MoE Kernel (#2465)
* add moe grouped gemm kernel * add benchmark, README * remove formatting from __init__.py
This commit is contained in:
jeromeku
GitHub
parent
9edbe23259
commit
b7fc12c8be
24 changed files with 6584 additions and 1 deletions
175
.gitignore
vendored
Normal file
175
.gitignore
vendored
Normal file
|
|
@ -0,0 +1,175 @@
|
|||
# Byte-compiled / optimized / DLL files
|
||||
__pycache__/
|
||||
*.py[cod]
|
||||
*.class
|
||||
|
||||
# C extensions
|
||||
*.so
|
||||
|
||||
# Distribution / packaging
|
||||
.Python
|
||||
build/
|
||||
develop-eggs/
|
||||
dist/
|
||||
downloads/
|
||||
eggs/
|
||||
.eggs/
|
||||
lib/
|
||||
lib64/
|
||||
parts/
|
||||
sdist/
|
||||
var/
|
||||
wheels/
|
||||
share/python-wheels/
|
||||
*.egg-info/
|
||||
.installed.cfg
|
||||
*.egg
|
||||
MANIFEST
|
||||
|
||||
# PyInstaller
|
||||
# Usually these files are written by a python script from a template
|
||||
# before PyInstaller builds the exe, so as to inject date/other infos into it.
|
||||
*.manifest
|
||||
*.spec
|
||||
|
||||
# Installer logs
|
||||
pip-log.txt
|
||||
pip-delete-this-directory.txt
|
||||
|
||||
# Unit test / coverage reports
|
||||
htmlcov/
|
||||
.tox/
|
||||
.nox/
|
||||
.coverage
|
||||
.coverage.*
|
||||
.cache
|
||||
nosetests.xml
|
||||
coverage.xml
|
||||
*.cover
|
||||
*.py,cover
|
||||
.hypothesis/
|
||||
.pytest_cache/
|
||||
cover/
|
||||
|
||||
# Translations
|
||||
*.mo
|
||||
*.pot
|
||||
|
||||
# Django stuff:
|
||||
*.log
|
||||
local_settings.py
|
||||
db.sqlite3
|
||||
db.sqlite3-journal
|
||||
|
||||
# Flask stuff:
|
||||
instance/
|
||||
.webassets-cache
|
||||
|
||||
# Scrapy stuff:
|
||||
.scrapy
|
||||
|
||||
# Sphinx documentation
|
||||
docs/_build/
|
||||
|
||||
# PyBuilder
|
||||
.pybuilder/
|
||||
target/
|
||||
|
||||
# Jupyter Notebook
|
||||
.ipynb_checkpoints
|
||||
|
||||
# IPython
|
||||
profile_default/
|
||||
ipython_config.py
|
||||
|
||||
# pyenv
|
||||
# For a library or package, you might want to ignore these files since the code is
|
||||
# intended to run in multiple environments; otherwise, check them in:
|
||||
# .python-version
|
||||
|
||||
# pipenv
|
||||
# According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
|
||||
# However, in case of collaboration, if having platform-specific dependencies or dependencies
|
||||
# having no cross-platform support, pipenv may install dependencies that don't work, or not
|
||||
# install all needed dependencies.
|
||||
#Pipfile.lock
|
||||
|
||||
# UV
|
||||
# Similar to Pipfile.lock, it is generally recommended to include uv.lock in version control.
|
||||
# This is especially recommended for binary packages to ensure reproducibility, and is more
|
||||
# commonly ignored for libraries.
|
||||
#uv.lock
|
||||
|
||||
# poetry
|
||||
# Similar to Pipfile.lock, it is generally recommended to include poetry.lock in version control.
|
||||
# This is especially recommended for binary packages to ensure reproducibility, and is more
|
||||
# commonly ignored for libraries.
|
||||
# https://python-poetry.org/docs/basic-usage/#commit-your-poetrylock-file-to-version-control
|
||||
#poetry.lock
|
||||
|
||||
# pdm
|
||||
# Similar to Pipfile.lock, it is generally recommended to include pdm.lock in version control.
|
||||
#pdm.lock
|
||||
# pdm stores project-wide configurations in .pdm.toml, but it is recommended to not include it
|
||||
# in version control.
|
||||
# https://pdm.fming.dev/latest/usage/project/#working-with-version-control
|
||||
.pdm.toml
|
||||
.pdm-python
|
||||
.pdm-build/
|
||||
|
||||
# PEP 582; used by e.g. github.com/David-OConnor/pyflow and github.com/pdm-project/pdm
|
||||
__pypackages__/
|
||||
|
||||
# Celery stuff
|
||||
celerybeat-schedule
|
||||
celerybeat.pid
|
||||
|
||||
# SageMath parsed files
|
||||
*.sage.py
|
||||
|
||||
# Environments
|
||||
.env
|
||||
.venv
|
||||
env/
|
||||
venv/
|
||||
ENV/
|
||||
env.bak/
|
||||
venv.bak/
|
||||
|
||||
# Spyder project settings
|
||||
.spyderproject
|
||||
.spyproject
|
||||
|
||||
# Rope project settings
|
||||
.ropeproject
|
||||
|
||||
# mkdocs documentation
|
||||
/site
|
||||
|
||||
# mypy
|
||||
.mypy_cache/
|
||||
.dmypy.json
|
||||
dmypy.json
|
||||
|
||||
# Pyre type checker
|
||||
.pyre/
|
||||
|
||||
# pytype static type analyzer
|
||||
.pytype/
|
||||
|
||||
# Cython debug symbols
|
||||
cython_debug/
|
||||
|
||||
# PyCharm
|
||||
# JetBrains specific template is maintained in a separate JetBrains.gitignore that can
|
||||
# be found at https://github.com/github/gitignore/blob/main/Global/JetBrains.gitignore
|
||||
# and can be added to the global gitignore or merged into this file. For a more nuclear
|
||||
# option (not recommended) you can uncomment the following to ignore the entire idea folder.
|
||||
#.idea/
|
||||
|
||||
# Ruff stuff:
|
||||
.ruff_cache/
|
||||
|
||||
# PyPI configuration file
|
||||
.pypirc
|
||||
.vscode
|
||||
|
|
@ -225,4 +225,4 @@ from .tokenizer_utils import *
|
|||
from .trainer import *
|
||||
|
||||
# Patch TRL trainers for backwards compatibility
|
||||
_patch_trl_trainer()
|
||||
_patch_trl_trainer()
|
||||
661
unsloth/kernels/moe/LICENSE
Normal file
661
unsloth/kernels/moe/LICENSE
Normal file
|
|
@ -0,0 +1,661 @@
|
|||
GNU AFFERO GENERAL PUBLIC LICENSE
|
||||
Version 3, 19 November 2007
|
||||
|
||||
Copyright (C) 2007 Free Software Foundation, Inc. <https://fsf.org/>
|
||||
Everyone is permitted to copy and distribute verbatim copies
|
||||
of this license document, but changing it is not allowed.
|
||||
|
||||
Preamble
|
||||
|
||||
The GNU Affero General Public License is a free, copyleft license for
|
||||
software and other kinds of works, specifically designed to ensure
|
||||
cooperation with the community in the case of network server software.
|
||||
|
||||
The licenses for most software and other practical works are designed
|
||||
to take away your freedom to share and change the works. By contrast,
|
||||
our General Public Licenses are intended to guarantee your freedom to
|
||||
share and change all versions of a program--to make sure it remains free
|
||||
software for all its users.
|
||||
|
||||
When we speak of free software, we are referring to freedom, not
|
||||
price. Our General Public Licenses are designed to make sure that you
|
||||
have the freedom to distribute copies of free software (and charge for
|
||||
them if you wish), that you receive source code or can get it if you
|
||||
want it, that you can change the software or use pieces of it in new
|
||||
free programs, and that you know you can do these things.
|
||||
|
||||
Developers that use our General Public Licenses protect your rights
|
||||
with two steps: (1) assert copyright on the software, and (2) offer
|
||||
you this License which gives you legal permission to copy, distribute
|
||||
and/or modify the software.
|
||||
|
||||
A secondary benefit of defending all users' freedom is that
|
||||
improvements made in alternate versions of the program, if they
|
||||
receive widespread use, become available for other developers to
|
||||
incorporate. Many developers of free software are heartened and
|
||||
encouraged by the resulting cooperation. However, in the case of
|
||||
software used on network servers, this result may fail to come about.
|
||||
The GNU General Public License permits making a modified version and
|
||||
letting the public access it on a server without ever releasing its
|
||||
source code to the public.
|
||||
|
||||
The GNU Affero General Public License is designed specifically to
|
||||
ensure that, in such cases, the modified source code becomes available
|
||||
to the community. It requires the operator of a network server to
|
||||
provide the source code of the modified version running there to the
|
||||
users of that server. Therefore, public use of a modified version, on
|
||||
a publicly accessible server, gives the public access to the source
|
||||
code of the modified version.
|
||||
|
||||
An older license, called the Affero General Public License and
|
||||
published by Affero, was designed to accomplish similar goals. This is
|
||||
a different license, not a version of the Affero GPL, but Affero has
|
||||
released a new version of the Affero GPL which permits relicensing under
|
||||
this license.
|
||||
|
||||
The precise terms and conditions for copying, distribution and
|
||||
modification follow.
|
||||
|
||||
TERMS AND CONDITIONS
|
||||
|
||||
0. Definitions.
|
||||
|
||||
"This License" refers to version 3 of the GNU Affero General Public License.
|
||||
|
||||
"Copyright" also means copyright-like laws that apply to other kinds of
|
||||
works, such as semiconductor masks.
|
||||
|
||||
"The Program" refers to any copyrightable work licensed under this
|
||||
License. Each licensee is addressed as "you". "Licensees" and
|
||||
"recipients" may be individuals or organizations.
|
||||
|
||||
To "modify" a work means to copy from or adapt all or part of the work
|
||||
in a fashion requiring copyright permission, other than the making of an
|
||||
exact copy. The resulting work is called a "modified version" of the
|
||||
earlier work or a work "based on" the earlier work.
|
||||
|
||||
A "covered work" means either the unmodified Program or a work based
|
||||
on the Program.
|
||||
|
||||
To "propagate" a work means to do anything with it that, without
|
||||
permission, would make you directly or secondarily liable for
|
||||
infringement under applicable copyright law, except executing it on a
|
||||
computer or modifying a private copy. Propagation includes copying,
|
||||
distribution (with or without modification), making available to the
|
||||
public, and in some countries other activities as well.
|
||||
|
||||
To "convey" a work means any kind of propagation that enables other
|
||||
parties to make or receive copies. Mere interaction with a user through
|
||||
a computer network, with no transfer of a copy, is not conveying.
|
||||
|
||||
An interactive user interface displays "Appropriate Legal Notices"
|
||||
to the extent that it includes a convenient and prominently visible
|
||||
feature that (1) displays an appropriate copyright notice, and (2)
|
||||
tells the user that there is no warranty for the work (except to the
|
||||
extent that warranties are provided), that licensees may convey the
|
||||
work under this License, and how to view a copy of this License. If
|
||||
the interface presents a list of user commands or options, such as a
|
||||
menu, a prominent item in the list meets this criterion.
|
||||
|
||||
1. Source Code.
|
||||
|
||||
The "source code" for a work means the preferred form of the work
|
||||
for making modifications to it. "Object code" means any non-source
|
||||
form of a work.
|
||||
|
||||
A "Standard Interface" means an interface that either is an official
|
||||
standard defined by a recognized standards body, or, in the case of
|
||||
interfaces specified for a particular programming language, one that
|
||||
is widely used among developers working in that language.
|
||||
|
||||
The "System Libraries" of an executable work include anything, other
|
||||
than the work as a whole, that (a) is included in the normal form of
|
||||
packaging a Major Component, but which is not part of that Major
|
||||
Component, and (b) serves only to enable use of the work with that
|
||||
Major Component, or to implement a Standard Interface for which an
|
||||
implementation is available to the public in source code form. A
|
||||
"Major Component", in this context, means a major essential component
|
||||
(kernel, window system, and so on) of the specific operating system
|
||||
(if any) on which the executable work runs, or a compiler used to
|
||||
produce the work, or an object code interpreter used to run it.
|
||||
|
||||
The "Corresponding Source" for a work in object code form means all
|
||||
the source code needed to generate, install, and (for an executable
|
||||
work) run the object code and to modify the work, including scripts to
|
||||
control those activities. However, it does not include the work's
|
||||
System Libraries, or general-purpose tools or generally available free
|
||||
programs which are used unmodified in performing those activities but
|
||||
which are not part of the work. For example, Corresponding Source
|
||||
includes interface definition files associated with source files for
|
||||
the work, and the source code for shared libraries and dynamically
|
||||
linked subprograms that the work is specifically designed to require,
|
||||
such as by intimate data communication or control flow between those
|
||||
subprograms and other parts of the work.
|
||||
|
||||
The Corresponding Source need not include anything that users
|
||||
can regenerate automatically from other parts of the Corresponding
|
||||
Source.
|
||||
|
||||
The Corresponding Source for a work in source code form is that
|
||||
same work.
|
||||
|
||||
2. Basic Permissions.
|
||||
|
||||
All rights granted under this License are granted for the term of
|
||||
copyright on the Program, and are irrevocable provided the stated
|
||||
conditions are met. This License explicitly affirms your unlimited
|
||||
permission to run the unmodified Program. The output from running a
|
||||
covered work is covered by this License only if the output, given its
|
||||
content, constitutes a covered work. This License acknowledges your
|
||||
rights of fair use or other equivalent, as provided by copyright law.
|
||||
|
||||
You may make, run and propagate covered works that you do not
|
||||
convey, without conditions so long as your license otherwise remains
|
||||
in force. You may convey covered works to others for the sole purpose
|
||||
of having them make modifications exclusively for you, or provide you
|
||||
with facilities for running those works, provided that you comply with
|
||||
the terms of this License in conveying all material for which you do
|
||||
not control copyright. Those thus making or running the covered works
|
||||
for you must do so exclusively on your behalf, under your direction
|
||||
and control, on terms that prohibit them from making any copies of
|
||||
your copyrighted material outside their relationship with you.
|
||||
|
||||
Conveying under any other circumstances is permitted solely under
|
||||
the conditions stated below. Sublicensing is not allowed; section 10
|
||||
makes it unnecessary.
|
||||
|
||||
3. Protecting Users' Legal Rights From Anti-Circumvention Law.
|
||||
|
||||
No covered work shall be deemed part of an effective technological
|
||||
measure under any applicable law fulfilling obligations under article
|
||||
11 of the WIPO copyright treaty adopted on 20 December 1996, or
|
||||
similar laws prohibiting or restricting circumvention of such
|
||||
measures.
|
||||
|
||||
When you convey a covered work, you waive any legal power to forbid
|
||||
circumvention of technological measures to the extent such circumvention
|
||||
is effected by exercising rights under this License with respect to
|
||||
the covered work, and you disclaim any intention to limit operation or
|
||||
modification of the work as a means of enforcing, against the work's
|
||||
users, your or third parties' legal rights to forbid circumvention of
|
||||
technological measures.
|
||||
|
||||
4. Conveying Verbatim Copies.
|
||||
|
||||
You may convey verbatim copies of the Program's source code as you
|
||||
receive it, in any medium, provided that you conspicuously and
|
||||
appropriately publish on each copy an appropriate copyright notice;
|
||||
keep intact all notices stating that this License and any
|
||||
non-permissive terms added in accord with section 7 apply to the code;
|
||||
keep intact all notices of the absence of any warranty; and give all
|
||||
recipients a copy of this License along with the Program.
|
||||
|
||||
You may charge any price or no price for each copy that you convey,
|
||||
and you may offer support or warranty protection for a fee.
|
||||
|
||||
5. Conveying Modified Source Versions.
|
||||
|
||||
You may convey a work based on the Program, or the modifications to
|
||||
produce it from the Program, in the form of source code under the
|
||||
terms of section 4, provided that you also meet all of these conditions:
|
||||
|
||||
a) The work must carry prominent notices stating that you modified
|
||||
it, and giving a relevant date.
|
||||
|
||||
b) The work must carry prominent notices stating that it is
|
||||
released under this License and any conditions added under section
|
||||
7. This requirement modifies the requirement in section 4 to
|
||||
"keep intact all notices".
|
||||
|
||||
c) You must license the entire work, as a whole, under this
|
||||
License to anyone who comes into possession of a copy. This
|
||||
License will therefore apply, along with any applicable section 7
|
||||
additional terms, to the whole of the work, and all its parts,
|
||||
regardless of how they are packaged. This License gives no
|
||||
permission to license the work in any other way, but it does not
|
||||
invalidate such permission if you have separately received it.
|
||||
|
||||
d) If the work has interactive user interfaces, each must display
|
||||
Appropriate Legal Notices; however, if the Program has interactive
|
||||
interfaces that do not display Appropriate Legal Notices, your
|
||||
work need not make them do so.
|
||||
|
||||
A compilation of a covered work with other separate and independent
|
||||
works, which are not by their nature extensions of the covered work,
|
||||
and which are not combined with it such as to form a larger program,
|
||||
in or on a volume of a storage or distribution medium, is called an
|
||||
"aggregate" if the compilation and its resulting copyright are not
|
||||
used to limit the access or legal rights of the compilation's users
|
||||
beyond what the individual works permit. Inclusion of a covered work
|
||||
in an aggregate does not cause this License to apply to the other
|
||||
parts of the aggregate.
|
||||
|
||||
6. Conveying Non-Source Forms.
|
||||
|
||||
You may convey a covered work in object code form under the terms
|
||||
of sections 4 and 5, provided that you also convey the
|
||||
machine-readable Corresponding Source under the terms of this License,
|
||||
in one of these ways:
|
||||
|
||||
a) Convey the object code in, or embodied in, a physical product
|
||||
(including a physical distribution medium), accompanied by the
|
||||
Corresponding Source fixed on a durable physical medium
|
||||
customarily used for software interchange.
|
||||
|
||||
b) Convey the object code in, or embodied in, a physical product
|
||||
(including a physical distribution medium), accompanied by a
|
||||
written offer, valid for at least three years and valid for as
|
||||
long as you offer spare parts or customer support for that product
|
||||
model, to give anyone who possesses the object code either (1) a
|
||||
copy of the Corresponding Source for all the software in the
|
||||
product that is covered by this License, on a durable physical
|
||||
medium customarily used for software interchange, for a price no
|
||||
more than your reasonable cost of physically performing this
|
||||
conveying of source, or (2) access to copy the
|
||||
Corresponding Source from a network server at no charge.
|
||||
|
||||
c) Convey individual copies of the object code with a copy of the
|
||||
written offer to provide the Corresponding Source. This
|
||||
alternative is allowed only occasionally and noncommercially, and
|
||||
only if you received the object code with such an offer, in accord
|
||||
with subsection 6b.
|
||||
|
||||
d) Convey the object code by offering access from a designated
|
||||
place (gratis or for a charge), and offer equivalent access to the
|
||||
Corresponding Source in the same way through the same place at no
|
||||
further charge. You need not require recipients to copy the
|
||||
Corresponding Source along with the object code. If the place to
|
||||
copy the object code is a network server, the Corresponding Source
|
||||
may be on a different server (operated by you or a third party)
|
||||
that supports equivalent copying facilities, provided you maintain
|
||||
clear directions next to the object code saying where to find the
|
||||
Corresponding Source. Regardless of what server hosts the
|
||||
Corresponding Source, you remain obligated to ensure that it is
|
||||
available for as long as needed to satisfy these requirements.
|
||||
|
||||
e) Convey the object code using peer-to-peer transmission, provided
|
||||
you inform other peers where the object code and Corresponding
|
||||
Source of the work are being offered to the general public at no
|
||||
charge under subsection 6d.
|
||||
|
||||
A separable portion of the object code, whose source code is excluded
|
||||
from the Corresponding Source as a System Library, need not be
|
||||
included in conveying the object code work.
|
||||
|
||||
A "User Product" is either (1) a "consumer product", which means any
|
||||
tangible personal property which is normally used for personal, family,
|
||||
or household purposes, or (2) anything designed or sold for incorporation
|
||||
into a dwelling. In determining whether a product is a consumer product,
|
||||
doubtful cases shall be resolved in favor of coverage. For a particular
|
||||
product received by a particular user, "normally used" refers to a
|
||||
typical or common use of that class of product, regardless of the status
|
||||
of the particular user or of the way in which the particular user
|
||||
actually uses, or expects or is expected to use, the product. A product
|
||||
is a consumer product regardless of whether the product has substantial
|
||||
commercial, industrial or non-consumer uses, unless such uses represent
|
||||
the only significant mode of use of the product.
|
||||
|
||||
"Installation Information" for a User Product means any methods,
|
||||
procedures, authorization keys, or other information required to install
|
||||
and execute modified versions of a covered work in that User Product from
|
||||
a modified version of its Corresponding Source. The information must
|
||||
suffice to ensure that the continued functioning of the modified object
|
||||
code is in no case prevented or interfered with solely because
|
||||
modification has been made.
