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
};

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

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
            RouterMode::Direct => Self::Direct,
61
62
63
        }
    }
}
64

65
mod context;
66
mod engine;
67
mod http;
68
mod kserve_grpc;
69
mod llm;
70
mod parsers;
71
mod planner;
72
mod prometheus_metrics;
73
74
75
76
77
78
79

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);
80
81
const SKIP_PYTHON_LOG_INIT_ENV: &str = "DYNAMO_SKIP_PYTHON_LOG_INIT";

82
83
// Helper to get appropriate span for instrumentation - always emit spans
fn get_span_for_context(context: &context::Context, operation: &str) -> tracing::Span {
84
85
86
87
88
89
    logging::make_client_request_span(
        operation,
        context.inner().id(),
        context.trace_context(),
        None,
    )
90
91
92
93
94
95
96
97
}

// 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 {
98
99
100
101
102
103
    logging::make_client_request_span(
        operation,
        context.inner().id(),
        context.trace_context(),
        Some(instance_id),
    )
104
105
}

106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
// 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(),
    }
}

129
130
131
132
133
/// 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<()> {
134
    // Initialize logging early unless OTEL export is enabled (which requires tokio runtime)
135
    if config::env_is_truthy(env_otlp::OTEL_EXPORT_ENABLED) {
136
        eprintln!(
137
            "Warning: OTEL_EXPORT_ENABLED detected. Logging initialization deferred until runtime is available. Early logs may be dropped."
138
        );
139
    } else if std::env::var_os(SKIP_PYTHON_LOG_INIT_ENV).is_none() {
140
141
142
        rs::logging::init();
    }

Yan Ru Pei's avatar
Yan Ru Pei committed
143
    m.add_function(wrap_pyfunction!(llm::kv::compute_block_hash_for_seq_py, m)?)?;
144
    m.add_function(wrap_pyfunction!(lora_name_to_id, m)?)?;
Ryan Olson's avatar
Ryan Olson committed
145
    m.add_function(wrap_pyfunction!(log_message, m)?)?;
146
147
148
    m.add_function(wrap_pyfunction!(register_model, m)?)?;
    m.add_function(wrap_pyfunction!(unregister_model, m)?)?;
    m.add_function(wrap_pyfunction!(fetch_model, m)?)?;
149
    m.add_function(wrap_pyfunction!(run_kv_indexer, m)?)?;
150
151
    m.add_function(wrap_pyfunction!(llm::entrypoint::make_engine, m)?)?;
    m.add_function(wrap_pyfunction!(llm::entrypoint::run_input, m)?)?;
152

153
154
    m.add_class::<DistributedRuntime>()?;
    m.add_class::<Endpoint>()?;
155
    m.add_class::<ModelCardInstanceId>()?;
156
157
    m.add_class::<Client>()?;
    m.add_class::<AsyncResponseStream>()?;
158
159
160
    m.add_class::<llm::entrypoint::EntrypointArgs>()?;
    m.add_class::<llm::entrypoint::EngineConfig>()?;
    m.add_class::<llm::entrypoint::EngineType>()?;
161
162
    m.add_class::<llm::entrypoint::RouterConfig>()?;
    m.add_class::<llm::entrypoint::KvRouterConfig>()?;
163
    m.add_class::<llm::kv::WorkerMetricsPublisher>()?;
164
    m.add_class::<llm::model_card::ModelDeploymentCard>()?; // Internal: only in _internal, not public API
165
    m.add_class::<llm::local_model::ModelRuntimeConfig>()?;
166
167
    m.add_class::<llm::preprocessor::MediaDecoder>()?;
    m.add_class::<llm::preprocessor::MediaFetcher>()?;
168
    m.add_class::<llm::kv::OverlapScores>()?;
169
    m.add_class::<llm::kv::KvEventPublisher>()?;
Yan Ru Pei's avatar
Yan Ru Pei committed
170
    m.add_class::<llm::kv::RadixTree>()?;
171
172
    m.add_class::<llm::fpm::FpmEventRelay>()?;
    m.add_class::<llm::fpm::FpmEventSubscriber>()?;
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
    m.add_class::<llm::kv::KvRouter>()?;
180
    m.add_class::<RouterMode>()?;
181
    m.add_class::<kserve_grpc::KserveGrpcService>()?;
182
    m.add("__version__", env!("CARGO_PKG_VERSION"))?;
183
184
185
    m.add_class::<planner::VirtualConnectorCoordinator>()?;
    m.add_class::<planner::VirtualConnectorClient>()?;
    m.add_class::<planner::PlannerDecision>()?;
186
187

