redact_ner 0.1.2 → 0.1.3

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: b5865619e758c5b2d9959ecc613d1ae7830c9ecef1ab6d69370d513117f9fa46
-  data.tar.gz: ae41cb709cd3b1eb7e51458c42eb3256e13fa857e0c788891585765132091afe
+  metadata.gz: 99ee93baed8bd1952265ca34c45657d848041a2afa3656e96d97bb07c455d1f9
+  data.tar.gz: f6d3b0f18c8e392fe8851cf951f3ed091a7d5b9a3d624c7943250b6f7f2736dc
 SHA512:
-  metadata.gz: 576bdab1c4023c482d802c46df56b1651f8e92e1d9f7b5d3160baf98da59d0aaa626484b885369730cf0e5bbb0ea7e247cb3cca59fefed3346e59b0b14facba4
-  data.tar.gz: 3b82ee15c23bdbf1df3c8b68cc746bbc03272997406d0b50f10f370f62a2630dffac6df9c547184e60893a3324f002d2dbc1ebd486358ca28c2ba186e515906d
+  metadata.gz: 9028c72248a875fb83429a7d0d0a4363e9361eb9725d86c24f66f616487be05c43752e8f6ab5b8c8be0d112d9026180b65845f3082e37279fd3352c348712e6f
+  data.tar.gz: cf82b82abe1e921f2da99dd6d872b020317504f5ccf23f88fceb848bb2143730ee914d873a89244d54e316c70365e43b2a4d89ef03f89a6c86f8049bcd9791a7

data/CHANGELOG.md

@@ -7,6 +7,15 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+## [0.1.3] - 2026-06-05
+
+## What's Changed
+* feat: release GVL during ONNX inference by @mitsuru in https://github.com/mitsuru/redact-ner-ruby/pull/7
+
+
+**Full Changelog**: https://github.com/mitsuru/redact-ner-ruby/compare/v0.1.2...v0.1.3
+
+
 ## [0.1.2] - 2026-05-17
 
 ## What's Changed
@@ -52,7 +61,8 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
   - The precompiled musl gems (Alpine / distroless) link `libstdc++`
     dynamically; install it at runtime, e.g. `apk add --no-cache libstdc++`.
 
-[Unreleased]: https://github.com/mitsuru/redact-ner-ruby/compare/v0.1.2...HEAD
+[Unreleased]: https://github.com/mitsuru/redact-ner-ruby/compare/v0.1.3...HEAD
 [0.1.1]: https://github.com/mitsuru/redact-ner-ruby/releases/tag/v0.1.1
 [0.1.0]: https://github.com/mitsuru/redact-ner-ruby/releases/tag/v0.1.0
 [0.1.2]: https://github.com/mitsuru/redact-ner-ruby/releases/tag/v0.1.2
+[0.1.3]: https://github.com/mitsuru/redact-ner-ruby/releases/tag/v0.1.3

data/Cargo.lock

@@ -1450,6 +1450,7 @@ version = "0.1.0"
 dependencies = [
  "magnus",
  "openssl-sys",
+ "rb-sys",
  "redact-core",
  "redact-ner",
 ]

data/ext/redact_ner/Cargo.toml

@@ -13,6 +13,12 @@ magnus = "0.8"
 redact-ner = "0.8.3"
 redact-core = "0.8.3"
 
+# Direct access to rb_thread_call_without_gvl so analyze() can release the GVL
+# during ONNX inference. magnus 0.8 exposes no without_gvl wrapper. rb-sys is
+# already pulled in transitively by magnus; default-features = false keeps
+# Cargo's resolved feature set identical to magnus's (no ABI drift).
+rb-sys = { version = ">=0.9.113", default-features = false }
+
 # Force a vendored, statically-linked OpenSSL so cross-compilation does not
 # depend on a per-target system OpenSSL. openssl-sys is pulled transitively
 # via ort -> ort-sys -> ureq -> native-tls; we only pin its build mode here.

