Zgc/ditorch support overflow apply to third party interface

README.md

@@ -10,8 +10,22 @@ ditorch 是设备无关 torch， 旨在屏蔽各硬件厂商 torch 差异，为
 
 只需添加两行代码，即可在国产芯片上像官方 pytorch 一样使用。
 ```
-import torch
-import ditorch
+>>> import torch
+>>> import ditorch
+ditorch.framework: torch_npu:2.1.0.post3  pid: 1729023
+>>> x = torch.randn(3,4,device="cuda")
+>>>
+>>> y = x + x
+>>> x
+Warning: Device do not support double dtype now, dtype cast repalce with float.
+tensor([[ 1.3310,  1.0011, -1.0679, -1.5444],
+        [-0.7345, -0.9888, -1.7310, -0.3305],
+        [-0.6676, -1.7792,  0.7108, -0.9981]], device='cuda:0')
+>>> y
+tensor([[ 2.6619,  2.0023, -2.1359, -3.0887],
+        [-1.4691, -1.9777, -3.4620, -0.6609],
+        [-1.3353, -3.5583,  1.4216, -1.9962]], device='cuda:0')
+>>>
 ```
 
 [ditorch + Ascend910 pytorch原生测例通过情况](ditorch/test/ascend_summary_of_pytorch_test_case_testing.csv.tar)

op_tools/custom_apply_hook.py

@@ -7,6 +7,7 @@
 from .op_observe_hook import OpObserveHook
 from .op_time_measure_hook import OpTimeMeasureHook
 from .op_dtype_cast_hook import OpDtypeCastHook
+from .op_overflow_check_hook import OpOverflowCheckHook
 from .base_hook import BaseHook
 
 
@@ -79,6 +80,7 @@ def apply_feature(ops, feature, condition_func=lambda *args, **kwargs: True):
         "dump_op_args",
         "cast_dtype",
         "op_capture",
+        "overflow_check",
     ]
     assert feature in feature_options, f"feature must be one of {feature_options}, but got {feature}"
     assert callable(condition_func)
@@ -94,6 +96,8 @@ def apply_feature(ops, feature, condition_func=lambda *args, **kwargs: True):
         hook_cls = OpDtypeCastHook
     elif feature == "op_capture":
         hook_cls = OpCaptureHook
+    elif feature == "overflow_check":
+        hook_cls = OpOverflowCheckHook
 
     if isinstance(ops, str):
         apply_hook_to_ops(ops, hook_cls, condition_func)

op_tools/op_autocompare_hook.py

@@ -16,7 +16,8 @@
     is_random_number_gen_op,
     garbage_collect,
     get_dtype_cast_dict_form_str,
-    set_env_if_env_is_empty
+    set_option_if_empty,
+    get_option,
 )
 from .save_op_args import save_op_args, serialize_args_to_dict
 
@@ -37,7 +38,7 @@ def append(self, forward_id, compare_info):
         for result in compare_info["result_list"]:
             self.global_autocompare_result.append({"forward_id": forward_id, **result})
 
-        if len(self.global_autocompare_result) > int(os.getenv("OP_TOOLS_MAX_CACHE_SIZE", "100")):
+        if len(self.global_autocompare_result) > int(get_option("OP_TOOLS_MAX_CACHE_SIZE", "100")):
             self.write_to_file()
 
     def write_to_file(self):
@@ -85,13 +86,13 @@ def cleanup(self):
             self.hook_handle = None
 
