lib.rs 43.6 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
    LeastLoaded,
55
56
57
58
59
60
61
}

impl From<RouterMode> for RsRouterMode {
    fn from(mode: RouterMode) -> Self {
        match mode {
            RouterMode::RoundRobin => Self::RoundRobin,
            RouterMode::Random => Self::Random,
62
            RouterMode::PowerOfTwoChoices => Self::PowerOfTwoChoices,
63
            RouterMode::KV => Self::KV,
64
            RouterMode::Direct => Self::Direct,
65
            RouterMode::LeastLoaded => Self::LeastLoaded,
66
67
68
        }
    }
}
69

70
mod context;
71
mod engine;
72
pub mod errors;
73
mod http;
74
mod kserve_grpc;
75
mod llm;
76
mod parsers;
77
mod planner;
78
mod prometheus_metrics;
79
80
81
82
83
84
85

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);
86
87
const SKIP_PYTHON_LOG_INIT_ENV: &str = "DYNAMO_SKIP_PYTHON_LOG_INIT";

88
89
// Helper to get appropriate span for instrumentation - always emit spans
fn get_span_for_context(context: &context::Context, operation: &str) -> tracing::Span {
90
91
92
93
94
95
    logging::make_client_request_span(
        operation,
        context.inner().id(),
        context.trace_context(),
        None,
    )
96
97
98
99
100
101
102
103
}

// 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 {
104
105
106
107
108
109
    logging::make_client_request_span(
        operation,
        context.inner().id(),
        context.trace_context(),
        Some(instance_id),
    )
110
111
}

112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
// 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(),
    }
}

135
136
137
138
139
/// 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<()> {
140
    // Initialize logging early unless OTEL export is enabled (which requires tokio runtime)
141
    if config::env_is_truthy(env_otlp::OTEL_EXPORT_ENABLED) {
142
        eprintln!(
143
            "Warning: OTEL_EXPORT_ENABLED detected. Logging initialization deferred until runtime is available. Early logs may be dropped."
144
        );
145
    } else if std::env::var_os(SKIP_PYTHON_LOG_INIT_ENV).is_none() {
146
147
148
        rs::logging::init();
    }

Yan Ru Pei's avatar
Yan Ru Pei committed
149
    m.add_function(wrap_pyfunction!(llm::kv::compute_block_hash_for_seq_py, m)?)?;
150
    m.add_function(wrap_pyfunction!(lora_name_to_id, m)?)?;
Ryan Olson's avatar
Ryan Olson committed
151
    m.add_function(wrap_pyfunction!(log_message, m)?)?;
152
153
154
    m.add_function(wrap_pyfunction!(register_model, m)?)?;
    m.add_function(wrap_pyfunction!(unregister_model, m)?)?;
    m.add_function(wrap_pyfunction!(fetch_model, m)?)?;
155
    m.add_function(wrap_pyfunction!(run_kv_indexer, m)?)?;
156
    m.add_function(wrap_pyfunction!(llm::entrypoint::make_engine, m)?)?;
157
    m.add_function(wrap_pyfunction!(llm::replay::run_mocker_trace_replay, m)?)?;
158
    m.add_function(wrap_pyfunction!(
159
        llm::replay::run_mocker_synthetic_trace_replay,
160
161
        m
    )?)?;
162
    m.add_function(wrap_pyfunction!(llm::entrypoint::run_input, m)?)?;
163

