Non-autoregressive System 1 decision engine. Typed choice, score and yes/no decisions over any text in a single forward pass, in 100+ languages, with a router that picks the right checkpoint per request.
Python calibrationclassificationdecision-modelhuggingfacejevmodernbertmultilingualnlppythonpytorchroutingtyped-decisionszero-shot
fix(serve): send Retry-After with the 503 admission refusal
bunnysayzza4494a2f31cf
diff --git a/laya/serve.py b/laya/serve.pyindex 2cd30922..cc0abe37 100644--- a/laya/serve.py+++ b/laya/serve.py@@ -321,8 +321,11 @@ def create_app(router: Optional[Any] = None):admission = asyncio.Semaphore(max_concurrent)if admission.locked():# Non-blocking: excess load is refused rather than queued, so the- # buffered bodies stay within the bound above.- raise HTTPException(status_code=503, detail="server busy, try again later")+ # buffered bodies stay within the bound above. Retry-After tells+ # well-behaved clients when a slot is likely free: admission turns+ # over at inference speed, so one second is the honest hint.+ raise HTTPException(status_code=503, detail="server busy, try again later",+ headers={"Retry-After": "1"})await admission.acquire()try:return await _systemone_inner(request)diff --git a/tests/test_serve.py b/tests/test_serve.pyindex ae52aae2..8eb0008f 100644--- a/tests/test_serve.py+++ b/tests/test_serve.py@@ -512,13 +512,16 @@ def test_admission_bound_refuses_with_503_when_full(monkeypatch):# Give the first request a moment to settle past the gate too, so the# second request deterministically finds the slot taken.await asyncio.sleep(0.2)- seen["second"] = (await client.post("/v1/systemone", json=REQ)).status_code+ second = await client.post("/v1/systemone", json=REQ)+ seen["second"] = second.status_code+ seen["retry_after"] = second.headers.get("retry-after")fake.release.set()seen["first"] = (await first).status_codeasyncio.run(drive())assert seen["second"] == 503, seen+ assert seen["retry_after"] == "1", seenassert seen["first"] == 200, seen