QVAC-18422 [TTS GGML] Optimize cpp backend multilingual for CPU

CMakeLists.txt

@@ -205,6 +205,24 @@ if (TTS_CPP_BUILD_TESTS)
     target_link_libraries(test-streaming PRIVATE ggml)
     target_include_directories(test-streaming PRIVATE ggml/include src include)
 
+    # CPU-side persistent-cache validation (QVAC-18422).
+    # Exercises the time_mlp / time_emb / cfm_estimator / weight_mirror
+    # caches that amortise per-synth overhead on the multilingual CPU
+    # path.  Links the chatterbox_tts.cpp directly so it can reach the
+    # internal test-hook entrypoints.
+    add_executable(test-cpu-caches
+        src/test_cpu_caches.cpp
+        src/chatterbox_tts.cpp)
+    target_link_libraries(test-cpu-caches PRIVATE ggml)
+    target_include_directories(test-cpu-caches PRIVATE ggml/include src include)
+
+    # T3 step-graph cache validation (QVAC-18422 round 4).  Links
+    # against the full tts-cpp library so it gets t3_mtl.cpp's
+    # cached eval_step_mtl alongside the test-hook entrypoints.
+    add_executable(test-t3-caches src/test_t3_caches.cpp)
+    target_link_libraries(test-t3-caches PRIVATE tts-cpp ggml)
+    target_include_directories(test-t3-caches PRIVATE ggml/include src include)
+
     add_executable(test-metal-ops src/test_metal_ops.cpp)
     target_link_libraries(test-metal-ops PRIVATE ggml)
     target_include_directories(test-metal-ops PRIVATE ggml/include src)

src/chatterbox_cli.cpp

@@ -1183,6 +1183,12 @@ int tts_cpp_cli_main(int argc, char ** argv) {
                     tts_cpp::chatterbox::detail::t3_stack_unregister(
                         model.buffer_stack, model.ctx_stack);
                 }
+                // QVAC-18422 round 4: drop the T3 step-graph cache
+                // BEFORE freeing the backend.  The cache holds
+                // gallocators that carry backend references; freeing
+                // them against a dead backend would assert inside the
+                // ggml-metal / ggml-vulkan / ggml-cuda dylib finalisers.
+                tts_cpp::chatterbox::detail::t3_release_caches();
                 ggml_backend_buffer_free(model.buffer_w);
                 ggml_backend_buffer_free(model.buffer_kv);
                 if (model.buffer_stack)    ggml_backend_buffer_free(model.buffer_stack);
@@ -2332,6 +2338,9 @@ int tts_cpp_cli_main(int argc, char ** argv) {
                 (long long)t3_total_ms, t3_tokens_total);
 
         ggml_gallocr_free(allocr);
+        // QVAC-18422 round 4: drop T3 step-graph cache BEFORE freeing
+        // the backend (gallocators in cached entries reference it).
+        tts_cpp::chatterbox::detail::t3_release_caches();
         ggml_backend_buffer_free(model.buffer_w);
         ggml_backend_buffer_free(model.buffer_kv);
         if (model.buffer_override) ggml_backend_buffer_free(model.buffer_override);

src/chatterbox_engine.cpp

@@ -165,6 +165,11 @@ struct Engine::Impl {
         if (model.buffer_stack || model.ctx_stack) {
             t3_stack_unregister(model.buffer_stack, model.ctx_stack);
         }
+        // QVAC-18422 round 4: drop the T3 step-graph cache BEFORE
+        // freeing the backend.  Cached gallocators carry backend
+        // references; freeing them against a dead backend asserts
+        // inside the GPU-backend dylib finalisers.
+        tts_cpp::chatterbox::detail::t3_release_caches();
         if (model.buffer_w)        { ggml_backend_buffer_free(model.buffer_w);        model.buffer_w        = nullptr; }
         if (model.buffer_kv)       { ggml_backend_buffer_free(model.buffer_kv);       model.buffer_kv       = nullptr; }
         if (model.buffer_stack)    { ggml_backend_buffer_free(model.buffer_stack);    model.buffer_stack    = nullptr; }

src/chatterbox_t3_internal.h

@@ -347,6 +347,16 @@ bool eval_step_mtl(
     std::vector<float> &     logits_cond_out,
     std::vector<float> &     logits_uncond_out);
 
+// Release every persistent T3-side cache held in this translation
+// unit (currently the round-4 step-graph cache).  Idempotent.
+//
+// Production callers (CLI free_t3 lambda, Engine::Impl::free_model)
+// MUST call this BEFORE `ggml_backend_free(model.backend)` because
+// the cached gallocators carry backend references; freeing them
+// against a freed backend would assert inside ggml-metal /
+// ggml-vulkan / ggml-cuda dylib finalisers.
+void t3_release_caches();
+
 // On a degenerate logits distribution (everything -inf after the sampling
 // cascade), returns `stop_token` so the caller's stop check fires cleanly
 // instead of emitting a pseudo-random in-vocab id.  Pass