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
ci(cpp): TSan cannot read the prebuilt ONNX Runtime, so stop reporting it
Nandakishor7fa53b3ebe67
diff --git a/.github/workflows/cpp.yml b/.github/workflows/cpp.ymlindex 9bfba099..37b1fd92 100644--- a/.github/workflows/cpp.yml+++ b/.github/workflows/cpp.yml@@ -225,6 +225,15 @@ jobs:env:LAYA_ONNX_GRAPH: ${{ env.ONNX_ROOT }}/multilingual/laya.onnxLAYA_ONNX_GRAPH_ENGLISH: ${{ env.ONNX_ROOT }}/english/laya.onnx+ # The ONNX Runtime this links is a prebuilt binary, so TSan cannot see the+ # synchronization inside it and reads every buffer its thread pool returns as a race.+ # laya-cpp/tsan-onnxruntime.supp records the measurement and what the suppression costs.+ # Unused by the g++/debug and clang++/asan-ubsan cells, which link no TSan runtime to+ # read it. An absolute path on purpose: TSan aborts the process with exit 134 on a+ # suppressions file it cannot read, so a path relative to the wrong directory would take+ # out every test rather than fail quietly. Inherited by laya_onnx_static_holder, whose+ # exit status AStaticSessionAndAgentExitCleanly asserts and which TSan leaves at 66.+ TSAN_OPTIONS: suppressions=${{ github.workspace }}/laya-cpp/tsan-onnxruntime.supprun: ctest --preset "${{ matrix.preset }}" --output-junit "$RUNNER_TEMP/ctest-onnx.xml"- name: Count tests (fail on skip, enforce the onnx floor)if: ${{ !cancelled() }}diff --git a/laya-cpp/tsan-onnxruntime.supp b/laya-cpp/tsan-onnxruntime.suppnew file mode 100644index 00000000..79cecbb2--- /dev/null+++ b/laya-cpp/tsan-onnxruntime.supp@@ -0,0 +1,24 @@+# ThreadSanitizer suppressions for the prebuilt ONNX Runtime.+#+# laya-cpp links the official ONNX Runtime binary that cmake/onnxruntime.cmake downloads and+# sha256-verifies. That library is not built with -fsanitize=thread, so TSan cannot see its+# internal synchronization: the happens-before edges its own mutexes and atomics establish are+# invisible, and every buffer its intra-op thread pool hands back to the main thread therefore+# looks like an unsynchronized access.+#+# Measured, not assumed: the `onnx (clang++, tsan)` job reported 342 races, every single one with+# a libonnxruntime.so frame, and all of them in `memcpy` or `operator delete`. 20 of 479 tests+# failed as a result, including four that only assert the static-holder helper's exit status and+# saw TSan's own exitcode=66.+#+# Scope, and what it costs. `called_from_lib` drops a report only when the accessing stack runs+# inside this library, so a race between two laya threads on laya-owned state is still reported.+# What it does give up is a race where laya hands memory to an ONNX Runtime worker thread, since+# that thread's stack is inside the library. TSan cannot adjudicate those either way while the+# library is uninstrumented, so the choice is between suppressing them and switching the whole+# cell off; this keeps the laya-vs-laya coverage the cell exists for, which is what+# OnnxRig.ConcurrentPredictAgreesWithASerialCall and CloseInterruptsARunInFlight exercise.+#+# Revisit if ONNX Runtime ever ships a TSan-instrumented build, or if laya starts passing a+# callback into the library.+called_from_lib:libonnxruntime.so