 
-set_env_if_env_is_empty("LINEAR_OP_DTYPE_CAST_DICT", "torch.float16->torch.float64,torch.bfloat16->torch.float64,torch.float32->torch.float64")  # noqa: E501
-set_env_if_env_is_empty("EMBEDDING_OP_DTYPE_CAST_DICT", "torch.float16->torch.float64,torch.bfloat16->torch.float64,torch.float32->torch.float64")  # noqa: E501
-set_env_if_env_is_empty("NORMALIZE_OP_DTYPE_CAST_DICT", "torch.float16->torch.float64,torch.bfloat16->torch.float64,torch.float32->torch.float64")  # noqa: E501
-set_env_if_env_is_empty("NORM_OP_DTYPE_CAST_DICT", "torch.float16->torch.float64,torch.bfloat16->torch.float64,torch.float32->torch.float64")  # noqa: E501
-set_env_if_env_is_empty("CROSS_ENTROPY_OP_DTYPE_CAST_DICT", "torch.float16->torch.float64,torch.bfloat16->torch.float64,torch.float32->torch.float64")  # noqa: E501
-set_env_if_env_is_empty("MUL_OP_DTYPE_CAST_DICT", "torch.float16->torch.float64,torch.bfloat16->torch.float64,torch.float32->torch.float64")  # noqa: E501
-set_env_if_env_is_empty("MATMUL_OP_DTYPE_CAST_DICT", "torch.float16->torch.float64,torch.bfloat16->torch.float64,torch.float32->torch.float64")  # noqa: E501
+set_option_if_empty("LINEAR_OP_DTYPE_CAST_DICT", "torch.float16->torch.float64,torch.bfloat16->torch.float64,torch.float32->torch.float64")  # noqa: E501
+set_option_if_empty("EMBEDDING_OP_DTYPE_CAST_DICT", "torch.float16->torch.float64,torch.bfloat16->torch.float64,torch.float32->torch.float64")  # noqa: E501
+set_option_if_empty("NORMALIZE_OP_DTYPE_CAST_DICT", "torch.float16->torch.float64,torch.bfloat16->torch.float64,torch.float32->torch.float64")  # noqa: E501
+set_option_if_empty("NORM_OP_DTYPE_CAST_DICT", "torch.float16->torch.float64,torch.bfloat16->torch.float64,torch.float32->torch.float64")  # noqa: E501
+set_option_if_empty("CROSS_ENTROPY_OP_DTYPE_CAST_DICT", "torch.float16->torch.float64,torch.bfloat16->torch.float64,torch.float32->torch.float64")  # noqa: E501
+set_option_if_empty("MUL_OP_DTYPE_CAST_DICT", "torch.float16->torch.float64,torch.bfloat16->torch.float64,torch.float32->torch.float64")  # noqa: E501
+set_option_if_empty("MATMUL_OP_DTYPE_CAST_DICT", "torch.float16->torch.float64,torch.bfloat16->torch.float64,torch.float32->torch.float64")  # noqa: E501
 
 
 class OpAutoCompareHook(BaseHook):
@@ -117,8 +118,8 @@ def update_dtype_cast_dict(self):
         heigher_priority_env_name = op_name_suffix + "_OP_DTYPE_CAST_DICT"
         lower_priority_env_name = "OP_DTYPE_CAST_DICT"
         default_env_value = "torch.float16->torch.float32,torch.bfloat16->torch.float32"
-        heigher_priority_env_value = os.environ.get(heigher_priority_env_name, None)
-        lower_priority_env_value = os.environ.get(lower_priority_env_name, default_env_value)
+        heigher_priority_env_value = get_option(heigher_priority_env_name, None)
+        lower_priority_env_value = get_option(lower_priority_env_name, default_env_value)
         self.dtype_cast_config_str = heigher_priority_env_value or lower_priority_env_value
 
         self.dtype_cast_dict = get_dtype_cast_dict_form_str(self.dtype_cast_config_str)
@@ -407,10 +408,10 @@ def is_should_apply(self, *args, **kwargs):
         if self.name.startswith("torch.empty"):
             return False
 
-        if is_opname_match(self.name, os.getenv("OP_AUTOCOMPARE_DISABLE_LIST", "")):
+        if is_opname_match(self.name, get_option("OP_AUTOCOMPARE_DISABLE_LIST", "")):
             return False
 
-        return is_opname_match(self.name, os.getenv("OP_AUTOCOMPARE_LIST", ".*"))
+        return is_opname_match(self.name, get_option("OP_AUTOCOMPARE_LIST", ".*"))
 
 
 atexit.register(dump_all_autocompare_info)

op_tools/op_capture_hook.py

@@ -1,9 +1,8 @@
 # Copyright (c) 2024, DeepLink.
-import os
 import torch
 import time
 from .base_hook import BaseHook, DisableHookGuard
-from .utils import traverse_container, is_opname_match, garbage_collect
+from .utils import traverse_container, is_opname_match, garbage_collect, get_option
 from .save_op_args import save_op_args, serialize_args_to_dict
 from .pretty_print import dict_data_list_to_table, packect_data_to_dict_list
 
