mindstudio-probe 1.0.1__py3-none-any.whl → 1.0.4__py3-none-any.whl

msprobe/pytorch/free_benchmark/compare/single_benchmark.py

@@ -1,104 +1,104 @@
-import math
-
-import torch
-from msprobe.pytorch.free_benchmark import logger
-from msprobe.pytorch.free_benchmark.common.constant import ThresholdConfig
-from msprobe.pytorch.free_benchmark.common.utils import TorchC
-
-
-class SingleCompare:
-    def __init__(self) -> None:
-        self.relative_err = None
-        self.absolute_err = None
-        self.eb = None
-        self.threshold = None
-
-    @staticmethod
-    def filter_overflow(tensor) -> int:
-        inf_num = TorchC.sum(TorchC.isinf(tensor))
-        nan_num = TorchC.sum(TorchC.isnan(tensor))
-        return inf_num + nan_num
-
-    @staticmethod
-    def replace_inf_or_nan(tensor):
-        finite_mask = TorchC.isfinite(tensor)
-        inf_or_nan_mask = TorchC.logical_not(finite_mask)
-        inf_or_nan_num = TorchC.sum(inf_or_nan_mask).item()
-        if inf_or_nan_num > 0:
-            tensor[inf_or_nan_mask] = 1
-        return tensor
-
-    @staticmethod
-    def compare_float_seq(actual, golden):
-        return math.isclose(actual, golden)
-
-    @staticmethod
-    def compare_other_seq(actual, golden):
-        return actual == golden
-
-    def compare_dict_seq(self, actual, golden):
-        if len(actual) != len(golden):
-            return False
-        for key, value in golden.items():
-            if not self.compare_seq(value, actual.get(key)):
-                return False
-        return True
-
-    def compare_list_seq(self, actual, golden):
-        if len(actual) != len(golden):
-            return False
-        for index_, value in enumerate(golden):
-            if not self.compare_seq(value, actual[index_]):
-                return False
-        return True
-
-    def compare_seq(self, actual, golden):
-        if isinstance(golden, torch.Tensor):
-            return self.compare_tensor_seq(actual, golden)
-        elif isinstance(golden, dict):
-            return self.compare_dict_seq(actual, golden)
-        elif isinstance(golden, (tuple, list)):
-            return self.compare_list_seq(actual, golden)
-        elif isinstance(golden, float):
-            return self.compare_float_seq(actual, golden)
-        else:
-            return self.compare_other_seq(actual, golden)
-
-    def compare_tensor_seq(self, actual, golden):
-        self.threshold = ThresholdConfig.BENCHMARK_THD_DICT.get(
-            actual.dtype, ThresholdConfig.BENCHMARK_THD_DICT.get(torch.float32)
-        )
-        if self.filter_overflow(golden) > 0:
-            logger.warning_on_rank_0("[msprobe] Free Benchmark: inf and nan"
-                                  "in golden tensor is not supported.")
-            return True
-        actual = self.replace_inf_or_nan(actual)
-        actual = actual.to(torch.float64)
-        golden = golden.to(torch.float64).to(actual.device)
-        self._cal_compare_metrics(actual, golden)
-        if self.absolute_err > self.threshold.small_value_atol:
-            return False
-        if self.relative_err > self.threshold.rtol:
-            return False
-        if self.eb > self.threshold.err_balance:
-            return False
-        return True
-
-    def _cal_compare_metrics(self, actual, golden):
-        diff_value = TorchC.subtract(actual, golden)
-        diff_abs = TorchC.abs(diff_value)
-        golden_abs = TorchC.abs(golden)
-        # 使用绝对误差的元素
-        self.absolute_err = TorchC.max(TorchC.where(
-            TorchC.lt(TorchC.abs(actual), self.threshold.small_value), diff_abs, 0
-        ))
-        diff_rel = TorchC.div(diff_abs, golden_abs)
-        # 使用相对误差的元素
-        self.relative_err = TorchC.max(TorchC.where(
-            TorchC.ge(TorchC.abs(actual), self.threshold.small_value), diff_rel, 0
-        ))
-        # 获取误差均衡性
-        divided = TorchC.where(
-            TorchC.ge(TorchC.abs(golden), self.threshold.small_value), golden_abs, 1
-        )
-        self.eb = TorchC.mean(TorchC.div(diff_value, divided))
+import math
+
+import torch
