[Core] Rework dtype resolution (#18751)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
This commit is contained in:
Cyrus Leung
2025-06-01 11:04:23 +08:00
committed by GitHub
parent 1bc86a3da1
commit 6aa8f9a4e7
13 changed files with 314 additions and 119 deletions
+49 -2
View File
@@ -37,8 +37,8 @@ from argparse import (Action, ArgumentDefaultsHelpFormatter, ArgumentParser,
_ArgumentGroup)
from asyncio import FIRST_COMPLETED, AbstractEventLoop, Task
from collections import UserDict, defaultdict
from collections.abc import (AsyncGenerator, Awaitable, Generator, Hashable,
Iterable, Iterator, KeysView, Mapping)
from collections.abc import (AsyncGenerator, Awaitable, Collection, Generator,
Hashable, Iterable, Iterator, KeysView, Mapping)
from concurrent.futures.process import ProcessPoolExecutor
from dataclasses import dataclass, field
from functools import cache, lru_cache, partial, wraps
@@ -979,6 +979,53 @@ def get_dtype_size(dtype: torch.dtype) -> int:
return torch.tensor([], dtype=dtype).element_size()
# bool = 0, int = 1, float = 2, complex = 3
def _get_precision_level(dtype: torch.dtype) -> int:
# NOTE: Complex dtypes return `is_floating_point=False`
return ((dtype != torch.bool) + dtype.is_floating_point +
dtype.is_complex * 2)
def is_lossless_cast(src_dtype: torch.dtype, tgt_dtype: torch.dtype):
"""
Test whether it is lossless to cast a tensor from
`src_dtype` to `tgt_dtype`.
"""
if src_dtype == tgt_dtype:
return True
src_level = _get_precision_level(src_dtype)
tgt_level = _get_precision_level(tgt_dtype)
if src_level < tgt_level:
return True
if src_level > tgt_level:
return False
# Compare integral types
if not src_dtype.is_floating_point and not src_dtype.is_complex:
src_info = torch.iinfo(src_dtype)
tgt_info = torch.iinfo(tgt_dtype)
return src_info.min >= tgt_info.min and src_info.max <= tgt_info.max
# Compare floating-point types
src_info = torch.finfo(src_dtype)
tgt_info = torch.finfo(tgt_dtype)
return (src_info.min >= tgt_info.min and src_info.max <= tgt_info.max
and src_info.resolution >= tgt_info.resolution)
def common_broadcastable_dtype(dtypes: Collection[torch.dtype]):
"""
Get the common `dtype` where all of the other `dtypes` can be
cast to it without losing any information.
"""
return max(
dtypes,
key=lambda dtype: sum(is_lossless_cast(dt, dtype) for dt in dtypes),
)
# `collections` helpers
def is_list_of(
value: object,