adapt to sglang v0.5.2rc1 on dcu
This commit is contained in:
maxiao
211
sgl-router/py_src/sglang_router/router.py
Normal file
211
sgl-router/py_src/sglang_router/router.py
Normal file
@@ -0,0 +1,211 @@
|
||||
from typing import Dict, List, Optional
|
||||
|
||||
from sglang_router_rs import PolicyType
|
||||
from sglang_router_rs import Router as _Router
|
||||
|
||||
|
||||
class Router:
|
||||
"""
|
||||
A high-performance router for distributing requests across worker nodes.
|
||||
|
||||
Args:
|
||||
worker_urls: List of URLs for worker nodes that will handle requests. Each URL should include
|
||||
the protocol, host, and port (e.g., ['http://worker1:8000', 'http://worker2:8000'])
|
||||
policy: Load balancing policy to use. Options:
|
||||
- PolicyType.Random: Randomly select workers
|
||||
- PolicyType.RoundRobin: Distribute requests in round-robin fashion
|
||||
- PolicyType.CacheAware: Distribute requests based on cache state and load balance
|
||||
- PolicyType.PowerOfTwo: Select best of two random workers based on load (PD mode only)
|
||||
host: Host address to bind the router server. Default: '127.0.0.1'
|
||||
port: Port number to bind the router server. Default: 3001
|
||||
worker_startup_timeout_secs: Timeout in seconds for worker startup. Default: 300
|
||||
worker_startup_check_interval: Interval in seconds between checks for worker initialization. Default: 10
|
||||
cache_threshold: Cache threshold (0.0-1.0) for cache-aware routing. Routes to cached worker
|
||||
if the match rate exceeds threshold, otherwise routes to the worker with the smallest
|
||||
tree. Default: 0.5
|
||||
balance_abs_threshold: Load balancing is triggered when (max_load - min_load) > abs_threshold
|
||||
AND max_load > min_load * rel_threshold. Otherwise, use cache aware. Default: 32
|
||||
balance_rel_threshold: Load balancing is triggered when (max_load - min_load) > abs_threshold
|
||||
AND max_load > min_load * rel_threshold. Otherwise, use cache aware. Default: 1.0001
|
||||
eviction_interval_secs: Interval in seconds between cache eviction operations in cache-aware
|
||||
routing. Default: 60
|
||||
max_payload_size: Maximum payload size in bytes. Default: 256MB
|
||||
max_tree_size: Maximum size of the approximation tree for cache-aware routing. Default: 2^24
|
||||
dp_aware: Enable data parallelism aware schedule. Default: False
|
||||
api_key: The api key used for the authorization with the worker.
|
||||
Useful when the dp aware scheduling strategy is enabled.
|
||||
Default: None
|
||||
log_dir: Directory to store log files. If None, logs are only output to console. Default: None
|
||||
log_level: Logging level. Options: 'debug', 'info', 'warning', 'error', 'critical'.
|
||||
service_discovery: Enable Kubernetes service discovery. When enabled, the router will
|
||||
automatically discover worker pods based on the selector. Default: False
|
||||
selector: Dictionary mapping of label keys to values for Kubernetes pod selection.
|
||||
Example: {"app": "sglang-worker"}. Default: {}
|
||||
service_discovery_port: Port to use for service discovery. The router will generate
|
||||
worker URLs using this port. Default: 80
|
||||
service_discovery_namespace: Kubernetes namespace to watch for pods. If not provided,
|
||||
watches pods across all namespaces (requires cluster-wide permissions). Default: None
|
||||
prefill_selector: Dictionary mapping of label keys to values for Kubernetes pod selection
|
||||
for prefill servers (PD mode only). Default: {}
|
||||
decode_selector: Dictionary mapping of label keys to values for Kubernetes pod selection
|
||||
for decode servers (PD mode only). Default: {}
|
||||
prometheus_port: Port to expose Prometheus metrics. Default: None
|
||||
prometheus_host: Host address to bind the Prometheus metrics server. Default: None
|
||||
pd_disaggregation: Enable PD (Prefill-Decode) disaggregated mode. Default: False
|
||||
prefill_urls: List of (url, bootstrap_port) tuples for prefill servers (PD mode only)
|
||||
decode_urls: List of URLs for decode servers (PD mode only)
|
||||
prefill_policy: Specific load balancing policy for prefill nodes (PD mode only).
