"docs/components/vscode:/vscode.git/clone" did not exist on "96237ba199a1a8d0fa1ac4b5df61207400c864fe"
lib.rs 41.4 KB
Newer Older
1
// SPDX-FileCopyrightText: Copyright (c) 2024-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
2
3
// SPDX-License-Identifier: Apache-2.0

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

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

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

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

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

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

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

type JsonServerStreamingIngress =
    Ingress<SingleIn<serde_json::Value>, ManyOut<RsAnnotated<serde_json::Value>>>;

static INIT: OnceCell<()> = OnceCell::new();

const DEFAULT_ANNOTATED_SETTING: Option<bool> = Some(true);

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

// Helper to create span for direct method with instance_id
fn get_span_for_direct_context(
    context: &context::Context,
    operation: &str,
    instance_id: &str,
) -> tracing::Span {
93
94
95
96
97
98
    logging::make_client_request_span(
        operation,
        context.inner().id(),
        context.trace_context(),
        Some(instance_id),
    )
99
100
}

101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
// Helper to create request context with proper linking and cancellation handling
fn create_request_context(
    request: serde_json::Value,
    parent_ctx: &Option<context::Context>,
) -> RsContext<serde_json::Value> {
    match parent_ctx {
        // If there is a parent context, link the request as a child context of it
        Some(parent_ctx) => {
            let child_ctx = RsContext::with_id(request, parent_ctx.inner().id().to_string());
            parent_ctx.inner().link_child(child_ctx.context());
            if parent_ctx.inner().is_stopped() || parent_ctx.inner().is_killed() {
                // Let the server handle the cancellation for now since not all backends are
                // properly handling request exceptions
                // TODO: (DIS-830) Return an error if context is cancelled
                child_ctx.context().stop_generating();
            }
            child_ctx
        }
        // Otherwise if there is no parent context, use the request as-is
        _ => request.into(),
    }
}

124
125
126
127
128
/// A Python module implemented in Rust. The name of this function must match
/// the `lib.name` setting in the `Cargo.toml`, else Python will not be able to
/// import the module.
#[pymodule]
fn _core(m: &Bound<'_, PyModule>) -> PyResult<()> {
129
    // Initialize logging early unless OTEL export is enabled (which requires tokio runtime)
130
    if config::env_is_truthy(env_otlp::OTEL_EXPORT_ENABLED) {
131
        eprintln!(
132
            "Warning: OTEL_EXPORT_ENABLED detected. Logging initialization deferred until runtime is available. Early logs may be dropped."
133
134
135
136
137
        );
    } else {
        rs::logging::init();
    }

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

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

    engine::add_to_module(m)?;
186
    parsers::add_to_module(m)?;
187

188
    m.add_class::<prometheus_metrics::RuntimeMetrics>()?;
189
190
191
192
    let prometheus_metrics = PyModule::new(m.py(), "prometheus_metrics")?;
    prometheus_metrics::add_to_module(&prometheus_metrics)?;
    m.add_submodule(&prometheus_metrics)?;

193
194
195
196
197
198
199
200
201
202
    Ok(())
}

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

Ryan Olson's avatar
Ryan Olson committed
203
204
205
206
207
208
209
/// 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);
}

210
211
212
213
214
215
216
/// 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)
}

217
218
/// 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.
219
220
221
222
223
224
225
///
/// 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.
226
#[pyfunction]
227
#[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))]
228
#[allow(clippy::too_many_arguments)]
229
230
fn register_llm<'p>(
    py: Python<'p>,
231
    model_input: ModelInput,
232
    model_type: ModelType,
233
234
235
    endpoint: Endpoint,
    model_path: &str,
    model_name: Option<&str>,
236
237
    context_length: Option<u32>,
    kv_cache_block_size: Option<u32>,
238
    router_mode: Option<RouterMode>,
239
    runtime_config: Option<ModelRuntimeConfig>,
240
    user_data: Option<&Bound<'p, PyDict>>,
241
    custom_template_path: Option<&str>,
242
243
    media_decoder: Option<MediaDecoder>,
    media_fetcher: Option<MediaFetcher>,
244
245
    lora_name: Option<&str>,
    base_model_path: Option<&str>,
246
) -> PyResult<Bound<'p, PyAny>> {
247
    // Validate Prefill model type requirements
248
249
250
251
252
253
    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",
        ));
254
255
    }

