SocketAddrV4};
13
use std::path::PathBuf;
14
use std::time::Duration;
15
16
use std::{fmt::Display, sync::Arc};
use tokio::sync::Mutex;
17
use tracing::{Instrument, info_span};
18

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
        EngineStream, ManyOut, SingleIn, context::Context as RsContext,
        network::egress::push_router::RouterMode as RsRouterMode,
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
mod planner;
58
mod prometheus_metrics;
59
60
61
62
63
64
65
66

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);

67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
// Helper to create client span - always emit spans for consistency
fn create_client_span(
    operation: &str,
    request_id: &str,
    trace_context: Option<&dynamo_runtime::logging::DistributedTraceContext>,
) -> tracing::Span {
    if let Some(ctx) = trace_context {
        info_span!(
            "client_request",
            operation = operation,
            request_id = request_id,
            trace_id = ctx.trace_id.as_str(),
            parent_id = ctx.span_id.as_str(),
            x_request_id = ctx.x_request_id.as_deref().unwrap_or(""),
            x_dynamo_request_id = ctx.x_dynamo_request_id.as_deref().unwrap_or(""),
            tracestate = ctx.tracestate.as_deref().unwrap_or("")
        )
    } else {
        info_span!(
            "client_request",
            operation = operation,
            request_id = request_id,
        )
    }
}

// Helper to get appropriate span for instrumentation - always emit spans
fn get_span_for_context(context: &context::Context, operation: &str) -> tracing::Span {
    create_client_span(operation, context.inner().id(), context.trace_context())
}

// Helper to create span for direct method with instance_id
fn get_span_for_direct_context(
    context: &context::Context,
    operation: &str,
    instance_id: &str,
) -> tracing::Span {
    if let Some(trace_ctx) = context.trace_context() {
        info_span!(
            "client_request",
            operation = operation,
            request_id = context.inner().id(),
            instance_id = instance_id,
            trace_id = trace_ctx.trace_id.as_str(),
            parent_id = trace_ctx.span_id.as_str(),
            x_request_id = trace_ctx.x_request_id.as_deref().unwrap_or(""),
            x_dynamo_request_id = trace_ctx.x_dynamo_request_id.as_deref().unwrap_or(""),
            tracestate = trace_ctx.tracestate.as_deref().unwrap_or("")
        )
    } else {
        info_span!(
            "client_request",
            operation = operation,
            request_id = context.inner().id(),
            instance_id = instance_id,
        )
    }
}

126
127
128
129
130
/// 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
131
    logging::init();
Yan Ru Pei's avatar
Yan Ru Pei committed
132
    m.add_function(wrap_pyfunction!(llm::kv::compute_block_hash_for_seq_py, m)?)?;
Ryan Olson's avatar
Ryan Olson committed
133
    m.add_function(wrap_pyfunction!(log_message, m)?)?;
134
    m.add_function(wrap_pyfunction!(register_llm, m)?)?;
135
136
    m.add_function(wrap_pyfunction!(llm::entrypoint::make_engine, m)?)?;
    m.add_function(wrap_pyfunction!(llm::entrypoint::run_input, m)?)?;
137

