router_manager.rs 17.3 KB
Newer Older
1
2
3
4
5
6
//! Router Manager for coordinating multiple routers and workers
//!
//! Provides centralized management based on enable_igw flag:
//! - Single Router Mode (enable_igw=false): Router owns workers directly
//! - Multi-Router Mode (enable_igw=true): RouterManager coordinates everything

7
8
use crate::config::{ConnectionMode, RoutingMode};
use crate::core::{WorkerRegistry, WorkerType};
9
use crate::protocols::spec::{
10
11
    ChatCompletionRequest, CompletionRequest, EmbeddingRequest, GenerateRequest, RerankRequest,
    ResponsesRequest,
12
};
13
use crate::routers::RouterTrait;
14
use crate::server::{AppContext, ServerConfig};
15
16
17
18
19
20
21
22
23
use async_trait::async_trait;
use axum::{
    body::Body,
    extract::Request,
    http::{HeaderMap, StatusCode},
    response::{IntoResponse, Response},
};
use dashmap::DashMap;
use std::sync::Arc;
24
use tracing::{debug, info, warn};
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41

#[derive(Debug, Clone, Hash, Eq, PartialEq)]
pub struct RouterId(String);

impl RouterId {
    pub fn new(id: String) -> Self {
        Self(id)
    }

    pub fn as_str(&self) -> &str {
        &self.0
    }
}

pub struct RouterManager {
    worker_registry: Arc<WorkerRegistry>,
    routers: Arc<DashMap<RouterId, Arc<dyn RouterTrait>>>,
42
    default_router: Arc<std::sync::RwLock<Option<RouterId>>>,
43
    enable_igw: bool,
44
45
46
}

impl RouterManager {
47
    pub fn new(worker_registry: Arc<WorkerRegistry>) -> Self {
48
49
50
        Self {
            worker_registry,
            routers: Arc::new(DashMap::new()),
51
            default_router: Arc::new(std::sync::RwLock::new(None)),
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
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
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
            enable_igw: false, // Will be set properly in from_config
        }
    }

    pub async fn from_config(
        config: &ServerConfig,
        app_context: &Arc<AppContext>,
    ) -> Result<Arc<Self>, String> {
        use crate::routers::RouterFactory;

        let mut manager = Self::new(app_context.worker_registry.clone());
        manager.enable_igw = config.router_config.enable_igw;
        let manager = Arc::new(manager);

        if config.router_config.enable_igw {
            info!("Initializing RouterManager in multi-router mode (IGW)");

            match RouterFactory::create_regular_router(app_context).await {
                Ok(http_regular) => {
                    info!("Created HTTP Regular router");
                    manager.register_router(
                        RouterId::new("http-regular".to_string()),
                        Arc::from(http_regular),
                    );
                }
                Err(e) => {
                    warn!("Failed to create HTTP Regular router: {e}");
                }
            }

            match RouterFactory::create_pd_router(
                None,
                None,
                &config.router_config.policy,
                app_context,
            )
            .await
            {
                Ok(http_pd) => {
                    info!("Created HTTP PD router");
                    manager
                        .register_router(RouterId::new("http-pd".to_string()), Arc::from(http_pd));
                }
                Err(e) => {
                    warn!("Failed to create HTTP PD router: {e}");
                }
            }

            // TODO: Add gRPC routers once we have dynamic tokenizer loading

            info!(
                "RouterManager initialized with {} routers for multi-router mode",
                manager.router_count()
            );
        } else {
            info!("Initializing RouterManager in single-router mode");

            let single_router = Arc::from(RouterFactory::create_router(app_context).await?);
            let router_id = Self::determine_router_id(
                &config.router_config.mode,
                &config.router_config.connection_mode,
            );

            info!("Created single router with ID: {}", router_id.as_str());
            manager.register_router(router_id.clone(), single_router);
            manager.set_default_router(router_id);
        }

        if manager.router_count() == 0 {
            return Err("No routers could be initialized".to_string());
        }

