mindstudio-probe 1.1.0__py3-none-any.whl → 1.2.1__py3-none-any.whl

msprobe/core/data_dump/data_processor/pytorch_processor.py

@@ -13,19 +13,24 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+import hashlib
 import zlib
 from dataclasses import asdict
 from typing import List
 
 import numpy as np
 import torch
+from torch import distributed as dist
+
 from msprobe.core.common.const import Const
 from msprobe.core.common.file_utils import path_len_exceeds_limit
 from msprobe.core.common.log import logger
+from msprobe.core.common.utils import convert_tuple
 from msprobe.core.data_dump.data_processor.base import BaseDataProcessor, ModuleBackwardInputsOutputs, \
     ModuleForwardInputsOutputs, TensorStatInfo
 from msprobe.pytorch.common.utils import save_pt, load_pt
 from msprobe.pytorch.free_benchmark import FreeBenchmarkCheck, UnequalRow
+from msprobe.core.common.utils import recursion_depth_decorator
 
 is_gpu = False
 try:
@@ -35,7 +40,13 @@ except ImportError:
 
 
 class PytorchDataProcessor(BaseDataProcessor):
-    pytorch_special_type = (torch.device, torch.dtype, torch.Size, torch.Tensor)
+    pytorch_special_type = (torch.device, torch.dtype, torch.Size, torch.Tensor, torch.memory_format, dist.ProcessGroup)
+    memory_format = {
+        torch.contiguous_format: "contiguous_format",
+        torch.channels_last: "channels_last",
+        torch.channels_last_3d: "channels_last_3d",
+        torch.preserve_format: "preserve_format"
+    }
 
     def __init__(self, config, data_writer):
         super().__init__(config, data_writer)
@@ -43,6 +54,7 @@ class PytorchDataProcessor(BaseDataProcessor):
             "device": self.analyze_device_in_kwargs,
             "dtype": self.analyze_dtype_in_kwargs
         }
+        self._async_dump_cache = {}
 
     @staticmethod
     def get_md5_for_tensor(x):
@@ -71,53 +83,114 @@ class PytorchDataProcessor(BaseDataProcessor):
         return {"type": "torch.dtype", "value": str(element)}
 
     @staticmethod
-    def get_stat_info(data):
+    def get_stat_info_async(data):
         tensor_stat = TensorStatInfo()
-        if data.is_meta:
-            return tensor_stat
-        data_clone = data.detach()
-        if data_clone.numel() == 0:
+        if torch.is_complex(data):
+            logger.warning("Async dump do not support complex data!")
             return tensor_stat
-        elif data_clone.dtype == torch.bool:
-            tensor_stat.max = True in data_clone
-            tensor_stat.min = False not in data_clone
-        elif not data_clone.shape:
-            tensor_stat.max = tensor_stat.min = tensor_stat.mean = tensor_stat.norm = data_clone.item()
-        elif torch.is_complex(data_clone):
-            data_np = data_clone.cpu().numpy()
+        elif data.dtype == torch.bool:
+            tensor_stat.stack_tensor_stat = (["Max", "Min"], torch.stack(
+                [torch.any(data), torch.all(data)]))
+        elif not data.shape:
+            tensor_stat.stack_tensor_stat = (["Max", "Min", "Mean", "Norm"], torch.stack([data, data, data, data]))
+        else:
+            if not data.is_floating_point() or data.dtype == torch.float64:
+                data = data.float()
+            tensor_stat.stack_tensor_stat = (["Max", "Min", "Mean", "Norm"], torch.stack([
+                torch.max(data),
+                torch.min(data),
+                torch.mean(data),
+                torch.norm(data)
+            ]))
+        return tensor_stat
+
+    @staticmethod
+    def get_stat_info_sync(data):
+        tensor_stat = TensorStatInfo()
+        if torch.is_complex(data):
+            data_np = data.cpu().numpy()
             data_abs = np.abs(data_np)
             tensor_stat.max = np.max(data_abs).item()
             tensor_stat.min = np.min(data_abs).item()
             tensor_stat.mean = np.mean(data_abs).item()
+        elif data.dtype == torch.bool:
+            tensor_stat.max = torch.any(data).item()
+            tensor_stat.min = torch.all(data).item()
+        elif not data.shape:
+            tensor_stat.max = tensor_stat.min = tensor_stat.mean = tensor_stat.norm = data.item()
         else:
-            if not data_clone.is_floating_point() or data_clone.dtype == torch.float64:
-                data_clone = data_clone.float()
-            tensor_stat.max = torch._C._VariableFunctionsClass.max(data_clone).item()
-            tensor_stat.min = torch._C._VariableFunctionsClass.min(data_clone).item()
-            tensor_stat.mean = torch._C._VariableFunctionsClass.mean(data_clone).item()
-            tensor_stat.norm = torch._C._VariableFunctionsClass.norm(data_clone).item()
+            if not data.is_floating_point() or data.dtype == torch.float64:
+                data = data.float()
+            tensor_stat.max = torch.max(data).item()
+            tensor_stat.min = torch.min(data).item()
+            tensor_stat.mean = torch.mean(data).item()
+            tensor_stat.norm = torch.norm(data).item()
         return tensor_stat
 
