mindstudio-probe 1.0.1__py3-none-any.whl → 1.0.3__py3-none-any.whl

msprobe/mindspore/ms_config.py

@@ -1,6 +1,12 @@
 import json
+
 from msprobe.core.common_config import CommonConfig, BaseConfig
 from msprobe.core.common.file_check import FileOpen
+from msprobe.core.common.const import Const
+from msprobe.mindspore.common.const import FreeBenchmarkConst
+from msprobe.mindspore.common.log import logger
+from msprobe.core.grad_probe.constant import level_adp
+from msprobe.core.grad_probe.utils import check_numeral_list_ascend
 
 
 class TensorConfig(BaseConfig):
@@ -31,39 +37,81 @@ class StatisticsConfig(BaseConfig):
         if self.data_mode is not None and len(self.data_mode) > 0:
             if len(self.data_mode) > 1 or self.data_mode[0] not in ["all", "input", "output"]:
                 raise Exception("data_mode must be all, input or output")
+        if self.summary_mode and self.summary_mode not in ["statistics", "md5"]:
+            raise Exception("summary_mode is invalid")
 
 
-class OverflowCheck(BaseConfig):
+class OverflowCheckConfig(BaseConfig):
     def __init__(self, json_config):
         super().__init__(json_config)
-        self.file_format = None
-        self.check_mode = json_config.get("check_mode")
+        self.data_mode = ["all"]
         self._check_config()
 
     def _check_config(self):
-        if self.data_mode is not None and len(self.data_mode) > 0:
-            if len(self.data_mode) > 1 or self.data_mode[0] not in ["all", "input", "output"]:
-                raise Exception("data_mode must be all, input or output")
+        if self.overflow_nums is not None and not isinstance(self.overflow_nums, int):
+            raise Exception("overflow_nums is invalid, it should be an integer")
+        if self.overflow_nums is not None and self.overflow_nums != -1 and self.overflow_nums <= 0:
+            raise Exception("overflow_nums should be -1 or positive integer")
         if self.check_mode and self.check_mode not in ["all", "aicore", "atomic"]:
             raise Exception("check_mode is invalid")
 
 
+class FreeBenchmarkConfig(BaseConfig):
+    def __init__(self, task_config):
+        super().__init__(task_config)
+        self._check_config()
+
+    def _check_config(self):
+        if self.fuzz_device and self.fuzz_device not in FreeBenchmarkConst.DEVICE_LIST:
+            raise Exception("fuzz_device must be npu or empty")
+        if self.pert_mode and self.pert_mode not in FreeBenchmarkConst.PERT_TYPE_LIST:
+            raise Exception("pert_mode must be improve_precision, add_noise, bit_noise, no_change or empty")
+        if self.handler_type and self.handler_type not in FreeBenchmarkConst.HANDLER_TYPE_LIST:
+            raise Exception("handler_type must be check, fix or empty")
+        if self.fuzz_level and self.fuzz_level not in FreeBenchmarkConst.DUMP_LEVEL_LIST:
+            raise Exception("fuzz_level must be L1 or empty")
+        if self.fuzz_stage and self.fuzz_stage not in FreeBenchmarkConst.STAGE_LIST:
+            raise Exception("fuzz_stage must be forward or empty")
+        if self.if_preheat or self.preheat_step or self.max_sample:
+            logger.warning("'if_preheat', 'preheat_step' and 'max_sample' settings "
+                           "are not supported for mindspore free benchmark task.")
+
+
+class GradProbeConfig(BaseConfig):
+    def __init__(self, json_config):
+        super().__init__(json_config)
+        self.grad_level = json_config.get("grad_level", "L1")
+        self.param_list = json_config.get("param_list", [])
+        self.bounds = json_config.get("bounds", [])
+    
+    def _check_config(self):
+        if self.grad_level not in level_adp.keys():
+            raise Exception(f"grad_level must be one of {level_adp.keys()}")
+        if not isinstance(self.param_list, list):
+            raise Exception(f"param_list must be a list")
+        check_numeral_list_ascend(self.bounds)
+
+
+TaskDict = {
+    Const.TENSOR: TensorConfig,
+    Const.STATISTICS: StatisticsConfig,
+    Const.OVERFLOW_CHECK: OverflowCheckConfig,
+    Const.FREE_BENCHMARK: FreeBenchmarkConfig,
+    Const.GRAD_PROBE: GradProbeConfig,
+}
+
+
 def parse_common_config(json_config):
     return CommonConfig(json_config)
 