        Ok(manager)
    }

    pub fn determine_router_id(
        routing_mode: &RoutingMode,
        connection_mode: &ConnectionMode,
    ) -> RouterId {
        match (connection_mode, routing_mode) {
            (ConnectionMode::Http, RoutingMode::Regular { .. }) => {
                RouterId::new("http-regular".to_string())
            }
            (ConnectionMode::Http, RoutingMode::PrefillDecode { .. }) => {
                RouterId::new("http-pd".to_string())
            }
            (ConnectionMode::Http, RoutingMode::OpenAI { .. }) => {
                RouterId::new("http-openai".to_string())
            }
            (ConnectionMode::Grpc, RoutingMode::Regular { .. }) => {
                RouterId::new("grpc-regular".to_string())
            }
            (ConnectionMode::Grpc, RoutingMode::PrefillDecode { .. }) => {
                RouterId::new("grpc-pd".to_string())
            }
            (ConnectionMode::Grpc, RoutingMode::OpenAI { .. }) => {
                RouterId::new("grpc-regular".to_string())
            }
150
151
152
        }
    }

153
    pub fn register_router(&self, id: RouterId, router: Arc<dyn RouterTrait>) {
154
155
        self.routers.insert(id.clone(), router);

156
157
158
        let mut default_router = self.default_router.write().unwrap();
        if default_router.is_none() {
            *default_router = Some(id.clone());
159
160
161
162
            info!("Set default router to {}", id.as_str());
        }
    }

163
164
165
    pub fn set_default_router(&self, id: RouterId) {
        let mut default_router = self.default_router.write().unwrap();
        *default_router = Some(id);
166
167
168
169
170
171
172
    }

    pub fn router_count(&self) -> usize {
        self.routers.len()
    }

    pub fn get_router_for_model(&self, model_id: &str) -> Option<Arc<dyn RouterTrait>> {
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
        let workers = self.worker_registry.get_by_model(model_id);

        if !workers.is_empty() {
            let has_pd_workers = workers.iter().any(|w| {
                matches!(
                    w.worker_type(),
                    WorkerType::Prefill { .. } | WorkerType::Decode
                )
            });

            let router_id = if has_pd_workers {
                RouterId::new("http-pd".to_string())
            } else {
                RouterId::new("http-regular".to_string())
            };

            if let Some(router) = self.routers.get(&router_id) {
                return Some(router.clone());
191
192
193
            }
        }

194
195
        let default_router = self.default_router.read().unwrap();
        if let Some(ref default_id) = *default_router {
196
197
198
199
200
201
202
203
204
205
206
            self.routers.get(default_id).map(|r| r.clone())
        } else {
            None
        }
    }

    pub fn select_router_for_request(
        &self,
        headers: Option<&HeaderMap>,
        model_id: Option<&str>,
    ) -> Option<Arc<dyn RouterTrait>> {
207
208
209
210
211
212
213
214
215
216
217
218
219
220
        // In single-router mode (enable_igw=false), always use the default router
        if !self.enable_igw {
            let default_router = self.default_router.read().unwrap();
            if let Some(ref default_id) = *default_router {
                debug!(
                    "Single-router mode: using default router {} for model {:?}",
                    default_id.as_str(),
                    model_id
                );
                return self.routers.get(default_id).map(|r| r.clone());
            }
        }

        // Multi-router mode logic follows
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
        let _priority_threshold = headers.and_then(|h| {
            h.get("x-worker-priority")
                .and_then(|v| v.to_str().ok())
                .and_then(|s| s.parse::<u32>().ok())
        });

        let _max_cost = headers.and_then(|h| {
            h.get("x-max-cost")
                .and_then(|v| v.to_str().ok())
                .and_then(|s| s.parse::<f32>().ok())
        });

        let prefer_pd = headers
            .and_then(|h| {
                h.get("x-prefer-pd")
                    .and_then(|v| v.to_str().ok())
                    .map(|s| s == "true" || s == "1")
            })
            .unwrap_or(false);

        let candidate_routers = if let Some(model) = model_id {
242
243
            if let Some(router) = self.get_router_for_model(model) {
                vec![router]
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
            } else {
                Vec::new()
            }
        } else {
            self.routers
                .iter()
                .map(|entry| entry.value().clone())
                .collect::<Vec<_>>()
        };

        if candidate_routers.is_empty() {
            return None;
        }

        let mut best_router = None;
        let mut best_score = 0.0;

        for router in candidate_routers {
            let mut score = 1.0;

            let is_pd = router.is_pd_mode();
            if prefer_pd && is_pd {
266
                score += 2.0;
267
            } else if !prefer_pd && !is_pd {
268
                score += 1.0;
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
            }

