model_manager.rs 18.9 KB
Newer Older
1
2
3
// SPDX-FileCopyrightText: Copyright (c) 2024-2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
// SPDX-License-Identifier: Apache-2.0

4
5
use std::{
    collections::{HashMap, HashSet},
6
    sync::Arc,
7
8
};

9
use parking_lot::{Mutex, RwLock};
10
use tokio::sync::oneshot;
11

12
13
14
15
16
17
18
use dynamo_runtime::{
    component::{Endpoint, TransportType},
    discovery::DiscoverySpec,
    prelude::DistributedRuntimeProvider,
    protocols::EndpointId,
    transports::nats,
};
19
20

use crate::{
21
    kv_router::{KvRouter, KvRouterConfig, router_endpoint_id, scheduler::DefaultWorkerSelector},
22
    model_card::ModelDeploymentCard,
23
    model_type::ModelType,
24
25
26
27
28
29
30
31
    types::{
        generic::tensor::TensorStreamingEngine,
        openai::{
            chat_completions::OpenAIChatCompletionsStreamingEngine,
            completions::OpenAICompletionsStreamingEngine,
            embeddings::OpenAIEmbeddingsStreamingEngine,
        },
    },
32
};
33

34
35
36
37
38
39
40
41
/// State for prefill router activation rendezvous
enum PrefillActivationState {
    /// Decode model registered, waiting for prefill endpoint
    DecodeWaiting(oneshot::Sender<Endpoint>),
    /// Prefill endpoint arrived, waiting for decode model to register
    PrefillReady(oneshot::Receiver<Endpoint>),
}

42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
#[derive(Debug, thiserror::Error)]
pub enum ModelManagerError {
    #[error("Model not found: {0}")]
    ModelNotFound(String),

    #[error("Model already exists: {0}")]
    ModelAlreadyExists(String),
}

// Don't implement Clone for this, put it in an Arc instead.
pub struct ModelManager {
    // We read a lot and write rarely, so these three are RwLock
    completion_engines: RwLock<ModelEngines<OpenAICompletionsStreamingEngine>>,
    chat_completion_engines: RwLock<ModelEngines<OpenAIChatCompletionsStreamingEngine>>,
    embeddings_engines: RwLock<ModelEngines<OpenAIEmbeddingsStreamingEngine>>,
57
    tensor_engines: RwLock<ModelEngines<TensorStreamingEngine>>,
58
59
    // Prefill models don't have engines - they're only tracked for discovery/lifecycle
    prefill_engines: RwLock<ModelEngines<()>>,
60

61
    // These are Mutex because we read and write rarely and equally
62
    cards: Mutex<HashMap<String, ModelDeploymentCard>>,
63
    kv_choosers: Mutex<HashMap<EndpointId, Arc<KvRouter>>>,
64
    prefill_router_activators: Mutex<HashMap<String, PrefillActivationState>>,
65
66
67
68
69
70
71
72
73
74
75
76
77
78
}

impl Default for ModelManager {
    fn default() -> Self {
        Self::new()
    }
}

impl ModelManager {
    pub fn new() -> Self {
        Self {
            completion_engines: RwLock::new(ModelEngines::default()),
            chat_completion_engines: RwLock::new(ModelEngines::default()),
            embeddings_engines: RwLock::new(ModelEngines::default()),
79
            tensor_engines: RwLock::new(ModelEngines::default()),
80
            prefill_engines: RwLock::new(ModelEngines::default()),
81
            cards: Mutex::new(HashMap::new()),
82
            kv_choosers: Mutex::new(HashMap::new()),
83
            prefill_router_activators: Mutex::new(HashMap::new()),
84
85
86
        }
    }