@@ -71,10 +70,10 @@ def after_call_op(self, result):
             garbage_collect()
 
     def is_should_apply(self, *args, **kwargs):
-        if is_opname_match(self.name, os.getenv("OP_CAPTURE_DISABLE_LIST", "")):
+        if is_opname_match(self.name, get_option("OP_CAPTURE_DISABLE_LIST", "")):
             return False
 
-        if not is_opname_match(self.name, os.getenv("OP_CAPTURE_LIST", ".*")):
+        if not is_opname_match(self.name, get_option("OP_CAPTURE_LIST", ".*")):
             return False
 
         return True

op_tools/op_dtype_cast_hook.py

@@ -11,7 +11,8 @@
     is_opname_match,
     is_view_op,
     is_dtype_cast_op,
-    garbage_collect
+    garbage_collect,
+    get_option,
 )
 from .pretty_print import dict_data_list_to_table
 
@@ -122,7 +123,7 @@ def after_call_op(self, result):
             return result
 
     def is_should_apply(self, *args, **kwargs):
-        if is_opname_match(self.name, os.getenv("OP_DTYPE_CAST_DISABLE_LIST", "")):
+        if is_opname_match(self.name, get_option("OP_DTYPE_CAST_DISABLE_LIST", "")):
             return False
 
         if is_view_op(self.name):
@@ -131,4 +132,4 @@ def is_should_apply(self, *args, **kwargs):
         if is_dtype_cast_op(self.name, *args, **kwargs):
             return False
 
-        return is_opname_match(self.name, os.getenv("OP_DTYPE_CAST_LIST", ".*"))
+        return is_opname_match(self.name, get_option("OP_DTYPE_CAST_LIST", ".*"))

op_tools/op_fallback_hook.py

@@ -1,9 +1,8 @@
 # Copyright (c) 2024, DeepLink.
 import torch
-import os
 
 from .base_hook import BaseHook, DisableHookGuard
-from .utils import to_device, is_cpu_op, is_opname_match, garbage_collect
+from .utils import to_device, is_cpu_op, is_opname_match, garbage_collect, get_option
 from .save_op_args import serialize_args_to_dict
 from .pretty_print import packect_data_to_dict_list, dict_data_list_to_table
 
@@ -114,9 +113,9 @@ def is_should_apply(self, *args, **kwargs):
         if self.name in BLACK_OP_LIST:
             return False
 
-        if is_opname_match(self.name, os.getenv("OP_FALLBACK_DISABLE_LIST", "")):
+        if is_opname_match(self.name, get_option("OP_FALLBACK_DISABLE_LIST", "")):
             return False
         # if name in VIEW_OPS:
         #    return False
 
-        return is_opname_match(self.name, os.getenv("OP_FALLBACK_LIST", ".*"))
+        return is_opname_match(self.name, get_option("OP_FALLBACK_LIST", ".*"))

op_tools/op_observe_hook.py

@@ -1,6 +1,5 @@
 # Copyright (c) 2024, DeepLink.
-import os
-from .utils import is_opname_match
+from .utils import is_opname_match, get_option
 from .base_hook import BaseHook, DisableHookGuard
 
 from .save_op_args import serialize_args_to_dict
@@ -26,6 +25,6 @@ def after_call_op(self, result):
             print("\n" * 2)
 
     def is_should_apply(self, *args, **kwargs):
-        if is_opname_match(self.name, os.getenv("OP_OBSERVE_DISABLE_LIST", "")):
+        if is_opname_match(self.name, get_option("OP_OBSERVE_DISABLE_LIST", "")):
             return False
-        return is_opname_match(self.name, os.getenv("OP_OBSERVE_LIST", ".*"))
+        return is_opname_match(self.name, get_option("OP_OBSERVE_LIST", ".*"))

op_tools/op_overflow_check_hook.py

@@ -1,6 +1,5 @@
 # Copyright (c) 2024, DeepLink.
-import os
-from .utils import is_opname_match, traverse_container, is_inf_or_nan, garbage_collect, compute_tensor_features
+from .utils import is_opname_match, traverse_container, is_inf_or_nan, garbage_collect, compute_tensor_features, get_option
 from .base_hook import BaseHook, DisableHookGuard
 import torch
 
