saymrwulf/pytorch
1
0
Fork 0
mirror of https://github.com/saymrwulf/pytorch.git synced 2026-05-14 20:57:59 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

[Docs] Add extended debugging options for troubleshooting (#122028)

Browse source 
Fixes #120889

Co-authored-by: Nikita Shulga <2453524+malfet@users.noreply.github.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/122028
Approved by: https://github.com/ezyang, https://github.com/malfet
This commit is contained in:
Sahdev Zala 2024-03-21 17:00:43 +00:00 committed by PyTorch MergeBot
parent c20bc18d59
commit 17175cdbc7
1 changed files with 15 additions and 0 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

15
docs/source/torch.compiler_troubleshooting.rst
View file

@ -707,3 +707,18 @@ to detect bugs in our codegen or with a backend compiler.
If you'd like to ensure that random number generation is the same across both torch
and triton then you can enable ``torch._inductor.config.fallback_random = True``
Extended Debugging
~~~~~~~~~~~~~~~~~~
Extended debugging can be enabled by using the following experimental flags.
``TORCHDYNAMO_EXTENDED_DEBUG_GUARD_ADDED`` - provides extended debug information if the
string representation of a guard matches this flag value. For example, set it to
"Ne(s0, 10)" to generate full Python and C++ backtrace whenever guard was issued.
``TORCHDYNAMO_EXTENDED_DEBUG_CREATE_SYMBOL`` - provides extended debug information when
a particular symbol is allocated. For example, set this to "u2" to generate full Python
and C++ backtrace whenever this symbol was created.
``TORCHDYNAMO_EXTENDED_DEBUG_CPP`` - provides extended debug information (C++ backtrace)
for all extended debug settings as well as errors. For example, set this to "1". The C++
backtrace is slow and very spammy so it is not included by default with extended debugging.