main.rs 19.2 KB
Newer Older
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
1
use clap::Parser;
2
use serde_json::Value;
Nicolas Patry's avatar
Nicolas Patry committed
3
use std::env;
4
use std::ffi::OsString;
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
5
6
7
8
9
10
11
12
13
14
15
use std::io::{BufRead, BufReader, Read};
use std::path::Path;
use std::process::ExitCode;
use std::sync::atomic::{AtomicBool, Ordering};
use std::sync::mpsc::TryRecvError;
use std::sync::Arc;
use std::sync::{mpsc, Mutex};
use std::thread;
use std::thread::sleep;
use std::time::{Duration, Instant};
use std::{fs, io};
16
use subprocess::{ExitStatus, Popen, PopenConfig, PopenError, Redirection};
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
17
18
19
20
21
22

/// App Configuration
#[derive(Parser, Debug)]
#[clap(author, version, about, long_about = None)]
struct Args {
    #[clap(default_value = "bigscience/bloom-560m", long, env)]
23
    model_id: String,
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
24
    #[clap(long, env)]
25
    revision: Option<String>,
26
27
    #[clap(default_value = "1", long, env)]
    num_shard: usize,
28
29
    #[clap(long, env)]
    quantize: bool,
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
30
31
32
33
34
35
    #[clap(default_value = "128", long, env)]
    max_concurrent_requests: usize,
    #[clap(default_value = "1000", long, env)]
    max_input_length: usize,
    #[clap(default_value = "32", long, env)]
    max_batch_size: usize,
36
37
    #[clap(default_value = "20", long, env)]
    max_waiting_tokens: usize,
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
38
39
40
41
    #[clap(default_value = "3000", long, short, env)]
    port: u16,
    #[clap(default_value = "/tmp/text-generation-server", long, env)]
    shard_uds_path: String,
42
    #[clap(default_value = "localhost", long, env)]
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
43
    master_addr: String,
44
    #[clap(default_value = "29500", long, env)]
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
45
    master_port: usize,
46
    #[clap(long, env)]
47
48
49
50
    huggingface_hub_cache: Option<String>,
    #[clap(long, env)]
    weights_cache_override: Option<String>,
    #[clap(long, env)]
51
52
    disable_custom_kernels: bool,
    #[clap(long, env)]
53
    json_output: bool,
54
55
    #[clap(long, env)]
    otlp_endpoint: Option<String>,
56
57
    #[clap(long, env)]
    cors_allow_origin: Vec<String>,
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
58
59
60
61
}

fn main() -> ExitCode {
    // Pattern match configuration
62
63
64
65
66
67
68
69
70
71
    let args = Args::parse();

    if args.json_output {
        tracing_subscriber::fmt().json().init();
    } else {
        tracing_subscriber::fmt().compact().init();
    }

    tracing::info!("{:?}", args);

Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
72
    let Args {
73
        model_id,
74
        revision,
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
75
        num_shard,
76
        quantize,
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
77
78
79
        max_concurrent_requests,
        max_input_length,
        max_batch_size,
80
        max_waiting_tokens,
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
81
82
83
84
        port,
        shard_uds_path,
        master_addr,
        master_port,
85
86
        huggingface_hub_cache,
        weights_cache_override,
87
        disable_custom_kernels,
88
        json_output,
89
        otlp_endpoint,
90
        cors_allow_origin,
91
    } = args;
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
92
93
94
95
96
97
98
99
100

    // Signal handler
    let running = Arc::new(AtomicBool::new(true));
    let r = running.clone();
    ctrlc::set_handler(move || {
        r.store(false, Ordering::SeqCst);
    })
    .expect("Error setting Ctrl-C handler");

101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
    // Download weights
    if weights_cache_override.is_none() {
        let mut download_argv = vec![
            "text-generation-server".to_string(),
            "download-weights".to_string(),
            model_id.clone(),
            "--logger-level".to_string(),
            "INFO".to_string(),
            "--json-output".to_string(),
        ];
        if num_shard == 1 {
            download_argv.push("--extension".to_string());
            download_argv.push(".bin".to_string());
        } else {
            download_argv.push("--extension".to_string());
            download_argv.push(".safetensors".to_string());
        }