+    @staticmethod
+    def get_stat_info(data, async_dump=False):
+        tensor_stat = TensorStatInfo()
+        if data.is_meta:
+            return tensor_stat
+        data_clone = data.detach()
+        if data_clone.numel() == 0:
+            return tensor_stat
+        else:
+            if data_clone.device.type == Const.CPU_LOWERCASE or not async_dump:
+                return PytorchDataProcessor.get_stat_info_sync(data_clone)
+            else:
+                return PytorchDataProcessor.get_stat_info_async(data_clone)
+
     @staticmethod
     def handle_tensor_extremum_nan_inf(tensor, operator):
         data_clone = tensor.detach()
-        data_nan = torch._C._VariableFunctionsClass.isnan(data_clone)
-        if int(torch._C._VariableFunctionsClass.sum(data_nan)) == data_clone.numel():
+        data_nan = torch.isnan(data_clone)
+        if int(torch.sum(data_nan)) == data_clone.numel():
             return float('nan')
-        finite_mask = torch._C._VariableFunctionsClass.isfinite(data_clone)
-        if int(torch._C._VariableFunctionsClass.sum(finite_mask)) > 0:
+
+        finite_mask = torch.isfinite(data_clone)
+        if int(torch.sum(finite_mask)) > 0:
             finite_values = data_clone[finite_mask]
-            return torch._C._VariableFunctionsClass.max(finite_values).item() if operator == 'max' else \
-                torch._C._VariableFunctionsClass.min(finite_values).item()
+            return torch.max(finite_values).item() if operator == 'max' else \
+                torch.min(finite_values).item()
         else:
             data_no_nan = data_clone[~data_nan]
-            return torch._C._VariableFunctionsClass.max(data_no_nan).item() if operator == 'max' else \
-                torch._C._VariableFunctionsClass.min(data_no_nan).item()
+            return torch.max(data_no_nan).item() if operator == 'max' else \
+                torch.min(data_no_nan).item()
+
+    @staticmethod
+    def process_group_hash(arg):
+        group_ranks = dist.get_process_group_ranks(arg)
+        group_ranks_hash = hashlib.md5(str(group_ranks).encode('utf-8')).hexdigest()
+        return group_ranks_hash
+
+    @staticmethod
+    def is_distributed_op(module):
+        return getattr(module, "op_is_distributed", False)
 
     @staticmethod
     def _analyze_torch_size(arg):
         return {"type": "torch.Size", "value": list(arg)}
 
+    @staticmethod
+    def _analyze_memory_format(arg):
+        # 获取内存格式
+        format_type = PytorchDataProcessor.memory_format.get(arg)
+
+        return {"type": "torch.memory_format", "format": format_type}
+
+    @staticmethod
+    def _analyze_process_group(arg):
+        group_info = {"type": "torch.ProcessGroup"}
+        try:
+            group_ranks = dist.get_process_group_ranks(arg)
+            group_info.update({"group_ranks": group_ranks})
+            group_id = PytorchDataProcessor.process_group_hash(arg)
+            group_info.update({"group_id": group_id})
+        except Exception as e:
+            logger.warning(f"Failed to get process group(id: {group_id}) ranks info with error info: {e}.")
+        return group_info
+
     @classmethod
     def get_special_types(cls):
         return super().get_special_types() + cls.pytorch_special_type
@@ -127,6 +200,10 @@ class PytorchDataProcessor(BaseDataProcessor):
             return self.torch_object_key[suffix_stack[-1]](element)
         if isinstance(element, torch.Size):
             return self._analyze_torch_size(element)
+        if isinstance(element, torch.memory_format):
+            return self._analyze_memory_format(element)
+        if isinstance(element, dist.ProcessGroup):
+            return self._analyze_process_group(element)
         converted_numpy, numpy_type = self._convert_numpy_to_builtin(element)
         if converted_numpy is not element:
             return self._analyze_numpy(converted_numpy, numpy_type)
@@ -136,26 +213,35 @@ class PytorchDataProcessor(BaseDataProcessor):
             return self._analyze_builtin(element)
         return {}
 
