diff --git a/plugins/tensorboard-plugins/tb_graph_ascend/server/app/service/db_graph_service.py b/plugins/tensorboard-plugins/tb_graph_ascend/server/app/service/db_graph_service.py deleted file mode 100644 index 814624848553259b8d0b550b8948f12a62f735b5..0000000000000000000000000000000000000000 --- a/plugins/tensorboard-plugins/tb_graph_ascend/server/app/service/db_graph_service.py +++ /dev/null @@ -1,54 +0,0 @@ -# Copyright (c) 2025, Huawei Technologies. -# All Rights Reserved. -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. -# ============================================================================== -from .base_graph_service import GraphServiceStrategy - - -class DbGraphService(GraphServiceStrategy): - def __init__(self, run_path, tag): - super().__init__(run_path, tag) - - def load_graph_data(self): - pass - - def load_graph_config_info(self): - pass - - def load_graph_all_node_list(self, meta_data): - pass - - def change_node_expand_state(self, node_info, meta_data): - pass - - def get_node_info(self, node_info, meta_data): - pass - - def add_match_nodes(self, npu_node_name, bench_node_name, meta_data, is_match_children): - pass - - def add_match_nodes_by_config(self, config_file, meta_data): - pass - - def delete_match_nodes(self, npu_node_name, bench_node_name, meta_data, is_unmatch_children): - pass - - def save_data(self, meta_data): - pass - - def update_colors(self, colors): - pass - - def save_matched_relations(self): - pass diff --git a/plugins/tensorboard-plugins/tb_graph_ascend/server/app/service/graph_service_db.py b/plugins/tensorboard-plugins/tb_graph_ascend/server/app/service/graph_service_db.py new file mode 100644 index 0000000000000000000000000000000000000000..cef3fc65eb1d65e99afa993985d8022e40365dc5 --- /dev/null +++ b/plugins/tensorboard-plugins/tb_graph_ascend/server/app/service/graph_service_db.py @@ -0,0 +1,464 @@ +# Copyright (c) 2025, Huawei Technologies. +# All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================== +import os +from tensorboard.util import tb_logging + +from .graph_service_base import GraphServiceStrategy +from ..repositories.graph_repo_db import GraphRepoDB +from ..utils.global_state import GraphState +from ..utils.graph_utils import GraphUtils +from ..utils.global_state import NPU, BENCH, SINGLE +from ..model.layout_hierarchy_model import LayoutHierarchyModel +from ..model.match_nodes_model import MatchNodesController +from ..utils.global_state import MAX_RELATIVE_ERR, MIN_RELATIVE_ERR, MEAN_RELATIVE_ERR, NORM_RELATIVE_ERR + +logger = tb_logging.get_logger() + + +class DbGraphService(GraphServiceStrategy): + + def __init__(self, run_path, tag): + super().__init__(run_path, tag) + runs = GraphState.get_global_value('runs') + db_path = os.path.join(runs.get(self.run), f"{tag}.vis.db") + self.repo = GraphRepoDB(db_path) + self.conn = self.repo.get_db_connection() + self.config_info = {} + + def load_graph_data(self): + if not self.repo: + return {'success': False, 'error': 'database not init'} + if not self.conn: + self.conn = self.repo.get_db_connection() + GraphState.set_global_value("all_node_info_cache", {}) # 切换文件清缓存 + return {'success': True} + + def load_graph_config_info(self): + try: + self.config_info = self.repo.query_config_info() + # 读取目录下配置文件列表 + modify_matched_files = GraphUtils.find_config_files(self.run) + self.config_info['matchedConfigFiles'] = modify_matched_files or [] + return {'success': True, 'data': self.config_info} + except Exception as e: + logger.error(f"load graph config info failed, {e}") + return {'success': False, 'error': f'load graph config info failed, {e}'} + + def load_graph_all_node_list(self, meta_data): + try: + if not self.conn: + return {'success': False, 'error': 'database connection not init'} + rank = meta_data.get('rank') + step = meta_data.get('step') + micro_step = meta_data.get('microStep') + result = {} + if