server.rs 11.7 KB
Newer Older
1
2
3
4
5
6
7
8
9
10
11
// SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
// SPDX-License-Identifier: Apache-2.0

use std::collections::HashMap;
use std::sync::Arc;

use axum::extract::State;
use axum::http::StatusCode;
use axum::response::IntoResponse;
use axum::routing::{get, post};
use axum::{Json, Router};
12
13
#[cfg(feature = "metrics")]
use prometheus::Encoder;
14
15
use serde::{Deserialize, Serialize};

16
use crate::protocols::{LocalBlockHash, WorkerId, compute_block_hash_for_seq};
17

18
use super::registry::{IndexerKey, WorkerRegistry};
19
20
21

pub struct AppState {
    pub registry: WorkerRegistry,
22
23
    #[cfg(feature = "metrics")]
    pub prom_registry: prometheus::Registry,
24
25
26
27
}

fn default_tenant() -> String {
    "default".to_string()
28
29
30
31
32
33
}

#[derive(Deserialize)]
pub struct RegisterRequest {
    pub instance_id: WorkerId,
    pub endpoint: String,
34
35
36
37
    pub model_name: String,
    #[serde(default = "default_tenant")]
    pub tenant_id: String,
    pub block_size: u32,
38
39
    #[serde(default)]
    pub dp_rank: Option<u32>,
40
41
    #[serde(default)]
    pub replay_endpoint: Option<String>,
42
43
44
45
46
}

#[derive(Deserialize)]
pub struct UnregisterRequest {
    pub instance_id: WorkerId,
47
48
49
    pub model_name: String,
    #[serde(default)]
    pub tenant_id: Option<String>,
50
51
52
53
54
55
56
57
58
59
60
61
62
    #[serde(default)]
    pub dp_rank: Option<u32>,
}

#[derive(Serialize)]
struct WorkerInfo {
    instance_id: WorkerId,
    endpoints: HashMap<u32, String>,
}

#[derive(Deserialize)]
pub struct QueryRequest {
    pub token_ids: Vec<u32>,
63
64
65
    pub model_name: String,
    #[serde(default = "default_tenant")]
    pub tenant_id: String,
66
67
68
69
70
71
72
    #[serde(default)]
    pub lora_name: Option<String>,
}

#[derive(Deserialize)]
pub struct QueryByHashRequest {
    pub block_hashes: Vec<i64>,
73
74
75
    pub model_name: String,
    #[serde(default = "default_tenant")]
    pub tenant_id: String,
76
77
78
79
80
81
82
83
84
85
86
87
88
}

#[derive(Serialize)]
struct ScoreResponse {
    scores: HashMap<String, HashMap<String, u32>>,
    frequencies: Vec<usize>,
    tree_sizes: HashMap<String, HashMap<String, usize>>,
}

async fn register(
    State(state): State<Arc<AppState>>,
    Json(req): Json<RegisterRequest>,
) -> impl IntoResponse {
89
90
91
92
93
94
95
96
97
98
99
100
101
    match state
        .registry
        .register(
            req.instance_id,
            req.endpoint,
            req.dp_rank.unwrap_or(0),
            req.model_name,
            req.tenant_id,
            req.block_size,
            req.replay_endpoint,
        )
        .await
    {
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
        Ok(()) => (
            StatusCode::CREATED,
            Json(serde_json::json!({"status": "ok"})),
        ),
        Err(e) => (
            StatusCode::CONFLICT,
            Json(serde_json::json!({"error": e.to_string()})),
        ),
    }
}

async fn unregister(
    State(state): State<Arc<AppState>>,
    Json(req): Json<UnregisterRequest>,
) -> impl IntoResponse {
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
    let result = match req.tenant_id {
        Some(tenant_id) => match req.dp_rank {
            Some(dp_rank) => {
                state
                    .registry
                    .deregister_dp_rank(req.instance_id, dp_rank, &req.model_name, &tenant_id)
                    .await
            }
            None => {
                state
                    .registry
                    .deregister(req.instance_id, &req.model_name, &tenant_id)
                    .await
            }
        },
        None => {
133
134
            state
                .registry
135
                .deregister_all_tenants(req.instance_id, &req.model_name)
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
                .await
        }
    };
    match result {
        Ok(()) => (StatusCode::OK, Json(serde_json::json!({"status": "ok"}))),
        Err(e) => (
            StatusCode::NOT_FOUND,
            Json(serde_json::json!({"error": e.to_string()})),
        ),
    }
}

async fn list_workers(State(state): State<Arc<AppState>>) -> impl IntoResponse {
    let workers: Vec<WorkerInfo> = state
        .registry
        .list()
        .into_iter()
        .map(|(instance_id, endpoints)| WorkerInfo {
            instance_id,
            endpoints,
        })
        .collect();
    Json(workers)
}