164
165
    m.add_class::<DistributedRuntime>()?;
    m.add_class::<Endpoint>()?;
166
    m.add_class::<ModelCardInstanceId>()?;
167
168
    m.add_class::<Client>()?;
    m.add_class::<AsyncResponseStream>()?;
169
170
171
    m.add_class::<llm::entrypoint::EntrypointArgs>()?;
    m.add_class::<llm::entrypoint::EngineConfig>()?;
    m.add_class::<llm::entrypoint::EngineType>()?;
172
    m.add_class::<llm::entrypoint::AicPerfConfig>()?;
173
174
    m.add_class::<llm::entrypoint::RouterConfig>()?;
    m.add_class::<llm::entrypoint::KvRouterConfig>()?;
175
176
177
    m.add_class::<llm::replay::ReasoningConfig>()?;
    m.add_class::<llm::replay::SglangArgs>()?;
    m.add_class::<llm::replay::MockEngineArgs>()?;
178
    m.add_class::<llm::kv::WorkerMetricsPublisher>()?;
179
    m.add_class::<llm::model_card::ModelDeploymentCard>()?; // Internal: only in _internal, not public API
180
    m.add_class::<llm::local_model::ModelRuntimeConfig>()?;
181
182
    m.add_class::<llm::preprocessor::MediaDecoder>()?;
    m.add_class::<llm::preprocessor::MediaFetcher>()?;
183
    m.add_class::<llm::kv::OverlapScores>()?;
184
    m.add_class::<llm::kv::KvEventPublisher>()?;
Yan Ru Pei's avatar
Yan Ru Pei committed
185
    m.add_class::<llm::kv::RadixTree>()?;
186
187
    m.add_class::<llm::fpm::FpmEventRelay>()?;
    m.add_class::<llm::fpm::FpmEventSubscriber>()?;
188
    m.add_class::<llm::lora::LoRADownloader>()?;
189
190
    m.add_class::<http::HttpService>()?;
    m.add_class::<http::HttpAsyncEngine>()?;
191
    m.add_class::<context::Context>()?;
192
    m.add_class::<ModelType>()?;
193
    m.add_class::<ModelInput>()?;
194
    m.add_class::<llm::kv::KvRouter>()?;
195
    m.add_class::<RouterMode>()?;
196
    m.add_class::<kserve_grpc::KserveGrpcService>()?;
197
    m.add("__version__", env!("CARGO_PKG_VERSION"))?;
198
199
200
    m.add_class::<planner::VirtualConnectorCoordinator>()?;
    m.add_class::<planner::VirtualConnectorClient>()?;
    m.add_class::<planner::PlannerDecision>()?;
201
202

    engine::add_to_module(m)?;
203
    errors::register_exceptions(m)?;
204
    parsers::add_to_module(m)?;
205

206
    m.add_class::<prometheus_metrics::RuntimeMetrics>()?;
207
208
209
210
    let prometheus_metrics = PyModule::new(m.py(), "prometheus_metrics")?;
    prometheus_metrics::add_to_module(&prometheus_metrics)?;
    m.add_submodule(&prometheus_metrics)?;

211
212
213
214
215
216
217
218
219
220
    Ok(())
}

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

221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
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
239
240
241
242
243
244
245
/// 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);
}

246
247
248
249
250
251
252
/// 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)
}

253
254
/// 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.
255
256
257
258
259
260
261
///
/// 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.
262
#[pyfunction]
263
#[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))]
264
#[allow(clippy::too_many_arguments)]
265
fn register_model<'p>(
266
    py: Python<'p>,
267
    model_input: ModelInput,
268
    model_type: ModelType,
269
270
271
    endpoint: Endpoint,
    model_path: &str,
    model_name: Option<&str>,
272
273
    context_length: Option<u32>,
    kv_cache_block_size: Option<u32>,
274
    router_mode: Option<RouterMode>,
275
    runtime_config: Option<ModelRuntimeConfig>,
276
    user_data: Option<&Bound<'p, PyDict>>,
277
    custom_template_path: Option<&str>,
278
279
    media_decoder: Option<MediaDecoder>,
    media_fetcher: Option<MediaFetcher>,
280
281
    lora_name: Option<&str>,
    base_model_path: Option<&str>,
282
) -> PyResult<Bound<'p, PyAny>> {
283
    // Validate Prefill model type requirements
284
285
286
287
288
289
    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",
        ));
290
291
    }

