lib.rs 42.6 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
49
#[pyclass(eq, eq_int)]
#[derive(Clone, Debug, PartialEq)]
pub enum RouterMode {
    RoundRobin,
    Random,
    KV,
50
51
52
    /// Direct routing - reads worker ID from each request's routing hints.
    /// Used when an external orchestrator (e.g., EPP) handles worker selection.
    Direct,
53
54
55
56
57
58
59
60
}

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

66
mod context;
67
mod engine;
68
mod http;
69
mod kserve_grpc;
70
mod llm;
71
mod parsers;
72
mod planner;
73
mod prometheus_metrics;
74
75
76
77
78
79
80

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);
81
82
const SKIP_PYTHON_LOG_INIT_ENV: &str = "DYNAMO_SKIP_PYTHON_LOG_INIT";

83
84
// Helper to get appropriate span for instrumentation - always emit spans
fn get_span_for_context(context: &context::Context, operation: &str) -> tracing::Span {
85
86
87
88
89
90
    logging::make_client_request_span(
        operation,
        context.inner().id(),
        context.trace_context(),
        None,
    )
91
92
93
94
95
96
97
98
}

// 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 {
99
100
101
102
103
104
    logging::make_client_request_span(
        operation,
        context.inner().id(),
        context.trace_context(),
        Some(instance_id),
    )
105
106
}

107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
// 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(),
    }
}

130
131
132
133
134
/// 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<()> {
135
    // Initialize logging early unless OTEL export is enabled (which requires tokio runtime)
136
    if config::env_is_truthy(env_otlp::OTEL_EXPORT_ENABLED) {
137
        eprintln!(
138
            "Warning: OTEL_EXPORT_ENABLED detected. Logging initialization deferred until runtime is available. Early logs may be dropped."
139
        );
140
    } else if std::env::var_os(SKIP_PYTHON_LOG_INIT_ENV).is_none() {
141
142
143
        rs::logging::init();
    }

Yan Ru Pei's avatar
Yan Ru Pei committed
144
    m.add_function(wrap_pyfunction!(llm::kv::compute_block_hash_for_seq_py, m)?)?;
145
    m.add_function(wrap_pyfunction!(lora_name_to_id, m)?)?;
Ryan Olson's avatar
Ryan Olson committed
146
    m.add_function(wrap_pyfunction!(log_message, m)?)?;
147
148
149
    m.add_function(wrap_pyfunction!(register_model, m)?)?;
    m.add_function(wrap_pyfunction!(unregister_model, m)?)?;
    m.add_function(wrap_pyfunction!(fetch_model, m)?)?;
150
    m.add_function(wrap_pyfunction!(run_kv_indexer, m)?)?;
151
152
    m.add_function(wrap_pyfunction!(llm::entrypoint::make_engine, m)?)?;
    m.add_function(wrap_pyfunction!(llm::entrypoint::run_input, m)?)?;
153

154
155
    m.add_class::<DistributedRuntime>()?;
    m.add_class::<Endpoint>()?;
156
    m.add_class::<ModelCardInstanceId>()?;
157
158
    m.add_class::<Client>()?;
    m.add_class::<AsyncResponseStream>()?;
159
160
161
    m.add_class::<llm::entrypoint::EntrypointArgs>()?;
    m.add_class::<llm::entrypoint::EngineConfig>()?;
    m.add_class::<llm::entrypoint::EngineType>()?;
162
163
    m.add_class::<llm::entrypoint::RouterConfig>()?;
    m.add_class::<llm::entrypoint::KvRouterConfig>()?;
164
    m.add_class::<llm::kv::WorkerMetricsPublisher>()?;
165
    m.add_class::<llm::model_card::ModelDeploymentCard>()?; // Internal: only in _internal, not public API
166
    m.add_class::<llm::local_model::ModelRuntimeConfig>()?;
167
168
    m.add_class::<llm::preprocessor::MediaDecoder>()?;
    m.add_class::<llm::preprocessor::MediaFetcher>()?;
169
    m.add_class::<llm::kv::OverlapScores>()?;
170
    m.add_class::<llm::kv::KvEventPublisher>()?;
Yan Ru Pei's avatar
Yan Ru Pei committed
171
    m.add_class::<llm::kv::RadixTree>()?;
172
173
    m.add_class::<llm::fpm::FpmEventRelay>()?;
    m.add_class::<llm::fpm::FpmEventSubscriber>()?;
174
    m.add_class::<llm::lora::LoRADownloader>()?;
175
176
    m.add_class::<http::HttpService>()?;
    m.add_class::<http::HttpAsyncEngine>()?;
177
    m.add_class::<context::Context>()?;
178
    m.add_class::<ModelType>()?;
179
    m.add_class::<ModelInput>()?;
180
    m.add_class::<llm::kv::KvRouter>()?;
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))
}

