Skip to content

GitLab

Unverified Commit 49345a68 authored by Simo Lin's avatar Simo Lin Committed by GitHub
Browse files

[router] update router readme to latest features (#11619)

parent 94d26d85
Hide whitespace changes
Inline Side-by-side
Showing with 393 additions and 503 deletions
sgl-router/README.md
View file @ 49345a68
# SGLang Router
SGLang router is a standalone Rust module that enables data parallelism across SGLang instances, providing high-performance request routing and advanced load balancing. The router supports multiple load balancing algorithms including cache-aware, power of two, random, and round robin, and acts as a specialized load balancer for prefill-decode disaggregated serving architectures.
# SGLang Model Gateway
High-performance model routing control and data plane for large-scale LLM deployments. The gateway orchestrates fleets of workers, balances traffic across HTTP and gRPC backends, and exposes OpenAI-compatible APIs with pluggable history storage and tool integrations—while remaining deeply optimized for the SGLang serving runtime.
## Overview
- Unified control plane for registering, monitoring, and orchestrating prefill, decode, and regular workers across heterogeneous model fleets.
- Data plane that routes requests across HTTP, PD (prefill/decode), gRPC, and OpenAI-compatible backends with shared reliability features.
- Industry-first gRPC pipeline with native Rust tokenization, reasoning, and tool-call execution for high-throughput OpenAI-compatible serving.
- Multi-model inference gateway mode (`--enable-igw`) that runs several routers at once and applies per-model policies.
- Conversation, response, and chat-history connectors that centralize state at the router, enabling compliant sharing across models/MCP loops with in-memory, no-op, or Oracle ATP storage options.
- Built-in reliability primitives: retries with exponential backoff, circuit breakers, token-bucket rate limiting, and queuing.
- First-class observability with structured logging and Prometheus metrics.
### Architecture at a Glance
**Control Plane**
- Worker Manager validates workers, discovers capabilities, and keeps the registry in sync.
- Job Queue serializes background operations (add/remove) and exposes status via `/workers/{url}`.
- Background health checker and load monitor keep circuit breakers and policies informed.
- Optional Kubernetes service discovery keeps the registry aligned with pods.
**Data Plane**
- SGLang HTTP routers for regular and PD (prefill/decode) traffic with policy-aware selection.
- SGLang gRPC router and pipeline that stream tokenized requests through SRT gRPC workers with fully Rust tokenizer, reasoning parser, and tool parser implementations for maximal OpenAI API performance, supporting both single-stage and PD serving topologies.
- OpenAI router that proxies OpenAI-style requests, responses, and conversations to remote vendors (OpenAI, xAI, Gemini, and other OpenAI-compatible providers) while preserving streaming/SSE semantics.
- Router Manager coordinates multiple router implementations when IGW is enabled.
- Resilience layer delivers token-bucket rate limiting, request queuing, retry executor, and per-worker circuit breakers to keep traffic flowing through failures.
- Advanced load balancing with cache-aware request reuse, load-aware (power-of-two) selection, and per-model policy overrides.
## Feature Highlights
- Multiple load balancing strategies (`random`, `round_robin`, `cache_aware`, `power_of_two`) with DP-aware scheduling.
- Multi-model HTTP serving and inference gateway routing with model-specific policies.
- Prefill/decode disaggregation, including bootstrap port handling and cache-aware merging.
- gRPC routing with fully Rust tokenizer loading, reasoning parser selection, and tool parser integration for OpenAI-compatible endpoints—supporting streaming and non-streaming modes across DeepSeek, Llama, Kimi K2, Qwen, GPT-OSS, Mistral, Step-3, GLM4, and other reasoning-capable models.
- OpenAI-compatible `/v1/chat/completions`, `/v1/responses`, `/v1/conversations`, `/v1/embeddings`, and `/v1/rerank` endpoints.
- Native MCP client integration supporting all MCP transport protocols (STDIO, HTTP, SSE, and Streamable) for tool execution loops.
- Pluggable history connectors: in-memory, disabled, or Oracle ATP (with pooling and credential support).
- Reliability controls: retry with jitter, worker-scoped circuit breakers, token bucket limiter with optional queue, and cache flush APIs.
- Service discovery for regular and PD workloads with independent selectors.
- Prometheus metrics and structured tracing for every stage of routing.
## Documentation
- **User Guide**: [docs.sglang.ai/advanced_features/router.html](https://docs.sglang.ai/advanced_features/router.html)
- Additional guides, API references, and deployment patterns are continuously updated alongside SGLang releases.
## Quick Start
## Installation
### Prerequisites
- **Rust and Cargo**
```bash
# Install rustup (Rust installer and version manager)
curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh
**Rust and Cargo:**
```bash
# Install rustup (Rust installer and version manager)
curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh
# Reload shell environment
source "$HOME/.cargo/env"
# Follow the installation prompts, then reload your shell
source $HOME/.cargo/env
# Verify installation
rustc --version
cargo --version
```
- **Python** with `pip` and virtualenv tooling available.