292
293
294
    let model_input = match model_input {
        ModelInput::Text => llm_rs::model_type::ModelInput::Text,
        ModelInput::Tokens => llm_rs::model_type::ModelInput::Tokens,
295
        ModelInput::Tensor => llm_rs::model_type::ModelInput::Tensor,
296
297
    };

298
    let is_tensor_based = model_type.inner.supports_tensor();
299
    let is_images = model_type.inner.supports_images();
300
    let is_videos = model_type.inner.supports_videos();
301

302
303
    let model_type_obj = model_type.inner;

304
    let inner_path = model_path.to_string();
305
    let model_name = model_name.map(|n| n.to_string());
306
307
308
    let router_mode = router_mode.unwrap_or(RouterMode::RoundRobin);
    let router_config = RouterConfig::new(router_mode.into(), KvRouterConfig::default());

309
310
311
312
313
314
315
316
317
318
319
320
321
    // 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()?;

322
323
324
325
326
327
328
    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))
        })?;

329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
    // 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()));

348
    pyo3_async_runtimes::tokio::future_into_py(py, async move {
349
350
351
        // 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 {
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
            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)
377
378
        let model_path = if fs::exists(&source_path)? {
            PathBuf::from(&source_path)
379
        } else {
380
            LocalModel::fetch(&source_path, false)
381
382
383
384
                .await
                .map_err(to_pyerr)?
        };

385
386
        let mut builder = dynamo_llm::local_model::LocalModelBuilder::default();
        builder
387
            // model path is the physical path on disk of the downloaded model
388
            .model_path(model_path)
389
390
391
392
            // 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
393
            .model_name(model_name.clone())
394
            .context_length(context_length)
395
            .kv_cache_block_size(kv_cache_block_size)
396
            .router_config(Some(router_config))
397
            .runtime_config(runtime_config.unwrap_or_default().inner)
398
            .user_data(user_data_json)
399
400
401
            .custom_template_path(custom_template_path_owned)
            .media_decoder(media_decoder.map(|m| m.inner))
            .media_fetcher(media_fetcher.map(|m| m.inner));
402

403
        let mut local_model = builder.build().await.map_err(to_pyerr)?;
404
405
406
407
408
409
410
411
412

        // 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,
            });

413
        local_model
414
            .attach(&endpoint.inner, model_type_obj, model_input, lora_info)
415
416
417
            .await
            .map_err(to_pyerr)?;

418
419
420
        if let Some(lora_name) = lora_identifier {
            tracing::info!("Registered LoRA '{}' MDC", lora_name);
        } else {
421
422
423
424
            tracing::info!(
                "Registered base model '{}' MDC",
                model_name.unwrap_or(source_path)
            );
425
426
        }

427
428
429
430
        Ok(())
    })
}

431
432
433
434
435
436
437
438
439
440
441
442
443
/// 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}`
444
#[pyfunction]
445
#[pyo3(signature = (endpoint, lora_name=None))]
446
fn unregister_model<'p>(
447
448
449
450
451
452
    py: Python<'p>,
    endpoint: Endpoint,
    lora_name: Option<&str>,
) -> PyResult<Bound<'p, PyAny>> {
    let lora_name_owned = lora_name.map(|s| s.to_string());

453
    pyo3_async_runtimes::tokio::future_into_py(py, async move {
454
455
        // Unified detach method handles both base models and LoRA adapters
        LocalModel::detach_from_endpoint(&endpoint.inner, lora_name_owned.as_deref())
456
457
458
459
460
461
            .await
            .map_err(to_pyerr)?;
        Ok(())
    })
}

462
463
/// Download a model from Hugging Face, returning its local path
/// Example: `model_path = await fetch_model("Qwen/Qwen3-0.6B")`
464
#[pyfunction]
465
#[pyo3(signature = (remote_name, ignore_weights=false))]
466
fn fetch_model<'p>(
467
468
469
470
    py: Python<'p>,
    remote_name: &str,
    ignore_weights: bool,
) -> PyResult<Bound<'p, PyAny>> {
471
472
    let repo = remote_name.to_string();
    pyo3_async_runtimes::tokio::future_into_py(py, async move {
473
474
475
        LocalModel::fetch(&repo, ignore_weights)
            .await
            .map_err(to_pyerr)
476
477
478
    })
}

