service_v2.rs 18.2 KB
Newer Older
1
2
// SPDX-FileCopyrightText: Copyright (c) 2024-2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
// SPDX-License-Identifier: Apache-2.0
3

4
use std::collections::HashMap;
5
use std::env::var;
Graham King's avatar
Graham King committed
6
use std::path::PathBuf;
7
use std::sync::Arc;
8
9
use std::sync::atomic::AtomicBool;
use std::sync::atomic::Ordering;
10
use std::time::Duration;
11

12
use super::Metrics;
13
use super::RouteDoc;
14
use super::metrics;
15
use crate::discovery::ModelManager;
16
use crate::endpoint_type::EndpointType;
17
use crate::request_template::RequestTemplate;
18
use anyhow::Result;
Graham King's avatar
Graham King committed
19
use axum_server::tls_rustls::RustlsConfig;
20
use derive_builder::Builder;
21
use dynamo_runtime::logging::make_request_span;
22
use dynamo_runtime::metrics::prometheus_names::name_prefix;
23
24
25
use dynamo_runtime::storage::key_value_store::EtcdStore;
use dynamo_runtime::storage::key_value_store::KeyValueStore;
use dynamo_runtime::storage::key_value_store::MemoryStore;
26
use dynamo_runtime::transports::etcd;
Graham King's avatar
Graham King committed
27
use std::net::SocketAddr;
28
use tokio::task::JoinHandle;
29
use tokio_util::sync::CancellationToken;
30
use tower_http::trace::TraceLayer;
31

32
33
34
35
/// HTTP service shared state
pub struct State {
    metrics: Arc<Metrics>,
    manager: Arc<ModelManager>,
36
    etcd_client: Option<etcd::Client>,
37
    store: Arc<dyn KeyValueStore>,
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
    flags: StateFlags,
}

#[derive(Default, Debug)]
struct StateFlags {
    chat_endpoints_enabled: AtomicBool,
    cmpl_endpoints_enabled: AtomicBool,
    embeddings_endpoints_enabled: AtomicBool,
    responses_endpoints_enabled: AtomicBool,
}

impl StateFlags {
    pub fn get(&self, endpoint_type: &EndpointType) -> bool {
        match endpoint_type {
            EndpointType::Chat => self.chat_endpoints_enabled.load(Ordering::Relaxed),
            EndpointType::Completion => self.cmpl_endpoints_enabled.load(Ordering::Relaxed),
            EndpointType::Embedding => self.embeddings_endpoints_enabled.load(Ordering::Relaxed),
            EndpointType::Responses => self.responses_endpoints_enabled.load(Ordering::Relaxed),
        }
    }

    pub fn set(&self, endpoint_type: &EndpointType, enabled: bool) {
        match endpoint_type {
            EndpointType::Chat => self
                .chat_endpoints_enabled
                .store(enabled, Ordering::Relaxed),
            EndpointType::Completion => self
                .cmpl_endpoints_enabled
                .store(enabled, Ordering::Relaxed),
            EndpointType::Embedding => self
                .embeddings_endpoints_enabled
                .store(enabled, Ordering::Relaxed),
            EndpointType::Responses => self
                .responses_endpoints_enabled
                .store(enabled, Ordering::Relaxed),
        }
    }
75
76
77
78
79
80
81
}

impl State {
    pub fn new(manager: Arc<ModelManager>) -> Self {
        Self {
            manager,
            metrics: Arc::new(Metrics::default()),
82
            etcd_client: None,
83
            store: Arc::new(MemoryStore::new()),
84
85
86
87
88
89
            flags: StateFlags {
                chat_endpoints_enabled: AtomicBool::new(false),
                cmpl_endpoints_enabled: AtomicBool::new(false),
                embeddings_endpoints_enabled: AtomicBool::new(false),
                responses_endpoints_enabled: AtomicBool::new(false),
            },
90
91
92
        }
    }

