feat(bindings/python): smg-as-tokenspeed-dependency surface (working serve_oai)

bindings/python/Cargo.toml

@@ -9,9 +9,16 @@ crate-type = ["cdylib"]
 
 [dependencies]
 pyo3 = { version = "0.28.2", features = ["extension-module", "abi3-py38"] }
+pyo3-async-runtimes = { version = "0.28", features = ["tokio-runtime"] }
 tokio = { version = "1.42.0", features = ["full"] }
 once_cell = "1.19"
 serde_yaml = "0.9"
+axum = { workspace = true }
+serde = { workspace = true }
+serde_json = { workspace = true }
+futures = { workspace = true }
+uuid = { workspace = true, features = ["v4"] }
+tracing = { workspace = true }
 
 [dependencies.smg]
 path = "../../model_gateway"
@@ -32,6 +39,12 @@ workspace = true
 [dependencies.reasoning-parser]
 workspace = true
 
+[dependencies.llm-tokenizer]
+workspace = true
+
+[dependencies.openai-protocol]
+workspace = true
+
 [features]
 default = ["pyo3/extension-module"]
 vendored-openssl = ["smg/vendored-openssl"]

bindings/python/README.md

@@ -2,6 +2,19 @@
 
 This directory contains the Python bindings for SMG (Shepherd Model Gateway), built using [maturin](https://github.com/PyO3/maturin) and [PyO3](https://github.com/PyO3/pyo3).
 
+## Two consumption modes
+
+These bindings serve two distinct use cases:
+
+1. **smg as a Python-launched binary** — the historical use, documented in
+   [Quick Start](#quick-start) below. `smg serve` boots a router in-process.
+2. **smg as a Python library** — used by inference engines that want smg's
+   protocol layer (tokenization, function calling, reasoning parser, OAI
+   server, MCP, response API) without owning the routing/gateway. The first
+   integration target is TokenSpeed; `tokenspeed serve` will `import smg_rs`
+   and call into the entry points in `src/serving.rs`. See the module
+   docstring there for the planned shape.
+
 ## Quick Start
 
 ### Installation

bindings/python/src/lib.rs

@@ -5,6 +5,8 @@ use pyo3::prelude::*;
 use smg::*;
 use smg_auth as auth;
 
+mod serving;
+
 // Define the enums with PyO3 bindings
 #[pyclass(eq, from_py_object)]
 #[derive(Clone, PartialEq, Debug)]
@@ -1274,5 +1276,6 @@ fn smg_rs(m: &Bound<'_, PyModule>) -> PyResult<()> {
     m.add_function(wrap_pyfunction!(print_banner, m)?)?;
     m.add_function(wrap_pyfunction!(get_available_tool_call_parsers, m)?)?;
     m.add_function(wrap_pyfunction!(get_available_reasoning_parsers, m)?)?;
+    serving::register(m)?;
     Ok(())
 }