server.rs 12.1 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::{BlockHashOptions, LocalBlockHash, WorkerId, compute_block_hash_for_seq};
17

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

pub struct AppState {
21
    pub registry: Arc<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
    #[serde(default)]
    pub dp_rank: Option<u32>,
}

#[derive(Deserialize)]
pub struct QueryRequest {
    pub token_ids: Vec<u32>,
57
58
59
    pub model_name: String,
    #[serde(default = "default_tenant")]
    pub tenant_id: String,
60
61
62
63
64
65
66
    #[serde(default)]
    pub lora_name: Option<String>,
}

#[derive(Deserialize)]
pub struct QueryByHashRequest {
    pub block_hashes: Vec<i64>,
67
68
69
    pub model_name: String,
    #[serde(default = "default_tenant")]
    pub tenant_id: String,
70
71
72
73
74
75
76
77
78
79
80
81
82
}

#[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 {
83
84
85
86
87
88
89
90
91
    if let Err(error) =
        super::validate_listener_endpoints(&req.endpoint, req.replay_endpoint.as_deref())
    {
        return (
            StatusCode::BAD_REQUEST,
            Json(serde_json::json!({"error": error.to_string()})),
        );
    }

92
93
94
95
96
97
98
99
100
101
102
103
104
    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
    {
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
        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 {
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
    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 => {
136
137
            state
                .registry
138
                .deregister_all_tenants(req.instance_id, &req.model_name)
139
140
141
142
143
144
145
146
147
148
149
150
151
                .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 {
152
    Json(state.registry.list())
153
154
}

155
fn build_score_response(
156
    overlap: crate::protocols::OverlapScores,
157
158
    block_size: u32,
) -> ScoreResponse {
159
160
161
162
163
    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()
164
            .insert(k.dp_rank.to_string(), v * block_size);
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
    }
    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 {
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
    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);

200
201
202
203
204
205
206
207
    let block_hashes = compute_block_hash_for_seq(
        &req.token_ids,
        block_size,
        BlockHashOptions {
            lora_name: req.lora_name.as_deref(),
            ..Default::default()
        },
    );
208
    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
#[derive(Deserialize)]
struct ListenerControlRequest {
    instance_id: WorkerId,
    #[serde(default)]
    dp_rank: Option<u32>,
}

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"}))),
273
        Err(error) => listener_control_error_response(error),
274
275
276
277
278
279
280
281
282
283
284
285
286
    }
}

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"}))),
287
        Err(error) => listener_control_error_response(error),
288
289
290
    }
}

291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
fn listener_control_error_response(
    error: ListenerControlError,
) -> (StatusCode, Json<serde_json::Value>) {
    let status = match &error {
        ListenerControlError::WorkerNotFound { .. }
        | ListenerControlError::ListenerNotFound { .. } => StatusCode::NOT_FOUND,
        ListenerControlError::DiscoveryManaged { .. }
        | ListenerControlError::InvalidPauseState { .. }
        | ListenerControlError::InvalidResumeState { .. } => StatusCode::CONFLICT,
    };
    (
        status,
        Json(serde_json::json!({"error": error.to_string()})),
    )
}

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
334
335
336
337
338
339
340
#[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())
}

341
async fn dump_events(State(state): State<Arc<AppState>>) -> impl IntoResponse {
342
343
    let all = state.registry.all_indexers_with_block_size();
    let mut handles = Vec::with_capacity(all.len());
344

345
    for (key, indexer, block_size) in all {
346
347
        handles.push(tokio::spawn(async move {
            let events = indexer.dump_events().await;
348
            (key, events, block_size)
349
350
351
352
353
354
        }));
    }

    let mut result: HashMap<String, serde_json::Value> = HashMap::new();
    for handle in handles {
        match handle.await {
355
            Ok((key, Ok(events), block_size)) => {
356
                let map_key = format!("{}:{}", key.model_name, key.tenant_id);
357
358
359
360
361
362
363
                result.insert(
                    map_key,
                    serde_json::json!({
                        "block_size": block_size,
                        "events": events,
                    }),
                );
364
            }
365
            Ok((key, Err(e), _)) => {
366
367
368
369
370
371
372
                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}");
            }
        }
373
    }
374
    (StatusCode::OK, Json(serde_json::json!(result)))
375
376
}

377
378
379
380
381
382
async fn handle_health() -> StatusCode {
    StatusCode::OK
}

#[cfg(feature = "metrics")]
async fn handle_metrics(State(state): State<Arc<AppState>>) -> impl IntoResponse {
383
    state.registry.refresh_metrics();
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
    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,
    )
}

399
pub fn create_router(state: Arc<AppState>) -> Router {
400
    let router = Router::new()
401
402
403
404
405
406
        .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))
407
408
        .route("/register_peer", post(register_peer))
        .route("/deregister_peer", post(deregister_peer))
409
410
        .route("/peers", get(list_peers))
        .route("/health", get(handle_health));
411
412
413

    let router = router
        .route("/test/pause_listener", post(test_pause_listener))
414
415
        .route("/test/resume_listener", post(test_resume_listener))
        .with_state(state.clone());
416
417
418
419
420
421
422
423
424
425
426
427
428
429

    #[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
430
}