 
 def parse_task_config(task, json_config):
-    task_map = json_config[task]
+    task_map = json_config.get(task)
     if not task_map:
         task_map = dict()
-    if task == "tensor":
-        return TensorConfig(task_map)
-    elif task == "statistics":
-        return StatisticsConfig(task_map)
-    elif task == "overflow_check":
-        return OverflowCheck(task_map)
-    else:
+    if task not in TaskDict:
         raise Exception("task is invalid.")
+    return TaskDict.get(task)(task_map)
 
 
 def parse_json_config(json_file_path):
@@ -73,6 +121,6 @@ def parse_json_config(json_file_path):
         json_config = json.load(file)
     common_config = parse_common_config(json_config)
     if not common_config.task:
-        common_config.task = "statistics"
+        common_config.task = Const.STATISTICS
     task_config = parse_task_config(common_config.task, json_config)
     return common_config, task_config

msprobe/mindspore/overflow_check/overflow_check_tool_factory.py

@@ -1,23 +1,24 @@
+from msprobe.mindspore.common.const import Const
 from msprobe.mindspore.debugger.debugger_config import DebuggerConfig
 from msprobe.mindspore.overflow_check.kernel_graph_overflow_check import KernelGraphOverflowCheck
 
 
 class OverflowCheckToolFactory:
     tools = {
-        "cell": {
-            "kbk": None,
-            "graph": None,
-            "pynative": None
+        Const.CELL: {
+            Const.GRAPH_KBYK_MODE: None,
+            Const.GRAPH_GE_MODE: None,
+            Const.PYNATIVE_MODE: None
         },
-        "api": {
-            "kbk": None,
-            "graph": None,
-            "pynative": None
+        Const.API: {
+            Const.GRAPH_KBYK_MODE: None,
+            Const.GRAPH_GE_MODE: None,
+            Const.PYNATIVE_MODE: None
         },
-        "kernel": {
-            "kbk": None,
-            "graph": KernelGraphOverflowCheck,
-            "pynative": None
+        Const.KERNEL: {
+            Const.GRAPH_KBYK_MODE: None,
+            Const.GRAPH_GE_MODE: KernelGraphOverflowCheck,
+            Const.PYNATIVE_MODE: None
         }
     }
 
@@ -25,8 +26,9 @@ class OverflowCheckToolFactory:
     def create(config: DebuggerConfig):
         tool = OverflowCheckToolFactory.tools.get(config.level)
         if not tool:
-            raise Exception("valid level is needed.")
-        tool = tool.get("graph")
+            raise Exception("Valid level is needed.")
+        tool = tool.get(config.execution_mode)
         if not tool:
-            raise Exception("Overflow check in not supported in this mode.")
+            raise Exception(f"Overflow check is not supported in {config.execution_mode} mode "
+                            f"when level is {config.level}.")
         return tool(config)

msprobe/mindspore/runtime.py

@@ -0,0 +1,4 @@
+class Runtime:
+    step_count: int = 0
+    rank_id: int = -1
+    is_running: bool = False