+from msprobe.pytorch.free_benchmark import logger
+from msprobe.pytorch.free_benchmark.common.constant import ThresholdConfig
+from msprobe.pytorch.free_benchmark.common.utils import TorchC
+
+
+class SingleCompare:
+    def __init__(self) -> None:
+        self.relative_err = None
+        self.absolute_err = None
+        self.eb = None
+        self.threshold = None
+
+    @staticmethod
+    def filter_overflow(tensor) -> int:
+        inf_num = TorchC.sum(TorchC.isinf(tensor))
+        nan_num = TorchC.sum(TorchC.isnan(tensor))
+        return inf_num + nan_num
+
+    @staticmethod
+    def replace_inf_or_nan(tensor):
+        finite_mask = TorchC.isfinite(tensor)
+        inf_or_nan_mask = TorchC.logical_not(finite_mask)
+        inf_or_nan_num = TorchC.sum(inf_or_nan_mask).item()
+        if inf_or_nan_num > 0:
+            tensor[inf_or_nan_mask] = 1
+        return tensor
+
+    @staticmethod
+    def compare_float_seq(actual, golden):
+        return math.isclose(actual, golden)
+
+    @staticmethod
+    def compare_other_seq(actual, golden):
+        return actual == golden
+
+    def compare_dict_seq(self, actual, golden):
+        if len(actual) != len(golden):
+            return False
+        for key, value in golden.items():
+            if not self.compare_seq(value, actual.get(key)):
+                return False
+        return True
+
+    def compare_list_seq(self, actual, golden):
+        if len(actual) != len(golden):
+            return False
+        for index_, value in enumerate(golden):
+            if not self.compare_seq(value, actual[index_]):
+                return False
+        return True
+
+    def compare_seq(self, actual, golden):
+        if isinstance(golden, torch.Tensor):
+            return self.compare_tensor_seq(actual, golden)
+        elif isinstance(golden, dict):
+            return self.compare_dict_seq(actual, golden)
+        elif isinstance(golden, (tuple, list)):
+            return self.compare_list_seq(actual, golden)
+        elif isinstance(golden, float):
+            return self.compare_float_seq(actual, golden)
+        else:
+            return self.compare_other_seq(actual, golden)
+
+    def compare_tensor_seq(self, actual, golden):
+        self.threshold = ThresholdConfig.BENCHMARK_THD_DICT.get(
+            actual.dtype, ThresholdConfig.BENCHMARK_THD_DICT.get(torch.float32)
+        )
+        if self.filter_overflow(golden) > 0:
+            logger.warning_on_rank_0("[msprobe] Free Benchmark: inf and nan"
+                                  "in golden tensor is not supported.")
+            return True
+        actual = self.replace_inf_or_nan(actual)
+        actual = actual.to(torch.float64)
+        golden = golden.to(torch.float64).to(actual.device)
+        self._cal_compare_metrics(actual, golden)
+        if self.absolute_err > self.threshold.small_value_atol:
+            return False
+        if self.relative_err > self.threshold.rtol:
+            return False
+        if self.eb > self.threshold.err_balance:
+            return False
+        return True
+
+    def _cal_compare_metrics(self, actual, golden):
+        diff_value = TorchC.subtract(actual, golden)
+        diff_abs = TorchC.abs(diff_value)
+        golden_abs = TorchC.abs(golden)
+        # 使用绝对误差的元素
+        self.absolute_err = TorchC.max(TorchC.where(
+            TorchC.lt(TorchC.abs(actual), self.threshold.small_value), diff_abs, 0
+        ))
+        diff_rel = TorchC.div(diff_abs, golden_abs)
+        # 使用相对误差的元素
+        self.relative_err = TorchC.max(TorchC.where(
+            TorchC.ge(TorchC.abs(actual), self.threshold.small_value), diff_rel, 0
+        ))
+        # 获取误差均衡性
+        divided = TorchC.where(
+            TorchC.ge(TorchC.abs(golden), self.threshold.small_value), golden_abs, 1
+        )
+        self.eb = TorchC.mean(TorchC.div(diff_value, divided))

msprobe/pytorch/free_benchmark/main.py

@@ -1,102 +1,105 @@
-from abc import ABC
-
-import torch
-from msprobe.core.common.const import Const
-from msprobe.pytorch.free_benchmark import logger
-from msprobe.pytorch.free_benchmark.common.constant import CommonField