256
257
258
    let model_input = match model_input {
        ModelInput::Text => llm_rs::model_type::ModelInput::Text,
        ModelInput::Tokens => llm_rs::model_type::ModelInput::Tokens,
259
        ModelInput::Tensor => llm_rs::model_type::ModelInput::Tensor,
260
261
    };

262
    let is_tensor_based = model_type.inner.supports_tensor();
263
    let is_images = model_type.inner.supports_images();
264

265
266
    let model_type_obj = model_type.inner;

267
    let inner_path = model_path.to_string();
268
    let model_name = model_name.map(|n| n.to_string());
269
270
271
    let router_mode = router_mode.unwrap_or(RouterMode::RoundRobin);
    let router_config = RouterConfig::new(router_mode.into(), KvRouterConfig::default());

272
273
274
275
276
277
278
279
280
281
282
283
284
    // 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()?;

285
286
287
288
289
290
291
    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))
        })?;

292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
    // 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()));

311
    pyo3_async_runtimes::tokio::future_into_py(py, async move {
312
        // For TensorBased and Images models, skip HuggingFace downloads and register directly
313
        // Images models (vLLM-Omni) handle model loading internally, no tokenizer extraction needed
314
        if is_tensor_based || is_images {
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
            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)
340
341
        let model_path = if fs::exists(&source_path)? {
            PathBuf::from(&source_path)
342
        } else {
343
            LocalModel::fetch(&source_path, false)
344
345
346
347
                .await
                .map_err(to_pyerr)?
        };

348
349
        let mut builder = dynamo_llm::local_model::LocalModelBuilder::default();
        builder
350
            // model path is the physical path on disk of the downloaded model
351
            .model_path(model_path)
352
353
354
355
            // 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
356
            .model_name(model_name.clone())
357
            .context_length(context_length)
358
            .kv_cache_block_size(kv_cache_block_size)
359
            .router_config(Some(router_config))
360
            .runtime_config(runtime_config.unwrap_or_default().inner)
361
            .user_data(user_data_json)
362
363
364
            .custom_template_path(custom_template_path_owned)
            .media_decoder(media_decoder.map(|m| m.inner))
            .media_fetcher(media_fetcher.map(|m| m.inner));
365

366
        let mut local_model = builder.build().await.map_err(to_pyerr)?;
367
368
369
370
371
372
373
374
375

        // 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,
            });

376
        local_model
377
            .attach(&endpoint.inner, model_type_obj, model_input, lora_info)
378
379
380
            .await
            .map_err(to_pyerr)?;

381
382
383
        if let Some(lora_name) = lora_identifier {
            tracing::info!("Registered LoRA '{}' MDC", lora_name);
        } else {
384
385
386
387
            tracing::info!(
                "Registered base model '{}' MDC",
                model_name.unwrap_or(source_path)
            );
388
389
        }

390
391
392
393
        Ok(())
    })
}

394
395
396
397
398
399
400
401
402
403
404
405
406
/// 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}`
407
#[pyfunction]
408
409
410
411
412
413
414
415
#[pyo3(signature = (endpoint, lora_name=None))]
fn unregister_llm<'p>(
    py: Python<'p>,
    endpoint: Endpoint,
    lora_name: Option<&str>,
) -> PyResult<Bound<'p, PyAny>> {
    let lora_name_owned = lora_name.map(|s| s.to_string());

416
    pyo3_async_runtimes::tokio::future_into_py(py, async move {
417
418
        // Unified detach method handles both base models and LoRA adapters
        LocalModel::detach_from_endpoint(&endpoint.inner, lora_name_owned.as_deref())
419
420
421
422
423
424
            .await
            .map_err(to_pyerr)?;
        Ok(())
    })
}

