pytorch

c039d2f5 - [release v2.0.1] Revisit `torch._six.string_classes` removal (#97737, #97789, #97863) (#98055)

[release v2.0.1] Revisit `torch._six.string_classes` removal (#97737, #97789, #97863) (#98055) * [DataLoader] Short circuit pin_memory recursion when operating on bytes (#97737) Slack thread: https://pytorch.slack.com/archives/GEEQ2K4MD/p1679962409906099 I was seeing some massive (~2x) slowdowns on a job after running it on PyTorch 2.0. From some profiling in `py-spy` it looked like the pin_memory thread was doing a lot more work than before. Looking at a trace in `nsys` I saw the thread doing the forward pass having a bunch of `pthread_cond_timedwait` with GIL reacquire calls in it’s call stack, and it seemed like the thread doing the forward pass was getting blocked (waiting for the GIL) by the pin memory thread (which was holding the GIL). After some debugging I found out the issue. If a `bytes` was passed into `pin_memory`, previously in 1.13 (before https://github.com/pytorch/pytorch/pull/94709) it would short-circuit and return here https://github.com/pytorch/pytorch/blob/d922c29a22e4bf0fba49526f7536395eb8cd66f4/torch/utils/data/_utils/pin_memory.py#L54-L55 since `bytes` was in `torch._six.string_classes`: ``` >>> from torch._six import string_classes >>> string_classes (<class 'str'>, <class 'bytes'>) >>> ``` However after https://github.com/pytorch/pytorch/pull/94709, if a `bytes` was passed into `pin_memory` it would fall into here instead https://github.com/pytorch/pytorch/blob/c263bd43e8e8502d4726643bc6fd046f0130ac0e/torch/utils/data/_utils/pin_memory.py#L68-L73 because the previous check is now doing `isinstance(data, str)` instead of `isinstance(data, (str, bytes))`! https://github.com/pytorch/pytorch/blob/c263bd43e8e8502d4726643bc6fd046f0130ac0e/torch/utils/data/_utils/pin_memory.py#L56-L57 As a result, `pin_memory` gets called recursively for each element in the `bytes` leading to a ton of wasted recursion. This also explains the slowdown / GIL contention I was seeing. This PR simply changes `isinstance(data, str)` to `isinstance(data, (str, bytes))` to match the behavior before https://github.com/pytorch/pytorch/pull/94709 Pull Request resolved: https://github.com/pytorch/pytorch/pull/97737 Approved by: https://github.com/albanD, https://github.com/NivekT * [DataLoader] Fix collation logic (#97789) Similar to #97737, a previous auto-refactor changed how `bytes` are handled during collation, which can potentially lead to performance regression. This PR undoes that. Pull Request resolved: https://github.com/pytorch/pytorch/pull/97789 Approved by: https://github.com/albanD * Revisit `torch._six.string_classes` removal (#94709) (#97863) Revisit `torch._six.string_classes` (which is `(str, bytes)`) removal: `isinstance(obj, string_classes) -> isinstance(obj, str)`. Both `str` and `bytes` are `Sequence` classes. ```python In [1]: from typing import Sequence In [2]: issubclass(bytes, Sequence) Out[2]: True In [3]: issubclass(str, Sequence) Out[3]: True ``` Re-add `bytes` to type guards like: ```python def is_seq(obj): return isinstance(obj, Sequence) and not isinstance(obj, (str, bytes)) ``` Ref: - https://github.com/pytorch/pytorch/pull/94709#issuecomment-1487282912 - #97737 - #97789 Pull Request resolved: https://github.com/pytorch/pytorch/pull/97863 Approved by: https://github.com/Skylion007, https://github.com/albanD --------- Co-authored-by: Eric Zhang <ezhang887@gmail.com> Co-authored-by: Kevin Tse <ktse@fb.com>