            // TODO: Once routers expose worker stats, we can evaluate:
            // - Average worker priority vs priority_threshold
            // - Average worker cost vs max_cost
            // - Current load and health status

            if score > best_score {
                best_score = score;
                best_router = Some(router);
            }
        }

        best_router
    }
}

#[async_trait]
impl RouterTrait for RouterManager {
    fn as_any(&self) -> &dyn std::any::Any {
        self
    }

    async fn health(&self, _req: Request<Body>) -> Response {
        (StatusCode::OK, "RouterManager is healthy").into_response()
    }

    async fn health_generate(&self, _req: Request<Body>) -> Response {
        // TODO: Should check if any router has healthy workers
        (
            StatusCode::SERVICE_UNAVAILABLE,
            "No routers with healthy workers available",
        )
            .into_response()
    }

    async fn get_server_info(&self, _req: Request<Body>) -> Response {
        // TODO: Aggregate info from all routers with healthy workers
        (
            StatusCode::OK,
            serde_json::json!({
                "router_manager": true,
                "routers_count": self.routers.len(),
                "workers_count": self.worker_registry.get_all().len()
            })
            .to_string(),
        )
            .into_response()
    }

    async fn get_models(&self, _req: Request<Body>) -> Response {
320
        let models = self.worker_registry.get_models();
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
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
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409

        if models.is_empty() {
            (StatusCode::SERVICE_UNAVAILABLE, "No models available").into_response()
        } else {
            (
                StatusCode::OK,
                serde_json::json!({
                    "models": models
                })
                .to_string(),
            )
                .into_response()
        }
    }

    async fn get_model_info(&self, _req: Request<Body>) -> Response {
        // TODO: Extract model from request and route to appropriate router
        (
            StatusCode::NOT_IMPLEMENTED,
            "Model info endpoint not yet implemented in RouterManager",
        )
            .into_response()
    }

    async fn route_generate(
        &self,
        headers: Option<&HeaderMap>,
        body: &GenerateRequest,
        _model_id: Option<&str>,
    ) -> Response {
        let router = self.select_router_for_request(headers, None);

        if let Some(router) = router {
            router.route_generate(headers, body, None).await
        } else {
            (
                StatusCode::NOT_FOUND,
                "No router available for this request",
            )
                .into_response()
        }
    }

    async fn route_chat(
        &self,
        headers: Option<&HeaderMap>,
        body: &ChatCompletionRequest,
        _model_id: Option<&str>,
    ) -> Response {
        let router = self.select_router_for_request(headers, Some(&body.model));

        if let Some(router) = router {
            router.route_chat(headers, body, Some(&body.model)).await
        } else {
            (
                StatusCode::NOT_FOUND,
                format!("Model '{}' not found or no router available", body.model),
            )
                .into_response()
        }
    }

    async fn route_completion(
        &self,
        headers: Option<&HeaderMap>,
        body: &CompletionRequest,
        _model_id: Option<&str>,
    ) -> Response {
        let router = self.select_router_for_request(headers, Some(&body.model));

        if let Some(router) = router {
            router
                .route_completion(headers, body, Some(&body.model))
                .await
        } else {
            (
                StatusCode::NOT_FOUND,
                format!("Model '{}' not found or no router available", body.model),
            )
                .into_response()
        }
    }

    async fn route_responses(
        &self,
        _headers: Option<&HeaderMap>,
        _body: &ResponsesRequest,
        _model_id: Option<&str>,
    ) -> Response {
410
411
412
413
414
415
416
        (
            StatusCode::NOT_IMPLEMENTED,
            "responses api not yet implemented in inference gateway mode",
        )
            .into_response()
    }

417
418
419
420
421
422
423
424
425
426
427
428
429
430
    async fn get_response(&self, headers: Option<&HeaderMap>, response_id: &str) -> Response {
        let router = self.select_router_for_request(headers, None);
        if let Some(router) = router {
            router.get_response(headers, response_id).await
        } else {
            (
                StatusCode::NOT_FOUND,
                format!("No router available to get response '{}'", response_id),
            )
                .into_response()
        }
    }

    async fn cancel_response(&self, headers: Option<&HeaderMap>, response_id: &str) -> Response {
431
        let router = self.select_router_for_request(headers, None);
432
433
434
435
436
437
438
439
440
441
442
        if let Some(router) = router {
            router.cancel_response(headers, response_id).await
        } else {
            (
                StatusCode::NOT_FOUND,
                format!("No router available to cancel response '{}'", response_id),
            )
                .into_response()
        }
    }

443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
    async fn delete_response(&self, _headers: Option<&HeaderMap>, _response_id: &str) -> Response {
        (
            StatusCode::NOT_IMPLEMENTED,
            "responses api not yet implemented in inference gateway mode",
        )
            .into_response()
    }

    async fn list_response_input_items(
        &self,
        _headers: Option<&HeaderMap>,
        _response_id: &str,
    ) -> Response {
        (
            StatusCode::NOT_IMPLEMENTED,
            "responses api not yet implemented in inference gateway mode",
        )
            .into_response()
    }

463
464
465
466
467
468
469
    async fn route_embeddings(
        &self,
        headers: Option<&HeaderMap>,
        body: &EmbeddingRequest,
        _model_id: Option<&str>,
    ) -> Response {
        let router = self.select_router_for_request(headers, Some(&body.model));
470
471

        if let Some(router) = router {
472
473
474
            router
                .route_embeddings(headers, body, Some(&body.model))
                .await
475
476
477
        } else {
            (
                StatusCode::NOT_FOUND,
478
                format!("Model '{}' not found or no router available", body.model),
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
            )
                .into_response()
        }
    }

    async fn route_rerank(
        &self,
        headers: Option<&HeaderMap>,
        body: &RerankRequest,
        model_id: Option<&str>,
    ) -> Response {
        let router = self.select_router_for_request(headers, None);

        if let Some(router) = router {
            router.route_rerank(headers, body, model_id).await
        } else {
            (
                StatusCode::NOT_FOUND,
                "No router available for rerank request",
            )
                .into_response()
        }
    }

    async fn flush_cache(&self) -> Response {
        // TODO: Call flush_cache on all routers that have workers
        if self.routers.is_empty() {
            (StatusCode::SERVICE_UNAVAILABLE, "No routers configured").into_response()
        } else {
            // TODO: Actually flush cache on all routers
            (StatusCode::OK, "Cache flush requested").into_response()
        }
    }

    async fn get_worker_loads(&self) -> Response {
        let workers = self.worker_registry.get_all();
        let loads: Vec<serde_json::Value> = workers
            .iter()
            .map(|w| {
                serde_json::json!({
                    "url": w.url(),
                    "model": w.model_id(),
                    "load": w.load(),
                    "is_healthy": w.is_healthy()
                })
            })
            .collect();

        (
            StatusCode::OK,
            serde_json::json!({
                "workers": loads
            })
            .to_string(),
        )
            .into_response()
    }

    fn router_type(&self) -> &'static str {
        "manager"
    }

    fn readiness(&self) -> Response {
        if self.routers.is_empty() {
            (StatusCode::SERVICE_UNAVAILABLE, "No routers configured").into_response()
        } else {
            // TODO: Check readiness of all routers
            (StatusCode::OK, "Ready").into_response()
        }
    }
}

impl std::fmt::Debug for RouterManager {
    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
        f.debug_struct("RouterManager")
            .field("routers_count", &self.routers.len())
            .field("workers_count", &self.worker_registry.get_all().len())
556
            .field("default_router", &*self.default_router.read().unwrap())
557
558
559
            .finish()
    }
}