87
88
89
90
91
92
93
94
95
96
97
98
99
    pub fn is_valid_checksum(
        &self,
        model_type: ModelType,
        model_name: &str,
        candidate_checksum: &str,
    ) -> Option<bool> {
        let mut results = vec![];
        for unit in model_type.units() {
            let maybe_valid_checksum = match unit {
                ModelType::Chat => self.chat_completion_engines.read().checksum(model_name),
                ModelType::Completions => self.completion_engines.read().checksum(model_name),
                ModelType::Embedding => self.embeddings_engines.read().checksum(model_name),
                ModelType::TensorBased => self.tensor_engines.read().checksum(model_name),
100
                ModelType::Prefill => self.prefill_engines.read().checksum(model_name),
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
                _ => {
                    continue;
                }
            };
            if let Some(is_valid) = maybe_valid_checksum.map(|valid_checksum| {
                tracing::debug!(
                    model_name,
                    valid_checksum,
                    candidate_checksum,
                    "is_valid_checksum: check case"
                );
                valid_checksum == candidate_checksum
            }) {
                results.push(is_valid)
            }
        }
        if results.is_empty() {
            None
        } else {
            // The checksum is valid if it is correct for all the ModelType in the bitflag.
            Some(results.into_iter().all(|x| x))
        }
    }

125
126
    pub fn get_model_cards(&self) -> Vec<ModelDeploymentCard> {
        self.cards.lock().values().cloned().collect()
127
128
    }

129
    pub fn has_model_any(&self, model: &str) -> bool {
130
131
        self.chat_completion_engines.read().contains(model)
            || self.completion_engines.read().contains(model)
132
133
    }

134
135
136
137
138
    pub fn model_display_names(&self) -> HashSet<String> {
        self.list_chat_completions_models()
            .into_iter()
            .chain(self.list_completions_models())
            .chain(self.list_embeddings_models())
139
            .chain(self.list_tensor_models())
140
            .chain(self.list_prefill_models())
141
142
143
            .collect()
    }

144
    pub fn list_chat_completions_models(&self) -> Vec<String> {
145
        self.chat_completion_engines.read().list()
146
147
148
    }

    pub fn list_completions_models(&self) -> Vec<String> {
149
        self.completion_engines.read().list()
150
151
152
    }

    pub fn list_embeddings_models(&self) -> Vec<String> {
153
        self.embeddings_engines.read().list()
154
155
    }

156
157
158
159
    pub fn list_tensor_models(&self) -> Vec<String> {
        self.tensor_engines.read().list()
    }

160
161
162
163
    pub fn list_prefill_models(&self) -> Vec<String> {
        self.prefill_engines.read().list()
    }

164
165
166
    pub fn add_completions_model(
        &self,
        model: &str,
167
        card_checksum: &str,
168
169
        engine: OpenAICompletionsStreamingEngine,
    ) -> Result<(), ModelManagerError> {
170
        let mut clients = self.completion_engines.write();
171
        clients.add(model, card_checksum, engine)
172
173
174
175
176
    }

    pub fn add_chat_completions_model(
        &self,
        model: &str,
177
        card_checksum: &str,
178
179
        engine: OpenAIChatCompletionsStreamingEngine,
    ) -> Result<(), ModelManagerError> {
180
        let mut clients = self.chat_completion_engines.write();
181
        clients.add(model, card_checksum, engine)
182
183
184
185
186
    }

    pub fn add_embeddings_model(
        &self,
        model: &str,
187
        card_checksum: &str,
188
189
        engine: OpenAIEmbeddingsStreamingEngine,
    ) -> Result<(), ModelManagerError> {
190
        let mut clients = self.embeddings_engines.write();
191
        clients.add(model, card_checksum, engine)
192
193
    }

194
195
196
    pub fn add_tensor_model(
        &self,
        model: &str,
197
        card_checksum: &str,
198
199
200
        engine: TensorStreamingEngine,
    ) -> Result<(), ModelManagerError> {
        let mut clients = self.tensor_engines.write();
201
        clients.add(model, card_checksum, engine)
202
203
    }

204
205
206
207
208
209
    pub fn add_prefill_model(
        &self,
        model: &str,
        card_checksum: &str,
    ) -> Result<(), ModelManagerError> {
        let mut clients = self.prefill_engines.write();
210
        clients.add(model, card_checksum, ())
211
212
    }

