* Copyright (c) 2025 Huawei Technologies Co., Ltd.
* This program is free software, you can redistribute it and/or modify it under the terms and conditions of
* CANN Open Software License Agreement Version 2.0 (the "License").
* Please refer to the License for details. You may not use this file except in compliance with the License.
* THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
* INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
* See LICENSE in the root of the software repository for the full text of the License.
*/
syntax = "proto3";
package deployer;
import "var_manager.proto";
import "ge_ir.proto";
service DeployerService {
rpc DeployerProcess(DeployerRequest) returns (DeployerResponse) {}
}
enum DeployerRequestType {
kInitRequest = 0;
kDisconnect = 1;
kLoadModel = 4;
kUnloadModel = 5;
kDownloadVarManager = 7;
kDownloadSharedContent = 8;
kInitProcessResource = 9;
kDownloadConf = 10;
kUpdateDeployPlan = 11;
kTransferFile = 12;
kAddFlowRoutePlan = 13;
kDatagwSchedInfo = 14;
kClearModelData = 15;
kDataFlowExceptionNotify = 16;
kUpdateProfilingInfo = 17;
kHeartbeat = 127;
}
enum DeviceConfigType {
kLogConfig = 0;
kDumpConfig = 1;
kProfilingConfig = 2;
kCommonConfig = 127;
}
message Options {
map<string, string> global_options = 1;
map<string, string> session_options = 2;
map<string, string> graph_options = 3;
}
message DeployerRequest {
int64 client_id = 1;
uint64 message_id = 2;
DeployerRequestType type = 3;
oneof body {
InitRequest init_request = 4;
LoadModelRequest load_model_request = 6;
UnloadModelRequest unload_model_request = 7;
MultiVarManagerRequest multi_var_manager_request = 9;
SharedContentDescRequest shared_content_desc_request = 10;
InitProcessResourceRequest init_process_resource_request = 11;
DownloadConfigRequest download_config_request = 12;
UpdateDeployPlanRequest update_deploy_plan_request = 13;
TransferFileRequest transfer_file_request = 14;
AddFlowRoutePlanRequest add_flow_route_plan_request = 15;
DataGwSchedInfos datagw_sched_info = 16;
ClearModelDataRequest model_data_clear = 17;
DataFlowExceptionNotifyRequest exception_notify_request = 18;
SendProfInfoRequest prof_info = 19;
}
}
message DeployerResponse {
uint64 message_id = 1;
uint32 error_code = 2;
string error_message = 3;
oneof body {
InitResponse init_response = 4;
HeartbeatResponse heartbeat_response = 5;
}
}
message InitRequest {
int32 logic_device_id = 2;
string type = 3;
string sign_data = 4;
map<string, string> envs = 5;
}
message InitResponse {
int64 client_id = 1;
string dgw_host_ip = 2;
int32 dgw_port = 3;
int32 dev_count = 4;
int32 dgw_port_offset = 5;
bool support_flowgw_merged = 6;
}
message DeviceStatus {
int32 device_id = 1;
int32 device_type = 2;
}
message SubmodelInstanceName {
map<string, bool> submodel_instance_name = 1;
}
message DeviceError {
repeated uint32 error_code = 1;
}
message HeartbeatResponse {
uint32 abnormal_type = 1;
repeated DeviceStatus device_status = 2;
map<uint32, SubmodelInstanceName> abnormal_submodel_instance_name = 3;
map<uint32, DeviceError> abnormal_device = 4;
}
message DataFlowExceptionNotify {
uint32 type = 1;
int32 exception_code = 2;
uint64 trans_id = 3;
string scope = 4;
uint64 user_context_id = 5;
bytes exception_context = 6;
}
message DataFlowExceptionNotifyRequest {
uint32 root_model_id = 1;
DataFlowExceptionNotify exception_notify = 2;
}
message DownloadConfigRequest {
DeviceConfigType sub_type = 1;
int32 device_id = 2;
bytes config_data = 3;
}
message ModelQueueIndices {
repeated int32 input_queue_indices = 1;
repeated int32 output_queue_indices = 2;
}
message InputAlignAttrs {
uint32 align_max_cache_num = 1;
int32 align_timeout = 2;
bool drop_when_not_align = 3;
}
message SubmodelDesc {
uint32 submodel_id = 1;
string model_name = 2;
string model_path = 3;
uint64 model_size = 4;
repeated int32 input_queue_indices = 5;
repeated int32 output_queue_indices = 6;
bool is_dynamic = 7;
string engine_name = 8;
uint32 replica_num = 9;
uint32 replica_idx = 10;
map<string, string> attrs = 11;
map<string, ModelQueueIndices> invoked_model_queues = 12;
repeated int32 status_input_queue_indices = 13;
repeated int32 status_output_queue_indices = 14;
string rank_id = 15;
int32 execute_times = 18;
int32 process_id = 19;
int32 phy_device_id = 21;
bool is_head = 23;
string model_instance_name = 24;
InputAlignAttrs input_align_attrs = 25;
string saved_model_file_path = 26;
bool is_remote_model = 27;
bool is_builtin_udf = 28;
uint32 replica_idx_on_node = 29;
bool enable_exception_catch = 30;
string scope = 31;
}
message QueueDesc {
int32 type = 1;
string name = 2;
uint32 depth = 3;
string enqueue_policy = 4;
int32 fusion_offset = 5;
int32 ref_index = 6;
}
message TagDesc {
string name = 1;
uint32 tag_id = 2;
uint32 rank_id = 3;
uint32 peer_rank_id = 4;
uint32 depth = 5;
int32 peer_node_id = 6;
int32 peer_device_id = 7;
int32 peer_device_type = 8;
}
message GroupDesc {
repeated int32 endpoint_indices = 1;
bool keep_out_of_order = 2;
}
message EndpointDesc {
int32 type = 1;
string name = 2;
oneof endpoint {
QueueDesc queue_desc = 3;
TagDesc tag_desc = 4;
GroupDesc group_desc = 5;
}
uint32 instance_num = 6;
uint32 instance_idx = 7;
int32 device_id = 8;
int32 device_type = 9;
uint32 model_id = 10;
bool is_dynamic_sched = 11;
uint32 root_model_id = 12;
int32 node_id = 13;
}
message EndpointBinding {
int32 src_index = 1;
int32 dst_index = 2;
}
message FlowRoutePlan {
repeated EndpointDesc endpoints = 1;
repeated EndpointBinding bindings = 2;
int64 route_id = 3;
bool has_proxy_q = 4;
repeated EndpointBinding bindings_before_load = 5;
}
message MultiVarManagerRequest {
MultiVarManagerInfo multi_var_manager_info = 1;
repeated int32 device_ids = 3;
}
message SharedContentDescRequest {
SharedContentDescription shared_content_desc = 1;
repeated int32 device_ids = 2;
FlowRoutePlan flow_route = 3;
}
message InitProcessResourceRequest {
int32 device_id = 1;
string rank_table = 2;
int32 rank_table_type = 3;
int32 rank_id = 4;
repeated int32 res_ids = 5;
int32 device_type = 6;
bool profiling_on = 7;
string remote_group_cache_alloc_config = 8;
}
message TransferFileRequest {
string path = 1;
bytes content = 2;
bool eof = 3;
}
message ModelMemSize {
uint32 std_mem_size = 1;
uint32 shared_mem_size = 2;
}
message UpdateDeployPlanRequest {
int32 device_id = 1;
uint64 session_id = 2;
uint32 root_model_id = 3;
repeated SubmodelDesc submodel_descs = 6;
uint32 graph_id = 11;
int32 device_type = 12;
Options options = 14;
ModelMemSize model_mem_size = 15;
}
message AddFlowRoutePlanRequest {
int32 node_id = 1;
uint32 root_model_id = 2;
FlowRoutePlan flow_route_plan = 3;
}
message LoadModelRequest {
uint32 root_model_id = 1;
}
message UnloadModelRequest {
uint32 model_id = 1;
int32 device_id = 2;
}
enum ExecutorRequestType {
kExecutorPreDownload = 0;
kExecutorDownload = 1;
kExecutorLoad = 2;
kExecutorUnload = 3;
kExecutorFinalize = 4;
kExecutorClear = 5;
kExecutorExceptionNotify = 6;
kNotify = 7;
}
message FileConstantDesc {
string name = 1;
ge.proto.TensorDescriptor tensor_desc = 2;
string path = 3;
uint64 offset = 4;
}
message ExecutorRequest {
message PreDownloadModelRequest {
uint32 model_id = 1;
uint32 root_model_id = 2;
uint64 model_size = 3;
}
message DownloadModelRequest {
uint32 model_id = 1;
uint32 root_model_id = 2;
uint64 offset = 3;
bytes model_data = 4;
}
message SyncVarManageRequest {
uint64 var_mem_base = 1;
uint64 var_mem_size = 2;
VarManagerInfo var_manager_info = 3;
repeated SharedContentDescription shared_content_descs = 4;
}
message ModelQueueIds {
repeated uint32 input_queues = 1;
repeated uint32 output_queues = 2;
}
message QueueAttrs {
uint32 queue_id = 1;
int32 device_type = 2;
int32 device_id = 3;
uint32 global_logic_id = 4;
}
message ModelQueuesAttrs {
repeated QueueAttrs input_queues_attrs = 1;
repeated QueueAttrs output_queues_attrs = 2;
}
message LoadModelRequest {
uint32 root_model_id = 1;
uint32 model_id = 2;
string model_path = 3;
bool is_dynamic = 4;
uint32 replica_num = 7;
uint32 replica_idx = 8;
map<string, string> attrs = 9;
ModelQueuesAttrs status_queues = 11;
uint32 model_uuid = 14;
bool is_dynamic_sched = 15;
bool need_report_status = 16;
int32 execute_times = 18;
repeated int32 input_fusion_offsets = 19;
int32 phy_device_id = 21;
bool is_dynamic_proxy_controlled = 22;
ModelQueuesAttrs model_queues_attrs = 23;
map<string, ModelQueuesAttrs> invoked_model_queues_attrs = 24;
bool is_head = 25;
InputAlignAttrs input_align_attrs = 26;
string model_instance_name = 27;
string saved_model_file_path = 28;
bool is_builtin_udf = 29;
bool enable_exception_catch = 30;
string scope = 31;
}
message BatchLoadModelMessage {
VarManagerInfo var_manager_info = 1;
repeated FileConstantDesc file_constant_descs = 2;
repeated LoadModelRequest models = 5;
uint32 root_model_id = 7;
uint32 graph_id = 10;
Options options = 12;
}
message UnloadModelMessage {
uint32 model_id = 1;
}
message ClearModelRequest {
uint32 model_id = 1;
int32 clear_msg_type = 2;
}
message UpdateProfRequest {
bool is_prof_start = 1;
string prof_data = 2;
}
oneof body {
SyncVarManageRequest sync_var_manager_message = 1;
BatchLoadModelMessage batch_load_model_message = 2;
LoadModelRequest load_model_message = 4;
UnloadModelMessage unload_model_message = 5;
ClearModelRequest clear_model_message = 6;
DataFlowExceptionNotifyRequest exception_notify_request = 7;
UpdateProfRequest update_prof_message = 8;
}
ExecutorRequestType type = 9;
uint64 message_id = 10;
}
message ExecutorResponse {
uint64 message_id = 1;
uint32 error_code = 2;
string error_message = 3;
}
message DataGwSchedInfos {
uint32 root_model_id = 1;
int32 device_id = 2;
int32 device_type = 3;
uint32 input_queue_indice = 4;
uint32 output_queue_indice = 5;
bool is_dynamic_sched = 6;
bool is_proxy = 7;
}
message QueueAttrs {
uint32 queue_id = 1;
int32 device_type = 2;
int32 device_id = 3;
uint32 logic_id = 4;
}
message QueueStatus {
QueueAttrs queue_attrs = 1;
uint32 queue_depth = 2;
uint32 input_consume_num = 3;
}
message SubmodelStatus {
uint32 model_uuid = 1;
repeated QueueStatus queue_statuses = 2;
}
message ExceptionDevInfo {
int32 device_id = 1;
int32 device_type = 2;
int32 node_id = 3;
}
message ClearModelDataRequest {
repeated uint32 root_model_ids = 1;
repeated ExceptionDevInfo exception_dev_info = 2;
int32 clear_type = 3;
int32 node_id = 4;
}
message SendProfInfoRequest {
bool is_prof_start = 1;
string prof_data = 2;
uint32 model_id = 3;
}
