no duplicated tensor while reading gguf

ggml.c

@@ -20824,6 +20824,14 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p
 
             gguf_tensor_info_sanitize(info);
 
+            // make sure there is no duplicated tensor names
+            for (uint64_t j = 0; j < i; ++j) {
+                if (strcmp(info->name.data, ctx->infos[j].name.data) == 0) {
+                    fprintf(stderr, "%s: duplicated tensor name %s\n", __func__, info->name.data);
+                    ok = false;
+                }
+            }
+
             if (!ok) {
                 fprintf(stderr, "%s: failed to read tensor info\n", __func__);
                 fclose(file);

gguf-py/gguf/gguf_reader.py

@@ -234,8 +234,14 @@ class GGUFReader:
 
     def _build_tensors(self, start_offs: int, fields: list[ReaderField]) -> None:
         tensors = []
+        tensor_names = [] # keep track of name to prevent duplicated tensors
         for field in fields:
             _name_len, name_data, _n_dims, dims, raw_dtype, offset_tensor = field.parts
+            # check if there's any tensor having same name already in the list
+            tensor_name = str(bytes(name_data), encoding = 'utf-8')
+            if tensor_name in tensor_names:
+                raise ValueError(f'Found duplicated tensor with name {tensor_name}')
+            tensor_names.add(tensor_name)
             ggml_type = GGMLQuantizationType(raw_dtype[0])
             n_elems = np.prod(dims)
             block_size, type_size = GGML_QUANT_SIZES[ggml_type]
@@ -267,7 +273,7 @@ class GGUFReader:
                 item_count = n_bytes
                 item_type = np.uint8
             tensors.append(ReaderTensor(
-                name = str(bytes(name_data), encoding = 'utf-8'),
+                name = tensor_name,
                 tensor_type = ggml_type,
                 shape = dims,
                 n_elements = n_elems,

llama.cpp

@@ -3110,9 +3110,17 @@ struct llama_model_loader {
 
         fver = (enum llama_fver) gguf_get_version(meta);
 
+        std::set<std::string> tensor_names;
         for (auto & w : weights) {
             n_elements += ggml_nelements(w.tensor);
             n_bytes    += ggml_nbytes(w.tensor);
+            // make sure there is no duplicated tensor names
+            const std::string name(w.tensor->name);
+            auto found = tensor_names.find(name);
+            if (found != tensor_names.end()) {
+                LLAMA_LOG_ERROR("%s: Found duplicated tensor name %s", __func__, w.tensor->name);
+            }
+            tensor_names.insert(name);
         }
 
         LLAMA_LOG_INFO("%s: loaded meta data with %d key-value pairs and %d tensors from %s (version %s)\n",