# Verify installation
rustc --version
cargo --version
### Rust Binary
```bash
# Build release binary
cargo build --release
```
**Python with pip installed**
### Installation
#### Option A: Build and Install Wheel (Recommended)
### Python Package
```bash
# Install build dependencies
pip install setuptools-rust wheel build
# Build the wheel package
python -m build
# Install the generated wheel
pip install dist/*.whl
# One-liner for development (rebuild + install)
# Rebuild & reinstall in one step during development
python -m build && pip install --force-reinstall dist/*.whl
```
> **Note:** Editable installs (`pip install -e .`) are currently not supported; prefer wheel builds for development.
#### Option B: Development Mode
```bash
# Currently broken
pip install -e .
```
⚠️ **Warning**: Editable installs may suffer performance degradation. Use wheel builds for performance testing.
### Basic Usage
```bash
# Build Rust components
cargo build
```
## Quick Start
#### Using the Rust Binary Directly (Alternative to Python)
### Regular HTTP Routing
- **Rust binary**
```bash
./target/release/sglang-router \
--worker-urls http://worker1:8000 http://worker2:8000 \
--policy cache_aware
```
`cargo run --release -- …` provides the same behavior during development.
- **Python launcher**
```bash
python3 -m sglang_router.launch_router \
--worker-urls http://worker1:8000 http://worker2:8000 \
--policy cache_aware
```
### Prefill/Decode Disaggregation (PD)
- **Rust binary**
```bash
./target/release/sglang-router \
--pd-disaggregation \
--prefill http://prefill1:30001 9001 \
--prefill http://prefill2:30002 \
--decode http://decode1:30011 \
--decode http://decode2:30012 \
--policy cache_aware \
--prefill-policy cache_aware \
--decode-policy power_of_two
```
- **Python launcher**
```bash
python3 -m sglang_router.launch_router \
--pd-disaggregation \
--prefill http://prefill1:30001 9001 \
--prefill http://prefill2:30002 \
--decode http://decode1:30011 \
--decode http://decode2:30012 \
--policy cache_aware
```
Prefill entries accept an optional bootstrap port. PD mode merges prefill metadata with decode outputs and streams results back to the client.
### Multi-Model Inference Gateway
Enable IGW mode to route multiple models through a single router while applying per-model policies:
```bash
# Build the Rust binary
cargo build --release
# Launch router with worker URLs in regular mode
./target/release/sglang-router \
--worker-urls http://worker1:8000 http://worker2:8000
# Or use cargo run
cargo run --release -- \
--worker-urls http://worker1:8000 http://worker2:8000
--enable-igw \
--policy cache_aware \
--max-concurrent-requests 512
# Register workers dynamically
curl -X POST http://localhost:30000/workers \
-H "Content-Type: application/json" \
-d '{
"url": "http://worker-a:8000",
"model_id": "mistral",
"priority": 10,
"labels": {"tier": "gold"}
}'
# Add another worker with a different model/policy hint
curl -X POST http://localhost:30000/workers \
-H "Content-Type: application/json" \
-d '{
"url": "http://worker-b:8000",
"model_id": "llama3",
"priority": 20,
"labels": {"policy": "power_of_two", "tier": "silver"}
}'
# Inspect registered workers
curl http://localhost:30000/workers | jq
```
#### Launch Router with Python (Original Method)
```bash
# Launch router with worker URLs
python -m sglang_router.launch_router \
--worker-urls http://worker1:8000 http://worker2:8000
Sample response (http workers):
```json
{
"workers": [
{"id":"http://0.0.0.0:31378","url":"http://0.0.0.0:31378","model_id":"mistral","priority":50,"cost":1.0,"worker_type":"regular","is_healthy":true,"load":0,"connection_mode":"Http"},
{"id":"http://0.0.0.0:34881","url":"http://0.0.0.0:34881","model_id":"llama3","priority":50,"cost":1.0,"worker_type":"regular","is_healthy":true,"load":0,"connection_mode":"Http"}
],
"total": 2,
"stats": {
"prefill_count": 0,
"decode_count": 0,
"regular_count": 2
}
}
```
Add more workers with the same API; include optional `labels` (for per-model policies) or `tokenizer_path` / `reasoning_parser` / `tool_parser` fields as needed. `/workers/{url}` exposes queued job status while background jobs finalize registration.
### gRPC Routing
- **Rust binary**
```bash
./target/release/sglang-router \
--worker-urls grpc://worker-grpc-0:31001 grpc://worker-grpc-1:31002 \
--tokenizer-path /path/to/tokenizer.json \
--reasoning-parser deepseek-r1 \
--tool-call-parser json
```
- **Python router**
```bash
python3 -m sglang_router.launch_router \
--worker-urls grpc://127.0.0.1:20000 \
--model-path meta-llama/Llama-3.1-8B-Instruct \
--host 0.0.0.0 \
--port 8080
```
The gRPC router tokenizes inputs locally, supports tool-call parsing, and streams responses. It supports both regular HTTP-equivalent serving and PD (prefill/decode) serving when the worker registry contains PD workers. Provide `--model-path` or `--tokenizer-path` (HuggingFace ID or local directory) whenever connection mode resolves to gRPC.
Use `--reasoning-parser` to select built-in reasoning pipelines (DeepSeek-R1, Qwen3, Step-3, GLM4, etc.) and `--tool-call-parser` for JSON/Pythonic/XML tool contracts in streaming or non-streaming modes.