-from msprobe.pytorch.free_benchmark.common.enums import (
-    DeviceType,
-    FuzzLevel,
-    HandlerType,
-    PerturbationMode,
-)
-from msprobe.pytorch.free_benchmark.common.params import data_pre_deal, make_handler_params
-from msprobe.pytorch.free_benchmark.compare.grad_saver import GradSaver
-from msprobe.pytorch.free_benchmark.perturbed_layers.layer_factory import LayerFactory
-from msprobe.pytorch.free_benchmark.result_handlers.handler_factory import (
-    FuzzHandlerFactory,
-)
-
-
-class FreeBenchmarkCheck(ABC):
-
-    def __init__(self, config) -> None:
-        super().__init__()
-        self.config = config
-        if self.config.pert_mode is None:
-            self.config.pert_mode = PerturbationMode.IMPROVE_PRECISION
-        if self.config.fuzz_level is None:
-            self.config.fuzz_level = FuzzLevel.BASE_LEVEL
-        if self.config.fuzz_device is None:
-            self.config.fuzz_device = DeviceType.NPU
-        self.current_iter = 0
-
-    def update_iter(self, update_iter):
-        self.current_iter = update_iter
-    
-    def if_fix(self):
-        if self.config.handler_type==HandlerType.FIX:
-            return True
-        return False
-
-    def pre_forward(self, name, module, data_processor, args, kwargs):
-        if not self.config.fuzz_stage == Const.BACKWARD:
-            return
-        origin_func = (
-            module._slow_forward if torch._C._get_tracing_state() else module.forward
-        )
-        handler_params = make_handler_params(name, self.config, self.current_iter)
-        grad_saver = GradSaver(origin_func, handler_params)
-        grad_saver.kwargs = kwargs
-        grad_saver.register_compare_func_for_inputs(args, data_processor)
-        grad_saver.cache_backward_input(args)
-        setattr(module, CommonField.GRADSAVER, grad_saver)
-
-    def forward(self, name, module, args, kwargs, output):
-        if not self.config.fuzz_stage == Const.FORWARD:
-            return output, []
-        origin_func = (
-            module._slow_forward if torch._C._get_tracing_state() else module.forward
-        )
-        data_params = data_pre_deal(name, origin_func, args, kwargs)
-        if data_params.valid_input_index == -1:
-            return output, []
-        data_params.original_result = output
-        data_params.fuzz_stage = self.config.fuzz_stage
-
-        layer = LayerFactory.create(
-            name, self.config.fuzz_device, self.config.pert_mode
-        )
-        layer.handle(data_params)
-        handler_params = make_handler_params(name, self.config, self.current_iter)
-        handler = FuzzHandlerFactory.create(handler_params)
-        handler.handle(data_params)
-        return data_params.perturbed_result, handler.get_unequal_rows()
-
-    def backward(self, name, module, grad_output):
-
-        if not self.config.fuzz_stage == Const.BACKWARD:
-            return
-        try:
-            grad_saver = getattr(module, CommonField.GRADSAVER)
-        except AttributeError:
-            logger.warning_on_rank_0(
-                f"[msprobe] Free benchmark:  get grad saver failed. api_name:{name}"
-            )
-            return
-
-        _new_grad_output = grad_output
-        try:
-            need_grad_tensors, _inner_args = grad_saver.get_vjp_input()
-            origin_grad_input = grad_saver.get_grad_input_from_vjp(
-                tuple(need_grad_tensors), _new_grad_output, _inner_args
-            )
-            grad_saver.origin_grad_input = tuple([x.cpu() for x in origin_grad_input])
-            grad_saver.calculate_perturbed_grad_input(
-                _new_grad_output, need_grad_tensors, _inner_args
-            )
-        except Exception as e:
-            logger.warning_on_rank_0(
-                f"[msprobe] Free benchmark: grad vjp calculate failed. api_name:{name} error: {e}"
-            )
-            return
+from abc import ABC
+
+import torch
+from msprobe.core.common.const import Const
+from msprobe.pytorch.free_benchmark import logger
+from msprobe.pytorch.free_benchmark.common.constant import CommonField
+from msprobe.pytorch.free_benchmark.common.enums import (
+    DeviceType,
+    FuzzLevel,
+    HandlerType,
+    PerturbationMode,
+)
+from msprobe.pytorch.free_benchmark.common.params import (
+    data_pre_deal,
+    make_handler_params,
+)
+from msprobe.pytorch.free_benchmark.compare.grad_saver import GradSaver
+from msprobe.pytorch.free_benchmark.perturbed_layers.layer_factory import LayerFactory
+from msprobe.pytorch.free_benchmark.result_handlers.handler_factory import (
+    FuzzHandlerFactory,
+)
+
+
+class FreeBenchmarkCheck(ABC):
+
+    def __init__(self, config) -> None:
+        super().__init__()
+        self.config = config
+        if self.config.pert_mode is None:
+            self.config.pert_mode = PerturbationMode.IMPROVE_PRECISION
+        if self.config.fuzz_level is None:
+            self.config.fuzz_level = FuzzLevel.BASE_LEVEL
+        if self.config.fuzz_device is None:
+            self.config.fuzz_device = DeviceType.NPU
+        self.current_iter = 0
+
+    def update_iter(self, update_iter):
+        self.current_iter = update_iter
+    
+    def if_fix(self):
+        if self.config.handler_type==HandlerType.FIX:
+            return True
+        return False
+
+    def pre_forward(self, name, module, data_processor, args, kwargs):
+        if not self.config.fuzz_stage == Const.BACKWARD:
+            return
+        origin_func = (
+            module._slow_forward if torch._C._get_tracing_state() else module.forward
+        )
+        handler_params = make_handler_params(name, self.config, self.current_iter)
+        grad_saver = GradSaver(origin_func, handler_params)
+        grad_saver.kwargs = kwargs
+        grad_saver.register_compare_func_for_inputs(args, data_processor)
+        grad_saver.cache_backward_input(args)
+        setattr(module, CommonField.GRADSAVER, grad_saver)
+
+    def forward(self, name, module, args, kwargs, output):
+        if not self.config.fuzz_stage == Const.FORWARD:
+            return output, []
+        origin_func = (
+            module._slow_forward if torch._C._get_tracing_state() else module.forward
+        )
+        data_params = data_pre_deal(name, origin_func, args, kwargs)
+        if data_params.valid_input_index == -1:
+            return output, []
+        data_params.original_result = output
+        data_params.fuzz_stage = self.config.fuzz_stage
+
+        layer = LayerFactory.create(
+            name, self.config.fuzz_device, self.config.pert_mode
+        )
+        layer.handle(data_params)
+        handler_params = make_handler_params(name, self.config, self.current_iter)
+        handler = FuzzHandlerFactory.create(handler_params)
+        perturbed_output = handler.handle(data_params) 
+        return perturbed_output, handler.get_unequal_rows()
+    
+    def backward(self, name, module, grad_output):
+
+        if not self.config.fuzz_stage == Const.BACKWARD:
+            return
+        try:
+            grad_saver = getattr(module, CommonField.GRADSAVER)
+        except AttributeError:
+            logger.warning_on_rank_0(
+                f"[msprobe] Free benchmark:  get grad saver failed. api_name:{name}"
+            )
+            return
+
+        _new_grad_output = grad_output
+        try:
+            need_grad_tensors, _inner_args = grad_saver.get_vjp_input()
+            origin_grad_input = grad_saver.get_grad_input_from_vjp(
+                tuple(need_grad_tensors), _new_grad_output, _inner_args
+            )
+            grad_saver.origin_grad_input = tuple([x.cpu() for x in origin_grad_input])
+            grad_saver.calculate_perturbed_grad_input(
+                _new_grad_output, need_grad_tensors, _inner_args
+            )
+        except Exception as e:
+            logger.warning_on_rank_0(
+                f"[msprobe] Free benchmark: grad vjp calculate failed. api_name:{name} error: {e}"
+            )
+            return

msprobe/pytorch/free_benchmark/perturbed_layers/base_layer.py

@@ -1,13 +1,13 @@
-from abc import ABC, abstractmethod
-from typing import Any
-
-from msprobe.pytorch.free_benchmark.common.params import DataParams