206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
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
224
225
226
227
228
229
230
/// 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);
}

231
232
233
234
235
236
237
/// 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)
}

238
239
/// 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.
240
241
242
243
244
245
246
///
/// 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.
247
#[pyfunction]
248
#[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))]
249
#[allow(clippy::too_many_arguments)]
250
fn register_model<'p>(
251
    py: Python<'p>,
252
    model_input: ModelInput,
253
    model_type: ModelType,
254
255
256
    endpoint: Endpoint,
    model_path: &str,
    model_name: Option<&str>,
257
258
    context_length: Option<u32>,
    kv_cache_block_size: Option<u32>,
259
    router_mode: Option<RouterMode>,
260
    runtime_config: Option<ModelRuntimeConfig>,
261
    user_data: Option<&Bound<'p, PyDict>>,
262
    custom_template_path: Option<&str>,
263
264
    media_decoder: Option<MediaDecoder>,
    media_fetcher: Option<MediaFetcher>,
265
266
    lora_name: Option<&str>,
    base_model_path: Option<&str>,
267
) -> PyResult<Bound<'p, PyAny>> {
268
    // Validate Prefill model type requirements
269
270
271
272
273
274
    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",
        ));
275
276
    }

277
278
279
    let model_input = match model_input {
        ModelInput::Text => llm_rs::model_type::ModelInput::Text,
        ModelInput::Tokens => llm_rs::model_type::ModelInput::Tokens,
280
        ModelInput::Tensor => llm_rs::model_type::ModelInput::Tensor,
281
282
    };

283
    let is_tensor_based = model_type.inner.supports_tensor();
284
    let is_images = model_type.inner.supports_images();
285
    let is_videos = model_type.inner.supports_videos();
286

287
288
    let model_type_obj = model_type.inner;

289
    let inner_path = model_path.to_string();
290
    let model_name = model_name.map(|n| n.to_string());
291
292
293
    let router_mode = router_mode.unwrap_or(RouterMode::RoundRobin);
    let router_config = RouterConfig::new(router_mode.into(), KvRouterConfig::default());

294
295
296
297
298
299
300
301
302
303
304
305
306
    // 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()?;

307
308
309
310
311
312
313
    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))
        })?;

314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
    // 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()));

333
    pyo3_async_runtimes::tokio::future_into_py(py, async move {
334
335
336
        // 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 {
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
            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)
362
363
        let model_path = if fs::exists(&source_path)? {
            PathBuf::from(&source_path)
364
        } else {
365
            LocalModel::fetch(&source_path, false)
366
367
368
369
                .await
                .map_err(to_pyerr)?
        };

370
371
        let mut builder = dynamo_llm::local_model::LocalModelBuilder::default();
        builder
372
            // model path is the physical path on disk of the downloaded model
373
            .model_path(model_path)
374
375
376
377
            // 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
378
            .model_name(model_name.clone())
379
            .context_length(context_length)
380
            .kv_cache_block_size(kv_cache_block_size)
381
            .router_config(Some(router_config))
382
            .runtime_config(runtime_config.unwrap_or_default().inner)
383
            .user_data(user_data_json)
384
385
386
            .custom_template_path(custom_template_path_owned)
            .media_decoder(media_decoder.map(|m| m.inner))
            .media_fetcher(media_fetcher.map(|m| m.inner));
387

388
        let mut local_model = builder.build().await.map_err(to_pyerr)?;
389
390
391
392
393
394
395
396
397

        // 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,
            });

398
        local_model
399
            .attach(&endpoint.inner, model_type_obj, model_input, lora_info)
400
401
402
            .await
            .map_err(to_pyerr)?;