    engine::add_to_module(m)?;
188
    parsers::add_to_module(m)?;
189

190
    m.add_class::<prometheus_metrics::RuntimeMetrics>()?;
191
192
193
194
    let prometheus_metrics = PyModule::new(m.py(), "prometheus_metrics")?;
    prometheus_metrics::add_to_module(&prometheus_metrics)?;
    m.add_submodule(&prometheus_metrics)?;

195
196
197
198
199
200
201
202
203
204
    Ok(())
}

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

205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
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
223
224
225
226
227
228
229
/// 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);
}

230
231
232
233
234
235
236
/// 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)
}

237
238
/// 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.
239
240
241
242
243
244
245
///
/// 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.
246
#[pyfunction]
247
#[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))]
248
#[allow(clippy::too_many_arguments)]
249
fn register_model<'p>(
250
    py: Python<'p>,
251
    model_input: ModelInput,
252
    model_type: ModelType,
253
254
255
    endpoint: Endpoint,
    model_path: &str,
    model_name: Option<&str>,
256
257
    context_length: Option<u32>,
    kv_cache_block_size: Option<u32>,
258
    router_mode: Option<RouterMode>,
259
    runtime_config: Option<ModelRuntimeConfig>,
260
    user_data: Option<&Bound<'p, PyDict>>,
261
    custom_template_path: Option<&str>,
262
263
    media_decoder: Option<MediaDecoder>,
    media_fetcher: Option<MediaFetcher>,
264
265
    lora_name: Option<&str>,
    base_model_path: Option<&str>,
266
) -> PyResult<Bound<'p, PyAny>> {
267
    // Validate Prefill model type requirements
268
269
270
271
272
273
    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",
        ));
274
275
    }

276
277
278
    let model_input = match model_input {
        ModelInput::Text => llm_rs::model_type::ModelInput::Text,
        ModelInput::Tokens => llm_rs::model_type::ModelInput::Tokens,
279
        ModelInput::Tensor => llm_rs::model_type::ModelInput::Tensor,
280
281
    };

282
    let is_tensor_based = model_type.inner.supports_tensor();
283
    let is_images = model_type.inner.supports_images();
284
    let is_videos = model_type.inner.supports_videos();
285

286
287
    let model_type_obj = model_type.inner;

288
    let inner_path = model_path.to_string();
289
    let model_name = model_name.map(|n| n.to_string());
290
291
292
    let router_mode = router_mode.unwrap_or(RouterMode::RoundRobin);
    let router_config = RouterConfig::new(router_mode.into(), KvRouterConfig::default());

293
294
295
296
297
298
299
300
301
302
303
304
305
    // 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()?;

306
307
308
309
310
311
312
    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))
        })?;

313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
    // 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()));

332
    pyo3_async_runtimes::tokio::future_into_py(py, async move {
333
334
335
        // 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 {
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
            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)
361
362
        let model_path = if fs::exists(&source_path)? {
            PathBuf::from(&source_path)
363
        } else {
364
            LocalModel::fetch(&source_path, false)
365
366
367
368
                .await
                .map_err(to_pyerr)?
        };

369
370
        let mut builder = dynamo_llm::local_model::LocalModelBuilder::default();
        builder
371
            // model path is the physical path on disk of the downloaded model
372
            .model_path(model_path)
373
374
375
376
            // 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
377
            .model_name(model_name.clone())
378
            .context_length(context_length)
379
            .kv_cache_block_size(kv_cache_block_size)
380
            .router_config(Some(router_config))
381
            .runtime_config(runtime_config.unwrap_or_default().inner)
382
            .user_data(user_data_json)
383
384
385
            .custom_template_path(custom_template_path_owned)
            .media_decoder(media_decoder.map(|m| m.inner))
            .media_fetcher(media_fetcher.map(|m| m.inner));
386

387
        let mut local_model = builder.build().await.map_err(to_pyerr)?;
388
389
390
391
392
393
394
395
396

        // 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,
            });

