service_v2.rs 17 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
use dynamo_runtime::storage::key_value_store::KeyValueStoreManager;
Graham King's avatar
Graham King committed
24
use std::net::SocketAddr;
25
use tokio::task::JoinHandle;
26
use tokio_util::sync::CancellationToken;
27
use tower_http::trace::TraceLayer;
28

29
30
31
32
/// HTTP service shared state
pub struct State {
    metrics: Arc<Metrics>,
    manager: Arc<ModelManager>,
33
    store: KeyValueStoreManager,
34
35
36
37
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
    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),
        }
    }
71
72
73
}

impl State {
74
    pub fn new(manager: Arc<ModelManager>, store: KeyValueStoreManager) -> Self {
75
76
77
        Self {
            manager,
            metrics: Arc::new(Metrics::default()),
78
            store,
79
80
81
82
83
84
            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),
            },
85
86
87
        }
    }

88
89
90
91
92
93
94
95
96
97
98
99
100
    /// 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()
    }

101
102
    pub fn store(&self) -> &KeyValueStoreManager {
        &self.store
103
104
    }

105
106
107
108
109
110
    // TODO
    pub fn sse_keep_alive(&self) -> Option<Duration> {
        None
    }
}

111
112
#[derive(Clone)]
pub struct HttpService {
113
114
115
    // The state we share with every request handler
    state: Arc<State>,

116
117
    router: axum::Router,
    port: u16,
118
    host: String,
Graham King's avatar
Graham King committed
119
120
121
    enable_tls: bool,
    tls_cert_path: Option<PathBuf>,
    tls_key_path: Option<PathBuf>,
122
    route_docs: Vec<RouteDoc>,
123
124
125
126
127
128

    // 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>>,
129
130
131
}

#[derive(Clone, Builder)]
132
#[builder(pattern = "owned", build_fn(private, name = "build_internal"))]
133
134
135
136
pub struct HttpServiceConfig {
    #[builder(default = "8787")]
    port: u16,

137
138
139
    #[builder(setter(into), default = "String::from(\"0.0.0.0\")")]
    host: String,

Graham King's avatar
Graham King committed
140
141
142
143
144
145
146
147
148
    #[builder(default = "false")]
    enable_tls: bool,

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

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

149
150
    // #[builder(default)]
    // custom: Vec<axum::Router>
151
    #[builder(default = "false")]
152
153
    enable_chat_endpoints: bool,

154
    #[builder(default = "false")]
155
    enable_cmpl_endpoints: bool,
156

157
    #[builder(default = "true")]
158
159
    enable_embeddings_endpoints: bool,

160
161
162
    #[builder(default = "true")]
    enable_responses_endpoints: bool,

163
164
    #[builder(default = "None")]
    request_template: Option<RequestTemplate>,
165

166
167
    #[builder(default)]
    store: KeyValueStoreManager,
168
169
170
171
172
173
174

    // 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>,
175
176
177
178
179
180
181
}

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

182
183
184
185
186
187
188
189
    pub fn state_clone(&self) -> Arc<State> {
        self.state.clone()
    }

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

190
    pub fn model_manager(&self) -> &ModelManager {
191
        self.state().manager()
192
193
    }

194
195
196
197
198
199
    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<()> {
200
        let address = format!("{}:{}", self.host, self.port);
Graham King's avatar
Graham King committed
201
202
        let protocol = if self.enable_tls { "HTTPS" } else { "HTTP" };
        tracing::info!(protocol, address, "Starting HTTP(S) service");
203
204
205
206

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

Graham King's avatar
Graham King committed
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
        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 {
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
            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
268
269
270
271
272
273

            axum::serve(listener, router)
                .with_graceful_shutdown(observer.cancelled_owned())
                .await
                .inspect_err(|_| cancel_token.cancel())?;
        }
274
275

        Ok(())
276
    }
277
278
279
280
281

    /// Documentation of exposed HTTP endpoints
    pub fn route_docs(&self) -> &[RouteDoc] {
        &self.route_docs
    }
282

283
    pub fn enable_model_endpoint(&self, endpoint_type: EndpointType, enable: bool) {
284
285
286
287
288
289
290
        self.state.flags.set(&endpoint_type, enable);
        tracing::info!(
            "{} endpoints {}",
            endpoint_type.as_str(),
            if enable { "enabled" } else { "disabled" }
        );
    }
291
292
}

293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
/// 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";

310
311
impl HttpServiceConfigBuilder {
    pub fn build(self) -> Result<HttpService, anyhow::Error> {
312
        let config: HttpServiceConfig = self.build_internal()?;
313

314
        let model_manager = Arc::new(ModelManager::new());
315
        let state = Arc::new(State::new(model_manager, config.store));
316
317
318
319
320
321
322
323
324
325
326
327
328
        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);

329
330
        // enable prometheus metrics
        let registry = metrics::Registry::new();
331
        state.metrics_clone().register(&registry)?;
332

333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
        // 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
            };
348

349
        let mut router = axum::Router::new();
350

351
352
353
        let mut all_docs = Vec::new();

        let mut routes = vec![
354
355
356
357
            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()),
358
359
        ];

360
361
362
363
        let endpoint_routes =
            HttpServiceConfigBuilder::get_endpoints_router(state.clone(), &config.request_template);
        routes.extend(endpoint_routes);
        for (route_docs, route) in routes {
364
365
366
367
            router = router.merge(route);
            all_docs.extend(route_docs);
        }

368
369
370
371
372
373
374
        // 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);

375
376
377
        // Add span for tracing
        router = router.layer(TraceLayer::new_for_http().make_span_with(make_request_span));

378
        Ok(HttpService {
379
            state,
380
381
            router,
            port: config.port,
382
            host: config.host,
Graham King's avatar
Graham King committed
383
384
385
            enable_tls: config.enable_tls,
            tls_cert_path: config.tls_cert_path,
            tls_key_path: config.tls_key_path,
386
            route_docs: all_docs,
387
388
389
390
            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,
391
392
        })
    }
393
394
395
396
397

    pub fn with_request_template(mut self, request_template: Option<RequestTemplate>) -> Self {
        self.request_template = Some(request_template);
        self
    }
398

399
400
401
402
403
404
405
406
407
408
409
    // 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
    }

410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
    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
    }
463
}