You'll never take us alive

We swore that death will do us part
They'll call our crimes a work of art

src/llama-impl.cpp

@@ -1,5 +1,6 @@
 #include "llama-impl.h"
 
+#include "gguf.h"
 #include "llama.h"
 
 #include <cinttypes>
@@ -138,7 +139,7 @@ std::string gguf_kv_to_str(const struct gguf_context * ctx_gguf, int i) {
             {
                 const enum gguf_type arr_type = gguf_get_arr_type(ctx_gguf, i);
                 int arr_n = gguf_get_arr_n(ctx_gguf, i);
-                const void * data = gguf_get_arr_data(ctx_gguf, i);
+                const void * data = arr_type == GGUF_TYPE_STRING ? nullptr : gguf_get_arr_data(ctx_gguf, i);
                 std::stringstream ss;
                 ss << "[";
                 for (int j = 0; j < arr_n; j++) {

src/llama-model-loader.cpp

@@ -22,7 +22,7 @@ const char * llama_file_version_name(llama_fver version) {
 }
 
 namespace GGUFMeta {
-    template <typename T, gguf_type gt_, T (*gfun)(const gguf_context *, const int)>
+    template <typename T, gguf_type gt_, T (*gfun)(const gguf_context *, const int64_t)>
     struct GKV_Base_Type {
         static constexpr gguf_type gt = gt_;
 
@@ -64,10 +64,11 @@ namespace GGUFMeta {
         public:
         static constexpr gguf_type gt = GGUF_TYPE_ARRAY;
         static ArrayInfo getter(const gguf_context *ctx, const int k) {
+            const enum gguf_type arr_type = gguf_get_arr_type(ctx, k);
             return ArrayInfo {
-                gguf_get_arr_type(ctx, k),
+                arr_type,
                 size_t(gguf_get_arr_n(ctx, k)),
-                gguf_get_arr_data(ctx, k),
+                arr_type == GGUF_TYPE_STRING ? nullptr : gguf_get_arr_data(ctx, k),
             };
         }
     };
@@ -558,7 +559,7 @@ llama_model_loader::llama_model_loader(const std::string & fname, bool use_mmap,
             const enum gguf_type type   = gguf_get_kv_type(meta.get(), i);
             const std::string type_name =
                 type == GGUF_TYPE_ARRAY
-                ? format("%s[%s,%d]", gguf_type_name(type), gguf_type_name(gguf_get_arr_type(meta.get(), i)), gguf_get_arr_n(meta.get(), i))
+                ? format("%s[%s,%zu]", gguf_type_name(type), gguf_type_name(gguf_get_arr_type(meta.get(), i)), gguf_get_arr_n(meta.get(), i))
                 : gguf_type_name(type);
 
             std::string value          = gguf_kv_to_str(meta.get(), i);
@@ -988,8 +989,10 @@ bool llama_model_loader::load_all_data(
         bool shouldoffload = (layernum>=0 && clblast_offload_fallback_layers>layernum);
         if(shouldoffload)
         {
-            cur->backend = GGML_BACKEND_TYPE_GPU;
+            cur->clblast_offload_gpu = true;
             ggml_cl_transform_tensor(cur->data, cur);
+        }else{
+            cur->clblast_offload_gpu = false;
         }
         #endif
 

src/llama-quant.cpp

@@ -875,7 +875,8 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
 
         // update the gguf meta data as we go
         gguf_set_tensor_type(ctx_outs[cur_split].get(), name.c_str(), new_type);
-        gguf_set_tensor_data(ctx_outs[cur_split].get(), name.c_str(), new_data, new_size);
+        GGML_ASSERT(gguf_get_tensor_size(ctx_outs[cur_split].get(), gguf_find_tensor(ctx_outs[cur_split].get(), name.c_str())) == new_size);
+        gguf_set_tensor_data(ctx_outs[cur_split].get(), name.c_str(), new_data);
 
         // write tensor data + padding
         fout.write((const char *) new_data, new_size);