425
426
427
/// Download a model from Hugging Face, returning it's local path
/// Example: `model_path = await fetch_llm("Qwen/Qwen3-0.6B")`
#[pyfunction]
428
429
430
431
432
433
#[pyo3(signature = (remote_name, ignore_weights=false))]
fn fetch_llm<'p>(
    py: Python<'p>,
    remote_name: &str,
    ignore_weights: bool,
) -> PyResult<Bound<'p, PyAny>> {
434
435
    let repo = remote_name.to_string();
    pyo3_async_runtimes::tokio::future_into_py(py, async move {
436
437
438
        LocalModel::fetch(&repo, ignore_weights)
            .await
            .map_err(to_pyerr)
439
440
441
    })
}

442
443
#[pyclass]
#[derive(Clone)]
Ryan Olson's avatar
Ryan Olson committed
444
pub struct DistributedRuntime {
445
446
447
448
    inner: rs::DistributedRuntime,
    event_loop: PyObject,
}

Ryan Olson's avatar
Ryan Olson committed
449
450
impl DistributedRuntime {
    #[allow(dead_code)]
451
    pub(crate) fn inner(&self) -> &rs::DistributedRuntime {
Ryan Olson's avatar
Ryan Olson committed
452
453
454
455
        &self.inner
    }
}

456
#[pyclass]
457
#[derive(Clone)]
458
459
460
461
462
struct CancellationToken {
    inner: rs::CancellationToken,
}

#[pyclass]
463
#[derive(Clone)]
464
465
466
467
468
469
struct Namespace {
    inner: rs::component::Namespace,
    event_loop: PyObject,
}

#[pyclass]
470
#[derive(Clone)]
471
472
473
474
475
476
struct Component {
    inner: rs::component::Component,
    event_loop: PyObject,
}

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

483
484
485
486
487
488
#[pyclass]
#[derive(Clone)]
struct ModelCardInstanceId {
    inner: rs::discovery::ModelCardInstanceId,
}

489
#[pyclass]
490
#[derive(Clone)]
491
struct Client {
492
    router: rs::pipeline::PushRouter<serde_json::Value, RsAnnotated<serde_json::Value>>,
493
494
}

495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
#[pyclass]
#[derive(Clone, PartialEq)]
struct ModelType {
    inner: llm_rs::model_type::ModelType,
}

#[pymethods]
#[allow(non_upper_case_globals)]
impl ModelType {
    #[classattr]
    const Chat: Self = ModelType {
        inner: llm_rs::model_type::ModelType::Chat,
    };
    #[classattr]
    const Completions: Self = ModelType {
        inner: llm_rs::model_type::ModelType::Completions,
    };
    #[classattr]
    const Embedding: Self = ModelType {
        inner: llm_rs::model_type::ModelType::Embedding,
    };
516
517
518
519
    #[classattr]
    const TensorBased: Self = ModelType {
        inner: llm_rs::model_type::ModelType::TensorBased,
    };
520
521
522
523
    #[classattr]
    const Prefill: Self = ModelType {
        inner: llm_rs::model_type::ModelType::Prefill,
    };
524
525
526
527
    #[classattr]
    const Images: Self = ModelType {
        inner: llm_rs::model_type::ModelType::Images,
    };
528

529
530
531
532
    fn supports_chat(&self) -> bool {
        self.inner.supports_chat()
    }

533
534
535
536
537
538
539
540
541
542
543
    fn __or__(&self, other: &Self) -> Self {
        ModelType {
            inner: self.inner | other.inner,
        }
    }

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

544
545
#[pyclass(eq, eq_int)]
#[derive(Clone, PartialEq)]
546
547
548
enum ModelInput {
    Text = 1,
    Tokens = 2,
549
    Tensor = 3,
550
551
}

552
553
554
#[pymethods]
impl DistributedRuntime {
    #[new]
555
556
557
558
559
560
561
    #[pyo3(signature = (event_loop, store_kv, request_plane, enable_nats=None))]
    fn new(
        event_loop: PyObject,
        store_kv: String,
        request_plane: String,
        enable_nats: Option<bool>,
    ) -> PyResult<Self> {
562
        let selected_kv_store: kv::Selector = store_kv.parse().map_err(to_pyerr)?;
563
        let request_plane: RequestPlaneMode = request_plane.parse().map_err(to_pyerr)?;
564

