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
docs(tl_kernels): tighten the route sentence, and let the row check accept a direct bucket_rows call
Aashishc3b0d3126133
diff --git a/laya/tl_kernels.py b/laya/tl_kernels.pyindex 617cfc2e..15382eff 100644--- a/laya/tl_kernels.py+++ b/laya/tl_kernels.py@@ -6,8 +6,8 @@ caller, `laya.fast.FastLaya._forward_unlocked`, pads the batch dim to the next p(`laya.backends.base.bucket_rows`, so a single question stays one row) and the sequence dim to a16-token bucket up to `DYNAMIC_MAX_L` and a 64-token bucket beyond -- the policy`laya.backends.base.bucket_tokens` implements, which `_forward_unlocked` repeats against its own-`FastLaya.DYNAMIC_MAX_L`/`LONG_BUCKET` instead of calling it. `laya.backends.base.pad_batch` is the-compile route's copy of the same policy, and that route never reaches these kernels. M is the flattened+`FastLaya.DYNAMIC_MAX_L`/`LONG_BUCKET` instead of calling it. `laya.backends.base.pad_batch` applies+that policy for the compile route, which never reaches these kernels. M is the flattened`rows * tokens`, so the multiple-of-16 comes from the token dim, not the row dim; out-of-boundsrows are predicated by TileLang's `T.ceildiv(M, bm)`, so the kernels themselves place noconstraint on M. Use compile_cpu for an explicit fp32 CPU specialization; the GPU defaults are unchanged.diff --git a/tests/test_confidence.py b/tests/test_confidence.pyindex 6ea751f9..204ca3ec 100644--- a/tests/test_confidence.py+++ b/tests/test_confidence.py@@ -2157,8 +2157,9 @@ def test_the_fast_path_still_inlines_the_token_bucket_the_docstring_says_it_repeassert "bucket_tokens(" not in body, ("`_forward_unlocked` calls `bucket_tokens` directly now; then there is one policy rather ""than two copies, and both the docstring and this check need re-reading")- assert "_bucket_n(" in body, (- "the row pad no longer comes from `laya.backends.base.bucket_rows` through `_bucket_n`")+ assert "_bucket_n(" in body or "bucket_rows(" in body, (+ "the row pad no longer comes from `laya.backends.base.bucket_rows`, through `_bucket_n` or "+ "directly, so the docstring's attribution needs re-reading")def test_the_two_bucket_policies_the_docstring_describes_are_still_the_same_numbers():