+    def analyze_forward_output(self, name, module, module_input_output: ModuleForwardInputsOutputs):
+        if self.is_distributed_op(module):
+            module_input_output.update_output_with_args_and_kwargs()
+        return super().analyze_forward_output(name, module, module_input_output)
+
     def _analyze_tensor(self, tensor, suffix):
-        tensor_stat = self.get_stat_info(tensor)
+        tensor_stat = self.get_stat_info(tensor, self.config.async_dump)
         tensor_json = {}
         tensor_json.update({'type': 'torch.Tensor'})
         tensor_json.update({'dtype': str(tensor.dtype)})
         tensor_json.update({"shape": tensor.shape})
-        tensor_json.update({"Max": tensor_stat.max})
-        tensor_json.update({"Min": tensor_stat.min})
-        tensor_json.update({"Mean": tensor_stat.mean})
-        tensor_json.update({"Norm": tensor_stat.norm})
-        tensor_json.update({"requires_grad": tensor.requires_grad})
-
-        if tensor_stat.max is not None:
-            if np.isinf(tensor_stat.max) or np.isnan(tensor_stat.max):
-                tensor_json['Max_except_inf_nan'] = self.handle_tensor_extremum_nan_inf(tensor, "max")
-        if tensor_stat.min is not None:
-            if np.isinf(tensor_stat.min) or np.isnan(tensor_stat.min):
-                tensor_json['Min_except_inf_nan'] = self.handle_tensor_extremum_nan_inf(tensor, "min")
-
-        if self.config.summary_mode == Const.MD5:
+        if tensor_stat.stack_tensor_stat is None:
+            tensor_json.update({"Max": tensor_stat.max})
+            tensor_json.update({"Min": tensor_stat.min})
+            tensor_json.update({"Mean": tensor_stat.mean})
+            tensor_json.update({"Norm": tensor_stat.norm})
+            tensor_json.update({"requires_grad": tensor.requires_grad})
+            if tensor_stat.max is not None:
+                if np.isinf(tensor_stat.max) or np.isnan(tensor_stat.max):
+                    tensor_json['Max_except_inf_nan'] = self.handle_tensor_extremum_nan_inf(tensor, "max")
+            if tensor_stat.min is not None:
+                if np.isinf(tensor_stat.min) or np.isnan(tensor_stat.min):
+                    tensor_json['Min_except_inf_nan'] = self.handle_tensor_extremum_nan_inf(tensor, "min")
+
+        else:
+            tensor_json.update({"requires_grad": tensor.requires_grad})
+            tensor_json.update({"tensor_stat": tensor_stat.stack_tensor_stat})
+
+        if self.config.summary_mode == Const.MD5 and not self.config.async_dump:
             tensor_md5 = self.get_md5_for_tensor(tensor)
             tensor_json.update({Const.MD5: tensor_md5})
         return tensor_json
@@ -166,12 +252,20 @@ class StatisticsDataProcessor(PytorchDataProcessor):
 
 
 class TensorDataProcessor(PytorchDataProcessor):
+    def dump_async_data(self):
+        for file_path, tensor in self._async_dump_cache.items():
+            save_pt(tensor.contiguous(), file_path)
+        self._async_dump_cache.clear()
+
     def _analyze_tensor(self, tensor, suffix):
         dump_data_name, file_path = self.get_save_file_path(suffix)
-        saved_tensor = tensor.clone().contiguous().detach()
-        save_pt(saved_tensor, file_path)
         single_arg = super()._analyze_tensor(tensor, suffix)
         single_arg.update({"data_name": dump_data_name})
+        if self.config.async_dump:
+            self._async_dump_cache[file_path] = tensor.clone().detach()
+        else:
+            saved_tensor = tensor.clone().contiguous().detach()
+            save_pt(saved_tensor, file_path)
         return single_arg
 
 
@@ -182,7 +276,7 @@ class OverflowCheckDataProcessor(PytorchDataProcessor):
         super().__init__(config, data_writer)
         self.has_overflow = False
         self.support_inf_nan = None
-        self.cached_inplace_api_info = {}
+        self.cached_api_info = {}
         self.cached_tensors_and_file_paths = {}
         self.bits_for_overflow = 8
         self.real_overflow_nums = 0
@@ -196,21 +290,21 @@ class OverflowCheckDataProcessor(PytorchDataProcessor):
             return True
         return False
 
-    def analyze_pre_forward_inplace(self, name, module_input_output: ModuleForwardInputsOutputs):
+    def analyze_forward_input(self, name, module, module_input_output: ModuleForwardInputsOutputs):
         self.has_overflow = False
         self._is_support_inf_nan()
-        self.cached_inplace_api_info = super().analyze_pre_forward_inplace(name, module_input_output)
+        self.cached_api_info = super().analyze_forward_input(name, module, module_input_output)
         return None
 
-    def analyze_forward_inplace(self, name, module_input_output: ModuleForwardInputsOutputs):
+    def analyze_forward_output(self, name, module, module_input_output: ModuleForwardInputsOutputs):
         self._is_support_inf_nan()
-        api_info_struct = super().analyze_forward_inplace(name, module_input_output)
-        if name in self.cached_inplace_api_info and name in api_info_struct:
-            self.cached_inplace_api_info[name].update(api_info_struct[name])
+        api_info_struct = super().analyze_forward_output(name, module, module_input_output)
+        if name in self.cached_api_info and name in api_info_struct:
+            self.cached_api_info[name].update(api_info_struct[name])
         elif name in api_info_struct:
-            self.cached_inplace_api_info = api_info_struct
+            self.cached_api_info = api_info_struct
         self.handle_overflow()
-        return self.cached_inplace_api_info if self.has_overflow else None
+        return self.cached_api_info if self.has_overflow else None
 
     def analyze_forward(self, name, module, module_input_output: ModuleForwardInputsOutputs):
         self.has_overflow = False
@@ -225,6 +319,13 @@ class OverflowCheckDataProcessor(PytorchDataProcessor):
         api_info_struct = super().analyze_backward(name, module, module_input_output)
         self.handle_overflow()
         return api_info_struct if self.has_overflow else None
+    
+    def analyze_params(self, name, param_name, grad):
+        self.has_overflow = False
+        self._is_support_inf_nan()
+        api_info_struct = super().analyze_params(name, param_name, grad)
+        self.handle_overflow()
+        return api_info_struct if self.has_overflow else None
 
     def handle_overflow(self):
         if not self.support_inf_nan:
@@ -299,10 +400,10 @@ class FreeBenchmarkDataProcessor(PytorchDataProcessor):
             )
         return
 
-    def analyze_pre_forward(self, name, module, module_input_output: ModuleForwardInputsOutputs):
+    def analyze_forward_input(self, name, module, module_input_output: ModuleForwardInputsOutputs):
         self.checker.pre_forward(name, module, self, module_input_output.args, module_input_output.kwargs)
 
-    def analyze_forward(self, name, module, module_input_output: ModuleForwardInputsOutputs):
+    def analyze_forward_output(self, name, module, module_input_output: ModuleForwardInputsOutputs):
         new_output, unequal_rows = self.checker.forward(
             name,
             module,
@@ -320,64 +421,120 @@ class FreeBenchmarkDataProcessor(PytorchDataProcessor):
 