397
        local_model
398
            .attach(&endpoint.inner, model_type_obj, model_input, lora_info)
399
400
401
            .await
            .map_err(to_pyerr)?;

402
403
404
        if let Some(lora_name) = lora_identifier {
            tracing::info!("Registered LoRA '{}' MDC", lora_name);
        } else {
405
406
407
408
            tracing::info!(
                "Registered base model '{}' MDC",
                model_name.unwrap_or(source_path)
            );
409
410
        }

411
412
413
414
        Ok(())
    })
}

415
416
417
418
419
420
421
422
423
424
425
426
427
/// 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}`
428
#[pyfunction]
429
#[pyo3(signature = (endpoint, lora_name=None))]
430
fn unregister_model<'p>(
431
432
433
434
435
436
    py: Python<'p>,
    endpoint: Endpoint,
    lora_name: Option<&str>,
) -> PyResult<Bound<'p, PyAny>> {
    let lora_name_owned = lora_name.map(|s| s.to_string());

437
    pyo3_async_runtimes::tokio::future_into_py(py, async move {
438
439
        // Unified detach method handles both base models and LoRA adapters
        LocalModel::detach_from_endpoint(&endpoint.inner, lora_name_owned.as_deref())
440
441
442
443
444
445
            .await
            .map_err(to_pyerr)?;
        Ok(())
    })
}

446
447
/// Download a model from Hugging Face, returning its local path
/// Example: `model_path = await fetch_model("Qwen/Qwen3-0.6B")`
448
#[pyfunction]
449
#[pyo3(signature = (remote_name, ignore_weights=false))]
450
fn fetch_model<'p>(
451
452
453
454
    py: Python<'p>,
    remote_name: &str,
    ignore_weights: bool,
) -> PyResult<Bound<'p, PyAny>> {
455
456
    let repo = remote_name.to_string();
    pyo3_async_runtimes::tokio::future_into_py(py, async move {
457
458
459
        LocalModel::fetch(&repo, ignore_weights)
            .await
            .map_err(to_pyerr)
460
461
462
    })
}

463
464
#[pyclass]
#[derive(Clone)]
Ryan Olson's avatar
Ryan Olson committed
465
pub struct DistributedRuntime {
466
467
468
469
    inner: rs::DistributedRuntime,
    event_loop: PyObject,
}

Ryan Olson's avatar
Ryan Olson committed
470
471
impl DistributedRuntime {
    #[allow(dead_code)]
472
    pub(crate) fn inner(&self) -> &rs::DistributedRuntime {
Ryan Olson's avatar
Ryan Olson committed
473
474
475
476
        &self.inner
    }
}

477
#[pyclass]
478
#[derive(Clone)]
479
480
481
482
483
struct CancellationToken {
    inner: rs::CancellationToken,
}

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

490
491
492
493
494
495
#[pyclass]
#[derive(Clone)]
struct ModelCardInstanceId {
    inner: rs::discovery::ModelCardInstanceId,
}

496
#[pyclass]
497
#[derive(Clone)]
498
struct Client {
499
    router: rs::pipeline::PushRouter<serde_json::Value, RsAnnotated<serde_json::Value>>,
500
501
}

502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
#[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,
    };
523
524
525
526
    #[classattr]
    const TensorBased: Self = ModelType {
        inner: llm_rs::model_type::ModelType::TensorBased,
    };
527
528
529
530
    #[classattr]
    const Prefill: Self = ModelType {
        inner: llm_rs::model_type::ModelType::Prefill,
    };
531
532
533
534
    #[classattr]
    const Images: Self = ModelType {
        inner: llm_rs::model_type::ModelType::Images,
    };
535
    #[classattr]
536
537
538
539
    const Audios: Self = ModelType {
        inner: llm_rs::model_type::ModelType::Audios,
    };
    #[classattr]
540
541
542
    const Videos: Self = ModelType {
        inner: llm_rs::model_type::ModelType::Videos,
    };
543

544
545
546
547
    fn supports_chat(&self) -> bool {
        self.inner.supports_chat()
    }

548
549
550
551
552
553
554
555
556
557
558
    fn __or__(&self, other: &Self) -> Self {
        ModelType {
            inner: self.inner | other.inner,
        }
    }