93
    pub fn new_with_etcd(manager: Arc<ModelManager>, etcd_client: etcd::Client) -> Self {
94
95
96
        Self {
            manager,
            metrics: Arc::new(Metrics::default()),
97
98
            store: Arc::new(EtcdStore::new(etcd_client.clone())),
            etcd_client: Some(etcd_client),
99
100
101
102
103
104
            flags: StateFlags {
                chat_endpoints_enabled: AtomicBool::new(false),
                cmpl_endpoints_enabled: AtomicBool::new(false),
                embeddings_endpoints_enabled: AtomicBool::new(false),
                responses_endpoints_enabled: AtomicBool::new(false),
            },
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
        }
    }
    /// Get the Prometheus [`Metrics`] object which tracks request counts and inflight requests
    pub fn metrics_clone(&self) -> Arc<Metrics> {
        self.metrics.clone()
    }

    pub fn manager(&self) -> &ModelManager {
        Arc::as_ref(&self.manager)
    }

    pub fn manager_clone(&self) -> Arc<ModelManager> {
        self.manager.clone()
    }

120
121
122
123
    pub fn etcd_client(&self) -> Option<&etcd::Client> {
        self.etcd_client.as_ref()
    }

124
125
126
127
    pub fn store(&self) -> Arc<dyn KeyValueStore> {
        self.store.clone()
    }

128
129
130
131
132
133
    // TODO
    pub fn sse_keep_alive(&self) -> Option<Duration> {
        None
    }
}

134
135
#[derive(Clone)]
pub struct HttpService {
136
137
138
    // The state we share with every request handler
    state: Arc<State>,

139
140
    router: axum::Router,
    port: u16,
141
    host: String,
Graham King's avatar
Graham King committed
142
143
144
    enable_tls: bool,
    tls_cert_path: Option<PathBuf>,
    tls_key_path: Option<PathBuf>,
145
    route_docs: Vec<RouteDoc>,
146
147
148
149
150
151

    // DEPRECATED: To be removed after custom backends migrate to Dynamo backend.
    pub(crate) custom_backend_namespace_component_endpoint: Option<String>,
    pub(crate) custom_backend_metrics_polling_interval: Option<f64>,
    pub(crate) custom_backend_registry:
        Option<Arc<super::custom_backend_metrics::CustomBackendMetricsRegistry>>,
152
153
154
}

#[derive(Clone, Builder)]
155
#[builder(pattern = "owned", build_fn(private, name = "build_internal"))]
156
157
158
159
pub struct HttpServiceConfig {
    #[builder(default = "8787")]
    port: u16,

160
161
162
    #[builder(setter(into), default = "String::from(\"0.0.0.0\")")]
    host: String,

Graham King's avatar
Graham King committed
163
164
165
166
167
168
169
170
171
    #[builder(default = "false")]
    enable_tls: bool,

    #[builder(default = "None")]
    tls_cert_path: Option<PathBuf>,

    #[builder(default = "None")]
    tls_key_path: Option<PathBuf>,

172
173
    // #[builder(default)]
    // custom: Vec<axum::Router>
174
    #[builder(default = "false")]
175
176
    enable_chat_endpoints: bool,

177
    #[builder(default = "false")]
178
    enable_cmpl_endpoints: bool,
179

180
    #[builder(default = "true")]
181
182
    enable_embeddings_endpoints: bool,

183
184
185
    #[builder(default = "true")]
    enable_responses_endpoints: bool,

186
187
    #[builder(default = "None")]
    request_template: Option<RequestTemplate>,
188
189
190

    #[builder(default = "None")]
    etcd_client: Option<etcd::Client>,
191
192
193
194
195
196
197

    // DEPRECATED: To be removed after custom backends migrate to Dynamo backend.
    #[builder(default = "None")]
    custom_backend_namespace_component_endpoint: Option<String>,

    #[builder(default = "None")]
    custom_backend_metrics_polling_interval: Option<f64>,
198
199
200
201
202
203
204
}

impl HttpService {
    pub fn builder() -> HttpServiceConfigBuilder {
        HttpServiceConfigBuilder::default()
    }

205
206
207
208
209
210
211
212
    pub fn state_clone(&self) -> Arc<State> {
        self.state.clone()
    }

    pub fn state(&self) -> &State {
        Arc::as_ref(&self.state)
    }

213
    pub fn model_manager(&self) -> &ModelManager {
214
        self.state().manager()
215
216
    }

