lib.rs 36.6 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
use dynamo_llm::local_model::LocalModel;
5
6
use futures::StreamExt;
use once_cell::sync::OnceCell;
7
use pyo3::IntoPyObjectExt;
8
use pyo3::exceptions::PyStopAsyncIteration;
9
use pyo3::types::{PyDict, PyString};
10
use pyo3::{exceptions::PyException, prelude::*};
11
use rand::seq::IteratorRandom as _;
12
use rs::pipeline::network::Ingress;
13
use std::fs;
14
use std::net::{IpAddr, Ipv4Addr, SocketAddr, SocketAddrV4};
15
use std::path::PathBuf;
16
use std::time::Duration;
17
18
use std::{fmt::Display, sync::Arc};
use tokio::sync::Mutex;
19
use tracing::Instrument;
20

Neelay Shah's avatar
Neelay Shah committed
21
use dynamo_runtime::{
Ryan Olson's avatar
Ryan Olson committed
22
    self as rs, logging,
23
    pipeline::{
24
        AsyncEngineContextProvider, EngineStream, ManyOut, SingleIn, context::Context as RsContext,
25
        network::egress::push_router::RouterMode as RsRouterMode,
26
    },
27
    protocols::annotated::Annotated as RsAnnotated,
28
    traits::DistributedRuntimeProvider,
29
30
};

Neelay Shah's avatar
Neelay Shah committed
31
use dynamo_llm::{self as llm_rs};
32
33
use dynamo_llm::{entrypoint::RouterConfig, kv_router::KvRouterConfig};

34
35
use crate::llm::local_model::ModelRuntimeConfig;

36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
#[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,
        }
    }
}
53

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

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);

69
70
// Helper to get appropriate span for instrumentation - always emit spans
fn get_span_for_context(context: &context::Context, operation: &str) -> tracing::Span {
71
72
73
74
75
76
    logging::make_client_request_span(
        operation,
        context.inner().id(),
        context.trace_context(),
        None,
    )
77
78
79
80
81
82
83
84
}

// 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 {
85
86
87
88
89
90
    logging::make_client_request_span(
        operation,
        context.inner().id(),
        context.trace_context(),
        Some(instance_id),
    )
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
// Helper to create request context with proper linking and cancellation handling
fn create_request_context(
    request: serde_json::Value,
    parent_ctx: &Option<context::Context>,
) -> RsContext<serde_json::Value> {
    match parent_ctx {
        // If there is a parent context, link the request as a child context of it
        Some(parent_ctx) => {
            let child_ctx = RsContext::with_id(request, parent_ctx.inner().id().to_string());
            parent_ctx.inner().link_child(child_ctx.context());
            if parent_ctx.inner().is_stopped() || parent_ctx.inner().is_killed() {
                // Let the server handle the cancellation for now since not all backends are
                // properly handling request exceptions
                // TODO: (DIS-830) Return an error if context is cancelled
                child_ctx.context().stop_generating();
            }
            child_ctx
        }
        // Otherwise if there is no parent context, use the request as-is
        _ => request.into(),
    }
}

116
117
118
119
120
/// 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<()> {
Yan Ru Pei's avatar
Yan Ru Pei committed
121
    m.add_function(wrap_pyfunction!(llm::kv::compute_block_hash_for_seq_py, m)?)?;
Ryan Olson's avatar
Ryan Olson committed
122
    m.add_function(wrap_pyfunction!(log_message, m)?)?;
123
    m.add_function(wrap_pyfunction!(register_llm, m)?)?;
124
    m.add_function(wrap_pyfunction!(fetch_llm, m)?)?;
125
126
    m.add_function(wrap_pyfunction!(llm::entrypoint::make_engine, m)?)?;
    m.add_function(wrap_pyfunction!(llm::entrypoint::run_input, m)?)?;
127

