vad : fix memory leak by storing ggml_context in vad context struct

danbev · danbev · commit aeafca437efa · 2025-10-04T08:51:23.000+02:00
This commit addresses a memory leak issue in the voice activity detection (VAD) where the ggml_context is not stored within the vad context structure. The motivation for this change that this is causing the context memory to stay allocated and the tensor still point to that memory but this memory is never freed. Resolves: ggml-org#3452
diff --git a/src/whisper.cpp b/src/whisper.cpp
@@ -4402,6 +4402,7 @@ struct whisper_vad_context {
     std::vector<ggml_backend_t> backends;
     ggml_backend_buffer_t       buffer = nullptr;
     whisper_context_params      params;
+    ggml_context *              ctx = nullptr;
     std::vector<uint8_t>        ctx_buf;
     whisper_sched               sched;
 
@@ -4661,21 +4662,21 @@ static bool whisper_vad_init_context(whisper_vad_context * vctx) {
         /*.no_alloc   =*/ true,
     };
 
-    ggml_context * ctx = ggml_init(params);
-    if (!ctx) {
+    vctx->ctx = ggml_init(params);
+    if (!vctx->ctx) {
         WHISPER_LOG_ERROR("%s: failed to init LSTM state ggml context\n", __func__);
         return false;
     }
 
     // LSTM Hidden state
-    vctx->h_state = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, lstm_hidden_size);
+    vctx->h_state = ggml_new_tensor_1d(vctx->ctx, GGML_TYPE_F32, lstm_hidden_size);
     ggml_set_name(vctx->h_state, "h_state");
 
     // LSTM Cell state
-    vctx->c_state = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, lstm_hidden_size);
+    vctx->c_state = ggml_new_tensor_1d(vctx->ctx, GGML_TYPE_F32, lstm_hidden_size);
     ggml_set_name(vctx->c_state, "c_state");
 
-    vctx->buffer = ggml_backend_alloc_ctx_tensors(ctx, vctx->backends[0]);
+    vctx->buffer = ggml_backend_alloc_ctx_tensors(vctx->ctx, vctx->backends[0]);
     if (!vctx->buffer) {
         WHISPER_LOG_ERROR("%s: failed to allocate memory for the VAD state\n", __func__);
         return false;
@@ -5433,7 +5434,7 @@ void whisper_vad_free(whisper_vad_context * ctx) {
         for (auto & backend : ctx->backends) {
             ggml_backend_free(backend);
         }
-
+        ggml_free(ctx->ctx);
 
         delete ctx;
     }
