lib.rs 40.9 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::{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
};

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

40
use crate::llm::local_model::ModelRuntimeConfig;
41
use crate::llm::preprocessor::{MediaDecoder, MediaFetcher};
42

43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
#[pyclass(eq, eq_int)]
#[derive(Clone, Debug, PartialEq)]
pub enum RouterMode {
    RoundRobin,
    Random,
    KV,
}

impl From<RouterMode> for RsRouterMode {
    fn from(mode: RouterMode) -> Self {
        match mode {
            RouterMode::RoundRobin => Self::RoundRobin,
            RouterMode::Random => Self::Random,
            RouterMode::KV => Self::KV,
        }
    }
}
60

61
mod context;
62
mod engine;
63
mod http;
64
mod kserve_grpc;
65
mod llm;
66
mod parsers;
67
mod planner;
68
mod prometheus_metrics;
69
70
71
72
73
74
75
76

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);

77
78
// Helper to get appropriate span for instrumentation - always emit spans
fn get_span_for_context(context: &context::Context, operation: &str) -> tracing::Span {
79
80
81
82
83
84
    logging::make_client_request_span(
        operation,
        context.inner().id(),
        context.trace_context(),
        None,
    )
85
86
87
88
89
90
91
92
}

// 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 {
93
94
95
96
97
98
    logging::make_client_request_span(
        operation,
        context.inner().id(),
        context.trace_context(),
        Some(instance_id),
    )
99
100
}

101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
// 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(),
    }
}

124
125
126
127
128
/// 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<()> {
129
    // Initialize logging early unless OTEL export is enabled (which requires tokio runtime)
130
    if config::env_is_truthy(env_otlp::OTEL_EXPORT_ENABLED) {
131
        eprintln!(
132
            "Warning: OTEL_EXPORT_ENABLED detected. Logging initialization deferred until runtime is available. Early logs may be dropped."
133
134
135
136
137
        );
    } else {
        rs::logging::init();
    }

Yan Ru Pei's avatar
Yan Ru Pei committed
138
    m.add_function(wrap_pyfunction!(llm::kv::compute_block_hash_for_seq_py, m)?)?;
139
    m.add_function(wrap_pyfunction!(lora_name_to_id, m)?)?;
Ryan Olson's avatar
Ryan Olson committed
140
    m.add_function(wrap_pyfunction!(log_message, m)?)?;
141
    m.add_function(wrap_pyfunction!(register_llm, m)?)?;
142
    m.add_function(wrap_pyfunction!(unregister_llm, m)?)?;
143
    m.add_function(wrap_pyfunction!(fetch_llm, m)?)?;
144
145
    m.add_function(wrap_pyfunction!(llm::entrypoint::make_engine, m)?)?;
    m.add_function(wrap_pyfunction!(llm::entrypoint::run_input, m)?)?;
146

