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
release: 0.4.9
Nandakishorca5045a3f1fa
diff --git a/README.md b/README.mdindex bf3196d2..c099b370 100644--- a/README.md+++ b/README.md@@ -94,31 +94,33 @@ The shipped checkpoints work zero-shot, but fine-tuning on decisions from your o**[nandhakishorm.github.io/laya](https://nandhakishorm.github.io/laya/)**: guides for [prediction hooks](https://nandhakishorm.github.io/laya/hooks/), [schema-driven decisions](https://nandhakishorm.github.io/laya/structured/), [Docker](https://nandhakishorm.github.io/laya/docker/) and [LangChain and LangGraph](https://nandhakishorm.github.io/laya/langchain/), plus a full [API reference](https://nandhakishorm.github.io/laya/reference/).-## What's new in 0.4.8--* **`laya-go` runs the model.** `agent.Open`, `Predict`, `PredictBatch` and `PredictLong` answer- typed questions over an `infer.Session`, and a nested `laya-go/onnx` module (its own `go.mod`,- cgo) runs a fused `laya.onnx` on ONNX Runtime 1.29+. The main module stays standard library- only: `go test ./...` there compiles no cgo, and `CGO_ENABLED=0` builds `onnx` too, where- `Open` returns `ErrNoCGO`. The goldens are regenerated from Python in CI (#1222, #1233).-* **A gate parameter that decides nothing is refused.** `--max ece=inf`, `--min accuracy=-inf` and- `--tolerance METRIC=1e999` passed every measurement and still exited 0, which is the code that- means the gate ran (#1225). A baseline that is not a metric name to number map, or that has no- metrics at all, now raises instead of being subtracted through, and `"choice_accuracy": true`- no longer compares as 100% (#1223). A report `laya-evals` cannot read is exit 2, not a traceback- and exit 1 (#1230). `bins` and `min_bucket_n` are positive integers (#1228), and `ece_score`- refuses a `bins` it cannot histogram, which used to return the perfect 0.0 (#1232).-* **`laya-serve` reads a blank `LAYA_*` value as unset**, which is what a `.env` line or a Compose- `KEY=` with nothing after it writes; `LAYA_PRELOAD=` was getting lazy loading nobody asked for- (#1226). A bad `LAYA_LOG_LEVEL` exits with a message naming the accepted values instead of a- `KeyError` from inside uvicorn (#1227).-* **Schema-driven decisions read two more shapes correctly.** A non-array `enum` is a- `SchemaError` naming the path, where a bare string silently split into per-character options- (#1220), and a single-value boolean constraint such as `Literal[True]` stays a one-option choice- instead of becoming a noul that could answer `False` (#1235).-* **`laya.jev_confidence` refuses a negative probability** rather than returning a number outside- `[0, 1]`, so the range is now a property of the function (#1238). `setup_laya.sh` detects the- platform and hands off to a Linux or macOS script (#1234).+## What's new in 0.4.9++* **A standalone Python HTTP client.** `sdk/python` holds `laya-client`: synchronous `predict()`+ and `health()` against a self-hosted `laya-serve`, written on the standard library with zero+ dependencies, so an application server no longer installs torch to call one (#1264, issue+ #1176). It refuses redirects so a bearer token is never forwarded, rejects a `base_url`+ carrying credentials or whitespace, and blocks non-finite JSON in both directions. Install it+ from this checkout with `python -m pip install ./sdk/python`; it is not on PyPI yet.+* **`laya-go` gains the confidence gate, hooks and digest verification.** `WithMinConfidence`+ applies the same abstention gate as the Python runtime, `OnPredictStart`/`End`/`Error` dispatch+ with timeouts and panic isolation, per-language temperatures are read from the checkpoint, and+ `laya-go/internal/digest` checks a checkpoint's SHA-256 before loading it, streamed, with path+ traversal refused (#1247, #1260).+* **`LAYA_DEVICE=cuda:258` is refused rather than silently asking for GPU 2.** torch keeps a CUDA+ index in its low 8 bits, so it parses that string and builds `cuda:2`, and `cuda:128` builds+ `cuda:-128`. A value whose parsed device is not the device it names now raises, naming the+ variable (#1242).+* **Three more gate parameters that could not fail.** A tolerance of `Infinity` accepts a drift of+ any size while the run still exits 0 (#1250), `--revision` naming two different commits for one+ checkpoint is refused instead of resolved to whichever came last (#1248), and a slice label+ containing `|` is escaped so a tag like `risk|high` cannot shift every number in its row of a+ `--markdown` report (#1241). A `Dataset` row built in Python now gets the checks a JSONL row+ already had, before a single inference call rather than after (#1253).+* **`LAYA_CPU_AMP=bf16` is documented as a hardware requirement, not a preference.** On a CPU+ without native BF16 instructions the forward dies with an illegal instruction rather than+ raising, so `_infer`'s autocast retry cannot catch it. `docs/reference/agent.md` now carries the+ capability reading to check and a child-process probe to qualify a host.Earlier releases are in the [GitHub releases](https://github.com/NandhaKishorM/laya/releases).diff --git a/laya/__init__.py b/laya/__init__.pyindex 3fb89c6a..4488cb75 100644--- a/laya/__init__.py+++ b/laya/__init__.py@@ -35,7 +35,7 @@ from .presets import (from .router import DEFAULT_MODELS, RouteDecision, Routerfrom .structured import DecisionResult, decide, decide_batch-__version__ = "0.4.8"+__version__ = "0.4.9"# Routing, language detection and email cleaning are pure Python. The torch-backed names are# resolved lazily so that `import laya` -- and therefore `from laya import Router` ordiff --git a/nix/package.nix b/nix/package.nixindex fbf5564a..b5fee1c3 100644--- a/nix/package.nix+++ b/nix/package.nix@@ -12,7 +12,7 @@ let# side, where tests/test_packaging.py compares all three declarations (this one,# pyproject.toml and laya.__version__) and fails CI when they diverge. Nothing# did, which is how this sat at 0.3.4 while the package reached 0.3.20.- version = "0.4.8";+ version = "0.4.9";src = ../.;format = "setuptools";propagatedBuildInputs = with python3Packages; [diff --git a/pyproject.toml b/pyproject.tomlindex c992923e..1c358b68 100644--- a/pyproject.toml+++ b/pyproject.toml@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"[project]name = "laya"-version = "0.4.8"+version = "0.4.9"description = "Fast, non-autoregressive System 1 decision engine with calibrated probabilities"readme = "README.md"requires-python = ">=3.10"