Deprecate torch_dtype

[Kaihui-intel]

torch_dtype -> dtype
#776

[wenhuach21]

This change cannot be applied directly, since users running Transformers versions below 4.57 would no longer be supported.

[wenhuach21]

as said in comment

[Kaihui-intel]

as said in comment

Do we need to upgrade transformers in requirement?

auto-round/requirements.txt

Line 12 in 3c1a678

transformers>=4.38

Will it bring other problems?

[wenhuach21]

as said in comment

Do we need to upgrade transformers in requirement?

auto-round/requirements.txt

Line 12 in 3c1a678

transformers>=4.38

Will it bring other problems?

not a good option for now

[wenhuach21]

1 one option is waiting for another 2 or 3 months to upgrade the trasnsformers
2 another is trying the way gpt suggested

import functools
import warnings

def support_legacy_keys(legacy_map: dict):
    """
    legacy_map: dict, key = 新参数名, value = 旧参数名
    """
    def decorator(func):
        @functools.wraps(func)
        def wrapper(*args, **kwargs):
            for new_key, old_key in legacy_map.items():
                if old_key in kwargs:
                    if new_key in kwargs:
                        raise ValueError(f"Cannot use both '{new_key}' and '{old_key}'")
                    warnings.warn(f"'{old_key}' is deprecated, use '{new_key}' instead", DeprecationWarning)
                    kwargs[new_key] = kwargs.pop(old_key)
            return func(*args, **kwargs)
        return wrapper
    return decorator

使用示例：

@support_legacy_keys({"dtype": "torch_dtype"})
def quan(*, dtype=None):
    print(f"dtype = {dtype}")

# 调用方式
quan(dtype="torch.float16")       # 新参数
quan(torch_dtype="torch.float16") # 老参数，会自动映射并提示

[Kaihui-intel]

1 one option is waiting for another 2 or 3 months to upgrade the trasnsformers 2 another is trying the way gpt suggested

import functools
import warnings

def support_legacy_keys(legacy_map: dict):
    """
    legacy_map: dict, key = 新参数名, value = 旧参数名
    """
    def decorator(func):
        @functools.wraps(func)
        def wrapper(*args, **kwargs):
            for new_key, old_key in legacy_map.items():
                if old_key in kwargs:
                    if new_key in kwargs:
                        raise ValueError(f"Cannot use both '{new_key}' and '{old_key}'")
                    warnings.warn(f"'{old_key}' is deprecated, use '{new_key}' instead", DeprecationWarning)
                    kwargs[new_key] = kwargs.pop(old_key)
            return func(*args, **kwargs)
        return wrapper
    return decorator

使用示例：

@support_legacy_keys({"dtype": "torch_dtype"})
def quan(*, dtype=None):
    print(f"dtype = {dtype}")

# 调用方式
quan(dtype="torch.float16")       # 新参数
quan(torch_dtype="torch.float16") # 老参数，会自动映射并提示

For 2, we also need to add a judgment on the transformers version. I prefer option 1.
Do we still have examples of dependencies v4.38-v4.57?

[wenhuach21]

1 one option is waiting for another 2 or 3 months to upgrade the trasnsformers 2 another is trying the way gpt suggested

import functools
import warnings

def support_legacy_keys(legacy_map: dict):
    """
    legacy_map: dict, key = 新参数名, value = 旧参数名
    """
    def decorator(func):
        @functools.wraps(func)
        def wrapper(*args, **kwargs):
            for new_key, old_key in legacy_map.items():
                if old_key in kwargs:
                    if new_key in kwargs:
                        raise ValueError(f"Cannot use both '{new_key}' and '{old_key}'")
                    warnings.warn(f"'{old_key}' is deprecated, use '{new_key}' instead", DeprecationWarning)
                    kwargs[new_key] = kwargs.pop(old_key)
            return func(*args, **kwargs)
        return wrapper
    return decorator

使用示例：

@support_legacy_keys({"dtype": "torch_dtype"})
def quan(*, dtype=None):
    print(f"dtype = {dtype}")

# 调用方式
quan(dtype="torch.float16")       # 新参数
quan(torch_dtype="torch.float16") # 老参数，会自动映射并提示

For 2, we also need to add a judgment on the transformers version. I prefer option 1. Do we still have examples of dependencies v4.38-v4.57?

This should not depend on our examples or perspective; we need to make decisions from the users’ point of view