147
148
149
150
151
152
153
    m.add_class::<DistributedRuntime>()?;
    m.add_class::<CancellationToken>()?;
    m.add_class::<Namespace>()?;
    m.add_class::<Component>()?;
    m.add_class::<Endpoint>()?;
    m.add_class::<Client>()?;
    m.add_class::<AsyncResponseStream>()?;
154
155
156
    m.add_class::<llm::entrypoint::EntrypointArgs>()?;
    m.add_class::<llm::entrypoint::EngineConfig>()?;
    m.add_class::<llm::entrypoint::EngineType>()?;
157
158
    m.add_class::<llm::entrypoint::RouterConfig>()?;
    m.add_class::<llm::entrypoint::KvRouterConfig>()?;
159
    m.add_class::<llm::kv::WorkerMetricsPublisher>()?;
160
    m.add_class::<llm::model_card::ModelDeploymentCard>()?;
161
    m.add_class::<llm::local_model::ModelRuntimeConfig>()?;
162
163
    m.add_class::<llm::preprocessor::MediaDecoder>()?;
    m.add_class::<llm::preprocessor::MediaFetcher>()?;
164
165
    m.add_class::<llm::kv::OverlapScores>()?;
    m.add_class::<llm::kv::KvIndexer>()?;
166
    m.add_class::<llm::kv::ApproxKvIndexer>()?;
167
    m.add_class::<llm::kv::KvEventPublisher>()?;
Yan Ru Pei's avatar
Yan Ru Pei committed
168
169
    m.add_class::<llm::kv::RadixTree>()?;
    m.add_class::<llm::kv::ZmqKvEventListener>()?;
170
171
    m.add_class::<llm::kv::ZmqKvEventPublisher>()?;
    m.add_class::<llm::kv::ZmqKvEventPublisherConfig>()?;
172
    m.add_class::<llm::kv::KvRecorder>()?;
173
    m.add_class::<llm::lora::LoRADownloader>()?;
174
175
    m.add_class::<http::HttpService>()?;
    m.add_class::<http::HttpAsyncEngine>()?;
176
    m.add_class::<context::Context>()?;
177
    m.add_class::<ModelType>()?;
178
    m.add_class::<ModelInput>()?;
179
180
    m.add_class::<llm::kv::KvPushRouter>()?;
    m.add_class::<llm::kv::KvPushRouterStream>()?;
181
    m.add_class::<RouterMode>()?;
182
    m.add_class::<kserve_grpc::KserveGrpcService>()?;
183
    m.add("__version__", env!("CARGO_PKG_VERSION"))?;
184
185
186
    m.add_class::<planner::VirtualConnectorCoordinator>()?;
    m.add_class::<planner::VirtualConnectorClient>()?;
    m.add_class::<planner::PlannerDecision>()?;
187
188

    engine::add_to_module(m)?;
189
    parsers::add_to_module(m)?;
190

191
    m.add_class::<prometheus_metrics::RuntimeMetrics>()?;
192
193
194
195
    let prometheus_metrics = PyModule::new(m.py(), "prometheus_metrics")?;
    prometheus_metrics::add_to_module(&prometheus_metrics)?;
    m.add_submodule(&prometheus_metrics)?;

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

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

Ryan Olson's avatar
Ryan Olson committed
206
207
208
209
210
211
212
/// Log a message from Python with file and line info
#[pyfunction]
#[pyo3(text_signature = "(level, message, module, file, line)")]
fn log_message(level: &str, message: &str, module: &str, file: &str, line: u32) {
    logging::log_message(level, message, module, file, line);
}

213
214
215
216
217
218
219
/// 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)
}

220
221
/// 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.
222
223
224
225
226
227
228
///
/// 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.
229
#[pyfunction]
230
#[pyo3(signature = (model_input, model_type, endpoint, model_path, model_name=None, context_length=None, kv_cache_block_size=None, router_mode=None, migration_limit=0, runtime_config=None, user_data=None, custom_template_path=None, media_decoder=None, media_fetcher=None, lora_name=None, base_model_path=None))]
231
#[allow(clippy::too_many_arguments)]
232
233
fn register_llm<'p>(
    py: Python<'p>,
234
    model_input: ModelInput,
235
    model_type: ModelType,
236
237
238
    endpoint: Endpoint,
    model_path: &str,
    model_name: Option<&str>,
239
240
    context_length: Option<u32>,
    kv_cache_block_size: Option<u32>,
241
    router_mode: Option<RouterMode>,
242
    migration_limit: u32,
243
    runtime_config: Option<ModelRuntimeConfig>,
244
    user_data: Option<&Bound<'p, PyDict>>,
245
    custom_template_path: Option<&str>,
246
247
    media_decoder: Option<MediaDecoder>,
    media_fetcher: Option<MediaFetcher>,
248
249
    lora_name: Option<&str>,
    base_model_path: Option<&str>,
250
) -> PyResult<Bound<'p, PyAny>> {
251
252
253
254
255
256
257
258
259
260
261
262
263
264
    // Validate Prefill model type requirements
    if model_type.inner == llm_rs::model_type::ModelType::Prefill {
        if !matches!(model_input, ModelInput::Tokens) {
            return Err(PyErr::new::<pyo3::exceptions::PyValueError, _>(
                "ModelType::Prefill requires model_input to be ModelInput::Tokens",
            ));
        }
        if migration_limit != 0 {
            return Err(PyErr::new::<pyo3::exceptions::PyValueError, _>(
                "ModelType::Prefill requires migration_limit to be 0",
            ));
        }
    }