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

559
560
#[pyclass(eq, eq_int)]
#[derive(Clone, PartialEq)]
561
562
563
enum ModelInput {
    Text = 1,
    Tokens = 2,
564
    Tensor = 3,
565
566
}

567
568
569
#[pymethods]
impl DistributedRuntime {
    #[new]
570
    #[pyo3(signature = (event_loop, discovery_backend, request_plane, enable_nats=None))]
571
572
    fn new(
        event_loop: PyObject,
573
        discovery_backend: String,
574
575
576
        request_plane: String,
        enable_nats: Option<bool>,
    ) -> PyResult<Self> {
577
578
579
580
581
582
583
        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)
            }
        };
584
        let request_plane: RequestPlaneMode = request_plane.parse().map_err(to_pyerr)?;
585

586
587
588
        // 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()
589
            .or_else(|_| -> anyhow::Result<rs::Runtime> {
590
591
592
593
                // No existing Worker, create new one
                let worker = rs::Worker::from_settings()?;

                // Initialize pyo3 bridge (only happens once per process)
594
                INIT.get_or_try_init(|| -> anyhow::Result<()> {
595
596
                    let primary = worker.tokio_runtime()?;
                    pyo3_async_runtimes::tokio::init_with_runtime(primary).map_err(|e| {
597
                        anyhow::anyhow!("failed to initialize pyo3 static runtime: {:?}", e)
598
                    })?;
599
                    Ok(())
600
601
                })?;

602
                Ok(worker.runtime().clone())
603
604
            })
            .map_err(to_pyerr)?;
605

606
607
        // Initialize logging in context where tokio runtime is available
        // otel exporter requires it
608
        if config::env_is_truthy(env_otlp::OTEL_EXPORT_ENABLED) {
609
610
611
612
            runtime.secondary().block_on(async {
                rs::logging::init();
            });
        }
613

614
615
616
617
618
619
620
621
622
623
        // 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

624
        let runtime_config = DistributedConfig {
625
            discovery_backend: discovery_backend_config,
626
            nats_config: if request_plane.is_nats() || enable_nats {
627
628
629
630
                Some(dynamo_runtime::transports::nats::ClientOptions::default())
            } else {
                None
            },
631
            request_plane,
632
633
634
635
636
        };
        let inner = runtime
            .secondary()
            .block_on(rs::DistributedRuntime::new(runtime, runtime_config))
            .map_err(to_pyerr)?;
637
638
639
640

        Ok(DistributedRuntime { inner, event_loop })
    }

Ryan Olson's avatar
Ryan Olson committed
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
    #[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(),
        })
    }

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
    /// 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,
684
685
686
687
688
            event_loop: self.event_loop.clone(),
        })
    }

    fn shutdown(&self) {
689
        self.inner.shutdown();
690
691
692
693
694
    }

    fn event_loop(&self) -> PyObject {
        self.event_loop.clone()
    }
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
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
    /// 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
774
775
776
777
778
779
780
781
782
783
784
785
786
787
    // 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))
    }
788
789
790
791
}

#[pymethods]
impl Endpoint {
792
    #[pyo3(signature = (generator, graceful_shutdown = true, metrics_labels = None, health_check_payload = None))]
793
794
795
796
    fn serve_endpoint<'p>(
        &self,
        py: Python<'p>,
        generator: PyObject,
797
        graceful_shutdown: Option<bool>,
798
        metrics_labels: Option<Vec<(String, String)>>,
799
        health_check_payload: Option<&Bound<'p, PyDict>>,
800
801
802
803
804
    ) -> PyResult<Bound<'p, PyAny>> {
        let engine = Arc::new(engine::PythonAsyncEngine::new(
            generator,
            self.event_loop.clone(),
        )?);
805
        let ingress = JsonServerStreamingIngress::for_engine(engine.clone()).map_err(to_pyerr)?;
806
807
808
809
810
811
812
813
814
815
816
817
818

        // 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
819
820
821
822
823
824
        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)",
            ));
825
826
827
        }

        let mut builder = self
828
829
830
831
            .inner
            .endpoint_builder()
            .metrics_labels(metrics_labels)
            .handler(ingress);
832
833
834
835
836

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

837
838
839
        // Register the engine in the local endpoint registry for in-process calls
        builder = builder.register_local_engine(engine).map_err(to_pyerr)?;