565
566
567
        // 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()
568
            .or_else(|_| -> anyhow::Result<rs::Runtime> {
569
570
571
572
                // No existing Worker, create new one
                let worker = rs::Worker::from_settings()?;

                // Initialize pyo3 bridge (only happens once per process)
573
                INIT.get_or_try_init(|| -> anyhow::Result<()> {
574
575
                    let primary = worker.tokio_runtime()?;
                    pyo3_async_runtimes::tokio::init_with_runtime(primary).map_err(|e| {
576
                        anyhow::anyhow!("failed to initialize pyo3 static runtime: {:?}", e)
577
                    })?;
578
                    Ok(())
579
580
                })?;

581
                Ok(worker.runtime().clone())
582
583
            })
            .map_err(to_pyerr)?;
584

585
586
        // Initialize logging in context where tokio runtime is available
        // otel exporter requires it
587
        if config::env_is_truthy(env_otlp::OTEL_EXPORT_ENABLED) {
588
589
590
591
            runtime.secondary().block_on(async {
                rs::logging::init();
            });
        }
592

593
594
595
596
597
598
599
600
601
602
        // 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

603
604
        let runtime_config = DistributedConfig {
            store_backend: selected_kv_store,
605
            nats_config: if request_plane.is_nats() || enable_nats {
606
607
608
609
                Some(dynamo_runtime::transports::nats::ClientOptions::default())
            } else {
                None
            },
610
            request_plane,
611
612
613
614
615
        };
        let inner = runtime
            .secondary()
            .block_on(rs::DistributedRuntime::new(runtime, runtime_config))
            .map_err(to_pyerr)?;
616
617
618
619

        Ok(DistributedRuntime { inner, event_loop })
    }

Ryan Olson's avatar
Ryan Olson committed
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
    #[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(),
        })
    }

635
636
637
638
639
640
641
642
    fn namespace(&self, name: String) -> PyResult<Namespace> {
        Ok(Namespace {
            inner: self.inner.namespace(name).map_err(to_pyerr)?,
            event_loop: self.event_loop.clone(),
        })
    }

    fn shutdown(&self) {
643
        self.inner.shutdown();
644
645
646
647
648
    }

    fn event_loop(&self) -> PyObject {
        self.event_loop.clone()
    }
649
650
651
652
653

    fn child_token(&self) -> CancellationToken {
        let inner = self.inner.runtime().child_token();
        CancellationToken { inner }
    }
Richard Huo's avatar
Richard Huo committed
654

655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
    /// 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
733
734
735
736
737
738
739
740
741
742
743
744
745
746
    // 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))
    }
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
}

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

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

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

#[pymethods]
impl Endpoint {
777
    #[pyo3(signature = (generator, graceful_shutdown = true, metrics_labels = None, health_check_payload = None))]
778
779
780
781
    fn serve_endpoint<'p>(
        &self,
        py: Python<'p>,
        generator: PyObject,
782
        graceful_shutdown: Option<bool>,
783
        metrics_labels: Option<Vec<(String, String)>>,
784
        health_check_payload: Option<&Bound<'p, PyDict>>,
785
786
787
788
789
    ) -> PyResult<Bound<'p, PyAny>> {
        let engine = Arc::new(engine::PythonAsyncEngine::new(
            generator,
            self.event_loop.clone(),
        )?);
790
        let ingress = JsonServerStreamingIngress::for_engine(engine.clone()).map_err(to_pyerr)?;
791
792
793
794
795
796
797
798
799
800
801
802
803

        // 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
804
805
806
807
808
809
        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)",
            ));
810
811
812
        }

        let mut builder = self
813
814
815
816
            .inner
            .endpoint_builder()
            .metrics_labels(metrics_labels)
            .handler(ingress);
817
818
819
820
821

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

822
823
824
        // Register the engine in the local endpoint registry for in-process calls
        builder = builder.register_local_engine(engine).map_err(to_pyerr)?;