msprobe/mindspore/service.py

@@ -0,0 +1,354 @@
+# Copyright 2024 Huawei Technologies Co., Ltd
+#
+# 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
+import copy
+from pathlib import Path
+import functools
+from collections import defaultdict
+
+import mindspore as ms
+from mindspore.common.tensor import Tensor
+from mindspore import ops
+from mindspore import nn
+try:
+    from mindspore.common._pijit_context import PIJitCaptureContext
+    pijit_label = True
+except ImportError:
+    pijit_label = False
+
+
+from msprobe.core.data_dump.data_collector import build_data_collector
+from msprobe.core.data_dump.scope import BaseScope
+from msprobe.mindspore.common.utils import get_rank_if_initialized
+from msprobe.core.common.file_check import FileChecker, FileCheckConst, check_path_before_create
+from msprobe.mindspore.common.log import logger
+from msprobe.core.common.utils import Const
+from msprobe.core.common.exceptions import DistributedNotInitializedError
+from msprobe.mindspore.dump.hook_cell.api_registry import api_register
+from msprobe.core.data_dump.data_processor.base import ModuleBackwardInputsOutputs, ModuleForwardInputsOutputs, \
+    ModuleBackwardInputs, ModuleBackwardOutputs
+from msprobe.core.common.exceptions import MsprobeException
+from msprobe.mindspore.dump.hook_cell.hook_cell import HOOKCell
+from msprobe.mindspore.cell_processor import CellProcessor
+from msprobe.mindspore.dump.jit_dump import JitDump
+
+
+class Service:
+    def __init__(self, config):
+        self.model = None
+        self.config = copy.deepcopy(config)
+        self.config.level = self.config.level_ori
+        self.data_collector = build_data_collector(self.config)
+        self.cell_processor = CellProcessor(self.data_collector.scope)
+        self.switch = False
+        self.current_iter = 0
+        self.first_start = True
+        self.current_rank = None
+        self.primitive_counters = {}
+        self.dump_iter_dir = None
+        self.start_call = False
+        self.check_level_valid()
+
+    @staticmethod
+    def check_model_valid(model):
+        if not model or isinstance(model, nn.Cell):
+            return model
+        raise MsprobeException(
+            MsprobeException.INVALID_PARAM_ERROR, "model 参数必须是 mindspore.nn.Cell 类型。"
+        )
+
+    def check_level_valid(self):
+        if self.config.level == "L2":
+            raise MsprobeException(
+                MsprobeException.INVALID_PARAM_ERROR, "L2 level dump function is currently not supported."
+            )
+
+    def build_hook(self, target_type, name):
+        def forward_hook(api_or_cell_name, cell, input, output):
+            if target_type == BaseScope.Module_Type_Module:
+                api_or_cell_name = cell.mindstudio_reserved_name
+            self.data_collector.visit_and_clear_overflow_status(api_or_cell_name)
+            if not self.switch:
+                return None
+            if self.data_collector:
+                if target_type == BaseScope.Module_Type_Module:
+                    module_input_output = ModuleForwardInputsOutputs(args=input, kwargs={}, output=output)
+                else:
+                    module_input_output = ModuleForwardInputsOutputs(args=input, kwargs=cell.input_kwargs, output=output)
+                self.data_collector.forward_data_collect(api_or_cell_name, cell, pid, module_input_output)
+                if self.data_collector.if_return_forward_new_output():
+                    return self.data_collector.get_forward_new_output()
+                if target_type == BaseScope.Module_Type_API:
+                    del cell.input_kwargs
+            return output
+
+        def backward_hook(api_or_cell_name, cell, grad_input, grad_output):
+            if target_type == BaseScope.Module_Type_Module:
+                api_or_cell_name = cell.mindstudio_reserved_name
+            self.data_collector.visit_and_clear_overflow_status(api_or_cell_name)
+            if not self.switch:
+                return
+            if self.data_collector:
+                # 框架最新接口变更，grad_input和grad_output的含义发生了变化，与torch含义保持一致，因此此处调换顺序传入
+                module_input_output = ModuleBackwardInputsOutputs(grad_input=grad_output, grad_output=grad_input)
+                self.data_collector.backward_data_collect(api_or_cell_name, cell, pid, module_input_output)
+
+        pid = os.getpid()
+        forward_name_template = name + Const.FORWARD
+        backward_name_template = name + Const.BACKWARD
+        forward_hook = functools.partial(forward_hook, forward_name_template)
+        backward_hook = functools.partial(backward_hook, backward_name_template)
+
+        def wrap_forward_hook(cell, input, output):
+            return forward_hook(cell, input, output)
+
+        def wrap_backward_hook(cell, grad_input, grad_output):
+            return backward_hook(cell, grad_input, grad_output)
+
+        return wrap_forward_hook, wrap_backward_hook
+
+    def wrap_primitive(self, origin_func, primitive_name):
+        service_instance = self
+
+        def create_backward_hook(captured_grads, num_tensors, updated_primitive_name, hook_type):
+            def backward_hook(grad):
+                captured_grads.append(grad)
+                backward_primitive_name = f"{updated_primitive_name}.{Const.BACKWARD}"
+                try:
+                    if len(captured_grads) == num_tensors and hook_type == Const.INPUT:
+                        service_instance.data_collector.visit_and_clear_overflow_status(backward_primitive_name)
+                        new_module_input_output = ModuleBackwardOutputs(grad_output=tuple(captured_grads))
+                        service_instance.data_collector.backward_output_data_collect(