 
 class KernelDumpDataProcessor(PytorchDataProcessor):
-    forward_init_status = False
-    multi_output_apis = ["_sort_", "npu_flash_attention"]
-
     def __init__(self, config, data_writer):
         super().__init__(config, data_writer)
+        self.enable_kernel_dump = True
+        self.is_found_output_tensor = False
+        self.is_found_grad_input_tensor = False
+        self.forward_args = None
+        self.forward_kwargs = None
+        self.forward_output_tensor = None
+        self.grad_input_tensor = None
+
+    @staticmethod
+    def start_kernel_dump(config_path):
+        torch_npu.npu.synchronize()
+        torch_npu.npu.init_dump()
+        torch_npu.npu.set_dump(config_path)
+        torch_npu.npu.synchronize()
+
+    @staticmethod
+    def stop_kernel_dump():
+        torch_npu.npu.synchronize()
+        torch_npu.npu.finalize_dump()
+        torch_npu.npu.synchronize()
+
+    @staticmethod
+    def _print_unsupported_log(api_name):
+        logger.warning(f"The kernel dump does not support the {api_name} API.")
+
+    def analyze_forward_input(self, name, module, module_input_output):
+        if not self.enable_kernel_dump:
+            return
+        if is_gpu:
+            logger.warning("The current environment is not a complete NPU environment, and kernel dump cannot be used.")
+            self.enable_kernel_dump = False
+            return
+
+        if self.config.is_backward_kernel_dump:
+            self.forward_args = self.clone_and_detach_tensor(module_input_output.args)
+            self.forward_kwargs = self.clone_and_detach_tensor(module_input_output.kwargs)
+            try:
+                output = module.forward(*self.forward_args, **self.forward_kwargs)
+            except Exception:
+                self._print_unsupported_log(name)
+                self.enable_kernel_dump = False
+                return
+
+            self.analyze_element(convert_tuple(output))
+            if not self.is_found_output_tensor:
+                self._print_unsupported_log(name)
+                self.enable_kernel_dump = False
+            return
+        self.start_kernel_dump(self.config.kernel_config_path)
+
+    def analyze_forward_output(self, name, module, module_input_output):
+        if not self.enable_kernel_dump:
+            return
+        if self.config.is_backward_kernel_dump:
+            return
+        self.enable_kernel_dump = False
+        self.stop_kernel_dump()
+        logger.info(f"The kernel data of {name} is dumped successfully.")
+
+    def analyze_backward(self, name, module, module_input_output):
+        if not self.enable_kernel_dump:
+            return
+        self.enable_kernel_dump = False
+
+        self.analyze_element(module_input_output.grad_input)
+        if not self.is_found_grad_input_tensor:
+            self._print_unsupported_log(name)
+            return
+        self.start_kernel_dump(self.config.kernel_config_path)
+
+        try:
+            self.forward_output_tensor.backward(self.grad_input_tensor, retain_graph=True)
+        except Exception:
+            self._print_unsupported_log(name)
+            self.stop_kernel_dump()
+            return
 
-    def analyze_forward(self, name, module, module_input_output):
-        if self.config.is_forward_acl_dump:
-            self.forward_acl_dump(name, module, module_input_output)
+        self.stop_kernel_dump()
+        logger.info(f"The kernel data of {name} is dumped successfully.")
+
+    @recursion_depth_decorator("KernelDump: KernelDumpDataProcessor.clone_and_detach_tensor")
+    def clone_and_detach_tensor(self, input_params):
+        if isinstance(input_params, torch.Tensor):
+            if input_params.requires_grad:
+                return input_params.clone().detach().requires_grad_()
+            return input_params.clone()
+        elif isinstance(input_params, tuple):
+            return tuple(self.clone_and_detach_tensor(x) for x in input_params)
+        elif isinstance(input_params, list):
+            return list(self.clone_and_detach_tensor(x) for x in input_params)
+        elif isinstance(input_params, dict):
+            return {k: self.clone_and_detach_tensor(v) for k, v in input_params.items()}
         else:
-            self.dump_mode_backward_acl_dump(name, module, module_input_output)
-
-    def forward_acl_dump(self, name, module, module_input_output):
-        if not KernelDumpDataProcessor.forward_init_status:
-            KernelDumpDataProcessor.forward_init_status = True
-            torch_npu.npu.synchronize()
-            torch_npu.npu.init_dump()
-            torch_npu.npu.set_dump(self.config.acl_config)
-            torch_npu.npu.synchronize()
-            if self.op_need_trigger(name):
-                module.forward(*module_input_output.args, **module_input_output.kwargs).cpu()
-            else:
-                module.forward(*module_input_output.args, **module_input_output.kwargs)
-            torch_npu.npu.synchronize()
-            torch_npu.npu.finalize_dump()
-            torch_npu.npu.synchronize()
-        KernelDumpDataProcessor.forward_init_status = False
-        logger.info("Dump %s op file." % name)
-
-    def acl_backward_dump_status(self, output, grad, module_name):
-        if isinstance(output, torch.Tensor):
-            output.backward(grad, retain_graph=True)
-            return True
+            return input_params
 
