fix(memory): eliminate memory leaks in Python bindings and inference pipeline

bindings/chatllm.py

@@ -562,6 +562,17 @@ def save_session(self, file_name: str) -> str:
     def load_session(self, file_name: str) -> str:
         return self._lib.load_session(self._chat, file_name)
 
+    def destroy(self) -> int:
+        if hasattr(self, "_chat") and self._chat:
+            if self.is_generating: self.abort()
+            obj_id = LibChatLLM._obj2id.get(self)
+            if obj_id is not None:
+                LibChatLLM._obj2id.pop(self, None)
+                LibChatLLM._id2obj.pop(obj_id, None)
+            self._lib.destroy(self._chat)
+            self._chat = None
+        return 0
+
     def callback_print_reference(self, s: str) -> None:
         self.references.append(s)
 

src/chat.cpp

@@ -627,6 +627,15 @@ namespace chatllm
             qa_encoder->set_tokenizer(this);
     }
 
+    BaseTokenizer::~BaseTokenizer()
+    {
+        if (tp)
+        {
+            delete tp;
+            tp = nullptr;
+        }
+    }
+
     void BaseTokenizer::set_chat_encoder(BaseHistoryEncoder *encoder)
     {
         chat_encoder = encoder;

src/chat.h

@@ -287,7 +287,7 @@ namespace chatllm
                         BaseHistoryEncoder *qa_encoder = nullptr,
                         BaseHistoryEncoder *completion_encoder = nullptr);
 
-        virtual ~BaseTokenizer() = default;
+        virtual ~BaseTokenizer();
 
         virtual size_t load(tokenizer::DataReader *buffer, int n_vocab) = 0;
 

src/layers.h

@@ -311,6 +311,14 @@ namespace chatllm
         PreludeCacheDisable(void): disabler(new BlockParams::DisableCache())
         {
         }
+        virtual ~PreludeCacheDisable()
+        {
+            if (disabler)
+            {
+                delete disabler;
+                disabler = nullptr;
+            }
+        }
     protected:
         BlockParams::DisableCache *disabler;
     };
@@ -1421,7 +1429,7 @@ namespace chatllm
               sinks(BlockParams::CoreAttentionUseSinks::get() > 0 ?
                     ggml::new_tensor_1d(ctx, ggml::type::GGML_TYPE_F32, BlockParams::CoreAttentionUseSinks::get())
                     : nullptr),
-              pos_helper(helper ? helper : &def_pos_helper)
+              pos_helper(helper ? helper : new BaseTensorPosHelper(max_length))
         {
             allocate_pos_tensor(ctx);
         }

src/models.cpp

@@ -925,6 +925,15 @@ namespace chatllm
             layer_ids.push_back(i);
     }
 
+    BaseModelForConditionalGeneration::~BaseModelForConditionalGeneration()
+    {
+        if (transformer)
+        {
+            delete transformer;
+            transformer = nullptr;
+        }
+    }
+
     void BaseModelForConditionalGeneration::set_layer_ids(const std::vector<int> &ids)
     {
         CHATLLM_CHECK((int)ids.size() == config_.num_hidden_layers) << "length(layer_ids) must be " << config_.num_hidden_layers;

src/models_priv.h

@@ -395,7 +395,7 @@ namespace chatllm
     {
     public:
         BaseModelForConditionalGeneration(ModelType model_type, BaseConfig config, const RuntimeConfig &runtime_config, size_t GRAPH_SIZE = 4096);
-        virtual ~BaseModelForConditionalGeneration() = default;
+        virtual ~BaseModelForConditionalGeneration();
 
         void set_layer_ids(const std::vector<int> &ids) override;
         int get_max_length(void) override;

src/tokenizer.h

@@ -106,6 +106,8 @@ class TextPrepAddLeadingSpace : public TextPreprocessor
 class DataReader
 {
 public:
+    virtual ~DataReader() {}
+
     virtual int64_t tell() = 0;
     virtual void seek(int64_t offset, int whence) = 0;
     virtual int64_t size(void) const { return _size; }
@@ -136,6 +138,8 @@ class Processor
         vocab_.byte_fallback_ready = false;
     }
 
+    virtual ~Processor() {}
+
     virtual size_t Load(DataReader *data_reader, int n_vocab) = 0;
 
     virtual int PieceToId(std::string_view piece) const;