+                            backward_primitive_name, service_instance, os.getpid(), new_module_input_output
+                        )
+                        captured_grads.clear()
+                    elif len(captured_grads) == num_tensors and hook_type == Const.OUTPUT:
+                        service_instance.data_collector.visit_and_clear_overflow_status(backward_primitive_name)
+                        new_module_input_output = ModuleBackwardInputs(grad_input=tuple(captured_grads))
+                        service_instance.data_collector.backward_input_data_collect(
+                            backward_primitive_name, service_instance, os.getpid(), new_module_input_output
+                        )
+                        captured_grads.clear()
+
+                except Exception as exception:
+                    raise Exception(f"This is a primitive op {hook_type}_backward dump error: {exception},"
+                                    f" updated_primitive_name: {updated_primitive_name}") from exception
+
+            return backward_hook
+
+        def hook_primitive_inputs(args, captured_grads_input, updated_primitive_name):
+            hooked_inputs = []
+            num_tensors = sum(isinstance(arg, Tensor) for arg in args)
+            input_backward_hook = create_backward_hook(captured_grads_input, num_tensors, updated_primitive_name,
+                                                       Const.INPUT)
+            for _, arg in enumerate(args):
+                if isinstance(arg, Tensor):
+                    arg_hooked = ops.HookBackward(input_backward_hook)(arg)
+                    hooked_inputs.append(arg_hooked)
+                else:
+                    hooked_inputs.append(arg)
+            return hooked_inputs
+
+        def hook_primitive_outputs(out, captured_grads_output, updated_primitive_name):
+            if isinstance(out, tuple):
+                num_output_tensors = sum(isinstance(tensor, Tensor) for tensor in out)
+            else:
+                num_output_tensors = 1
+            output_backward_hook = create_backward_hook(captured_grads_output, num_output_tensors,
+                                                        updated_primitive_name, Const.OUTPUT)
+
+            if isinstance(out, Tensor):
+                return ops.HookBackward(output_backward_hook)(out)
+            elif isinstance(out, tuple):
+                hooked_outputs = []
+                for tensor in out:
+                    if isinstance(tensor, Tensor):
+                        hooked_outputs.append(ops.HookBackward(output_backward_hook)(tensor))
+                    else:
+                        hooked_outputs.append(tensor)
+                return tuple(hooked_outputs)
+            return out
+
+        def wrapped_primitive_call(instance_self, *args, **kwargs):
+            service_instance.update_primitive_counters(primitive_name)
+            current_count = service_instance.primitive_counters.get(primitive_name, 0)
+            updated_primitive_name = f"{Const.PRIMITIVE_PREFIX}.{primitive_name}.{current_count}"
+
+            if not service_instance.switch:
+                return origin_func(*args, **kwargs)
+
+            captured_grads_input, captured_grads_output = [], []
+
+            try:
+                hooked_inputs = hook_primitive_inputs(args, captured_grads_input, updated_primitive_name)
+            except Exception as exception:
+                raise Exception("This is a primitive op dump error during input hooking: {},"
+                                " primitive_name: {}".format(exception, primitive_name)) from exception
+
+            try:
+                out = origin_func(*hooked_inputs, **kwargs)
+            except Exception as exception:
+                raise Exception("This is a primitive op dump error during function call: {},"
+                                " primitive_name: {}".format(exception, primitive_name)) from exception
+
+            forward_primitive_name = f"{updated_primitive_name}.{Const.FORWARD}"
+            service_instance.data_collector.visit_and_clear_overflow_status(forward_primitive_name)
+            if service_instance.data_collector:
+                module_input_output = ModuleForwardInputsOutputs(args=hooked_inputs, kwargs=kwargs, output=out)
+                try:
+                    service_instance.data_collector.forward_data_collect(forward_primitive_name, instance_self,
+                                                                         os.getpid(), module_input_output)
+                except Exception as exception:
+                    raise Exception("This is a primitive op dump error during forward data collection: {},"
+                                    " primitive_name: {}".format(exception, primitive_name)) from exception
+
+                if service_instance.data_collector.if_return_forward_new_output():
+                    out = service_instance.data_collector.get_forward_new_output()
+
+            try:
+                out = hook_primitive_outputs(out, captured_grads_output, updated_primitive_name)
+            except Exception as exception:
+                raise Exception("This is a primitive op dump error during output hooking: {},"
+                                " primitive_name: {}".format(exception, primitive_name)) from exception
+
+            return out
+
+        return wrapped_primitive_call
+
+    def update_primitive_counters(self, primitive_name):
+        if primitive_name not in self.primitive_counters:
+            self.primitive_counters[primitive_name] = 0
+        else:
+            self.primitive_counters[primitive_name] += 1
+
+    def register_hooks(self):
+        primitive_set = set()
+        for _, cell in self.model.cells_and_names():