128
129
130
131
132
133
134
    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>()?;
135
    m.add_class::<llm::disagg_router::DisaggregatedRouter>()?;
136
137
138
    m.add_class::<llm::entrypoint::EntrypointArgs>()?;
    m.add_class::<llm::entrypoint::EngineConfig>()?;
    m.add_class::<llm::entrypoint::EngineType>()?;
139
140
    m.add_class::<llm::entrypoint::RouterConfig>()?;
    m.add_class::<llm::entrypoint::KvRouterConfig>()?;
141
    m.add_class::<llm::kv::WorkerMetricsPublisher>()?;
142
    m.add_class::<llm::model_card::ModelDeploymentCard>()?;
143
    m.add_class::<llm::local_model::ModelRuntimeConfig>()?;
144
145
    m.add_class::<llm::preprocessor::OAIChatPreprocessor>()?;
    m.add_class::<llm::backend::Backend>()?;
146
147
    m.add_class::<llm::kv::OverlapScores>()?;
    m.add_class::<llm::kv::KvIndexer>()?;
148
    m.add_class::<llm::kv::ApproxKvIndexer>()?;
149
150
151
    m.add_class::<llm::kv::EndpointKvMetrics>()?;
    m.add_class::<llm::kv::AggregatedMetrics>()?;
    m.add_class::<llm::kv::KvMetricsAggregator>()?;
152
    m.add_class::<llm::kv::KvEventPublisher>()?;
Yan Ru Pei's avatar
Yan Ru Pei committed
153
154
    m.add_class::<llm::kv::RadixTree>()?;
    m.add_class::<llm::kv::ZmqKvEventListener>()?;
155
156
    m.add_class::<llm::kv::ZmqKvEventPublisher>()?;
    m.add_class::<llm::kv::ZmqKvEventPublisherConfig>()?;
157
    m.add_class::<llm::kv::KvRecorder>()?;
158
159
    m.add_class::<http::HttpService>()?;
    m.add_class::<http::HttpAsyncEngine>()?;
160
    m.add_class::<context::Context>()?;
161
    m.add_class::<ModelType>()?;
162
    m.add_class::<ModelInput>()?;
163
164
165
166
    m.add_class::<llm::kv::ForwardPassMetrics>()?;
    m.add_class::<llm::kv::WorkerStats>()?;
    m.add_class::<llm::kv::KvStats>()?;
    m.add_class::<llm::kv::SpecDecodeStats>()?;
167
168
    m.add_class::<llm::kv::KvPushRouter>()?;
    m.add_class::<llm::kv::KvPushRouterStream>()?;
169
    m.add_class::<RouterMode>()?;
170
    m.add("__version__", env!("CARGO_PKG_VERSION"))?;
171
172
173
    m.add_class::<planner::VirtualConnectorCoordinator>()?;
    m.add_class::<planner::VirtualConnectorClient>()?;
    m.add_class::<planner::PlannerDecision>()?;
174
175

    engine::add_to_module(m)?;
176
    parsers::add_to_module(m)?;
177

178
    m.add_class::<prometheus_metrics::RuntimeMetrics>()?;
179
180
181
182
    let prometheus_metrics = PyModule::new(m.py(), "prometheus_metrics")?;
    prometheus_metrics::add_to_module(&prometheus_metrics)?;
    m.add_submodule(&prometheus_metrics)?;

183
184
185
    #[cfg(feature = "block-manager")]
    llm::block_manager::add_to_module(m)?;

186
187
188
189
190
191
192
193
194
195
    Ok(())
}

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

Ryan Olson's avatar
Ryan Olson committed
196
197
198
199
200
201
202
/// 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);
}

203
204
/// Create an engine and attach it to an endpoint to make it visible to the frontend.
/// This is the main way you create a Dynamo worker / backend.
205
#[pyfunction]
206
#[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))]
207
#[allow(clippy::too_many_arguments)]
208
209
fn register_llm<'p>(
    py: Python<'p>,
210
    model_input: ModelInput,
211
    model_type: ModelType,
212
213
214
    endpoint: Endpoint,
    model_path: &str,
    model_name: Option<&str>,
215
216
    context_length: Option<u32>,
    kv_cache_block_size: Option<u32>,
217
    router_mode: Option<RouterMode>,
218
    migration_limit: u32,
219
    runtime_config: Option<ModelRuntimeConfig>,
220
    user_data: Option<&Bound<'p, PyDict>>,
221
    custom_template_path: Option<&str>,
222
) -> PyResult<Bound<'p, PyAny>> {
223
224
225
    let model_input = match model_input {
        ModelInput::Text => llm_rs::model_type::ModelInput::Text,
        ModelInput::Tokens => llm_rs::model_type::ModelInput::Tokens,
226
        ModelInput::Tensor => llm_rs::model_type::ModelInput::Tensor,
227
228
    };

229
230
    let model_type_obj = model_type.inner;