#### OpenAI Backend Mode
### OpenAI Backend Mode
Route requests to OpenAI or OpenAI-compatible endpoints:
```bash
# Route to OpenAI API
python -m sglang_router.launch_router \
--backend openai \
--worker-urls https://api.openai.com
# Route to custom OpenAI-compatible endpoint
python -m sglang_router.launch_router \
--backend openai \
--worker-urls http://my-openai-compatible-service:8000
python3 -m sglang_router.launch_router \
--backend openai \
--worker-urls https://api.openai.com \
--api-key "$OPENAI_API_KEY"
# Route to custom OpenAI-compatible endpoint (Gemini, xAI, etc.)
python3 -m sglang_router.launch_router \
--backend openai \
--worker-urls http://my-openai-compatible-service:8000 \
--api-key "tenant-api-key"
```
**Note**:
- OpenAI backend mode acts as a simple proxy. Load balancing is not applicable in this mode.
- This mode now requires exactly one `--worker-urls` entry.
**Notes**
- OpenAI backend mode acts as a proxy to a single remote endpoint; load balancing is not applied.
- Provide exactly one `--worker-urls` entry per router instance.
- The Rust binary supports the same flags (`./target/release/sglang-router --backend openai ...`).
#### Launch Router with Worker URLs in prefill-decode mode
### Python Launcher (Router + Workers)
Launch router and SGLang worker processes together; `launch_server` spins up workers (HTTP or gRPC) and the router in one shot.
```bash
# Note that the prefill and decode URLs must be provided in the following format:
# http://<ip>:<port> for decode nodes
# http://<ip>:<port> bootstrap-port for prefill nodes, where bootstrap-port is optional
# Using Rust binary directly
./target/release/sglang-router \
--pd-disaggregation \
--policy cache_aware \
--prefill http://127.0.0.1:30001 9001 \
--prefill http://127.0.0.2:30002 9002 \
--prefill http://127.0.0.3:30003 9003 \
--prefill http://127.0.0.4:30004 9004 \
--decode http://127.0.0.5:30005 \
--decode http://127.0.0.6:30006 \
--decode http://127.0.0.7:30007 \
--host 0.0.0.0 \
--port 8080
# Or using Python launcher
python -m sglang_router.launch_router \
--pd-disaggregation \
--policy cache_aware \
--prefill http://127.0.0.1:30001 9001 \
--prefill http://127.0.0.2:30002 9002 \
--prefill http://127.0.0.3:30003 9003 \
--prefill http://127.0.0.4:30004 9004 \
--decode http://127.0.0.5:30005 \
--decode http://127.0.0.6:30006 \
--decode http://127.0.0.7:30007 \
--host 0.0.0.0 \
--port 8080
````
## Configuration
### Logging
Enable structured logging with optional file output:
```python
from sglang_router import Router
# Console logging (default)
router = Router(worker_urls=["http://worker1:8000", "http://worker2:8000"])
# File logging enabled
router = Router(
worker_urls=["http://worker1:8000", "http://worker2:8000"],
log_dir="./logs" # Daily log files created here
)
python3 -m sglang_router.launch_server --host 0.0.0.0
```
Set log level with `--log-level` flag ([documentation](https://docs.sglang.ai/backend/server_arguments.html#logging)).
### Metrics
Prometheus metrics endpoint available at `127.0.0.1:29000` by default.
Add flags as needed for production deployments:
```bash
# Custom metrics configuration
python -m sglang_router.launch_router \
--worker-urls http://localhost:8080 http://localhost:8081 \
--prometheus-host 0.0.0.0 \
--prometheus-port 9000
python3 -m sglang_router.launch_server \
--host 0.0.0.0 \
--port 8080 \
--model /raid/models/meta-llama/Llama-3.1-8B-Instruct \
--tp-size 1 \
--dp-size 8 \
--grpc-mode
```
Omit `--grpc-mode` to start HTTP workers; the router automatically configures worker URLs and schedules them based on the provided DP size.
### Retries and Circuit Breakers
- Retries (regular router) are enabled by default with exponential backoff and jitter. You can tune them via CLI:
### Mini Load Balancer (Debug)
```bash
python -m sglang_router.launch_router \
--worker-urls http://localhost:8080 http://localhost:8081 \
--retry-max-retries 3 \
--retry-initial-backoff-ms 100 \
--retry-max-backoff-ms 10000 \
--retry-backoff-multiplier 2.0 \
--retry-jitter-factor 0.1
python3 -m sglang_router.launch_router \
--mini-lb \
--pd-disaggregation \
--prefill http://localhost:30001 \
--decode http://localhost:30011
```
- Circuit Breaker defaults protect workers and auto-recover. Tune thresholds/timeouts:
MiniLB forwards PD requests using simple random routing and is intended for local debugging only.
### Running Worker Servers
Use upstream SGLang binaries to start dedicated worker processes.
- **Prefill worker server (gRPC mode)**:
```bash
python3 -m sglang.launch_server \
--model /raid/models/meta-llama/Llama-3.1-8B-Instruct \
--port 20000 \
--tp-size 1 \
--grpc-mode
```
Remove `--grpc-mode` for HTTP workers. Combine with the router commands above to register the worker via CLI flags or the control-plane API.
## Control Plane
### Worker Lifecycle & Job Queue
- `JobQueue` handles asynchronous add/remove operations to avoid blocking clients.