+            for pname, primitive in cell._primitives.items():
+                primitive_set.add((pname, primitive))
+
+        for pname, primitive in primitive_set:
+            NewPrimitive = type('NewPrimitive', (primitive.__class__,),
+                                {'__call__': self.wrap_primitive(primitive.__call__, pname)})
+            primitive.__class__ = NewPrimitive
+
+    def step(self):
+        self.current_iter += 1
+        self.data_collector.update_iter(self.current_iter)
+        HOOKCell.cell_count = defaultdict(int)
+        CellProcessor.cell_count = {}
+        self.primitive_counters.clear()
+
+    def start(self, model=None):
+        self.model = self.check_model_valid(model)
+        self.start_call = True
+        logger.info("msprobe: debugger.start() is set successfully")
+        if self.config.step and self.current_iter > max(self.config.step):
+            self.stop()
+            raise Exception("msprobe: exit after iteration {}".format(max(self.config.step)))
+        if self.config.step and self.current_iter not in self.config.step:
+            return
+        if self.first_start:
+            try:
+                self.current_rank = get_rank_if_initialized()
+            except DistributedNotInitializedError:
+                self.current_rank = None
+
+            if self.config.rank and self.current_rank not in self.config.rank:
+                return
+            self.register_hook_new()
+            self.first_start = False
+        self.switch = True
+        logger.info(f"Dump switch is turned on at step {self.current_iter}. ")
+        self.create_dirs()
+        logger.info(f"Dump data will be saved in {self.dump_iter_dir}.")
+        if self.config.level == "L1":
+            JitDump.set_config(self.config)
+            JitDump.set_data_collector(self.data_collector)
+            ms.common.api._MindsporeFunctionExecutor =  JitDump
+            ms.common.api._PyNativeExecutor.grad =  JitDump.grad
+            if pijit_label:
+                PIJitCaptureContext.__enter__ = self.empty
+                PIJitCaptureContext.__exit__ = self.empty
+
+    def stop(self):
+        logger.info("msprobe: debugger.stop() is set successfully. "
+                    "Please set debugger.start() to turn on the dump switch again. ")
+        if not self.start_call:
+            logger.error("msprobe: debugger.start() is not set in the current scope.")
+            raise Exception("debugger.start() is not set in the current scope.")
+        if self.config.step and self.current_iter not in self.config.step:
+            return
+        if self.config.rank and self.current_rank not in self.config.rank:
+            return
+        self.switch = False
+        self.start_call = False
+        self.data_collector.write_json()
+
+    def create_dirs(self):
+        check_path_before_create(self.config.dump_path)
+        if not os.path.exists(self.config.dump_path):
+            Path(self.config.dump_path).mkdir(mode=0o750, exist_ok=True)
+        file_check = FileChecker(self.config.dump_path, FileCheckConst.DIR)
+        file_check.common_check()
+        self.dump_iter_dir = os.path.join(self.config.dump_path, f"step{self.current_iter}")
+        cur_rank = self.current_rank if self.current_rank is not None else ''
+        dump_dir = os.path.join(self.dump_iter_dir, f"rank{cur_rank}")
+        if not os.path.exists(dump_dir):
+            Path(dump_dir).mkdir(mode=0o750, parents=True, exist_ok=True)
+        if self.config.task in self.data_collector.tasks_need_tensor_data:
+            dump_data_dir = os.path.join(dump_dir, "dump_tensor_data")
+            Path(dump_data_dir).mkdir(mode=0o750, exist_ok=True)
+        else:
+            dump_data_dir = None
+
+        dump_file_path = os.path.join(dump_dir, "dump.json")
+        stack_file_path = os.path.join(dump_dir, "stack.json")
+        construct_file_path = os.path.join(dump_dir, "construct.json")
+        self.data_collector.update_dump_paths(
+            dump_file_path, stack_file_path, construct_file_path, dump_data_dir, None)
+        
+    def empty(self, *args, **kwargs):
+        pass
+
+    def register_hook_new(self):
+        logger.info("The {} hook function is successfully mounted to the model.".format(self.config.task))
+        if self.config.level == "L1":
+            api_register.initialize_hook(functools.partial(self.build_hook, BaseScope.Module_Type_API))
+            api_register.api_set_hook_func()
+            if self.model:
+                self.register_hooks()
+
+        if self.config.level == "L0":
+            if not self.model:
+                raise MsprobeException(MsprobeException.INVALID_PARAM_ERROR, "The current level is L0, the model cannot be None")
+            for name, cell in self.model.cells_and_names():
+                if cell == self.model:
+                    continue
+                prefix = 'Cell' + Const.SEP + name + Const.SEP + \
+                         cell.__class__.__name__ + Const.SEP
+                forward_hook, backward_hook = self.build_hook(BaseScope.Module_Type_Module, prefix)
+                cell.register_forward_hook(forward_hook)
+                cell.register_backward_hook(backward_hook)
+
+                cell.register_forward_pre_hook(
+                    self.cell_processor.node_hook(prefix + Const.FORWARD, Const.START))
+                cell.register_forward_hook(
+                    self.cell_processor.node_hook(prefix + Const.FORWARD, Const.STOP))
+                cell.register_backward_pre_hook(
+                    self.cell_processor.node_hook(prefix + Const.BACKWARD, Const.START))
+                cell.register_backward_hook(
+                    self.cell_processor.node_hook(prefix + Const.BACKWARD, Const.STOP))