265
266
267
    let model_input = match model_input {
        ModelInput::Text => llm_rs::model_type::ModelInput::Text,
        ModelInput::Tokens => llm_rs::model_type::ModelInput::Tokens,
268
        ModelInput::Tensor => llm_rs::model_type::ModelInput::Tensor,
269
270
    };

271
    let is_tensor_based = model_type.inner.supports_tensor();
272
    let is_images = model_type.inner.supports_images();
273

274
275
    let model_type_obj = model_type.inner;

276
    let inner_path = model_path.to_string();
277
    let model_name = model_name.map(|n| n.to_string());
278
279
280
    let router_mode = router_mode.unwrap_or(RouterMode::RoundRobin);
    let router_config = RouterConfig::new(router_mode.into(), KvRouterConfig::default());

281
282
283
284
285
286
287
288
289
290
291
292
293
    // 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()?;

294
295
296
297
298
299
300
    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))
        })?;

301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
    // 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()));

320
    pyo3_async_runtimes::tokio::future_into_py(py, async move {
321
322
323
        // For TensorBased and Images models, skip HuggingFace downloads and register directly
        // These model types don't require tokenizers
        if is_tensor_based || is_images {
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
            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)
349
350
        let model_path = if fs::exists(&source_path)? {
            PathBuf::from(&source_path)
351
        } else {
352
            LocalModel::fetch(&source_path, false)
353
354
355
356
                .await
                .map_err(to_pyerr)?
        };

357
358
        let mut builder = dynamo_llm::local_model::LocalModelBuilder::default();
        builder
359
            // model path is the physical path on disk of the downloaded model
360
            .model_path(model_path)
361
362
363
364
            // 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
365
            .model_name(model_name.clone())
366
            .context_length(context_length)
367
            .kv_cache_block_size(kv_cache_block_size)
368
            .router_config(Some(router_config))
369
            .migration_limit(Some(migration_limit))
370
            .runtime_config(runtime_config.unwrap_or_default().inner)
371
            .user_data(user_data_json)
372
373
374
            .custom_template_path(custom_template_path_owned)
            .media_decoder(media_decoder.map(|m| m.inner))
            .media_fetcher(media_fetcher.map(|m| m.inner));
375

376
        let mut local_model = builder.build().await.map_err(to_pyerr)?;
377
        local_model
378
379
380
381
382
383
            .attach(
                &endpoint.inner,
                model_type_obj,
                model_input,
                lora_identifier.as_deref(),
            )
384
385
386
            .await
            .map_err(to_pyerr)?;

387
388
389
        if let Some(lora_name) = lora_identifier {
            tracing::info!("Registered LoRA '{}' MDC", lora_name);
        } else {
390
391
392
393
            tracing::info!(
                "Registered base model '{}' MDC",
                model_name.unwrap_or(source_path)
            );
394
395
        }

396
397
398
399
        Ok(())
    })
}

400
401
402
403
404
405
406
407
408
409
410
411
412
/// 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}`
413
#[pyfunction]
414
415
416
417
418
419
420
421
#[pyo3(signature = (endpoint, lora_name=None))]
fn unregister_llm<'p>(
    py: Python<'p>,
    endpoint: Endpoint,
    lora_name: Option<&str>,
) -> PyResult<Bound<'p, PyAny>> {
    let lora_name_owned = lora_name.map(|s| s.to_string());