not self.config_info: + self.config_info = self.repo.query_config_info() + # 单图 + if self.config_info.get('isSingleGraph'): + # DB：根据graphType rank step micro_step 查询节点列表 + node_name_list = self.repo.query_node_name_list(rank, step, micro_step) + result['npuNodeList'] = node_name_list + return {'success': True, 'data': result} + # 双图 + else: + config_data = GraphState.get_global_value("config_data") + all_node_info = self.repo.query_all_node_info_in_one(rank, step, micro_step) + config_data['npuMatchNodes'] = all_node_info.get('npu_match_node') + config_data['benchMatchNodes'] = all_node_info.get('bench_match_node') + config_data['npuUnMatchNodes'] = all_node_info.get('npu_unmatch_node') + config_data['benchUnMatchNodes'] = all_node_info.get('bench_unmatch_node') + + result['npuMatchNodes'] = all_node_info.get('npu_match_node') + result['benchMatchNodes'] = all_node_info.get('bench_match_node') + result['npuUnMatchNodes'] = all_node_info.get('npu_unmatch_node') + result['benchUnMatchNodes'] = all_node_info.get('bench_unmatch_node') + result['npuNodeList'] = all_node_info.get('npu_node_list') + result['benchNodeList'] = all_node_info.get('bench_node_list') + + return {'success': True, 'data': result} + except Exception as e: + logger.error(f"load graph all node list failed, {e}") + return {'success': False, 'error': f'load graph all node list failed, {e}'} + + def change_node_expand_state(self, node_info, meta_data): + try: + if not self.conn: + return {'success': False, 'error': 'database connection not init'} + graph_type = node_info.get('nodeType') + node_name = node_info.get('nodeName') + rank = meta_data.get('rank') + step = meta_data.get('step') + micro_step = meta_data.get('microStep') + # 单图 + rank_step = { + 'rank': rank, + 'step': step + } + if self.config_info.get('isSingleGraph'): + hierarchy = LayoutHierarchyModel.change_expand_state(node_name, SINGLE, self.repo, micro_step, + rank_step) + # NPU + elif graph_type == NPU: + hierarchy = LayoutHierarchyModel.change_expand_state(node_name, NPU, self.repo, micro_step, rank_step) + # 标杆 + elif graph_type == BENCH: + hierarchy = LayoutHierarchyModel.change_expand_state(node_name, BENCH, self.repo, micro_step, rank_step) + else: + return {'success': True, 'data': {}} + return {'success': True, 'data': hierarchy} + except Exception as e: + logger.error('节点展开或收起发生错误:' + str(e)) + return {'success': False, 'error': f'节点展开或收起发生错误', 'data': None} + + def search_node_by_precision(self, meta_data, values): + try: + if not self.conn: + return {'success': False, 'error': 'database connection not init'} + rank = meta_data.get('rank') + step = meta_data.get('step') + micro_step = meta_data.get('microStep') + + is_filter_unmatch_nodes = True if '无匹配节点' in values else False + if is_filter_unmatch_nodes: + values.remove('无匹配节点') + + update_precision_cache = GraphState.get_global_value("update_precision_cache", {}) + node_name_list = [] + # 先查全局缓存 + if update_precision_cache: + for node_name, node_info in update_precision_cache.items(): + if not node_info.get("is_leaf_nodes"): + continue + matched_node_link = node_info.get('matched_node_link', None) + if is_filter_unmatch_nodes and (not matched_node_link): + node_name_list.append(node_name) + + if isinstance(node_info.get("precision_index"), (int, float, complex)) and any(low <= node_info. + get("precision_index", -1) <= high for low, high in values): + node_name_list.append(node_name) + # 在查数据库 + else: + node_name_list = self.repo.query_node_list_by_precision(step, rank, micro_step, values, + is_filter_unmatch_nodes) + return {'success': True, 'data': node_name_list} + except Exception as e: + logger.error('节点搜索发生错误:' + str(e)) + return {'success': False, 'error': f'节点搜索发生错误', 'data': None} + + def search_node_by_overflow(self, meta_data, values): + try: + if not self.conn: + return {'success': False, 'error': 'database connection not init'} + rank = meta_data.get('rank') + step = meta_data.get('step') + micro_step = meta_data.get('microStep') + + node_name_list = self.repo.query_node_list_by_overflow(step, rank, micro_step, values) + return {'success': True, 'data': node_name_list} + except Exception as e: + logger.error('节点搜索发生错误:' + str(e)) + return {'success': False, 'error': f'节点搜索发生错误', 'data': None} + + def update_hierarchy_data(self, graph_type): + if (graph_type == NPU or graph_type == BENCH): + hierarchy = LayoutHierarchyModel.update_hierarchy_data(graph_type) + return {'success': True, 'data': hierarchy} + else: + return {'success': False, 'error': '节点类型错误'} + + def get_node_info(self, node_info, meta_data): + try: + if not self.conn: + return {'success': False, 'error': 'database connection not init'} + result = {} + graph_type = node_info.get('nodeType') + node_name = node_info.get('nodeName') + rank = meta_data.get('rank') + step = meta_data.get('step') + + if self.config_info.get('isSingleGraph') or graph_type == SINGLE: + result['npu'] = self.repo.query_node_info(node_name, graph_type, rank, step) + else: + matched_node_type = BENCH if graph_type == NPU else NPU + node = self.repo.query_node_info(node_name, graph_type, rank, step) + matched_node_link = node.get('matched_node_link', []) + if isinstance(node.get('matched_node_link', []), list) and len(matched_node_link) > 0: + matched_node = self.repo.query_node_info(matched_node_link[-1], matched_node_type, rank, step) + else: + matched_node = None + result['npu'] = node if graph_type == NPU else matched_node + result['bench'] = node if graph_type == BENCH else matched_node + return {'success': True, 'data': result} + except Exception as e: + logger.error('获取节点信息失败:' + str(e)) + return {'success': False, 'error': '获取节点信息失败:' + str(e), 'data': None} + + def add_match_nodes(self, npu_node_name, bench_node_name, meta_data, is_match_children): + try: + if not self.conn: + return {'success': False, 'error': 'database connection not init'} + rank = meta_data.get('rank') + step = meta_data.get('step') + task = self.config_info.get('task') + # 根据任务类型计算误差 + if task == 'md5' or task == 'summary': + if is_match_children: + graph_data = self.repo.query_node_and_sub_nodes(npu_node_name, bench_node_name, rank, step) + match_result = MatchNodesController.process_task_add_child_layer(graph_data, npu_node_name, + bench_node_name, task) + else: + graph_data = self.repo.query_matched_nodes_info(npu_node_name, bench_node_name, rank, step) + match_result = MatchNodesController.process_task_add(graph_data, npu_node_name, + bench_node_name, task) + # 处理匹配结果 + update_data = [] + for item in match_result: + if item.get('success') is True: + nodes = item.get('data', []) + for node in nodes: + update_data.append(node) + if len(update_data) > 0: + # DB：更新数据库节点信息 + update_db_res = self.repo.update_nodes_info(update_data, rank, step) + if not update_db_res: + return {'success': False, 'error': '更新数据库失败(Update database failed) '} + # 视图：调用更新update_hirarchy方法，同步更新图 + LayoutHierarchyModel.update_current_hierarchy_data(update_data) + # 返回：返回更新后的节点信息 + config_data = GraphState.get_global_value("config_data") + result = { + 'success': True, + 'data': { + 'npuMatchNodes': config_data.get('npuMatchNodes', {}), + 'benchMatchNodes': config_data.get('benchMatchNodes', {}), + 'npuUnMatchNodes': config_data.get('npuUnMatchNodes', []), + 'benchUnMatchNodes': config_data.get('benchUnMatchNodes', []) + } + } + else: + result = {'success': False, 'error': '选择的节点不可匹配(Selected nodes do not match) '} + return result + else: + return {'success': False, 'error': '任务类型不支持(Task type not supported)'} + except Exception as e: + logger.error(str(e)) + return {'success': False, 'error': str(e), 'data': None} + + def add_match_nodes_by_config(self, config_file_name, meta_data): + try: + if not self.conn: + return {'success': False, 'error': 'database connection