161
fn build_score_response(
162
    overlap: crate::protocols::OverlapScores,
163
164
    block_size: u32,
) -> ScoreResponse {
165
166
167
168
169
    let mut scores: HashMap<String, HashMap<String, u32>> = HashMap::new();
    for (k, v) in &overlap.scores {
        scores
            .entry(k.worker_id.to_string())
            .or_default()
170
            .insert(k.dp_rank.to_string(), v * block_size);
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
    }
    let mut tree_sizes: HashMap<String, HashMap<String, usize>> = HashMap::new();
    for (k, v) in &overlap.tree_sizes {
        tree_sizes
            .entry(k.worker_id.to_string())
            .or_default()
            .insert(k.dp_rank.to_string(), *v);
    }
    ScoreResponse {
        scores,
        frequencies: overlap.frequencies,
        tree_sizes,
    }
}

async fn query(
    State(state): State<Arc<AppState>>,
    Json(req): Json<QueryRequest>,
) -> impl IntoResponse {
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
    let key = IndexerKey {
        model_name: req.model_name,
        tenant_id: req.tenant_id,
    };
    let Some(ie) = state.registry.get_indexer(&key) else {
        return (
            StatusCode::NOT_FOUND,
            Json(serde_json::json!({
                "error": format!("no indexer for model={} tenant={}", key.model_name, key.tenant_id)
            })),
        );
    };
    let block_size = ie.block_size;
    let indexer = ie.indexer.clone();
    drop(ie);

    let block_hashes =
        compute_block_hash_for_seq(&req.token_ids, block_size, None, req.lora_name.as_deref());
    match indexer.find_matches(block_hashes).await {
209
210
        Ok(overlap) => (
            StatusCode::OK,
211
            Json(serde_json::json!(build_score_response(overlap, block_size))),
212
213
214
215
216
217
218
219
220
221
222
223
        ),
        Err(e) => (
            StatusCode::INTERNAL_SERVER_ERROR,
            Json(serde_json::json!({"error": e.to_string()})),
        ),
    }
}

async fn query_by_hash(
    State(state): State<Arc<AppState>>,
    Json(req): Json<QueryByHashRequest>,
) -> impl IntoResponse {
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
    let key = IndexerKey {
        model_name: req.model_name,
        tenant_id: req.tenant_id,
    };
    let Some(ie) = state.registry.get_indexer(&key) else {
        return (
            StatusCode::NOT_FOUND,
            Json(serde_json::json!({
                "error": format!("no indexer for model={} tenant={}", key.model_name, key.tenant_id)
            })),
        );
    };
    let block_size = ie.block_size;
    let indexer = ie.indexer.clone();
    drop(ie);

240
241
242
243
244
    let block_hashes: Vec<LocalBlockHash> = req
        .block_hashes
        .iter()
        .map(|h| LocalBlockHash(*h as u64))
        .collect();
245
    match indexer.find_matches(block_hashes).await {
246
247
        Ok(overlap) => (
            StatusCode::OK,
248
            Json(serde_json::json!(build_score_response(overlap, block_size))),
249
250
251
252
253
254
255
256
        ),
        Err(e) => (
            StatusCode::INTERNAL_SERVER_ERROR,
            Json(serde_json::json!({"error": e.to_string()})),
        ),
    }
}

257
258
259
260
261
262
263
264
265
266
267
268
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
#[cfg(feature = "test-endpoints")]
#[derive(Deserialize)]
struct ListenerControlRequest {
    instance_id: WorkerId,
    #[serde(default)]
    dp_rank: Option<u32>,
}

#[cfg(feature = "test-endpoints")]
async fn test_pause_listener(
    State(state): State<Arc<AppState>>,
    Json(req): Json<ListenerControlRequest>,
) -> impl IntoResponse {
    match state
        .registry
        .pause_listener(req.instance_id, req.dp_rank.unwrap_or(0))
    {
        Ok(()) => (StatusCode::OK, Json(serde_json::json!({"status": "ok"}))),
        Err(e) => (
            StatusCode::NOT_FOUND,
            Json(serde_json::json!({"error": e.to_string()})),
        ),
    }
}