422
    pyo3_async_runtimes::tokio::future_into_py(py, async move {
423
424
        // Unified detach method handles both base models and LoRA adapters
        LocalModel::detach_from_endpoint(&endpoint.inner, lora_name_owned.as_deref())
425
426
427
428
429
430
            .await
            .map_err(to_pyerr)?;
        Ok(())
    })
}

431
432
433
/// Download a model from Hugging Face, returning it's local path
/// Example: `model_path = await fetch_llm("Qwen/Qwen3-0.6B")`
#[pyfunction]
434
435
436
437
438
439
#[pyo3(signature = (remote_name, ignore_weights=false))]
fn fetch_llm<'p>(
    py: Python<'p>,
    remote_name: &str,
    ignore_weights: bool,
) -> PyResult<Bound<'p, PyAny>> {
440
441
    let repo = remote_name.to_string();
    pyo3_async_runtimes::tokio::future_into_py(py, async move {
442
443
444
        LocalModel::fetch(&repo, ignore_weights)
            .await
            .map_err(to_pyerr)
445
446
447
    })
}

448
449
#[pyclass]
#[derive(Clone)]
Ryan Olson's avatar
Ryan Olson committed
450
pub struct DistributedRuntime {
451
452
453
454
    inner: rs::DistributedRuntime,
    event_loop: PyObject,
}

Ryan Olson's avatar
Ryan Olson committed
455
456
impl DistributedRuntime {
    #[allow(dead_code)]
457
    pub(crate) fn inner(&self) -> &rs::DistributedRuntime {
Ryan Olson's avatar
Ryan Olson committed
458
459
460
461
        &self.inner
    }
}

462
#[pyclass]
463
#[derive(Clone)]
464
465
466
467
468
struct CancellationToken {
    inner: rs::CancellationToken,
}

#[pyclass]
469
#[derive(Clone)]
470
471
472
473
474
475
struct Namespace {
    inner: rs::component::Namespace,
    event_loop: PyObject,
}

#[pyclass]
476
#[derive(Clone)]
477
478
479
480
481
482
struct Component {
    inner: rs::component::Component,
    event_loop: PyObject,
}

#[pyclass]
483
#[derive(Clone)]
484
485
486
487
488
489
struct Endpoint {
    inner: rs::component::Endpoint,
    event_loop: PyObject,
}

#[pyclass]
490
#[derive(Clone)]
491
struct Client {
492
    router: rs::pipeline::PushRouter<serde_json::Value, RsAnnotated<serde_json::Value>>,
493
494
}

495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
#[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,
    };
516
517
518
519
    #[classattr]
    const TensorBased: Self = ModelType {
        inner: llm_rs::model_type::ModelType::TensorBased,
    };
520
521
522
523
    #[classattr]
    const Prefill: Self = ModelType {
        inner: llm_rs::model_type::ModelType::Prefill,
    };
524
525
526
527
    #[classattr]
    const Images: Self = ModelType {
        inner: llm_rs::model_type::ModelType::Images,
    };
528
529
530
531
532
533
534
535
536
537
538
539

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

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

540
541
#[pyclass(eq, eq_int)]
#[derive(Clone, PartialEq)]
542
543
544
enum ModelInput {
    Text = 1,
    Tokens = 2,
545
    Tensor = 3,
546
547
}

548
549
550
#[pymethods]
impl DistributedRuntime {
    #[new]
551
552
553
554
555
556
557
    #[pyo3(signature = (event_loop, store_kv, request_plane, enable_nats=None))]
    fn new(
        event_loop: PyObject,
        store_kv: String,
        request_plane: String,
        enable_nats: Option<bool>,
    ) -> PyResult<Self> {
558
        let selected_kv_store: kv::Selector = store_kv.parse().map_err(to_pyerr)?;
559
        let request_plane: RequestPlaneMode = request_plane.parse().map_err(to_pyerr)?;
560

