lib.rs 43.4 KB
Newer Older
1
// SPDX-FileCopyrightText: Copyright (c) 2024-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
2
3
// SPDX-License-Identifier: Apache-2.0

4
use dynamo_llm::local_model::LocalModel;
5
use dynamo_runtime::distributed::{DiscoveryBackend, DistributedConfig, RequestPlaneMode};
6
use dynamo_runtime::storage::kv;
7
8
use futures::StreamExt;
use once_cell::sync::OnceCell;
9
use pyo3::IntoPyObjectExt;
10
use pyo3::exceptions::PyStopAsyncIteration;
Richard Huo's avatar
Richard Huo committed
11
use pyo3::types::PyCapsule;
12
use pyo3::types::{PyDict, PyString};
13
14
use pyo3::{exceptions::PyException, prelude::*};
use rs::pipeline::network::Ingress;
Richard Huo's avatar
Richard Huo committed
15
use std::ffi::CString;
16
use std::fs;
17
use std::path::PathBuf;
Richard Huo's avatar
Richard Huo committed
18
19
20
21
use std::{
    fmt::Display,
    sync::{Arc, Weak},
};
22
use tokio::sync::Mutex;
23
use tracing::Instrument;
24

25
use dynamo_runtime::config;
26
use dynamo_runtime::config::environment_names::logging::otlp as env_otlp;
Neelay Shah's avatar
Neelay Shah committed
27
use dynamo_runtime::{
Ryan Olson's avatar
Ryan Olson committed
28
    self as rs, logging,
29
    pipeline::{
30
        AsyncEngineContextProvider, EngineStream, ManyOut, SingleIn, context::Context as RsContext,
31
        network::egress::push_router::RouterMode as RsRouterMode,
32
    },
33
    protocols::annotated::Annotated as RsAnnotated,
34
    traits::DistributedRuntimeProvider,
35
36
};

37
38
use dynamo_kv_router::config::KvRouterConfig;
use dynamo_llm::entrypoint::RouterConfig;
Neelay Shah's avatar
Neelay Shah committed
39
use dynamo_llm::{self as llm_rs};
40

41
use crate::llm::local_model::ModelRuntimeConfig;
42
use crate::llm::preprocessor::{MediaDecoder, MediaFetcher};
43

44
45
46
47
48
#[pyclass(eq, eq_int)]
#[derive(Clone, Debug, PartialEq)]
pub enum RouterMode {
    RoundRobin,
    Random,
49
    PowerOfTwoChoices,
50
    KV,
51
52
53
    /// Direct routing - reads worker ID from each request's routing hints.
    /// Used when an external orchestrator (e.g., EPP) handles worker selection.
    Direct,
54
55
56
57
58
59
60
}

impl From<RouterMode> for RsRouterMode {
    fn from(mode: RouterMode) -> Self {
        match mode {
            RouterMode::RoundRobin => Self::RoundRobin,
            RouterMode::Random => Self::Random,
61
            RouterMode::PowerOfTwoChoices => Self::PowerOfTwoChoices,
62
            RouterMode::KV => Self::KV,
63
            RouterMode::Direct => Self::Direct,
64
65
66
        }
    }
}
67

68
mod context;
69
mod engine;
70
mod http;
71
mod kserve_grpc;
72
mod llm;
73
mod parsers;
74
mod planner;
75
mod prometheus_metrics;
76
77
78
79
80
81
82

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);
83
84
const SKIP_PYTHON_LOG_INIT_ENV: &str = "DYNAMO_SKIP_PYTHON_LOG_INIT";

85
86
// Helper to get appropriate span for instrumentation - always emit spans
fn get_span_for_context(context: &context::Context, operation: &str) -> tracing::Span {
87
88
89
90
91
92
    logging::make_client_request_span(
        operation,
        context.inner().id(),
        context.trace_context(),
        None,
    )
93
94
95
96
97
98
99
100
}

// 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 {
101
102
103
104
105
106
    logging::make_client_request_span(
        operation,
        context.inner().id(),
        context.trace_context(),
        Some(instance_id),
    )
107
108
}

109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
// 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(),
    }
}

132
133
134
135
136
/// 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<()> {
137
    // Initialize logging early unless OTEL export is enabled (which requires tokio runtime)
138
    if config::env_is_truthy(env_otlp::OTEL_EXPORT_ENABLED) {
139
        eprintln!(
140
            "Warning: OTEL_EXPORT_ENABLED detected. Logging initialization deferred until runtime is available. Early logs may be dropped."
141
        );
142
    } else if std::env::var_os(SKIP_PYTHON_LOG_INIT_ENV).is_none() {
143
144
145
        rs::logging::init();
    }

Yan Ru Pei's avatar
Yan Ru Pei committed
146
    m.add_function(wrap_pyfunction!(llm::kv::compute_block_hash_for_seq_py, m)?)?;
147
    m.add_function(wrap_pyfunction!(lora_name_to_id, m)?)?;
Ryan Olson's avatar
Ryan Olson committed
148
    m.add_function(wrap_pyfunction!(log_message, m)?)?;
149
150
151
    m.add_function(wrap_pyfunction!(register_model, m)?)?;
    m.add_function(wrap_pyfunction!(unregister_model, m)?)?;
    m.add_function(wrap_pyfunction!(fetch_model, m)?)?;
152
    m.add_function(wrap_pyfunction!(run_kv_indexer, m)?)?;
153
    m.add_function(wrap_pyfunction!(llm::entrypoint::make_engine, m)?)?;
154
    m.add_function(wrap_pyfunction!(llm::replay::run_mocker_trace_replay, m)?)?;
155
    m.add_function(wrap_pyfunction!(
156
        llm::replay::run_mocker_synthetic_trace_replay,
157
158
        m
    )?)?;
159
    m.add_function(wrap_pyfunction!(llm::entrypoint::run_input, m)?)?;
160