|
||||
If not specified, uses the main policy. Default: None
|
||||
decode_policy: Specific load balancing policy for decode nodes (PD mode only).
|
||||
If not specified, uses the main policy. Default: None
|
||||
request_id_headers: List of HTTP headers to check for request IDs. If not specified,
|
||||
uses common defaults: ['x-request-id', 'x-correlation-id', 'x-trace-id', 'request-id'].
|
||||
Example: ['x-my-request-id', 'x-custom-trace-id']. Default: None
|
||||
bootstrap_port_annotation: Kubernetes annotation name for bootstrap port (PD mode).
|
||||
Default: 'sglang.ai/bootstrap-port'
|
||||
request_timeout_secs: Request timeout in seconds. Default: 600
|
||||
max_concurrent_requests: Maximum number of concurrent requests allowed for rate limiting. Default: 256
|
||||
queue_size: Queue size for pending requests when max concurrent limit reached (0 = no queue, return 429 immediately). Default: 100
|
||||
queue_timeout_secs: Maximum time (in seconds) a request can wait in queue before timing out. Default: 60
|
||||
rate_limit_tokens_per_second: Token bucket refill rate (tokens per second). If not set, defaults to max_concurrent_requests. Default: None
|
||||
cors_allowed_origins: List of allowed origins for CORS. Empty list allows all origins. Default: []
|
||||
health_failure_threshold: Number of consecutive health check failures before marking worker unhealthy. Default: 3
|
||||
health_success_threshold: Number of consecutive health check successes before marking worker healthy. Default: 2
|
||||
health_check_timeout_secs: Timeout in seconds for health check requests. Default: 5
|
||||
health_check_interval_secs: Interval in seconds between runtime health checks. Default: 60
|
||||
health_check_endpoint: Health check endpoint path. Default: '/health'
|
||||
model_path: Model path for loading tokenizer (HuggingFace model ID or local path). Default: None
|
||||
tokenizer_path: Explicit tokenizer path (overrides model_path tokenizer if provided). Default: None
|
||||
"""
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
worker_urls: List[str],
|
||||
policy: PolicyType = PolicyType.RoundRobin,
|
||||
host: str = "127.0.0.1",
|
||||
port: int = 3001,
|
||||
worker_startup_timeout_secs: int = 600,
|
||||
worker_startup_check_interval: int = 30,
|
||||
cache_threshold: float = 0.3,
|
||||
balance_abs_threshold: int = 64,
|
||||
balance_rel_threshold: float = 1.5,
|
||||
eviction_interval_secs: int = 120,
|
||||
max_tree_size: int = 2**26,
|
||||
max_payload_size: int = 512 * 1024 * 1024, # 512MB
|
||||
dp_aware: bool = False,
|
||||
api_key: Optional[str] = None,
|
||||
log_dir: Optional[str] = None,
|
||||
log_level: Optional[str] = None,
|
||||
service_discovery: bool = False,
|
||||
selector: Dict[str, str] = None,
|
||||
service_discovery_port: int = 80,
|
||||
service_discovery_namespace: Optional[str] = None,
|
||||
prefill_selector: Dict[str, str] = None,
|
||||
decode_selector: Dict[str, str] = None,
|
||||
bootstrap_port_annotation: str = "sglang.ai/bootstrap-port",
|
||||
prometheus_port: Optional[int] = None,
|
||||
prometheus_host: Optional[str] = None,
|
||||
request_timeout_secs: int = 1800,
|
||||
request_id_headers: Optional[List[str]] = None,
|
||||
pd_disaggregation: bool = False,
|
||||
prefill_urls: Optional[List[tuple]] = None,
|
||||
decode_urls: Optional[List[str]] = None,
|
||||
prefill_policy: Optional[PolicyType] = None,
|
||||
decode_policy: Optional[PolicyType] = None,
|
||||
max_concurrent_requests: int = 256,
|
||||
queue_size: int = 100,
|
||||
queue_timeout_secs: int = 60,
|
||||
rate_limit_tokens_per_second: Optional[int] = None,
|
||||
cors_allowed_origins: List[str] = None,
|
||||
retry_max_retries: int = 5,
|
||||
retry_initial_backoff_ms: int = 50,
|
||||
retry_max_backoff_ms: int = 30_000,
|
||||
retry_backoff_multiplier: float = 1.5,
|
||||
retry_jitter_factor: float = 0.2,
|
||||
cb_failure_threshold: int = 10,
|
||||
cb_success_threshold: int = 3,
|
||||
cb_timeout_duration_secs: int = 60,
|
||||
cb_window_duration_secs: int = 120,
|
||||
disable_retries: bool = False,
|
||||
disable_circuit_breaker: bool = False,
|
||||
health_failure_threshold: int = 3,
|
||||
health_success_threshold: int = 2,
|
||||
health_check_timeout_secs: int = 5,
|
||||
health_check_interval_secs: int = 60,
|
||||
health_check_endpoint: str = "/health",
|
||||
model_path: Optional[str] = None,
|
||||
tokenizer_path: Optional[str] = None,
|
||||
):
|
||||
if selector is None:
|
||||
selector = {}
|
||||
if prefill_selector is None:
|
||||
prefill_selector = {}
|
||||
if decode_selector is None:
|
||||
decode_selector = {}
|
||||
if cors_allowed_origins is None:
|
||||
cors_allowed_origins = []
|
||||
|
||||
self._router = _Router(
|
||||
worker_urls=worker_urls,
|
||||
policy=policy,
|
||||
host=host,
|
||||
port=port,
|
||||
worker_startup_timeout_secs=worker_startup_timeout_secs,
|
||||
worker_startup_check_interval=worker_startup_check_interval,
|
||||
cache_threshold=cache_threshold,
|
||||
balance_abs_threshold=balance_abs_threshold,
|
||||
balance_rel_threshold=balance_rel_threshold,
|
||||
eviction_interval_secs=eviction_interval_secs,
|
||||
max_tree_size=max_tree_size,
|
||||
max_payload_size=max_payload_size,
|
||||
dp_aware=dp_aware,
|
||||
api_key=api_key,
|
||||
log_dir=log_dir,
|
||||
log_level=log_level,
|
||||
service_discovery=service_discovery,
|
||||
selector=selector,
|
||||
service_discovery_port=service_discovery_port,
|
||||
service_discovery_namespace=service_discovery_namespace,
|
||||
prefill_selector=prefill_selector,
|
||||
decode_selector=decode_selector,
|
||||
bootstrap_port_annotation=bootstrap_port_annotation,
|
||||
prometheus_port=prometheus_port,
|
||||
prometheus_host=prometheus_host,
|
||||
request_timeout_secs=request_timeout_secs,
|
||||
request_id_headers=request_id_headers,
|
||||
pd_disaggregation=pd_disaggregation,
|
||||
prefill_urls=prefill_urls,
|
||||
decode_urls=decode_urls,
|
||||
prefill_policy=prefill_policy,
|
||||
decode_policy=decode_policy,
|
||||
max_concurrent_requests=max_concurrent_requests,
|
||||
queue_size=queue_size,
|
||||
queue_timeout_secs=queue_timeout_secs,
|
||||
rate_limit_tokens_per_second=rate_limit_tokens_per_second,
|
||||
cors_allowed_origins=cors_allowed_origins,
|
||||
retry_max_retries=retry_max_retries,
|
||||
retry_initial_backoff_ms=retry_initial_backoff_ms,
|
||||
retry_max_backoff_ms=retry_max_backoff_ms,
|
||||
retry_backoff_multiplier=retry_backoff_multiplier,
|
||||
retry_jitter_factor=retry_jitter_factor,
|
||||
cb_failure_threshold=cb_failure_threshold,
|
||||
cb_success_threshold=cb_success_threshold,
|
||||
cb_timeout_duration_secs=cb_timeout_duration_secs,
|
||||
cb_window_duration_secs=cb_window_duration_secs,
|
||||
disable_retries=disable_retries,
|
||||
disable_circuit_breaker=disable_circuit_breaker,
|
||||
health_failure_threshold=health_failure_threshold,
|
||||
health_success_threshold=health_success_threshold,
|
||||
health_check_timeout_secs=health_check_timeout_secs,
|
||||
health_check_interval_secs=health_check_interval_secs,
|
||||
health_check_endpoint=health_check_endpoint,
|
||||
model_path=model_path,
|
||||
tokenizer_path=tokenizer_path,
|
||||
)
|
||||
|
||||
def start(self) -> None:
|
||||
"""Start the router server.
|
||||
|
||||
This method blocks until the server is shut down.
|
||||
"""
|
||||
self._router.start()
|
||||
Reference in New Issue
Block a user