231
    let inner_path = model_path.to_string();
232
    let mut model_name = model_name.map(|n| n.to_string());
233
234
235
    let router_mode = router_mode.unwrap_or(RouterMode::RoundRobin);
    let router_config = RouterConfig::new(router_mode.into(), KvRouterConfig::default());

236
237
238
239
240
241
242
243
244
245
246
247
248
    // 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()?;

249
250
251
252
253
254
255
    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))
        })?;

256
    pyo3_async_runtimes::tokio::future_into_py(py, async move {
257
258
259
260
261
262
263
264
265
266
267
268
269
        let model_path = if fs::exists(&inner_path)? {
            PathBuf::from(inner_path)
        } else {
            // Preserve the model name
            if model_name.is_none() {
                model_name = Some(inner_path.clone());
            }
            // Likely it's a Hugging Face repo, download it
            LocalModel::fetch(&inner_path, false)
                .await
                .map_err(to_pyerr)?
        };

270
271
        let mut builder = dynamo_llm::local_model::LocalModelBuilder::default();
        builder
272
            .model_path(model_path)
273
274
            .model_name(model_name)
            .context_length(context_length)
275
            .kv_cache_block_size(kv_cache_block_size)
276
            .router_config(Some(router_config))
277
            .migration_limit(Some(migration_limit))
278
            .runtime_config(runtime_config.unwrap_or_default().inner)
279
280
            .user_data(user_data_json)
            .custom_template_path(custom_template_path_owned);
281
        // Load the ModelDeploymentCard
282
        let mut local_model = builder.build().await.map_err(to_pyerr)?;
283
284
        // Advertise ourself on etcd so ingress can find us
        local_model
285
            .attach(&endpoint.inner, model_type_obj, model_input)
286
287
288
289
290
291
292
            .await
            .map_err(to_pyerr)?;

        Ok(())
    })
}

293
294
295
296
297
298
299
300
301
302
303
/// Download a model from Hugging Face, returning it's local path
/// Example: `model_path = await fetch_llm("Qwen/Qwen3-0.6B")`
#[pyfunction]
#[pyo3(signature = (remote_name))]
fn fetch_llm<'p>(py: Python<'p>, remote_name: &str) -> PyResult<Bound<'p, PyAny>> {
    let repo = remote_name.to_string();
    pyo3_async_runtimes::tokio::future_into_py(py, async move {
        LocalModel::fetch(&repo, false).await.map_err(to_pyerr)
    })
}

304
305
#[pyclass]
#[derive(Clone)]
Ryan Olson's avatar
Ryan Olson committed
306
pub struct DistributedRuntime {
307
308
309
310
    inner: rs::DistributedRuntime,
    event_loop: PyObject,
}

Ryan Olson's avatar
Ryan Olson committed
311
312
impl DistributedRuntime {
    #[allow(dead_code)]
313
    pub(crate) fn inner(&self) -> &rs::DistributedRuntime {
Ryan Olson's avatar
Ryan Olson committed
314
315
316
317
        &self.inner
    }
}

318
#[pyclass]
319
#[derive(Clone)]
320
321
322
323
324
struct CancellationToken {
    inner: rs::CancellationToken,
}

#[pyclass]
325
#[derive(Clone)]
326
327
328
329
330
331
struct Namespace {
    inner: rs::component::Namespace,
    event_loop: PyObject,
}

#[pyclass]
332
#[derive(Clone)]
333
334
335
336
337
338
struct Component {
    inner: rs::component::Component,
    event_loop: PyObject,
}

#[pyclass]
339
#[derive(Clone)]
340
341
342
343
344
345
struct Endpoint {
    inner: rs::component::Endpoint,
    event_loop: PyObject,
}

#[pyclass]
346
#[derive(Clone)]
347
struct Client {
348
    router: rs::pipeline::PushRouter<serde_json::Value, RsAnnotated<serde_json::Value>>,
349
350
}

351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
#[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,
    };
372
373
374
375
    #[classattr]
    const TensorBased: Self = ModelType {
        inner: llm_rs::model_type::ModelType::TensorBased,
    };
376
377
378
379
380
381
382
383
384
385
386
387

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

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

388
389
#[pyclass(eq, eq_int)]
#[derive(Clone, PartialEq)]
390
391
392
enum ModelInput {
    Text = 1,
    Tokens = 2,
393
    Tensor = 3,
394
395
}