561
562
563
        // 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()
564
            .or_else(|_| -> anyhow::Result<rs::Runtime> {
565
566
567
568
                // No existing Worker, create new one
                let worker = rs::Worker::from_settings()?;

                // Initialize pyo3 bridge (only happens once per process)
569
                INIT.get_or_try_init(|| -> anyhow::Result<()> {
570
571
                    let primary = worker.tokio_runtime()?;
                    pyo3_async_runtimes::tokio::init_with_runtime(primary).map_err(|e| {
572
                        anyhow::anyhow!("failed to initialize pyo3 static runtime: {:?}", e)
573
                    })?;
574
                    Ok(())
575
576
                })?;

577
                Ok(worker.runtime().clone())
578
579
            })
            .map_err(to_pyerr)?;
580

581
582
        // Initialize logging in context where tokio runtime is available
        // otel exporter requires it
583
        if config::env_is_truthy(env_otlp::OTEL_EXPORT_ENABLED) {
584
585
586
587
            runtime.secondary().block_on(async {
                rs::logging::init();
            });
        }
588

589
590
591
592
593
594
595
596
597
598
        // 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

599
600
        let runtime_config = DistributedConfig {
            store_backend: selected_kv_store,
601
            nats_config: if request_plane.is_nats() || enable_nats {
602
603
604
605
                Some(dynamo_runtime::transports::nats::ClientOptions::default())
            } else {
                None
            },
606
            request_plane,
607
608
609
610
611
        };
        let inner = runtime
            .secondary()
            .block_on(rs::DistributedRuntime::new(runtime, runtime_config))
            .map_err(to_pyerr)?;
612
613
614
615

        Ok(DistributedRuntime { inner, event_loop })
    }

Ryan Olson's avatar
Ryan Olson committed
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
    #[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(),
        })
    }

631
632
633
634
635
636
637
638
    fn namespace(&self, name: String) -> PyResult<Namespace> {
        Ok(Namespace {
            inner: self.inner.namespace(name).map_err(to_pyerr)?,
            event_loop: self.event_loop.clone(),
        })
    }

    fn shutdown(&self) {
639
        self.inner.shutdown();
640
641
642
643
644
    }

    fn event_loop(&self) -> PyObject {
        self.event_loop.clone()
    }
645
646
647
648
649

    fn child_token(&self) -> CancellationToken {
        let inner = self.inner.runtime().child_token();
        CancellationToken { inner }
    }
Richard Huo's avatar
Richard Huo committed
650

651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
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
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
    /// 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(())
    }

Richard Huo's avatar
Richard Huo committed
729
730
731
732
733
734
735
736
737
738
739
740
741
742
    // 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))
    }
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
}

#[pymethods]
impl CancellationToken {
    fn cancel(&self) {
        self.inner.cancel();
    }

    fn cancelled<'p>(&self, py: Python<'p>) -> PyResult<Bound<'p, PyAny>> {
        let token = self.inner.clone();
        pyo3_async_runtimes::tokio::future_into_py(py, async move {
            token.cancelled().await;
            Ok(())
        })
    }
}

#[pymethods]
impl Component {
    fn endpoint(&self, name: String) -> PyResult<Endpoint> {
        let inner = self.inner.endpoint(name);
        Ok(Endpoint {
            inner,
            event_loop: self.event_loop.clone(),
        })
    }
}

#[pymethods]
impl Endpoint {
773
    #[pyo3(signature = (generator, graceful_shutdown = true, metrics_labels = None, health_check_payload = None))]
774
775
776
777
    fn serve_endpoint<'p>(
        &self,
        py: Python<'p>,
        generator: PyObject,
778
        graceful_shutdown: Option<bool>,
779
        metrics_labels: Option<Vec<(String, String)>>,
780
        health_check_payload: Option<&Bound<'p, PyDict>>,
781
782
783
784
785
    ) -> PyResult<Bound<'p, PyAny>> {
        let engine = Arc::new(engine::PythonAsyncEngine::new(
            generator,
            self.event_loop.clone(),
        )?);
786
        let ingress = JsonServerStreamingIngress::for_engine(engine.clone()).map_err(to_pyerr)?;
787
788
789
790
791
792
793
794
795
796
797
798
799

        // 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
800
801
802
803
804
805
        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)",
            ));
806
807
808
        }

        let mut builder = self
809
810
811
812
            .inner
            .endpoint_builder()
            .metrics_labels(metrics_labels)
            .handler(ingress);
813
814
815
816
817

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

818
819
820
        // Register the engine in the local endpoint registry for in-process calls
        builder = builder.register_local_engine(engine).map_err(to_pyerr)?;

821
        let graceful_shutdown = graceful_shutdown.unwrap_or(true);
822
        pyo3_async_runtimes::tokio::future_into_py(py, async move {
823
824
825
826
827
            builder
                .graceful_shutdown(graceful_shutdown)
                .start()
                .await
                .map_err(to_pyerr)?;
828
829
830
831
            Ok(())
        })
    }

832
833
834
    fn client<'p>(&self, py: Python<'p>) -> PyResult<Bound<'p, PyAny>> {
        let inner = self.inner.clone();
        pyo3_async_runtimes::tokio::future_into_py(py, async move {
835
            let client = inner.client().await.map_err(to_pyerr)?;
836
837
838
839
840
841
            let push_router = rs::pipeline::PushRouter::<
                serde_json::Value,
                RsAnnotated<serde_json::Value>,
            >::from_client(client, Default::default())
            .await
            .map_err(to_pyerr)?;
842
843
844
            Ok(Client {
                router: push_router,
            })
845
846
        })
    }
847

848
849
850
    // Opaque unique ID for this worker. May change over worker lifetime.
    fn connection_id(&self) -> u64 {
        self.inner.drt().connection_id()
851
    }
852
853
854

    /// Get a RuntimeMetrics helper for creating Prometheus metrics
    #[getter]
855
856
    fn metrics(&self) -> prometheus_metrics::RuntimeMetrics {
        prometheus_metrics::RuntimeMetrics::from_endpoint(self.inner.clone())
857
    }
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886

    /// 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(())
        })
    }
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
}

#[pymethods]
impl Namespace {
    fn component(&self, name: String) -> PyResult<Component> {
        let inner = self.inner.component(name).map_err(to_pyerr)?;
        Ok(Component {
            inner,
            event_loop: self.event_loop.clone(),
        })
    }
}

#[pymethods]
impl Client {
902
903
    /// Get list of current instances.
    /// Replaces endpoint_ids.
904
    fn instance_ids(&self) -> Vec<u64> {
905
        self.router.client.instance_ids()
906
907
    }

908
909
910
    /// Wait for an instance to be available for work.
    /// Replaces wait_for_endpoints.
    fn wait_for_instances<'p>(&self, py: Python<'p>) -> PyResult<Bound<'p, PyAny>> {
911
        let inner = self.router.client.clone();
912
        pyo3_async_runtimes::tokio::future_into_py(py, async move {
913
            inner
914
                .wait_for_instances()
915
                .await
916
                .map(|v| v.into_iter().map(|cei| cei.id()).collect::<Vec<u64>>())
917
                .map_err(to_pyerr)
918
919
920
921
        })
    }

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

    /// Send a request to the next endpoint in a round-robin fashion.
934
    #[pyo3(signature = (request, annotated=DEFAULT_ANNOTATED_SETTING, context=None))]
935
936
937
938
939
    fn round_robin<'p>(
        &self,
        py: Python<'p>,
        request: PyObject,
        annotated: Option<bool>,
940
        context: Option<context::Context>,
941
942
    ) -> PyResult<Bound<'p, PyAny>> {
        let request: serde_json::Value = pythonize::depythonize(&request.into_bound(py))?;
943
        let request_ctx = create_request_context(request, &context);
944
945
946
        let annotated = annotated.unwrap_or(false);