161
162
    m.add_class::<DistributedRuntime>()?;
    m.add_class::<Endpoint>()?;
163
    m.add_class::<ModelCardInstanceId>()?;
164
165
    m.add_class::<Client>()?;
    m.add_class::<AsyncResponseStream>()?;
166
167
168
    m.add_class::<llm::entrypoint::EntrypointArgs>()?;
    m.add_class::<llm::entrypoint::EngineConfig>()?;
    m.add_class::<llm::entrypoint::EngineType>()?;
169
170
    m.add_class::<llm::entrypoint::RouterConfig>()?;
    m.add_class::<llm::entrypoint::KvRouterConfig>()?;
171
172
173
    m.add_class::<llm::replay::ReasoningConfig>()?;
    m.add_class::<llm::replay::SglangArgs>()?;
    m.add_class::<llm::replay::MockEngineArgs>()?;
174
    m.add_class::<llm::kv::WorkerMetricsPublisher>()?;
175
    m.add_class::<llm::model_card::ModelDeploymentCard>()?; // Internal: only in _internal, not public API
176
    m.add_class::<llm::local_model::ModelRuntimeConfig>()?;
177
178
    m.add_class::<llm::preprocessor::MediaDecoder>()?;
    m.add_class::<llm::preprocessor::MediaFetcher>()?;
179
    m.add_class::<llm::kv::OverlapScores>()?;
180
    m.add_class::<llm::kv::KvEventPublisher>()?;
Yan Ru Pei's avatar
Yan Ru Pei committed
181
    m.add_class::<llm::kv::RadixTree>()?;
182
183
    m.add_class::<llm::fpm::FpmEventRelay>()?;
    m.add_class::<llm::fpm::FpmEventSubscriber>()?;
184
    m.add_class::<llm::lora::LoRADownloader>()?;
185
186
    m.add_class::<http::HttpService>()?;
    m.add_class::<http::HttpAsyncEngine>()?;
187
    m.add_class::<context::Context>()?;
188
    m.add_class::<ModelType>()?;
189
    m.add_class::<ModelInput>()?;
190
    m.add_class::<llm::kv::KvRouter>()?;
191
    m.add_class::<RouterMode>()?;
192
    m.add_class::<kserve_grpc::KserveGrpcService>()?;
193
    m.add("__version__", env!("CARGO_PKG_VERSION"))?;
194
195
196
    m.add_class::<planner::VirtualConnectorCoordinator>()?;
    m.add_class::<planner::VirtualConnectorClient>()?;
    m.add_class::<planner::PlannerDecision>()?;
197
198

    engine::add_to_module(m)?;
199
    parsers::add_to_module(m)?;
200

201
    m.add_class::<prometheus_metrics::RuntimeMetrics>()?;
202
203
204
205
    let prometheus_metrics = PyModule::new(m.py(), "prometheus_metrics")?;
    prometheus_metrics::add_to_module(&prometheus_metrics)?;
    m.add_submodule(&prometheus_metrics)?;

206
207
208
209
210
211
212
213
214
215
    Ok(())
}

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

216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
fn kv_indexer_to_pyerr(err: anyhow::Error) -> PyErr {
    #[cfg(feature = "kv-indexer")]
    if let Some(clap_error) = err.downcast_ref::<clap::Error>() {
        let _ = clap_error.print();
        return pyo3::exceptions::PySystemExit::new_err(clap_error.exit_code());
    }

    to_pyerr(err)
}

#[pyfunction(name = "run_kv_indexer")]
#[pyo3(signature = (argv=None))]
fn run_kv_indexer(py: Python<'_>, argv: Option<Vec<String>>) -> PyResult<()> {
    let argv = argv.unwrap_or_default();
    py.allow_threads(move || llm::kv::run_kv_indexer_cli(argv))
        .map_err(kv_indexer_to_pyerr)
}

Ryan Olson's avatar
Ryan Olson committed
234
235
236
237
238
239
240
/// 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);
}

241
242
243
244
245
246
247
/// Generate a deterministic signed int32 ID from a LoRA name using blake3 hash.
#[pyfunction]
#[pyo3(text_signature = "(lora_name)")]
fn lora_name_to_id(lora_name: &str) -> i32 {
    llm_rs::utils::lora_name_to_id(lora_name)
}

