[Ryan VanderMeulen (rvandermeulen)]

Description

Adds a CMake option, onnxruntime_DISABLE_STREAM (default OFF), that keeps onnxruntime_set_compile_flags from defining ORT_ENABLE_STREAM in a non-minimal build. Minimal builds already compile without it, and the framework and session code it controls is guarded by #ifdef ORT_ENABLE_STREAM, so nothing else needs to change.

Execution providers that use device streams (CUDA, TensorRT, WebGPU, ...) need stream support, so the option is meant for builds that do not use them. It does not refuse those providers at configure time, since plugin execution providers can be loaded at runtime anyway; if you would like a configure-time check for the built-in ones, or a build.py flag next to --disable_rtti, I can add either.

Full training builds (onnxruntime_ENABLE_TRAINING) also need streams, so that combination is rejected at configure time.

The stream-disabled planner path now honors the session's execution order (priority-based ordering was ignored there, in minimal builds too), covered by a new planner test.

Motivation and Context

Firefox ships ONNX Runtime as a CPU-only shared library (non-minimal, --disable_exceptions, --disable_rtti, lite protobuf) for its on-device ML features, and carries a local patch that comments out the ORT_ENABLE_STREAM definition. This option would replace that patch.

Measured on main, Linux x86-64, MinSizeRel with LTO: compiling streams in makes the library's code about 104 KB (0.7%) larger, and our CPU inference benchmarks on Linux, Windows and macOS show no benefit from it. Most differences are within noise; the repeatable ones are small costs (a few MiB more memory in one benchmark, slightly slower short inference on Windows).

A CPU-only non-minimal Release build of main with streams disabled (GCC 15.2, Linux x86-64) passes the same ctest suite as the default build, except for the eight stream-specific tests that are compiled out (multi-stream PlannerTest cases, DeviceStreamCollection.TestOverride, StreamAwareArenaTest.TwoStreamAllocation). ActivationOpNoInfTest.Softsign fails locally in both configurations.

Earlier reports about this define (#16345, #20180) were closed as stale without a decision on supporting the configuration.

[azure-pipelines]

Azure Pipelines:
There may be pipelines that require an authorized user to comment /azp run to run.

[Copilot]

🟡 Changes recommended

CPU-only full-training builds still require streams, so the incompatible option combination needs rejection or implementation support.

1 open issue

• Reject disabling streams with ONNX Runtime training enabled · New

What changed in this PR

Adds an opt-out for stream support in CPU-only, non-minimal ONNX Runtime builds, replacing Firefox’s local size-reduction patch.

Changes:

• Adds onnxruntime_DISABLE_STREAM, defaulting to OFF.
• Omits ORT_ENABLE_STREAM when requested, preserving existing defaults and minimal-build behavior.

File	Description
cmake/​CMakeLists.txt	Declares the option and gates the stream compile definition.

_{Review effort: Balanced}

[Copilot]

🔵 Needs a closer look

The stream-disabled path silently ignores priority-based CPU execution ordering.

No open issues 🎉

1 previously missed issue

In code that hasn't changed since last review

Preserve priority-based execution order when streams are disabled

cmake/​CMakeLists.txt:1208

With onnxruntime_DISABLE_STREAM=ON, a non-minimal CPU session configured with execution_order=PRIORITY_BASED silently executes in the default order. The no-stream branch in onnxruntime/core/framework/allocation_planner.cc:1964 calls GetNodesInTopologicalOrder() without an argument, then builds kernel steps in that order. The stream-enabled branch forwards the requested execution order. Pass context_->GetExecutionOrder() to the no-stream call as well so this build option preserves priority-based CPU scheduling.

1 resolved since last review

• Reject disabling streams with ONNX Runtime training enabled

_{Review effort: Balanced}

[Copilot]

🟡 Changes recommended

The stream-disabled planner execution-order change needs focused regression coverage.

1 open issue

• Add regression test for priority-based execution order without streams · New

_{Review effort: Balanced}

[Copilot]

🟢 Approval recommended

The focused changes preserve defaults, address the known training incompatibility, and include targeted regression coverage with no unresolved findings.

No open issues 🎉

1 resolved since last review

• Add regression test for priority-based execution order without streams

_{Review effort: Balanced}

[Ryan VanderMeulen (rvandermeulen)]

Edward Chen (@edgchen1) This is the last one, I swear 😅! Totally fine if y'all don't want to take this, but having an explicit option for this would let us drop manually patching CMakeLists.txt otherwise.