- `WorkerManager` inspects worker metadata (`/get_server_info`, `/get_model_info`), tracks load, and exposes `flush_cache` and `get_loads`.
- Per-worker circuit breakers and health probes keep the registry healthy; load monitor feeds metrics to cache-aware and power-of-two policies.
### Administrative & Worker APIs
| Method | Path | Description |
|----------|------------------|-----------------------------------------------------------------------------------------------------------------------------------------------------------|
| `POST` | `/workers` | Queue worker registration (prefill/decode/regular). Body matches `WorkerConfigRequest`. Returns `202 Accepted` while the job queue processes the request. |
| `GET` | `/workers` | List workers with health, load, policy metadata, and queued job status. |
| `GET` | `/workers/{url}` | Inspect a specific worker or job queue entry. |
| `DELETE` | `/workers/{url}` | Queue worker removal. |
| `POST` | `/add_worker` | Legacy immediate worker registration using query params. Returns synchronously. **Deprecated soon**—use `POST /workers` instead. |
| `POST` | `/remove_worker` | Legacy immediate removal. **Deprecated soon**—use `DELETE /workers/{url}` instead. |
| `GET` | `/list_workers` | Legacy list of worker URLs. **Deprecated soon**—use `GET /workers` instead. |
| `POST` | `/flush_cache` | Trigger cache flush across HTTP workers with success/failure breakdown. |
| `GET` | `/get_loads` | Sample current load reported by each worker. |
All administrative routes inherit router API-key protection when `--api-key` is supplied. Job status includes `pending`, `processing`, and `failed` phases with timestamps.
### Service Discovery
Enable Kubernetes discovery to reconcile workers automatically:
```bash
python -m sglang_router.launch_router \
--worker-urls http://localhost:8080 http://localhost:8081 \
--cb-failure-threshold 5 \
--cb-success-threshold 2 \
--cb-timeout-duration-secs 30 \
--cb-window-duration-secs 60
./target/release/sglang-router \
--service-discovery \
--selector app=sglang-worker role=inference \
--service-discovery-namespace sglang-system \
--service-discovery-port 8000
```
Behavior summary:
- Closed → Open after N consecutive failures (failure-threshold)
- Open → HalfOpen after timeout (timeout-duration-secs)
- HalfOpen → Closed after M consecutive successes (success-threshold)
- Any failure in HalfOpen reopens immediately
Retry predicate (regular router): retry on 408/429/500/502/503/504, otherwise return immediately. Backoff/jitter observed between attempts.
### Request ID Tracking
Track requests across distributed systems with configurable headers:
PD mode accepts dedicated selectors:
```bash
# Use custom request ID headers
python -m sglang_router.launch_router \
--worker-urls http://localhost:8080 \
--request-id-headers x-trace-id x-request-id
--pd-disaggregation \
--prefill-selector app=sglang component=prefill \
--decode-selector app=sglang component=decode \
--service-discovery
```
Default headers: `x-request-id`, `x-correlation-id`, `x-trace-id`, `request-id`
### History Backend (Conversation Storage)
Prefill pods can expose bootstrap ports via the `sglang.ai/bootstrap-port` annotation. RBAC must allow `get`, `list`, and `watch` on pods.
## Data Plane
### Router Capabilities (HTTP & gRPC)
Both router stacks:
- Share load-balancing policies (random, round-robin, cache-aware, power-of-two) with DP-aware scheduling, retries, circuit breakers, and rate limiting.
- Record metrics per request, track running load, and integrate with the router-wide policy registry.
The HTTP router exposes the full OpenAI-compatible surface area (`/generate`, `/v1/chat/completions`, `/v1/completions`, `/v1/embeddings`, `/v1/responses`, `/v1/rerank`, etc.). The gRPC router delivers blazing-fast `/generate` and `/v1/chat/completions` today, with the remaining endpoints returning `501 Not Implemented` until their pipelines are finalised.
#### HTTP Router specifics
- **Regular router** handles classic single-stage workers with per-model policy overrides.
- **Prefill/Decode router** coordinates disaggregated prefill and decode workers, merges metadata, and manages streaming fan-in.
#### gRPC Router specifics
- Industry-first fully Rust implementation of an OpenAI-compatible gRPC inference gateway, including tokenizer, reasoning parser, and tool parser execution in-process for maximum throughput.
- Supports both single-stage and PD (prefill/decode) worker topologies; the router automatically selects the appropriate pipeline per model.
- Provides the same `/v1/*` APIs as the HTTP router while streaming tokenized requests/responses directly to SRT gRPC workers.
- Built-in reasoning parsers for DeepSeek, Qwen, Llama, Mistral, GPT-OSS, Step-3, GLM4, Kimi K2, and other structured-thought models.
- Tool-call parsers for JSON, Pythonic, XML, and custom schemas with streaming and non-streaming execution loops.
- Tokenizer factory supporting HuggingFace models, local tokenizer.json files, and chat template overrides (see `src/tokenizer`).
- Explore the code paths in `src/reasoning_parser`, `src/tool_parser`, and `src/tokenizer` for the end-to-end Rust implementations that power gRPC mode.
### OpenAI Router
- Proxies OpenAI-compatible chat completions and responses APIs, preserving headers and SSE streams end-to-end.
- Supports `/v1/responses` background jobs with cancellation, deletion, and listing input items—enabling agentic, multi-turn orchestration without persisting data at remote vendor endpoints.
- Conversation APIs (`/v1/conversations` and `/items`) interact with the configured conversation storage backend for compliant chat-history management. Conversation state lives at the router tier, so the same history can drive different models or MCP loops without leaking data to upstream vendors.
- Chat history, agentic multi-turn `/v1/responses`, and the native MCP client (STDIO/HTTP/SSE/Streamable transports) are designed to satisfy enterprise data-privacy requirements by keeping sensitive state within the router.
### Request Endpoints
| Endpoint | Notes |
|----------------------------------------------------------------------------------|------------------------------------------------------------|
| `POST /generate` | SGLang generate API. |
| `POST /v1/chat/completions` | OpenAI-compatible chat. Supports streaming and tool calls. |
| `POST /v1/completions` | OpenAI-compatible text completions. |
| `POST /v1/responses` | Create background responses, returns response IDs. |
| `GET /v1/responses/{id}` | Retrieve stored responses. |
| `POST /v1/responses/{id}/cancel` | Cancel in-flight background jobs. |
| `DELETE /v1/responses/{id}` | Delete stored response. |
| `GET /v1/responses/{id}/input` | List captured input items. |
| Conversation endpoints (`/v1/conversations`, `/v1/conversations/{id}`, `/items`) | Manage chat history. |
| `POST /v1/embeddings` | Forward embedding requests. |
| `POST /v1/rerank`, `POST /rerank` | Ranking APIs. |
Public health endpoints (`/liveness`, `/readiness`, `/health`, `/health_generate`) reflect registry state; readiness ensures PD workers are paired and IGW has at least one healthy route.
## Conversations, Responses, and Data Connectors
- `--history-backend memory` (default) stores responses and conversations in-process.
- `--history-backend none` disables persistence while keeping APIs.
- `--history-backend oracle` uses Oracle Autonomous Database; provide credentials via flags or environment variables.
- Conversation item storage mirrors the history backend (Oracle or memory). The same storage powers OpenAI `/responses` and conversation APIs.
### History Backend (OpenAI Router Mode)
Store conversation and response data for tracking, debugging, or analytics.
**NOTE: This feature is currently only supported in OpenAI router mode.**
#### Available options for history storage
> **Note:** History backends are currently supported only when running with `--backend openai`. gRPC mode support for the `/v1/responses` API is planned.
- **Memory** (default): In-memory storage, fast but ephemeral
- **None**: No storage, minimal overhead
- **Oracle**: Persistent storage using Oracle ATP
#### Available storage options
- **Memory** (default): In-memory storage, fast but ephemeral.
- **None**: No storage, minimal overhead.
- **Oracle**: Persistent storage backed by Oracle Autonomous Database.
```bash
# Memory backend (default)
python -m sglang_router.launch_router --backend openai \
--worker-urls https://api.openai.com \
--history-backend memory
python3 -m sglang_router.launch_router \
--backend openai \
--worker-urls https://api.openai.com \
--history-backend memory
# No storage for maximum performance
python -m sglang_router.launch_router --backend openai \
--worker-urls https://api.openai.com \
--history-backend none
# Oracle ATP backend
# Install Oracle Instant Client
# https://download.oracle.com/otn_software/linux/instantclient/2390000/instantclient-basic-linux.x64-23.9.0.25.07.zip
export LD_LIBRARY_PATH=/home/ubuntu/instant-client/instantclient_23_9
python3 -m sglang_router.launch_router \
--backend openai \
--worker-urls https://api.openai.com \
--history-backend none
# Oracle ATP backend (see configuration below)
python3 -m sglang_router.launch_router \
--backend openai \
--worker-urls https://api.openai.com \
--history-backend oracle
```
# choose ONE of the following connection methods:
# Option 1: Using full connection descriptor
#### Oracle configuration
Install the Oracle Instant Client and set `LD_LIBRARY_PATH` accordingly. Choose **one** connection method:
```bash
# Option 1: Full connection descriptor
export ATP_DSN="(description=(address=(protocol=tcps)(port=1522)(host=adb.region.oraclecloud.com))(connect_data=(service_name=service_name)))"
# Option 2: Using TNS alias (requires wallet)
# Option 2: TNS alias (requires wallet)
export ATP_TNS_ALIAS="sglroutertestatp_high"
export ATP_WALLET_PATH="/path/to/wallet"
# service user config
export ATP_USER="admin"
export ATP_PASSWORD="YourPassword123"
python -m sglang_router.launch_router \
--worker-urls https://api.openai.com \
--backend openai \
--history-backend oracle
```
**Oracle Configuration Parameters:**
- `--oracle-tns-alias`: TNS alias from tnsnames.ora (env: `ATP_TNS_ALIAS`)
- Requires `--oracle-wallet-path` to locate tnsnames.ora
- `--oracle-username`: Database username (env: `ATP_USER`)
- `--oracle-password`: Database password (env: `ATP_PASSWORD`)
- `--oracle-wallet-path`: Path to wallet directory (env: `ATP_WALLET_PATH`)
- Required when using TNS alias
- `--oracle-pool-min`: Minimum connections (default: 1, env: `ATP_POOL_MIN`)
- `--oracle-pool-max`: Maximum connections (default: 16, env: `ATP_POOL_MAX`)
**Note**: You must provide **either** `--oracle-tns-alias` **or** `--oracle-connect-descriptor`, but not both.
## Advanced Features
### Kubernetes Service Discovery
Automatic worker discovery and management in Kubernetes environments.