396
397
398
#[pymethods]
impl DistributedRuntime {
    #[new]
399
    fn new(event_loop: PyObject, is_static: bool) -> PyResult<Self> {
400
        let worker = rs::Worker::from_settings().map_err(to_pyerr)?;
401
        INIT.get_or_try_init(|| {
402
            let primary = worker.tokio_runtime()?;
403
404
405
406
407
408
            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)?;

409
        let runtime = worker.runtime().clone();
410

411
412
413
414
415
416
        // Initialize logging in context where tokio runtime is available
        // otel exporter requires it
        runtime.secondary().block_on(async {
            rs::logging::init();
        });

417
418
419
420
421
422
423
424
425
426
427
        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)?;
428
429
430
431

        Ok(DistributedRuntime { inner, event_loop })
    }

Ryan Olson's avatar
Ryan Olson committed
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
    #[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(),
        })
    }

447
448
449
450
451
452
453
    fn namespace(&self, name: String) -> PyResult<Namespace> {
        Ok(Namespace {
            inner: self.inner.namespace(name).map_err(to_pyerr)?,
            event_loop: self.event_loop.clone(),
        })
    }

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
511
512
513
514
515
516
517
518
519
520
521
522
    /// 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!(
523
524
525
                                "Failed to bind to port block starting at {start_port} (attempt {}): {e}",
                                attempt_idx + 1,
                            );
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
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
                            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!(
581
582
                "Failed to allocate and reserve a port block of size {block_size} from range {min}-{max} after {candidate_count} attempts"
            )))
583
584
585
        })
    }

586
587
588
589
590
591
592
    fn shutdown(&self) {
        self.inner.runtime().shutdown();
    }

    fn event_loop(&self) -> PyObject {
        self.event_loop.clone()
    }
593
594
595
596
597

    fn child_token(&self) -> CancellationToken {
        let inner = self.inner.runtime().child_token();
        CancellationToken { inner }
    }
598
599
}

600
601
602
603
604
605
606
607
608
609
610
611
612
613
// 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> {
614
    Ok(format!("v1/{namespace}/ports/{node_ip}/{port}"))
615
616
617
618
619
620
621
622
623
624
625
626
}

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),
    })
}

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
#[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(())
        })
    }
659
660
661

    /// Get a RuntimeMetrics helper for creating Prometheus metrics
    #[getter]
662
663
    fn metrics(&self) -> prometheus_metrics::RuntimeMetrics {
        prometheus_metrics::RuntimeMetrics::from_component(self.inner.clone())
664
    }
665
666
667
668
}

#[pymethods]
impl Endpoint {
669
    #[pyo3(signature = (generator, graceful_shutdown = true, metrics_labels = None, health_check_payload = None))]
670
671
672
673
    fn serve_endpoint<'p>(
        &self,
        py: Python<'p>,
        generator: PyObject,
674
        graceful_shutdown: Option<bool>,
675
        metrics_labels: Option<Vec<(String, String)>>,
676
        health_check_payload: Option<&Bound<'p, PyDict>>,
677
678
679
680
681
682
    ) -> 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)?;
683
684
685
686
687
688
689
690
691
692
693
694
695

        // 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
696
697
698
699
700
701
        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)",
            ));
702
703
704
        }

        let mut builder = self
705
706
707
708
            .inner
            .endpoint_builder()
            .metrics_labels(metrics_labels)
            .handler(ingress);
709
710
711
712
713

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

714
        let graceful_shutdown = graceful_shutdown.unwrap_or(true);
715
        pyo3_async_runtimes::tokio::future_into_py(py, async move {
716
717
718
719
720
            builder
                .graceful_shutdown(graceful_shutdown)
                .start()
                .await
                .map_err(to_pyerr)?;
721
722
723
724
            Ok(())
        })
    }

725
726
727
    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 {
728
            let client = inner.client().await.map_err(to_pyerr)?;
729
730
731
732
733
734
            let push_router = rs::pipeline::PushRouter::<
                serde_json::Value,
                RsAnnotated<serde_json::Value>,
            >::from_client(client, Default::default())
            .await
            .map_err(to_pyerr)?;
735
736
737
            Ok(Client {
                router: push_router,
            })
738
739
        })
    }
740
741

    fn lease_id(&self) -> i64 {
742
743
744
745
746
        self.inner
            .drt()
            .primary_lease()
            .map(|l| l.id())
            .unwrap_or(0)
747
    }
748
749
750