data/ext/redact_ner/src/lib.rs

@@ -1,6 +1,58 @@
 use magnus::{function, method, prelude::*, Error, RArray, Ruby};
 use redact_core::Recognizer as _;
 use redact_ner::NerRecognizer;
+use std::os::raw::c_void;
+
+/// Run `f` with the GVL released, then re-acquire it before returning.
+///
+/// SAFETY CONTRACT: `f` MUST NOT touch any Ruby object or call any Ruby API —
+/// no other Ruby C function is safe to call while the GVL is released. It is
+/// for pure-Rust, CPU-bound (or sleeping) work only.
+///
+/// `f` runs synchronously on the *same* OS thread (the call blocks until it
+/// returns), so borrows captured by `f` stay valid and no `Send` bound is
+/// needed. Panics are caught and resumed after the GVL is re-acquired so we
+/// never unwind across the C boundary. No unblock function is passed (NULL
+/// ubf), so the native call is not interruptible by Thread#kill/Timeout —
+/// acceptable because ONNX inference has no clean cancellation point.
+fn nogvl<F, R>(f: F) -> R
+where
+    F: FnOnce() -> R,
+{
+    struct Data<F, R> {
+        func: Option<F>,
+        result: Option<std::thread::Result<R>>,
+    }
+
+    unsafe extern "C" fn trampoline<F, R>(arg: *mut c_void) -> *mut c_void
+    where
+        F: FnOnce() -> R,
+    {
+        let data = &mut *(arg as *mut Data<F, R>);
+        let func = data.func.take().expect("nogvl closure run twice");
+        data.result = Some(std::panic::catch_unwind(std::panic::AssertUnwindSafe(func)));
+        std::ptr::null_mut()
+    }
+
+    let mut data: Data<F, R> = Data {
+        func: Some(f),
+        result: None,
+    };
+
+    unsafe {
+        rb_sys::rb_thread_call_without_gvl(
+            Some(trampoline::<F, R>),
+            &mut data as *mut _ as *mut c_void,
+            None,
+            std::ptr::null_mut(),
+        );
+    }
+
+    match data.result.take().expect("nogvl closure did not run") {
+        Ok(value) => value,
+        Err(panic) => std::panic::resume_unwind(panic),
+    }
+}
 
 #[magnus::wrap(class = "RedactNer::Recognizer", free_immediately, size)]
 struct RbRecognizer {
@@ -26,8 +78,13 @@ impl RbRecognizer {
     }
 
     fn analyze_raw(&self, text: String, language: String) -> Result<RArray, Error> {
+        // ONNX inference is CPU-bound and touches no Ruby objects, so run it
+        // with the GVL released to let other Ruby threads make progress.
+        let results = nogvl(|| self.inner.analyze(&text, &language)).map_err(runtime_error)?;
+
+        // Building the Ruby Array below DOES touch Ruby, so it runs with the
+        // GVL held (we are back in normal extension context here).
         let ruby = Ruby::get().map_err(runtime_error)?;
-        let results = self.inner.analyze(&text, &language).map_err(runtime_error)?;
 
         let arr = ruby.ary_new_capa(results.len());
         for r in results {
@@ -74,6 +131,13 @@ impl RbRecognizer {
             .collect()
     }
 
+    // Test-only probe used by test/test_gvl_release.rb to prove the GVL is
+    // released during native CPU-bound work. Sleeps inside the same code path
+    // that wraps inference. NOT part of the public API (leading underscore).
+    fn nogvl_sleep_ms(&self, ms: u64) {
+        nogvl(|| std::thread::sleep(std::time::Duration::from_millis(ms)));
+    }
+
     fn min_confidence(&self) -> f32 {
         self.inner.config().min_confidence
     }
@@ -105,6 +169,7 @@ fn init(ruby: &Ruby) -> Result<(), Error> {
         "supported_entities",
         method!(RbRecognizer::supported_entities, 0),
     )?;
+    class.define_method("_nogvl_sleep_ms", method!(RbRecognizer::nogvl_sleep_ms, 1))?;
     class.define_method("min_confidence", method!(RbRecognizer::min_confidence, 0))?;
     class.define_method("max_seq_length", method!(RbRecognizer::max_seq_length, 0))?;
     class.define_method("model_path", method!(RbRecognizer::model_path, 0))?;

data/lib/redact_ner/version.rb

@@ -1,5 +1,5 @@
 # frozen_string_literal: true
 
 module RedactNer
-  VERSION = "0.1.2"
+  VERSION = "0.1.3"
 end

metadata

@@ -1,7 +1,7 @@
 --- !ruby/object:Gem::Specification
 name: redact_ner
 version: !ruby/object:Gem::Version
-  version: 0.1.2
+  version: 0.1.3
 platform: ruby
 authors:
 - Mitsuru Hayasaka
@@ -95,7 +95,7 @@ metadata:
   source_code_uri: https://github.com/mitsuru/redact-ner-ruby
   bug_tracker_uri: https://github.com/mitsuru/redact-ner-ruby/issues
   changelog_uri: https://github.com/mitsuru/redact-ner-ruby/blob/main/CHANGELOG.md
-  documentation_uri: https://rubydoc.info/gems/redact_ner/0.1.2
+  documentation_uri: https://rubydoc.info/gems/redact_ner/0.1.3
   rubygems_mfa_required: 'true'
 rdoc_options: []
 require_paths: