已合并
profiler: set the msmonitor status to uninitialized when the polling subprocess ends #27273
xfeng创建于 2025年11月29日
profiler: set the msmonitor status to uninitialized when the polling subprocess ends #27273
已合并
共 2 个文件变更+11-1
| @@ -244,4 +244,12 @@ def worker_dyno_func(params_dict): | |||
| 244 | except Exception as ex: | 244 | except Exception as ex: |
| 245 | dynamic_profiler_utils.out_log("Dynamic profiler cfg bytes write failed, {} has occur!".format(str(ex)), | 245 | dynamic_profiler_utils.out_log("Dynamic profiler cfg bytes write failed, {} has occur!".format(str(ex)), |
| 246 | dynamic_profiler_utils.LoggerLevelEnum.ERROR) | 246 | dynamic_profiler_utils.LoggerLevelEnum.ERROR) |
| 247 | + if hasattr(py_dyno_monitor, "update_profiler_status"): | ||
| 248 | + profiler_status = { | ||
O | |||
| 249 | + DynamicProfilerUtils.PROFILER_STATUS: str(DynamicProfilerUtils.ProfilerStatus.UNINITIALIZED.value), | ||
| 250 | + DynamicProfilerUtils.CURRENT_STEP: "-1", | ||
| 251 | + DynamicProfilerUtils.START_STEP: "-1", | ||
| 252 | + DynamicProfilerUtils.STOP_STEP: "-1", | ||
| 253 | + } | ||
| 254 | + py_dyno_monitor.update_profiler_status(profiler_status) | ||
| 247 | dynamic_profiler_utils.out_log("Dynolog profiler process done", dynamic_profiler_utils.LoggerLevelEnum.INFO) | 255 | dynamic_profiler_utils.out_log("Dynolog profiler process done", dynamic_profiler_utils.LoggerLevelEnum.INFO) |
| @@ -40,6 +40,7 @@ class _DynamicProfile: | |||
| 40 | self._step_record_time = None | 40 | self._step_record_time = None |
| 41 | self._step_time = 0 | 41 | self._step_time = 0 |
| 42 | self._min_poll_interval = 1 | 42 | self._min_poll_interval = 1 |
| 43 | + self._max_poll_interval = 50 | ||
| 43 | self._step_mstx_range_id = 0 | 44 | self._step_mstx_range_id = 0 |
| 44 | 45 | ||
| 45 | def init(self): | 46 | def init(self): |
| @@ -78,7 +79,8 @@ class _DynamicProfile: | |||
| 78 | if self.cur_step == self.RECORD_TIME_STEP: | 79 | if self.cur_step == self.RECORD_TIME_STEP: |
| 79 | self._step_record_time = time.time() | 80 | self._step_record_time = time.time() |
| 80 | elif self.cur_step - self.RECORD_TIME_STEP == 1: | 81 | elif self.cur_step - self.RECORD_TIME_STEP == 1: |
| 81 | - self._step_time = max(self._min_poll_interval, int(time.time() - self._step_record_time)) | 82 | + self._step_time = min(self._max_poll_interval, |
| 83 | + max(self._min_poll_interval, int(time.time() - self._step_record_time))) | ||
| 82 | self._dynamic_monitor.modify_step_time(self._step_time) | 84 | self._dynamic_monitor.modify_step_time(self._step_time) |
| 83 | if self._step_mstx_range_id: | 85 | if self._step_mstx_range_id: |
| 84 | mstx.range_end(self._step_mstx_range_id) | 86 | mstx.range_end(self._step_mstx_range_id) |
代码逻辑和结构: 在worker_dyno_func函数中，profiler_status字典的创建存在硬编码问题。该字典的结构与DynamicProfilerMonitor类中第26-31行定义的_profiler_status字典结构相同，但代码中重复定义了相同的字典结构，违反了DRY原则。
问题类型: 代码逻辑和结构 文件路径:
torch_npu/profiler/_dynamic_profiler/_dynamic_profiler_monitor.py行号: 248 问题代码:profiler_status = { DynamicProfilerUtils.PROFILER_STATUS: str(DynamicProfilerUtils.ProfilerStatus.UNINITIALIZED.value), DynamicProfilerUtils.CURRENT_STEP: "-1", DynamicProfilerUtils.START_STEP: "-1", DynamicProfilerUtils.STOP_STEP: "-1", }修改建议:
此评论由代码审查工具自动生成