479
480
#[pyclass]
#[derive(Clone)]
Ryan Olson's avatar
Ryan Olson committed
481
pub struct DistributedRuntime {
482
483
484
485
    inner: rs::DistributedRuntime,
    event_loop: PyObject,
}

Ryan Olson's avatar
Ryan Olson committed
486
487
impl DistributedRuntime {
    #[allow(dead_code)]
488
    pub(crate) fn inner(&self) -> &rs::DistributedRuntime {
Ryan Olson's avatar
Ryan Olson committed
489
490
491
492
        &self.inner
    }
}

493
#[pyclass]
494
#[derive(Clone)]
495
496
497
498
499
struct CancellationToken {
    inner: rs::CancellationToken,
}

#[pyclass]
500
#[derive(Clone)]
501
502
503
504
505
struct Endpoint {
    inner: rs::component::Endpoint,
    event_loop: PyObject,
}

506
507
508
509
510
511
#[pyclass]
#[derive(Clone)]
struct ModelCardInstanceId {
    inner: rs::discovery::ModelCardInstanceId,
}

512
#[pyclass]
513
#[derive(Clone)]
514
struct Client {
515
    router: rs::pipeline::PushRouter<serde_json::Value, RsAnnotated<serde_json::Value>>,
516
517
}

518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
#[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,
    };
539
540
541
542
    #[classattr]
    const TensorBased: Self = ModelType {
        inner: llm_rs::model_type::ModelType::TensorBased,
    };
543
544
545
546
    #[classattr]
    const Prefill: Self = ModelType {
        inner: llm_rs::model_type::ModelType::Prefill,
    };
547
548
549
550
    #[classattr]
    const Images: Self = ModelType {
        inner: llm_rs::model_type::ModelType::Images,
    };
551
    #[classattr]
552
553
554
555
    const Audios: Self = ModelType {
        inner: llm_rs::model_type::ModelType::Audios,
    };
    #[classattr]
556
557
558
    const Videos: Self = ModelType {
        inner: llm_rs::model_type::ModelType::Videos,
    };
559

560
561
562
563
    fn supports_chat(&self) -> bool {
        self.inner.supports_chat()
    }

564
565
566
567
568
569
570
571
572
573
574
    fn __or__(&self, other: &Self) -> Self {
        ModelType {
            inner: self.inner | other.inner,
        }
    }

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

575
576
#[pyclass(eq, eq_int)]
#[derive(Clone, PartialEq)]
577
578
579
enum ModelInput {
    Text = 1,
    Tokens = 2,
580
    Tensor = 3,
581
582
}

583
584
585
#[pymethods]
impl DistributedRuntime {
    #[new]
586
    #[pyo3(signature = (event_loop, discovery_backend, request_plane, enable_nats=None))]
587
588
    fn new(
        event_loop: PyObject,
589
        discovery_backend: String,
590
591
592
        request_plane: String,
        enable_nats: Option<bool>,
    ) -> PyResult<Self> {
593
594
595
596
597
598
599
        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)
            }
        };
600
        let request_plane: RequestPlaneMode = request_plane.parse().map_err(to_pyerr)?;
601

602
603
604
        // 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()
605
            .or_else(|_| -> anyhow::Result<rs::Runtime> {
606
607
608
609
                // No existing Worker, create new one
                let worker = rs::Worker::from_settings()?;

                // Initialize pyo3 bridge (only happens once per process)
610
                INIT.get_or_try_init(|| -> anyhow::Result<()> {
611
612
                    let primary = worker.tokio_runtime()?;
                    pyo3_async_runtimes::tokio::init_with_runtime(primary).map_err(|e| {
613
                        anyhow::anyhow!("failed to initialize pyo3 static runtime: {:?}", e)
614
                    })?;
615
                    Ok(())
616
617
                })?;

618
                Ok(worker.runtime().clone())
619
620
            })
            .map_err(to_pyerr)?;
621

622
623
        // Initialize logging in context where tokio runtime is available
        // otel exporter requires it
624
        if config::env_is_truthy(env_otlp::OTEL_EXPORT_ENABLED) {
625
626
627
628
            runtime.secondary().block_on(async {
                rs::logging::init();
            });
        }
629

630
631
632
633
634
635
636
637
638
639
        // 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

640
        let runtime_config = DistributedConfig {
641
            discovery_backend: discovery_backend_config,
642
            nats_config: if request_plane.is_nats() || enable_nats {
643
644
645
646
                Some(dynamo_runtime::transports::nats::ClientOptions::default())
            } else {
                None
            },
647
            request_plane,
648
649
650
651
652
        };
        let inner = runtime
            .secondary()
            .block_on(rs::DistributedRuntime::new(runtime, runtime_config))
            .map_err(to_pyerr)?;
653
654
655
656

        Ok(DistributedRuntime { inner, event_loop })
    }

Ryan Olson's avatar
Ryan Olson committed
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
    #[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(),
        })
    }

672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
    /// 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,
700
701
702
703
704
            event_loop: self.event_loop.clone(),
        })
    }

    fn shutdown(&self) {
705
        self.inner.shutdown();
706
707
708
709
710
    }

    fn event_loop(&self) -> PyObject {
        self.event_loop.clone()
    }
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
785
786
787
788
789
    /// 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(())
    }

790
791
792
793
794
795
796
797
798
799
800
    /// 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
801
802
803
804
805
806
807
808
809
810
811
812
813
814
    // 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))
    }
815
816
817
818
}

#[pymethods]
impl Endpoint {
819
    #[pyo3(signature = (generator, graceful_shutdown = true, metrics_labels = None, health_check_payload = None))]
820
821
822
823
    fn serve_endpoint<'p>(
        &self,
        py: Python<'p>,
        generator: PyObject,
824
        graceful_shutdown: Option<bool>,
825
        metrics_labels: Option<Vec<(String, String)>>,
826
        health_check_payload: Option<&Bound<'p, PyDict>>,
827
828
829
830
831
    ) -> PyResult<Bound<'p, PyAny>> {
        let engine = Arc::new(engine::PythonAsyncEngine::new(
            generator,
            self.event_loop.clone(),
        )?);
832
        let ingress = JsonServerStreamingIngress::for_engine(engine.clone()).map_err(to_pyerr)?;
833
834
835
836
837
838
839
840
841
842
843
844
845

        // 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
846
847
848
849
850
851
        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)",
            ));
852
853
854
        }

        let mut builder = self
855
856
857
858
            .inner
            .endpoint_builder()
            .metrics_labels(metrics_labels)
            .handler(ingress);
859
860
861
862
863

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

864
865
866
        // Register the engine in the local endpoint registry for in-process calls
        builder = builder.register_local_engine(engine).map_err(to_pyerr)?;

867
        let graceful_shutdown = graceful_shutdown.unwrap_or(true);
868
        pyo3_async_runtimes::tokio::future_into_py(py, async move {
869
870
871
872
873
            builder
                .graceful_shutdown(graceful_shutdown)
                .start()
                .await
                .map_err(to_pyerr)?;
874
875
876
877
            Ok(())
        })
    }

878
879
880
881
882
883
884
    #[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);
885
886
        let inner = self.inner.clone();
        pyo3_async_runtimes::tokio::future_into_py(py, async move {
887
            let client = inner.client().await.map_err(to_pyerr)?;
888
889
890
            let push_router = rs::pipeline::PushRouter::<
                serde_json::Value,
                RsAnnotated<serde_json::Value>,
891
            >::from_client(client, router_mode.into())
892
893
            .await
            .map_err(to_pyerr)?;
894
895
896
            Ok(Client {
                router: push_router,
            })
897
898
        })
    }
899