840
        let graceful_shutdown = graceful_shutdown.unwrap_or(true);
841
        pyo3_async_runtimes::tokio::future_into_py(py, async move {
842
843
844
845
846
            builder
                .graceful_shutdown(graceful_shutdown)
                .start()
                .await
                .map_err(to_pyerr)?;
847
848
849
850
            Ok(())
        })
    }

851
852
853
854
855
856
857
    #[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);
858
859
        let inner = self.inner.clone();
        pyo3_async_runtimes::tokio::future_into_py(py, async move {
860
            let client = inner.client().await.map_err(to_pyerr)?;
861
862
863
            let push_router = rs::pipeline::PushRouter::<
                serde_json::Value,
                RsAnnotated<serde_json::Value>,
864
            >::from_client(client, router_mode.into())
865
866
            .await
            .map_err(to_pyerr)?;
867
868
869
            Ok(Client {
                router: push_router,
            })
870
871
        })
    }
872

873
874
875
    // Opaque unique ID for this worker. May change over worker lifetime.
    fn connection_id(&self) -> u64 {
        self.inner.drt().connection_id()
876
    }
877
878
879

    /// Get a RuntimeMetrics helper for creating Prometheus metrics
    #[getter]
880
881
    fn metrics(&self) -> prometheus_metrics::RuntimeMetrics {
        prometheus_metrics::RuntimeMetrics::from_endpoint(self.inner.clone())
882
    }
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

    /// 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(())
        })
    }
912
913
}

914
915
916
917
918
919
920
921
922
923
924
925
926
#[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(),
        )
    }
}

927
928
#[pymethods]
impl Client {
929
930
    /// Get list of current instances.
    /// Replaces endpoint_ids.
931
    fn instance_ids(&self) -> Vec<u64> {
932
        self.router.client.instance_ids()
933
934
    }

935
936
937
    /// 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>> {
938
        let inner = self.router.client.clone();
939
        pyo3_async_runtimes::tokio::future_into_py(py, async move {
940
            inner
941
                .wait_for_instances()
942
                .await
943
                .map(|v| v.into_iter().map(|cei| cei.id()).collect::<Vec<u64>>())
944
                .map_err(to_pyerr)
945
946
947
948
        })
    }

    /// Issue a request to the endpoint using the default routing strategy.
949
    #[pyo3(signature = (request, annotated=DEFAULT_ANNOTATED_SETTING, context=None))]
950
951
952
953
954
    fn generate<'p>(
        &self,
        py: Python<'p>,
        request: PyObject,
        annotated: Option<bool>,
955
        context: Option<context::Context>,
956
    ) -> PyResult<Bound<'p, PyAny>> {
957
        self.random(py, request, annotated, context)
958
959
960
    }

    /// Send a request to the next endpoint in a round-robin fashion.
961
    #[pyo3(signature = (request, annotated=DEFAULT_ANNOTATED_SETTING, context=None))]
962
963
964
965
966
    fn round_robin<'p>(
        &self,
        py: Python<'p>,
        request: PyObject,
        annotated: Option<bool>,
967
        context: Option<context::Context>,
968
969
    ) -> PyResult<Bound<'p, PyAny>> {
        let request: serde_json::Value = pythonize::depythonize(&request.into_bound(py))?;
970
        let request_ctx = create_request_context(request, &context);
971
972
973
        let annotated = annotated.unwrap_or(false);

        let (tx, rx) = tokio::sync::mpsc::channel(32);
974
        let client = self.router.clone();
975
976

        pyo3_async_runtimes::tokio::future_into_py(py, async move {
977
978
            let stream = match context {
                Some(context) => {
979
980
                    // Always instrument with appropriate span (none if no trace context)
                    let span = get_span_for_context(&context, "round_robin");
981
982
983
984
985
                    client
                        .round_robin(request_ctx)
                        .instrument(span)
                        .await
                        .map_err(to_pyerr)?
986
                }
987
                _ => client.round_robin(request_ctx).await.map_err(to_pyerr)?,
988
            };
989
            tokio::spawn(process_stream(stream, tx));
990
            Ok(AsyncResponseStream::new(rx, annotated))
991
992
993
994
        })
    }

