#pragma once // clang-format off #include "ggml.h" #include "ggml-backend-impl.h" #include #include #include #include // clang-format on // ggml_tensor is serialized into apir_rpc_tensor struct apir_rpc_tensor { uint64_t id; uint32_t type; uint64_t buffer; uint32_t ne[GGML_MAX_DIMS]; uint32_t nb[GGML_MAX_DIMS]; uint32_t op; int32_t op_params[GGML_MAX_OP_PARAMS / sizeof(int32_t)]; int32_t flags; uint64_t src[GGML_MAX_SRC]; uint64_t view_src; uint64_t view_offs; uint64_t data; char name[GGML_MAX_NAME]; char padding[4]; }; /* frontend */ apir_rpc_tensor apir_serialize_tensor(const ggml_tensor * tensor); void apir_serialize_graph(const ggml_cgraph * cgraph, std::vector & output); /* backend */ void apir_track_backend_buffer(ggml_backend_buffer_t buffer); bool apir_untrack_backend_buffer(ggml_backend_buffer_t buffer); std::unordered_set apir_get_track_backend_buffers(); void apir_add_tensor(ggml_tensor * tensor, std::vector & tensors, std::unordered_set & visited); ggml_tensor * apir_deserialize_tensor(ggml_context * ctx, const apir_rpc_tensor * tensor); ggml_tensor * apir_create_node(uint64_t id, ggml_context * ctx, const std::unordered_map & tensor_ptrs, std::unordered_map & tensor_map); ggml_cgraph * apir_deserialize_graph(uint32_t n_nodes, uint32_t n_tensors, const apir_rpc_tensor * tensors, const uint64_t * nodes);