403
404
405
        if let Some(lora_name) = lora_identifier {
            tracing::info!("Registered LoRA '{}' MDC", lora_name);
        } else {
406
407
408
409
            tracing::info!(
                "Registered base model '{}' MDC",
                model_name.unwrap_or(source_path)
            );
410
411
        }

412
413
414
415
        Ok(())
    })
}

416
417
418
419
420
421
422
423
424
425
426
427
428
/// 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}`
429
#[pyfunction]
430
#[pyo3(signature = (endpoint, lora_name=None))]
431
fn unregister_model<'p>(
432
433
434
435
436
437
    py: Python<'p>,
    endpoint: Endpoint,
    lora_name: Option<&str>,
) -> PyResult<Bound<'p, PyAny>> {
    let lora_name_owned = lora_name.map(|s| s.to_string());

438
    pyo3_async_runtimes::tokio::future_into_py(py, async move {
439
440
        // Unified detach method handles both base models and LoRA adapters
        LocalModel::detach_from_endpoint(&endpoint.inner, lora_name_owned.as_deref())
441
442
443
444
445
446
            .await
            .map_err(to_pyerr)?;
        Ok(())
    })
}

447
448
/// Download a model from Hugging Face, returning its local path
/// Example: `model_path = await fetch_model("Qwen/Qwen3-0.6B")`
449
#[pyfunction]
450
#[pyo3(signature = (remote_name, ignore_weights=false))]
451
fn fetch_model<'p>(
452
453
454
455
    py: Python<'p>,
    remote_name: &str,
    ignore_weights: bool,
) -> PyResult<Bound<'p, PyAny>> {
456
457
    let repo = remote_name.to_string();
    pyo3_async_runtimes::tokio::future_into_py(py, async move {
458
459
460
        LocalModel::fetch(&repo, ignore_weights)
            .await
            .map_err(to_pyerr)
461
462
463
    })
}

464
465
#[pyclass]
#[derive(Clone)]
Ryan Olson's avatar
Ryan Olson committed
466
pub struct DistributedRuntime {
467
468
469
470
    inner: rs::DistributedRuntime,
    event_loop: PyObject,
}

Ryan Olson's avatar
Ryan Olson committed
471
472
impl DistributedRuntime {
    #[allow(dead_code)]
473
    pub(crate) fn inner(&self) -> &rs::DistributedRuntime {
Ryan Olson's avatar
Ryan Olson committed
474
475
476
477
        &self.inner
    }
}

478
#[pyclass]
479
#[derive(Clone)]
480
481
482
483
484
struct CancellationToken {
    inner: rs::CancellationToken,
}

#[pyclass]
485
#[derive(Clone)]
486
487
488
489
490
struct Endpoint {
    inner: rs::component::Endpoint,
    event_loop: PyObject,
}

491
492
493
494
495
496
#[pyclass]
#[derive(Clone)]
struct ModelCardInstanceId {
    inner: rs::discovery::ModelCardInstanceId,
}

497
#[pyclass]
498
#[derive(Clone)]
499
struct Client {
500
    router: rs::pipeline::PushRouter<serde_json::Value, RsAnnotated<serde_json::Value>>,
501
502
}

503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
#[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,
    };
524
525
526
527
    #[classattr]
    const TensorBased: Self = ModelType {
        inner: llm_rs::model_type::ModelType::TensorBased,
    };
528
529
530
531
    #[classattr]
    const Prefill: Self = ModelType {
        inner: llm_rs::model_type::ModelType::Prefill,
    };
532
533
534
535
    #[classattr]
    const Images: Self = ModelType {
        inner: llm_rs::model_type::ModelType::Images,
    };
536
    #[classattr]
537
538
539
540
    const Audios: Self = ModelType {
        inner: llm_rs::model_type::ModelType::Audios,
    };
    #[classattr]
541
542
543
    const Videos: Self = ModelType {
        inner: llm_rs::model_type::ModelType::Videos,
    };
544

545
546
547
548
    fn supports_chat(&self) -> bool {
        self.inner.supports_chat()
    }

549
550
551
552
553
554
555
556
557
558
559
    fn __or__(&self, other: &Self) -> Self {
        ModelType {
            inner: self.inner | other.inner,
        }
    }

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

560
561
#[pyclass(eq, eq_int)]
#[derive(Clone, PartialEq)]
562
563
564
enum ModelInput {
    Text = 1,
    Tokens = 2,
565
    Tensor = 3,
566
567
}

568
569
570
#[pymethods]
impl DistributedRuntime {
    #[new]
571
    #[pyo3(signature = (event_loop, discovery_backend, request_plane, enable_nats=None))]
572
573
    fn new(
        event_loop: PyObject,
574
        discovery_backend: String,
575
576
577
        request_plane: String,
        enable_nats: Option<bool>,
    ) -> PyResult<Self> {
578
579
580
581
582
583
584
        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)
            }
        };
585
        let request_plane: RequestPlaneMode = request_plane.parse().map_err(to_pyerr)?;
586

587
588
589
        // 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()
590
            .or_else(|_| -> anyhow::Result<rs::Runtime> {
591
592
593
594
                // No existing Worker, create new one
                let worker = rs::Worker::from_settings()?;

                // Initialize pyo3 bridge (only happens once per process)
595
                INIT.get_or_try_init(|| -> anyhow::Result<()> {
596
597
                    let primary = worker.tokio_runtime()?;
                    pyo3_async_runtimes::tokio::init_with_runtime(primary).map_err(|e| {
598
                        anyhow::anyhow!("failed to initialize pyo3 static runtime: {:?}", e)
599
                    })?;
600
                    Ok(())
601
602
                })?;

603
                Ok(worker.runtime().clone())
604
605
            })
            .map_err(to_pyerr)?;
606

607
608
        // Initialize logging in context where tokio runtime is available
        // otel exporter requires it
609
        if config::env_is_truthy(env_otlp::OTEL_EXPORT_ENABLED) {
610
611
612
613
            runtime.secondary().block_on(async {
                rs::logging::init();
            });
        }
614

615
616
617
618
619
620
621
622
623
624
        // 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

625
        let runtime_config = DistributedConfig {
626
            discovery_backend: discovery_backend_config,
627
            nats_config: if request_plane.is_nats() || enable_nats {
628
629
630
631
                Some(dynamo_runtime::transports::nats::ClientOptions::default())
            } else {
                None
            },
632
            request_plane,
633
634
635
636
637
        };
        let inner = runtime
            .secondary()
            .block_on(rs::DistributedRuntime::new(runtime, runtime_config))
            .map_err(to_pyerr)?;
638
639
640
641

        Ok(DistributedRuntime { inner, event_loop })
    }

Ryan Olson's avatar
Ryan Olson committed
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
    #[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(),
        })
    }

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
    /// 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,
685
686
687
688
689
            event_loop: self.event_loop.clone(),
        })
    }

    fn shutdown(&self) {
690
        self.inner.shutdown();
691
692
693
694
695
    }

    fn event_loop(&self) -> PyObject {
        self.event_loop.clone()
    }
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
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
    /// 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
775
776
777
778
779
780
781
782
783
784
785
786
787
788
    // 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))
    }
789
790
791
792
}

#[pymethods]
impl Endpoint {
793
    #[pyo3(signature = (generator, graceful_shutdown = true, metrics_labels = None, health_check_payload = None))]
794
795
796
797
    fn serve_endpoint<'p>(
        &self,
        py: Python<'p>,
        generator: PyObject,
798
        graceful_shutdown: Option<bool>,
799
        metrics_labels: Option<Vec<(String, String)>>,
800
        health_check_payload: Option<&Bound<'p, PyDict>>,
801
802
803
804
805
    ) -> PyResult<Bound<'p, PyAny>> {
        let engine = Arc::new(engine::PythonAsyncEngine::new(
            generator,
            self.event_loop.clone(),
        )?);
806
        let ingress = JsonServerStreamingIngress::for_engine(engine.clone()).map_err(to_pyerr)?;
807
808
809
810
811
812
813
814
815
816
817
818
819

        // 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
820
821
822
823
824
825
        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)",
            ));
826
827
828
        }

        let mut builder = self
829
830
831
832
            .inner
            .endpoint_builder()
            .metrics_labels(metrics_labels)
            .handler(ingress);
833
834
835
836
837

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

838
839
840
        // Register the engine in the local endpoint registry for in-process calls
        builder = builder.register_local_engine(engine).map_err(to_pyerr)?;

841
        let graceful_shutdown = graceful_shutdown.unwrap_or(true);
842
        pyo3_async_runtimes::tokio::future_into_py(py, async move {
843
844
845
846
847
            builder
                .graceful_shutdown(graceful_shutdown)
                .start()
                .await
                .map_err(to_pyerr)?;
848
849
850
851
            Ok(())
        })
    }

852
853
854
855
856
857
858
    #[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);
859
860
        let inner = self.inner.clone();
        pyo3_async_runtimes::tokio::future_into_py(py, async move {
861
            let client = inner.client().await.map_err(to_pyerr)?;
862
863
864
            let push_router = rs::pipeline::PushRouter::<
                serde_json::Value,
                RsAnnotated<serde_json::Value>,
865
            >::from_client(client, router_mode.into())
866
867
            .await
            .map_err(to_pyerr)?;
868
869
870
            Ok(Client {
                router: push_router,
            })
871
872
        })
    }
873

874
875
876
    // Opaque unique ID for this worker. May change over worker lifetime.
    fn connection_id(&self) -> u64 {
        self.inner.drt().connection_id()
877
    }
878
879
880

    /// Get a RuntimeMetrics helper for creating Prometheus metrics
    #[getter]
881
882
    fn metrics(&self) -> prometheus_metrics::RuntimeMetrics {
        prometheus_metrics::RuntimeMetrics::from_endpoint(self.inner.clone())
883
    }
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912

    /// 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(())
        })
    }
913
914
}

915
916
917
918
919
920
921
922
923
924
925
926
927
#[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(),
        )
    }
}

928
929
#[pymethods]
impl Client {
930
931
    /// Get list of current instances.
    /// Replaces endpoint_ids.
932
    fn instance_ids(&self) -> Vec<u64> {
933
        self.router.client.instance_ids()
934
935
    }

936
937
938
    /// 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>> {
939
        let inner = self.router.client.clone();
940
        pyo3_async_runtimes::tokio::future_into_py(py, async move {
941
            inner
942
                .wait_for_instances()
943
                .await
944
                .map(|v| v.into_iter().map(|cei| cei.id()).collect::<Vec<u64>>())
945
                .map_err(to_pyerr)
946
947
948
949
        })
    }

    /// Issue a request to the endpoint using the default routing strategy.
950
    #[pyo3(signature = (request, annotated=DEFAULT_ANNOTATED_SETTING, context=None))]
951
952
953
954
955
    fn generate<'p>(
        &self,
        py: Python<'p>,
        request: PyObject,
        annotated: Option<bool>,
956
        context: Option<context::Context>,
957
    ) -> PyResult<Bound<'p, PyAny>> {
958
        self.random(py, request, annotated, context)
959
960
961
    }

    /// Send a request to the next endpoint in a round-robin fashion.
962
    #[pyo3(signature = (request, annotated=DEFAULT_ANNOTATED_SETTING, context=None))]
963
964
965
966
967
    fn round_robin<'p>(
        &self,
        py: Python<'p>,
        request: PyObject,
        annotated: Option<bool>,
968
        context: Option<context::Context>,
969
970
    ) -> PyResult<Bound<'p, PyAny>> {
        let request: serde_json::Value = pythonize::depythonize(&request.into_bound(py))?;
971
        let request_ctx = create_request_context(request, &context);
972
973
974
        let annotated = annotated.unwrap_or(false);

        let (tx, rx) = tokio::sync::mpsc::channel(32);
975
        let client = self.router.clone();
976
977

        pyo3_async_runtimes::tokio::future_into_py(py, async move {
978
979
            let stream = match context {
                Some(context) => {
980
981
                    // Always instrument with appropriate span (none if no trace context)
                    let span = get_span_for_context(&context, "round_robin");
982
983
984
985
986
                    client
                        .round_robin(request_ctx)
                        .instrument(span)
                        .await
                        .map_err(to_pyerr)?
987
                }
988
                _ => client.round_robin(request_ctx).await.map_err(to_pyerr)?,
989
            };
990
            tokio::spawn(process_stream(stream, tx));
991
            Ok(AsyncResponseStream::new(rx, annotated))
992
993
994
995
        })
    }

    /// Send a request to a random endpoint.
996
    #[pyo3(signature = (request, annotated=DEFAULT_ANNOTATED_SETTING, context=None))]
997
998
999
1000
1001
    fn random<'p>(
        &self,
        py: Python<'p>,
        request: PyObject,
        annotated: Option<bool>,
1002
        context: Option<context::Context>,
1003
1004
    ) -> PyResult<Bound<'p, PyAny>> {
        let request: serde_json::Value = pythonize::depythonize(&request.into_bound(py))?;
1005
        let request_ctx = create_request_context(request, &context);
1006
1007
1008
        let annotated = annotated.unwrap_or(false);

        let (tx, rx) = tokio::sync::mpsc::channel(32);
1009
        let client = self.router.clone();
1010
1011

        pyo3_async_runtimes::tokio::future_into_py(py, async move {
1012
1013
            let stream = match context {
                Some(context) => {
1014
                    // Always instrument with appropriate span (none if no trace context)
1015
                    let span = get_span_for_context(&context, "random");
1016
1017
1018
1019
1020
                    client
                        .random(request_ctx)
                        .instrument(span)
                        .await
                        .map_err(to_pyerr)?
1021
                }
1022
                _ => client.random(request_ctx).await.map_err(to_pyerr)?,
1023
            };
1024
            tokio::spawn(process_stream(stream, tx));
1025
            Ok(AsyncResponseStream::new(rx, annotated))
1026
1027
1028
1029
        })
    }

    /// Directly send a request to a specific endpoint.
1030
    #[pyo3(signature = (request, instance_id, annotated=DEFAULT_ANNOTATED_SETTING, context=None))]
1031
1032
1033
1034
    fn direct<'p>(
        &self,
        py: Python<'p>,
        request: PyObject,
1035
        instance_id: u64,
1036
        annotated: Option<bool>,
1037
        context: Option<context::Context>,
1038
1039
    ) -> PyResult<Bound<'p, PyAny>> {
        let request: serde_json::Value = pythonize::depythonize(&request.into_bound(py))?;
1040
        let request_ctx = create_request_context(request, &context);
1041
1042
1043
        let annotated = annotated.unwrap_or(false);

        let (tx, rx) = tokio::sync::mpsc::channel(32);
1044
        let client = self.router.clone();
1045
1046

        pyo3_async_runtimes::tokio::future_into_py(py, async move {
1047
1048
            let stream = match context {
                Some(context) => {
1049
                    // Always instrument with appropriate span (none if no trace context)
1050
1051
                    let span =
                        get_span_for_direct_context(&context, "direct", &instance_id.to_string());
1052
1053
1054
1055
1056
                    client
                        .direct(request_ctx, instance_id)
                        .instrument(span)
                        .await
                        .map_err(to_pyerr)?
1057
1058
                }
                _ => client
1059
                    .direct(request_ctx, instance_id)
1060
1061
1062
                    .await
                    .map_err(to_pyerr)?,
            };
1063
1064
1065

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

1066
            Ok(AsyncResponseStream::new(rx, annotated))
1067
1068
1069
1070
1071
        })
    }
}

async fn process_stream(
1072
    stream: EngineStream<RsAnnotated<serde_json::Value>>,
1073
1074
1075
1076
1077
    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
1078
        let annotated: RsAnnotated<serde_json::Value> = response;
1079
        let annotated: RsAnnotated<PyObject> = annotated.map_data(|data| {
1080
            Python::with_gil(|py| match pythonize::pythonize(py, &data) {
1081
1082
                Ok(pyobj) => Ok(pyobj.into()),
                Err(e) => Err(e.to_string()),
1083
            })
1084
1085
1086
1087
1088
1089
        });

        let is_error = annotated.is_error();

        // Send the PyObject through the channel or log an error
        if let Err(e) = tx.send(annotated).await {
1090
            tracing::error!("Failed to send response: {:?}", e);
1091
            break;
1092
1093
1094
1095
1096
1097
1098
1099
1100
        }

        if is_error {
            break;
        }
    }
}

#[pyclass]
1101
pub(crate) struct AsyncResponseStream {
1102
1103
1104
    rx: Arc<Mutex<tokio::sync::mpsc::Receiver<RsAnnotated<PyObject>>>>,
    annotated: bool,
}
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116

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

#[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")
        )
    }
}