mstar.api_server.rust_frontend#

The Rust axum frontend in front of APIServer.

mstar-server (rust/server/) owns the HTTP surface — routes, OpenAI translation, SSE/NDJSON streaming, uploads, CORS — and speaks a flattened msgpack protocol over the Rust ZMQ transport to this bridge, which drives the existing Python data plane (APIServer.submit_request / iter_result_chunks / abort_request; preprocessing and the conductor protocol are untouched):

frontend -> bridge: {t:”submit”, rid, text, file_paths,
input_modalities, output_modalities, model_kwargs, streaming} | {t:”abort”, rid}

bridge -> frontend: {t:”chunk”, rid, modality, data(bin), metadata}

{t:”err”, rid, msg} | {t:”done”, rid}

The submit shape is APIServer.submit_request’s signature verbatim — the protocol was designed by flattening it. The bridge mesh lives in its own socket dir (private to the frontend/bridge pair), so entity names never collide with the conductor/worker mesh.

Opt-in: mstar-serve --rust-frontend replaces uvicorn.run with the mstar-server binary + this loop. Everything else in the process (preprocess worker, conductor spawn, tensor transport) is identical.

Module Attributes

BRIDGE_RECV_WAIT_S

Upper bound on one blocking receive, so stop() is noticed promptly and the loop stays responsive to shutdown.

Functions

`find_server_binary`([explicit])	Resolve the `mstar-server` binary: explicit arg, `MSTAR_SERVER_BIN`, `$PATH`, then the in-repo release build.
`launch_rust_server`(binary, model_name, port, ...)

Classes

RustFrontendBridge(server, bridge_dir)

The serve loop between the Rust frontend and APIServer.

mstar.api_server.rust_frontend.BRIDGE_RECV_WAIT_S = 0.05#: Upper bound on one blocking receive, so stop() is noticed promptly and the loop stays responsive to shutdown. Not a latency knob: an arriving message ends the wait immediately.

class mstar.api_server.rust_frontend.RustFrontendBridge(server, bridge_dir)[source]#