-        for api_name in KernelDumpDataProcessor.multi_output_apis:
-            if api_name in module_name:
-                output[0].backward(grad, retain_graph=True)
-                return True
-        return False
+    def analyze_single_element(self, element, suffix_stack):
+        if isinstance(element, torch.Tensor):
+            if not self.is_found_output_tensor:
+                if element.requires_grad:
+                    self.forward_output_tensor = element
+                    self.is_found_output_tensor = True
+                return {}
+            if not self.is_found_grad_input_tensor:
+                self.grad_input_tensor = element.clone()
+                self.is_found_grad_input_tensor = True
+        return {}
 
-    def dump_mode_backward_acl_dump(self, name, module, module_input_output):
-        grad_path = self.config.backward_input.get(name)
-        if not KernelDumpDataProcessor.forward_init_status:
-            KernelDumpDataProcessor.forward_init_status = True
-            output = module.forward(*module_input_output.args, **module_input_output.kwargs)
-            pt = load_pt(grad_path)
-            grad = pt.to("npu").requires_grad_()
-            torch_npu.npu.init_dump()
-            torch_npu.npu.set_dump(self.config.acl_config)
-            torch_npu.npu.synchronize()
-            if not self.acl_backward_dump_status(output, grad, name):
-                logger.warning("The output of {} is not of tensor type and cannot be automatically derived. "
-                               "you can manually construct a single API backward case for ACL dump.".format(
-                    name))
-            torch_npu.npu.synchronize()
-            torch_npu.npu.finalize_dump()
-        KernelDumpDataProcessor.forward_init_status = False
-        logger.info("Dump %s op file." % name)
-
-    def op_need_trigger(self, module_name):
-        return 'Tensor.__getitem__.' in module_name
+    def reset_status(self):
+        self.enable_kernel_dump = True
+        self.is_found_output_tensor = False
+        self.is_found_grad_input_tensor = False
+        self.forward_args = None
+        self.forward_kwargs = None
+        self.forward_output_tensor = None
+        self.grad_input_tensor = None

msprobe/core/data_dump/json_writer.py

@@ -15,10 +15,12 @@
 
 import csv
 import os
+import numpy as np
 
 from msprobe.core.common.const import Const, FileCheckConst
-from msprobe.core.common.file_utils import change_mode, FileOpen, save_json
+from msprobe.core.common.file_utils import change_mode, FileOpen, save_json, load_json
 from msprobe.core.common.log import logger
+from msprobe.core.common.exceptions import MsprobeException
 
 
 class DataWriter:
@@ -115,3 +117,29 @@ class DataWriter:
             self.write_stack_info_json(self.stack_file_path)
         if self.cache_construct:
             self.write_construct_info_json(self.construct_file_path)
+
+    def fill_stack_tensor_data(self):
+        self.process_stat_data_recursive(self.cache_data)
+
+    def process_stat_data_recursive(self, data, depth=0):
+        if depth > Const.MAX_DEPTH:
+            logger.error(f"The maximum depth of recursive process stat data, {Const.MAX_DEPTH} is reached.")
+            raise MsprobeException(MsprobeException.RECURSION_LIMIT_ERROR)
+        if isinstance(data, dict):
+            if "tensor_stat" in data.keys():
+                tensor_stat = data["tensor_stat"]
+                if len(tensor_stat) != Const.TENSOR_STAT_LEN or len(tensor_stat[0]) != len(tensor_stat[1]):
+                    logger.warning("Some bad data in async dump")
+                else:
+                    tensor_stat_index, tensor_stat_data = tensor_stat[0], tensor_stat[1]
+                    if hasattr(tensor_stat_data, "device") and tensor_stat_data.device != Const.CPU_LOWERCASE:
+                        tensor_stat_data = tensor_stat_data.cpu()
+                    for index, stat in zip(tensor_stat_index, tensor_stat_data):
+                        data.update({index, stat.item()})
+                del data["tensor_stat"]
+            else:
+                for key in data.keys():
+                    self.process_stat_data_recursive(data[key], depth + 1)
+        elif isinstance(data, (list, tuple)):
+            for i in data:
+                self.process_stat_data_recursive(i, depth + 1)