248
249
/// 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.
250
251
252
253
254
255
256
///
/// If `lora_name` is provided, this function will publish a LoRA adapter instead of a base model:
/// - LoRA path: v1/mdc/{namespace}/{component}/{endpoint}/{instance_id}/{lora_slug}
/// - Base model path: v1/mdc/{namespace}/{component}/{endpoint}/{instance_id}
///
/// For LoRA mode, both `lora_name` and `base_model_path` must be provided together.
/// Providing only one of them will result in an error.
257
#[pyfunction]
258
#[pyo3(signature = (model_input, model_type, endpoint, model_path, model_name=None, context_length=None, kv_cache_block_size=None, router_mode=None, runtime_config=None, user_data=None, custom_template_path=None, media_decoder=None, media_fetcher=None, lora_name=None, base_model_path=None))]
259
#[allow(clippy::too_many_arguments)]
260
fn register_model<'p>(
261
    py: Python<'p>,
262
    model_input: ModelInput,
263
    model_type: ModelType,
264
265
266
    endpoint: Endpoint,
    model_path: &str,
    model_name: Option<&str>,
267
268
    context_length: Option<u32>,
    kv_cache_block_size: Option<u32>,
269
    router_mode: Option<RouterMode>,
270
    runtime_config: Option<ModelRuntimeConfig>,
271
    user_data: Option<&Bound<'p, PyDict>>,
272
    custom_template_path: Option<&str>,
273
274
    media_decoder: Option<MediaDecoder>,
    media_fetcher: Option<MediaFetcher>,
275
276
    lora_name: Option<&str>,
    base_model_path: Option<&str>,
277
) -> PyResult<Bound<'p, PyAny>> {
278
    // Validate Prefill model type requirements
279
280
281
282
283
284
    if model_type.inner == llm_rs::model_type::ModelType::Prefill
        && !matches!(model_input, ModelInput::Tokens)
    {
        return Err(PyErr::new::<pyo3::exceptions::PyValueError, _>(
            "ModelType::Prefill requires model_input to be ModelInput::Tokens",
        ));
285
286
    }

287
288
289
    let model_input = match model_input {
        ModelInput::Text => llm_rs::model_type::ModelInput::Text,
        ModelInput::Tokens => llm_rs::model_type::ModelInput::Tokens,
290
        ModelInput::Tensor => llm_rs::model_type::ModelInput::Tensor,
291
292
    };

293
    let is_tensor_based = model_type.inner.supports_tensor();
294
    let is_images = model_type.inner.supports_images();
295
    let is_videos = model_type.inner.supports_videos();
296

297
298
    let model_type_obj = model_type.inner;

299
    let inner_path = model_path.to_string();
300
    let model_name = model_name.map(|n| n.to_string());
301
302
303
    let router_mode = router_mode.unwrap_or(RouterMode::RoundRobin);
    let router_config = RouterConfig::new(router_mode.into(), KvRouterConfig::default());

304
305
306
307
308
309
310
311
312
313
314
315
316
    // 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()?;

317
318
319
320
321
322
323
    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))
        })?;

324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
    // Validate LoRA parameters: both or neither must be provided
    if lora_name.is_some() ^ base_model_path.is_some() {
        return Err(PyErr::new::<pyo3::exceptions::PyValueError, _>(
            "lora_name and base_model_path must both be provided together, or neither",
        ));
    }

    // Determine source_path and lora_identifier based on registration mode
    let (source_path, lora_identifier) = match (lora_name, base_model_path) {
        (Some(lora), Some(base)) => (base.to_string(), Some(lora.to_string())),
        _ => (inner_path, None),
    };

    // Model name: use lora name if present, otherwise provided name or default to source path
    let model_name = lora_identifier
        .clone()
        .or(model_name)
        .or_else(|| Some(source_path.clone()));

343
    pyo3_async_runtimes::tokio::future_into_py(py, async move {
344
345
346
        // For TensorBased, Images, and Videos models, skip HuggingFace downloads and register directly
        // These model types handle model loading internally, no tokenizer extraction needed
        if is_tensor_based || is_images || is_videos {
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
            let model_name = model_name.unwrap_or_else(|| source_path.clone());
            let mut card = llm_rs::model_card::ModelDeploymentCard::with_name_only(&model_name);
            card.model_type = model_type_obj;
            card.model_input = model_input;
            card.user_data = user_data_json;

            if let Some(cfg) = runtime_config {
                card.runtime_config = cfg.inner;
            }

            // Register the Model Deployment Card via discovery interface
            let discovery = endpoint.inner.drt().discovery();
            let spec = rs::discovery::DiscoverySpec::from_model(
                endpoint.inner.component().namespace().name().to_string(),
                endpoint.inner.component().name().to_string(),
                endpoint.inner.name().to_string(),
                &card,
            )
            .map_err(to_pyerr)?;
            discovery.register(spec).await.map_err(to_pyerr)?;

            return Ok(());
        }

        // For non-TensorBased models, resolve the model path (local or fetch from HuggingFace)
372
373
        let model_path = if fs::exists(&source_path)? {
            PathBuf::from(&source_path)
374
        } else {
375
            LocalModel::fetch(&source_path, false)
376
377
378
379
                .await
                .map_err(to_pyerr)?
        };

380
381
        let mut builder = dynamo_llm::local_model::LocalModelBuilder::default();
        builder
382
            // model path is the physical path on disk of the downloaded model
383
            .model_path(model_path)
384
385
386
387
            // source path is what the user gave as `--model-path`, either a real path (in which
            // case it matches model_path above), or an HF repo.
            .source_path(source_path.clone().into())
            // --served_model_name
388
            .model_name(model_name.clone())
389
            .context_length(context_length)
390
            .kv_cache_block_size(kv_cache_block_size)
391
            .router_config(Some(router_config))
392
            .runtime_config(runtime_config.unwrap_or_default().inner)
393
            .user_data(user_data_json)
394
395
396
            .custom_template_path(custom_template_path_owned)
            .media_decoder(media_decoder.map(|m| m.inner))
            .media_fetcher(media_fetcher.map(|m| m.inner));
397

398
        let mut local_model = builder.build().await.map_err(to_pyerr)?;
399
400
401
402
403
404
405
406
407

        // Convert lora_identifier (Option<String>) to Option<LoraInfo>
        let lora_info = lora_identifier
            .as_ref()
            .map(|name| llm_rs::model_card::LoraInfo {
                name: name.clone(),
                max_gpu_lora_count: None,
            });

408
        local_model
409
            .attach(&endpoint.inner, model_type_obj, model_input, lora_info)
410
411
412
            .await
            .map_err(to_pyerr)?;

413
414
415
        if let Some(lora_name) = lora_identifier {
            tracing::info!("Registered LoRA '{}' MDC", lora_name);
        } else {
416
417
418
419
            tracing::info!(
                "Registered base model '{}' MDC",
                model_name.unwrap_or(source_path)
            );
420
421
        }

422
423
424
425
        Ok(())
    })
}

426
427
428
429
430
431
432
433
434
435
436
437
438
/// Unregister a Model Deployment Card (MDC) from the service registry
///
/// This removes an LLM deployment from the discovery system.
///
/// # Arguments
///
/// * `endpoint` - The endpoint where the model is registered
/// * `lora_name` - Optional LoRA adapter name (if unregistering a LoRA deployment)
///
/// # MDC Path Format
///
/// - Base model: `v1/mdc/{namespace}/{component}/{endpoint}/{instance_id}`
/// - LoRA model: `v1/mdc/{namespace}/{component}/{endpoint}/{instance_id}/{lora_slug}`
439
#[pyfunction]
440
#[pyo3(signature = (endpoint, lora_name=None))]
441
fn unregister_model<'p>(
442
443
444
445
446
447
    py: Python<'p>,
    endpoint: Endpoint,
    lora_name: Option<&str>,
) -> PyResult<Bound<'p, PyAny>> {
    let lora_name_owned = lora_name.map(|s| s.to_string());

448
    pyo3_async_runtimes::tokio::future_into_py(py, async move {
449
450
        // Unified detach method handles both base models and LoRA adapters
        LocalModel::detach_from_endpoint(&endpoint.inner, lora_name_owned.as_deref())
451
452
453
454
455
456
            .await
            .map_err(to_pyerr)?;
        Ok(())
    })
}

457
458
/// Download a model from Hugging Face, returning its local path
/// Example: `model_path = await fetch_model("Qwen/Qwen3-0.6B")`
459
#[pyfunction]
460
#[pyo3(signature = (remote_name, ignore_weights=false))]
461
fn fetch_model<'p>(
462
463
464
465
    py: Python<'p>,
    remote_name: &str,
    ignore_weights: bool,
) -> PyResult<Bound<'p, PyAny>> {
466
467
    let repo = remote_name.to_string();
    pyo3_async_runtimes::tokio::future_into_py(py, async move {
468
469
470
        LocalModel::fetch(&repo, ignore_weights)
            .await
            .map_err(to_pyerr)
471
472
473
    })
}

474
475
#[pyclass]
#[derive(Clone)]
Ryan Olson's avatar
Ryan Olson committed
476
pub struct DistributedRuntime {
477
478
479
480
    inner: rs::DistributedRuntime,
    event_loop: PyObject,
}

Ryan Olson's avatar
Ryan Olson committed
481
482
impl DistributedRuntime {
    #[allow(dead_code)]
483
    pub(crate) fn inner(&self) -> &rs::DistributedRuntime {
Ryan Olson's avatar
Ryan Olson committed
484
485
486
487
        &self.inner
    }
}

488
#[pyclass]
489
#[derive(Clone)]
490
491
492
493
494
struct CancellationToken {
    inner: rs::CancellationToken,
}

#[pyclass]
495
#[derive(Clone)]
496
497
498
499
500
struct Endpoint {
    inner: rs::component::Endpoint,
    event_loop: PyObject,
}

501
502
503
504
505
506
#[pyclass]
#[derive(Clone)]
struct ModelCardInstanceId {
    inner: rs::discovery::ModelCardInstanceId,
}

507
#[pyclass]
508
#[derive(Clone)]
509
struct Client {
510
    router: rs::pipeline::PushRouter<serde_json::Value, RsAnnotated<serde_json::Value>>,
511
512
}

513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
#[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,
    };
534
535
536
537
    #[classattr]
    const TensorBased: Self = ModelType {
        inner: llm_rs::model_type::ModelType::TensorBased,
    };
538
539
540
541
    #[classattr]
    const Prefill: Self = ModelType {
        inner: llm_rs::model_type::ModelType::Prefill,
    };
542
543
544
545
    #[classattr]
    const Images: Self = ModelType {
        inner: llm_rs::model_type::ModelType::Images,
    };
546
    #[classattr]
547
548
549
550
    const Audios: Self = ModelType {
        inner: llm_rs::model_type::ModelType::Audios,
    };
    #[classattr]
551
552
553
    const Videos: Self = ModelType {
        inner: llm_rs::model_type::ModelType::Videos,
    };
554

555
556
557
558
    fn supports_chat(&self) -> bool {
        self.inner.supports_chat()
    }

559
560
561
562
563
564
565
566
567
568
569
    fn __or__(&self, other: &Self) -> Self {
        ModelType {
            inner: self.inner | other.inner,
        }
    }

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

570
571
#[pyclass(eq, eq_int)]
#[derive(Clone, PartialEq)]
572
573
574
enum ModelInput {
    Text = 1,
    Tokens = 2,
575
    Tensor = 3,
576
577
}

578
579
580
#[pymethods]
impl DistributedRuntime {
    #[new]
581
    #[pyo3(signature = (event_loop, discovery_backend, request_plane, enable_nats=None))]
582
583
    fn new(
        event_loop: PyObject,
584
        discovery_backend: String,
585
586
587
        request_plane: String,
        enable_nats: Option<bool>,
    ) -> PyResult<Self> {
588
589
590
591
592
593
594
        let discovery_backend_config = match discovery_backend.as_str() {
            "kubernetes" => DiscoveryBackend::Kubernetes,
            other => {
                let selector: kv::Selector = other.parse().map_err(to_pyerr)?;
                DiscoveryBackend::KvStore(selector)
            }
        };
595
        let request_plane: RequestPlaneMode = request_plane.parse().map_err(to_pyerr)?;
596

597
598
599
        // Try to get existing runtime first, create new Worker only if needed
        // This allows multiple DistributedRuntime instances to share the same tokio runtime
        let runtime = rs::Worker::runtime_from_existing()
600
            .or_else(|_| -> anyhow::Result<rs::Runtime> {
601
602
603
604
                // No existing Worker, create new one
                let worker = rs::Worker::from_settings()?;

                // Initialize pyo3 bridge (only happens once per process)
605
                INIT.get_or_try_init(|| -> anyhow::Result<()> {
606
607
                    let primary = worker.tokio_runtime()?;
                    pyo3_async_runtimes::tokio::init_with_runtime(primary).map_err(|e| {
608
                        anyhow::anyhow!("failed to initialize pyo3 static runtime: {:?}", e)
609
                    })?;
610
                    Ok(())
611
612
                })?;

613
                Ok(worker.runtime().clone())
614
615
            })
            .map_err(to_pyerr)?;
616

617
618
        // Initialize logging in context where tokio runtime is available
        // otel exporter requires it
619
        if config::env_is_truthy(env_otlp::OTEL_EXPORT_ENABLED) {
620
621
622
623
            runtime.secondary().block_on(async {
                rs::logging::init();
            });
        }
624

625
626
627
628
629
630
631
632
633
634
        // NATS is used for more than just the NATS request-plane:
        // - KV router events (JetStream or NATS core + local indexer)
        // - inter-router replica sync (NATS core)
        //
        // NATS initialization logic:
        // 1. If request_plane is NATS, always enable NATS
        // 2. Otherwise, use enable_nats parameter (defaults to true for backward compat)
        //    Pass false to disable NATS (e.g., for approximate KV routing mode)
        let enable_nats = enable_nats.unwrap_or(true); // Default to true

635
        let runtime_config = DistributedConfig {
636
            discovery_backend: discovery_backend_config,
637
            nats_config: if request_plane.is_nats() || enable_nats {
638
639
640
641
                Some(dynamo_runtime::transports::nats::ClientOptions::default())
            } else {
                None
            },
642
            request_plane,
643
644
645
646
647
        };
        let inner = runtime
            .secondary()
            .block_on(rs::DistributedRuntime::new(runtime, runtime_config))
            .map_err(to_pyerr)?;
648
649
650
651

        Ok(DistributedRuntime { inner, event_loop })
    }

Ryan Olson's avatar
Ryan Olson committed
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
    #[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(),
        })
    }

667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
    /// Get an endpoint directly by path (e.g., "namespace.component.endpoint" or "dyn://...").
    fn endpoint(&self, path: String) -> PyResult<Endpoint> {
        let trimmed_path = path.trim_start_matches("dyn://");
        let parts: Vec<&str> = trimmed_path.split('.').collect();

        if parts.len() != 3 {
            return Err(PyErr::new::<pyo3::exceptions::PyValueError, _>(format!(
                "Invalid endpoint path '{}'. Expected format: 'namespace.component.endpoint' or 'dyn://namespace.component.endpoint'",
                path
            )));
        }

        let namespace_name = parts[0];
        let component_name = parts[1];
        let endpoint_name = parts[2];

        // Get endpoint using existing chain
        let namespace = self
            .inner
            .namespace(namespace_name.to_string())
            .map_err(to_pyerr)?;
        let component = namespace
            .component(component_name.to_string())
            .map_err(to_pyerr)?;
        let endpoint = component.endpoint(endpoint_name.to_string());

        Ok(Endpoint {
            inner: endpoint,
695
696
697
698
699
            event_loop: self.event_loop.clone(),
        })
    }

    fn shutdown(&self) {
700
        self.inner.shutdown();
701
702
703
704
705
    }

    fn event_loop(&self) -> PyObject {
        self.event_loop.clone()
    }
706

707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
    /// Register an async Python callback for /engine/{route_name}
    ///
    /// Args:
    ///     route_name: Route path (e.g., "start_profile" → /engine/start_profile)
    ///     callback: Async function with signature: async def(body: dict) -> dict
    ///
    /// Example:
    /// ```python
    /// async def start_profile(body: dict) -> dict:
    ///     await engine.start_profile(**body)
    ///     return {"status": "ok"}
    ///
    /// runtime.register_engine_route("start_profile", start_profile)
    /// ```
    #[pyo3(signature = (route_name, callback))]
    fn register_engine_route(
        &self,
        py: Python<'_>,
        route_name: String,
        callback: PyObject,
    ) -> PyResult<()> {
        // Capture TaskLocals at registration time when Python's event loop is running.
        // This is needed because later, when the callback is invoked from an HTTP request,
        // we'll be on a Rust thread without a running Python event loop.
        let locals =
            Arc::new(pyo3_async_runtimes::tokio::get_current_locals(py).map_err(to_pyerr)?);
        let callback = Arc::new(callback);

        // Wrap Python async callback in Rust async closure
        let rust_callback: rs::engine_routes::EngineRouteCallback =
            Arc::new(move |body: serde_json::Value| {
                let callback = callback.clone();
                let locals = locals.clone();

                // Return a boxed future
                Box::pin(async move {
                    // Acquire GIL to call Python callback and convert coroutine to future
                    let py_future = Python::with_gil(|py| {
                        // Convert body to Python dict
                        let py_body = pythonize::pythonize(py, &body).map_err(|e| {
                            anyhow::anyhow!("Failed to convert request body to Python: {}", e)
                        })?;

                        // Call Python async function to get a coroutine
                        let coroutine = callback.call1(py, (py_body,)).map_err(|e| {
                            anyhow::anyhow!("Failed to call Python callback: {}", e)
                        })?;

                        // Use the TaskLocals captured at registration time
                        pyo3_async_runtimes::into_future_with_locals(
                            &locals,
                            coroutine.into_bound(py),
                        )
                        .map_err(|e| {
                            anyhow::anyhow!("Failed to convert coroutine to future: {}", e)
                        })
                    })?;

                    // Await the Python coroutine (GIL is released during await)
                    let py_result = py_future
                        .await
                        .map_err(|e| anyhow::anyhow!("Python callback failed: {}", e))?;

                    // Convert result back to serde_json::Value
                    Python::with_gil(|py| {
                        pythonize::depythonize::<serde_json::Value>(py_result.bind(py))
                            .map_err(|e| anyhow::anyhow!("Failed to serialize response: {}", e))
                    })
                })
            });

        self.inner
            .engine_routes()
            .register(&route_name, rust_callback);
        tracing::debug!("Registered engine route: /engine/{}", route_name);
        Ok(())
    }

785
786
787
788
789
790
791
792
793
794
795
    /// Set the system-level health status (Ready / NotReady).
    fn set_health_status(&self, ready: bool) -> PyResult<()> {
        let status = if ready {
            config::HealthStatus::Ready
        } else {
            config::HealthStatus::NotReady
        };
        self.inner.system_health().lock().set_health_status(status);
        Ok(())
    }

Richard Huo's avatar
Richard Huo committed
796
797
798
799
800
801
802
803
804
805
806
807
808
809
    // This is used to pass the DistributedRuntime from the dynamo-runtime bindings
    // to the KVBM bindings, since KVBM cannot directly use the struct from this cdylib.
    // TODO: Create a separate crate "dynamo-python" so that all binding crates can import
    // from it and share the same crate path. This will allow PyO3 to automatically
    // recognize that both bindings use the same PyClass.
    #[pyo3(name = "to_capsule")]
    fn to_capsule<'py>(&self, py: Python<'py>) -> PyResult<Bound<'py, PyCapsule>> {
        let arc: Arc<rs::DistributedRuntime> = Arc::new(self.inner.clone());
        let weak: Weak<rs::DistributedRuntime> = Arc::downgrade(&arc);

        let name = CString::new("dynamo.runtime.weak").expect("valid capsule name");

        PyCapsule::new(py, weak, Some(name))
    }
810
811
812
813
}

#[pymethods]
impl Endpoint {
814
    #[pyo3(signature = (generator, graceful_shutdown = true, metrics_labels = None, health_check_payload = None))]
815
816
817
818
    fn serve_endpoint<'p>(
        &self,
        py: Python<'p>,
        generator: PyObject,
819
        graceful_shutdown: Option<bool>,
820
        metrics_labels: Option<Vec<(String, String)>>,
821
        health_check_payload: Option<&Bound<'p, PyDict>>,
822
823
824
825
826
    ) -> PyResult<Bound<'p, PyAny>> {
        let engine = Arc::new(engine::PythonAsyncEngine::new(
            generator,
            self.event_loop.clone(),
        )?);
827
        let ingress = JsonServerStreamingIngress::for_engine(engine.clone()).map_err(to_pyerr)?;
828
829
830
831
832
833
834
835
836
837
838
839
840

        // 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
841
842
843
844
845
846
        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)",
            ));
847
848
849
        }

        let mut builder = self
850
851
852
853
            .inner
            .endpoint_builder()
            .metrics_labels(metrics_labels)
            .handler(ingress);
854
855
856
857
858

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

859
860
861
        // Register the engine in the local endpoint registry for in-process calls
        builder = builder.register_local_engine(engine).map_err(to_pyerr)?;