138
139
140
141
142
143
144
    m.add_class::<DistributedRuntime>()?;
    m.add_class::<CancellationToken>()?;
    m.add_class::<Namespace>()?;
    m.add_class::<Component>()?;
    m.add_class::<Endpoint>()?;
    m.add_class::<Client>()?;
    m.add_class::<AsyncResponseStream>()?;
145
    m.add_class::<llm::disagg_router::DisaggregatedRouter>()?;
146
147
148
    m.add_class::<llm::entrypoint::EntrypointArgs>()?;
    m.add_class::<llm::entrypoint::EngineConfig>()?;
    m.add_class::<llm::entrypoint::EngineType>()?;
149
150
    m.add_class::<llm::entrypoint::RouterConfig>()?;
    m.add_class::<llm::entrypoint::KvRouterConfig>()?;
151
    m.add_class::<llm::kv::WorkerMetricsPublisher>()?;
152
    m.add_class::<llm::model_card::ModelDeploymentCard>()?;
153
    m.add_class::<llm::local_model::ModelRuntimeConfig>()?;
154
155
    m.add_class::<llm::preprocessor::OAIChatPreprocessor>()?;
    m.add_class::<llm::backend::Backend>()?;
156
157
    m.add_class::<llm::kv::OverlapScores>()?;
    m.add_class::<llm::kv::KvIndexer>()?;
158
    m.add_class::<llm::kv::ApproxKvIndexer>()?;
159
160
161
    m.add_class::<llm::kv::EndpointKvMetrics>()?;
    m.add_class::<llm::kv::AggregatedMetrics>()?;
    m.add_class::<llm::kv::KvMetricsAggregator>()?;
162
    m.add_class::<llm::kv::KvEventPublisher>()?;
Yan Ru Pei's avatar
Yan Ru Pei committed
163
164
    m.add_class::<llm::kv::RadixTree>()?;
    m.add_class::<llm::kv::ZmqKvEventListener>()?;
165
166
    m.add_class::<llm::kv::ZmqKvEventPublisher>()?;
    m.add_class::<llm::kv::ZmqKvEventPublisherConfig>()?;
167
    m.add_class::<llm::kv::KvRecorder>()?;
168
169
    m.add_class::<http::HttpService>()?;
    m.add_class::<http::HttpAsyncEngine>()?;
170
    m.add_class::<context::Context>()?;
171
    m.add_class::<ModelType>()?;
172
    m.add_class::<ModelInput>()?;
173
174
175
176
    m.add_class::<llm::kv::ForwardPassMetrics>()?;
    m.add_class::<llm::kv::WorkerStats>()?;
    m.add_class::<llm::kv::KvStats>()?;
    m.add_class::<llm::kv::SpecDecodeStats>()?;
177
178
    m.add_class::<llm::kv::KvPushRouter>()?;
    m.add_class::<llm::kv::KvPushRouterStream>()?;
179
    m.add_class::<RouterMode>()?;
180
    m.add("__version__", env!("CARGO_PKG_VERSION"))?;
181
182
183
    m.add_class::<planner::VirtualConnectorCoordinator>()?;
    m.add_class::<planner::VirtualConnectorClient>()?;
    m.add_class::<planner::PlannerDecision>()?;
184
185

    engine::add_to_module(m)?;
186
    parsers::add_to_module(m)?;
187

188
189
190
191
192
    m.add_class::<prometheus_metrics::PyRuntimeMetrics>()?;
    let prometheus_metrics = PyModule::new(m.py(), "prometheus_metrics")?;
    prometheus_metrics::add_to_module(&prometheus_metrics)?;
    m.add_submodule(&prometheus_metrics)?;

193
194
195
    #[cfg(feature = "block-manager")]
    llm::block_manager::add_to_module(m)?;

196
197
198
199
200
201
202
203
204
205
    Ok(())
}

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

Ryan Olson's avatar
Ryan Olson committed
206
207
208
209
210
211
212
/// 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);
}

213
#[pyfunction]
214
#[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))]
215
#[allow(clippy::too_many_arguments)]
216
217
fn register_llm<'p>(
    py: Python<'p>,
218
    model_input: ModelInput,
219
    model_type: ModelType,
220
221
222
    endpoint: Endpoint,
    model_path: &str,
    model_name: Option<&str>,
223
224
    context_length: Option<u32>,
    kv_cache_block_size: Option<u32>,
225
    router_mode: Option<RouterMode>,
226
    migration_limit: u32,
227
    runtime_config: Option<ModelRuntimeConfig>,
228
    user_data: Option<&Bound<'p, PyDict>>,
229
    custom_template_path: Option<&str>,
230
) -> PyResult<Bound<'p, PyAny>> {
231
232
233
    let model_input = match model_input {
        ModelInput::Text => llm_rs::model_type::ModelInput::Text,
        ModelInput::Tokens => llm_rs::model_type::ModelInput::Tokens,
234
        ModelInput::Tensor => llm_rs::model_type::ModelInput::Tensor,
235
236
    };

237
238
    let model_type_obj = model_type.inner;

239
240
    let inner_path = model_path.to_string();
    let model_name = model_name.map(|n| n.to_string());
241
242
243
    let router_mode = router_mode.unwrap_or(RouterMode::RoundRobin);
    let router_config = RouterConfig::new(router_mode.into(), KvRouterConfig::default());

244
245
246
247
248
249
250
251
252
253
254
255
256
    // 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()?;

257
258
259
260
261
262
263
    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))
        })?;

264
    pyo3_async_runtimes::tokio::future_into_py(py, async move {
265
266
267
268
269
        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)
270
            .kv_cache_block_size(kv_cache_block_size)
271
            .router_config(Some(router_config))
272
            .migration_limit(Some(migration_limit))
273
            .runtime_config(runtime_config.unwrap_or_default().inner)
274
275
            .user_data(user_data_json)
            .custom_template_path(custom_template_path_owned);
276
        // Download from HF, load the ModelDeploymentCard
277
        let mut local_model = builder.build().await.map_err(to_pyerr)?;
278
279
        // Advertise ourself on etcd so ingress can find us
        local_model
280
            .attach(&endpoint.inner, model_type_obj, model_input)
281
282
283
284
285
286
287
            .await
            .map_err(to_pyerr)?;

        Ok(())
    })
}

288
289
#[pyclass]
#[derive(Clone)]
Ryan Olson's avatar
Ryan Olson committed
290
pub struct DistributedRuntime {
291
292
293
294
    inner: rs::DistributedRuntime,
    event_loop: PyObject,
}

Ryan Olson's avatar
Ryan Olson committed
295
296
impl DistributedRuntime {
    #[allow(dead_code)]
297
    pub(crate) fn inner(&self) -> &rs::DistributedRuntime {
Ryan Olson's avatar
Ryan Olson committed
298
299
300
301
        &self.inner
    }
}

302
#[pyclass]
303
#[derive(Clone)]
304
305
306
307
308
struct CancellationToken {
    inner: rs::CancellationToken,
}

#[pyclass]
309
#[derive(Clone)]
310
311
312
313
314
315
struct Namespace {
    inner: rs::component::Namespace,
    event_loop: PyObject,
}

#[pyclass]
316
#[derive(Clone)]
317
318
319
320
321
322
struct Component {
    inner: rs::component::Component,
    event_loop: PyObject,
}

#[pyclass]
323
#[derive(Clone)]
324
325
326
327
328
329
struct Endpoint {
    inner: rs::component::Endpoint,
    event_loop: PyObject,
}

#[pyclass]
330
#[derive(Clone)]
331
struct Client {
332
    router: rs::pipeline::PushRouter<serde_json::Value, RsAnnotated<serde_json::Value>>,
333
334
}

335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
#[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,
    };
356
357
358
359
    #[classattr]
    const TensorBased: Self = ModelType {
        inner: llm_rs::model_type::ModelType::TensorBased,
    };
360
361
362
363
364
365
366
367
368
369
370
371

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

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

372
373
#[pyclass(eq, eq_int)]
#[derive(Clone, PartialEq)]
374
375
376
enum ModelInput {
    Text = 1,
    Tokens = 2,
377
    Tensor = 3,
378
379
}

380
381
382
#[pymethods]
impl DistributedRuntime {
    #[new]
383
    fn new(event_loop: PyObject, is_static: bool) -> PyResult<Self> {
384
        let worker = rs::Worker::from_settings().map_err(to_pyerr)?;
385
        INIT.get_or_try_init(|| {
386
            let primary = worker.tokio_runtime()?;
387
388
389
390
391
392
            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)?;

393
        let runtime = worker.runtime().clone();
394

395
396
397
398
399
400
401
402
403
404
405
        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)?;
406
407
408
409

