launch_router.py 23 KB
Newer Older
1
2
import argparse
import dataclasses
3
import logging
4
import sys
5
from typing import Dict, List, Optional
6
7
8
9
10

from sglang_router import Router
from sglang_router_rs import PolicyType


11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
def setup_logger():
    logger = logging.getLogger("router")
    logger.setLevel(logging.INFO)

    formatter = logging.Formatter(
        "[Router (Python)] %(asctime)s - %(levelname)s - %(message)s",
        datefmt="%Y-%m-%d %H:%M:%S",
    )

    handler = logging.StreamHandler()
    handler.setFormatter(formatter)
    logger.addHandler(handler)

    return logger


27
28
29
@dataclasses.dataclass
class RouterArgs:
    # Worker configuration
30
    worker_urls: List[str] = dataclasses.field(default_factory=list)
31
32
33
    host: str = "127.0.0.1"
    port: int = 30000

34
    # PD-specific configuration
35
    pd_disaggregation: bool = False  # Enable PD disaggregated mode
36
37
38
39
40
    prefill_urls: List[tuple] = dataclasses.field(
        default_factory=list
    )  # List of (url, bootstrap_port)
    decode_urls: List[str] = dataclasses.field(default_factory=list)

41
42
    # Routing policy
    policy: str = "cache_aware"
43
44
    prefill_policy: Optional[str] = None  # Specific policy for prefill nodes in PD mode
    decode_policy: Optional[str] = None  # Specific policy for decode nodes in PD mode
45
    worker_startup_timeout_secs: int = 300
46
    worker_startup_check_interval: int = 10
47
    cache_threshold: float = 0.5
48
49
    balance_abs_threshold: int = 32
    balance_rel_threshold: float = 1.0001
50
51
    eviction_interval: int = 60
    max_tree_size: int = 2**24
52
    max_payload_size: int = 256 * 1024 * 1024  # 256MB default for large batches
53
54
    dp_aware: bool = False
    api_key: Optional[str] = None
55
    log_dir: Optional[str] = None
56
    log_level: Optional[str] = None
57
58
59
60
61
    # Service discovery configuration
    service_discovery: bool = False
    selector: Dict[str, str] = dataclasses.field(default_factory=dict)
    service_discovery_port: int = 80
    service_discovery_namespace: Optional[str] = None
62
63
64
65
    # PD service discovery configuration
    prefill_selector: Dict[str, str] = dataclasses.field(default_factory=dict)
    decode_selector: Dict[str, str] = dataclasses.field(default_factory=dict)
    bootstrap_port_annotation: str = "sglang.ai/bootstrap-port"
66
67
68
    # Prometheus configuration
    prometheus_port: Optional[int] = None
    prometheus_host: Optional[str] = None
69
70
    # Request ID headers configuration
    request_id_headers: Optional[List[str]] = None
71
72
73
74
75
76
    # Request timeout in seconds
    request_timeout_secs: int = 600
    # Max concurrent requests for rate limiting
    max_concurrent_requests: int = 64
    # CORS allowed origins
    cors_allowed_origins: List[str] = dataclasses.field(default_factory=list)
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

    @staticmethod
    def add_cli_args(
        parser: argparse.ArgumentParser,
        use_router_prefix: bool = False,
        exclude_host_port: bool = False,
    ):
        """
        Add router-specific arguments to an argument parser.

        Args:
            parser: The argument parser to add arguments to
            use_router_prefix: If True, prefix all arguments with 'router-' to avoid conflicts
            exclude_host_port: If True, don't add host and port arguments (used when inheriting from server)
        """
        prefix = "router-" if use_router_prefix else ""

        # Worker configuration
        if not exclude_host_port:
            parser.add_argument(
                "--host",
                type=str,
                default=RouterArgs.host,
                help="Host address to bind the router server",
            )
            parser.add_argument(
                "--port",
                type=int,
                default=RouterArgs.port,
                help="Port number to bind the router server",
            )

        parser.add_argument(
            "--worker-urls",
            type=str,
112
113
            nargs="*",
            default=[],
114
115
116
117
118
119
120
121
            help="List of worker URLs (e.g., http://worker1:8000 http://worker2:8000)",
        )

        # Routing policy configuration
        parser.add_argument(
            f"--{prefix}policy",
            type=str,
            default=RouterArgs.policy,
122
            choices=["random", "round_robin", "cache_aware", "power_of_two"],
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
            help="Load balancing policy to use. In PD mode, this is used for both prefill and decode unless overridden",
        )
        parser.add_argument(
            f"--{prefix}prefill-policy",
            type=str,
            default=None,
            choices=["random", "round_robin", "cache_aware", "power_of_two"],
            help="Specific policy for prefill nodes in PD mode. If not specified, uses the main policy",
        )
        parser.add_argument(
            f"--{prefix}decode-policy",
            type=str,
            default=None,
            choices=["random", "round_robin", "cache_aware", "power_of_two"],
            help="Specific policy for decode nodes in PD mode. If not specified, uses the main policy",
138
139
140
141
        )

        # PD-specific arguments
        parser.add_argument(
142
            f"--{prefix}pd-disaggregation",
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
            action="store_true",
            help="Enable PD (Prefill-Decode) disaggregated mode",
        )
        parser.add_argument(
            f"--{prefix}prefill",
            nargs=2,
            action="append",
            metavar=("URL", "BOOTSTRAP_PORT"),
            help="Prefill server URL and bootstrap port. Can be specified multiple times. BOOTSTRAP_PORT can be 'none' for no bootstrap port.",
        )
        parser.add_argument(
            f"--{prefix}decode",
            nargs=1,
            action="append",
            metavar=("URL",),
            help="Decode server URL. Can be specified multiple times.",
159
        )
160
161
162
163
164
165
        parser.add_argument(
            f"--{prefix}worker-startup-timeout-secs",
            type=int,
            default=RouterArgs.worker_startup_timeout_secs,
            help="Timeout in seconds for worker startup",
        )
166
167
168
169
170
171
        parser.add_argument(
            f"--{prefix}worker-startup-check-interval",
            type=int,
            default=RouterArgs.worker_startup_check_interval,
            help="Interval in seconds between checks for worker startup",
        )
172
173
174
175
176
177
178
        parser.add_argument(
            f"--{prefix}cache-threshold",
            type=float,
            default=RouterArgs.cache_threshold,
            help="Cache threshold (0.0-1.0) for cache-aware routing",
        )
        parser.add_argument(
179
180
181
182
183
184
185
            f"--{prefix}balance-abs-threshold",
            type=int,
            default=RouterArgs.balance_abs_threshold,
            help="Load balancing is triggered when (max_load - min_load) > abs_threshold AND max_load > min_load * rel_threshold. Otherwise, use cache aware",
        )
        parser.add_argument(
            f"--{prefix}balance-rel-threshold",
186
            type=float,
187
188
            default=RouterArgs.balance_rel_threshold,
            help="Load balancing is triggered when (max_load - min_load) > abs_threshold AND max_load > min_load * rel_threshold. Otherwise, use cache aware",
189
190
191
192
193
194
195
196
197
198
199
200
201
        )
        parser.add_argument(
            f"--{prefix}eviction-interval",
            type=int,
            default=RouterArgs.eviction_interval,
            help="Interval in seconds between cache eviction operations",
        )
        parser.add_argument(
            f"--{prefix}max-tree-size",
            type=int,
            default=RouterArgs.max_tree_size,
            help="Maximum size of the approximation tree for cache-aware routing",
        )
202
203
204
205
206
207
        parser.add_argument(
            f"--{prefix}max-payload-size",
            type=int,
            default=RouterArgs.max_payload_size,
            help="Maximum payload size in bytes",
        )
208
209
210
211
212
213
214
215
216
217
218
        parser.add_argument(
            f"--{prefix}dp-aware",
            action="store_true",
            help="Enable data parallelism aware schedule",
        )
        parser.add_argument(
            f"--{prefix}api-key",
            type=str,
            default=None,
            help="The api key used for the authorization with the worker.  Useful when the dp aware scheduling strategy is enaled.",
        )
219
220
221
222
223
224
        parser.add_argument(
            f"--{prefix}log-dir",
            type=str,
            default=None,
            help="Directory to store log files. If not specified, logs are only output to console.",
        )
225
226
227
228
229
230
231
        parser.add_argument(
            f"--{prefix}log-level",
            type=str,
            default="info",
            choices=["debug", "info", "warning", "error", "critical"],
            help="Set the logging level. If not specified, defaults to INFO.",
        )
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
        parser.add_argument(
            f"--{prefix}service-discovery",
            action="store_true",
            help="Enable Kubernetes service discovery",
        )
        parser.add_argument(
            f"--{prefix}selector",
            type=str,
            nargs="+",
            help="Label selector for Kubernetes service discovery (format: key1=value1 key2=value2)",
        )
        parser.add_argument(
            f"--{prefix}service-discovery-port",
            type=int,
            default=RouterArgs.service_discovery_port,
            help="Port to use for discovered worker pods",
        )
        parser.add_argument(
            f"--{prefix}service-discovery-namespace",
            type=str,
            help="Kubernetes namespace to watch for pods. If not provided, watches all namespaces (requires cluster-wide permissions)",
        )
254
255
256
257
258
259
260
261
262
263
264
265
        parser.add_argument(
            f"--{prefix}prefill-selector",
            type=str,
            nargs="+",
            help="Label selector for prefill server pods in PD mode (format: key1=value1 key2=value2)",
        )
        parser.add_argument(
            f"--{prefix}decode-selector",
            type=str,
            nargs="+",
            help="Label selector for decode server pods in PD mode (format: key1=value1 key2=value2)",
        )
266
267
268
269
270
271
272
273
274
275
276
277
278
        # Prometheus configuration
        parser.add_argument(
            f"--{prefix}prometheus-port",
            type=int,
            default=29000,
            help="Port to expose Prometheus metrics. If not specified, Prometheus metrics are disabled",
        )
        parser.add_argument(
            f"--{prefix}prometheus-host",
            type=str,
            default="127.0.0.1",
            help="Host address to bind the Prometheus metrics server",
        )
279
280
281
282
283
284
        parser.add_argument(
            f"--{prefix}request-id-headers",
            type=str,
            nargs="*",
            help="Custom HTTP headers to check for request IDs (e.g., x-request-id x-trace-id). If not specified, uses common defaults.",
        )
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
        parser.add_argument(
            f"--{prefix}request-timeout-secs",
            type=int,
            default=RouterArgs.request_timeout_secs,
            help="Request timeout in seconds",
        )
        parser.add_argument(
            f"--{prefix}max-concurrent-requests",
            type=int,
            default=RouterArgs.max_concurrent_requests,
            help="Maximum number of concurrent requests allowed (for rate limiting)",
        )
        parser.add_argument(
            f"--{prefix}cors-allowed-origins",
            type=str,
            nargs="*",
            default=[],
            help="CORS allowed origins (e.g., http://localhost:3000 https://example.com)",
        )
304
305
306
307
308
309
310
311
312
313
314
315
316

    @classmethod
    def from_cli_args(
        cls, args: argparse.Namespace, use_router_prefix: bool = False
    ) -> "RouterArgs":
        """
        Create RouterArgs instance from parsed command line arguments.

        Args:
            args: Parsed command line arguments
            use_router_prefix: If True, look for arguments with 'router-' prefix
        """
        prefix = "router_" if use_router_prefix else ""
317
318
319
320
321
322
        worker_urls = getattr(args, "worker_urls", [])

        # Parse PD URLs
        prefill_urls = cls._parse_prefill_urls(getattr(args, f"{prefix}prefill", None))
        decode_urls = cls._parse_decode_urls(getattr(args, f"{prefix}decode", None))

323
        return cls(
324
            worker_urls=worker_urls,
325
326
            host=args.host,
            port=args.port,
327
            pd_disaggregation=getattr(args, f"{prefix}pd_disaggregation", False),
328
329
            prefill_urls=prefill_urls,
            decode_urls=decode_urls,
330
            policy=getattr(args, f"{prefix}policy"),
331
332
            prefill_policy=getattr(args, f"{prefix}prefill_policy", None),
            decode_policy=getattr(args, f"{prefix}decode_policy", None),
333
334
335
            worker_startup_timeout_secs=getattr(
                args, f"{prefix}worker_startup_timeout_secs"
            ),
336
337
338
            worker_startup_check_interval=getattr(
                args, f"{prefix}worker_startup_check_interval"
            ),
339
            cache_threshold=getattr(args, f"{prefix}cache_threshold"),
340
341
            balance_abs_threshold=getattr(args, f"{prefix}balance_abs_threshold"),
            balance_rel_threshold=getattr(args, f"{prefix}balance_rel_threshold"),
342
343
            eviction_interval=getattr(args, f"{prefix}eviction_interval"),
            max_tree_size=getattr(args, f"{prefix}max_tree_size"),
344
            max_payload_size=getattr(args, f"{prefix}max_payload_size"),
345
346
            dp_aware=getattr(args, f"{prefix}dp_aware", False),
            api_key=getattr(args, f"{prefix}api_key", None),
347
            log_dir=getattr(args, f"{prefix}log_dir", None),
348
            log_level=getattr(args, f"{prefix}log_level", None),
349
350
351
352
353
354
            service_discovery=getattr(args, f"{prefix}service_discovery", False),
            selector=cls._parse_selector(getattr(args, f"{prefix}selector", None)),
            service_discovery_port=getattr(args, f"{prefix}service_discovery_port"),
            service_discovery_namespace=getattr(
                args, f"{prefix}service_discovery_namespace", None
            ),
355
356
357
358
359
360
361
            prefill_selector=cls._parse_selector(
                getattr(args, f"{prefix}prefill_selector", None)
            ),
            decode_selector=cls._parse_selector(
                getattr(args, f"{prefix}decode_selector", None)
            ),
            bootstrap_port_annotation="sglang.ai/bootstrap-port",  # Mooncake-specific annotation
362
363
            prometheus_port=getattr(args, f"{prefix}prometheus_port", None),
            prometheus_host=getattr(args, f"{prefix}prometheus_host", None),
364
            request_id_headers=getattr(args, f"{prefix}request_id_headers", None),
365
366
367
368
369
370
371
372
373
            request_timeout_secs=getattr(
                args, f"{prefix}request_timeout_secs", RouterArgs.request_timeout_secs
            ),
            max_concurrent_requests=getattr(
                args,
                f"{prefix}max_concurrent_requests",
                RouterArgs.max_concurrent_requests,
            ),
            cors_allowed_origins=getattr(args, f"{prefix}cors_allowed_origins", []),
374
375
        )