#[cfg(feature = "test-endpoints")]
async fn test_resume_listener(
    State(state): State<Arc<AppState>>,
    Json(req): Json<ListenerControlRequest>,
) -> impl IntoResponse {
    match state
        .registry
        .resume_listener(req.instance_id, req.dp_rank.unwrap_or(0))
        .await
    {
        Ok(()) => (StatusCode::OK, Json(serde_json::json!({"status": "ok"}))),
        Err(e) => (
            StatusCode::CONFLICT,
            Json(serde_json::json!({"error": e.to_string()})),
        ),
    }
}

300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
#[derive(Deserialize)]
struct PeerRequest {
    url: String,
}

async fn register_peer(
    State(state): State<Arc<AppState>>,
    Json(req): Json<PeerRequest>,
) -> impl IntoResponse {
    state.registry.register_peer(req.url);
    (
        StatusCode::CREATED,
        Json(serde_json::json!({"status": "ok"})),
    )
}

async fn deregister_peer(
    State(state): State<Arc<AppState>>,
    Json(req): Json<PeerRequest>,
) -> impl IntoResponse {
    if state.registry.deregister_peer(&req.url) {
        (StatusCode::OK, Json(serde_json::json!({"status": "ok"})))
    } else {
        (
            StatusCode::NOT_FOUND,
            Json(serde_json::json!({"error": "peer not found"})),
        )
    }
}

async fn list_peers(State(state): State<Arc<AppState>>) -> impl IntoResponse {
    Json(state.registry.list_peers())
}

334
async fn dump_events(State(state): State<Arc<AppState>>) -> impl IntoResponse {
335
336
    let all = state.registry.all_indexers_with_block_size();
    let mut handles = Vec::with_capacity(all.len());
337

338
    for (key, indexer, block_size) in all {
339
340
        handles.push(tokio::spawn(async move {
            let events = indexer.dump_events().await;
341
            (key, events, block_size)
342
343
344
345
346
347
        }));
    }

    let mut result: HashMap<String, serde_json::Value> = HashMap::new();
    for handle in handles {
        match handle.await {
348
            Ok((key, Ok(events), block_size)) => {
349
                let map_key = format!("{}:{}", key.model_name, key.tenant_id);
350
351
352
353
354
355
356
                result.insert(
                    map_key,
                    serde_json::json!({
                        "block_size": block_size,
                        "events": events,
                    }),
                );
357
            }
358
            Ok((key, Err(e), _)) => {
359
360
361
362
363
364
365
                let map_key = format!("{}:{}", key.model_name, key.tenant_id);
                result.insert(map_key, serde_json::json!({"error": e.to_string()}));
            }
            Err(e) => {
                tracing::warn!("dump task join error: {e}");
            }
        }
366
    }
367
    (StatusCode::OK, Json(serde_json::json!(result)))
368
369
}

370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
async fn handle_health() -> StatusCode {
    StatusCode::OK
}

#[cfg(feature = "metrics")]
async fn handle_metrics(State(state): State<Arc<AppState>>) -> impl IntoResponse {
    let encoder = prometheus::TextEncoder::new();
    let mut buf = Vec::new();
    encoder
        .encode(&state.prom_registry.gather(), &mut buf)
        .unwrap();
    (
        StatusCode::OK,
        [(
            axum::http::header::CONTENT_TYPE,
            prometheus::TEXT_FORMAT.to_string(),
        )],
        buf,
    )
}

391
pub fn create_router(state: Arc<AppState>) -> Router {
392
    let router = Router::new()
393
394
395
396
397
398
        .route("/register", post(register))
        .route("/unregister", post(unregister))
        .route("/workers", get(list_workers))
        .route("/query", post(query))
        .route("/query_by_hash", post(query_by_hash))
        .route("/dump", get(dump_events))
399
400
        .route("/register_peer", post(register_peer))
        .route("/deregister_peer", post(deregister_peer))
401
402
        .route("/peers", get(list_peers))
        .route("/health", get(handle_health));
403
404
405
406
407
408

    #[cfg(feature = "test-endpoints")]
    let router = router
        .route("/test/pause_listener", post(test_pause_listener))
        .route("/test/resume_listener", post(test_resume_listener));

409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
    let router = router.with_state(state.clone());

    #[cfg(feature = "metrics")]
    let router = {
        let metrics_route = Router::new()
            .route("/metrics", get(handle_metrics))
            .with_state(state);
        router
            .layer(axum::middleware::from_fn(
                super::metrics::metrics_middleware,
            ))
            .merge(metrics_route)
    };

    router
424
}