        // Model optional revision
        if let Some(ref revision) = revision {
            download_argv.push("--revision".to_string());
            download_argv.push(revision.to_string())
        }

125
126
        // Copy current process env
        let mut env: Vec<(OsString, OsString)> = env::vars_os().collect();
127

128
        // If huggingface_hub_cache is set, pass it to the shard
129
130
131
132
133
134
        // Useful when running inside a docker container
        if let Some(ref huggingface_hub_cache) = huggingface_hub_cache {
            env.push(("HUGGINGFACE_HUB_CACHE".into(), huggingface_hub_cache.into()));
        };

        // Start process
135
        tracing::info!("Starting download process.");
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
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
        let mut download_process = match Popen::create(
            &download_argv,
            PopenConfig {
                stdout: Redirection::Pipe,
                stderr: Redirection::Pipe,
                // Needed for the shutdown procedure
                setpgid: true,
                env: Some(env),
                ..Default::default()
            },
        ) {
            Ok(p) => p,
            Err(err) => {
                if let PopenError::IoError(ref err) = err {
                    if err.kind() == io::ErrorKind::NotFound {
                        tracing::error!("text-generation-server not found in PATH");
                        tracing::error!("Please install it with `make install-server`")
                    }
                }
                return ExitCode::FAILURE;
            }
        };

        // Redirect STDOUT to the console
        let download_stdout = download_process.stdout.take().unwrap();
        thread::spawn(move || {
            // Enter download tracing span
            let stdout = BufReader::new(download_stdout);
            let _span = tracing::span!(tracing::Level::INFO, "download").entered();
            for line in stdout.lines() {
                // Parse loguru logs
                if let Ok(value) = serde_json::from_str::<Value>(&line.unwrap()) {
                    if let Some(text) = value.get("text") {
                        // Format escaped newlines
                        tracing::info!("{}", text.to_string().replace("\\n", ""));
                    }
                }
            }
        });

        loop {
            if let Some(status) = download_process.poll() {
                match status {
                    ExitStatus::Exited(exit_code) => {
                        if exit_code == 0 {
                            tracing::info!("Successfully downloaded weights.");
                            break;
                        } else {
                            let mut err = String::new();
                            download_process
                                .stderr
                                .take()
                                .unwrap()
                                .read_to_string(&mut err)
                                .unwrap();
                            tracing::error!("Download encountered an error: {err}");
                            return ExitCode::FAILURE;
                        }
                    }
                    _ => {
196
                        tracing::error!("Download process exited with an unknown status.");
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
                        return ExitCode::FAILURE;
                    }
                }
            }
            if !running.load(Ordering::SeqCst) {
                download_process.terminate().unwrap();
                tracing::info!("Waiting for download process to gracefully shutdown");
                download_process
                    .wait_timeout(Duration::from_secs(90))
                    .unwrap();
                tracing::info!("Download process terminated");
                return ExitCode::SUCCESS;
            }
            sleep(Duration::from_millis(100));
        }
    } else {
        tracing::info!(
            "weights_cache_override is set to {:?}.",
            weights_cache_override
        );
        tracing::info!("Skipping download.")
    }

Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
220
221
222
223
224
225
226
227
228
229
230
    // Shared shutdown bool
    let shutdown = Arc::new(Mutex::new(false));
    // Shared shutdown channel
    // When shutting down, the main thread will wait for all senders to be dropped
    let (shutdown_sender, shutdown_receiver) = mpsc::channel();

    // Shared channel to track shard status
    let (status_sender, status_receiver) = mpsc::channel();

    // Start shard processes
    for rank in 0..num_shard {
231
        let model_id = model_id.clone();
232
        let revision = revision.clone();
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
233
234
        let uds_path = shard_uds_path.clone();
        let master_addr = master_addr.clone();
235
236
        let huggingface_hub_cache = huggingface_hub_cache.clone();
        let weights_cache_override = weights_cache_override.clone();
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
237
238
239
        let status_sender = status_sender.clone();
        let shutdown = shutdown.clone();
        let shutdown_sender = shutdown_sender.clone();
240
        let otlp_endpoint = otlp_endpoint.clone();
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
241
242
        thread::spawn(move || {
            shard_manager(
243
                model_id,
244
                revision,
245
                quantize,
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
246
247
248
249
250
                uds_path,
                rank,
                num_shard,
                master_addr,
                master_port,
251
252
                huggingface_hub_cache,
                weights_cache_override,
253
                disable_custom_kernels,
254
                otlp_endpoint,
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
255
256
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
                status_sender,
                shutdown,
                shutdown_sender,
            )
        });
    }
    drop(shutdown_sender);