862
        let graceful_shutdown = graceful_shutdown.unwrap_or(true);
863
        pyo3_async_runtimes::tokio::future_into_py(py, async move {
864
865
866
867
868
            builder
                .graceful_shutdown(graceful_shutdown)
                .start()
                .await
                .map_err(to_pyerr)?;
869
870
871
872
            Ok(())
        })
    }

873
874
875
876
877
878
879
    #[pyo3(signature = (router_mode = None))]
    fn client<'p>(
        &self,
        py: Python<'p>,
        router_mode: Option<RouterMode>,
    ) -> PyResult<Bound<'p, PyAny>> {
        let router_mode = router_mode.unwrap_or(RouterMode::RoundRobin);
880
881
        let inner = self.inner.clone();
        pyo3_async_runtimes::tokio::future_into_py(py, async move {
882
            let client = inner.client().await.map_err(to_pyerr)?;
883
884
885
            let push_router = rs::pipeline::PushRouter::<
                serde_json::Value,
                RsAnnotated<serde_json::Value>,
886
            >::from_client(client, router_mode.into())
887
888
            .await
            .map_err(to_pyerr)?;
889
890
891
            Ok(Client {
                router: push_router,
            })
892
893
        })
    }
894

895
896
897
    // Opaque unique ID for this worker. May change over worker lifetime.
    fn connection_id(&self) -> u64 {
        self.inner.drt().connection_id()
898
    }
899
900
901

    /// Get a RuntimeMetrics helper for creating Prometheus metrics
    #[getter]
902
903
    fn metrics(&self) -> prometheus_metrics::RuntimeMetrics {
        prometheus_metrics::RuntimeMetrics::from_endpoint(self.inner.clone())
904
    }
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933

    /// Unregister this endpoint instance from discovery.
    ///
    /// This removes the endpoint from the instances bucket, preventing the router
    /// from sending requests to this worker. Use this when a worker is sleeping
    /// and should not receive any requests.
    fn unregister_endpoint_instance<'p>(&self, py: Python<'p>) -> PyResult<Bound<'p, PyAny>> {
        let inner = self.inner.clone();
        pyo3_async_runtimes::tokio::future_into_py(py, async move {
            inner
                .unregister_endpoint_instance()
                .await
                .map_err(to_pyerr)?;
            Ok(())
        })
    }

    /// Re-register this endpoint instance to discovery.
    ///
    /// This adds the endpoint back to the instances bucket, allowing the router
    /// to send requests to this worker again. Use this when a worker wakes up
    /// and should start receiving requests.
    fn register_endpoint_instance<'p>(&self, py: Python<'p>) -> PyResult<Bound<'p, PyAny>> {
        let inner = self.inner.clone();
        pyo3_async_runtimes::tokio::future_into_py(py, async move {
            inner.register_endpoint_instance().await.map_err(to_pyerr)?;
            Ok(())
        })
    }
934
935
}

936
937
938
939
940
941
942
943
944
945
946
947
948
#[pymethods]
impl ModelCardInstanceId {
    // (namespace, component, endpoint)
    // TODO: Can these be borrowed as &str?
    fn triple(&self) -> (String, String, String) {
        (
            self.inner.namespace.clone(),
            self.inner.component.clone(),
            self.inner.endpoint.clone(),
        )
    }
}

949
950
#[pymethods]
impl Client {
951
952
    /// Get list of current instances.
    /// Replaces endpoint_ids.
953
    fn instance_ids(&self) -> Vec<u64> {
954
        self.router.client.instance_ids()
955
956
    }

957
958
959
    /// 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>> {
960
        let inner = self.router.client.clone();
961
        pyo3_async_runtimes::tokio::future_into_py(py, async move {
962
            inner
963
                .wait_for_instances()
964
                .await
965
                .map(|v| v.into_iter().map(|cei| cei.id()).collect::<Vec<u64>>())
966
                .map_err(to_pyerr)
967
968
969
970
        })
    }

    /// Issue a request to the endpoint using the default routing strategy.
971
    #[pyo3(signature = (request, annotated=DEFAULT_ANNOTATED_SETTING, context=None))]
972
973
974
975
976
    fn generate<'p>(
        &self,
        py: Python<'p>,
        request: PyObject,
        annotated: Option<bool>,
977
        context: Option<context::Context>,
978
    ) -> PyResult<Bound<'p, PyAny>> {
979
        self.random(py, request, annotated, context)
980
981
982
    }

    /// Send a request to the next endpoint in a round-robin fashion.
983
    #[pyo3(signature = (request, annotated=DEFAULT_ANNOTATED_SETTING, context=None))]
984
985
986
987
988
    fn round_robin<'p>(
        &self,
        py: Python<'p>,
        request: PyObject,
        annotated: Option<bool>,
989
        context: Option<context::Context>,
990
991
    ) -> PyResult<Bound<'p, PyAny>> {
        let request: serde_json::Value = pythonize::depythonize(&request.into_bound(py))?;
992
        let request_ctx = create_request_context(request, &context);
993
994
995
        let annotated = annotated.unwrap_or(false);

        let (tx, rx) = tokio::sync::mpsc::channel(32);
996
        let client = self.router.clone();
997
998

        pyo3_async_runtimes::tokio::future_into_py(py, async move {
999
1000
            let stream = match context {
                Some(context) => {
1001
1002
                    // Always instrument with appropriate span (none if no trace context)
                    let span = get_span_for_context(&context, "round_robin");
1003
1004
1005
1006
1007
                    client
                        .round_robin(request_ctx)
                        .instrument(span)
                        .await
                        .map_err(to_pyerr)?
1008
                }
1009
                _ => client.round_robin(request_ctx).await.map_err(to_pyerr)?,
1010
            };
1011
            tokio::spawn(process_stream(stream, tx));
1012
            Ok(AsyncResponseStream::new(rx, annotated))
1013
1014
1015
1016
        })
    }