217
218
219
220
221
222
    pub async fn spawn(&self, cancel_token: CancellationToken) -> JoinHandle<Result<()>> {
        let this = self.clone();
        tokio::spawn(async move { this.run(cancel_token).await })
    }

    pub async fn run(&self, cancel_token: CancellationToken) -> Result<()> {
223
        let address = format!("{}:{}", self.host, self.port);
Graham King's avatar
Graham King committed
224
225
        let protocol = if self.enable_tls { "HTTPS" } else { "HTTP" };
        tracing::info!(protocol, address, "Starting HTTP(S) service");
226
227
228
229

        let router = self.router.clone();
        let observer = cancel_token.child_token();

Graham King's avatar
Graham King committed
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
        let addr: SocketAddr = address
            .parse()
            .map_err(|e| anyhow::anyhow!("Invalid address '{}': {}", address, e))?;

        if self.enable_tls {
            let cert_path = self
                .tls_cert_path
                .as_ref()
                .ok_or_else(|| anyhow::anyhow!("TLS certificate path not provided"))?;
            let key_path = self
                .tls_key_path
                .as_ref()
                .ok_or_else(|| anyhow::anyhow!("TLS private key path not provided"))?;

            // aws_lc_rs is the default but other crates pull in `ring` also,
            // so rustls doesn't know which one to use. Tell it.
            if let Err(e) = rustls::crypto::aws_lc_rs::default_provider().install_default() {
                tracing::debug!("TLS crypto provider already installed: {e:?}");
            }

            let config = RustlsConfig::from_pem_file(cert_path, key_path)
                .await
                .map_err(|e| anyhow::anyhow!("Failed to create TLS config: {}", e))?;

            let handle = axum_server::Handle::new();
            let server = axum_server::bind_rustls(addr, config)
                .handle(handle.clone())
                .serve(router.into_make_service());

            tokio::select! {
                result = server => {
                    result.map_err(|e| anyhow::anyhow!("HTTPS server error: {}", e))?;
                }
                _ = observer.cancelled() => {
                    tracing::info!("HTTPS server shutdown requested");
                    handle.graceful_shutdown(Some(Duration::from_secs(5)));
                    // TODO: Do we need to wait?
                }
            }
        } else {
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
            let listener = tokio::net::TcpListener::bind(addr).await.map_err(|e| {
                tracing::error!(
                    protocol = %protocol,
                    address = %address,
                    error = %e,
                    "Failed to bind server to address"
                );
                match e.kind() {
                    std::io::ErrorKind::AddrInUse => anyhow::anyhow!(
                        "Failed to start {} server: port {} already in use. Use --http-port to specify a different port.",
                        protocol,
                        self.port
                    ),
                    _ => anyhow::anyhow!(
                        "Failed to start {} server on {}: {}",
                        protocol,
                        address,
                        e
                    ),
                }
            })?;
Graham King's avatar
Graham King committed
291
292
293
294
295
296

            axum::serve(listener, router)
                .with_graceful_shutdown(observer.cancelled_owned())
                .await
                .inspect_err(|_| cancel_token.cancel())?;
        }
297
298

        Ok(())
299
    }
300
301
302
303
304

    /// Documentation of exposed HTTP endpoints
    pub fn route_docs(&self) -> &[RouteDoc] {
        &self.route_docs
    }
305

306
    pub fn enable_model_endpoint(&self, endpoint_type: EndpointType, enable: bool) {
307
308
309
310
311
312
313
        self.state.flags.set(&endpoint_type, enable);
        tracing::info!(
            "{} endpoints {}",
            endpoint_type.as_str(),
            if enable { "enabled" } else { "disabled" }
        );
    }
314
315
}

316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
/// Environment variable to set the metrics endpoint path (default: `/metrics`)
static HTTP_SVC_METRICS_PATH_ENV: &str = "DYN_HTTP_SVC_METRICS_PATH";
/// Environment variable to set the models endpoint path (default: `/v1/models`)
static HTTP_SVC_MODELS_PATH_ENV: &str = "DYN_HTTP_SVC_MODELS_PATH";
/// Environment variable to set the health endpoint path (default: `/health`)
static HTTP_SVC_HEALTH_PATH_ENV: &str = "DYN_HTTP_SVC_HEALTH_PATH";
/// Environment variable to set the live endpoint path (default: `/live`)
static HTTP_SVC_LIVE_PATH_ENV: &str = "DYN_HTTP_SVC_LIVE_PATH";
/// Environment variable to set the chat completions endpoint path (default: `/v1/chat/completions`)
static HTTP_SVC_CHAT_PATH_ENV: &str = "DYN_HTTP_SVC_CHAT_PATH";
/// Environment variable to set the completions endpoint path (default: `/v1/completions`)
static HTTP_SVC_CMP_PATH_ENV: &str = "DYN_HTTP_SVC_CMP_PATH";
/// Environment variable to set the embeddings endpoint path (default: `/v1/embeddings`)
static HTTP_SVC_EMB_PATH_ENV: &str = "DYN_HTTP_SVC_EMB_PATH";
/// Environment variable to set the responses endpoint path (default: `/v1/responses`)
static HTTP_SVC_RESPONSES_PATH_ENV: &str = "DYN_HTTP_SVC_RESPONSES_PATH";

333
334
impl HttpServiceConfigBuilder {
    pub fn build(self) -> Result<HttpService, anyhow::Error> {
335
        let config: HttpServiceConfig = self.build_internal()?;
336

337
        let model_manager = Arc::new(ModelManager::new());
338
339
340
341
        let state = match config.etcd_client {
            Some(etcd_client) => Arc::new(State::new_with_etcd(model_manager, etcd_client)),
            None => Arc::new(State::new(model_manager)),
        };
342
343
344
345
346
347
348
349
350
351
352
353
354
        state
            .flags
            .set(&EndpointType::Chat, config.enable_chat_endpoints);
        state
            .flags
            .set(&EndpointType::Completion, config.enable_cmpl_endpoints);
        state
            .flags
            .set(&EndpointType::Embedding, config.enable_embeddings_endpoints);
        state
            .flags
            .set(&EndpointType::Responses, config.enable_responses_endpoints);

355
356
        // enable prometheus metrics
        let registry = metrics::Registry::new();
357
        state.metrics_clone().register(&registry)?;
358

359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
        // DEPRECATED: To be removed after custom backends migrate to Dynamo backend.
        // Setup custom backend metrics if configured
        let custom_backend_registry =
            if config.custom_backend_namespace_component_endpoint.is_some()
                && config.custom_backend_metrics_polling_interval.is_some()
            {
                Some(Arc::new(
                    super::custom_backend_metrics::CustomBackendMetricsRegistry::new(
                        name_prefix::COMPONENT.to_string(),
                        registry.clone(),
                    ),
                ))
            } else {
                None
            };
374

375
        let mut router = axum::Router::new();
376

377
378
379
        let mut all_docs = Vec::new();

        let mut routes = vec![
380
381
382
383
            metrics::router(registry, var(HTTP_SVC_METRICS_PATH_ENV).ok()),
            super::openai::list_models_router(state.clone(), var(HTTP_SVC_MODELS_PATH_ENV).ok()),
            super::health::health_check_router(state.clone(), var(HTTP_SVC_HEALTH_PATH_ENV).ok()),
            super::health::live_check_router(state.clone(), var(HTTP_SVC_LIVE_PATH_ENV).ok()),
384
385
        ];

386
387
388
389
        let endpoint_routes =
            HttpServiceConfigBuilder::get_endpoints_router(state.clone(), &config.request_template);
        routes.extend(endpoint_routes);
        for (route_docs, route) in routes {
390
391
392
393
            router = router.merge(route);
            all_docs.extend(route_docs);
        }

394
395
396
397
398
399
400
        // Add OpenAPI documentation routes (must be after all other routes so it can document them)
        // Note: The path parameter is currently unused as SwaggerUi requires static paths
        let (openapi_docs, openapi_route) =
            super::openapi_docs::openapi_router(all_docs.clone(), None);
        router = router.merge(openapi_route);
        all_docs.extend(openapi_docs);

401
402
403
        // Add span for tracing
        router = router.layer(TraceLayer::new_for_http().make_span_with(make_request_span));

404
        Ok(HttpService {
405
            state,
406
407
            router,
            port: config.port,
408
            host: config.host,
Graham King's avatar
Graham King committed
409
410
411
            enable_tls: config.enable_tls,
            tls_cert_path: config.tls_cert_path,
            tls_key_path: config.tls_key_path,
412
            route_docs: all_docs,
413
414
415
416
            custom_backend_namespace_component_endpoint: config
                .custom_backend_namespace_component_endpoint,
            custom_backend_metrics_polling_interval: config.custom_backend_metrics_polling_interval,
            custom_backend_registry,
417
418
        })
    }
419
420
421
422
423

    pub fn with_request_template(mut self, request_template: Option<RequestTemplate>) -> Self {
        self.request_template = Some(request_template);
        self
    }
424
425
426
427
428

    pub fn with_etcd_client(mut self, etcd_client: Option<etcd::Client>) -> Self {
        self.etcd_client = Some(etcd_client);
        self
    }
429

430
431
432
433
434
435
436
437
438
439
440
    // DEPRECATED: To be removed after custom backends migrate to Dynamo backend.
    pub fn with_custom_backend_config(
        mut self,
        namespace_component_endpoint: Option<String>,
        polling_interval: Option<f64>,
    ) -> Self {
        self.custom_backend_namespace_component_endpoint = Some(namespace_component_endpoint);
        self.custom_backend_metrics_polling_interval = Some(polling_interval);
        self
    }

441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
    fn get_endpoints_router(
        state: Arc<State>,
        request_template: &Option<RequestTemplate>,
    ) -> Vec<(Vec<RouteDoc>, axum::Router)> {
        let mut routes = Vec::new();
        // Add chat completions route with conditional middleware
        let (chat_docs, chat_route) = super::openai::chat_completions_router(
            state.clone(),
            request_template.clone(),
            var(HTTP_SVC_CHAT_PATH_ENV).ok(),
        );
        let (cmpl_docs, cmpl_route) =
            super::openai::completions_router(state.clone(), var(HTTP_SVC_CMP_PATH_ENV).ok());
        let (embed_docs, embed_route) =
            super::openai::embeddings_router(state.clone(), var(HTTP_SVC_EMB_PATH_ENV).ok());
        let (responses_docs, responses_route) = super::openai::responses_router(
            state.clone(),
            request_template.clone(),
            var(HTTP_SVC_RESPONSES_PATH_ENV).ok(),
        );

        let mut endpoint_routes = HashMap::new();
        endpoint_routes.insert(EndpointType::Chat, (chat_docs, chat_route));
        endpoint_routes.insert(EndpointType::Completion, (cmpl_docs, cmpl_route));
        endpoint_routes.insert(EndpointType::Embedding, (embed_docs, embed_route));
        endpoint_routes.insert(EndpointType::Responses, (responses_docs, responses_route));

        for endpoint_type in EndpointType::all() {
            let state_route = state.clone();
            if !endpoint_routes.contains_key(&endpoint_type) {
                tracing::debug!("{} endpoints are disabled", endpoint_type.as_str());
                continue;
            }
            let (docs, route) = endpoint_routes.get(&endpoint_type).cloned().unwrap();
            let route = route.route_layer(axum::middleware::from_fn(
                move |req: axum::http::Request<axum::body::Body>, next: axum::middleware::Next| {
                    let state: Arc<State> = state_route.clone();
                    async move {
                        // Check if the endpoint is enabled
                        let enabled = state.flags.get(&endpoint_type);
                        if enabled {
                            Ok(next.run(req).await)
                        } else {
                            tracing::debug!("{} endpoints are disabled", endpoint_type.as_str());
                            Err(axum::http::StatusCode::SERVICE_UNAVAILABLE)
                        }
                    }
                },
            ));
            routes.push((docs, route));
        }
        routes
    }
494
}