900
901
902
    // Opaque unique ID for this worker. May change over worker lifetime.
    fn connection_id(&self) -> u64 {
        self.inner.drt().connection_id()
903
    }
904
905
906

    /// Get a RuntimeMetrics helper for creating Prometheus metrics
    #[getter]
907
908
    fn metrics(&self) -> prometheus_metrics::RuntimeMetrics {
        prometheus_metrics::RuntimeMetrics::from_endpoint(self.inner.clone())
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
934
935
936
937
938

    /// 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(())
        })
    }
939
940
}

941
942
943
944
945
946
947
948
949
950
951
952
953
#[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(),
        )
    }
}

954
955
#[pymethods]
impl Client {
956
957
    /// Get list of current instances.
    /// Replaces endpoint_ids.
958
    fn instance_ids(&self) -> Vec<u64> {
959
        self.router.client.instance_ids()
960
961
    }

962
963
964
    /// 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>> {
965
        let inner = self.router.client.clone();
966
        pyo3_async_runtimes::tokio::future_into_py(py, async move {
967
            inner
968
                .wait_for_instances()
969
                .await
970
                .map(|v| v.into_iter().map(|cei| cei.id()).collect::<Vec<u64>>())
971
                .map_err(to_pyerr)
972
973
974
975
        })
    }

    /// Issue a request to the endpoint using the default routing strategy.
976
    #[pyo3(signature = (request, annotated=DEFAULT_ANNOTATED_SETTING, context=None))]
977
978
979
980
981
    fn generate<'p>(
        &self,
        py: Python<'p>,
        request: PyObject,
        annotated: Option<bool>,
982
        context: Option<context::Context>,
983
    ) -> PyResult<Bound<'p, PyAny>> {
984
        self.random(py, request, annotated, context)
985
986
987
    }

    /// Send a request to the next endpoint in a round-robin fashion.
988
    #[pyo3(signature = (request, annotated=DEFAULT_ANNOTATED_SETTING, context=None))]
989
990
991
992
993
    fn round_robin<'p>(
        &self,
        py: Python<'p>,
        request: PyObject,
        annotated: Option<bool>,
994
        context: Option<context::Context>,
995
996
    ) -> PyResult<Bound<'p, PyAny>> {
        let request: serde_json::Value = pythonize::depythonize(&request.into_bound(py))?;
997
        let request_ctx = create_request_context(request, &context);
998
999
1000
        let annotated = annotated.unwrap_or(false);

        let (tx, rx) = tokio::sync::mpsc::channel(32);
1001
        let client = self.router.clone();
1002
1003

        pyo3_async_runtimes::tokio::future_into_py(py, async move {
1004
1005
            let stream = match context {
                Some(context) => {
1006
1007
                    // Always instrument with appropriate span (none if no trace context)
                    let span = get_span_for_context(&context, "round_robin");
1008
1009
1010
1011
1012
                    client
                        .round_robin(request_ctx)
                        .instrument(span)
                        .await
                        .map_err(to_pyerr)?
1013
                }
1014
                _ => client.round_robin(request_ctx).await.map_err(to_pyerr)?,
1015
            };
1016
            tokio::spawn(process_stream(stream, tx));
1017
            Ok(AsyncResponseStream::new(rx, annotated))
1018
1019
1020
1021
        })
    }

    /// Send a request to a random endpoint.
1022
    #[pyo3(signature = (request, annotated=DEFAULT_ANNOTATED_SETTING, context=None))]
1023
1024
1025
1026
1027
    fn random<'p>(
        &self,
        py: Python<'p>,
        request: PyObject,
        annotated: Option<bool>,
1028
        context: Option<context::Context>,
1029
1030
    ) -> PyResult<Bound<'p, PyAny>> {
        let request: serde_json::Value = pythonize::depythonize(&request.into_bound(py))?;
1031
        let request_ctx = create_request_context(request, &context);
1032
1033
1034
        let annotated = annotated.unwrap_or(false);