825
        let graceful_shutdown = graceful_shutdown.unwrap_or(true);
826
        pyo3_async_runtimes::tokio::future_into_py(py, async move {
827
828
829
830
831
            builder
                .graceful_shutdown(graceful_shutdown)
                .start()
                .await
                .map_err(to_pyerr)?;
832
833
834
835
            Ok(())
        })
    }

836
837
838
839
840
841
842
    #[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);
843
844
        let inner = self.inner.clone();
        pyo3_async_runtimes::tokio::future_into_py(py, async move {
845
            let client = inner.client().await.map_err(to_pyerr)?;
846
847
848
            let push_router = rs::pipeline::PushRouter::<
                serde_json::Value,
                RsAnnotated<serde_json::Value>,
849
            >::from_client(client, router_mode.into())
850
851
            .await
            .map_err(to_pyerr)?;
852
853
854
            Ok(Client {
                router: push_router,
            })
855
856
        })
    }
857

858
859
860
    // Opaque unique ID for this worker. May change over worker lifetime.
    fn connection_id(&self) -> u64 {
        self.inner.drt().connection_id()
861
    }
862
863
864

    /// Get a RuntimeMetrics helper for creating Prometheus metrics
    #[getter]
865
866
    fn metrics(&self) -> prometheus_metrics::RuntimeMetrics {
        prometheus_metrics::RuntimeMetrics::from_endpoint(self.inner.clone())
867
    }
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896

    /// 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(())
        })
    }
897
898
899
900
901
902
903
904
905
906
907
908
909
}

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

910
911
912
913
914
915
916
917
918
919
920
921
922
#[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(),
        )
    }
}

923
924
#[pymethods]
impl Client {
925
926
    /// Get list of current instances.
    /// Replaces endpoint_ids.
927
    fn instance_ids(&self) -> Vec<u64> {
928
        self.router.client.instance_ids()
929
930
    }

931
932
933
    /// 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>> {
934
        let inner = self.router.client.clone();
935
        pyo3_async_runtimes::tokio::future_into_py(py, async move {
936
            inner
937
                .wait_for_instances()
938
                .await
939
                .map(|v| v.into_iter().map(|cei| cei.id()).collect::<Vec<u64>>())
940
                .map_err(to_pyerr)
941
942
943
944
        })
    }

    /// Issue a request to the endpoint using the default routing strategy.
945
    #[pyo3(signature = (request, annotated=DEFAULT_ANNOTATED_SETTING, context=None))]
946
947
948
949
950
    fn generate<'p>(
        &self,
        py: Python<'p>,
        request: PyObject,
        annotated: Option<bool>,
951
        context: Option<context::Context>,
952
    ) -> PyResult<Bound<'p, PyAny>> {
953
        self.random(py, request, annotated, context)
954
955
956
    }

    /// Send a request to the next endpoint in a round-robin fashion.
957
    #[pyo3(signature = (request, annotated=DEFAULT_ANNOTATED_SETTING, context=None))]
958
959
960
961
962
    fn round_robin<'p>(
        &self,
        py: Python<'p>,
        request: PyObject,
        annotated: Option<bool>,
963
        context: Option<context::Context>,
964
965
    ) -> PyResult<Bound<'p, PyAny>> {
        let request: serde_json::Value = pythonize::depythonize(&request.into_bound(py))?;
966
        let request_ctx = create_request_context(request, &context);
967
968
969
        let annotated = annotated.unwrap_or(false);

        let (tx, rx) = tokio::sync::mpsc::channel(32);
970
        let client = self.router.clone();
971
972

        pyo3_async_runtimes::tokio::future_into_py(py, async move {
973
974
            let stream = match context {
                Some(context) => {
975
976
                    // Always instrument with appropriate span (none if no trace context)
                    let span = get_span_for_context(&context, "round_robin");
977
978
979
980
981
                    client
                        .round_robin(request_ctx)
                        .instrument(span)
                        .await
                        .map_err(to_pyerr)?
982
                }
983
                _ => client.round_robin(request_ctx).await.map_err(to_pyerr)?,
984
            };
985
986
987
988
989
990
991
992
993
            tokio::spawn(process_stream(stream, tx));
            Ok(AsyncResponseStream {
                rx: Arc::new(Mutex::new(rx)),
                annotated,
            })
        })
    }

    /// Send a request to a random endpoint.
994
    #[pyo3(signature = (request, annotated=DEFAULT_ANNOTATED_SETTING, context=None))]
995
996
997
998
999
    fn random<'p>(
        &self,
        py: Python<'p>,
        request: PyObject,
        annotated: Option<bool>,
1000
        context: Option<context::Context>,
1001
1002
    ) -> PyResult<Bound<'p, PyAny>> {
        let request: serde_json::Value = pythonize::depythonize(&request.into_bound(py))?;
1003
        let request_ctx = create_request_context(request, &context);
1004
1005
1006
        let annotated = annotated.unwrap_or(false);

        let (tx, rx) = tokio::sync::mpsc::channel(32);
1007
        let client = self.router.clone();
1008
1009

        pyo3_async_runtimes::tokio::future_into_py(py, async move {
1010
1011
            let stream = match context {
                Some(context) => {
1012
                    // Always instrument with appropriate span (none if no trace context)
1013
                    let span = get_span_for_context(&context, "random");
1014
1015
1016
1017
1018
                    client
                        .random(request_ctx)
                        .instrument(span)
                        .await
                        .map_err(to_pyerr)?
1019
                }
1020
                _ => client.random(request_ctx).await.map_err(to_pyerr)?,
1021
            };
1022
1023
1024
1025
1026
1027
1028
1029
1030
            tokio::spawn(process_stream(stream, tx));
            Ok(AsyncResponseStream {
                rx: Arc::new(Mutex::new(rx)),
                annotated,
            })
        })
    }

    /// Directly send a request to a specific endpoint.
1031
    #[pyo3(signature = (request, instance_id, annotated=DEFAULT_ANNOTATED_SETTING, context=None))]
1032
1033
1034
1035
    fn direct<'p>(
        &self,
        py: Python<'p>,
        request: PyObject,
1036
        instance_id: u64,
1037
        annotated: Option<bool>,
1038
        context: Option<context::Context>,
1039
1040
    ) -> PyResult<Bound<'p, PyAny>> {
        let request: serde_json::Value = pythonize::depythonize(&request.into_bound(py))?;
1041
        let request_ctx = create_request_context(request, &context);
1042
1043
1044
        let annotated = annotated.unwrap_or(false);

        let (tx, rx) = tokio::sync::mpsc::channel(32);
1045
        let client = self.router.clone();
1046
1047

        pyo3_async_runtimes::tokio::future_into_py(py, async move {
1048
1049
            let stream = match context {
                Some(context) => {
1050
                    // Always instrument with appropriate span (none if no trace context)
1051
1052
                    let span =
                        get_span_for_direct_context(&context, "direct", &instance_id.to_string());
1053
1054
1055
1056
1057
                    client
                        .direct(request_ctx, instance_id)
                        .instrument(span)
                        .await
                        .map_err(to_pyerr)?
1058
1059
                }
                _ => client
1060
                    .direct(request_ctx, instance_id)
1061
1062
1063
                    .await
                    .map_err(to_pyerr)?,
            };
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075

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

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

async fn process_stream(
1076
    stream: EngineStream<RsAnnotated<serde_json::Value>>,
1077
1078
1079
1080
1081
    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
1082
        let annotated: RsAnnotated<serde_json::Value> = response;
1083
        let annotated: RsAnnotated<PyObject> = annotated.map_data(|data| {
1084
            Python::with_gil(|py| match pythonize::pythonize(py, &data) {
1085
1086
                Ok(pyobj) => Ok(pyobj.into()),
                Err(e) => Err(e.to_string()),
1087
            })
1088
1089
1090
1091
1092
1093
        });

        let is_error = annotated.is_error();

        // Send the PyObject through the channel or log an error
        if let Err(e) = tx.send(annotated).await {
1094
            tracing::error!("Failed to send response: {:?}", e);
1095
            break;
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
        }

        if is_error {
            break;
        }
    }
}

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

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

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

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

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

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

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

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

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

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

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

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

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