#### Basic Service Discovery
```bash
python -m sglang_router.launch_router \
--service-discovery \
--selector app=sglang-worker role=inference \
--service-discovery-namespace default
```
#### PD (Prefill-Decode) Mode
For disaggregated prefill/decode routing:
Provide database credentials and optional pool sizing:
```bash
python -m sglang_router.launch_router \
--pd-disaggregation \
--policy cache_aware \
--service-discovery \
--prefill-selector app=sglang component=prefill \
--decode-selector app=sglang component=decode \
--service-discovery-namespace sglang-system
# With separate routing policies:
python -m sglang_router.launch_router \
--pd-disaggregation \
--prefill-policy cache_aware \
--decode-policy power_of_two \
--service-discovery \
--prefill-selector app=sglang component=prefill \
--decode-selector app=sglang component=decode \
--service-discovery-namespace sglang-system
# in lws case, such as tp16(1 leader pod, 1 worker pod)
python -m sglang_router.launch_router \
--pd-disaggregation \
--policy cache_aware \
--service-discovery \
--prefill-selector app=sglang component=prefill role=leader\
--decode-selector app=sglang component=decode role=leader\
--service-discovery-namespace sglang-system
```
#### Kubernetes Pod Configuration
**Prefill Server Pod:**
```yaml
apiVersion: v1
kind: Pod
metadata:
name: sglang-prefill-1
labels:
app: sglang
component: prefill
annotations:
sglang.ai/bootstrap-port: "9001" # Optional: Bootstrap port
spec:
containers:
- name: sglang
image: lmsys/sglang:latest
ports:
- containerPort: 8000 # Main API port
- containerPort: 9001 # Optional: Bootstrap port
```
**Decode Server Pod:**
```yaml
apiVersion: v1
kind: Pod
metadata:
name: sglang-decode-1
labels:
app: sglang
component: decode
spec:
containers:
- name: sglang
image: lmsys/sglang:latest
ports:
- containerPort: 8000
```
#### RBAC Configuration
**Namespace-scoped (recommended):**
```yaml
apiVersion: v1
kind: ServiceAccount
metadata:
name: sglang-router
namespace: sglang-system
---
apiVersion: rbac.authorization.k8s.io/v1
kind: Role
metadata:
namespace: sglang-system
name: sglang-router
rules:
- apiGroups: [""]
resources: ["pods"]
verbs: ["get", "list", "watch"]
---
apiVersion: rbac.authorization.k8s.io/v1
kind: RoleBinding
metadata:
name: sglang-router
namespace: sglang-system
subjects:
- kind: ServiceAccount
name: sglang-router
namespace: sglang-system
roleRef:
kind: Role
name: sglang-router
apiGroup: rbac.authorization.k8s.io
```
#### Complete PD Example
```bash
python -m sglang_router.launch_router \
--pd-disaggregation \
--policy cache_aware \
--service-discovery \
--prefill-selector app=sglang component=prefill environment=production \
--decode-selector app=sglang component=decode environment=production \
--service-discovery-namespace production \
--host 0.0.0.0 \
--port 8080 \
--prometheus-host 0.0.0.0 \
--prometheus-port 9090
```
### API Key Authentication
The router supports multi-level API key authentication for both the router itself and individual workers:
#### Router API Key
Protect access to the router endpoints:
```bash
python -m sglang_router.launch_router \
--api-key "your-router-api-key" \
--worker-urls http://worker1:8000 http://worker2:8000
```
When router API key is set, clients must include the Bearer token:
```bash
curl -H "Authorization: Bearer your-router-api-key" http://localhost:8080/v1/chat/completions
export ATP_USER="admin"
export ATP_PASSWORD="YourPassword123"
export ATP_POOL_MIN=4
export ATP_POOL_MAX=32
```
#### Worker API Keys
Workers can have their own API keys for authentication:
Router flags map to these values:
- `--oracle-dsn` (env: `ATP_DSN`) or `--oracle-tns-alias` with `--oracle-wallet-path`.
- `--oracle-user` / `--oracle-password` (`ATP_USER` / `ATP_PASSWORD`).
- `--oracle-wallet-path` (`ATP_WALLET_PATH`) when using TNS alias.
- `--oracle-pool-min`, `--oracle-pool-max`, `--oracle-pool-timeout-secs`.
Only one of `--oracle-dsn` or `--oracle-tns-alias` should be supplied.
## Reliability & Flow Control
- **Retries**: Default max retries = 5 with exponential backoff (`--retry-max-retries`, `--retry-initial-backoff-ms`, `--retry-max-backoff-ms`, `--retry-backoff-multiplier`, `--retry-jitter-factor`). Retries trigger on 408/429/500/502/503/504.
- **Circuit Breakers**: Per worker thresholds (`--cb-failure-threshold`, `--cb-success-threshold`, `--cb-timeout-duration-secs`, `--cb-window-duration-secs`). Disable via `--disable-circuit-breaker`.