        let (tx, rx) = tokio::sync::mpsc::channel(32);
947
        let client = self.router.clone();
948
949

        pyo3_async_runtimes::tokio::future_into_py(py, async move {
950
951
            let stream = match context {
                Some(context) => {
952
953
                    // Always instrument with appropriate span (none if no trace context)
                    let span = get_span_for_context(&context, "round_robin");
954
955
956
957
958
                    client
                        .round_robin(request_ctx)
                        .instrument(span)
                        .await
                        .map_err(to_pyerr)?
959
                }
960
                _ => client.round_robin(request_ctx).await.map_err(to_pyerr)?,
961
            };
962
963
964
965
966
967
968
969
970
            tokio::spawn(process_stream(stream, tx));
            Ok(AsyncResponseStream {
                rx: Arc::new(Mutex::new(rx)),
                annotated,
            })
        })
    }

    /// Send a request to a random endpoint.
971
    #[pyo3(signature = (request, annotated=DEFAULT_ANNOTATED_SETTING, context=None))]
972
973
974
975
976
    fn random<'p>(
        &self,
        py: Python<'p>,
        request: PyObject,
        annotated: Option<bool>,
977
        context: Option<context::Context>,
978
979
    ) -> PyResult<Bound<'p, PyAny>> {
        let request: serde_json::Value = pythonize::depythonize(&request.into_bound(py))?;
980
        let request_ctx = create_request_context(request, &context);
981
982
983
        let annotated = annotated.unwrap_or(false);

        let (tx, rx) = tokio::sync::mpsc::channel(32);
984
        let client = self.router.clone();
985
986

        pyo3_async_runtimes::tokio::future_into_py(py, async move {
987
988
            let stream = match context {
                Some(context) => {
989
                    // Always instrument with appropriate span (none if no trace context)
990
                    let span = get_span_for_context(&context, "random");
991
992
993
994
995
                    client
                        .random(request_ctx)
                        .instrument(span)
                        .await
                        .map_err(to_pyerr)?
996
                }
997
                _ => client.random(request_ctx).await.map_err(to_pyerr)?,
998
            };
999
1000
1001
1002
1003
1004
1005
1006
1007
            tokio::spawn(process_stream(stream, tx));
            Ok(AsyncResponseStream {
                rx: Arc::new(Mutex::new(rx)),
                annotated,
            })
        })
    }

    /// Directly send a request to a specific endpoint.
1008
    #[pyo3(signature = (request, instance_id, annotated=DEFAULT_ANNOTATED_SETTING, context=None))]
1009
1010
1011
1012
    fn direct<'p>(
        &self,
        py: Python<'p>,
        request: PyObject,
1013
        instance_id: u64,
1014
        annotated: Option<bool>,
1015
        context: Option<context::Context>,
1016
1017
    ) -> PyResult<Bound<'p, PyAny>> {
        let request: serde_json::Value = pythonize::depythonize(&request.into_bound(py))?;
1018
        let request_ctx = create_request_context(request, &context);
1019
1020
1021
        let annotated = annotated.unwrap_or(false);

        let (tx, rx) = tokio::sync::mpsc::channel(32);
1022
        let client = self.router.clone();
1023
1024

        pyo3_async_runtimes::tokio::future_into_py(py, async move {
1025
1026
            let stream = match context {
                Some(context) => {
1027
                    // Always instrument with appropriate span (none if no trace context)
1028
1029
                    let span =
                        get_span_for_direct_context(&context, "direct", &instance_id.to_string());
1030
1031
1032
1033
1034
                    client
                        .direct(request_ctx, instance_id)
                        .instrument(span)
                        .await
                        .map_err(to_pyerr)?
1035
1036
                }
                _ => client
1037
                    .direct(request_ctx, instance_id)
1038
1039
1040
                    .await
                    .map_err(to_pyerr)?,
            };
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052

