llama: fix leaked buffers for mmap + split files (ggml-org#16765)

Author: JohannesGaessler

src/llama-model.cpp

@@ -17,7 +17,6 @@
 
 #include <algorithm>
 #include <cassert>
-#include <cmath>
 #include <cfloat>
 #include <cstring>
 #include <cmath>
@@ -440,7 +439,7 @@ struct llama_model::impl {
     llama_mlocks mlock_mmaps;
 
     // contexts where the model tensors metadata is stored as well ass the corresponding buffers:
-    std::vector<std::pair<ggml_context_ptr, ggml_backend_buffer_ptr>> ctxs_bufs;
+    std::vector<std::pair<ggml_context_ptr, std::vector<ggml_backend_buffer_ptr>>> ctxs_bufs;
 
     buft_list_t cpu_buft_list;
     std::map<ggml_backend_dev_t, buft_list_t> gpu_buft_list;
@@ -6188,7 +6187,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
         bool buffer_from_host_ptr_supported = props.caps.buffer_from_host_ptr;
         bool is_default_buft = buft == ggml_backend_dev_buffer_type(dev);
 
-        ggml_backend_buffer_t buf = nullptr;
+        std::vector<ggml_backend_buffer_ptr> bufs;
         if (ml.use_mmap && use_mmap_buffer && buffer_from_host_ptr_supported && is_default_buft) {
             for (uint32_t idx = 0; idx < ml.files.size(); idx++) {
                 // only the mmap region containing the tensors in the model is mapped to the backend buffer
@@ -6201,15 +6200,16 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                     continue;
                 }
                 const size_t max_size = ggml_get_max_tensor_size(ctx);
-                buf = ggml_backend_dev_buffer_from_host_ptr(dev, (char *) addr + first, last - first, max_size);
+                ggml_backend_buffer_t buf = ggml_backend_dev_buffer_from_host_ptr(dev, (char *) addr + first, last - first, max_size);
                 if (buf == nullptr) {
                     throw std::runtime_error(format("unable to allocate %s buffer", ggml_backend_buft_name(buft)));
                 }
+                bufs.emplace_back(buf);
                 buf_map.emplace(idx, buf);
             }
         }
         else {
-            buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx, buft);
+            ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx, buft);
             if (buf == nullptr) {
                 throw std::runtime_error(format("unable to allocate %s buffer", ggml_backend_buft_name(buft)));
             }
@@ -6219,11 +6219,12 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                 mlock_buf->init   (ggml_backend_buffer_get_base(buf));
                 mlock_buf->grow_to(ggml_backend_buffer_get_size(buf));
             }
+            bufs.emplace_back(buf);
             for (uint32_t idx = 0; idx < ml.files.size(); idx++) {
                 buf_map.emplace(idx, buf);
             }
         }
-        pimpl->ctxs_bufs.emplace_back(std::move(ctx_ptr), buf);
+        pimpl->ctxs_bufs.emplace_back(std::move(ctx_ptr), std::move(bufs));
 
         for (auto & buf : buf_map) {
             // indicate that this buffer contains weights
@@ -6249,8 +6250,11 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
     }
 
     // print memory requirements per buffer type
-    for (auto & [_, buf] : pimpl->ctxs_bufs) {
-        LLAMA_LOG_INFO("%s: %12s model buffer size = %8.2f MiB\n", __func__, ggml_backend_buffer_name(buf.get()), ggml_backend_buffer_get_size(buf.get()) / 1024.0 / 1024.0);
+    for (auto & [_, bufs] : pimpl->ctxs_bufs) {
+        for (auto & buf: bufs) {
+            LLAMA_LOG_INFO("%s: %12s model buffer size = %8.2f MiB\n",
+                __func__, ggml_backend_buffer_name(buf.get()), ggml_backend_buffer_get_size(buf.get()) / 1024.0 / 1024.0);
+        }
     }
 
     // populate tensors_by_name
@@ -6302,8 +6306,10 @@ size_t llama_model::n_devices() const {
 
 std::map<ggml_backend_buffer_type_t, size_t> llama_model::memory_breakdown() const {
     std::map<ggml_backend_buffer_type_t, size_t> ret;
-    for (const auto & [_, buf] : pimpl->ctxs_bufs) {
-        ret[ggml_backend_buffer_get_type(buf.get())] += ggml_backend_buffer_get_size(buf.get());
+    for (const auto & [_, bufs] : pimpl->ctxs_bufs) {
+        for (const auto & buf : bufs) {
+            ret[ggml_backend_buffer_get_type(buf.get())] += ggml_backend_buffer_get_size(buf.get());
+        }
     }
     return ret;
 }