- **Rate Limiting**: Token bucket driven by `--max-concurrent-requests`. Set `--rate-limit-tokens-per-second` to override refill rate. Configure request queue via `--queue-size` and `--queue-timeout-secs`; queued requests observe FIFO order and respect cancellation.
- **Health Checks**: Runtime probes via `--health-check-interval-secs`, `--health-check-timeout-secs`, failure/success thresholds, and `--health-check-endpoint`.
- **Cache Management**: `/flush_cache` ensures LRU eviction when redeploying PD workers.
## Load Balancing Policies
- `random`: uniform random worker selection.
- `round_robin`: sequential rotation with atomic counters.
- `cache_aware`: maintains a prefix tree of prompts to route repeat traffic and evens load with configurable thresholds (`--cache-threshold`, `--balance-abs-threshold`, `--balance-rel-threshold`, `--eviction-interval`, `--max-tree-size`).
- `power_of_two`: chooses the lighter worker among two random candidates; integrates with `LoadMonitor`.
Per-model overrides are available in PD mode (`--prefill-policy`, `--decode-policy`) and IGW mode via the worker registry.
## Observability
- **Logging**: Structured tracing through `tracing` with optional file sink (`--log-dir`) and `--log-level` (`debug`, `info`, `warn`, `error`).
- **Prometheus Metrics**: Enable with `--prometheus-host`/`--prometheus-port` (defaults to `0.0.0.0:29000`). Metrics cover request latency, retry behavior, circuit breaker states, worker health/load, queue depth, PD pipeline stats, tokenizer timings, and MCP activity.
- **Request IDs**: Configurable headers via `--request-id-headers`; responses include `x-request-id`.
- **CORS**: Set `--cors-allowed-origins` for browser access.
## Security
### Router and Worker API Keys
- **Router API key (`--api-key`)** protects client access to router endpoints; all protected routes expect `Authorization: Bearer <key>`.
- Workers listed in `--worker-urls` inherit the router API key automatically.
- When adding workers dynamically, provide explicit API keys via payload or query string; they do **not** inherit automatically.
```bash
# Workers specified in --worker-urls automatically inherit the router's API key
python -m sglang_router.launch_router \
--api-key "shared-api-key" \
--worker-urls http://worker1:8000 http://worker2:8000
# Both workers will use "shared-api-key" for authentication
# Router and initial workers share the same key
python3 -m sglang_router.launch_router \
--api-key "shared-api-key" \
--worker-urls http://worker1:8000 http://worker2:8000
# Adding workers dynamically WITHOUT inheriting router's key
# Adding a worker without key while router has one triggers a warning and leaves the worker unprotected
curl -X POST http://localhost:8080/add_worker?url=http://worker3:8000
# WARNING: This worker has NO API key even though router has one!
# Adding workers with specific API keys dynamically
curl -X POST http://localhost:8080/add_worker?url=http://worker3:8000&api_key=worker3-specific-key
# Add worker with explicit key
curl -X POST "http://localhost:8080/add_worker?url=http://worker3:8000&api_key=worker3-specific-key"
```
#### Security Configurations
1. **No Authentication** (default):
- Router and workers accessible without keys
- Suitable for trusted environments
2. **Router-only Authentication**:
- Clients need key to access router
- Router can access workers freely
3. **Worker-only Authentication**:
- Router accessible without key
- Each worker requires authentication
### Security Configurations
1. **No Authentication** (default): Router and workers accept requests without keys—use only in trusted environments.
2. **Router-only Authentication**: Provide `--api-key`; clients must present the key, router accesses workers without credentials.
3. **Worker-only Authentication**: Router open to clients; each worker requires its own key. Supply keys when calling `/workers` or `/add_worker`.