    /// Send a request to a random endpoint.
1017
    #[pyo3(signature = (request, annotated=DEFAULT_ANNOTATED_SETTING, context=None))]
1018
1019
1020
1021
1022
    fn random<'p>(
        &self,
        py: Python<'p>,
        request: PyObject,
        annotated: Option<bool>,
1023
        context: Option<context::Context>,
1024
1025
    ) -> PyResult<Bound<'p, PyAny>> {
        let request: serde_json::Value = pythonize::depythonize(&request.into_bound(py))?;
1026
        let request_ctx = create_request_context(request, &context);
1027
1028
1029
        let annotated = annotated.unwrap_or(false);

        let (tx, rx) = tokio::sync::mpsc::channel(32);
1030
        let client = self.router.clone();
1031
1032

        pyo3_async_runtimes::tokio::future_into_py(py, async move {
1033
1034
            let stream = match context {
                Some(context) => {
1035
                    // Always instrument with appropriate span (none if no trace context)
1036
                    let span = get_span_for_context(&context, "random");
1037
1038
1039
1040
1041
                    client
                        .random(request_ctx)
                        .instrument(span)
                        .await
                        .map_err(to_pyerr)?
1042
                }
1043
                _ => client.random(request_ctx).await.map_err(to_pyerr)?,
1044
            };
1045
            tokio::spawn(process_stream(stream, tx));
1046
            Ok(AsyncResponseStream::new(rx, annotated))
1047
1048
1049
1050
        })
    }

    /// Directly send a request to a specific endpoint.
1051
    #[pyo3(signature = (request, instance_id, annotated=DEFAULT_ANNOTATED_SETTING, context=None))]
1052
1053
1054
1055
    fn direct<'p>(
        &self,
        py: Python<'p>,
        request: PyObject,
1056
        instance_id: u64,
1057
        annotated: Option<bool>,
1058
        context: Option<context::Context>,
1059
1060
    ) -> PyResult<Bound<'p, PyAny>> {
        let request: serde_json::Value = pythonize::depythonize(&request.into_bound(py))?;
1061
        let request_ctx = create_request_context(request, &context);
1062
1063
1064
        let annotated = annotated.unwrap_or(false);

        let (tx, rx) = tokio::sync::mpsc::channel(32);
1065
        let client = self.router.clone();
1066
1067

        pyo3_async_runtimes::tokio::future_into_py(py, async move {
1068
1069
            let stream = match context {
                Some(context) => {
1070
                    // Always instrument with appropriate span (none if no trace context)
1071
1072
                    let span =
                        get_span_for_direct_context(&context, "direct", &instance_id.to_string());
1073
1074
1075
1076
1077
                    client
                        .direct(request_ctx, instance_id)
                        .instrument(span)
                        .await
                        .map_err(to_pyerr)?
1078
1079
                }
                _ => client
1080
                    .direct(request_ctx, instance_id)
1081
1082
1083
                    .await
                    .map_err(to_pyerr)?,
            };
1084
1085
1086

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

1087
            Ok(AsyncResponseStream::new(rx, annotated))
1088
1089
1090
1091
1092
        })
    }
}

async fn process_stream(
1093
    stream: EngineStream<RsAnnotated<serde_json::Value>>,
1094
1095
1096
1097
1098
    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
1099
        let annotated: RsAnnotated<serde_json::Value> = response;
1100
        let annotated: RsAnnotated<PyObject> = annotated.map_data(|data| {
1101
            Python::with_gil(|py| match pythonize::pythonize(py, &data) {
1102
1103
                Ok(pyobj) => Ok(pyobj.into()),
                Err(e) => Err(e.to_string()),
1104
            })
1105
1106
1107
1108
1109
1110
        });

        let is_error = annotated.is_error();

        // Send the PyObject through the channel or log an error
        if let Err(e) = tx.send(annotated).await {
1111
            tracing::error!("Failed to send response: {:?}", e);
1112
            break;
1113
1114
1115
1116
1117
1118
1119
1120
1121
        }

        if is_error {
            break;
        }
    }
}

#[pyclass]
1122
pub(crate) struct AsyncResponseStream {
1123
1124
1125
    rx: Arc<Mutex<tokio::sync::mpsc::Receiver<RsAnnotated<PyObject>>>>,
    annotated: bool,
}
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137

impl AsyncResponseStream {
    pub(crate) fn new(
        rx: tokio::sync::mpsc::Receiver<RsAnnotated<PyObject>>,
        annotated: bool,
    ) -> Self {
        Self {
            rx: Arc::new(Mutex::new(rx)),
            annotated,
        }
    }
}
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
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234

#[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")
        )
    }
}