msprobe/mindspore/task_handler_factory.py

@@ -1,20 +1,23 @@
+from msprobe.core.common.const import Const
 from msprobe.mindspore.debugger.debugger_config import DebuggerConfig
 from msprobe.mindspore.dump.dump_tool_factory import DumpToolFactory
 from msprobe.mindspore.overflow_check.overflow_check_tool_factory import OverflowCheckToolFactory
+from msprobe.mindspore.free_benchmark.self_check_tool_factory import SelfCheckToolFactory
 
 
 class TaskHandlerFactory:
     tasks = {
-        "tensor": DumpToolFactory,
-        "statistics": DumpToolFactory,
-        "overflow_check": OverflowCheckToolFactory
+        Const.TENSOR: DumpToolFactory,
+        Const.STATISTICS: DumpToolFactory,
+        Const.OVERFLOW_CHECK: OverflowCheckToolFactory,
+        Const.FREE_BENCHMARK: SelfCheckToolFactory
     }
 
     @staticmethod
     def create(config: DebuggerConfig):
         task = TaskHandlerFactory.tasks.get(config.task)
         if not task:
-            raise Exception("valid task is needed.")
+            raise Exception("Valid task is needed.")
         handler = task.create(config)
         if not handler:
             raise Exception("Can not find task handler")