    // Wait for shard to start
    let mut shard_ready = 0;
    while running.load(Ordering::SeqCst) {
        match status_receiver.try_recv() {
            Ok(ShardStatus::Ready) => {
                shard_ready += 1;
                if shard_ready == num_shard {
                    break;
                }
            }
            Err(TryRecvError::Empty) => {
                sleep(Duration::from_millis(100));
            }
            Ok(ShardStatus::Failed((rank, err))) => {
                tracing::error!("Shard {} failed to start:\n{}", rank, err);
                shutdown_shards(shutdown, &shutdown_receiver);
                return ExitCode::FAILURE;
            }
            Err(TryRecvError::Disconnected) => {
                tracing::error!("Shard status channel disconnected");
                shutdown_shards(shutdown, &shutdown_receiver);
                return ExitCode::FAILURE;
            }
        }
    }

    // We might have received a termination signal
    if !running.load(Ordering::SeqCst) {
        shutdown_shards(shutdown, &shutdown_receiver);
        return ExitCode::SUCCESS;
    }

    // All shard started
    // Start webserver
    tracing::info!("Starting Webserver");
298
299
300
301
302
303
304
305
306
307
308
309
310
    let mut argv = vec![
        "text-generation-router".to_string(),
        "--max-concurrent-requests".to_string(),
        max_concurrent_requests.to_string(),
        "--max-input-length".to_string(),
        max_input_length.to_string(),
        "--max-batch-size".to_string(),
        max_batch_size.to_string(),
        "--max-waiting-tokens".to_string(),
        max_waiting_tokens.to_string(),
        "--port".to_string(),
        port.to_string(),
        "--master-shard-uds-path".to_string(),
311
        format!("{shard_uds_path}-0"),
312
        "--tokenizer-name".to_string(),
313
        model_id,
314
315
316
317
318
319
    ];

    if json_output {
        argv.push("--json-output".to_string());
    }

320
321
322
323
324
325
    // OpenTelemetry
    if let Some(otlp_endpoint) = otlp_endpoint {
        argv.push("--otlp-endpoint".to_string());
        argv.push(otlp_endpoint);
    }

326
327
328
329
330
331
    // CORS origins
    for origin in cors_allow_origin.into_iter() {
        argv.push("--cors-allow-origin".to_string());
        argv.push(origin);
    }

Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
332
    let mut webserver = match Popen::create(
333
        &argv,
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
        PopenConfig {
            stdout: Redirection::Pipe,
            stderr: Redirection::Pipe,
            // Needed for the shutdown procedure
            setpgid: true,
            ..Default::default()
        },
    ) {
        Ok(p) => p,
        Err(err) => {
            tracing::error!("Failed to start webserver: {}", err);
            if let PopenError::IoError(err) = err {
                if err.kind() == io::ErrorKind::NotFound {
                    tracing::error!("text-generation-router not found in PATH");
                    tracing::error!("Please install it with `make install-router`")
                }
350
351
            } else {
                tracing::error!("{}", err);
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
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
            }

            shutdown_shards(shutdown, &shutdown_receiver);
            return ExitCode::FAILURE;
        }
    };

    // Redirect STDOUT and STDERR to the console
    let webserver_stdout = webserver.stdout.take().unwrap();
    let webserver_stderr = webserver.stderr.take().unwrap();

    thread::spawn(move || {
        let stdout = BufReader::new(webserver_stdout);
        let stderr = BufReader::new(webserver_stderr);
        for line in stdout.lines() {
            println!("{}", line.unwrap());
        }
        for line in stderr.lines() {
            println!("{}", line.unwrap());
        }
    });

    // Default exit code
    let mut exit_code = ExitCode::SUCCESS;

    while running.load(Ordering::SeqCst) {
        if let Ok(ShardStatus::Failed((rank, err))) = status_receiver.try_recv() {
379
            tracing::error!("Shard {rank} failed:\n{err}");
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
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
            exit_code = ExitCode::FAILURE;
            break;
        };

        match webserver.poll() {
            Some(_) => {
                tracing::error!("Webserver Crashed");
                shutdown_shards(shutdown, &shutdown_receiver);
                return ExitCode::FAILURE;
            }
            None => {
                sleep(Duration::from_millis(100));
            }
        };
    }

    // Graceful termination
    webserver.terminate().unwrap();
    tracing::info!("Waiting for webserver to gracefully shutdown");
    webserver.wait_timeout(Duration::from_secs(90)).unwrap();
    tracing::info!("Webserver terminated");
    shutdown_shards(shutdown, &shutdown_receiver);

    exit_code
}

#[derive(Debug)]
enum ShardStatus {
    Ready,
    Failed((usize, String)),
}

#[allow(clippy::too_many_arguments)]
fn shard_manager(
414
    model_id: String,
415
    revision: Option<String>,
416
    quantize: bool,
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
417
418
419
420
421
    uds_path: String,
    rank: usize,
    world_size: usize,
    master_addr: String,
    master_port: usize,
422
423
    huggingface_hub_cache: Option<String>,
    weights_cache_override: Option<String>,
424
    disable_custom_kernels: bool,
425
    otlp_endpoint: Option<String>,
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
426
427
428
429
430
    status_sender: mpsc::Sender<ShardStatus>,
    shutdown: Arc<Mutex<bool>>,
    _shutdown_sender: mpsc::Sender<()>,
) {
    // Get UDS path
431
    let uds_string = format!("{uds_path}-{rank}");
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
432
433
434
435
436
437
    let uds = Path::new(&uds_string);
    // Clean previous runs
    fs::remove_file(uds).unwrap_or_default();

    // Process args
    let mut shard_argv = vec![
438
        "text-generation-server".to_string(),
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
439
        "serve".to_string(),
440
        model_id,
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
441
442
        "--uds-path".to_string(),
        uds_path,
443
444
445
        "--logger-level".to_string(),
        "ERROR".to_string(),
        "--json-output".to_string(),
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
446
447
    ];

448
    // Activate tensor parallelism
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
449
450
451
452
    if world_size > 1 {
        shard_argv.push("--sharded".to_string());
    }

453
454
455
456
    if quantize {
        shard_argv.push("--quantize".to_string())
    }

457
    // Model optional revision
458
459
460
461
462
    if let Some(revision) = revision {
        shard_argv.push("--revision".to_string());
        shard_argv.push(revision)
    }

463
464
465
466
467
468
    // OpenTelemetry
    if let Some(otlp_endpoint) = otlp_endpoint {
        shard_argv.push("--otlp-endpoint".to_string());
        shard_argv.push(otlp_endpoint);
    }

469
470
471
472
473
474
475
476
477
478
479
480
    // Copy current process env
    let mut env: Vec<(OsString, OsString)> = env::vars_os().collect();

    // Torch Distributed Env vars
    env.push(("RANK".into(), rank.to_string().into()));
    env.push(("WORLD_SIZE".into(), world_size.to_string().into()));
    env.push(("MASTER_ADDR".into(), master_addr.into()));
    env.push(("MASTER_PORT".into(), master_port.to_string().into()));
    env.push(("NCCL_ASYNC_ERROR_HANDLING".into(), "1".into()));

    // Safetensors load fast
    env.push(("SAFETENSORS_FAST_GPU".into(), "1".into()));
Nicolas Patry's avatar
Nicolas Patry committed
481

482
    // If huggingface_hub_cache is some, pass it to the shard
Nicolas Patry's avatar
Nicolas Patry committed
483
    // Useful when running inside a docker container
484
    if let Some(huggingface_hub_cache) = huggingface_hub_cache {
485
        env.push(("HUGGINGFACE_HUB_CACHE".into(), huggingface_hub_cache.into()));
Nicolas Patry's avatar
Nicolas Patry committed
486
    };
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
487