@@ -106,6 +105,6 @@ def after_call_op(self, result):
             garbage_collect()
 
     def is_should_apply(self, *args, **kwargs):
-        if is_opname_match(self.name, os.getenv("OP_OVERFLOW_CHECK_DISABLE_LIST", "")):
+        if is_opname_match(self.name, get_option("OP_OVERFLOW_CHECK_DISABLE_LIST", "")):
             return False
-        return is_opname_match(self.name, os.getenv("OP_OVERFLOW_CHECK_LIST", ".*"))
+        return is_opname_match(self.name, get_option("OP_OVERFLOW_CHECK_LIST", ".*"))

op_tools/op_time_measure_hook.py

@@ -6,7 +6,7 @@
 from .base_hook import BaseHook, DisableHookGuard
 
 from .save_op_args import serialize_args_to_dict
-from .utils import is_opname_match, traverse_container, garbage_collect
+from .utils import is_opname_match, traverse_container, garbage_collect, get_option
 from .pretty_print import (
     dict_data_list_to_table,
     packect_data_to_dict_list,
@@ -39,7 +39,7 @@ def append(self, forward_id, elasped_info):
         else:
             self.global_elasped_info_dict[forward_id].update(elasped_info)
 
-        if len(self.global_elasped_info_dict) > int(os.getenv("OP_TOOLS_MAX_CACHE_SIZE", "5000")):
+        if len(self.global_elasped_info_dict) > int(get_option("OP_TOOLS_MAX_CACHE_SIZE", "5000")):
             self.write_to_file()
 
     def write_to_file(self):
@@ -153,10 +153,10 @@ def after_call_op(self, result):
             garbage_collect()
 
     def is_should_apply(self, *args, **kwargs):
-        if is_opname_match(self.name, os.getenv("OP_TIME_MEASURE_DISABLE_LIST", "")):
+        if is_opname_match(self.name, get_option("OP_TIME_MEASURE_DISABLE_LIST", "")):
             return False
 
-        return is_opname_match(self.name, os.getenv("OP_TIME_MEASURE_LIST", ".*"))
+        return is_opname_match(self.name, get_option("OP_TIME_MEASURE_LIST", ".*"))
 
 
 atexit.register(dump_all_op_elasped_info)

op_tools/process_monitor.py

@@ -5,6 +5,7 @@
 import os
 from prettytable import PrettyTable
 from op_tools.pretty_print import dict_data_list_to_table
+from op_tools.utils import get_option
 
 is_ascend_npu_env = subprocess.run("npu-smi info", shell=True, capture_output=True, text=True).returncode == 0
 is_camb_mlu_env = subprocess.run("cnmon", shell=True, capture_output=True, text=True).returncode == 0
@@ -104,13 +105,13 @@ def __init__(self, pid) -> None:
             device_name = "ascend"
         if is_camb_mlu_env:
             device_name = "camb"
-        self.file_name = f"op_tools_results/process_monitor_result/process_monitor_result_{device_name}_pid{pid}_{os.getenv('label', '')}_{time.strftime('%Y-%m-%d-%H-%M-%S', time.localtime())}.csv"  # noqa: E501
+        self.file_name = f"op_tools_results/process_monitor_result/process_monitor_result_{device_name}_pid{pid}_{get_option('label', '')}_{time.strftime('%Y-%m-%d-%H-%M-%S', time.localtime())}.csv"  # noqa: E501
         self.dir = self.file_name[0 : self.file_name.rfind("/")]
 
     def append(self, info):
         self.global_result.append(info)
 
-        if len(self.global_result) > int(os.getenv("OP_TOOLS_MAX_CACHE_SIZE", "1")):
+        if len(self.global_result) > int(get_option("OP_TOOLS_MAX_CACHE_SIZE", "1")):
             self.write_to_file()
 
     def write_to_file(self):

op_tools/test/test_custom_apply_hook.py

@@ -88,6 +88,15 @@ def test_cast_dtype(self):
         y = torch.rand(4, 5, dtype=torch.float16, device="cuda", requires_grad=True)
         _test_function(x, y)
 
