perf: enable concurrent classification via Arc+clone

Makefile

@@ -141,7 +141,7 @@ check-go-mod-tidy:
 test-binding: rust
 	@echo "Running Go tests with static library..."
 	@export LD_LIBRARY_PATH=${PWD}/candle-binding/target/release && \
-		cd candle-binding && CGO_ENABLED=1 go test -v
+		cd candle-binding && CGO_ENABLED=1 go test -v -race
 
 # Test with the candle-binding library
 test-category-classifier: rust

candle-binding/semantic-router_test.go

@@ -1,10 +1,13 @@
 package candle_binding
 
 import (
+	"context"
+	"fmt"
 	"math"
 	"runtime"
 	"strings"
 	"sync"
+	"sync/atomic"
 	"testing"
 	"time"
 )
@@ -406,6 +409,99 @@ func TestModernBERTClassifiers(t *testing.T) {
 	})
 }
 
+func TestModernBertClassifier_ConcurrentClassificationSafety(t *testing.T) {
+	// init
+	if err := InitModernBertClassifier(CategoryClassifierModelPath, true); err != nil {
+		t.Skipf("ModernBERT classifier not available: %v", err)
+	}
+
+	texts := []string{
+		"This is a test sentence for classification",
+		"Another example text to classify with ModernBERT",
+		"The quick brown fox jumps over the lazy dog",
+		"Machine learning models are becoming more efficient",
+		"Natural language processing is a fascinating field",
+	}
+
+	// Baseline (single-threaded)
+	baseline := make(map[string]ClassResult, len(texts))
+	for _, txt := range texts {
+		res, err := ClassifyModernBertText(txt)
+		if err != nil {
+			t.Fatalf("baseline call failed for %q: %v", txt, err)
+		}
+		baseline[txt] = res
+	}
+
+	const numGoroutines = 10
+	const iterationsPerGoroutine = 5
+
+	var wg sync.WaitGroup
+	errCh := make(chan error, numGoroutines*iterationsPerGoroutine)
+	var total int64
+
+	ctx, cancel := context.WithCancel(context.Background())
+	defer cancel()
+
+	for g := range numGoroutines {
+		wg.Add(1)
+		go func(id int) {
+			defer wg.Done()
+			for i := range iterationsPerGoroutine {
+				select {
+				case <-ctx.Done():
+					return
+				default:
+				}
+
+				txt := texts[(id+i)%len(texts)]
+				res, err := ClassifyModernBertText(txt)
+				if err != nil {
+					errCh <- fmt.Errorf("gor %d iter %d classify error: %v", id, i, err)
+					cancel() // stop early
+					return
+				}
+
+				// Strict: class must match baseline
+				base := baseline[txt]
+				if res.Class != base.Class {
+					errCh <- fmt.Errorf("gor %d iter %d: class mismatch for %q: got %d expected %d", id, i, txt, res.Class, base.Class)
+					cancel()
+					return
+				}
+
+				// Allow small FP differences
+				if math.Abs(float64(res.Confidence)-float64(base.Confidence)) > 0.05 {
+					errCh <- fmt.Errorf("gor %d iter %d: confidence mismatch for %q: got %f expected %f", id, i, txt, res.Confidence, base.Confidence)
+					cancel()
+					return
+				}
+
+				atomic.AddInt64(&total, 1)
+			}
+		}(g)
+	}
+
+	wg.Wait()
+	close(errCh)
+
+	errs := 0
+	for e := range errCh {
+		t.Error(e)
+		errs++
+	}
+	if errs > 0 {
+		t.Fatalf("concurrency test failed with %d errors", errs)
+	}
+
+	expected := int64(numGoroutines * iterationsPerGoroutine)
+	if total != expected {
+		t.Fatalf("expected %d successful results, got %d", expected, total)
+	}
+
+	t.Logf("concurrent test OK: goroutines=%d iterations=%d", numGoroutines, iterationsPerGoroutine)
+}
+
 // TestModernBERTPIITokenClassification tests the PII token classification functionality
 func TestModernBERTPIITokenClassification(t *testing.T) {
 	// Test data with various PII entities

candle-binding/src/modernbert.rs

@@ -268,7 +268,7 @@ pub struct ModernBertClassifier {
 }
 
 lazy_static::lazy_static! {
-    static ref MODERNBERT_CLASSIFIER: Arc<Mutex<Option<ModernBertClassifier>>> = Arc::new(Mutex::new(None));
+   static ref MODERNBERT_CLASSIFIER: Arc<Mutex<Option<Arc<ModernBertClassifier>>>> = Arc::new(Mutex::new(None));
     static ref MODERNBERT_PII_CLASSIFIER: Arc<Mutex<Option<ModernBertClassifier>>> = Arc::new(Mutex::new(None));
     static ref MODERNBERT_JAILBREAK_CLASSIFIER: Arc<Mutex<Option<ModernBertClassifier>>> = Arc::new(Mutex::new(None));
 }
@@ -830,7 +830,7 @@ pub extern "C" fn init_modernbert_classifier(model_id: *const c_char, use_cpu: b
     match ModernBertClassifier::new(model_id, use_cpu) {
         Ok(classifier) => {
             let mut bert_opt = MODERNBERT_CLASSIFIER.lock().unwrap();
-            *bert_opt = Some(classifier);
+            *bert_opt = Some(Arc::new(classifier));
             true
         }
         Err(e) => {
@@ -930,18 +930,23 @@ pub extern "C" fn classify_modernbert_text(text: *const c_char) -> ModernBertCla
         }
     };
 
-    let bert_opt = MODERNBERT_CLASSIFIER.lock().unwrap();
-    match &*bert_opt {
-        Some(classifier) => match classifier.classify_text(text) {
-            Ok((class_idx, confidence)) => ModernBertClassificationResult {
-                class: class_idx as i32,
-                confidence,
-            },
-            Err(e) => {
-                eprintln!("Error classifying text with ModernBERT: {e}");
-                default_result
+    let classifier_arc_opt = MODERNBERT_CLASSIFIER.lock().unwrap();
+    match &*classifier_arc_opt {
+        Some(classifier_arc) => {
+            let classifier = classifier_arc.clone();
+            drop(classifier_arc_opt);
+
+            match classifier.classify_text(text) {
+                Ok((class_idx, confidence)) => ModernBertClassificationResult {
+                    class: class_idx as i32,
+                    confidence,
+                },
+                Err(e) => {
+                    eprintln!("Error classifying text with ModernBERT: {e}");
+                    default_result
+                }
             }
-        },
+        }
         None => {
             eprintln!("ModernBERT classifier not initialized");
             default_result
@@ -968,26 +973,31 @@ pub extern "C" fn classify_modernbert_text_with_probabilities(
         }
     };
 
-    let bert_opt = MODERNBERT_CLASSIFIER.lock().unwrap();
-    match &*bert_opt {
-        Some(classifier) => match classifier.classify_text_with_probs(text) {
-            Ok((class_idx, confidence, probabilities)) => {
-                // Allocate memory for probabilities array
-                let prob_len = probabilities.len();
-                let prob_ptr = Box::into_raw(probabilities.into_boxed_slice()) as *mut f32;
-
-                ModernBertClassificationResultWithProbs {
-                    class: class_idx as i32,
-                    confidence,
-                    probabilities: prob_ptr,
-                    num_classes: prob_len as i32,
+    let classifier_arc_opt = MODERNBERT_CLASSIFIER.lock().unwrap();
+    match &*classifier_arc_opt {
+        Some(classifier_arc) => {
+            let classifier = classifier_arc.clone();
+            drop(classifier_arc_opt);
+
+            match classifier.classify_text_with_probs(text) {
+                Ok((class_idx, confidence, probabilities)) => {
+                    // Allocate memory for probabilities array
+                    let prob_len = probabilities.len();
+                    let prob_ptr = Box::into_raw(probabilities.into_boxed_slice()) as *mut f32;
+
+                    ModernBertClassificationResultWithProbs {
+                        class: class_idx as i32,
+                        confidence,
+                        probabilities: prob_ptr,
+                        num_classes: prob_len as i32,
+                    }
+                }
+                Err(e) => {
+                    eprintln!("Error classifying text with probabilities using ModernBERT: {e}");
+                    default_result
                 }
             }
-            Err(e) => {
-                eprintln!("Error classifying text with probabilities using ModernBERT: {e}");
-                default_result
-            }
-        },
+        }
         None => {
             eprintln!("ModernBERT classifier not initialized");
             default_result