-
-
-class BaseLayer(ABC):
-    def __init__(self, api_name: str) -> None:
-        self.api_name = api_name
-
-    @abstractmethod
-    def handle(self, params: DataParams) -> Any:
-        pass
+from abc import ABC, abstractmethod
+from typing import Any
+
+from msprobe.pytorch.free_benchmark.common.params import DataParams
+
+
+class BaseLayer(ABC):
+    def __init__(self, api_name: str) -> None:
+        self.api_name = api_name
+
+    @abstractmethod
+    def handle(self, params: DataParams) -> Any:
+        pass

msprobe/pytorch/free_benchmark/perturbed_layers/layer_factory.py

@@ -1,41 +1,41 @@
-from msprobe.pytorch.free_benchmark import FreeBenchmarkException
-from msprobe.pytorch.free_benchmark.common.enums import DeviceType, PerturbationMode
-from msprobe.pytorch.free_benchmark.perturbed_layers.npu.improve_precision import (
-    ImprovePrecisionLayer,
-)
-from msprobe.pytorch.free_benchmark.perturbed_layers.npu.add_noise import AddNoiseLayer
-from msprobe.pytorch.free_benchmark.perturbed_layers.npu.bit_noise import BitNoiseLayer
-from msprobe.pytorch.free_benchmark.perturbed_layers.npu.no_change import NoChangeLayer
-from msprobe.pytorch.free_benchmark.perturbed_layers.npu.change_value import (
-    ChangeValueLayer,
-)
-from msprobe.pytorch.free_benchmark.perturbed_layers.run_cpu import CpuLayer
-
-
-class LayerFactory:
-    layers = {
-        DeviceType.NPU: {
-            PerturbationMode.ADD_NOISE: AddNoiseLayer,
-            PerturbationMode.CHANGE_VALUE: ChangeValueLayer,
-            PerturbationMode.NO_CHANGE: NoChangeLayer,
-            PerturbationMode.BIT_NOISE: BitNoiseLayer,
-            PerturbationMode.IMPROVE_PRECISION: ImprovePrecisionLayer,
-        },
-        DeviceType.CPU: {PerturbationMode.TO_CPU: CpuLayer},
-    }
-
-    @staticmethod
-    def create(api_name: str, device_type: str, mode: str):
-        layer = LayerFactory.layers.get(device_type)
-        if not layer:
-            raise FreeBenchmarkException(
-                FreeBenchmarkException.UnsupportedType,
-                f"无标杆工具不支持当前设备 {device_type}",
-            )
-        layer = layer.get(mode)
-        if not layer:
-            raise FreeBenchmarkException(
-                FreeBenchmarkException.UnsupportedType,
-                f"无标杆工具无法识别该扰动因子 {mode} on {device_type}",
-            )
-        return layer(api_name)
+from msprobe.pytorch.free_benchmark import FreeBenchmarkException
+from msprobe.pytorch.free_benchmark.common.enums import DeviceType, PerturbationMode
+from msprobe.pytorch.free_benchmark.perturbed_layers.npu.improve_precision import (
+    ImprovePrecisionLayer,
+)
+from msprobe.pytorch.free_benchmark.perturbed_layers.npu.add_noise import AddNoiseLayer
+from msprobe.pytorch.free_benchmark.perturbed_layers.npu.bit_noise import BitNoiseLayer
+from msprobe.pytorch.free_benchmark.perturbed_layers.npu.no_change import NoChangeLayer
+from msprobe.pytorch.free_benchmark.perturbed_layers.npu.change_value import (
+    ChangeValueLayer,
+)
+from msprobe.pytorch.free_benchmark.perturbed_layers.run_cpu import CpuLayer
+
+
+class LayerFactory:
+    layers = {
+        DeviceType.NPU: {
+            PerturbationMode.ADD_NOISE: AddNoiseLayer,
+            PerturbationMode.CHANGE_VALUE: ChangeValueLayer,
+            PerturbationMode.NO_CHANGE: NoChangeLayer,
+            PerturbationMode.BIT_NOISE: BitNoiseLayer,
+            PerturbationMode.IMPROVE_PRECISION: ImprovePrecisionLayer,
+        },
+        DeviceType.CPU: {PerturbationMode.TO_CPU: CpuLayer},
+    }
+
+    @staticmethod
+    def create(api_name: str, device_type: str, mode: str):
+        layer = LayerFactory.layers.get(device_type)
+        if not layer:
+            raise FreeBenchmarkException(
+                FreeBenchmarkException.UnsupportedType,
+                f"无标杆工具不支持当前设备 {device_type}",
+            )
+        layer = layer.get(mode)
+        if not layer:
+            raise FreeBenchmarkException(
+                FreeBenchmarkException.UnsupportedType,
+                f"无标杆工具无法识别该扰动因子 {mode} on {device_type}",
+            )
+        return layer(api_name)