213
    pub fn remove_completions_model(&self, model: &str) -> Result<(), ModelManagerError> {
214
        let mut clients = self.completion_engines.write();
215
216
217
218
        clients.remove(model)
    }

    pub fn remove_chat_completions_model(&self, model: &str) -> Result<(), ModelManagerError> {
219
        let mut clients = self.chat_completion_engines.write();
220
221
222
223
        clients.remove(model)
    }

    pub fn remove_embeddings_model(&self, model: &str) -> Result<(), ModelManagerError> {
224
        let mut clients = self.embeddings_engines.write();
225
226
227
        clients.remove(model)
    }

228
229
230
231
232
    pub fn remove_tensor_model(&self, model: &str) -> Result<(), ModelManagerError> {
        let mut clients = self.tensor_engines.write();
        clients.remove(model)
    }

233
234
235
236
237
    pub fn remove_prefill_model(&self, model: &str) -> Result<(), ModelManagerError> {
        let mut clients = self.prefill_engines.write();
        clients.remove(model)
    }

238
    pub fn get_embeddings_engine(
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
270
        &self,
        model: &str,
    ) -> Result<OpenAIEmbeddingsStreamingEngine, ModelManagerError> {
        self.embeddings_engines
            .read()
            .get(model)
            .cloned()
            .ok_or(ModelManagerError::ModelNotFound(model.to_string()))
    }

    pub fn get_completions_engine(
        &self,
        model: &str,
    ) -> Result<OpenAICompletionsStreamingEngine, ModelManagerError> {
        self.completion_engines
            .read()
            .get(model)
            .cloned()
            .ok_or(ModelManagerError::ModelNotFound(model.to_string()))
    }

    pub fn get_chat_completions_engine(
        &self,
        model: &str,
    ) -> Result<OpenAIChatCompletionsStreamingEngine, ModelManagerError> {
        self.chat_completion_engines
            .read()
            .get(model)
            .cloned()
            .ok_or(ModelManagerError::ModelNotFound(model.to_string()))
    }

271
272
273
274
275
276
277
278
279
280
281
    pub fn get_tensor_engine(
        &self,
        model: &str,
    ) -> Result<TensorStreamingEngine, ModelManagerError> {
        self.tensor_engines
            .read()
            .get(model)
            .cloned()
            .ok_or(ModelManagerError::ModelNotFound(model.to_string()))
    }

282
283
284
285
286
    /// Save a ModelDeploymentCard from an instance's ModelDeploymentCard key so we can fetch it later when the key is
    /// deleted.
    pub fn save_model_card(&self, key: &str, card: ModelDeploymentCard) -> anyhow::Result<()> {
        self.cards.lock().insert(key.to_string(), card);
        Ok(())
287
288
    }

289
290
291
    /// Remove and return model card for this instance's etcd key. We do this when the instance stops.
    pub fn remove_model_card(&self, key: &str) -> Option<ModelDeploymentCard> {
        self.cards.lock().remove(key)
292
293
294
295
    }

    pub async fn kv_chooser_for(
        &self,
296
        endpoint: &Endpoint,
297
        kv_cache_block_size: u32,
298
        kv_router_config: Option<KvRouterConfig>,
299
    ) -> anyhow::Result<Arc<KvRouter>> {
300
        let endpoint_id = endpoint.id();
301

302
        if let Some(kv_chooser) = self.get_kv_chooser(&endpoint_id) {
303
304
305
            // Check if the existing router has a different block size
            if kv_chooser.block_size() != kv_cache_block_size {
                tracing::warn!(
306
                    endpoint = %endpoint_id,
307
308
                    existing_block_size = %kv_chooser.block_size(),
                    requested_block_size = %kv_cache_block_size,
309
                    "KV Router block size mismatch! Endpoint is requesting a different kv_cache_block_size than the existing router. \
310
311
312
                     This will cause routing to fail silently. Consider using the same block size or restarting the router."
                );
            }
313
314
315
            return Ok(kv_chooser);
        }