        Ok(DistributedRuntime { inner, event_loop })
    }

Ryan Olson's avatar
Ryan Olson committed
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
    #[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(),
        })
    }

425
426
427
428
429
430
431
    fn namespace(&self, name: String) -> PyResult<Namespace> {
        Ok(Namespace {
            inner: self.inner.namespace(name).map_err(to_pyerr)?,
            event_loop: self.event_loop.clone(),
        })
    }

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
    /// 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!(
501
502
503
                                "Failed to bind to port block starting at {start_port} (attempt {}): {e}",
                                attempt_idx + 1,
                            );
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
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
554
555
556
557
558
                            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!(
559
560
                "Failed to allocate and reserve a port block of size {block_size} from range {min}-{max} after {candidate_count} attempts"
            )))
561
562
563
        })
    }

564
565
566
567
568
569
570
    fn shutdown(&self) {
        self.inner.runtime().shutdown();
    }

    fn event_loop(&self) -> PyObject {
        self.event_loop.clone()
    }
571
572
573
574
575

    fn child_token(&self) -> CancellationToken {
        let inner = self.inner.runtime().child_token();
        CancellationToken { inner }
    }
576
577
}

578
579
580
581
582
583
584
585
586
587
588
589
590
591
// 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> {
592
    Ok(format!("v1/{namespace}/ports/{node_ip}/{port}"))
593
594
595
596
597
598
599
600
601
602
603
604
}

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),
    })
}

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
#[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(())
        })
    }
637
638
639
640
641
642

    /// Get a RuntimeMetrics helper for creating Prometheus metrics
    #[getter]
    fn metrics(&self) -> prometheus_metrics::PyRuntimeMetrics {
        prometheus_metrics::PyRuntimeMetrics::from_component(self.inner.clone())
    }
643
644
645
646
}

#[pymethods]
impl Endpoint {
647
    #[pyo3(signature = (generator, graceful_shutdown = true, metrics_labels = None, health_check_payload = None))]
648
649
650
651
    fn serve_endpoint<'p>(
        &self,
        py: Python<'p>,
        generator: PyObject,
652
        graceful_shutdown: Option<bool>,
653
        metrics_labels: Option<Vec<(String, String)>>,
654
        health_check_payload: Option<&Bound<'p, PyDict>>,
655
656
657
658
659
660
    ) -> 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)?;
661
662
663
664
665
666
667
668
669
670
671
672
673

        // 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
674
675
676
677
678
679
        if let Some(ref payload) = health_payload_json
            && !payload.is_object()
        {
            return Err(pyo3::exceptions::PyTypeError::new_err(
                "health_check_payload must be a JSON object (dict)",
            ));
680
681
682
        }

        let mut builder = self
683
684
685
686
            .inner
            .endpoint_builder()
            .metrics_labels(metrics_labels)
            .handler(ingress);
687
688
689
690
691

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

692
        let graceful_shutdown = graceful_shutdown.unwrap_or(true);
693
        pyo3_async_runtimes::tokio::future_into_py(py, async move {
694
695
696
697
698
            builder
                .graceful_shutdown(graceful_shutdown)
                .start()
                .await
                .map_err(to_pyerr)?;
699
700
701
702
            Ok(())
        })
    }

703
704
705
    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 {
706
            let client = inner.client().await.map_err(to_pyerr)?;
707
708
709
710
711
712
            let push_router = rs::pipeline::PushRouter::<
                serde_json::Value,
                RsAnnotated<serde_json::Value>,
            >::from_client(client, Default::default())
            .await
            .map_err(to_pyerr)?;
713
714
715
            Ok(Client {
                router: push_router,
            })
716
717
        })
    }
718
719

    fn lease_id(&self) -> i64 {
720
721
722
723
724
        self.inner
            .drt()
            .primary_lease()
            .map(|l| l.id())
            .unwrap_or(0)
725
    }
726
727
728
729
730
731

    /// Get a RuntimeMetrics helper for creating Prometheus metrics
    #[getter]
    fn metrics(&self) -> prometheus_metrics::PyRuntimeMetrics {
        prometheus_metrics::PyRuntimeMetrics::from_endpoint(self.inner.clone())
    }
732
733
734
735
736
737
738
739
740
741
742
}

#[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(),
        })
    }
743
744
745
746
747
748

    /// Get a RuntimeMetrics helper for creating Prometheus metrics
    #[getter]
    fn metrics(&self) -> prometheus_metrics::PyRuntimeMetrics {
        prometheus_metrics::PyRuntimeMetrics::from_namespace(self.inner.clone())
    }
749
750
751
752
}

#[pymethods]
impl Client {
753
754
755
756
    /// Get list of current instances.
    /// Replaces endpoint_ids.
    fn instance_ids(&self) -> Vec<i64> {
        self.router.client.instance_ids()
757
758
    }

759
760
761
    /// 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>> {
762
        let inner = self.router.client.clone();
763
        pyo3_async_runtimes::tokio::future_into_py(py, async move {
764
            inner
765
                .wait_for_instances()
766
767
768
                .await
                .map(|v| v.into_iter().map(|cei| cei.id()).collect::<Vec<i64>>())
                .map_err(to_pyerr)
769
770
771
772
        })
    }

    /// Issue a request to the endpoint using the default routing strategy.
773
    #[pyo3(signature = (request, annotated=DEFAULT_ANNOTATED_SETTING, context=None))]
774
775
776
777
778
    fn generate<'p>(
        &self,
        py: Python<'p>,
        request: PyObject,
        annotated: Option<bool>,
779
        context: Option<context::Context>,
780
    ) -> PyResult<Bound<'p, PyAny>> {
781
        if self.router.client.is_static() {
782
            self.r#static(py, request, annotated, context)
783
        } else {
784
            self.random(py, request, annotated, context)
785
        }
786
787
788
    }

    /// Send a request to the next endpoint in a round-robin fashion.
789
    #[pyo3(signature = (request, annotated=DEFAULT_ANNOTATED_SETTING, context=None))]
790
791
792
793
794
    fn round_robin<'p>(
        &self,
        py: Python<'p>,
        request: PyObject,
        annotated: Option<bool>,
795
        context: Option<context::Context>,
796
797
798
799
800
    ) -> 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);
801
        let client = self.router.clone();
802
803

        pyo3_async_runtimes::tokio::future_into_py(py, async move {
804
805
            let stream = match context {
                Some(context) => {
806
807
808
809
810
811
812
                    let request_ctx = RsContext::with_id(request, context.inner().id().to_string());

                    // Always instrument with appropriate span (none if no trace context)
                    let span = get_span_for_context(&context, "round_robin");
                    let stream_future = client.round_robin(request_ctx).instrument(span);

                    let stream = stream_future.await.map_err(to_pyerr)?;
813
814
815
816
817
                    context.inner().link_child(stream.context());
                    stream
                }
                _ => client.round_robin(request.into()).await.map_err(to_pyerr)?,
            };
818
819
820
821
822
823
824
825
826
            tokio::spawn(process_stream(stream, tx));
            Ok(AsyncResponseStream {
                rx: Arc::new(Mutex::new(rx)),
                annotated,
            })
        })
    }

    /// Send a request to a random endpoint.
827
    #[pyo3(signature = (request, annotated=DEFAULT_ANNOTATED_SETTING, context=None))]
828
829
830
831
832
    fn random<'p>(
        &self,
        py: Python<'p>,
        request: PyObject,
        annotated: Option<bool>,
833
        context: Option<context::Context>,
834
835
836
837
838
    ) -> 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);
839
        let client = self.router.clone();
840
841

        pyo3_async_runtimes::tokio::future_into_py(py, async move {
842
843
            let stream = match context {
                Some(context) => {
844
845
846
847
848
849
                    let request_ctx = RsContext::with_id(request, context.inner().id().to_string());

                    let span = get_span_for_context(&context, "random");
                    let stream_future = client.random(request_ctx).instrument(span);

                    let stream = stream_future.await.map_err(to_pyerr)?;
850
851
852
853
854
                    context.inner().link_child(stream.context());
                    stream
                }
                _ => client.random(request.into()).await.map_err(to_pyerr)?,
            };
855
856
857
858
859
860
861
862
863
            tokio::spawn(process_stream(stream, tx));
            Ok(AsyncResponseStream {
                rx: Arc::new(Mutex::new(rx)),
                annotated,
            })
        })
    }