4. **Full Authentication**: Set router API key and provide per-worker keys. Example:
```bash
# Add workers with their API keys
curl -X POST http://localhost:8080/add_worker?url=http://worker:8000&api_key=worker-key
```
4. **Full Authentication**:
- Router requires key from clients
- Each worker requires its own key
```bash
# Start router with its key
python -m sglang_router.launch_router --api-key "router-key"
# Add workers with their keys
python3 -m sglang_router.launch_router --api-key "router-key"
curl -H "Authorization: Bearer router-key" \
-X POST http://localhost:8080/add_worker?url=http://worker:8000&api_key=worker-key
-X POST http://localhost:8080/add_worker?url=http://worker:8000&api_key=worker-key
```
#### Important Notes
- **Initial Workers**: Workers specified in `--worker-urls` automatically inherit the router's API key
- **Dynamic Workers**: When adding workers via API, you must explicitly specify their API keys - they do NOT inherit the router's key
- **Security Warning**: When adding workers without API keys while the router has one configured, a warning will be logged
- **Common Pitfall**: If router and workers use the same API key, you must still specify the key when adding workers dynamically
### Command Line Arguments Reference
#### Backend Selection
- `--backend`: Backend runtime to use (default: `sglang`)
- `sglang`: SGLang workers (default)
- `openai`: OpenAI or OpenAI-compatible endpoints
#### History Backend
- `--history-backend`: Storage backend for conversations (default: `memory`)
- `memory`: In-memory storage (default)
- `none`: No storage
- `oracle`: Oracle ATP persistent storage
- `--oracle-tns-alias`: Oracle TNS alias from tnsnames.ora (env: `ATP_TNS_ALIAS`, mutually exclusive with `--oracle-connect-descriptor`)
- `--oracle-connect-descriptor`: Oracle full connection string (env: `ATP_DSN`, mutually exclusive with `--oracle-tns-alias`)
- `--oracle-username`: Oracle username (env: `ATP_USER`)
- `--oracle-password`: Oracle password (env: `ATP_PASSWORD`)
- `--oracle-wallet-path`: Oracle wallet directory (env: `ATP_WALLET_PATH`, required for TNS alias)
- `--oracle-pool-min`: Min pool connections (default: 1, env: `ATP_POOL_MIN`)
- `--oracle-pool-max`: Max pool connections (default: 16, env: `ATP_POOL_MAX`)
#### Service Discovery
- `--service-discovery`: Enable Kubernetes service discovery
- `--service-discovery-port`: Port for worker URLs (default: 8000)
- `--service-discovery-namespace`: Kubernetes namespace to watch
- `--selector`: Label selectors for regular mode (format: `key1=value1 key2=value2`)
#### PD Mode
- `--pd-disaggregation`: Enable Prefill-Decode disaggregated mode
- `--prefill`: Initial prefill server (format: `URL BOOTSTRAP_PORT`)
- `--decode`: Initial decode server URL
- `--prefill-selector`: Label selector for prefill pods
- `--decode-selector`: Label selector for decode pods
- `--policy`: Routing policy (`cache_aware`, `random`, `power_of_two`, `round_robin`)
- `--prefill-policy`: Separate routing policy for prefill nodes (optional, overrides `--policy` for prefill)
- `--decode-policy`: Separate routing policy for decode nodes (optional, overrides `--policy` for decode)
#### Authentication
- `--api-key`: API key for router authentication (clients must provide this as Bearer token)
#### Concurrency and Rate Limiting
- `--queue-size`: Size of the pending-request queue when concurrency limits are reached (default: 100; set to 0 to disable queuing)
- `--queue-timeout-secs`: Maximum time a request may wait in the queue before timing out (default: 60; must be > 0 when queue is enabled)
- `--rate-limit-tokens-per-second`: Override token bucket refill rate for rate limiting (defaults to `--max-concurrent-requests` when omitted)
## Development
### Build Process
### Important Notes
- Initial workers declared via CLI inherit the router key; dynamic workers must supply keys explicitly.
- Router logs a warning when a worker is registered without a key while the router expects authentication.
- When router and workers share the same key, still include the key when invoking dynamic registration APIs.
## Development & Testing
```bash
# Build Rust project
# Build Rust components
cargo build
# Build Python binding (see Installation section above)
```
# Run Rust tests
cargo test
**Note**: When modifying Rust code, you must rebuild the wheel for changes to take effect.
### Troubleshooting
**VSCode Rust Analyzer Issues:**
Set `rust-analyzer.linkedProjects` to the absolute path of `Cargo.toml`:
# Build & install Python bindings
python -m build
pip install --force-reinstall dist/*.whl
```json
{
"rust-analyzer.linkedProjects": ["/workspaces/sglang/sgl-router/Cargo.toml"]
}
# Run Python tests
pytest
```
When modifying runtime behavior, rebuild the wheel or run the binary directly. Use `python -m sglang_router.launch_server` to co-launch router and SGLang workers in small clusters for local validation.
---
### CI/CD Pipeline
The continuous integration pipeline includes comprehensive testing, benchmarking, and publishing:
#### Build & Test
1. **Build Wheels**: Uses `cibuildwheel` for manylinux x86_64 packages
2. **Build Source Distribution**: Creates source distribution for pip fallback
3. **Rust HTTP Server Benchmarking**: Performance testing of router overhead
4. **Basic Inference Testing**: End-to-end validation through the router
5. **PD Disaggregation Testing**: Benchmark and sanity checks for prefill-decode load balancing
#### Publishing
- **PyPI Publishing**: Wheels and source distributions are published only when the version changes in `pyproject.toml`
- **Container Images**: Docker images published using `/docker/Dockerfile.router`
## Features
- **High Performance**: Rust-based routing with connection pooling and optimized request handling
- **Advanced Load Balancing**: Multiple algorithms including:
- **Cache-Aware**: Intelligent routing based on cache locality for optimal performance
- **Power of Two**: Chooses the less loaded of two randomly selected workers
- **Random**: Distributes requests randomly across available workers
- **Round Robin**: Sequential distribution across workers in rotation
- **Prefill-Decode Disaggregation**: Specialized load balancing for separated prefill and decode servers
- **Multiple Backend Support**: Route to SGLang, OpenAI, or other OpenAI-compatible services
- **Conversation Storage**: Track conversations with memory, Oracle ATP, or disable for minimal overhead
- **Service Discovery**: Automatic Kubernetes worker discovery and health management
- **Monitoring**: Comprehensive Prometheus metrics and structured logging
- **Scalability**: Handles thousands of concurrent connections with efficient resource utilization
SGLang Model Gateway continues to evolve alongside the core SGLang runtime. Contributions should keep CLI flags, documentation, and Python bindings in sync with the Rust implementation.
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment