Fix Windows shutdown deadlock and leak with the Ruy backend

src/cpu/backend.cc

@@ -106,9 +106,24 @@ namespace ctranslate2 {
     }
 
 #ifdef CT2_WITH_RUY
+    // The per-thread ruy::Context is heap-allocated so that its lifetime is not tied
+    // to thread-local destruction. Its destructor joins ruy's internal thread pool;
+    // on Windows that join, if it runs while the owning thread is exiting (under the
+    // loader lock), deadlocks ThreadPool shutdown (jkawamoto/ctranslate2-rs#64).
+    // Instead we destroy it explicitly via clear_ruy_context() from ReplicaWorker::
+    // finalize(), which runs on the worker thread in a normal context (not thread
+    // exit), so the join completes and no memory (incl. ruy's prepacked cache) leaks.
+    static thread_local ruy::Context* ruy_context = nullptr;
+
     ruy::Context *get_ruy_context() {
-      static thread_local ruy::Context context;
-      return &context;
+      if (!ruy_context)
+        ruy_context = new ruy::Context();
+      return ruy_context;
+    }
+
+    void clear_ruy_context() {
+      delete ruy_context;
+      ruy_context = nullptr;
     }
 #endif
   }

src/cpu/backend.h

@@ -28,6 +28,9 @@ namespace ctranslate2 {
     bool pack_gemm_weights(ComputeType compute_type);
 #ifdef CT2_WITH_RUY
     ruy::Context *get_ruy_context();
+    // Destroy the calling thread's ruy::Context (joins ruy's thread pool). Must be
+    // called from a normal execution context, not during thread exit — see backend.cc.
+    void clear_ruy_context();
 #endif
 
   }

src/devices.cc

@@ -4,6 +4,9 @@
 #  include "cuda/utils.h"
 #  include "cuda/random.h"
 #endif
+#ifdef CT2_WITH_RUY
+#  include "cpu/backend.h"
+#endif
 #ifdef CT2_WITH_TENSOR_PARALLEL
 #  include <unistd.h>
 #endif
@@ -125,9 +128,16 @@ namespace ctranslate2 {
       if (device == Device::CUDA) {
           cuda::free_curand_states();
       }
-#else
-      (void)device;
 #endif
+#ifdef CT2_WITH_RUY
+      if (device == Device::CPU) {
+          // Release this worker thread's ruy::Context here (a normal execution
+          // context) rather than at thread exit, where joining ruy's internal
+          // threads deadlocks ThreadPool shutdown on Windows (ctranslate2-rs#64).
+          cpu::clear_ruy_context();
+      }
+#endif
+      (void)device;
   }
 
   // Initialize the static member variable