        let (tx, rx) = tokio::sync::mpsc::channel(32);
1035
        let client = self.router.clone();
1036
1037

        pyo3_async_runtimes::tokio::future_into_py(py, async move {
1038
1039
            let stream = match context {
                Some(context) => {
1040
                    // Always instrument with appropriate span (none if no trace context)
1041
                    let span = get_span_for_context(&context, "random");
1042
1043
1044
1045
1046
                    client
                        .random(request_ctx)
                        .instrument(span)
                        .await
                        .map_err(to_pyerr)?
1047
                }
1048
                _ => client.random(request_ctx).await.map_err(to_pyerr)?,
1049
            };
1050
            tokio::spawn(process_stream(stream, tx));
1051
            Ok(AsyncResponseStream::new(rx, annotated))
1052
1053
1054
1055
        })
    }

    /// Directly send a request to a specific endpoint.
1056
    #[pyo3(signature = (request, instance_id, annotated=DEFAULT_ANNOTATED_SETTING, context=None))]
1057
1058
1059
1060
    fn direct<'p>(
        &self,
        py: Python<'p>,
        request: PyObject,
1061
        instance_id: u64,
1062
        annotated: Option<bool>,
1063
        context: Option<context::Context>,
1064
1065
    ) -> PyResult<Bound<'p, PyAny>> {
        let request: serde_json::Value = pythonize::depythonize(&request.into_bound(py))?;
1066
        let request_ctx = create_request_context(request, &context);
1067
1068
1069
        let annotated = annotated.unwrap_or(false);

        let (tx, rx) = tokio::sync::mpsc::channel(32);
1070
        let client = self.router.clone();
1071
1072

        pyo3_async_runtimes::tokio::future_into_py(py, async move {
1073
1074
            let stream = match context {
                Some(context) => {
1075
                    // Always instrument with appropriate span (none if no trace context)
1076
1077
                    let span =
                        get_span_for_direct_context(&context, "direct", &instance_id.to_string());
1078
1079
1080
1081
1082
                    client
                        .direct(request_ctx, instance_id)
                        .instrument(span)
                        .await
                        .map_err(to_pyerr)?
1083
1084
                }
                _ => client
1085
                    .direct(request_ctx, instance_id)
1086
1087
1088
                    .await
                    .map_err(to_pyerr)?,
            };
1089
1090
1091

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

1092
            Ok(AsyncResponseStream::new(rx, annotated))
1093
1094
1095
1096
1097
        })
    }
}

async fn process_stream(
1098
    stream: EngineStream<RsAnnotated<serde_json::Value>>,
1099
1100
1101
1102
1103
    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
1104
        let annotated: RsAnnotated<serde_json::Value> = response;
1105
        let annotated: RsAnnotated<PyObject> = annotated.map_data(|data| {
1106
            Python::with_gil(|py| match pythonize::pythonize(py, &data) {
1107
1108
                Ok(pyobj) => Ok(pyobj.into()),
                Err(e) => Err(e.to_string()),
1109
            })
1110
1111
1112
1113
1114
1115
        });

        let is_error = annotated.is_error();

        // Send the PyObject through the channel or log an error
        if let Err(e) = tx.send(annotated).await {
1116
            tracing::error!("Failed to send response: {:?}", e);
1117
            break;
1118
1119
1120
1121
1122
1123
1124
1125
1126
        }

        if is_error {
            break;
        }
    }
}

#[pyclass]
1127
pub(crate) struct AsyncResponseStream {
1128
1129
1130
    rx: Arc<Mutex<tokio::sync::mpsc::Receiver<RsAnnotated<PyObject>>>>,
    annotated: bool,
}
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142

impl AsyncResponseStream {
    pub(crate) fn new(
        rx: tokio::sync::mpsc::Receiver<RsAnnotated<PyObject>>,
        annotated: bool,
    ) -> Self {
        Self {
            rx: Arc::new(Mutex::new(rx)),
            annotated,
        }
    }
}
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
1235
1236
1237
1238
1239

#[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")
        )
    }
}