376
377
378
379
380
381
382
383
384
385
386
387
    @staticmethod
    def _parse_selector(selector_list):
        if not selector_list:
            return {}

        selector = {}
        for item in selector_list:
            if "=" in item:
                key, value = item.split("=", 1)
                selector[key] = value
        return selector

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
    @staticmethod
    def _parse_prefill_urls(prefill_list):
        """Parse prefill URLs from --prefill arguments.

        Format: --prefill URL BOOTSTRAP_PORT
        Example: --prefill http://prefill1:8080 9000 --prefill http://prefill2:8080 none
        """
        if not prefill_list:
            return []

        prefill_urls = []
        for url, bootstrap_port_str in prefill_list:
            # Handle 'none' as None
            if bootstrap_port_str.lower() == "none":
                bootstrap_port = None
            else:
                try:
                    bootstrap_port = int(bootstrap_port_str)
                except ValueError:
                    raise ValueError(
                        f"Invalid bootstrap port: {bootstrap_port_str}. Must be a number or 'none'"
                    )

            prefill_urls.append((url, bootstrap_port))

        return prefill_urls

    @staticmethod
    def _parse_decode_urls(decode_list):
        """Parse decode URLs from --decode arguments.

        Format: --decode URL
        Example: --decode http://decode1:8081 --decode http://decode2:8081
        """
        if not decode_list:
            return []

        # decode_list is a list of single-element lists due to nargs=1
        return [url[0] for url in decode_list]

428
429
430
431
432
433
434

def policy_from_str(policy_str: str) -> PolicyType:
    """Convert policy string to PolicyType enum."""
    policy_map = {
        "random": PolicyType.Random,
        "round_robin": PolicyType.RoundRobin,
        "cache_aware": PolicyType.CacheAware,
435
        "power_of_two": PolicyType.PowerOfTwo,
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
    }
    return policy_map[policy_str]


def launch_router(args: argparse.Namespace) -> Optional[Router]:
    """
    Launch the SGLang router with the configuration from parsed arguments.

    Args:
        args: Namespace object containing router configuration
            Can be either raw argparse.Namespace or converted RouterArgs

    Returns:
        Router instance if successful, None if failed
    """
451
    logger = logging.getLogger("router")
452
453
454
455
456
457
458
    try:
        # Convert to RouterArgs if needed
        if not isinstance(args, RouterArgs):
            router_args = RouterArgs.from_cli_args(args)
        else:
            router_args = args

459
        # Validate configuration based on mode
460
461
462
463
464
465
466
        if router_args.pd_disaggregation:
            # Validate PD configuration - skip URL requirements if using service discovery
            if not router_args.service_discovery:
                if not router_args.prefill_urls:
                    raise ValueError("PD disaggregation mode requires --prefill")
                if not router_args.decode_urls:
                    raise ValueError("PD disaggregation mode requires --decode")
467

468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
            # Warn about policy usage in PD mode
            if (
                router_args.prefill_policy
                and router_args.decode_policy
                and router_args.policy
            ):
                logger.warning(
                    "Both --prefill-policy and --decode-policy are specified. "
                    "The main --policy flag will be ignored for PD mode."
                )
            elif (
                router_args.prefill_policy
                and not router_args.decode_policy
                and router_args.policy
            ):
                logger.info(
                    f"Using --prefill-policy '{router_args.prefill_policy}' for prefill nodes "
                    f"and --policy '{router_args.policy}' for decode nodes."
                )
            elif (
                router_args.decode_policy
                and not router_args.prefill_policy
                and router_args.policy
            ):
                logger.info(
                    f"Using --policy '{router_args.policy}' for prefill nodes "
                    f"and --decode-policy '{router_args.decode_policy}' for decode nodes."
                )

497
        # Create router with unified constructor
498
        router = Router(
499
            worker_urls=(
500
501
502
                []
                if router_args.service_discovery or router_args.pd_disaggregation
                else router_args.worker_urls
503
            ),
504
505
            host=router_args.host,
            port=router_args.port,
506
507
            policy=policy_from_str(router_args.policy),
            worker_startup_timeout_secs=router_args.worker_startup_timeout_secs,
508
            worker_startup_check_interval=router_args.worker_startup_check_interval,
509
            cache_threshold=router_args.cache_threshold,
510
511
            balance_abs_threshold=router_args.balance_abs_threshold,
            balance_rel_threshold=router_args.balance_rel_threshold,
512
513
            eviction_interval_secs=router_args.eviction_interval,
            max_tree_size=router_args.max_tree_size,
514
            max_payload_size=router_args.max_payload_size,
515
516
            dp_aware=router_args.dp_aware,
            api_key=router_args.api_key,
517
            log_dir=router_args.log_dir,
518
            log_level=router_args.log_level,
519
520
521
522
            service_discovery=router_args.service_discovery,
            selector=router_args.selector,
            service_discovery_port=router_args.service_discovery_port,
            service_discovery_namespace=router_args.service_discovery_namespace,
523
524
            prefill_selector=router_args.prefill_selector,
            decode_selector=router_args.decode_selector,
525
526
            prometheus_port=router_args.prometheus_port,
            prometheus_host=router_args.prometheus_host,
527
            request_timeout_secs=router_args.request_timeout_secs,
528
            pd_disaggregation=router_args.pd_disaggregation,
529
            prefill_urls=(
530
                router_args.prefill_urls if router_args.pd_disaggregation else None
531
532
            ),
            decode_urls=(
533
                router_args.decode_urls if router_args.pd_disaggregation else None
534
            ),
535
536
537
538
539
540
541
542
543
544
            prefill_policy=(
                policy_from_str(router_args.prefill_policy)
                if router_args.prefill_policy
                else None
            ),
            decode_policy=(
                policy_from_str(router_args.decode_policy)
                if router_args.decode_policy
                else None
            ),
545
            request_id_headers=router_args.request_id_headers,
546
547
            max_concurrent_requests=router_args.max_concurrent_requests,
            cors_allowed_origins=router_args.cors_allowed_origins,
548
549
550
551
552
553
        )

        router.start()
        return router

    except Exception as e:
554
        logger.error(f"Error starting router: {e}")
555
        raise e
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575


class CustomHelpFormatter(
    argparse.RawDescriptionHelpFormatter, argparse.ArgumentDefaultsHelpFormatter
):
    """Custom formatter that preserves both description formatting and shows defaults"""

    pass


def parse_router_args(args: List[str]) -> RouterArgs:
    """Parse command line arguments and return RouterArgs instance."""
    parser = argparse.ArgumentParser(
        description="""SGLang Router - High-performance request distribution across worker nodes

Usage:
This launcher enables starting a router with individual worker instances. It is useful for
multi-node setups or when you want to start workers and router separately.

Examples:
576
  # Regular mode
577
  python -m sglang_router.launch_router --worker-urls http://worker1:8000 http://worker2:8000
578

579
  # PD disaggregated mode with same policy for both
580
  python -m sglang_router.launch_router --pd-disaggregation \\
581
582
583
    --prefill http://prefill1:8000 9000 --prefill http://prefill2:8000 none \\
    --decode http://decode1:8001 --decode http://decode2:8001 \\
    --policy cache_aware
584

585
586
587
588
589
590
  # PD mode with different policies for prefill and decode
  python -m sglang_router.launch_router --pd-disaggregation \\
    --prefill http://prefill1:8000 9000 --prefill http://prefill2:8000 none \\
    --decode http://decode1:8001 --decode http://decode2:8001 \\
    --prefill-policy cache_aware --decode-policy power_of_two

591
592
593
594
595
596
597
598
599
600
    """,
        formatter_class=CustomHelpFormatter,
    )

    RouterArgs.add_cli_args(parser, use_router_prefix=False)
    return RouterArgs.from_cli_args(parser.parse_args(args), use_router_prefix=False)


def main() -> None:
    router_args = parse_router_args(sys.argv[1:])
601
    launch_router(router_args)
602
603
604
605


if __name__ == "__main__":
    main()