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
test(runtime-fixes): stage markers, so the Windows crash names its section
Nandakishor083a229b5def
diff --git a/tests/test_runtime_fixes.py b/tests/test_runtime_fixes.pyindex 7684bdb0..e7d3e2e0 100644--- a/tests/test_runtime_fixes.py+++ b/tests/test_runtime_fixes.py@@ -32,6 +32,16 @@ def check(name, got, want):FAIL.append("%s: got %r, want %r" % (name, got, want))+# `tests (windows)` has killed this suite about three runs in thirty-five, and now that the runner+# prints the exit status (#1172) the status says what it is: `exit 3221225501 (0xC000001D)`,+# STATUS_ILLEGAL_INSTRUCTION. Every `check` above collects into `PASS`/`FAIL` and nothing prints+# until the report at the foot of the file, so a crash took the whole log with it: the lane showed+# the suite's banner and then the next suite's. These markers go to stderr, unbuffered and+# unconditional, so the next occurrence names the section it died in. They cost one write each.+def stage(name):+ print("[stage] %s" % name, file=sys.stderr, flush=True)++# ------------------------------------------------------------------ fake tokenizerclass FakeTok:"""Just enough of a tokenizer for `build_sequence` / `system_one`."""@@ -77,6 +87,7 @@ def _bare_agent(model, dtype=torch.float32, amp=False, tok=None):# ------------------------------------------------------------------ state reused once+stage("state reused once")STATE = {"subject": "Duplicate charge", "body": "x" * 400}QUESTION = {"t": "choice", "ins": "pick", "crit": {"a": "x", "b": "y"}}@@ -122,6 +133,7 @@ check("build_sequence/option segments stay <= 49 tokens",# ------------------------------------------------------------------ autocast+stage("autocast")class DummyEnc(nn.Module):def __init__(self, d=16):super().__init__()@@ -215,6 +227,7 @@ check("infer/non-autocast error propagates", raised, True)# ------------------------------------------------------------------ MPS autocast gating+stage("MPS autocast gating")# MPS fp16 is slower than fp32 on one small row and only wins once the batch grows, so it is# gated by row count. These check the gate without an MPS device.def _mps_agent(amp=True, min_rows=None):@@ -291,6 +304,7 @@ del os.environ["LAYA_CPU_AMP"]# ------------------------------------------------------------------ amp context shape+stage("amp context shape")# A disabled gate must never construct torch.autocast: on torch builds without an MPS# autocast backend, entering it raises even with enabled=False, which broke MPS predict().mps = torch.device("mps")@@ -311,6 +325,7 @@ check("amp-context/_infer disabled completes", True, True)# ------------------------------------------------------------------ OOM fallback observability (#351)+stage("OOM fallback observability (#351)")class FakeCUDAInput:"""An input_ids that fails the way a real CUDA OOM does when moved off CPU.@@ -383,6 +398,7 @@ check("oom-fallback/plain CPU reason stays None", cpu_disabled.last_fallback_rea# ------------------------------------------------------------------ example 35 must teach this policy+stage("example 35 must teach this policy")# examples/35 is the page a reader opens to ask "does my machine get mixed precision?", and on main# it answered no twice: "CPU and MPS run **fp32**; `torch.autocast` has no MPS backend here, so Laya# only wraps the forward pass on CUDA" in the banner, then, printed under the timings, "both legs