relayserve 1.1

Relay: minimal LLM inference server for heterogeneous devices

RelayServe

RelayServe is a minimal LLM inference server that adapts to heterogeneous devices.

Quick start

Install from PyPI:

pip install relayserve
relayserve

Defaults:

• HTTP server: :8080
• Endpoints:
  • GET /healthz
  • GET /v1/models
  • POST /v1/chat/completions
  • GET /metrics
  • GET /debug/shard
  • POST /v1/chat/pretty (colorized text response)
• Backends: set RELAYSERVE_BACKENDS to comma-separated llama.cpp servers

Environment

• RELAYSERVE_PORT (default 8080)
• RELAYSERVE_MODEL_ID (default relay-gguf)
• RELAYSERVE_BACKENDS (comma-separated, e.g. http://localhost:8081,http://localhost:8082)
• RELAYSERVE_BATCH_SIZE (default 4)
• RELAYSERVE_BATCH_WAIT_MS (default 10)
• RELAYSERVE_METRICS_MAX_ITEMS (default 1000)
• RELAYSERVE_TOTAL_LAYERS (default 32)
• RELAYSERVE_PRETTY_JSON (set 1 for readable JSON responses)
• RELAYSERVE_PRETTY_DEFAULT (default 1, set 0 for JSON by default)

Spawning llama.cpp backends

export LLAMA_SERVER_PATH=/path/to/llama.cpp/server
export LLAMA_MODEL_PATH=/path/to/models/phi-3-mini.gguf
export LLAMA_PORTS=8081,8082
python scripts/spawn_backends.py

Then run the RelayServe server with:

export RELAYSERVE_BACKENDS=http://localhost:8081,http://localhost:8082
relayserve