316
        let client = endpoint.client().await?;
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338

        // Register router via discovery mechanism
        let discovery = endpoint.component().drt().discovery();
        let instance_id = discovery.instance_id();

        // Build NATS transport subject for the router endpoint
        // Use KV_ROUTER_COMPONENT as the component name to distinguish from the generate endpoint's component
        let router_endpoint_id = router_endpoint_id(endpoint.id().namespace);
        // Placeholder subject - router is not callable, only registered for lifecycle coordination
        let nats_subject = nats::instance_subject(&router_endpoint_id, instance_id);

        let discovery_spec = DiscoverySpec::Endpoint {
            namespace: router_endpoint_id.namespace.clone(),
            component: router_endpoint_id.component.clone(),
            endpoint: router_endpoint_id.name.clone(),
            transport: TransportType::Nats(nats_subject),
        };

        discovery.register(discovery_spec).await?;

        // Use instance_id (hex) as the consumer ID for NATS consumer coordination
        let consumer_id = instance_id.to_string();
339

340
        let selector = Box::new(DefaultWorkerSelector::new(kv_router_config));
341
        let chooser = KvRouter::new(
342
343
            endpoint.clone(),
            client,
344
345
            kv_cache_block_size,
            Some(selector),
346
            kv_router_config,
347
            consumer_id,
348
349
        )
        .await?;
350
351
352
        let new_kv_chooser = Arc::new(chooser);
        self.kv_choosers
            .lock()
353
            .insert(endpoint_id, new_kv_chooser.clone());
354
355
        Ok(new_kv_chooser)
    }
356

357
358
    fn get_kv_chooser(&self, id: &EndpointId) -> Option<Arc<KvRouter>> {
        self.kv_choosers.lock().get(id).cloned()
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
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
    }

    /// Register a prefill router for a decode model. Returns a receiver that will be
    /// activated when the corresponding prefill model is discovered.
    /// Returns None if the decode model was already registered.
    pub fn register_prefill_router(
        &self,
        model_name: String,
    ) -> Option<oneshot::Receiver<Endpoint>> {
        let mut activators = self.prefill_router_activators.lock();

        match activators.remove(&model_name) {
            Some(PrefillActivationState::PrefillReady(rx)) => {
                // Prefill endpoint already arrived - rx will immediately resolve
                tracing::debug!(
                    model_name = %model_name,
                    "Prefill endpoint already available, returning receiver with endpoint"
                );
                Some(rx)
            }
            Some(PrefillActivationState::DecodeWaiting(tx)) => {
                // Decode already registered - this shouldn't happen, restore state and return None
                tracing::error!(
                    model_name = %model_name,
                    "Decode model already registered for this prefill router"
                );
                activators.insert(model_name, PrefillActivationState::DecodeWaiting(tx));
                None
            }
            None => {
                // New registration: create tx/rx pair, store sender and return receiver
                let (tx, rx) = oneshot::channel();
                activators.insert(
                    model_name.clone(),
                    PrefillActivationState::DecodeWaiting(tx),
                );
                tracing::debug!(
                    model_name = %model_name,
                    "No prefill endpoint available yet, storing sender for future activation"
                );
                Some(rx)
            }
        }
    }

    /// Activate a prefill router by sending the endpoint through the oneshot channel.
    /// If no decode model has registered yet, stores the endpoint for future retrieval.
    pub fn activate_prefill_router(
        &self,
        model_name: &str,
        endpoint: Endpoint,
    ) -> anyhow::Result<()> {
        let mut activators = self.prefill_router_activators.lock();

        match activators.remove(model_name) {
            Some(PrefillActivationState::DecodeWaiting(sender)) => {
                // Decode model already registered
                sender.send(endpoint).map_err(|_| {
                    anyhow::anyhow!(
                        "Failed to send endpoint to prefill router activator for model: {}",
                        model_name
                    )
                })?;

                tracing::info!(
                    model_name = %model_name,
                    "Activated prefill router for already-registered decode model"
                );

                Ok(())
            }
            Some(PrefillActivationState::PrefillReady(_)) => {
                // Prefill already activated - this shouldn't happen
                anyhow::bail!("Prefill router for model {} already activated", model_name);
            }
            None => {
                // Decode model not registered yet - create pair and immediately send endpoint
                let (tx, rx) = oneshot::channel();

                tx.send(endpoint).map_err(|_| {
                    anyhow::anyhow!("Failed to send endpoint for prefill model: {}", model_name)
                })?;

                // Store the receiver for when decode model registers
                activators.insert(
                    model_name.to_string(),
                    PrefillActivationState::PrefillReady(rx),
                );

                tracing::info!(
                    model_name = %model_name,
                    "Stored prefill endpoint for future decode model registration"
                );

                Ok(())
            }
        }
456
457
    }

458
    pub fn get_model_tool_call_parser(&self, model: &str) -> Option<String> {
459
        self.cards
460
461
            .lock()
            .values()
462
463
            .find(|c| c.display_name == model)
            .and_then(|c| c.runtime_config.tool_call_parser.as_ref())
464
            .map(|parser| parser.to_string())
465
    }
466
467
468
469
470
471
472
473
474

    /// Creates parsing options with tool call parser and reasoning parser for the specified model.
    /// Currently reasoning parser is not implemented (returns None).
    pub fn get_parsing_options(&self, model: &str) -> crate::protocols::openai::ParsingOptions {
        let tool_call_parser = self.get_model_tool_call_parser(model);
        let reasoning_parser = None; // TODO: Implement reasoning parser

        crate::protocols::openai::ParsingOptions::new(tool_call_parser, reasoning_parser)
    }
475
476
477
478
479
480
}

pub struct ModelEngines<E> {
    /// Optional default model name
    default: Option<String>,
    engines: HashMap<String, E>,
481
482
483
    /// Key: Model name, value: Checksum of the ModelDeploymentCard. New instances must have the
    /// same card.
    checksums: HashMap<String, String>,
484
485
486
487
488
489
490
}

impl<E> Default for ModelEngines<E> {
    fn default() -> Self {
        Self {
            default: None,
            engines: HashMap::new(),
491
            checksums: HashMap::new(),
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
        }
    }
}

impl<E> ModelEngines<E> {
    #[allow(dead_code)]
    fn set_default(&mut self, model: &str) {
        self.default = Some(model.to_string());
    }

    #[allow(dead_code)]
    fn clear_default(&mut self) {
        self.default = None;
    }

507
    fn add(&mut self, model: &str, checksum: &str, engine: E) -> Result<(), ModelManagerError> {
508
509
510
511
        if self.engines.contains_key(model) {
            return Err(ModelManagerError::ModelAlreadyExists(model.to_string()));
        }
        self.engines.insert(model.to_string(), engine);
512
513
        self.checksums
            .insert(model.to_string(), checksum.to_string());
514
515
516
517
518
519
520
        Ok(())
    }

    fn remove(&mut self, model: &str) -> Result<(), ModelManagerError> {
        if self.engines.remove(model).is_none() {
            return Err(ModelManagerError::ModelNotFound(model.to_string()));
        }
521
        let _ = self.checksums.remove(model);
522
523
524
525
526
527
528
529
530
531
532
533
534
535
        Ok(())
    }

    fn get(&self, model: &str) -> Option<&E> {
        self.engines.get(model)
    }

    fn contains(&self, model: &str) -> bool {
        self.engines.contains_key(model)
    }

    pub fn list(&self) -> Vec<String> {
        self.engines.keys().map(|k| k.to_owned()).collect()
    }
536
537
538
539
540
541

    /// Returns a newly allocated String for called convenience. All the places I use
    /// this I need a String.
    pub fn checksum(&self, model: &str) -> Option<String> {
        self.checksums.get(model).map(|s| s.to_string())
    }
542
}