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

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

async fn process_stream(
1053
    stream: EngineStream<RsAnnotated<serde_json::Value>>,
1054
1055
1056
1057
1058
    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
1059
        let annotated: RsAnnotated<serde_json::Value> = response;
1060
        let annotated: RsAnnotated<PyObject> = annotated.map_data(|data| {
1061
            Python::with_gil(|py| match pythonize::pythonize(py, &data) {
1062
1063
                Ok(pyobj) => Ok(pyobj.into()),
                Err(e) => Err(e.to_string()),
1064
            })
1065
1066
1067
1068
1069
1070
        });

        let is_error = annotated.is_error();

        // Send the PyObject through the channel or log an error
        if let Err(e) = tx.send(annotated).await {
1071
            tracing::error!("Failed to send response: {:?}", e);
1072
            break;
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
        }

        if is_error {
            break;
        }
    }
}

#[pyclass]
struct AsyncResponseStream {
    rx: Arc<Mutex<tokio::sync::mpsc::Receiver<RsAnnotated<PyObject>>>>,
    annotated: bool,
}

#[pymethods]
impl AsyncResponseStream {
    /// This method is required to implement the `AsyncIterator` protocol.
    #[pyo3(name = "__aiter__")]
    fn aiter(slf: PyRef<Self>, py: Python) -> PyResult<Py<PyAny>> {
        slf.into_py_any(py)
    }
    /// This method is required to implement the `AsyncIterator` protocol.
    #[pyo3(name = "__anext__")]
    fn next<'p>(&self, py: Python<'p>) -> PyResult<Bound<'p, PyAny>> {
        let rx = self.rx.clone();
        let annotated = self.annotated;

        pyo3_async_runtimes::tokio::future_into_py(py, async move {
            loop {
                let value = rx.lock().await.recv().await;
                match value {
                    Some(pyobj) => {
                        let pyobj = match pyobj.ok() {
                            Ok(pyobj) => pyobj,
                            Err(e) => {
                                return Err(PyErr::new::<pyo3::exceptions::PyValueError, _>(e));
                            }
                        };

                        if annotated {
                            let object = Annotated { inner: pyobj };
                            #[allow(deprecated)]
                            let object = Python::with_gil(|py| object.into_py(py));
                            return Ok(object);
                        } else {
                            match pyobj.data {
                                Some(data) => return Ok(data),
                                None => continue,
                            }
                        }
                    }
                    None => return Err(PyStopAsyncIteration::new_err("Stream exhausted")),
                }
            }
        })
    }
}

#[pyclass]
struct Annotated {
    inner: RsAnnotated<PyObject>,
}

#[pymethods]
impl Annotated {
    #[new]
    fn new(data: PyObject) -> Self {
        Annotated {
            inner: RsAnnotated::from_data(data),
        }
    }

    fn is_error(&self) -> bool {
        self.inner.is_error()
    }

    fn data(&self) -> Option<PyObject> {
        self.inner.data.clone()
    }

    fn event(&self) -> Option<String> {
        self.inner.event.clone()
    }

    fn comments(&self) -> Option<Vec<String>> {
        self.inner.comment.clone()
    }

    fn id(&self) -> Option<String> {
        self.inner.id.clone()
    }

    #[pyo3(name = "__repr__")]
    fn _repr(&self, py: Python) -> String {
        let data = self.inner.data.clone().map(|obj| {
            obj.call_method0(py, "__repr__")
                .and_then(|repr_obj| repr_obj.extract::<Py<PyString>>(py))
                .map(|py_str| py_str.to_string_lossy(py).into_owned())
                .unwrap_or_else(|_| "<failed_repr>".to_string())
        });

        format!(
            "Annotated(data={}, event={}, comment={:?}, id={})",
            data.unwrap_or_else(|| "<no_data>".to_string()),
            self.inner.event.as_deref().unwrap_or("None"),
            self.inner.comment.as_deref().unwrap_or(&[]),
            self.inner.id.as_deref().unwrap_or("None")
        )
    }
}