|
||||
|
||||
If you convey an object code work under this section in, or with, or
|
||||
specifically for use in, a User Product, and the conveying occurs as
|
||||
part of a transaction in which the right of possession and use of the
|
||||
User Product is transferred to the recipient in perpetuity or for a
|
||||
fixed term (regardless of how the transaction is characterized), the
|
||||
Corresponding Source conveyed under this section must be accompanied
|
||||
by the Installation Information. But this requirement does not apply
|
||||
if neither you nor any third party retains the ability to install
|
||||
modified object code on the User Product (for example, the work has
|
||||
been installed in ROM).
|
||||
|
||||
The requirement to provide Installation Information does not include a
|
||||
requirement to continue to provide support service, warranty, or updates
|
||||
for a work that has been modified or installed by the recipient, or for
|
||||
the User Product in which it has been modified or installed. Access to a
|
||||
network may be denied when the modification itself materially and
|
||||
adversely affects the operation of the network or violates the rules and
|
||||
protocols for communication across the network.
|
||||
|
||||
Corresponding Source conveyed, and Installation Information provided,
|
||||
in accord with this section must be in a format that is publicly
|
||||
documented (and with an implementation available to the public in
|
||||
source code form), and must require no special password or key for
|
||||
unpacking, reading or copying.
|
||||
|
||||
7. Additional Terms.
|
||||
|
||||
"Additional permissions" are terms that supplement the terms of this
|
||||
License by making exceptions from one or more of its conditions.
|
||||
Additional permissions that are applicable to the entire Program shall
|
||||
be treated as though they were included in this License, to the extent
|
||||
that they are valid under applicable law. If additional permissions
|
||||
apply only to part of the Program, that part may be used separately
|
||||
under those permissions, but the entire Program remains governed by
|
||||
this License without regard to the additional permissions.
|
||||
|
||||
When you convey a copy of a covered work, you may at your option
|
||||
remove any additional permissions from that copy, or from any part of
|
||||
it. (Additional permissions may be written to require their own
|
||||
removal in certain cases when you modify the work.) You may place
|
||||
additional permissions on material, added by you to a covered work,
|
||||
for which you have or can give appropriate copyright permission.
|
||||
|
||||
Notwithstanding any other provision of this License, for material you
|
||||
add to a covered work, you may (if authorized by the copyright holders of
|
||||
that material) supplement the terms of this License with terms:
|
||||
|
||||
a) Disclaiming warranty or limiting liability differently from the
|
||||
terms of sections 15 and 16 of this License; or
|
||||
|
||||
b) Requiring preservation of specified reasonable legal notices or
|
||||
author attributions in that material or in the Appropriate Legal
|
||||
Notices displayed by works containing it; or
|
||||
|
||||
c) Prohibiting misrepresentation of the origin of that material, or
|
||||
requiring that modified versions of such material be marked in
|
||||
reasonable ways as different from the original version; or
|
||||
|
||||
d) Limiting the use for publicity purposes of names of licensors or
|
||||
authors of the material; or
|
||||
|
||||
e) Declining to grant rights under trademark law for use of some
|
||||
trade names, trademarks, or service marks; or
|
||||
|
||||
f) Requiring indemnification of licensors and authors of that
|
||||
material by anyone who conveys the material (or modified versions of
|
||||
it) with contractual assumptions of liability to the recipient, for
|
||||
any liability that these contractual assumptions directly impose on
|
||||
those licensors and authors.
|
||||
|
||||
All other non-permissive additional terms are considered "further
|
||||
restrictions" within the meaning of section 10. If the Program as you
|
||||
received it, or any part of it, contains a notice stating that it is
|
||||
governed by this License along with a term that is a further
|
||||
restriction, you may remove that term. If a license document contains
|
||||
a further restriction but permits relicensing or conveying under this
|
||||
License, you may add to a covered work material governed by the terms
|
||||
of that license document, provided that the further restriction does
|
||||
not survive such relicensing or conveying.
|
||||
|
||||
If you add terms to a covered work in accord with this section, you
|
||||
must place, in the relevant source files, a statement of the
|
||||
additional terms that apply to those files, or a notice indicating
|
||||
where to find the applicable terms.
|
||||
|
||||
Additional terms, permissive or non-permissive, may be stated in the
|
||||
form of a separately written license, or stated as exceptions;
|
||||
the above requirements apply either way.
|
||||
|
||||
8. Termination.
|
||||
|
||||
You may not propagate or modify a covered work except as expressly
|
||||
provided under this License. Any attempt otherwise to propagate or
|
||||
modify it is void, and will automatically terminate your rights under
|
||||
this License (including any patent licenses granted under the third
|
||||
paragraph of section 11).
|
||||
|
||||
However, if you cease all violation of this License, then your
|
||||
license from a particular copyright holder is reinstated (a)
|
||||
provisionally, unless and until the copyright holder explicitly and
|
||||
finally terminates your license, and (b) permanently, if the copyright
|
||||
holder fails to notify you of the violation by some reasonable means
|
||||
prior to 60 days after the cessation.
|
||||
|
||||
Moreover, your license from a particular copyright holder is
|
||||
reinstated permanently if the copyright holder notifies you of the
|
||||
violation by some reasonable means, this is the first time you have
|
||||
received notice of violation of this License (for any work) from that
|
||||
copyright holder, and you cure the violation prior to 30 days after
|
||||
your receipt of the notice.
|
||||
|
||||
Termination of your rights under this section does not terminate the
|
||||
licenses of parties who have received copies or rights from you under
|
||||
this License. If your rights have been terminated and not permanently
|
||||
reinstated, you do not qualify to receive new licenses for the same
|
||||
material under section 10.
|
||||
|
||||
9. Acceptance Not Required for Having Copies.
|
||||
|
||||
You are not required to accept this License in order to receive or
|
||||
run a copy of the Program. Ancillary propagation of a covered work
|
||||
occurring solely as a consequence of using peer-to-peer transmission
|
||||
to receive a copy likewise does not require acceptance. However,
|
||||
nothing other than this License grants you permission to propagate or
|
||||
modify any covered work. These actions infringe copyright if you do
|
||||
not accept this License. Therefore, by modifying or propagating a
|
||||
covered work, you indicate your acceptance of this License to do so.
|
||||
|
||||
10. Automatic Licensing of Downstream Recipients.
|
||||
|
||||
Each time you convey a covered work, the recipient automatically
|
||||
receives a license from the original licensors, to run, modify and
|
||||
propagate that work, subject to this License. You are not responsible
|
||||
for enforcing compliance by third parties with this License.
|
||||
|
||||
An "entity transaction" is a transaction transferring control of an
|
||||
organization, or substantially all assets of one, or subdividing an
|
||||
organization, or merging organizations. If propagation of a covered
|
||||
work results from an entity transaction, each party to that
|
||||
transaction who receives a copy of the work also receives whatever
|
||||
licenses to the work the party's predecessor in interest had or could
|
||||
give under the previous paragraph, plus a right to possession of the
|
||||
Corresponding Source of the work from the predecessor in interest, if
|
||||
the predecessor has it or can get it with reasonable efforts.
|
||||
|
||||
You may not impose any further restrictions on the exercise of the
|
||||
rights granted or affirmed under this License. For example, you may
|
||||
not impose a license fee, royalty, or other charge for exercise of
|
||||
rights granted under this License, and you may not initiate litigation
|
||||
(including a cross-claim or counterclaim in a lawsuit) alleging that
|
||||
any patent claim is infringed by making, using, selling, offering for
|
||||
sale, or importing the Program or any portion of it.
|
||||
|
||||
11. Patents.
|
||||
|
||||
A "contributor" is a copyright holder who authorizes use under this
|
||||
License of the Program or a work on which the Program is based. The
|
||||
work thus licensed is called the contributor's "contributor version".
|
||||
|
||||
A contributor's "essential patent claims" are all patent claims
|
||||
owned or controlled by the contributor, whether already acquired or
|
||||
hereafter acquired, that would be infringed by some manner, permitted
|
||||
by this License, of making, using, or selling its contributor version,
|
||||
but do not include claims that would be infringed only as a
|
||||
consequence of further modification of the contributor version. For
|
||||
purposes of this definition, "control" includes the right to grant
|
||||
patent sublicenses in a manner consistent with the requirements of
|
||||
this License.
|
||||
|
||||
Each contributor grants you a non-exclusive, worldwide, royalty-free
|
||||
patent license under the contributor's essential patent claims, to
|
||||
make, use, sell, offer for sale, import and otherwise run, modify and
|
||||
propagate the contents of its contributor version.
|
||||
|
||||
In the following three paragraphs, a "patent license" is any express
|
||||
agreement or commitment, however denominated, not to enforce a patent
|
||||
(such as an express permission to practice a patent or covenant not to
|
||||
sue for patent infringement). To "grant" such a patent license to a
|
||||
party means to make such an agreement or commitment not to enforce a
|
||||
patent against the party.
|
||||
|
||||
If you convey a covered work, knowingly relying on a patent license,
|
||||
and the Corresponding Source of the work is not available for anyone
|
||||
to copy, free of charge and under the terms of this License, through a
|
||||
publicly available network server or other readily accessible means,
|
||||
then you must either (1) cause the Corresponding Source to be so
|
||||
available, or (2) arrange to deprive yourself of the benefit of the
|
||||
patent license for this particular work, or (3) arrange, in a manner
|
||||
consistent with the requirements of this License, to extend the patent
|
||||
license to downstream recipients. "Knowingly relying" means you have
|
||||
actual knowledge that, but for the patent license, your conveying the
|
||||
covered work in a country, or your recipient's use of the covered work
|
||||
in a country, would infringe one or more identifiable patents in that
|
||||
country that you have reason to believe are valid.
|
||||
|
||||
If, pursuant to or in connection with a single transaction or
|
||||
arrangement, you convey, or propagate by procuring conveyance of, a
|
||||
covered work, and grant a patent license to some of the parties
|
||||
receiving the covered work authorizing them to use, propagate, modify
|
||||
or convey a specific copy of the covered work, then the patent license
|
||||
you grant is automatically extended to all recipients of the covered
|
||||
work and works based on it.
|
||||
|
||||
A patent license is "discriminatory" if it does not include within
|
||||
the scope of its coverage, prohibits the exercise of, or is
|
||||
conditioned on the non-exercise of one or more of the rights that are
|
||||
specifically granted under this License. You may not convey a covered
|
||||
work if you are a party to an arrangement with a third party that is
|
||||
in the business of distributing software, under which you make payment
|
||||
to the third party based on the extent of your activity of conveying
|
||||
the work, and under which the third party grants, to any of the
|
||||
parties who would receive the covered work from you, a discriminatory
|
||||
patent license (a) in connection with copies of the covered work
|
||||
conveyed by you (or copies made from those copies), or (b) primarily
|
||||
for and in connection with specific products or compilations that
|
||||
contain the covered work, unless you entered into that arrangement,
|
||||
or that patent license was granted, prior to 28 March 2007.
|
||||
|
||||
Nothing in this License shall be construed as excluding or limiting
|
||||
any implied license or other defenses to infringement that may
|
||||
otherwise be available to you under applicable patent law.
|
||||
|
||||
12. No Surrender of Others' Freedom.
|
||||
|
||||
If conditions are imposed on you (whether by court order, agreement or
|
||||
otherwise) that contradict the conditions of this License, they do not
|
||||
excuse you from the conditions of this License. If you cannot convey a
|
||||
covered work so as to satisfy simultaneously your obligations under this
|
||||
License and any other pertinent obligations, then as a consequence you may
|
||||
not convey it at all. For example, if you agree to terms that obligate you
|
||||
to collect a royalty for further conveying from those to whom you convey
|
||||
the Program, the only way you could satisfy both those terms and this
|
||||
License would be to refrain entirely from conveying the Program.
|
||||
|
||||
13. Remote Network Interaction; Use with the GNU General Public License.
|
||||
|
||||
Notwithstanding any other provision of this License, if you modify the
|
||||
Program, your modified version must prominently offer all users
|
||||
interacting with it remotely through a computer network (if your version
|
||||
supports such interaction) an opportunity to receive the Corresponding
|
||||
Source of your version by providing access to the Corresponding Source
|
||||
from a network server at no charge, through some standard or customary
|
||||
means of facilitating copying of software. This Corresponding Source
|
||||
shall include the Corresponding Source for any work covered by version 3
|
||||
of the GNU General Public License that is incorporated pursuant to the
|
||||
following paragraph.
|
||||
|
||||
Notwithstanding any other provision of this License, you have
|
||||
permission to link or combine any covered work with a work licensed
|
||||
under version 3 of the GNU General Public License into a single
|
||||
combined work, and to convey the resulting work. The terms of this
|
||||
License will continue to apply to the part which is the covered work,
|
||||
but the work with which it is combined will remain governed by version
|
||||
3 of the GNU General Public License.
|
||||
|
||||
14. Revised Versions of this License.
|
||||
|
||||
The Free Software Foundation may publish revised and/or new versions of
|
||||
the GNU Affero General Public License from time to time. Such new versions
|
||||
will be similar in spirit to the present version, but may differ in detail to
|
||||
address new problems or concerns.
|
||||
|
||||
Each version is given a distinguishing version number. If the
|
||||
Program specifies that a certain numbered version of the GNU Affero General
|
||||
Public License "or any later version" applies to it, you have the
|
||||
option of following the terms and conditions either of that numbered
|
||||
version or of any later version published by the Free Software
|
||||
Foundation. If the Program does not specify a version number of the
|
||||
GNU Affero General Public License, you may choose any version ever published
|
||||
by the Free Software Foundation.
|
||||
|
||||
If the Program specifies that a proxy can decide which future
|
||||
versions of the GNU Affero General Public License can be used, that proxy's
|
||||
public statement of acceptance of a version permanently authorizes you
|
||||
to choose that version for the Program.
|
||||
|
||||
Later license versions may give you additional or different
|
||||
permissions. However, no additional obligations are imposed on any
|
||||
author or copyright holder as a result of your choosing to follow a
|
||||
later version.
|
||||
|
||||
15. Disclaimer of Warranty.
|
||||
|
||||
THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY
|
||||
APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT
|
||||
HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY
|
||||
OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO,
|
||||
THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
|
||||
PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM
|
||||
IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF
|
||||
ALL NECESSARY SERVICING, REPAIR OR CORRECTION.
|
||||
|
||||
16. Limitation of Liability.
|
||||
|
||||
IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING
|
||||
WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS
|
||||
THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY
|
||||
GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE
|
||||
USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF
|
||||
DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD
|
||||
PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS),
|
||||
EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF
|
||||
SUCH DAMAGES.
|
||||
|
||||
17. Interpretation of Sections 15 and 16.
|
||||
|
||||
If the disclaimer of warranty and limitation of liability provided
|
||||
above cannot be given local legal effect according to their terms,
|
||||
reviewing courts shall apply local law that most closely approximates
|
||||
an absolute waiver of all civil liability in connection with the
|
||||
Program, unless a warranty or assumption of liability accompanies a
|
||||
copy of the Program in return for a fee.
|
||||
|
||||
END OF TERMS AND CONDITIONS
|
||||
|
||||
How to Apply These Terms to Your New Programs
|
||||
|
||||
If you develop a new program, and you want it to be of the greatest
|
||||
possible use to the public, the best way to achieve this is to make it
|
||||
free software which everyone can redistribute and change under these terms.
|
||||
|
||||
To do so, attach the following notices to the program. It is safest
|
||||
to attach them to the start of each source file to most effectively
|
||||
state the exclusion of warranty; and each file should have at least
|
||||
the "copyright" line and a pointer to where the full notice is found.
|
||||
|
||||
<one line to give the program's name and a brief idea of what it does.>
|
||||
Copyright (C) <year> <name of author>
|
||||
|
||||
This program is free software: you can redistribute it and/or modify
|
||||
it under the terms of the GNU Affero General Public License as published
|
||||
by the Free Software Foundation, either version 3 of the License, or
|
||||
(at your option) any later version.
|
||||
|
||||
This program is distributed in the hope that it will be useful,
|
||||
but WITHOUT ANY WARRANTY; without even the implied warranty of
|
||||
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
||||
GNU Affero General Public License for more details.
|
||||
|
||||
You should have received a copy of the GNU Affero General Public License
|
||||
along with this program. If not, see <https://www.gnu.org/licenses/>.
|
||||
|
||||
Also add information on how to contact you by electronic and paper mail.
|
||||
|
||||
If your software can interact with users remotely through a computer
|
||||
network, you should also make sure that it provides a way for users to
|
||||
get its source. For example, if your program is a web application, its
|
||||
interface could display a "Source" link that leads users to an archive
|
||||
of the code. There are many ways you could offer source, and different
|
||||
solutions will be better for different programs; see section 13 for the
|
||||
specific requirements.
|
||||
|
||||
You should also get your employer (if you work as a programmer) or school,
|
||||
if any, to sign a "copyright disclaimer" for the program, if necessary.
|
||||
For more information on this, and how to apply and follow the GNU AGPL, see
|
||||
<https://www.gnu.org/licenses/>.
|
||||
72
unsloth/kernels/moe/README.md
Normal file
72
unsloth/kernels/moe/README.md
Normal file
|
|
@ -0,0 +1,72 @@
|
|||
## MoE Grouped GEMM
|
||||
|
||||
Optimized implementation of `MoE MLP Block`.
|
||||
|
||||
### Background
|
||||
|
||||
`MoE MLP` requires the following steps:
|
||||
- Calculate `topk_weights` and `topk_indices`
|
||||
- If using a grouped gemm implementation, calculate permutation indices needed to rearrange tokens grouped by expert
|
||||
- For each expert:
|
||||
- `expert_tokens`: gather the tokens assigned to the expert
|
||||
- `first_gemm`: `gate / up proj` @ `expert_tokens`
|
||||
- `silu_and_mul`: `silu` and `mul` of `first_gemm`
|
||||
- `second_gemm`: `silu_and_mul` @ `down proj`
|
||||
- `scatter_second_gemm`: scatter the `second_gemm` to the original token order
|
||||
- `topk_weight_mul`: `second_gemm` @ `topk_weights`
|
||||
- `final_output`: if `topk > 1`, `topk_weight_mul.view(num_tokens, topk, -1).sum(dim=1)` else `topk_weight_mul`
|
||||
|
||||
One way to eliminate the loop is to use a grouped GEMM, where all expert GEMMs are computed within a single kernel, which iterates over tiles of the expert GEMMs as individual GEMMs, where each GEMM, the `A` matrix is `M' x K` and the `B` matrix is `K x N`, where `M'` is the number of tokens assigned to the expert and `B` is the weight matrix for that expert.
|
||||
|
||||
This requires an additional permute (and subsequent copy) of the hidden states such that the tokens assigned to each expert are contiguous in memory before running the first grouped GEMM within the Expert MLP.
|
||||
Additionally, after the second grouped GEMM, the hidden states must be permuted back to the original token order and multiplied by `topk_weights` to get the final output.
|
||||
|
||||
### Optimizations
|
||||
This repo implements a grouped GEMM-based MoE MLP with the following optimizations:
|
||||
- Eliminates the loop over experts by performing gemms as a grouped GEMM, computing the expert gemms within a single fused triton kernel
|
||||
- Fuses the permutation of hidden states from token order (original input order) to expert order (tokens grouped by expert) within the prologue of first the first grouped GEMM
|
||||
- Fuses the (un)permutation of hidden states from expert order back to token order in second GEMM
|
||||
- Fuses the mul of hidden states by expert weights within epilogue of second GEMM (only implemented for inference, not for training)
|
||||
|
||||
### Structure
|
||||
- `grouped_gemm/interface.py`: wrappers for the individual forward / backward kernels as well as the `torch.autograd.Function`
|
||||
- `grouped_gemm/kernels/forward.py`: forward kernel
|
||||
- `grouped_gemm/kernels/backward.py`: backward dX and dW kernels
|
||||
- `grouped_gemm/kernels/tuning.py`: manual tuning utils
|
||||
- `grouped_gemm/kernels/autotuning.py`: autotuning utils
|
||||
- `grouped_gemm/reference/moe_block.py`: contains `Qwen3MoeFusedGroupedGEMMBlock`, a reference implementation of Huggingface `Qwen3SparseMOEBlock` with fused triton kernel in-place of original HF expert computation
|
||||
- `grouped_gemm/reference/moe_ops.py`: supporting ops (routing, token sorting, etc.) and reference MoE block using a torch-native grouped gemm approach.
|
||||
|
||||
### Tests
|
||||
- `grouped_gemm/tests/test_grouped_gemm.py`: unit tests for forward, backward grouped gemm kernels as well as the wrapped grouped gemm autograd.Function. Best not to run this entire test suite at once due to the large number of parametrized unit tests. Rather, use filters to run specific
|
||||
sets of tests. E.g., to run forward tests with autotune turned on: `pytest -sv -k "forward and autotune" --tb=short tests/test_grouped_gemm.py`. Use the test function names and parameter ids for words to filter on.
|
||||
- `grouped_gemm/tests/test_qwen3_moe.py`: end to end test for Qwen3 MoE block. IMPORTANT: read `tests/run_qwen3_moe_tests.sh` as well as notes in the test itself for complications when running parametrized pytest test suites and triton / autotune. TLDR: use the test script and NOT pytest to run the tests.