    /// Get a RuntimeMetrics helper for creating Prometheus metrics
    #[getter]
751
752
    fn metrics(&self) -> prometheus_metrics::RuntimeMetrics {
        prometheus_metrics::RuntimeMetrics::from_endpoint(self.inner.clone())
753
    }
754
755
756
757
758
759
760
761
762
763
764
}

#[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(),
        })
    }
765
766
767

    /// Get a RuntimeMetrics helper for creating Prometheus metrics
    #[getter]
768
769
    fn metrics(&self) -> prometheus_metrics::RuntimeMetrics {
        prometheus_metrics::RuntimeMetrics::from_namespace(self.inner.clone())
770
    }
771
772
773
774
}

#[pymethods]
impl Client {
775
776
777
778
    /// Get list of current instances.
    /// Replaces endpoint_ids.
    fn instance_ids(&self) -> Vec<i64> {
        self.router.client.instance_ids()
779
780
    }

781
782
783
    /// 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>> {
784
        let inner = self.router.client.clone();
785
        pyo3_async_runtimes::tokio::future_into_py(py, async move {
786
            inner
787
                .wait_for_instances()
788
789
790
                .await
                .map(|v| v.into_iter().map(|cei| cei.id()).collect::<Vec<i64>>())
                .map_err(to_pyerr)
791
792
793
794
        })
    }

    /// Issue a request to the endpoint using the default routing strategy.
795
    #[pyo3(signature = (request, annotated=DEFAULT_ANNOTATED_SETTING, context=None))]
796
797
798
799
800
    fn generate<'p>(
        &self,
        py: Python<'p>,
        request: PyObject,
        annotated: Option<bool>,
801
        context: Option<context::Context>,
802
    ) -> PyResult<Bound<'p, PyAny>> {
803
        if self.router.client.is_static() {
804
            self.r#static(py, request, annotated, context)
805
        } else {
806
            self.random(py, request, annotated, context)
807
        }
808
809
810
    }

    /// Send a request to the next endpoint in a round-robin fashion.
811
    #[pyo3(signature = (request, annotated=DEFAULT_ANNOTATED_SETTING, context=None))]
812
813
814
815
816
    fn round_robin<'p>(
        &self,
        py: Python<'p>,
        request: PyObject,
        annotated: Option<bool>,
817
        context: Option<context::Context>,
818
819
    ) -> PyResult<Bound<'p, PyAny>> {
        let request: serde_json::Value = pythonize::depythonize(&request.into_bound(py))?;
820
        let request_ctx = create_request_context(request, &context);
821
822
823
        let annotated = annotated.unwrap_or(false);

        let (tx, rx) = tokio::sync::mpsc::channel(32);
824
        let client = self.router.clone();
825
826

        pyo3_async_runtimes::tokio::future_into_py(py, async move {
827
828
            let stream = match context {
                Some(context) => {
829
830
                    // Always instrument with appropriate span (none if no trace context)
                    let span = get_span_for_context(&context, "round_robin");
831
832
833
834
835
                    client
                        .round_robin(request_ctx)
                        .instrument(span)
                        .await
                        .map_err(to_pyerr)?
836
                }
837
                _ => client.round_robin(request_ctx).await.map_err(to_pyerr)?,
838
            };
839
840
841
842
843
844
845
846
847
            tokio::spawn(process_stream(stream, tx));
            Ok(AsyncResponseStream {
                rx: Arc::new(Mutex::new(rx)),
                annotated,
            })
        })
    }

    /// Send a request to a random endpoint.
848
    #[pyo3(signature = (request, annotated=DEFAULT_ANNOTATED_SETTING, context=None))]
849
850
851
852
853
    fn random<'p>(
        &self,
        py: Python<'p>,
        request: PyObject,
        annotated: Option<bool>,
854
        context: Option<context::Context>,
855
856
    ) -> PyResult<Bound<'p, PyAny>> {
        let request: serde_json::Value = pythonize::depythonize(&request.into_bound(py))?;
857
        let request_ctx = create_request_context(request, &context);
858
859
860
        let annotated = annotated.unwrap_or(false);

        let (tx, rx) = tokio::sync::mpsc::channel(32);
861
        let client = self.router.clone();
862
863

        pyo3_async_runtimes::tokio::future_into_py(py, async move {
864
865
            let stream = match context {
                Some(context) => {
866
                    // Always instrument with appropriate span (none if no trace context)
867
                    let span = get_span_for_context(&context, "random");
868
869
870
871
872
                    client
                        .random(request_ctx)
                        .instrument(span)
                        .await
                        .map_err(to_pyerr)?
873
                }
874
                _ => client.random(request_ctx).await.map_err(to_pyerr)?,
875
            };
876
877
878
879
880
881
882
883
884
            tokio::spawn(process_stream(stream, tx));
            Ok(AsyncResponseStream {
                rx: Arc::new(Mutex::new(rx)),
                annotated,
            })
        })
    }