    /// Send a request to a random endpoint.
995
    #[pyo3(signature = (request, annotated=DEFAULT_ANNOTATED_SETTING, context=None))]
996
997
998
999
1000
    fn random<'p>(
        &self,
        py: Python<'p>,
        request: PyObject,
        annotated: Option<bool>,
1001
        context: Option<context::Context>,
1002
1003
    ) -> PyResult<Bound<'p, PyAny>> {
        let request: serde_json::Value = pythonize::depythonize(&request.into_bound(py))?;
1004
        let request_ctx = create_request_context(request, &context);
1005
1006
1007
        let annotated = annotated.unwrap_or(false);

        let (tx, rx) = tokio::sync::mpsc::channel(32);
1008
        let client = self.router.clone();
1009
1010

        pyo3_async_runtimes::tokio::future_into_py(py, async move {
1011
1012
            let stream = match context {
                Some(context) => {
1013
                    // Always instrument with appropriate span (none if no trace context)
1014
                    let span = get_span_for_context(&context, "random");
1015
1016
1017
1018
1019
                    client
                        .random(request_ctx)
                        .instrument(span)
                        .await
                        .map_err(to_pyerr)?
1020
                }
1021
                _ => client.random(request_ctx).await.map_err(to_pyerr)?,
1022
            };
1023
            tokio::spawn(process_stream(stream, tx));
1024
            Ok(AsyncResponseStream::new(rx, annotated))
1025
1026
1027
1028
        })
    }

    /// Directly send a request to a specific endpoint.
1029
    #[pyo3(signature = (request, instance_id, annotated=DEFAULT_ANNOTATED_SETTING, context=None))]
1030
1031
1032
1033
    fn direct<'p>(
        &self,
        py: Python<'p>,
        request: PyObject,
1034
        instance_id: u64,
1035
        annotated: Option<bool>,
1036
        context: Option<context::Context>,
1037
1038
    ) -> PyResult<Bound<'p, PyAny>> {
        let request: serde_json::Value = pythonize::depythonize(&request.into_bound(py))?;
1039
        let request_ctx = create_request_context(request, &context);
1040
1041
1042
        let annotated = annotated.unwrap_or(false);

        let (tx, rx) = tokio::sync::mpsc::channel(32);
1043
        let client = self.router.clone();
1044
1045

        pyo3_async_runtimes::tokio::future_into_py(py, async move {
1046
1047
            let stream = match context {
                Some(context) => {
1048
                    // Always instrument with appropriate span (none if no trace context)
1049
1050
                    let span =
                        get_span_for_direct_context(&context, "direct", &instance_id.to_string());
1051
1052
1053
1054
1055
                    client
                        .direct(request_ctx, instance_id)
                        .instrument(span)
                        .await
                        .map_err(to_pyerr)?
1056
1057
                }
                _ => client
1058
                    .direct(request_ctx, instance_id)
1059
1060
1061
                    .await
                    .map_err(to_pyerr)?,
            };
1062
1063
1064

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

1065
            Ok(AsyncResponseStream::new(rx, annotated))
1066
1067
1068
1069
1070
        })
    }
}

async fn process_stream(
1071
    stream: EngineStream<RsAnnotated<serde_json::Value>>,
1072
1073
1074
1075
1076
    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
1077
        let annotated: RsAnnotated<serde_json::Value> = response;
1078
        let annotated: RsAnnotated<PyObject> = annotated.map_data(|data| {
1079
            Python::with_gil(|py| match pythonize::pythonize(py, &data) {
1080
1081
                Ok(pyobj) => Ok(pyobj.into()),
                Err(e) => Err(e.to_string()),
1082
            })
1083
1084
1085
1086
1087
1088
        });

        let is_error = annotated.is_error();

        // Send the PyObject through the channel or log an error
        if let Err(e) = tx.send(annotated).await {
1089
            tracing::error!("Failed to send response: {:?}", e);
1090
            break;
1091
1092
1093
1094
1095
1096
1097
1098
1099
        }

        if is_error {
            break;
        }
    }
}

#[pyclass]
1100
pub(crate) struct AsyncResponseStream {
1101
1102
1103
    rx: Arc<Mutex<tokio::sync::mpsc::Receiver<RsAnnotated<PyObject>>>>,
    annotated: bool,
}
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115

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

#[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")
        )
    }
}