    /// Directly send a request to a specific endpoint.
864
    #[pyo3(signature = (request, instance_id, annotated=DEFAULT_ANNOTATED_SETTING, context=None))]
865
866
867
868
    fn direct<'p>(
        &self,
        py: Python<'p>,
        request: PyObject,
869
        instance_id: i64,
870
        annotated: Option<bool>,
871
        context: Option<context::Context>,
872
873
874
875
876
    ) -> 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);
877
        let client = self.router.clone();
878
879

        pyo3_async_runtimes::tokio::future_into_py(py, async move {
880
881
            let stream = match context {
                Some(context) => {
882
883
884
885
886
887
888
                    let request_ctx = RsContext::with_id(request, context.inner().id().to_string());

                    let span =
                        get_span_for_direct_context(&context, "direct", &instance_id.to_string());
                    let stream_future = client.direct(request_ctx, instance_id).instrument(span);

                    let stream = stream_future.await.map_err(to_pyerr)?;
889
890
891
892
893
894
895
896
                    context.inner().link_child(stream.context());
                    stream
                }
                _ => client
                    .direct(request.into(), instance_id)
                    .await
                    .map_err(to_pyerr)?,
            };
897
898
899
900
901
902
903
904
905

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

            Ok(AsyncResponseStream {
                rx: Arc::new(Mutex::new(rx)),
                annotated,
            })
        })
    }
906
907

    /// Directly send a request to a pre-defined static worker
908
    #[pyo3(signature = (request, annotated=DEFAULT_ANNOTATED_SETTING, context=None))]
909
910
911
912
913
    fn r#static<'p>(
        &self,
        py: Python<'p>,
        request: PyObject,
        annotated: Option<bool>,
914
        context: Option<context::Context>,
915
916
917
918
919
    ) -> 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);
920
        let client = self.router.clone();
921
922

        pyo3_async_runtimes::tokio::future_into_py(py, async move {
923
924
            let stream = match context {
                Some(context) => {
925
926
927
928
929
930
                    let request_ctx = RsContext::with_id(request, context.inner().id().to_string());

                    let span = get_span_for_context(&context, "static");
                    let stream_future = client.r#static(request_ctx).instrument(span);

                    let stream = stream_future.await.map_err(to_pyerr)?;
931
932
933
934
935
                    context.inner().link_child(stream.context());
                    stream
                }
                _ => client.r#static(request.into()).await.map_err(to_pyerr)?,
            };
936
937
938
939
940
941
942
943
944

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

            Ok(AsyncResponseStream {
                rx: Arc::new(Mutex::new(rx)),
                annotated,
            })
        })
    }
945
946
947
}

async fn process_stream(
948
    stream: EngineStream<RsAnnotated<serde_json::Value>>,
949
950
951
952
953
    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
954
        let annotated: RsAnnotated<serde_json::Value> = response;
955
        let annotated: RsAnnotated<PyObject> = annotated.map_data(|data| {
956
            Python::with_gil(|py| match pythonize::pythonize(py, &data) {
957
958
                Ok(pyobj) => Ok(pyobj.into()),
                Err(e) => Err(e.to_string()),
959
            })
960
961
962
963
964
965
        });

        let is_error = annotated.is_error();

        // Send the PyObject through the channel or log an error
        if let Err(e) = tx.send(annotated).await {
966
            tracing::error!("Failed to send response: {:?}", e);
967
            break;
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
        }

        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")
        )
    }
}