    /// Directly send a request to a specific endpoint.
885
    #[pyo3(signature = (request, instance_id, annotated=DEFAULT_ANNOTATED_SETTING, context=None))]
886
887
888
889
    fn direct<'p>(
        &self,
        py: Python<'p>,
        request: PyObject,
890
        instance_id: i64,
891
        annotated: Option<bool>,
892
        context: Option<context::Context>,
893
894
    ) -> PyResult<Bound<'p, PyAny>> {
        let request: serde_json::Value = pythonize::depythonize(&request.into_bound(py))?;
895
        let request_ctx = create_request_context(request, &context);
896
897
898
        let annotated = annotated.unwrap_or(false);

        let (tx, rx) = tokio::sync::mpsc::channel(32);
899
        let client = self.router.clone();
900
901

        pyo3_async_runtimes::tokio::future_into_py(py, async move {
902
903
            let stream = match context {
                Some(context) => {
904
                    // Always instrument with appropriate span (none if no trace context)
905
906
                    let span =
                        get_span_for_direct_context(&context, "direct", &instance_id.to_string());
907
908
909
910
911
                    client
                        .direct(request_ctx, instance_id)
                        .instrument(span)
                        .await
                        .map_err(to_pyerr)?
912
913
                }
                _ => client
914
                    .direct(request_ctx, instance_id)
915
916
917
                    .await
                    .map_err(to_pyerr)?,
            };
918
919
920
921
922
923
924
925
926

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

            Ok(AsyncResponseStream {
                rx: Arc::new(Mutex::new(rx)),
                annotated,
            })
        })
    }
927
928

    /// Directly send a request to a pre-defined static worker
929
    #[pyo3(signature = (request, annotated=DEFAULT_ANNOTATED_SETTING, context=None))]
930
931
932
933
934
    fn r#static<'p>(
        &self,
        py: Python<'p>,
        request: PyObject,
        annotated: Option<bool>,
935
        context: Option<context::Context>,
936
937
    ) -> PyResult<Bound<'p, PyAny>> {
        let request: serde_json::Value = pythonize::depythonize(&request.into_bound(py))?;
938
        let request_ctx = create_request_context(request, &context);
939
940
941
        let annotated = annotated.unwrap_or(false);

        let (tx, rx) = tokio::sync::mpsc::channel(32);
942
        let client = self.router.clone();
943
944

        pyo3_async_runtimes::tokio::future_into_py(py, async move {
945
946
            let stream = match context {
                Some(context) => {
947
                    // Always instrument with appropriate span (none if no trace context)
948
                    let span = get_span_for_context(&context, "static");
949
950
951
952
953
                    client
                        .r#static(request_ctx)
                        .instrument(span)
                        .await
                        .map_err(to_pyerr)?
954
                }
955
                _ => client.r#static(request_ctx).await.map_err(to_pyerr)?,
956
            };
957
958
959
960
961
962
963
964
965

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

            Ok(AsyncResponseStream {
                rx: Arc::new(Mutex::new(rx)),
                annotated,
            })
        })
    }
966
967
968
}

async fn process_stream(
969
    stream: EngineStream<RsAnnotated<serde_json::Value>>,
970
971
972
973
974
    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
975
        let annotated: RsAnnotated<serde_json::Value> = response;
976
        let annotated: RsAnnotated<PyObject> = annotated.map_data(|data| {
977
            Python::with_gil(|py| match pythonize::pythonize(py, &data) {
978
979
                Ok(pyobj) => Ok(pyobj.into()),
                Err(e) => Err(e.to_string()),
980
            })
981
982
983
984
985
986
        });

        let is_error = annotated.is_error();

        // Send the PyObject through the channel or log an error
        if let Err(e) = tx.send(annotated).await {
987
            tracing::error!("Failed to send response: {:?}", e);
988
            break;
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
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
        }

        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")
        )
    }
}