488
    // If weights_cache_override is some, pass it to the shard
489
    // Useful when running inside a HuggingFace Inference Endpoint
490
    if let Some(weights_cache_override) = weights_cache_override {
491
492
493
494
        env.push((
            "WEIGHTS_CACHE_OVERRIDE".into(),
            weights_cache_override.into(),
        ));
495
496
    };

497
498
499
500
501
    // If disable_custom_kernels is true, pass it to the shard as an env var
    if disable_custom_kernels {
        env.push(("DISABLE_CUSTOM_KERNELS".into(), "True".into()))
    }

Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
502
    // Start process
503
    tracing::info!("Starting shard {rank}");
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
504
505
506
507
508
509
510
511
    let mut p = match Popen::create(
        &shard_argv,
        PopenConfig {
            stdout: Redirection::Pipe,
            stderr: Redirection::Pipe,
            // Needed for the shutdown procedure
            setpgid: true,
            // NCCL env vars
Nicolas Patry's avatar
Nicolas Patry committed
512
            env: Some(env),
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
513
514
515
516
517
518
519
            ..Default::default()
        },
    ) {
        Ok(p) => p,
        Err(err) => {
            if let PopenError::IoError(ref err) = err {
                if err.kind() == io::ErrorKind::NotFound {
520
                    tracing::error!("text-generation-server not found in PATH");
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
521
522
523
524
525
526
527
528
529
530
                    tracing::error!("Please install it with `make install-server`")
                }
            }
            status_sender
                .send(ShardStatus::Failed((rank, err.to_string())))
                .unwrap();
            return;
        }
    };

531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
    // Redirect STDOUT to the console
    let shard_stdout = p.stdout.take().unwrap();

    thread::spawn(move || {
        // Enter shard-manager tracing span
        let stdout = BufReader::new(shard_stdout);
        let _span = tracing::span!(tracing::Level::INFO, "shard-manager", rank = rank).entered();
        for line in stdout.lines() {
            // Parse loguru logs
            if let Ok(value) = serde_json::from_str::<Value>(&line.unwrap()) {
                if let Some(text) = value.get("text") {
                    // Format escaped newlines
                    tracing::error!("{}", text.to_string().replace("\\n", "\n"));
                }
            }
        }
    });

Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
549
550
    let mut ready = false;
    let start_time = Instant::now();
Nicolas Patry's avatar
Nicolas Patry committed
551
    let mut wait_time = Instant::now();
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
    loop {
        // Process exited
        if p.poll().is_some() {
            let mut err = String::new();
            p.stderr.take().unwrap().read_to_string(&mut err).unwrap();
            status_sender
                .send(ShardStatus::Failed((rank, err)))
                .unwrap();
            return;
        }

        // We received a shutdown signal
        if *shutdown.lock().unwrap() {
            p.terminate().unwrap();
            let _ = p.wait_timeout(Duration::from_secs(90));
567
            tracing::info!("Shard {rank} terminated");
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
568
569
570
571
572
            return;
        }

        // Shard is ready
        if uds.exists() && !ready {
573
            tracing::info!("Shard {rank} ready in {:?}", start_time.elapsed());
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
574
575
            status_sender.send(ShardStatus::Ready).unwrap();
            ready = true;
576
        } else if !ready && wait_time.elapsed() > Duration::from_secs(10) {
577
            tracing::info!("Waiting for shard {rank} to be ready...");
578
            wait_time = Instant::now();
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
579
        }
Nicolas Patry's avatar
Nicolas Patry committed
580
        sleep(Duration::from_millis(100));
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
    }
}

fn shutdown_shards(shutdown: Arc<Mutex<bool>>, shutdown_receiver: &mpsc::Receiver<()>) {
    tracing::info!("Shutting down shards");
    // Update shutdown value to true
    // This will be picked up by the shard manager
    {
        let mut shutdown = shutdown.lock().unwrap();
        *shutdown = true;
    }

    // Wait for shards to shutdown
    // This will block till all shutdown_sender are dropped
    let _ = shutdown_receiver.recv();
}