+    def test_overflow_check(self):
+        op_tools.apply_feature(
+            ops=["torch.add", "torch.sub", "torch.mul", "torch.div"],
+            feature="overflow_check",
+        )
+        x = torch.tensor([1, 2, 3], device="cuda", dtype=torch.float16, requires_grad=True)
+        y = torch.tensor([4, 5, 6], device="cuda", dtype=torch.float16, requires_grad=True)
+        _test_function(x, y)
+
     def test_condition_fallback(self):
         def condition_func(a, b, **kwargs):
             if a.dtype == torch.float16:

op_tools/utils.py

@@ -299,12 +299,12 @@ def get_error_tolerance_for_type(dtype_name, atol, rtol):
         op_name_processed = op_name.split(".")[-1].upper() + "_"
         env_name = "AUTOCOMPARE_ERROR_TOLERANCE_" + dtype_name.upper()
         high_priority_env_name = op_name_processed + env_name
-        if os.getenv(high_priority_env_name) is not None:
-            atol, rtol = map(float, os.getenv(high_priority_env_name).split(","))
-        elif os.getenv(env_name) is not None:
-            atol, rtol = map(float, os.getenv(env_name).split(","))
-        elif os.getenv("AUTOCOMPARE_ERROR_TOLERANCE") is not None:
-            atol, rtol = map(float, os.getenv("AUTOCOMPARE_ERROR_TOLERANCE").split(","))
+        if get_option(high_priority_env_name) is not None:
+            atol, rtol = map(float, get_option(high_priority_env_name).split(","))
+        elif get_option(env_name) is not None:
+            atol, rtol = map(float, get_option(env_name).split(","))
+        elif get_option("AUTOCOMPARE_ERROR_TOLERANCE") is not None:
+            atol, rtol = map(float, get_option("AUTOCOMPARE_ERROR_TOLERANCE").split(","))
         return atol, rtol
 
     if dtype == torch.float16:
@@ -317,8 +317,8 @@ def get_error_tolerance_for_type(dtype_name, atol, rtol):
         return get_error_tolerance_for_type("FLOAT64", 1e-8, 1e-8)
     else:
         atol, rtol = 1e-3, 1e-3
-        if os.getenv("AUTOCOMPARE_ERROR_TOLERANCE") is not None:
-            atol, rtol = map(float, os.getenv("AUTOCOMPARE_ERROR_TOLERANCE").split(","))
+        if get_option("AUTOCOMPARE_ERROR_TOLERANCE") is not None:
+            atol, rtol = map(float, get_option("AUTOCOMPARE_ERROR_TOLERANCE").split(","))
         return atol, rtol
 
 
@@ -510,7 +510,7 @@ def current_location(name=None, stack_depth=-1, print_stack=False):
         else:
             break
 
-    if print_stack or int(os.getenv("OP_TOOLS_PRINT_STACK", "0")) > 0:
+    if print_stack or int(get_option("OP_TOOLS_PRINT_STACK", "0")) > 0:
         for i in range(len(stack) - 2):
             file, line, func, text = stack[i]
             print(f"{file}:{line} {func} {text}")
@@ -519,6 +519,39 @@ def current_location(name=None, stack_depth=-1, print_stack=False):
     return f"{file}:{line} {func}: {text}"
 
 
-def set_env_if_env_is_empty(env_name, env_value):
-    if os.environ.get(env_name, None) is None:
-        os.environ[env_name] = env_value
+class OptionManger:
+    def __init__(self):
+        self.options = {}
+
+    def set_option(self, option_name, value):
+        if option_name in self.options:
+            if self.options[option_name] == value:
+                return
+        print(f"option: {option_name}={value}")
+        self.options[option_name] = value
+        os.environ[option_name] = str(value)
+
+    def get_option(self, option_name, default_value=None):
+        if option_name in os.environ:
+            value = os.environ[option_name]
+            self.set_option(option_name, value)
+            return value
+        else:
+            if default_value is not None:
+                self.set_option(option_name, default_value)
+            return default_value
+
+    def is_option_set(self, option_name):
+        return option_name in os.environ
+
+
+global_option_manger = OptionManger()
+
+
+def set_option_if_empty(option_name, value):
+    if not global_option_manger.is_option_set(option_name):
+        global_option_manger.set_option(option_name, value)
+
+
+def get_option(option_name, default_value=None):
+    return global_option_manger.get_option(option_name, default_value)