Skip to content

YC (S26) | Open Computer History | Continuously record your company computer work, map your workflows, help you find work worth automating, and power your agents' context

fix(redact): stop ORT thread pool from burning 4 cores while redaction drains

Louis Beaumont68099e5773c1
diff --git a/crates/screenpipe-redact/src/adapters/onnx.rs b/crates/screenpipe-redact/src/adapters/onnx.rs
index 4e35f98f0..482adff5b 100644
--- a/crates/screenpipe-redact/src/adapters/onnx.rs
+++ b/crates/screenpipe-redact/src/adapters/onnx.rs
@@ -778,23 +778,20 @@ mod runtime {
|| -> Result<Session, ort::Error> {
let builder = Session::builder()?
.with_optimization_level(GraphOptimizationLevel::Level3)?
- .with_intra_threads(num_cpus_physical())?;
- #[cfg(feature = "onnx-coreml")]
- let builder = builder.with_execution_providers([
- ort::execution_providers::CoreMLExecutionProvider::default()
- // MLProgram + ComputeUnits::All offloads to the Apple Neural
- // Engine; measured ~3.4x faster than the legacy NeuralNetwork
- // default (which is slower than CPU) and lower CPU/GPU power.
- .with_model_format(
- ort::execution_providers::coreml::CoreMLModelFormat::MLProgram,
- )
- .with_compute_units(
- ort::execution_providers::coreml::CoreMLComputeUnits::All,
- )
- .with_subgraphs(true)
- .build(),
- ort::execution_providers::CPUExecutionProvider::default().build(),
- ])?;
+ // This session serves a background batch worker — never let the
+ // intra-op pool busy-spin between ops. A spinning full-width
+ // pool burned ~4 cores in ThreadPoolTempl::WorkerLoop while the
+ // redaction backlog drained (340% CPU regression after 3b9a1a105).
+ .with_intra_op_spinning(false)?
+ // Half the cores is enough for a background batch worker and
+ // keeps the drain from monopolizing the machine.
+ .with_intra_threads((num_cpus_physical() / 2).max(2))?;
+ // NO CoreML EP here, deliberately: this text model is int8-quantized
+ // RoBERTa with dynamic sequence lengths, and the ANE compiler rejects
+ // every layer ("E5RT: unbounded dimension is not supported"), so
+ // CoreML contributes nothing but per-partition compile attempts and
+ // EP handoff overhead. The image model (rfdetr.rs, fixed-size input)
+ // is the one that actually runs on the ANE.
#[cfg(feature = "onnx-directml")]
let builder = builder.with_execution_providers([
ort::execution_providers::DirectMLExecutionProvider::default()
diff --git a/crates/screenpipe-redact/src/adapters/rfdetr.rs b/crates/screenpipe-redact/src/adapters/rfdetr.rs
index 663d85192..b40cb0e87 100644
--- a/crates/screenpipe-redact/src/adapters/rfdetr.rs
+++ b/crates/screenpipe-redact/src/adapters/rfdetr.rs
@@ -245,7 +245,18 @@ mod imp {
|| -> Result<Session, ort::Error> {
let builder = Session::builder()?
.with_optimization_level(GraphOptimizationLevel::Level3)?
- .with_intra_threads(num_cpus_physical())?;
+ // Background batch worker — never busy-spin between ops. With
+ // CoreML active the graph is partitioned ANE/CPU and a spinning
+ // full-width pool burned ~4 cores in WorkerLoop while the
+ // redaction backlog drained (340% CPU regression after 3b9a1a105).
+ .with_intra_op_spinning(false)?
+ // With CoreML the CPU pool only runs CoreML-rejected fallback
+ // ops; 2 threads is plenty. CPU-only builds keep the full pool.
+ .with_intra_threads(if cfg!(feature = "onnx-coreml") {
+ 2
+ } else {
+ num_cpus_physical()
+ })?;
// Offload to the Apple Neural Engine (Mac) / NPU (Windows) instead of
// running CPU-only. CoreML MLProgram + ComputeUnits::All measured ~3.4x
// faster than the legacy default and keeps the work off the CPU/GPU.