not init'} + result = {} + rank = meta_data.get('rank') + step = meta_data.get('step') + task = self.config_info.get('task') + match_node_links, error = GraphUtils.safe_load_data(meta_data.get('run'), config_file_name) + graph_data = self.repo.query_matched_nodes_info_by_config(match_node_links, rank, step) + if error: + return {'success': False, 'error': '配置文件失败'} + # 根据任务类型计算误差 + if task == 'md5' or task == 'summary': + match_result = MatchNodesController.process_task_add_child_layer_by_config(graph_data, + match_node_links, task) + update_data = [] + for item in match_result: + if item.get('success') is True: + nodes = item.get('data', []) + for node in nodes: + update_data.append(node) + + if len(update_data) > 0: + update_db_res = self.repo.update_nodes_info(update_data, rank, step) + if not update_db_res: + return {'success': False, 'error': '更新数据库失败(Update database failed) '} + # 视图：调用更新update_hirarchy方法，同步更新图 + LayoutHierarchyModel.update_current_hierarchy_data(update_data) + # 返回：返回更新后的节点信息 + config_data = GraphState.get_global_value("config_data") + result['success'] = True + result['data'] = { + 'matchReslut': match_result, + 'npuMatchNodes': config_data.get('npuMatchNodes', {}), + 'benchMatchNodes': config_data.get('benchMatchNodes', {}), + 'npuUnMatchNodes': config_data.get('npuUnMatchNodes', []), + 'benchUnMatchNodes': config_data.get('benchUnMatchNodes', []) + } + else: + result = {'success': False, 'error': '选择的节点不可匹配(Selected nodes do not match) '} + return result + else: + return {'success': False, 'error': '任务类型不支持(Task type not supported)'} + except Exception as e: + logger.error(str(e)) + return {'success': False, 'error': str(e), 'data': None} + + def delete_match_nodes(self, npu_node_name, bench_node_name, meta_data, is_unmatch_children): + try: + if not self.conn: + return {'success': False, 'error': 'database connection not init'} + rank = meta_data.get('rank') + step = meta_data.get('step') + task = self.config_info.get('task') + + # 根据任务类型计算误差 + if task == 'md5' or task == 'summary': + if is_unmatch_children: + # DB：当前节点及其所有的子节点信息 + graph_data = self.repo.query_node_and_sub_nodes(npu_node_name, bench_node_name, rank, step) + match_result = MatchNodesController.process_task_delete_child_layer(graph_data, npu_node_name, + bench_node_name, task) + else: + graph_data = self.repo.query_matched_nodes_info(npu_node_name, bench_node_name, rank, step) + match_result = MatchNodesController.process_task_delete(graph_data, npu_node_name, + bench_node_name, task) + # 遍历match_result，找到的success=true的节点，合并所有的data字段(数组类型)，合并到update_db_data + update_data = [] + for item in match_result: + if item.get('success') is True: + nodes = item.get('data', []) + for node in nodes: + update_data.append(node) + + if len(update_data) > 0: + # DB：更新数据库节点信息 + update_db_res = self.repo.update_nodes_info(update_data, rank, step) + if not update_db_res: + return {'success': False, 'error': '更新数据库失败(Update database failed) '} + # 视图：调用更新update_hirarchy方法，同步更新图 + LayoutHierarchyModel.update_current_hierarchy_data(update_data) + # 返回：返回更新后的节点信息 + config_data = GraphState.get_global_value("config_data") + result = { + 'success': True, + 'data': { + 'npuMatchNodes': config_data.get('npuMatchNodes', {}), + 'benchMatchNodes': config_data.get('benchMatchNodes', {}), + 'npuUnMatchNodes': config_data.get('npuUnMatchNodes', []), + 'benchUnMatchNodes': config_data.get('benchUnMatchNodes', []) + } + } + else: + result = {'success': False, 'error': '未找到可匹配的节点(Matched node not found) '} + return result + else: + return {'success': False, 'error': '任务类型不支持(Task type not supported) '} + except Exception as e: + logger.error('delete_match_nodes error: {}'.format(e)) + return {'success': False, '操作失败': str(e), 'data': None} + + def update_precision_error(self, meta_data, filter_value): + try: + if not self.conn: + return {'success': False, 'error': 'database connection not init'} + + rank = meta_data.get('rank') + step = meta_data.get('step') + npu_node_list = self.repo.query_node_info_by_data_source(step, rank, NPU) + update_data_hierarchy = {} + update_data_db = [] + + for _, node_info in npu_node_list.items(): + output_statistical_diff = node_info.get('output_data', None) + if not node_info.get('matched_node_link') or not output_statistical_diff: + update_data_hierarchy[node_info.get('node_name')] = { + "precision_index": node_info.get('data').get('precision_index'), + "node_name": node_info.get('node_name'), + "matched_node_link": node_info.get('matched_node_link'), + 'is_leaf_nodes': node_info.get('subnodes', []) == [], + 'graph_type': NPU + } + continue + max_rel_error = -1 + # 根据filter_value 的选择指标计算新的误差值 + for _, diff_values in output_statistical_diff.items(): + filter_diff_rel = [] + if MAX_RELATIVE_ERR in filter_value: + filter_diff_rel.append(diff_values.get('MaxRelativeErr')) + if MIN_RELATIVE_ERR in filter_value: + filter_diff_rel.append(diff_values.get('MinRelativeErr')) + if NORM_RELATIVE_ERR in filter_value: + filter_diff_rel.append(diff_values.get('NormRelativeErr')) + if MEAN_RELATIVE_ERR in filter_value: + filter_diff_rel.append(diff_values.get('MeanRelativeErr')) + # 过滤掉N/A + filter_diff_rel = [x for x in filter_diff_rel if x and x != 'N/A'] + # 如果output指标中存在 Nan/inf/-inf, 直接标记为最大值 + if "Nan" in filter_diff_rel or "inf" in filter_diff_rel or "-inf" in filter_diff_rel: + max_rel_error = 1 + break + filter_diff_rel = [GraphUtils.convert_to_float(x) for x in filter_diff_rel] + max_rel_error_for_key = max(filter_diff_rel) if filter_diff_rel else 0 + max_rel_error = max(max_rel_error, max_rel_error_for_key) + if max_rel_error != -1: + update_data_db.append(( + min(max_rel_error, 1), + step, + rank, + node_info.get('node_name') + )) + update_data_hierarchy[node_info.get('node_name')] = { + "precision_index": min(max_rel_error, 1), + "node_name": node_info.get('node_name'), + "matched_node_link": node_info.get('matched_node_link'), + 'is_leaf_nodes': node_info.get('subnodes', []) == [], + 'graph_type': NPU + } + if len(update_data_hierarchy) > 0: + # 视图：调用更新update_hirarchy方法，同步更新图 + LayoutHierarchyModel.update_current_hierarchy_data(list(update_data_hierarchy.values())) + # 更新全局缓存，使用update_data_hierarchy覆盖原来的缓存即可，注意，切换视图需要重置缓存信息 + GraphState.set_global_value("update_precision_cache", update_data_hierarchy) + return {'success': True, 'data': {}} + else: + return {'success': False, 'error': '未找到可更新的节点(Matched node not found) '} + except Exception as e: + logger.error('update_precision_error error: {}'.format(e)) + return {'success': False, 'error': str(e), 'data': None} + + def update_colors(self, colors): + try: + if not self.conn: + return {'success': False, 'error': 'database connection not init'} + # DB：更新颜色 + update_db_res = self.repo.update_config_colors(colors) + if not update_db_res: + return {'success': False, 'error': '更新数据库失败(Update database failed) '} + return {'success': True} + except Exception as e: + return {'success': False, 'error': str(e), 'data': None} + + def save_matched_relations(self, meta_data): + try: + if not self.conn: + return {'success': False, 'error': 'database connection not init'} + run = meta_data.get('run') + tag = meta_data.get('tag') + rank = meta_data.get('rank') + step = meta_data.get('step') + # DB：根据step rank modify match_node_link查询已经修改的匹配成功的节点关系 + modify_matched_nodes_list = self.repo.query_modify_matched_nodes_list(rank, step) + confilg_file_name = f"{tag}_{step}_{rank}.vis.config" + _, error = GraphUtils.safe_save_data(modify_matched_nodes_list, run, confilg_file_name) + if error: + return {'success': False, 'error': error} + else: + return {'success': True, 'data': confilg_file_name} + + except Exception as e: + logger.error('save_matched_relations error: {}'.format(e)) + return {'success': False, 'error': str(e), 'data': None} +