|
||||
|
||||
### Benchmarks
|
||||
- `grouped_gemm/benchmark/benchmark_fused_moe.py`: benchmarks HF `Qwen3SpareMOEBlock` against the fused implementation
|
||||
Running with these flags on an `H100` to bench forward pass (run with `--help` to see all available flags):
|
||||
```
|
||||
python benchmark/benchmark_fused_moe.py --mode forward --seqlen 1024 --permute_x --permute_y --autotune
|
||||
```
|
||||
|
||||
For the backward bench:
|
||||
```
|
||||
python benchmark/benchmark_fused_moe.py --mode backward --seqlen 1024 --permute_x --permute_y --autotune
|
||||
```
|
||||
|
||||
On my machine and env, I get speedups > 25x and 14x respectively.
|
||||
|
||||
### Notes
|
||||
- Tested and benched on `H100`, though should run on Ampere and possibly even earlier gpu generations though the autotuning configs will need to be adjusted.
|
||||
- The env I used to develop the kernel was `pytorch 2.7/2.8` and `pytorch-triton 3.3`.
|
||||
- The kernels can be run either as autotuned (see `autotuning.py`) or with manually specified config (see `tuning.py`). Recommended to run using autotuner since the MoE block requires 2 configs for the forward (2 grouped gemms) and 4 for the backwards (dX and dW per grouped gemm, 2 grouped gemms).
|
||||
- Running with autotuning turned off with the default manual kernel config will result is **highly** sub-optimal performance as it is only meant for testing / debugging purposes.
|
||||
- I've tried to strike a balance between compilation time and autotuning search space -- can probably squeeze even more performance for specific workloads.
|
||||
|
||||
TODO:
|
||||
- TMA store: implemented but not enabled currently due to non-determinism arising from triton pipelining bug.
|
||||
- Warp specialization: Hopper support for WS not yet enabled on triton 3.3x branch which ships with latest pytorch 2.7.
|
||||
- Additional optimizations:
|
||||
- Fused / optimized implementations of routing, token sorting, etc.
|
||||
- Better software pipelining within grouped gemm
|
||||
- Threadblock swizzling for better L2 caching
|
||||
0
unsloth/kernels/moe/__init__.py
Normal file
0
unsloth/kernels/moe/__init__.py
Normal file
297
unsloth/kernels/moe/benchmark/benchmark_fused_moe.py
Normal file
297
unsloth/kernels/moe/benchmark/benchmark_fused_moe.py
Normal file
|
|
@ -0,0 +1,297 @@
|
|||
import argparse
|
||||
import time
|
||||
|
||||
import torch
|
||||
from grouped_gemm.kernels.autotuning import (
|
||||
DEFAULT_K_BLOCK_SIZES,
|
||||
DEFAULT_M_BLOCK_SIZES,
|
||||
DEFAULT_N_BLOCK_SIZES,
|
||||
DEFAULT_NUM_STAGES,
|
||||
DEFAULT_NUM_WARPS,
|
||||
)
|
||||
from grouped_gemm.kernels.tuning import (
|
||||
KernelConfigBackward_dW,
|
||||
KernelConfigBackward_dX,
|
||||
KernelConfigForward,
|
||||
KernelResult,
|
||||
TritonTuningContext,
|
||||
)
|
||||
from grouped_gemm.reference.moe_block import Qwen3MoeFusedGroupedGEMMBlock
|
||||
from transformers import AutoConfig
|
||||
from transformers.models.qwen3_moe import Qwen3MoeConfig
|
||||
from transformers.models.qwen3_moe.modeling_qwen3_moe import Qwen3MoeSparseMoeBlock
|
||||
from triton.testing import do_bench
|
||||
from utils import (
|
||||
create_kernel_configs,
|
||||
post_process_results,
|
||||
save_results,
|
||||
)
|
||||
|
||||
SEED = 42
|
||||
|
||||
def run_benchmark_forward(
|
||||
config: Qwen3MoeConfig,
|
||||
seqlen: int,
|
||||
dtype: torch.dtype,
|
||||
permute_x: bool,
|
||||
permute_y: bool,
|
||||
autotune: bool,
|
||||
kernel_config_fwd: KernelConfigForward = None,
|
||||
kernel_config_bwd_dW: KernelConfigBackward_dW = None,
|
||||
kernel_config_bwd_dX: KernelConfigBackward_dX = None,
|
||||
):
|
||||
torch.manual_seed(SEED) # Should not be needed when running using pytest -- autouse fixture in conftest.py
|
||||
device = "cuda"
|
||||
hidden_size = config.hidden_size
|
||||
bs = 1
|
||||
|
||||
# Reference op -- HF
|
||||
moe_block = Qwen3MoeSparseMoeBlock(config).to(device, dtype)
|
||||
|
||||
# Triton kernel grouped gemm version of MoE Block -- this is what we're testing
|
||||
fused_gemm_block = Qwen3MoeFusedGroupedGEMMBlock.from_hf(
|
||||
moe_block,
|
||||
permute_x=permute_x,
|
||||
permute_y=permute_y,
|
||||
autotune=autotune,
|
||||
kernel_config_fwd=kernel_config_fwd,
|
||||
kernel_config_bwd_dW=kernel_config_bwd_dW,
|
||||
kernel_config_bwd_dX=kernel_config_bwd_dX,
|
||||
).to(device, dtype)
|
||||
X = torch.randn(bs, seqlen, hidden_size, dtype=dtype, device=device, requires_grad=True)
|
||||
|
||||
ref_output, _ = moe_block(X)
|
||||
# Forward
|
||||
bench_forward_ref = lambda: moe_block(X)
|
||||
bench_forward_fused = lambda: fused_gemm_block(X)
|
||||
|
||||
ref_forward_time = do_bench(bench_forward_ref)
|
||||
with TritonTuningContext(kernel_config_fwd) as ctx:
|
||||
fused_forward_time = do_bench(bench_forward_fused)
|
||||
|
||||
if not ctx.success:
|
||||
return 0, 1
|
||||
|
||||
print(
|
||||
f"Forward: ref {ref_forward_time:.4f}, fused {fused_forward_time:.4f}, speedup {ref_forward_time / fused_forward_time:.1f}x"
|
||||
)
|
||||
return ref_forward_time, fused_forward_time
|
||||
|
||||
def run_benchmark_backward(
|
||||
config: Qwen3MoeConfig,
|
||||
seqlen: int,
|
||||
dtype: torch.dtype,
|
||||
permute_x: bool,
|
||||
permute_y: bool,
|
||||
autotune: bool,
|
||||
kernel_config_fwd: KernelConfigForward = None,
|
||||
kernel_config_bwd_dW: KernelConfigBackward_dW = None,
|
||||
kernel_config_bwd_dX: KernelConfigBackward_dX = None,
|
||||
dX_only: bool = False,
|
||||
dW_only: bool = False,
|
||||
):
|
||||
torch.manual_seed(SEED) # Should not be needed when running using pytest -- autouse fixture in conftest.py
|
||||
device = "cuda"
|
||||
hidden_size = config.hidden_size
|
||||
bs = 1
|
||||
|
||||
# Reference op -- HF
|
||||
moe_block = Qwen3MoeSparseMoeBlock(config).to(device, dtype)
|
||||
|
||||
# Triton kernel grouped gemm version of MoE Block -- this is what we're testing
|
||||
fused_gemm_block = Qwen3MoeFusedGroupedGEMMBlock.from_hf(
|
||||
moe_block,
|
||||
permute_x=permute_x,
|
||||
permute_y=permute_y,
|
||||
autotune=autotune,
|
||||
kernel_config_fwd=kernel_config_fwd,
|
||||
kernel_config_bwd_dW=kernel_config_bwd_dW,
|
||||
kernel_config_bwd_dX=kernel_config_bwd_dX,
|
||||
dX_only=dX_only,
|
||||
dW_only=dW_only,
|
||||
).to(device, dtype)
|
||||
|
||||
X = torch.randn(bs, seqlen, hidden_size, dtype=dtype, device=device, requires_grad=True)
|
||||
X_test = X.detach().clone().requires_grad_(True)
|
||||
|
||||
output, _ = moe_block(X)
|
||||
|
||||
# Prevent autotuning forward pass
|
||||
from grouped_gemm.kernels.forward import _autotuned_grouped_gemm_forward_kernel
|
||||
_autotuned_grouped_gemm_forward_kernel.configs = _autotuned_grouped_gemm_forward_kernel.configs[:20]
|
||||
test_output, _ = fused_gemm_block(X_test)
|
||||
|
||||
# Bench
|
||||
grad_output = torch.randn_like(output)
|
||||
bench_backward_ref = lambda: output.backward(grad_output, retain_graph=True) # noqa: E731
|
||||
bench_backward_fused = lambda: test_output.backward(grad_output, retain_graph=True) # noqa: E731
|
||||
|
||||
ref_backward_time = do_bench(bench_backward_ref, grad_to_none=[X, *moe_block.parameters()])
|
||||
fused_backward_time = do_bench(bench_backward_fused, grad_to_none=[X_test, *fused_gemm_block.parameters()])
|
||||
print(
|
||||
f"Backward: ref {ref_backward_time:.4f}, fused {fused_backward_time:.4f}, speedup {ref_backward_time / fused_backward_time:.1f}x"
|
||||
)
|
||||
return ref_backward_time, fused_backward_time
|
||||
|
||||
|
||||
def run_benchmark(
|
||||
mode: str,
|
||||
model_config: Qwen3MoeConfig,
|
||||
seqlen: int,
|
||||
dtype: torch.dtype,
|
||||
permute_x: bool,
|
||||
permute_y: bool,
|
||||
autotune: bool,
|
||||
kernel_config_fwd: KernelConfigForward = None,
|
||||
kernel_config_bwd_dW: KernelConfigBackward_dW = None,
|
||||
kernel_config_bwd_dX: KernelConfigBackward_dX = None,
|
||||
):
|
||||
|
||||
if mode == "forward":
|
||||
|
||||
ref_time, fused_time = run_benchmark_forward(
|
||||
model_config,
|
||||
seqlen,
|
||||
dtype,
|
||||
permute_x,
|
||||
permute_y,
|
||||
autotune,
|
||||
kernel_config_fwd,
|
||||
kernel_config_bwd_dW,
|
||||
kernel_config_bwd_dX,
|
||||
)
|
||||
elif mode == "dW":
|
||||
ref_time, fused_time = run_benchmark_backward(
|
||||
model_config,
|
||||
seqlen,
|
||||
dtype,
|
||||
permute_x,
|
||||
permute_y,
|
||||
autotune,
|
||||
kernel_config_fwd,
|
||||
kernel_config_bwd_dW,
|
||||
kernel_config_bwd_dX,
|
||||
dW_only=True,
|
||||
)
|
||||
elif mode == "dX":
|
||||
ref_time, fused_time = run_benchmark_backward(
|
||||
model_config,
|
||||
seqlen,
|
||||
dtype,
|
||||
permute_x,
|
||||
permute_y,
|
||||
autotune,
|
||||
kernel_config_fwd,
|
||||
kernel_config_bwd_dW,
|
||||
kernel_config_bwd_dX,
|
||||
dX_only=True,
|
||||
)
|
||||
elif mode == "backward":
|
||||
ref_time, fused_time = run_benchmark_backward(
|
||||
model_config,
|
||||
seqlen,
|
||||
dtype,
|
||||
permute_x,
|
||||
permute_y,
|
||||
autotune,
|
||||
kernel_config_fwd,
|
||||
kernel_config_bwd_dW,
|
||||
kernel_config_bwd_dX,
|
||||
dX_only=False,
|
||||
dW_only=False,
|
||||
)
|
||||
|
||||
return ref_time, fused_time
|
||||
|
||||
# NOTE: better to use autotuner for now, since the MoE block needs 2 different kernel configs for forward (2 grouped gemms, gate_up_proj and down_proj)
|
||||
# and the backward pass needs 4 different kernel configs (2 grouped gemms each for dW and dX)
|
||||
# The benchmark only supports 1 kernel config at a time so the same config will be used for both grouped gemms, which is suboptimal.
|
||||
|
||||
if __name__ == "__main__":
|
||||
parser = argparse.ArgumentParser()
|
||||
parser.add_argument("--results_dir", type=str, default="benchmark_results")
|
||||
parser.add_argument("--seqlen", type=int, default=1024)
|
||||
parser.add_argument("--dtype", type=str, choices=["bfloat16", "float16"], default="bfloat16")
|
||||
parser.add_argument("--permute_x", action="store_true")
|
||||
parser.add_argument("--permute_y", action="store_true")
|
||||
parser.add_argument("--autotune", action="store_true")
|
||||
parser.add_argument("--BLOCK_SIZE_M", nargs=2, type=int, default=[DEFAULT_M_BLOCK_SIZES[0], DEFAULT_M_BLOCK_SIZES[-1]])
|
||||
parser.add_argument("--BLOCK_SIZE_N", nargs=2, type=int, default=[DEFAULT_N_BLOCK_SIZES[0], DEFAULT_N_BLOCK_SIZES[-1]])
|
||||
parser.add_argument("--BLOCK_SIZE_K", nargs=2, type=int, default=[DEFAULT_K_BLOCK_SIZES[0], DEFAULT_K_BLOCK_SIZES[-1]])
|
||||
parser.add_argument("--num_warps", nargs=2, type=int, default=[DEFAULT_NUM_WARPS[0], DEFAULT_NUM_WARPS[-1]])
|
||||
parser.add_argument("--num_stages", nargs=2, type=int, default=[DEFAULT_NUM_STAGES[0], DEFAULT_NUM_STAGES[-1]])
|
||||
parser.add_argument("--use_tma_load_w", action="store_true") # No need to specify, will automatically parametrize these for each kernel config
|
||||
parser.add_argument("--use_tma_load_x", action="store_true") # No need to specify, will automatically parametrize these for each kernel config
|
||||
parser.add_argument("--use_tma_load_dy", action="store_true") # No need to specify, will automatically parametrize these for each kernel config
|
||||
parser.add_argument("--mode", type=str, choices=["forward", "backward", "dW", "dX"], default="forward")
|
||||
args = parser.parse_args()
|
||||
args.dtype = getattr(torch, args.dtype)
|
||||
|
||||
model_id = "Qwen/Qwen3-30B-A3B"
|
||||
model_config = AutoConfig.from_pretrained(model_id)
|
||||
|
||||
mode = args.mode
|
||||
|
||||
if args.autotune:
|
||||
print(
|
||||
f"Benchmarking {model_id} {mode}: seqlen={args.seqlen}, dtype={args.dtype}, permute_x={args.permute_x}, permute_y={args.permute_y}, autotune"
|
||||
)
|
||||
start_time = time.time()
|
||||
ref_time, fused_time = run_benchmark(
|
||||
args.mode,
|
||||
model_config,
|
||||
seqlen=args.seqlen,
|
||||
dtype=args.dtype,
|
||||
permute_x=args.permute_x,
|
||||
permute_y=args.permute_y,
|
||||
autotune=args.autotune,
|
||||
)
|
||||
end_time = time.time()
|
||||
print(f"Total time: {end_time - start_time:.4f} seconds")
|
||||
|
||||
else:
|
||||
kernel_configs = create_kernel_configs(args, args.permute_x, args.permute_y)
|
||||
print(f"Running {len(kernel_configs)} kernel configs")
|
||||
default_kernel_config_fwd = KernelConfigForward(permute_x=args.permute_x, permute_y=args.permute_y)
|
||||
default_kernel_config_bwd_dW = KernelConfigBackward_dW(permute_x=args.permute_x, permute_y=args.permute_y)
|
||||
default_kernel_config_bwd_dX = KernelConfigBackward_dX(permute_x=args.permute_x, permute_y=args.permute_y)
|
||||
results = []
|
||||
for kernel_config in kernel_configs:
|
||||
if args.mode == "forward":
|
||||
kernel_config_fwd = kernel_config
|
||||
kernel_config_bwd_dW = default_kernel_config_bwd_dW
|
||||
kernel_config_bwd_dX = default_kernel_config_bwd_dX
|
||||
elif args.mode == "dW":
|
||||
kernel_config_fwd = default_kernel_config_fwd
|
||||
kernel_config_bwd_dW = kernel_config
|
||||
kernel_config_bwd_dX = default_kernel_config_bwd_dX
|
||||
elif args.mode == "dX":
|
||||
kernel_config_fwd = default_kernel_config_fwd
|
||||
kernel_config_bwd_dW = default_kernel_config_bwd_dW
|
||||
kernel_config_bwd_dX = kernel_config
|
||||
else:
|
||||
raise ValueError(f"Invalid mode: {args.mode}")
|
||||
print(
|
||||
f"Benchmarking {model_id} {args.mode} with seqlen={args.seqlen}, dtype={args.dtype}, permute_x={args.permute_x}, permute_y={args.permute_y}, kernel_config_fwd={kernel_config_fwd}, kernel_config_bwd_dW={kernel_config_bwd_dW}, kernel_config_bwd_dX={kernel_config_bwd_dX}"
|
||||
)
|
||||
|
||||
ref_time, fused_time = run_benchmark(
|
||||
args.mode,
|
||||
model_config,
|
||||
seqlen=args.seqlen,
|
||||
dtype=args.dtype,
|
||||
permute_x=kernel_config.permute_x,
|
||||
permute_y=kernel_config.permute_y,
|
||||
autotune=False,
|
||||
kernel_config_fwd=kernel_config_fwd,
|
||||
kernel_config_bwd_dW=kernel_config_bwd_dW,
|
||||
kernel_config_bwd_dX=kernel_config_bwd_dX,
|
||||
)
|
||||
results.append(KernelResult(
|
||||
torch_time=ref_time,
|
||||
triton_time=fused_time,
|
||||
speedup=ref_time / fused_time,
|
||||
kernel_config=kernel_config,
|
||||
))
|
||||
df = post_process_results(results, args.mode, args.seqlen, args.dtype, args.autotune)
|
||||
save_results(df, args.results_dir, args.mode, args.seqlen, args.dtype, args.autotune)
|
||||
143
unsloth/kernels/moe/benchmark/utils.py
Normal file
143
unsloth/kernels/moe/benchmark/utils.py
Normal file
|
|
@ -0,0 +1,143 @@
|
|||
import argparse
|
||||
import datetime
|
||||
import json
|
||||
import logging
|
||||
import math
|
||||
import os
|
||||
from itertools import product
|
||||
|
||||
import pandas as pd
|
||||
import torch
|
||||
from grouped_gemm.kernels.tuning import (
|
||||
KernelConfigBackward_dW,
|
||||
KernelConfigBackward_dX,
|
||||
KernelConfigForward,
|
||||
KernelResult,
|
||||
)
|
||||
|
||||
SEED = 42
|
||||
|
||||
|
||||
def create_merged_results(
|
||||
df: pd.DataFrame, mode: str, seqlen: int, dtype: torch.dtype, autotune: bool
|
||||
):
|
||||
kernel_result_cols = df.columns.to_list()
|
||||
test_config_dict = {"mode": mode, "seqlen": seqlen, "dtype": dtype, "autotune": autotune}
|
||||
test_config_cols = list(test_config_dict.keys())
|
||||
for col in test_config_cols:
|
||||
df[col] = test_config_dict[col]
|
||||
# Reorder columns so that test config cols are first
|
||||
df = df[test_config_cols + kernel_result_cols]
|
||||
return df
|
||||
|
||||
|
||||
def post_process_results(
|
||||
results: list[KernelResult],
|
||||
mode: str,
|
||||
seqlen: int,
|
||||
dtype: torch.dtype,
|
||||
autotune: bool,
|
||||
):
|
||||
df = KernelResult.to_dataframe(results, sort_by="speedup")
|
||||
df = create_merged_results(df, mode, seqlen, dtype, autotune)
|
||||
return df
|
||||
|
||||
|
||||
def save_results(
|
||||
df: pd.DataFrame,
|
||||
results_dir: str,
|
||||
mode: str,
|
||||
seqlen: int,
|
||||
dtype: torch.dtype,
|
||||
autotune: bool,
|
||||
):
|
||||
dt = datetime.datetime.now().strftime("%Y%m%d_%H%M")
|
||||
save_dir = f"{results_dir}/{mode}"
|
||||
save_path = f"{save_dir}/{dt}_{seqlen}_{str(dtype).split('.')[-1]}.csv"
|
||||
if not os.path.exists(save_dir):
|
||||
os.makedirs(save_dir)
|
||||
print(f"Saving results to {save_path}")
|
||||
df.to_csv(save_path, index=False)
|
||||
|
||||
|
||||
def create_kernel_configs(args: argparse.Namespace, permute_x: bool, permute_y: bool):
|
||||
block_m_range = power_of_two_range(args.BLOCK_SIZE_M[0], args.BLOCK_SIZE_M[1])
|
||||
block_n_range = power_of_two_range(args.BLOCK_SIZE_N[0], args.BLOCK_SIZE_N[1])
|
||||
block_k_range = power_of_two_range(args.BLOCK_SIZE_K[0], args.BLOCK_SIZE_K[1])
|
||||
num_warps_range = multiples_of_range(args.num_warps[0], args.num_warps[1], step=2)
|
||||
num_stages_range = multiples_of_range(args.num_stages[0], args.num_stages[1], step=1)
|
||||
|
||||
mode = args.mode
|
||||
kernel_configs = []
|
||||
for block_m, block_n, block_k, num_warps, num_stages, tma_load_a, tma_load_b in product(
|
||||
block_m_range, block_n_range, block_k_range, num_warps_range, num_stages_range, [True, False], [True, False]
|
||||
):
|
||||
if mode == "forward":
|
||||
kernel_config = KernelConfigForward(
|
||||
BLOCK_SIZE_M=block_m,
|
||||
BLOCK_SIZE_N=block_n,
|
||||
BLOCK_SIZE_K=block_k,
|
||||
num_warps=num_warps,
|
||||
num_stages=num_stages,
|
||||
use_tma_load_w=tma_load_a,
|
||||
use_tma_load_x=tma_load_b,
|
||||
permute_x=permute_x,
|
||||
permute_y=permute_y,
|
||||
)
|
||||
elif mode == "dW":
|
||||
kernel_config = KernelConfigBackward_dW(
|
||||
BLOCK_SIZE_M=block_m,
|
||||
BLOCK_SIZE_N=block_n,
|
||||
BLOCK_SIZE_K=block_k,
|
||||
num_warps=num_warps,
|
||||
num_stages=num_stages,
|
||||
use_tma_load_dy=tma_load_a,
|
||||
use_tma_load_x=tma_load_b,
|
||||
permute_x=permute_x,
|
||||
permute_y=permute_y,
|
||||
)
|
||||
elif mode == "dX":
|
||||
kernel_config = KernelConfigBackward_dX(
|
||||
BLOCK_SIZE_M=block_m,
|
||||
BLOCK_SIZE_N=block_n,
|
||||
BLOCK_SIZE_K=block_k,
|
||||
num_warps=num_warps,
|
||||
num_stages=num_stages,
|
||||
use_tma_load_dy=tma_load_a,
|
||||
use_tma_load_w=tma_load_b,
|
||||
permute_x=permute_x,
|
||||
permute_y=permute_y,
|
||||
)
|
||||
else:
|
||||
raise ValueError(f"Invalid mode: {mode}")
|
||||
kernel_configs.append(kernel_config)
|
||||
|
||||
logging.info(f"Pruning {len(kernel_configs)} kernel configs")
|
||||
|
||||
pruned_configs = []
|
||||
for config in kernel_configs:
|
||||
if mode == "forward":
|
||||
if permute_x and config.use_tma_load_x:
|
||||
continue
|
||||
elif mode == "dW":
|
||||
if permute_x and config.use_tma_load_x:
|
||||
continue
|
||||
if permute_y and config.use_tma_load_dy:
|
||||
continue
|
||||
elif mode == "dX":
|
||||
if permute_y and config.use_tma_load_dy:
|
||||
continue
|
||||
pruned_configs.append(config)
|
||||
logging.info(f"After pruning, {len(pruned_configs)} kernel configs")
|
||||
|
||||
return pruned_configs
|
||||
|
||||
|
||||
def power_of_two_range(start, end):
|
||||
start = math.log2(start)
|
||||
end = math.log2(end)
|
||||
return [2**i for i in range(int(start), int(end) + 1)]
|
||||
|
||||
|
||||
def multiples_of_range(start, end, step=1):
|
||||
return list(range(start, end + step, step))
|
||||
0
unsloth/kernels/moe/grouped_gemm/__init__.py
Normal file
0
unsloth/kernels/moe/grouped_gemm/__init__.py
Normal file
965
unsloth/kernels/moe/grouped_gemm/interface.py
Normal file
965
unsloth/kernels/moe/grouped_gemm/interface.py
Normal file
|
|
@ -0,0 +1,965 @@
|
|||
import logging
|
||||
import warnings
|
||||
from dataclasses import asdict
|
||||
|
||||
import torch
|
||||
import triton
|
||||
|
||||
from grouped_gemm.kernels.backward import (
|
||||
_autotuned_grouped_gemm_dW_kernel,
|
||||
_autotuned_grouped_gemm_dX_kernel,
|
||||
_grouped_gemm_dW_kernel,
|
||||
_grouped_gemm_dX_kernel,
|
||||
)
|
||||
from grouped_gemm.kernels.forward import (
|
||||
_autotuned_grouped_gemm_forward_kernel,
|
||||
_grouped_gemm_forward_kernel,
|
||||
)
|
||||
from grouped_gemm.kernels.tuning import (
|
||||
KernelConfigBackward_dW,
|
||||
KernelConfigBackward_dX,
|
||||
KernelConfigForward,
|
||||
)
|
||||
|
||||
logger = logging.getLogger(__name__)
|
||||
# Set formatter to include timestamp, pathname and lineno
|
||||
formatter = logging.Formatter(
|
||||
"%(asctime)s::%(levelname)s,%(pathname)s:%(lineno)d:: %(message)s"
|
||||
)
|
||||
|
||||
# Add console handler
|
||||
ch = logging.StreamHandler()
|
||||
ch.setFormatter(formatter)
|
||||
logger.addHandler(ch)
|
||||
|
||||
_FUSED_MUL_WARN = False
|
||||
_SUPPORTS_TMA = None
|
||||
|
||||
|
||||
def supports_tma():
|
||||
global _SUPPORTS_TMA
|
||||
if _SUPPORTS_TMA is None:
|
||||
_SUPPORTS_TMA = torch.cuda.get_device_capability()[0] >= 9
|
||||
return _SUPPORTS_TMA
|
||||
|
||||
|
||||
_per_device_alloc_fns = {}
|
||||
|
||||
|
||||
def get_per_device_per_stream_alloc_fn(device):
|
||||
if device not in _per_device_alloc_fns:
|
||||
_per_stream_tensors = {}
|
||||
|
||||
def alloc_fn(size: int, alignment: int, stream):
|
||||
assert alignment == 128
|
||||
if (
|
||||
stream not in _per_stream_tensors
|
||||
or _per_stream_tensors[stream].numel() < size
|
||||
):
|
||||
_per_stream_tensors[stream] = torch.empty(
|
||||
size, device=device, dtype=torch.int8
|
||||
)
|
||||
_per_stream_tensors[stream].__hibernate__ = {"type": "ignore"}
|
||||
return _per_stream_tensors[stream]
|
||||
|
||||
_per_device_alloc_fns[device] = alloc_fn
|
||||
return _per_device_alloc_fns[device]
|
||||
|
||||
|
||||
def log_kernel_info(
|
||||
compiled_kernel: triton.compiler.CompiledKernel, best_config: triton.Config = None
|
||||
):
|
||||
kernel_name = compiled_kernel.name
|
||||
nregs = compiled_kernel.n_regs
|
||||
nspills = compiled_kernel.n_spills
|
||||
metadata = compiled_kernel.metadata
|
||||
logger.debug(
|
||||
f"{kernel_name}: n_regs={nregs} n_spills={nspills} metadata={metadata}"
|
||||
)
|
||||
if best_config is not None:
|
||||
logger.debug(f"{kernel_name} autotuned best_config: {best_config}")
|
||||
|
||||
|
||||
def grouped_gemm_forward(
|
||||
X: torch.Tensor,
|
||||
W: torch.Tensor,
|
||||
topk: int,
|
||||
m_sizes: torch.Tensor,
|
||||
gather_indices: torch.Tensor = None,
|
||||
topk_weights: torch.Tensor = None,
|
||||
# Fusions
|
||||
permute_x: bool = False,
|
||||
permute_y: bool = False,
|
||||
fuse_mul_post: bool = False,
|
||||
# Autotuning - manual kernel params will be ignored if autotune is True
|
||||
autotune: bool = False,
|
||||
# Kernel tuning params if not autotuning -- NOTE: these params need to be tuned, otherwise performance will be poor
|
||||
BLOCK_SIZE_M: int = 32,
|
||||
BLOCK_SIZE_N: int = 32,
|
||||
BLOCK_SIZE_K: int = 32,
|
||||
num_warps: int = 4,
|
||||
num_stages: int = 2,
|
||||
use_tma_load_w: bool = False,
|
||||
use_tma_load_x: bool = False,
|
||||
use_tma_store: bool = False,
|
||||
# software pipelining -- set to True for now, won't impact until loop is re-written
|
||||
flatten: bool = True,
|
||||
# debugging
|
||||
debug: bool = False,
|
||||
) -> torch.Tensor:
|
||||
"""
|
||||
Grouped GEMM forward pass for MoE MLPs.
|
||||
|
||||
The implementation offers a number of fusions specific to MoE:
|
||||
- `permute_x`: fuse the permutation of hidden states from token order (original order) to grouped expert order, typically only needed for the first grouped GEMM in an MoE MLP.
|
||||
- When `permute_x` is True, `X` is expected to be of shape (num_tokens, K).
|
||||
- When `permute_x` is False, `X` is expected to be of shape (total_tokens, K) where `total_tokens = num_tokens * topk` AND already permuted to grouped expert order, i.e., hidden states are sorted such that tokens assigned to each expert are contiguous.
|
||||
- `permute_y`: fused the permuation of the output from expert grouped order back to original token order, typically only needed for the second grouped GEMM in an MoE MLP.
|
||||
- `fuse_mul_pre`: fuse the multiplication of the routed input with topk_weights, only done in the first grouped GEMM in an MoE MLP as for Llama4. Do not use, since results in performance regression as it interrupts the GEMM mainloop.
|
||||
- `fuse_mul_post`: fuse the multiplication of the routed output with topk_weights, used only when `permute_y` is True. NOTE: this should only be used when using this kernel for inference, not for training.
|
||||
|
||||
X: (M, K) hidden states where M is the num_tokens if `permute_x` is True, otherwise `total_tokens` where `total_tokens = num_tokens * topk`.
|
||||
W: (E, N, K) expert weights, where E is number of experts, N in the intermediate (output) dim, and K is the reduction dim
|
||||
m_sizes: tokens assigned to each expert which correspond to the size of M in the respective GEMMs in the grouped GEMM.
|
||||
gather_indices: (total_tokens,) indices of tokens assigned to each expert. E.g., slicing gather_indices by cumsum of m_sizes gives the indices of tokens assigned to each expert.
|
||||
topk_weights: (total_tokens,) weights to multiply routed output by in expert MLP calculation, used only when `fuse_mul_post` is True (see note on `fuse_mul_post`).
|
||||
use_fast_accum: currently unused; trade off faster accumulation dtype in GEMM for less precision.
|
||||
use_tma_load_x: use TMA for loading activations, incompatible with permute_x. TODO: add TMA gather / scatter support for Blackwell+.
|
||||
use_tma_load_w: use TMA for loading weights. If TMA supported, this should always be enabled as it is faster than global memory load.
|
||||
use_tma_store: use TMA for storing output, incompatible with permute_y. TODO: add TMA scatter support for Blackwell+.
|
||||
|
||||
Returns:
|
||||
y: (total_tokens, N) output of grouped GEMM
|
||||
"""
|
||||
|
||||
assert X.device.type == "cuda", "X and W must be on CUDA"
|
||||
assert m_sizes.device.type == "cuda", "m_sizes must be on CUDA"
|
||||
|
||||
X = X.contiguous()
|
||||
W = W.contiguous()
|
||||
m_sizes = m_sizes.contiguous()
|
||||
|
||||
# Preconditions
|
||||
assert not (permute_x and permute_y), "Cannot permute both X and Y"
|
||||
assert not (permute_y and use_tma_store), "Cannot use both TMA store and permute_y"
|
||||
|
||||
if use_tma_load_x:
|
||||
# TMA load for activations, TMA gather only supported on Blackwell+
|
||||
assert not permute_x, "Cannot use both use_tma_load_x and permute_x"
|
||||
|
||||
use_tma = use_tma_load_w or use_tma_load_x or use_tma_store
|
||||
if not supports_tma() and use_tma:
|
||||
warnings.warn("TMA not supported, tma_load will be set to False")
|
||||
use_tma_load_w = False
|
||||
use_tma_load_x = False
|
||||
use_tma_store = False
|
||||
|
||||
if use_tma or autotune:
|
||||
|
||||
def alloc_fn(size: int, alignment: int, stream: int):
|
||||
return torch.empty(size, device="cuda", dtype=torch.int8)
|
||||
|
||||
triton.set_allocator(alloc_fn)
|
||||
|
||||
X = X.view(-1, X.shape[-1])
|
||||
W = W.view(-1, W.shape[-1])
|
||||
|
||||
if permute_x or permute_y:
|
||||
assert gather_indices is not None, (
|
||||
"gather_indices must be provided when permute_x or permute_y is True"
|
||||
)
|
||||
assert gather_indices.is_contiguous()
|
||||
assert gather_indices.device.type == "cuda"
|
||||
assert gather_indices.ndim == 1
|
||||
total_tokens = gather_indices.shape[0]
|
||||
num_tokens = total_tokens // topk
|
||||
if permute_x:
|
||||
assert X.shape[0] == num_tokens, (
|
||||
f"X.shape[0] ({X.shape[0]}) must match num_tokens ({num_tokens})"
|
||||
)
|
||||
else:
|
||||
assert X.shape[0] == total_tokens, (
|
||||
f"X.shape[0] ({X.shape[0]}) must match total_tokens ({total_tokens})"
|
||||
)
|
||||
else:
|
||||
total_tokens = X.shape[0]
|
||||
num_tokens = total_tokens // topk
|
||||
|
||||
num_experts = m_sizes.shape[0]
|
||||
_, K = X.shape
|
||||
N = W.shape[0] // num_experts
|
||||
assert K == W.shape[1], f"K ({K}) must match W.shape[1] ({W.shape[1]})"
|
||||
|
||||
if fuse_mul_post:
|
||||
global _FUSED_MUL_WARN
|
||||
if not _FUSED_MUL_WARN:
|
||||
warnings.warn(
|
||||
"fused_mul should only be used for inference, not for training"
|
||||
)
|
||||
_FUSED_MUL_WARN = True
|
||||
assert permute_y, "FUSE_MUL requires PERMUTE_Y"
|
||||
assert topk_weights is not None
|
||||
assert topk_weights.numel() == total_tokens
|
||||
assert topk_weights.device.type == "cuda"
|
||||
assert topk_weights.is_contiguous()
|
||||
topk_weights = topk_weights.view(-1)
|
||||
if debug:
|
||||
print(
|
||||
f"DEBUG::GROUPED_GEMM {topk_weights.tolist()} {gather_indices.tolist()}"
|
||||
)
|
||||
|
||||
y = torch.empty((total_tokens, N), device=X.device, dtype=X.dtype)
|
||||
if total_tokens == 0 or N == 0:
|
||||
return y
|
||||
|
||||
NUM_SMS = torch.cuda.get_device_properties("cuda").multi_processor_count
|
||||
|
||||
def grid(META):
|
||||
return (NUM_SMS,)
|
||||
|
||||
if not autotune:
|
||||
BLOCK_SIZE_K = min(K, BLOCK_SIZE_K)
|
||||
BLOCK_SIZE_N = min(N, BLOCK_SIZE_N)
|
||||
BLOCK_SIZE_M = min(total_tokens, BLOCK_SIZE_M)
|
||||
|
||||
if debug:
|
||||
print(
|
||||
f"DEBUG::GROUPED_GEMM {num_tokens=} {topk=} {num_experts=} {N=} {K=} {BLOCK_SIZE_M=} {BLOCK_SIZE_N=} {BLOCK_SIZE_K=} {permute_x=}"
|
||||
)
|
||||
print(
|
||||
f"DEBUG::GROUPED_GEMM {m_sizes.tolist()} {(gather_indices // topk).tolist()}"
|
||||
)
|
||||
|
||||
kernel_args = {
|
||||
# Inputs
|
||||
"x_ptr": X,
|
||||
"w_ptr": W,
|
||||
"m_sizes_ptr": m_sizes,
|
||||
"gather_indices_ptr": gather_indices,
|
||||
"topk_weights_ptr": topk_weights,
|
||||
# Output
|
||||
"y_ptr": y,
|
||||
# Problem shapes
|
||||
"NUM_TOKENS": num_tokens,
|
||||
"NUM_EXPERTS": num_experts,
|
||||
"TOPK": topk,
|
||||
"N": N,
|
||||
"K": K,
|
||||
"NUM_SMS": NUM_SMS,
|
||||
# Gather / Scatter
|
||||
"PERMUTE_X": permute_x,
|
||||
"PERMUTE_Y": permute_y,
|
||||
# TopK weight merging
|
||||
"FUSE_MUL_POST": fuse_mul_post,
|
||||
# Loop pipelining
|
||||
"FLATTEN": flatten,
|
||||
}
|
||||
if not autotune:
|
||||
kernel_args.update(
|
||||
{
|
||||
"USE_TMA_LOAD_W": use_tma_load_w,
|
||||
"USE_TMA_LOAD_X": use_tma_load_x,
|
||||
"USE_TMA_STORE": use_tma_store,
|
||||
"BLOCK_SIZE_M": BLOCK_SIZE_M,
|
||||
"BLOCK_SIZE_N": BLOCK_SIZE_N,
|
||||
"BLOCK_SIZE_K": BLOCK_SIZE_K,
|
||||
"num_warps": num_warps,
|
||||
"num_stages": num_stages,
|
||||
}
|
||||
)
|
||||
|
||||
kernel = (
|
||||
_autotuned_grouped_gemm_forward_kernel
|
||||
if autotune
|
||||
else _grouped_gemm_forward_kernel
|
||||
)
|
||||
compiled_kernel: triton.compiler.CompiledKernel = kernel[grid](**kernel_args)
|
||||
|
||||
if autotune:
|
||||
log_kernel_info(compiled_kernel, kernel.best_config)
|
||||
else:
|
||||
log_kernel_info(compiled_kernel)
|
||||
|
||||
return y
|
||||
|
||||
|
||||
def grouped_gemm_dX(
|
||||
dY: torch.Tensor,
|
||||
W: torch.Tensor,
|
||||
gather_indices: torch.Tensor,
|
||||
m_sizes: torch.Tensor,
|
||||
topk: int,
|
||||
BLOCK_SIZE_M: int = 32,
|
||||
BLOCK_SIZE_N: int = 32,
|
||||
BLOCK_SIZE_K: int = 32,
|
||||
debug: bool = False,
|
||||
permute_x: bool = False,
|
||||
permute_y: bool = False,
|
||||
use_tma_load_w: bool = False,
|
||||
use_tma_load_dy: bool = False,
|
||||
use_tma_store: bool = False,
|
||||
num_warps: int = 4,
|
||||
num_stages: int = 2,
|
||||
flatten: bool = True,
|
||||
fuse_mul_pre: bool = False,
|
||||
fuse_mul_post: bool = False,
|
||||
autotune: bool = False,
|
||||
) -> torch.Tensor:
|
||||
"""
|
||||
dX backward kernel
|
||||
grad_output: (M, N)
|
||||
gather_indices: (total_tokens,), indices of tokens assigned to each expert. E.g., slicing gather_indices by cumsum of m_sizes gives the indices of tokens assigned to each expert.
|
||||
m_sizes: tokens assigned to each expert which correspond to the size of M in the respective GEMMs in the grouped GEMM.
|
||||
topk: number of experts chosen per token.
|
||||
`permute_x`: whether X was permuted on load in the forward pass, typically only used for the first grouped GEMM in an MoE MLP to group tokens by expert.
|
||||
- In the forward pass, if we permuted X on load, we need to permute store in the backward pass
|
||||
- Shapes
|
||||
- the forward pass input X shape is [NUM_TOKENS, K], reduce across K, output y is [NUM_TOKENS * TOPK, K]
|
||||
- the backward pass input dy shape is [NUM_TOKENS * TOPK, N], reduce across N, output dX is [NUM_TOKENS * TOPK, K]
|
||||
- Note that in the backward pass, the output size is still [NUM_TOKENS * TOPK, K] since we still need to accumulate gradients for each expert chosen by the token in a post-processing step.
|
||||
`permute_y`: whether the output was permuted on store in the forward pass, typically only used for the second grouped GEMM in an MoE MLP to restore to the original token order.
|
||||
- In the forward pass, if we permuted output on store (e.g., in the second grouped GEMM in fused MoE MLP), we need to permute on load to get from token order to expert grouped order
|
||||
- We still store in contiguous order since we are writing out dX which will be the input to the backwards pass of the first grouped GEMM
|
||||
`fuse_mul_{pre,post}`: always set to False since this should only be used for inference.
|
||||
use_tma_load_dy: use TMA for loading dy. use_tma_load_dy is incompatible with permute_y. TODO: add TMA gather / scatter support for Blackwell+ which will enable permute_y and use_tma_load_dy.
|
||||
use_tma_load_w: use TMA for loading weights. If TMA supported, this should always be enabled as it is faster than global memory load.
|
||||
use_tma_store: use TMA for storing dX. Incompatible with permute_x. TODO: add TMA gather / scatter support for Blackwell+ which will enable permute_x and use_tma_store.
|
||||
"""
|
||||
assert not fuse_mul_pre, (
|
||||
"fuse_mul_pre should only be used for inference, not for training"
|
||||
)
|
||||
assert not fuse_mul_post, (
|
||||
"fuse_mul_post should only be used for inference, not for training"
|
||||
)
|
||||
assert dY.is_contiguous()
|
||||
assert W.is_contiguous()
|
||||
assert m_sizes.is_contiguous()
|
||||
assert m_sizes.ndim == 1
|
||||
|
||||
# Preconditions
|
||||
assert not (permute_x and permute_y), "Cannot permute both X and Y"
|
||||
# Note that this is flipped from the forward pass
|
||||
# If we permuted y in the forward, we need to permute on load in the backward
|
||||
assert not (permute_y and use_tma_load_dy), "Cannot use both TMA load and permute_y"
|
||||
assert not (permute_x and use_tma_store), "Cannot use both TMA store and permute_x"
|
||||
|
||||
use_tma = use_tma_load_dy or use_tma_load_w or use_tma_store
|
||||
if not supports_tma() and use_tma:
|
||||
warnings.warn("TMA not supported, tma_load will be set to False")
|
||||
use_tma_load_w = False
|
||||
use_tma_load_dy = False
|
||||
use_tma_store = False
|
||||
|
||||
if use_tma or autotune:
|
||||
|
||||
def alloc_fn(size: int, alignment: int, stream: int):
|
||||
# print(f"DEBUG::GROUPED_GEMM alloc_fn {size=} {alignment=} {stream=}")
|
||||
return torch.empty(size, device="cuda", dtype=torch.int8)
|
||||
|
||||
triton.set_allocator(alloc_fn)
|
||||
|
||||
num_experts = m_sizes.shape[0]
|
||||
dY = dY.view(-1, dY.shape[-1])
|
||||
W = W.view(-1, W.shape[-1])
|
||||
|
||||
M_total, N_grad = dY.shape
|
||||
N_total, K = W.shape
|
||||
N = N_total // num_experts
|
||||
assert N_grad == N, f"Grad_output N ({N_grad}) must match weight N ({N})"
|
||||
|
||||
assert M_total % topk == 0, (
|
||||
f"M_total ({M_total}) must be divisible by topk ({topk})"
|
||||
)
|
||||
num_tokens = M_total // topk
|
||||
|
||||
total_tokens = gather_indices.shape[0]
|
||||
assert total_tokens == M_total, (
|
||||
f"Total tokens ({total_tokens}) must match M_total ({M_total})"
|
||||
)
|
||||
|
||||
# Note that the output shape is [NUM_TOKENS * TOPK, K] even when `permute_x` is True since we need to accumulate gradients across all experts chosen by the token.
|
||||
# This will be done in a post-processing step reduction step.
|
||||
output_shape = (total_tokens, K)
|
||||
dX = torch.zeros(output_shape, device=dY.device, dtype=dY.dtype)
|
||||
|
||||
NUM_SMS = torch.cuda.get_device_properties(
|
||||
"cuda"
|
||||
).multi_processor_count # if not debug else 1
|
||||
|
||||
def grid(META):
|
||||
return (NUM_SMS,)
|
||||
|
||||
if not autotune:
|
||||
BLOCK_SIZE_M = min(M_total, BLOCK_SIZE_M)
|
||||
BLOCK_SIZE_N = min(N_grad, BLOCK_SIZE_N)
|
||||
BLOCK_SIZE_K = min(K, BLOCK_SIZE_K)
|
||||
|
||||
if debug:
|
||||
print(
|
||||
f"DEBUG::GROUPED_GEMM {num_tokens=} {topk=} {output_shape=} {num_experts=} {N=} {K=} {BLOCK_SIZE_M=} {BLOCK_SIZE_N=} {BLOCK_SIZE_K=} {NUM_SMS=}"
|
||||
)
|
||||
print(f"DEBUG::GROUPED_GEMM {m_sizes.tolist()}")
|
||||
|
||||
kernel_args = {
|
||||
# Inputs
|
||||
"dY_ptr": dY,
|
||||
"w_ptr": W,
|
||||
"gather_indices_ptr": gather_indices,
|
||||
"m_sizes_ptr": m_sizes,
|
||||
# Output
|
||||
"dX_ptr": dX,
|
||||
# Problem sizes
|
||||
"NUM_EXPERTS": num_experts,
|
||||
"NUM_TOKENS": num_tokens,
|
||||
"TOPK": topk,
|
||||
"N": N,
|
||||
"K": K,
|
||||
"NUM_SMS": NUM_SMS,
|
||||
# Gather / Scatter
|
||||
"PERMUTE_X": permute_x,
|
||||
"PERMUTE_Y": permute_y,
|
||||
"FLATTEN": flatten,
|
||||
}
|
||||
if not autotune:
|
||||
kernel_args.update(
|
||||
{
|
||||
"BLOCK_SIZE_M": BLOCK_SIZE_M,
|
||||
"BLOCK_SIZE_N": BLOCK_SIZE_N,
|
||||
"BLOCK_SIZE_K": BLOCK_SIZE_K,
|
||||
"num_warps": num_warps,
|
||||
"num_stages": num_stages,
|
||||
"USE_TMA_LOAD_dY": use_tma_load_dy,
|
||||
"USE_TMA_LOAD_W": use_tma_load_w,
|
||||
"USE_TMA_STORE": use_tma_store,
|
||||
}
|
||||
)
|
||||
kernel = _autotuned_grouped_gemm_dX_kernel if autotune else _grouped_gemm_dX_kernel
|
||||
compiled_kernel: triton.compiler.CompiledKernel = kernel[grid](**kernel_args)
|
||||
|
||||
if autotune:
|
||||
log_kernel_info(compiled_kernel, kernel.best_config)
|
||||
else:
|
||||
log_kernel_info(compiled_kernel)
|
||||
return dX
|
||||
|
||||
|
||||
def grouped_gemm_dW(
|
||||
X: torch.Tensor,
|
||||
dY: torch.Tensor,
|
||||
m_sizes: torch.Tensor,
|
||||
gather_indices: torch.Tensor,
|
||||
topk: int,
|
||||
BLOCK_SIZE_M: int = 32,
|
||||
BLOCK_SIZE_N: int = 32,
|
||||
BLOCK_SIZE_K: int = 32,
|
||||
permute_x: bool = False,
|
||||
permute_y: bool = False,
|
||||
use_tma_load_dy: bool = False,
|
||||
use_tma_load_x: bool = False,
|
||||
use_tma_store: bool = False,
|
||||
fuse_mul_pre: bool = False,
|
||||
fuse_mul_post: bool = False,
|
||||
num_warps: int = 4,
|
||||
num_stages: int = 2,
|
||||
flatten: bool = True,
|
||||
autotune: bool = False,
|
||||
debug: bool = False,
|
||||
) -> torch.Tensor:
|
||||
"""
|
||||
X: (M, K) hidden states where M is the num_tokens if `permute_x` is True, otherwise `total_tokens` where `total_tokens = num_tokens * topk`.
|
||||
dY: (M, N)
|
||||
topk: number of experts to choose per token.
|
||||
m_sizes: tokens assigned to each expert which correspond to the size of M in the respective GEMMs in the grouped GEMM.
|
||||
gather_indices: (total_tokens,) indices of tokens assigned to each expert. E.g., slicing gather_indices by cumsum of m_sizes gives the indices of tokens assigned to each expert.
|
||||
permute_x: whether X was permuted on load in the forward pass, typically only used for the first grouped GEMM in an MoE MLP to group tokens by expert.
|
||||
- for the first grouped GEMM, we permuted on load -> X was [num_tokens, K] and stored y in expert grouped order [num_tokens * topk, K]
|
||||
- in the backwards pass, we need to permute on load of X while loading dy in contiguous (expert grouped) order
|
||||
- since we are writing out dW, there is no need to permute on store
|
||||
permute_y: whether the output was permuted on store in the forward pass, typically only used for the second grouped GEMM in an MoE MLP to restore to the original token order.
|
||||
- for the second grouped GEMM, we permuted on store -> y was permuted from expert grouped order to token order while X was loaded in expert grouped order since it was the output of the first grouped GEMM
|
||||
- in the backwards pass, we need to permute on load of dy to get from token order to expert grouped order to match the order of X
|
||||
- since we are writing out dW, there is no need to permute on store
|
||||
use_tma_load_dy: use TMA for loading dy. use_tma_load_dy is incompatible with permute_y. TODO: add TMA gather / scatter support for Blackwell+ which will enable permute_y and use_tma_load_dy.
|
||||
use_tma_load_x: use TMA for loading x. use_tma_load_x is incompatible with permute_x. TODO: add TMA gather / scatter support for Blackwell+ which will enable permute_x and use_tma_load_x.
|
||||
use_tma_store: use TMA for storing dW. If TMA supported, this should always be enabled as it is faster than global memory store.
|
||||
"""
|
||||
assert not fuse_mul_pre, "fuse_mul_pre not supported"
|
||||
assert not fuse_mul_post, "fuse_mul_post not supported"
|
||||
NUM_SMS = (
|
||||
torch.cuda.get_device_properties("cuda").multi_processor_count
|
||||
if not debug
|
||||
else 1
|
||||
)
|
||||
X = X.view(-1, X.shape[-1]).contiguous()
|
||||
dY = dY.contiguous()
|
||||
m_sizes = m_sizes.contiguous()
|
||||
|
||||
# Preconditions
|
||||
assert not (permute_x and permute_y), "Cannot permute both X and Y"
|
||||
assert not (permute_y and use_tma_load_dy), "Cannot use both TMA load and permute_y"
|
||||
assert not (permute_x and use_tma_load_x), "Cannot use both TMA load and permute_x"
|
||||
|
||||
use_tma = use_tma_load_dy or use_tma_load_x or use_tma_store
|
||||
if not supports_tma() and use_tma:
|
||||
warnings.warn("TMA not supported, tma_load will be set to False")
|
||||
use_tma_load_x = False
|
||||
use_tma_load_dy = False
|
||||
use_tma_store = False
|
||||
|
||||
if use_tma or autotune:
|
||||
|
||||
def alloc_fn(size: int, alignment: int, stream: int):
|
||||
return torch.empty(size, device="cuda", dtype=torch.int8)
|
||||
|
||||
triton.set_allocator(alloc_fn)
|
||||
|
||||
if permute_x or permute_y:
|
||||
assert gather_indices is not None
|
||||
assert gather_indices.is_contiguous()
|
||||
assert gather_indices.device.type == "cuda"
|
||||
assert gather_indices.ndim == 1
|
||||
total_tokens = gather_indices.shape[0]
|
||||
num_tokens = total_tokens // topk
|
||||
if permute_x:
|
||||
assert X.shape[0] == num_tokens
|
||||
else:
|
||||
assert X.shape[0] == total_tokens
|
||||
else:
|
||||
total_tokens = X.shape[0]
|
||||
num_tokens = total_tokens // topk
|
||||
|
||||
num_experts = m_sizes.shape[0]
|
||||
# Get dimensions
|
||||
_, K = X.shape
|
||||
M_grad, N = dY.shape
|
||||
|
||||
assert M_grad == total_tokens, f"dY M ({M_grad}) != total_tokens ({total_tokens})"
|
||||
|
||||
dW = torch.zeros((num_experts, N, K), device=X.device, dtype=X.dtype)
|
||||
|
||||
if not autotune:
|
||||
BLOCK_SIZE_M = min(total_tokens, BLOCK_SIZE_M)
|
||||
BLOCK_SIZE_N = min(N, BLOCK_SIZE_N)
|
||||
BLOCK_SIZE_K = min(K, BLOCK_SIZE_K)
|
||||
|
||||
def grid(META):
|
||||
return (NUM_SMS,)
|
||||
|
||||
if debug:
|
||||
print(
|
||||
f"DEBUG::GROUPED_GEMM_DW_TMA {num_experts=} {N=} {K=} {BLOCK_SIZE_M=} {BLOCK_SIZE_N=} {BLOCK_SIZE_K=} {NUM_SMS=}"
|
||||
)
|
||||
|
||||
print(f"DEBUG::GROUPED_GEMM_DW_TMA {m_sizes.tolist()=}")
|
||||
print(f"DEBUG::GROUPED_GEMM_DW_TMA {gather_indices.tolist()=}")
|
||||
m_start = 0
|
||||
for i in range(num_experts):
|
||||
expert_token_idx = gather_indices[m_start : m_start + m_sizes[i]]
|
||||
t_start = 0
|
||||
while t_start < m_sizes[i]:
|
||||
token_idx = expert_token_idx[t_start : t_start + BLOCK_SIZE_M]
|
||||
if permute_x:
|
||||
token_idx = token_idx // topk
|
||||
print(
|
||||
f"DEBUG::GROUPED_GEMM_DW_TMA Token expert {i} indices: {token_idx.tolist()}"
|
||||
)
|
||||
t_start += BLOCK_SIZE_M
|
||||
|
||||
m_start += m_sizes[i]
|
||||
|
||||
kernel_args = {
|
||||
# Inputs
|
||||
"x_ptr": X,
|
||||
"dY_ptr": dY,
|
||||
"m_sizes_ptr": m_sizes,
|
||||
"gather_indices_ptr": gather_indices,
|
||||
# Output
|
||||
"dW_ptr": dW,
|
||||
# Problem sizes
|
||||
"NUM_TOKENS": num_tokens,
|
||||
"TOPK": topk,
|
||||
"NUM_EXPERTS": num_experts,
|
||||
"N": N,
|
||||
"K": K,
|
||||
"NUM_SMS": NUM_SMS,
|
||||
# Gather / Scatter
|
||||
"PERMUTE_X": permute_x,
|
||||
"PERMUTE_Y": permute_y,
|
||||
# Loop pipelining
|
||||
"FLATTEN": flatten,
|
||||
}
|
||||
|
||||
if not autotune:
|
||||
kernel_args.update(
|
||||
{
|
||||
"BLOCK_SIZE_M": BLOCK_SIZE_M,
|
||||
"BLOCK_SIZE_N": BLOCK_SIZE_N,
|
||||
"BLOCK_SIZE_K": BLOCK_SIZE_K,
|
||||
"USE_TMA_LOAD_dY": use_tma_load_dy,
|
||||
"USE_TMA_LOAD_X": use_tma_load_x,
|
||||
"USE_TMA_STORE": use_tma_store,
|
||||
"num_warps": num_warps,
|
||||
"num_stages": num_stages,
|
||||
}
|
||||
)
|
||||
|
||||
kernel = _autotuned_grouped_gemm_dW_kernel if autotune else _grouped_gemm_dW_kernel
|
||||
compiled_kernel: triton.compiler.CompiledKernel = kernel[grid](**kernel_args)
|
||||
|
||||
if autotune:
|
||||
log_kernel_info(compiled_kernel, kernel.best_config)
|
||||
else:
|
||||
log_kernel_info(compiled_kernel)
|
||||
|
||||
return dW
|
||||
|
||||
|
||||
class GroupedGemm(torch.autograd.Function):
|
||||
@staticmethod
|
||||
def forward(
|
||||
ctx,
|
||||
X,
|
||||
W,
|
||||
m_sizes,
|
||||
topk,
|
||||
gather_indices,
|
||||
permute_x,
|
||||
permute_y,
|
||||
topk_weights,
|
||||
fuse_mul_post,
|
||||
kernel_config_fwd,
|
||||
kernel_config_bwd_dX,
|
||||
kernel_config_bwd_dW,
|
||||
autotune,
|
||||
dX_only,
|
||||
dW_only,
|
||||
):
|
||||
ctx.topk = topk
|
||||
ctx.permute_x = permute_x
|
||||
ctx.permute_y = permute_y
|
||||
ctx.fuse_mul_post = fuse_mul_post
|
||||
ctx.kernel_config_fwd = kernel_config_fwd
|
||||
ctx.kernel_config_bwd_dX = kernel_config_bwd_dX
|
||||
ctx.kernel_config_bwd_dW = kernel_config_bwd_dW
|
||||
ctx.autotune = autotune
|
||||
ctx.dX_only = dX_only
|
||||
ctx.dW_only = dW_only
|
||||
|
||||
# NOTE: we don't save topk_weights for backward since we do not support training with fused_mul
|
||||
ctx.save_for_backward(X, W, m_sizes, gather_indices)
|
||||
|
||||
fwd_config = {}
|
||||
if kernel_config_fwd is not None:
|
||||
fwd_config["BLOCK_SIZE_M"] = kernel_config_fwd.BLOCK_SIZE_M
|
||||
fwd_config["BLOCK_SIZE_N"] = kernel_config_fwd.BLOCK_SIZE_N
|
||||
fwd_config["BLOCK_SIZE_K"] = kernel_config_fwd.BLOCK_SIZE_K
|
||||
fwd_config["num_warps"] = kernel_config_fwd.num_warps
|
||||
fwd_config["num_stages"] = kernel_config_fwd.num_stages
|
||||
fwd_config["use_tma_load_x"] = kernel_config_fwd.use_tma_load_x
|
||||
fwd_config["use_tma_load_w"] = kernel_config_fwd.use_tma_load_w
|
||||
fwd_config["use_tma_store"] = kernel_config_fwd.use_tma_store
|
||||
|
||||
return grouped_gemm_forward(
|
||||
X=X,
|
||||
W=W,
|
||||
topk=topk,
|
||||
m_sizes=m_sizes,
|
||||
gather_indices=gather_indices,
|
||||
topk_weights=topk_weights,
|
||||
permute_x=permute_x,
|
||||
permute_y=permute_y,
|
||||
fuse_mul_post=fuse_mul_post,
|
||||
# Autotune -- this will override the manual kernel config if true
|
||||
autotune=autotune,
|
||||
# Manual kernel config
|
||||
**fwd_config,
|
||||
)
|
||||
|
||||
@staticmethod
|
||||
def backward(ctx, dY):
|
||||
X, W, m_sizes, gather_indices = ctx.saved_tensors
|
||||
topk = ctx.topk
|
||||
permute_x = ctx.permute_x
|
||||
permute_y = ctx.permute_y
|
||||
fuse_mul_post = ctx.fuse_mul_post
|
||||
kernel_config_bwd_dX = ctx.kernel_config_bwd_dX
|
||||
kernel_config_bwd_dW = ctx.kernel_config_bwd_dW
|
||||
autotune = ctx.autotune
|
||||
dX_only = ctx.dX_only
|
||||
dW_only = ctx.dW_only
|
||||
|
||||
if not autotune:
|
||||
if not dW_only:
|
||||
assert kernel_config_bwd_dX is not None, (
|
||||
"kernel_config_bwd_dX must be provided if autotune is False"
|
||||
)
|
||||
if not dX_only:
|
||||
assert kernel_config_bwd_dW is not None, (
|
||||
"kernel_config_bwd_dW must be provided if autotune is False"
|
||||
)
|
||||
|
||||
assert not fuse_mul_post, (
|
||||
"fused_mul should only be used for inference, not for training"
|
||||
)
|
||||
|
||||
if not dX_only:
|
||||
bwd_dW_config = {}
|
||||
|
||||
if kernel_config_bwd_dW is not None:
|
||||
bwd_dW_config["use_tma_load_dy"] = kernel_config_bwd_dW.use_tma_load_dy
|
||||
bwd_dW_config["use_tma_load_x"] = kernel_config_bwd_dW.use_tma_load_x
|
||||
bwd_dW_config["use_tma_store"] = kernel_config_bwd_dW.use_tma_store
|
||||
bwd_dW_config["BLOCK_SIZE_M"] = kernel_config_bwd_dW.BLOCK_SIZE_M
|
||||
bwd_dW_config["BLOCK_SIZE_N"] = kernel_config_bwd_dW.BLOCK_SIZE_N
|
||||
bwd_dW_config["BLOCK_SIZE_K"] = kernel_config_bwd_dW.BLOCK_SIZE_K
|
||||
bwd_dW_config["num_warps"] = kernel_config_bwd_dW.num_warps
|
||||
bwd_dW_config["num_stages"] = kernel_config_bwd_dW.num_stages
|
||||
|
||||
dW = grouped_gemm_dW(
|
||||
X=X,
|
||||
dY=dY,
|
||||
m_sizes=m_sizes,
|
||||
gather_indices=gather_indices,
|
||||
topk=topk,
|
||||
permute_x=permute_x,
|
||||
permute_y=permute_y,
|
||||
# Autotune -- this will override the manual kernel config if true
|
||||
autotune=autotune,
|
||||
# Manual kernel config
|
||||
**bwd_dW_config,
|
||||
)
|
||||
else:
|
||||
dW = None
|
||||
|
||||
if not dW_only:
|
||||
bwd_dX_config = {}
|
||||
if kernel_config_bwd_dX is not None:
|
||||
bwd_dX_config["use_tma_load_dy"] = kernel_config_bwd_dX.use_tma_load_dy
|
||||
bwd_dX_config["use_tma_load_w"] = kernel_config_bwd_dX.use_tma_load_w
|
||||
bwd_dX_config["use_tma_store"] = kernel_config_bwd_dX.use_tma_store
|
||||
bwd_dX_config["BLOCK_SIZE_M"] = kernel_config_bwd_dX.BLOCK_SIZE_M
|
||||
bwd_dX_config["BLOCK_SIZE_N"] = kernel_config_bwd_dX.BLOCK_SIZE_N
|
||||
bwd_dX_config["BLOCK_SIZE_K"] = kernel_config_bwd_dX.BLOCK_SIZE_K
|
||||
bwd_dX_config["num_warps"] = kernel_config_bwd_dX.num_warps
|
||||
bwd_dX_config["num_stages"] = kernel_config_bwd_dX.num_stages
|
||||
|
||||
dX = grouped_gemm_dX(
|
||||
dY=dY,
|
||||
W=W,
|
||||
m_sizes=m_sizes,
|
||||
gather_indices=gather_indices,
|
||||
topk=topk,
|
||||
permute_x=permute_x,
|
||||
permute_y=permute_y,
|
||||
# Autotune -- this will override the manual kernel config if true
|
||||
autotune=autotune,
|
||||
# Manual kernel config
|
||||
**bwd_dX_config,
|
||||
)
|
||||
|
||||
if topk > 1 and permute_x:
|
||||
dX = dX.view(X.shape[0], topk, -1).sum(dim=1)
|
||||
else:
|
||||
dX = None
|
||||
|
||||
return (
|
||||
dX,
|
||||
dW,
|
||||
None, # m_sizes
|
||||
None, # gather_indices
|
||||
None, # topk
|
||||
None, # permute_x
|
||||
None, # permute_y
|
||||
None, # topk_weights
|
||||
None, # fuse_mul_post
|
||||
None, # kernel_config_fwd
|
||||
None, # kernel_config_bwd_dX
|
||||
None, # kernel_config_bwd_dW
|
||||
None, # autotune
|
||||
None, # dX_only
|
||||
None, # dW_only
|
||||
)
|
||||
|
||||
|
||||
def check_valid_config_fwd(
|
||||
permute_x,
|
||||
permute_y,
|
||||
use_tma_load_x,
|
||||
use_tma_load_w,
|
||||
use_tma_store,
|
||||
fuse_mul_post,
|
||||
is_first_gemm,
|
||||
):
|
||||
"""
|
||||
Check if the configuration is valid for the forward pass.
|
||||
"""
|
||||
is_second_gemm = not is_first_gemm
|
||||
|
||||
assert not (permute_x and permute_y), "Cannot permute both X and Y"
|
||||
assert not (is_second_gemm and permute_x), (
|
||||
"Cannot permute X for the second grouped GEMM"
|
||||
)
|
||||
assert not (is_first_gemm and permute_y), (
|
||||
"Cannot permute Y for the first grouped GEMM"
|
||||
)
|
||||
assert not (fuse_mul_post and is_first_gemm), (
|
||||
"Cannot fuse mul for the first grouped GEMM"
|
||||
)
|
||||
assert not (use_tma_load_x and permute_x), (
|
||||
"Cannot use TMA load and permute X unless on sm100+ (Blackwell+)"
|
||||
)
|
||||
assert not (use_tma_store and permute_y and is_second_gemm), (
|
||||
"Cannot use TMA store and permute Y for the second grouped GEMM unless on sm100+ (Blackwell+)"
|
||||
)
|
||||
|
||||
|
||||
def check_valid_config_bwd_dW(
|
||||
permute_x,
|
||||
permute_y,
|
||||
use_tma_load_dY,
|
||||
use_tma_load_x,
|
||||
use_tma_store,
|
||||
fuse_mul_post,
|
||||
is_first_gemm,
|
||||
):
|
||||
"""
|
||||
Check if the configuration is valid for the backward pass of dW.
|
||||
"""
|
||||
is_second_gemm = not is_first_gemm
|
||||
if fuse_mul_post:
|
||||
assert False, "Cannot fuse_mul is not supported for backward pass"
|
||||
if is_second_gemm and permute_y and use_tma_load_dY:
|
||||
assert False, "Cannot use TMA load and permute Y for the second grouped GEMM"
|
||||
if is_first_gemm and permute_x and use_tma_load_x:
|
||||
assert False, "Cannot use TMA load and permute X for the first grouped GEMM"
|
||||
|
||||
|
||||
def check_valid_config_bwd_dX(
|
||||
permute_x,
|
||||
permute_y,
|
||||
use_tma_load_dY,
|
||||
use_tma_load_w,
|
||||
use_tma_store,
|
||||
fuse_mul_post,
|
||||
is_first_gemm,
|
||||
):
|
||||
"""
|
||||
Check if the configuration is valid for the backward pass of dW.
|
||||
"""
|
||||
is_second_gemm = not is_first_gemm
|
||||
if fuse_mul_post:
|
||||
assert False, "Cannot fuse_mul is not supported for backward pass"
|
||||
if is_second_gemm and permute_y and use_tma_load_dY:
|
||||
assert False, "Cannot use TMA load and permute Y for the second grouped GEMM"
|
||||
if use_tma_store and permute_x and is_first_gemm:
|
||||
assert False, "Cannot use TMA store and permute X for the first grouped GEMM"
|
||||
|
||||
|
||||
def grouped_gemm(
|
||||
X: torch.Tensor,
|
||||
W: torch.Tensor,
|
||||
m_sizes: torch.Tensor,
|
||||
topk: int,
|
||||
gather_indices: torch.Tensor = None,
|
||||
permute_x: bool = False,
|
||||
permute_y: bool = False,
|
||||
topk_weights=None,
|
||||
fuse_mul_post=False,
|
||||
kernel_config_fwd: KernelConfigForward = None,
|
||||
kernel_config_bwd_dX: KernelConfigBackward_dX = None,
|
||||
kernel_config_bwd_dW: KernelConfigBackward_dW = None,
|
||||
autotune: bool = False,
|
||||
is_first_gemm: bool = True,
|
||||
# Only for debugging
|
||||
dX_only: bool = False,
|
||||
dW_only: bool = False,
|
||||
):
|
||||
"""
|
||||
Grouped GEMM for MoE MLPs.
|
||||
|
||||
The implementation offers a number of fusions specific to MoE:
|
||||
- `permute_x`: fuse the permutation of hidden states from token order (original order) to grouped expert order, typically only needed for the first grouped GEMM in an MoE MLP.
|
||||
- When `permute_x` is True, `X` is expected to be of shape (num_tokens, K).
|
||||
- When `permute_x` is False, `X` is expected to be of shape (total_tokens, K) where `total_tokens = num_tokens * topk` AND already permuted to grouped expert order, i.e., hidden states are sorted such that tokens assigned to each expert are contiguous.
|
||||
- `permute_y`: fused the permuation of the output from expert grouped order back to original token order, typically only needed for the second grouped GEMM in an MoE MLP.
|
||||
- `fuse_mul`: fuse the multiplication of the routed output with topk_weights, used only when `permute_y` is True. NOTE: this should only be used when using this kernel for inference, not for training.
|
||||
|
||||
X: (M, K) hidden states where M is the num_tokens if `permute_x` is True, otherwise `total_tokens` where `total_tokens = num_tokens * topk`.
|
||||
W: (E, N, K) expert weights, where E is number of experts, N in the intermediate (output) dim, and K is the reduction dim
|
||||
m_sizes: tokens assigned to each expert which correspond to the size of M in the respective GEMMs in the grouped GEMM.
|
||||
gather_indices: (total_tokens,) indices of tokens assigned to each expert. E.g., slicing gather_indices by cumsum of m_sizes gives the indices of tokens assigned to each expert. Needed when either `permute_x` or `permute_y` is True.
|
||||
topk_weights: (total_tokens,) weights to multiply routed output by in expert MLP calculation, used only when `fuse_mul` is True (see note on `fuse_mul`).
|
||||
kernel_config_fwd: KernelConfigForward for forward pass.
|
||||
kernel_config_bwd_dX: KernelConfigBackward_dX for backward pass of dX.
|
||||
kernel_config_bwd_dW: KernelConfigBackward_dW for backward pass of dW.
|
||||
autotune: whether to autotune the kernel, if yes, kernel_config_fwd, kernel_config_bwd_dX, and kernel_config_bwd_dW will be ignored.
|
||||
is_first_gemm: whether this is the first grouped GEMM in an MoE MLP. This is needed to check whether kernel configs are valid. `permute_x` should only be used for first gemm; `permute_y` should only be used for second gemm.
|
||||
This will impact whether TMA can be used for loading and storing.
|
||||
|
||||
"""
|
||||
if not autotune:
|
||||
assert kernel_config_fwd is not None, (
|
||||
"kernel_config_fwd must be provided if autotune is False"
|
||||
)
|
||||
|
||||
check_valid_config_fwd(
|
||||
permute_x,
|
||||
permute_y,
|
||||
use_tma_load_x=kernel_config_fwd.use_tma_load_x,
|
||||
use_tma_load_w=kernel_config_fwd.use_tma_load_w,
|
||||
use_tma_store=kernel_config_fwd.use_tma_store,
|
||||
fuse_mul_post=fuse_mul_post,
|
||||
is_first_gemm=is_first_gemm,
|
||||
)
|
||||
if kernel_config_bwd_dW is not None and not dX_only:
|
||||
check_valid_config_bwd_dW(
|
||||
permute_x,
|
||||
permute_y,
|
||||
use_tma_load_dY=kernel_config_bwd_dW.use_tma_load_dy,
|
||||
use_tma_load_x=kernel_config_bwd_dW.use_tma_load_x,
|
||||
use_tma_store=kernel_config_bwd_dW.use_tma_store,
|
||||
fuse_mul_post=fuse_mul_post,
|
||||
is_first_gemm=is_first_gemm,
|
||||
)
|
||||
if kernel_config_bwd_dX is not None and not dW_only:
|
||||
check_valid_config_bwd_dX(
|
||||
permute_x,
|
||||
permute_y,
|
||||
use_tma_load_dY=kernel_config_bwd_dX.use_tma_load_dy,
|
||||
use_tma_load_w=kernel_config_bwd_dX.use_tma_load_w,
|
||||
use_tma_store=kernel_config_bwd_dX.use_tma_store,
|
||||
fuse_mul_post=fuse_mul_post,
|
||||
is_first_gemm=is_first_gemm,
|
||||
)
|
||||
|
||||
if permute_x or permute_y:
|
||||
assert gather_indices is not None, (
|
||||
"gather_indices is required when either permute_x or permute_y is True"
|
||||
)
|
||||
|
||||
if fuse_mul_post:
|
||||
assert topk_weights is not None, (
|
||||
"topk_weights is required when fuse_mul_post is True"
|
||||
)
|
||||
|
||||
X = X.view(-1, X.shape[-1])
|
||||
m_sizes = m_sizes.view(-1)
|
||||
gather_indices = gather_indices.view(-1)
|
||||
|
||||
return GroupedGemm.apply(
|
||||
X,
|
||||
W,
|
||||
m_sizes,
|
||||
topk,
|
||||
gather_indices,
|
||||
permute_x,
|
||||
permute_y,
|
||||
topk_weights,
|
||||
fuse_mul_post,
|
||||
kernel_config_fwd,
|
||||
kernel_config_bwd_dX,
|
||||
kernel_config_bwd_dW,
|
||||
autotune,
|
||||
dX_only,
|
||||
dW_only,
|
||||
)
|
||||
0
unsloth/kernels/moe/grouped_gemm/kernels/__init__.py
Normal file
0
unsloth/kernels/moe/grouped_gemm/kernels/__init__.py
Normal file
393
unsloth/kernels/moe/grouped_gemm/kernels/autotuning.py
Normal file
393
unsloth/kernels/moe/grouped_gemm/kernels/autotuning.py
Normal file
|
|
@ -0,0 +1,393 @@
|
|||
"""
|
||||
Autotuning utils
|
||||
"""
|
||||
|
||||
import logging
|
||||
from itertools import product
|
||||
from typing import List
|
||||
|
||||
import torch
|
||||
import triton
|
||||
|
||||
logger = logging.getLogger(__name__)
|
||||
|
||||
DEFAULT_M_BLOCK_SIZES = [64, 128]
|
||||
DEFAULT_N_BLOCK_SIZES = [64, 128, 256]
|
||||
DEFAULT_K_BLOCK_SIZES = [64, 128, 256]
|
||||
DEFAULT_NUM_CTAS = 1
|
||||
DEFAULT_NUM_WARPS = [4, 8]
|
||||
DEFAULT_NUM_STAGES = [3, 4, 5]
|
||||
BOOLS = [True, False]
|
||||
|
||||
|
||||
def val_to_list(val):
|
||||
if val is None:
|
||||
return None
|
||||
elif isinstance(val, list):
|
||||
return val
|
||||
else:
|
||||
return [val]
|
||||
|
||||
|
||||
def convert_args_to_list(args):
|
||||
return [val_to_list(arg) for arg in args]
|
||||
|
||||
|
||||
def get_forward_configs(
|
||||
BLOCK_M=DEFAULT_M_BLOCK_SIZES,
|
||||
BLOCK_N=DEFAULT_N_BLOCK_SIZES,
|
||||
BLOCK_K=DEFAULT_K_BLOCK_SIZES,
|
||||
TMA_LOAD_X=True,
|
||||
TMA_LOAD_W=True,
|
||||
TMA_STORE=False, # NOTE: TMA_STORE is disabled for now
|
||||
num_warps=DEFAULT_NUM_WARPS,
|
||||
num_stages=DEFAULT_NUM_STAGES,
|
||||
num_ctas=DEFAULT_NUM_CTAS,
|
||||
):
|
||||
(
|
||||
BLOCK_M,
|
||||
BLOCK_N,
|
||||
BLOCK_K,
|
||||
TMA_LOAD_X,
|
||||
TMA_LOAD_W,
|
||||
TMA_STORE,
|
||||
num_warps,
|
||||
num_stages,
|
||||
num_ctas,
|
||||
) = convert_args_to_list(
|
||||
[
|
||||
BLOCK_M,
|
||||
BLOCK_N,
|
||||
BLOCK_K,
|
||||
TMA_LOAD_X,
|
||||
TMA_LOAD_W,
|
||||
TMA_STORE,
|
||||
num_warps,
|
||||
num_stages,
|
||||
num_ctas,
|
||||
]
|
||||
)
|
||||
kernel_configs = []
|
||||
for (
|
||||
block_m,
|
||||
block_n,
|
||||
block_k,
|
||||
w,
|
||||
s,
|
||||
tma_load_x,
|
||||
tma_load_w,
|
||||
tma_store,
|
||||
num_ctas,
|
||||
) in product(
|
||||
BLOCK_M,
|
||||
BLOCK_N,
|
||||
BLOCK_K,
|
||||
num_warps,
|
||||
num_stages,
|
||||
TMA_LOAD_X,
|
||||
TMA_LOAD_W,
|
||||
TMA_STORE,
|
||||
num_ctas,
|
||||
):
|
||||
kernel_configs.append(
|
||||
triton.Config(
|
||||
dict(
|
||||
BLOCK_SIZE_M=block_m,
|
||||
BLOCK_SIZE_N=block_n,
|
||||
BLOCK_SIZE_K=block_k,
|
||||
USE_TMA_LOAD_X=tma_load_x,
|
||||
USE_TMA_LOAD_W=tma_load_w,
|
||||
USE_TMA_STORE=tma_store,
|
||||
),
|
||||
num_warps=w,
|
||||
num_stages=s,
|
||||
num_ctas=num_ctas,
|
||||
)
|
||||
)
|
||||
|
||||
return kernel_configs
|
||||
|
||||
|
||||
def get_dX_kernel_configs(
|
||||
BLOCK_M=DEFAULT_M_BLOCK_SIZES,
|
||||
BLOCK_N=DEFAULT_N_BLOCK_SIZES,
|
||||
BLOCK_K=DEFAULT_K_BLOCK_SIZES,
|
||||
TMA_LOAD_dY=True,
|
||||
TMA_LOAD_W=True,
|
||||
TMA_STORE=False, # NOTE: TMA_STORE is disabled for now
|
||||
num_warps=DEFAULT_NUM_WARPS,
|
||||
num_stages=DEFAULT_NUM_STAGES,
|
||||
num_ctas=DEFAULT_NUM_CTAS,
|
||||
):
|
||||
(
|
||||
BLOCK_M,
|
||||
BLOCK_N,
|
||||
BLOCK_K,
|
||||
TMA_LOAD_dY,
|
||||
TMA_LOAD_W,
|
||||
TMA_STORE,
|
||||
num_warps,
|
||||
num_stages,
|
||||
num_ctas,
|
||||
) = convert_args_to_list(
|
||||
[
|
||||
BLOCK_M,
|
||||
BLOCK_N,
|
||||
BLOCK_K,
|
||||
TMA_LOAD_dY,
|
||||
TMA_LOAD_W,
|
||||
TMA_STORE,
|
||||
num_warps,
|
||||
num_stages,
|
||||
num_ctas,
|
||||
]
|
||||
)
|
||||
kernel_configs = []
|
||||
for (
|
||||
block_m,
|
||||
block_n,
|
||||
block_k,
|
||||
w,
|
||||
s,
|
||||
tma_load_dy,
|
||||
tma_load_w,
|
||||
tma_store,
|
||||
num_ctas,
|
||||
) in product(
|
||||
BLOCK_M,
|
||||
BLOCK_N,
|
||||
BLOCK_K,
|
||||
num_warps,
|
||||
num_stages,
|
||||
TMA_LOAD_dY,
|
||||
TMA_LOAD_W,
|
||||
TMA_STORE,
|
||||
num_ctas,
|
||||
):
|
||||
kernel_configs.append(
|
||||
triton.Config(
|
||||
dict(
|
||||
BLOCK_SIZE_M=block_m,
|
||||
BLOCK_SIZE_N=block_n,
|
||||
BLOCK_SIZE_K=block_k,
|
||||
USE_TMA_LOAD_dY=tma_load_dy,
|
||||
USE_TMA_LOAD_W=tma_load_w,
|
||||
USE_TMA_STORE=tma_store,
|
||||
),
|
||||
num_warps=w,
|
||||
num_stages=s,
|
||||
num_ctas=num_ctas,
|
||||
)
|
||||
)
|
||||
|
||||
return kernel_configs
|
||||
|
||||
|
||||
def get_dW_kernel_configs(
|
||||
BLOCK_M=DEFAULT_M_BLOCK_SIZES,
|
||||
BLOCK_N=DEFAULT_N_BLOCK_SIZES,
|
||||
BLOCK_K=DEFAULT_K_BLOCK_SIZES,
|
||||
num_warps=DEFAULT_NUM_WARPS,
|
||||
num_stages=DEFAULT_NUM_STAGES,
|
||||
num_ctas=DEFAULT_NUM_CTAS,
|
||||
TMA_LOAD_dY=True,
|
||||
TMA_LOAD_X=True,
|
||||
TMA_STORE=False,
|
||||
):
|
||||
(
|
||||
BLOCK_M,
|
||||
BLOCK_N,
|
||||
BLOCK_K,
|
||||
num_warps,
|
||||
num_stages,
|
||||
num_ctas,
|
||||
TMA_LOAD_dY,
|
||||
TMA_LOAD_X,
|
||||
TMA_STORE,
|
||||
) = convert_args_to_list(
|
||||
[
|
||||
BLOCK_M,
|
||||
BLOCK_N,
|
||||
BLOCK_K,
|
||||
num_warps,
|
||||
num_stages,
|
||||
num_ctas,
|
||||
TMA_LOAD_dY,
|
||||
TMA_LOAD_X,
|
||||
TMA_STORE,
|
||||
]
|
||||
)
|
||||
kernel_configs = []
|
||||
for (
|
||||
block_m,
|
||||
block_n,
|
||||
block_k,
|
||||
w,
|
||||
s,
|
||||
tma_load_dy,
|
||||
tma_load_x,
|
||||
tma_store,
|
||||
num_ctas,
|
||||
) in product(
|
||||
BLOCK_M,
|
||||
BLOCK_N,
|
||||
BLOCK_K,
|
||||
num_warps,
|
||||
num_stages,
|
||||
TMA_LOAD_dY,
|
||||
TMA_LOAD_X,
|
||||
TMA_STORE,
|
||||
num_ctas,
|
||||
):
|
||||
kernel_configs.append(
|
||||
triton.Config(
|
||||
dict(
|
||||
BLOCK_SIZE_M=block_m,
|
||||
BLOCK_SIZE_N=block_n,
|
||||
BLOCK_SIZE_K=block_k,
|
||||
USE_TMA_LOAD_dY=tma_load_dy,
|
||||
USE_TMA_LOAD_X=tma_load_x,
|
||||
USE_TMA_STORE=tma_store,
|
||||
),
|
||||
num_warps=w,
|
||||
num_stages=s,
|
||||
num_ctas=num_ctas,
|
||||
)
|
||||
)
|
||||
|
||||
return kernel_configs
|
||||
|
||||
|
||||
def estimate_smem_reqs(
|
||||
num_stages: int,
|
||||
BLOCK_SIZE_M: int,
|
||||
BLOCK_SIZE_N: int,
|
||||
BLOCK_SIZE_K: int,
|
||||
dtype: torch.dtype,
|
||||
):
|
||||
num_bytes = dtype.itemsize
|
||||
return (
|
||||
num_stages * BLOCK_SIZE_K * (BLOCK_SIZE_M + BLOCK_SIZE_N)
|
||||
+ BLOCK_SIZE_M * BLOCK_SIZE_N
|
||||
) * num_bytes
|
||||
|
||||
|
||||
def exceeds_smem_capacity(
|
||||
num_stages: int,
|
||||
BLOCK_SIZE_M: int,
|
||||
BLOCK_SIZE_N: int,
|
||||
BLOCK_SIZE_K: int,
|
||||
dtype: torch.dtype,
|
||||
smem_size: int,
|
||||
slack: float = 50000,
|
||||
):
|
||||
smem_reqs = estimate_smem_reqs(
|
||||
num_stages, BLOCK_SIZE_M, BLOCK_SIZE_N, BLOCK_SIZE_K, dtype
|
||||
)
|
||||
return smem_reqs > smem_size + slack
|
||||
|
||||
|
||||
def common_prune_criteria(config: triton.Config, kwargs: dict, dtype):
|
||||
from grouped_gemm.interface import supports_tma
|
||||
from grouped_gemm.kernels.tuning import get_device_properties
|
||||
|
||||
smem_size = get_device_properties().SIZE_SMEM
|
||||
|
||||
num_stages = config.num_stages
|
||||
BLOCK_SIZE_M = config.kwargs["BLOCK_SIZE_M"]
|
||||
BLOCK_SIZE_N = config.kwargs["BLOCK_SIZE_N"]
|
||||
BLOCK_SIZE_K = config.kwargs["BLOCK_SIZE_K"]
|
||||
|
||||
num_tokens = kwargs["NUM_TOKENS"]
|
||||
num_experts = kwargs["NUM_EXPERTS"]
|
||||
permute_x = kwargs["PERMUTE_X"]
|
||||
permute_y = kwargs["PERMUTE_Y"]
|
||||
tokens_per_expert = num_tokens // num_experts
|
||||
|
||||
# use_tma = [k for k in config.kwargs.keys() if k.startswith("USE_TMA_")]
|
||||
MIN_BLOCK_SIZE_M = DEFAULT_M_BLOCK_SIZES[0]
|
||||
if exceeds_smem_capacity(
|
||||
num_stages, BLOCK_SIZE_M, BLOCK_SIZE_N, BLOCK_SIZE_K, dtype, smem_size
|
||||
):
|
||||
return True
|
||||
if BLOCK_SIZE_M > tokens_per_expert * 2 and tokens_per_expert > MIN_BLOCK_SIZE_M:
|
||||
return True
|
||||
if permute_x and permute_y:
|
||||
return True
|
||||
# if not supports_tma() and any(use_tma):
|
||||
# return True
|
||||
return False
|
||||
|
||||
|
||||
def maybe_disable_tma(config: triton.Config):
|
||||
from grouped_gemm.interface import supports_tma
|
||||
|
||||
tma_keys = [k for k in config.kwargs.keys() if k.startswith("USE_TMA_")]
|
||||
if not supports_tma():
|
||||
logger.info("Disabling TMA")
|
||||
for k in tma_keys:
|
||||
config.kwargs[k] = False
|
||||
|
||||
|
||||
def prune_kernel_configs_fwd(configs: list[triton.Config], args, **kwargs):
|
||||
x = kwargs["x_ptr"]
|
||||
dtype = x.dtype
|
||||
|
||||
logger.debug(f"Pruning configs: {len(configs)}")
|
||||
|
||||
pruned_configs = []
|
||||
for config in configs:
|
||||
# disable TMA if gpu does not support it
|
||||
maybe_disable_tma(config)
|
||||
|
||||
if common_prune_criteria(config, kwargs, dtype):
|
||||
continue
|
||||
if config.kwargs["USE_TMA_LOAD_X"] and kwargs["PERMUTE_X"]:
|
||||
# Dynamically disable TMA_LOAD_X for permuted X
|
||||
config.kwargs["USE_TMA_LOAD_X"] = False
|
||||
if config.kwargs["USE_TMA_STORE"] and kwargs["PERMUTE_Y"]:
|
||||
continue
|
||||
|
||||
pruned_configs.append(config)
|
||||
|
||||
logger.debug(f"Pruned configs: {len(pruned_configs)}")
|
||||
return pruned_configs
|
||||
|
||||
|
||||
def prune_dX_configs(configs: List[triton.Config], args, **kwargs):
|
||||
dtype = kwargs["w_ptr"].dtype
|
||||
|
||||
logger.debug(f"Pruning configs: {len(configs)}")
|
||||
pruned_configs = []
|
||||
|
||||
for config in configs:
|
||||
if common_prune_criteria(config, kwargs, dtype):
|
||||
continue
|
||||
if config.kwargs["USE_TMA_LOAD_dY"] and kwargs["PERMUTE_Y"]:
|
||||
# dynamically disable TMA_LOAD_dY for permuted Y
|
||||
config.kwargs["USE_TMA_LOAD_dY"] = False
|
||||
if config.kwargs["USE_TMA_STORE"] and kwargs["PERMUTE_X"]:
|
||||
continue
|
||||
pruned_configs.append(config)
|
||||
|
||||
logger.debug(f"Pruned configs: {len(pruned_configs)}")
|
||||
return pruned_configs
|
||||
|
||||
|
||||
def prune_kernel_configs_backward_dW(configs: list[triton.Config], args, **kwargs):
|
||||
dtype = kwargs["x_ptr"].dtype
|
||||
|
||||
pruned_configs = []
|
||||
logger.debug(f"Pruning configs: {len(configs)}")
|
||||
|
||||
for config in configs:
|
||||
if common_prune_criteria(config, kwargs, dtype):
|
||||
continue
|
||||
if config.kwargs["USE_TMA_LOAD_dY"] and kwargs["PERMUTE_Y"]:
|
||||
config.kwargs["USE_TMA_LOAD_dY"] = False
|
||||
if config.kwargs["USE_TMA_LOAD_X"] and kwargs["PERMUTE_X"]:
|
||||
config.kwargs["USE_TMA_LOAD_X"] = False
|
||||
pruned_configs.append(config)
|
||||
|
||||
logger.debug(f"Pruned configs: {len(pruned_configs)}")
|
||||
return pruned_configs
|
||||
499
unsloth/kernels/moe/grouped_gemm/kernels/backward.py
Normal file
499
unsloth/kernels/moe/grouped_gemm/kernels/backward.py
Normal file
|
|
@ -0,0 +1,499 @@
|
|||
import torch
|
||||
import triton
|
||||
import triton.language as tl
|
||||
|
||||
from grouped_gemm.kernels.autotuning import (
|
||||
get_dW_kernel_configs,
|
||||
get_dX_kernel_configs,
|
||||
prune_dX_configs,
|
||||
prune_kernel_configs_backward_dW,
|
||||
)
|
||||
|
||||
"""
|
||||
dX backward kernel
|
||||
|
||||
- Shapes
|
||||
- the forward pass input X shape is [NUM_TOKENS, K] if permute_x else [NUM_TOKENS * TOPK, K]; output y is [NUM_TOKENS * TOPK, N]
|
||||
- the backward pass input dy shape is [NUM_TOKENS * TOPK, N], reduce across N, output dX is [NUM_TOKENS * TOPK, K]
|
||||
- Note that in the backward pass, the output size is still [NUM_TOKENS * TOPK, K] since we still need to accumulate gradients for each expert chosen by the token in a post-processing step.
|
||||
|
||||
`permute_x` notes:
|
||||
- In the forward pass, if we permute X on load, we need to permute on store in the backward pass to restore to original token order
|
||||
- the output dX with have shape [NUM_TOKENS * TOPK, K] and we need to perform an additional reduction across topk to accumulate gradients
|
||||
- This is done as a post-processing step in autograd.Function.
|
||||
- If not `permute_x`, this postprocessing step should take place outside autograd.Function such that the gradient shape matches the input X shape.
|
||||
|
||||
`permute_y` notes:
|
||||
- In the forward pass, if we permuted output on store (e.g., in the second grouped GEMM in fused MoE MLP), we need to permute on load to get from token order to expert grouped order
|
||||
- We still store in contiguous order since we are writing out dX which will be the input to the backwards pass of the first grouped GEMM
|
||||
|
||||
`fused_mul` notes:
|
||||
- In the forward pass, if we used the multiplication of topk weights (e.g., in the second grouped GEMM in fused MoE MLP), we need to make a few additional changes:
|
||||
1) We load topk_weights in natural (token) order. Since we only enable `fuse_mul` when permuting on store (`permute_y`), we multiply grad_output by topk_weights before backpropagating
|
||||
2) We need to calculate the gradient of the topk_weights. This gets messy since we need do an additioanl elementwise multiplication in the GEMM main loop and then write out in unpermuted order. For now, we do not fuse this step but calculate as a simple
|
||||
|
||||
Invalid combinations:
|
||||
- permute_y and use_tma_load: permuting y on store in forward -> load in permuted order in backward, therefore can't use TMA load (unless Blackwell which supports gather / scatter TMA)
|
||||
- permute_x and use_tma_store: permuting x on load in forward -> store in permuted order in backward, therefore can't use TMA store (unless Blackwell which supports gather / scatter TMA)
|
||||
|
||||
TODO:
|
||||
- We define indices for all conditions and expect that unused indices will be DCE'd during compilation. Check that this is the case otherwise will result in unnecessary register usage.
|
||||
"""
|
||||
|
||||
|
||||
@triton.jit
|
||||
def _grouped_gemm_dX_kernel(
|
||||
dY_ptr, # [M_total, N]
|
||||
w_ptr, # [E, N, K]
|
||||
dX_ptr, # [M_total, K]
|
||||
gather_indices_ptr,
|
||||
m_sizes_ptr,
|
||||
# problem sizes
|
||||
NUM_EXPERTS: tl.constexpr,
|
||||
NUM_TOKENS: tl.constexpr,
|
||||
TOPK: tl.constexpr,
|
||||
N: tl.constexpr,
|
||||
K: tl.constexpr,
|
||||
NUM_SMS: tl.constexpr,
|
||||
# Tuning parameters
|
||||
BLOCK_SIZE_M: tl.constexpr,
|
||||
BLOCK_SIZE_N: tl.constexpr,
|
||||
BLOCK_SIZE_K: tl.constexpr,
|
||||
PERMUTE_X: tl.constexpr = False,
|
||||
PERMUTE_Y: tl.constexpr = False,
|
||||
USE_TMA_LOAD_W: tl.constexpr = False,
|
||||
USE_TMA_LOAD_dY: tl.constexpr = False,
|
||||
USE_TMA_STORE: tl.constexpr = False,
|
||||
FLATTEN: tl.constexpr = True,
|
||||
) -> None:
|
||||
TOTAL_TOKENS: tl.constexpr = NUM_TOKENS * TOPK
|
||||
output_dtype = dX_ptr.dtype.element_ty
|
||||
|
||||
tidx = tl.program_id(0)
|
||||
# This removes the need for predication along N in the GEMM main loop
|
||||
tl.static_assert(N % BLOCK_SIZE_N == 0, "N must be divisible by BLOCK_SIZE_N")
|
||||
tl.static_assert(K % BLOCK_SIZE_K == 0, "K must be divisible by BLOCK_SIZE_K")
|
||||
|
||||
# Create TMA descriptors for loading sorted tokens
|
||||
# When using TMA load, we don't permute_x, so shape should be [TOTAL_TOKENS, K]
|
||||
# Also, we are defining a single global descriptor with single block shape
|
||||
# Need to check that this does not result in errors when crossing expert boundaries
|
||||
if USE_TMA_LOAD_dY:
|
||||
dY_desc = tl._experimental_make_tensor_descriptor(
|
||||
dY_ptr,
|
||||
shape=[TOTAL_TOKENS, N],
|
||||
strides=[N, 1],
|
||||
block_shape=[BLOCK_SIZE_M, BLOCK_SIZE_N],
|
||||
)
|
||||
|
||||
if USE_TMA_LOAD_W:
|
||||
expert_stride = N * K
|
||||
w_desc = tl._experimental_make_tensor_descriptor(
|
||||
w_ptr,
|
||||
shape=[NUM_EXPERTS, N, K],
|
||||
strides=[expert_stride, K, 1],
|
||||
block_shape=[1, BLOCK_SIZE_N, BLOCK_SIZE_K],
|
||||
)
|
||||
|
||||
m_end = 0
|
||||
processed_tiles = 0
|
||||
m_block_range = tl.arange(0, BLOCK_SIZE_M)
|
||||
n_block_range = tl.arange(0, BLOCK_SIZE_N)
|
||||
k_block_range = tl.arange(0, BLOCK_SIZE_K)
|
||||
|
||||
for expert_idx in range(NUM_EXPERTS, flatten=FLATTEN):
|
||||
m_start = m_end
|
||||
m_size = tl.load(m_sizes_ptr + expert_idx).to(tl.int32)
|
||||
m_end = m_start + m_size
|
||||
|
||||
if m_size > 0:
|
||||
# Advance n offset to the weights for that respective expert
|
||||
n_start = expert_idx * N
|
||||
# N_start_offset = g.to(tl.int64) * N
|
||||
# tiles for this group's GEMM
|
||||
num_m_tiles = tl.cdiv(m_size, BLOCK_SIZE_M)
|
||||
num_k_tiles = tl.cdiv(K, BLOCK_SIZE_K)
|
||||
num_tiles_per_expert = num_m_tiles * num_k_tiles
|
||||
|
||||
if USE_TMA_STORE:
|
||||
# Need to define descript within loop to predicate store along M
|
||||
tl.static_assert(
|
||||
K % BLOCK_SIZE_K == 0, "K must be divisible by BLOCK_SIZE_K"
|
||||
)
|
||||
dX_desc = tl._experimental_make_tensor_descriptor(
|
||||
dX_ptr,
|
||||
shape=[m_end, K],
|
||||
strides=[K, 1],
|
||||
block_shape=[BLOCK_SIZE_M, BLOCK_SIZE_K],
|
||||
)
|
||||
|
||||
# Lower bound and upper bound are defined relative to the total tiles processed so far
|
||||
# This ensures that we are only processing tiles for the current expert group AND
|
||||
# we never exceed the total number of tiles for all expert groups
|
||||
while tidx >= processed_tiles and tidx < (
|
||||
processed_tiles + num_tiles_per_expert
|
||||
):
|
||||
group_index = tidx - processed_tiles
|
||||
|
||||
# Output tile for this thread block for this expert group
|
||||
tile_m_idx = group_index % num_m_tiles
|
||||
tile_k_idx = group_index // num_m_tiles
|
||||
|
||||
if PERMUTE_X or PERMUTE_Y:
|
||||
# These will be used for loading and storing in permuted order
|
||||
gather_offsets = tile_m_idx * BLOCK_SIZE_M + m_block_range
|
||||
# indices_to_gather = m_start + gather_offsets
|
||||
indices_to_gather = m_start + tl.max_contiguous(
|
||||
tl.multiple_of(gather_offsets % m_size, BLOCK_SIZE_M),
|
||||
BLOCK_SIZE_M,
|
||||
)
|
||||
expert_token_idx = tl.load(
|
||||
gather_indices_ptr + indices_to_gather,
|
||||
mask=indices_to_gather < TOTAL_TOKENS,
|
||||
)
|
||||
expert_token_offsets = expert_token_idx[:, None]
|
||||
|
||||
# Masks for permuted load and store
|
||||
row_mask = gather_offsets < m_size
|
||||
row_mask = row_mask[:, None]
|
||||
|
||||
# We only take into account the following two cases: (PERMUTE_X and NOT PERMUTE_Y) and (NOT PERMUTE_X and PERMUTE_Y)
|
||||
# Hence, we can make the following simplifying assumptions when loading and storing
|
||||
# Note the different strides between the two cases: the offsets for loading and storing are flipped and the strides must also be adjusted
|
||||
|
||||
if PERMUTE_X:
|
||||
# Case where we permuted on load in the forward pass (typically first grouped GEMM in MoE MLP)
|
||||
load_a_idx = (
|
||||
indices_to_gather[:, None] * N
|
||||
) # Load in contiguous (expert grouped) order
|
||||
store_idx = (
|
||||
expert_token_offsets * K
|
||||
) # Permute on store from expert -> token order
|
||||
else:
|
||||
# Case where we permuted on store in the forward pass (typically second grouped GEMM in MoE MLP)
|
||||
load_a_idx = (
|
||||
expert_token_offsets * N
|
||||
) # Permute on load from token -> expert order
|
||||
store_idx = (
|
||||
indices_to_gather[:, None] * K
|
||||
) # Store in contiguous order
|
||||
else:
|
||||
# # Position in full matrix - needed for TMA
|
||||
# m_offset = (M_start + (tile_m_idx * BLOCK_SIZE_M)).to(tl.int32)
|
||||
# k_offset = (tile_k_idx * BLOCK_SIZE_K).to(tl.int32)
|
||||
# Offsets *relative* to the *current* expert -- m_start will then advance to this expert's start token
|
||||
offs_am = tile_m_idx * BLOCK_SIZE_M + m_block_range
|
||||
|
||||
# [M, N] @ [N, K] -> [M, K] => Stride for A is N, stride for B is K
|
||||
# We need two additional offsets:
|
||||
# 1. For A, m_start to advance to this expert's start token
|
||||
# 2. For B, n_start to advance to this expert's weights since we are passing in an [E, N, K] weight matrix
|
||||
row_offsets_a = m_start + offs_am[:, None]
|
||||
load_a_idx = row_offsets_a * N
|
||||
store_idx = row_offsets_a * K
|
||||
row_mask = offs_am[:, None] < m_size
|
||||
|
||||
if not USE_TMA_LOAD_dY:
|
||||
dY_ptrs = dY_ptr + load_a_idx + n_block_range[None, :]
|
||||
|
||||
offs_bk = tile_k_idx * BLOCK_SIZE_K + k_block_range
|
||||
if not USE_TMA_LOAD_W:
|
||||
row_offsets_b = n_start + n_block_range
|
||||
# offs_bn = n_start + n_block_range
|
||||
# row_offsets_b = tl.max_contiguous(tl.multiple_of(offs_bn, BLOCK_SIZE_N), BLOCK_SIZE_N)
|
||||
w_ptrs = w_ptr + row_offsets_b[:, None] * K + offs_bk[None, :]
|
||||
|
||||
# TODO: check whether predication along K is needed since we checked that K is divisible by BLOCK_SIZE_K in the forward kernel
|
||||
# col_mask = offs_bk[None, :] < K
|
||||
store_mask = row_mask # & col_mask
|
||||
|
||||
accumulator = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_K), dtype=tl.float32)
|
||||
|
||||
# GEMM main loop
|
||||
for n_offset in range(0, N, BLOCK_SIZE_N):
|
||||
# dY block [M, N]
|
||||
if not USE_TMA_LOAD_dY:
|
||||
dY = tl.load(dY_ptrs, mask=row_mask)
|
||||
else:
|
||||
dY = dY_desc.load(
|
||||
[m_start + tile_m_idx * BLOCK_SIZE_M, n_offset]
|
||||
)
|
||||
|
||||
if not USE_TMA_LOAD_W:
|
||||
w = tl.load(w_ptrs) # , mask=col_mask)
|
||||
else:
|
||||
w = w_desc.load(
|
||||
[expert_idx, n_offset, tile_k_idx * BLOCK_SIZE_K]
|
||||
)
|
||||
w = tl.reshape(w, (BLOCK_SIZE_N, BLOCK_SIZE_K))
|
||||
# TODO: check if predication along K is needed since we checked that K is divisible by BLOCK_SIZE_K in the forward kernel
|
||||
|
||||
# [M, N] @ [N, K] -> [M, K]
|
||||
accumulator += tl.dot(dY, w) # NOTE: no transpose of b
|
||||
|
||||
# Advance A along contiguous dimension
|
||||
if not USE_TMA_LOAD_dY:
|
||||
dY_ptrs += BLOCK_SIZE_N
|
||||
# Note we are no longer advancing B along contiguous dimension since weights are arranged as [N, K]
|
||||
# Instead, we need to stride by K to advance to the [N_BLOCK_SIZE, K_BLOCK_SIZE] tile
|
||||
if not USE_TMA_LOAD_W:
|
||||
w_ptrs += BLOCK_SIZE_N * K
|
||||
|
||||
dX = accumulator.to(output_dtype)
|
||||
|
||||
# Writing out a BLOCK_M x BLOCK_K tile, so we need to stride by K
|
||||
if USE_TMA_STORE:
|
||||
offset_m = tile_m_idx * BLOCK_SIZE_M # .to(tl.int32)
|
||||
offset_k = tile_k_idx * BLOCK_SIZE_K # .to(tl.int32)
|
||||
dX_desc.store([m_start + offset_m, offset_k], dX)
|
||||
else:
|
||||
tl.store(
|
||||
dX_ptr + store_idx + offs_bk[None, :],
|
||||
dX,
|
||||
mask=store_mask,
|
||||
)
|
||||
|
||||
# Move to the next tile within this expert group
|
||||
tidx += NUM_SMS
|
||||
|
||||
# Update the total tiles count for the next expert group
|
||||
processed_tiles += num_tiles_per_expert
|
||||
|
||||
|
||||
_autotuned_grouped_gemm_dX_kernel = triton.autotune(
|
||||
configs=get_dX_kernel_configs(),
|
||||
prune_configs_by={"early_config_prune": prune_dX_configs},
|
||||
key=["NUM_EXPERTS", "NUM_TOKENS", "N", "K", "PERMUTE_X", "PERMUTE_Y"],
|
||||
)(_grouped_gemm_dX_kernel)
|
||||
|
||||
"""
|
||||
notes on permute_x:
|
||||
- for the first grouped GEMM, we permuted on load -> X was [num_tokens, K] and stored y in expert grouped order [num_tokens * topk, K]
|
||||
- in the backwards pass, we need to permute on load of X while loading dy in contiguous (expert grouped) order
|
||||
- since we are writing out dW, there is no need to permute on store
|
||||
|
||||
notes on permute_y:
|
||||
- for the second grouped GEMM, we permuted on store -> y was permuted from expert grouped order to token order, x was loaded in expert grouped order since it was the output of the first grouped GEMM
|
||||
- in the backwards pass, we need to permute on load of dy to get from token order to expert grouped order to match the order of X
|
||||
- since we are writing out dW, there is no need to permute on store
|
||||
|
||||
notes on TMA loading:
|
||||
- if we're TMA loading both X and dY, then we need to mask along the M dimension
|
||||
to account for expert boundaries
|
||||
- we can either
|
||||
- define TMA descriptors within the outer for loop to predicate loads
|
||||
or
|
||||
- mask along M after loading
|
||||
"""
|
||||
|
||||
|
||||
@triton.jit
|
||||
def _grouped_gemm_dW_kernel(
|
||||
x_ptr,
|
||||
dY_ptr,
|
||||
dW_ptr,
|
||||
m_sizes_ptr,
|
||||
gather_indices_ptr,
|
||||
# problem sizes
|
||||
NUM_TOKENS: tl.constexpr,
|
||||
TOPK: tl.constexpr,
|
||||
NUM_EXPERTS: tl.constexpr,
|
||||
N: tl.constexpr,
|
||||
K: tl.constexpr,
|
||||
NUM_SMS: tl.constexpr,
|
||||
BLOCK_SIZE_N: tl.constexpr,
|
||||
BLOCK_SIZE_K: tl.constexpr,
|
||||
BLOCK_SIZE_M: tl.constexpr,
|
||||
PERMUTE_X: tl.constexpr = False,
|
||||
PERMUTE_Y: tl.constexpr = False,
|
||||
USE_TMA_LOAD_dY: tl.constexpr = False,
|
||||
USE_TMA_LOAD_X: tl.constexpr = False,
|
||||
USE_TMA_STORE: tl.constexpr = False,
|
||||
FLATTEN: tl.constexpr = True,
|
||||
acc_dtype: tl.constexpr = tl.float32,
|
||||
) -> None:
|
||||
TOTAL_TOKENS: tl.constexpr = NUM_TOKENS * TOPK
|
||||
TMA_LOAD_BOTH: tl.constexpr = USE_TMA_LOAD_X and USE_TMA_LOAD_dY
|
||||
|
||||
tidx = tl.program_id(0)
|
||||
output_dtype = dW_ptr.dtype.element_ty
|
||||
|
||||
if USE_TMA_LOAD_dY and not TMA_LOAD_BOTH:
|
||||
dY_desc = tl._experimental_make_tensor_descriptor(
|
||||
dY_ptr,
|
||||
shape=[TOTAL_TOKENS, N],
|
||||
strides=[N, 1],
|
||||
block_shape=[BLOCK_SIZE_M, BLOCK_SIZE_N],
|
||||
)
|
||||
|
||||
if USE_TMA_LOAD_X and not TMA_LOAD_BOTH:
|
||||
x_desc = tl._experimental_make_tensor_descriptor(
|
||||
x_ptr,
|
||||
shape=[TOTAL_TOKENS, K],
|
||||
strides=[K, 1],
|
||||
block_shape=[BLOCK_SIZE_M, BLOCK_SIZE_K],
|
||||
)
|
||||
# Output tiles per expert, since each expert weight matrix is [N, K]
|
||||
num_n_tiles = tl.cdiv(N, BLOCK_SIZE_N)
|
||||
num_k_tiles = tl.cdiv(K, BLOCK_SIZE_K)
|
||||
output_tiles_per_expert = num_n_tiles * num_k_tiles
|
||||
|
||||
block_range_m = tl.arange(0, BLOCK_SIZE_M)
|
||||
block_range_n = tl.arange(0, BLOCK_SIZE_N)
|
||||
block_range_k = tl.arange(0, BLOCK_SIZE_K)
|
||||
|
||||
# NOTE: Important that N % BLOCK_SIZE_N == 0 and K % BLOCK_SIZE_K == 0 when using TMA store
|
||||
if USE_TMA_STORE:
|
||||
tl.static_assert(N % BLOCK_SIZE_N == 0, "N must be divisible by BLOCK_SIZE_N")
|
||||
tl.static_assert(K % BLOCK_SIZE_K == 0, "K must be divisible by BLOCK_SIZE_K")
|
||||
dW_desc = tl._experimental_make_tensor_descriptor(
|
||||
dW_ptr,
|
||||
shape=[NUM_EXPERTS, N, K],
|
||||
strides=[N * K, K, 1],
|
||||
block_shape=[1, BLOCK_SIZE_N, BLOCK_SIZE_K],
|
||||
)
|
||||
|
||||
for tile_idx in range(
|
||||
tidx, output_tiles_per_expert, NUM_SMS
|
||||
): # , flatten=FLATTEN):
|
||||
# Output tile index
|
||||
tile_n_idx = tile_idx % num_n_tiles
|
||||
tile_k_idx = tile_idx // num_n_tiles
|
||||
|
||||
# Output tile offsets
|
||||
n_offset = tile_n_idx * BLOCK_SIZE_N
|
||||
k_offset = tile_k_idx * BLOCK_SIZE_K
|
||||
|
||||
# For storing
|
||||
# TODO: Check whether the k mask is needed since we statically check that K is divisible by BLOCK_SIZE_K in the forward kernel
|
||||
# ditto for n_mask
|
||||
n_mask = block_range_n + n_offset < N
|
||||
k_mask = block_range_k + k_offset < K
|
||||
nk_mask = n_mask[:, None] & k_mask[None, :]
|
||||
|
||||
m_end = 0
|
||||
for expert_idx in range(NUM_EXPERTS):
|
||||
# We need to instantiate a fresh accumulator for each expert
|
||||
accumulator = tl.zeros((BLOCK_SIZE_N, BLOCK_SIZE_K), dtype=acc_dtype)
|
||||
|
||||
m_start = m_end
|
||||
# Need to figure out why this cast is needed, otherwise compiler complains about mismatching types
|
||||
m_size = tl.load(m_sizes_ptr + expert_idx).to(tl.int32)
|
||||
m_end = m_start + m_size
|
||||
|
||||
# NOTE: when storing the result, we need to offset by n_start since we are storing the result for this expert to the global [E, N, K] weight matrix
|
||||
n_start = expert_idx * N
|
||||
store_row_offs = n_start + n_offset + block_range_n
|
||||
|
||||
if m_size > 0:
|
||||
if TMA_LOAD_BOTH:
|
||||
dY_desc = tl._experimental_make_tensor_descriptor(
|
||||
dY_ptr,
|
||||
shape=[m_end, N],
|
||||
strides=[N, 1],
|
||||
block_shape=[BLOCK_SIZE_M, BLOCK_SIZE_N],
|
||||
)
|
||||
|
||||
x_desc = tl._experimental_make_tensor_descriptor(
|
||||
x_ptr,
|
||||
shape=[m_end, K],
|
||||
strides=[K, 1],
|
||||
block_shape=[BLOCK_SIZE_M, BLOCK_SIZE_K],
|
||||
)
|
||||
|
||||
for tile_m_idx in range(0, m_size, BLOCK_SIZE_M):
|
||||
m_block_size = tl.minimum(BLOCK_SIZE_M, m_size - tile_m_idx)
|
||||
|
||||
if m_block_size > 0:
|
||||
# Global offset for this chunk
|
||||
m_global_offset = m_start + tile_m_idx
|
||||
m_offsets = m_global_offset + block_range_m
|
||||
|
||||
if PERMUTE_X or PERMUTE_Y:
|
||||
# These will be used for loading and storing in permuted order
|
||||
gather_offsets = (
|
||||
tile_m_idx + block_range_m
|
||||
) # NOTE: tile_m_idx is already strided by BLOCK_SIZE_M
|
||||
|
||||
indices_to_gather = m_start + tl.max_contiguous(
|
||||
tl.multiple_of(gather_offsets % m_size, BLOCK_SIZE_M),
|
||||
BLOCK_SIZE_M,
|
||||
)
|
||||
# indices_to_gather = m_start + gather_offsets
|
||||
expert_token_idx = tl.load(
|
||||
gather_indices_ptr + indices_to_gather,
|
||||
mask=indices_to_gather < TOTAL_TOKENS,
|
||||
)
|
||||
expert_token_offsets = expert_token_idx[:, None]
|
||||
|
||||
# Masks for permuted load and store
|
||||
row_load_mask = gather_offsets < m_size
|
||||
|
||||
# We only take into account the following two cases: (PERMUTE_X and NOT PERMUTE_Y) and (NOT PERMUTE_X and PERMUTE_Y)
|
||||
# Hence, we can make the following simplifying assumptions when loading and storing
|
||||
# Note the different strides between the two cases: the offsets for loading and storing are flipped and the strides must also be adjusted
|
||||
if PERMUTE_X:
|
||||
x_row_load_idx = (
|
||||
(expert_token_offsets // TOPK) * K
|
||||
) # Permute on load from token -> expert order, divide by TOPK to index from original number of tokens
|
||||
dY_row_load_idx = m_offsets[:, None] * N
|
||||
else:
|
||||
x_row_load_idx = (
|
||||
indices_to_gather[:, None] * K
|
||||
) # Load in contiguous order (no permutation on load)
|
||||
dY_row_load_idx = expert_token_offsets * N
|
||||
|
||||
else:
|
||||
x_row_load_idx = m_offsets[:, None] * K
|
||||
dY_row_load_idx = m_offsets[:, None] * N
|
||||
row_load_mask = block_range_m < m_block_size
|
||||
|
||||
mk_mask = row_load_mask[:, None] & k_mask[None, :]
|
||||
mn_mask = row_load_mask[:, None] & n_mask[None, :]
|
||||
|
||||
if USE_TMA_LOAD_X:
|
||||
x = x_desc.load([m_global_offset, k_offset])
|
||||
else:
|
||||
x = tl.load(
|
||||
x_ptr
|
||||
+ x_row_load_idx
|
||||
+ (k_offset + block_range_k)[None, :],
|
||||
mask=mk_mask,
|
||||
)
|
||||
|
||||
if USE_TMA_LOAD_dY:
|
||||
dY = dY_desc.load([m_global_offset, n_offset])
|
||||
else:
|
||||
dY = tl.load(
|
||||
dY_ptr
|
||||
+ dY_row_load_idx
|
||||
+ (n_offset + block_range_n)[None, :],
|
||||
mask=mn_mask,
|
||||
)
|
||||
|
||||
accumulator += tl.dot(
|
||||
dY.T, # [BLOCK_N, BLOCK_M]
|
||||
x, # [BLOCK_M, BLOCK_K]
|
||||
)
|
||||
|
||||
y = accumulator.to(output_dtype)
|
||||
if USE_TMA_STORE:
|
||||
# Need to expand dims to match [E, N, K] shape
|
||||
y = tl.expand_dims(y, 0)
|
||||
dW_desc.store([expert_idx, n_offset, k_offset], y)
|
||||
else:
|
||||
tl.store(
|
||||
dW_ptr
|
||||
# + (n_offset + offs_n)[:, None] * K
|
||||
+ store_row_offs[:, None] * K
|
||||
+ (k_offset + block_range_k)[None, :],
|
||||
y,
|
||||
mask=nk_mask,
|
||||
)
|
||||
|
||||
|
||||
_autotuned_grouped_gemm_dW_kernel = triton.autotune(
|
||||
configs=get_dW_kernel_configs(),
|
||||
prune_configs_by={"early_config_prune": prune_kernel_configs_backward_dW},
|
||||
key=["NUM_EXPERTS", "NUM_TOKENS", "N", "K", "PERMUTE_X", "PERMUTE_Y"],
|
||||
)(_grouped_gemm_dW_kernel)
|
||||
262
unsloth/kernels/moe/grouped_gemm/kernels/forward.py
Normal file
262
unsloth/kernels/moe/grouped_gemm/kernels/forward.py
Normal file
|
|
@ -0,0 +1,262 @@
|
|||
import torch
|
||||
import triton
|
||||
import triton.language as tl
|
||||
|
||||
from grouped_gemm.kernels.autotuning import (
|
||||
get_forward_configs,
|
||||
prune_kernel_configs_fwd,
|
||||
)
|
||||
|
||||
|
||||
#
|
||||
# PERMUTE_X -> permute tokens so that they are ordered by expert
|
||||
# PERMUTE_Y -> permute output so that they are ordered by token
|
||||
# These are effectively the same thing: the former loads in permuted order, the latter stores in permuted order => we only need to define the permutation indices once
|
||||
# In the former, we use these row indices when loading X
|
||||
# For the latter, we use these row indices when storing Y
|
||||
# FUSE_MUL -> multiply routed outputs by their respective weights
|
||||
# topk_weights are in token order
|
||||
# Only account for the case when X is in expert order and we are permuting Y when fusing mul -- this precondition is checked in the interface
|
||||
@triton.jit
|
||||
def _grouped_gemm_forward_kernel(
|
||||
x_ptr,
|
||||
w_ptr,
|
||||
y_ptr,
|
||||
# Variable depending on routed probs
|
||||
m_sizes_ptr,
|
||||
gather_indices_ptr,
|
||||
topk_weights_ptr,
|
||||
# Constant problem shapes
|
||||
NUM_EXPERTS: tl.constexpr,
|
||||
NUM_TOKENS: tl.constexpr,
|
||||
TOPK: tl.constexpr,
|
||||
N: tl.constexpr,
|
||||
K: tl.constexpr,
|
||||
NUM_SMS: tl.constexpr,
|
||||
# Tuning params
|
||||
BLOCK_SIZE_M: tl.constexpr,
|
||||
BLOCK_SIZE_N: tl.constexpr,
|
||||
BLOCK_SIZE_K: tl.constexpr,
|
||||
PERMUTE_X: tl.constexpr = False,
|
||||
PERMUTE_Y: tl.constexpr = False,
|
||||
FUSE_MUL_PRE: tl.constexpr = False,
|
||||
FUSE_MUL_POST: tl.constexpr = False,
|
||||
USE_FAST_ACCUM: tl.constexpr = False,
|
||||
USE_TMA_LOAD_W: tl.constexpr = False,
|
||||
USE_TMA_LOAD_X: tl.constexpr = False,
|
||||
USE_TMA_STORE: tl.constexpr = False,
|
||||
acc_dtype: tl.constexpr = tl.float32,
|
||||
FLATTEN: tl.constexpr = True,
|
||||
) -> None:
|
||||
tl.static_assert(K % BLOCK_SIZE_K == 0)
|
||||
|
||||
TOTAL_TOKENS: tl.constexpr = NUM_TOKENS * TOPK
|
||||
SHOULD_PERMUTE: tl.constexpr = PERMUTE_X or PERMUTE_Y
|
||||
SHOULD_FUSE_MUL: tl.constexpr = FUSE_MUL_PRE or FUSE_MUL_POST
|
||||
SHOULD_PERMUTE_OR_FUSE: tl.constexpr = SHOULD_PERMUTE or SHOULD_FUSE_MUL
|
||||
# tl.static_print("SHOULD_PERMUTE", PERMUTE_X, PERMUTE_Y, FUSE_MUL_PRE, FUSE_MUL_POST, SHOULD_PERMUTE, SHOULD_FUSE, SHOULD_PERMUTE_OR_FUSE)
|
||||
tidx = tl.program_id(0)
|
||||
output_dtype: tl.dtype = y_ptr.dtype.element_ty
|
||||
|
||||
# Create TMA descriptors for loading sorted tokens
|
||||
# When using TMA load, we don't permute_x, so shape should be [TOTAL_TOKENS, K]
|
||||
# Also, we are defining a single global descriptor with single block shape
|
||||
# Need to check that this does not result in errors when crossing expert boundaries
|
||||
if USE_TMA_LOAD_X:
|
||||
x_desc = tl._experimental_make_tensor_descriptor(
|
||||
x_ptr,
|
||||
shape=[TOTAL_TOKENS, K],
|
||||
strides=[K, 1],
|
||||
block_shape=[BLOCK_SIZE_M, BLOCK_SIZE_K],
|
||||
)
|
||||
|
||||
if USE_TMA_LOAD_W:
|
||||
expert_stride = N * K
|
||||
w_desc = tl._experimental_make_tensor_descriptor(
|
||||
w_ptr,
|
||||
shape=[NUM_EXPERTS, N, K],
|
||||
strides=[expert_stride, K, 1],
|
||||
block_shape=[1, BLOCK_SIZE_N, BLOCK_SIZE_K],
|
||||
)
|
||||
|
||||
m_end = 0
|
||||
processed_tiles = 0
|
||||
m_block_range = tl.arange(0, BLOCK_SIZE_M)
|
||||
|
||||
for expert_idx in tl.range(NUM_EXPERTS, flatten=FLATTEN):
|
||||
m_start = m_end
|
||||
m_size = tl.load(m_sizes_ptr + expert_idx).to(tl.int32)
|
||||
m_end = m_start + m_size
|
||||
|
||||
if m_size > 0:
|
||||
n_start = expert_idx * N
|
||||
|
||||
num_m_tiles = tl.cdiv(m_size, BLOCK_SIZE_M)
|
||||
num_n_tiles = tl.cdiv(N, BLOCK_SIZE_N)
|
||||
num_tiles_per_expert = num_m_tiles * num_n_tiles
|
||||
|
||||
# Need to create tma_store within loop since we need to predicate stores based on m_size
|
||||
if USE_TMA_STORE:
|
||||
y_desc = tl._experimental_make_tensor_descriptor(
|
||||
y_ptr, # + m_start * N,
|
||||
shape=[m_end, N],
|
||||
strides=[N, 1],
|
||||
block_shape=[BLOCK_SIZE_M, BLOCK_SIZE_N],
|
||||
)
|
||||
|
||||
# Process tiles for this expert
|
||||
while (
|
||||
tidx >= processed_tiles
|
||||
and tidx < processed_tiles + num_tiles_per_expert
|
||||
):
|
||||
tile_idx = tidx - processed_tiles
|
||||
|
||||
# Check if L2 cache re-use for this order is optimal
|
||||
tile_m_idx = tile_idx % num_m_tiles
|
||||
tile_n_idx = tile_idx // num_m_tiles
|
||||
|
||||
if SHOULD_PERMUTE_OR_FUSE:
|
||||
# These will be used for loading and storing in permuted order
|
||||
gather_offsets = tile_m_idx * BLOCK_SIZE_M + m_block_range
|
||||
indices_to_gather = m_start + tl.max_contiguous(
|
||||
tl.multiple_of(gather_offsets % m_size, BLOCK_SIZE_M),
|
||||
BLOCK_SIZE_M,
|
||||
)
|
||||
expert_token_idx = tl.load(
|
||||
gather_indices_ptr + indices_to_gather,
|
||||
mask=indices_to_gather < TOTAL_TOKENS,
|
||||
)
|
||||
expert_token_offsets = expert_token_idx[:, None]
|
||||
|
||||
# Masks for permuted load and store
|
||||
|
||||
row_mask = gather_offsets < m_size
|
||||
row_mask = row_mask[:, None]
|
||||
|
||||
# row_mask = indices_to_gather < m_end
|
||||
# row_mask = row_mask[:, None]
|
||||
|
||||
# We only take into account the following two cases: (PERMUTE_X and NOT PERMUTE_Y) and (NOT PERMUTE_X and PERMUTE_Y)
|
||||
# Hence, we can make the following simplifying assumptions when loading and storing
|
||||
# Note the different strides between the two cases: the offsets for loading and storing are flipped and the strides must also be adjusted
|
||||
if PERMUTE_X:
|
||||
load_idx = (
|
||||
(expert_token_offsets // TOPK) * K
|
||||
) # Permute on load from token -> expert order, divide by TOPK to index from original number of tokens
|
||||
store_idx = (
|
||||
indices_to_gather[:, None] * N
|
||||
) # Store in contiguous order
|
||||
else:
|
||||
off_am = tile_m_idx * BLOCK_SIZE_M
|
||||
if not PERMUTE_Y:
|
||||
# These will already be computed if permuting y
|
||||
offs_am = off_am + m_block_range
|
||||
row_mask = offs_am[:, None] < m_size
|
||||
row_idx = m_start + offs_am[:, None]
|
||||
store_idx = row_idx * N
|
||||
if not USE_TMA_LOAD_X:
|
||||
load_idx = row_idx * K
|
||||
|
||||
if PERMUTE_Y:
|
||||
if not USE_TMA_LOAD_X:
|
||||
load_idx = (
|
||||
indices_to_gather[:, None] * K
|
||||
) # Load in contiguous order (no permutation on load)
|
||||
# offs_am = off_am + m_block_range
|
||||
# row_mask = offs_am[:, None] < m_size
|
||||
store_idx = (
|
||||
expert_token_offsets * N
|
||||
) # Permute on store from expert -> token order
|
||||
|
||||
# We always load topk weights in expert order
|
||||
# In the pre-multiplication case, we multiply permuted hidden states by weights before the first gemm
|
||||
# In the post-multiplication case, we multiply permuted hidden states by weights after the second gemm
|
||||
# In either case, the hidden states are grouped by expert, so we always permute on load of topk weights
|
||||
if SHOULD_FUSE_MUL:
|
||||
topk_load_idx = expert_token_offsets
|
||||
|
||||
accumulator = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N), dtype=acc_dtype)
|
||||
|
||||
offs_k = tl.arange(0, BLOCK_SIZE_K)
|
||||
|
||||
if not USE_TMA_LOAD_X:
|
||||
x_ptrs = x_ptr + load_idx + offs_k[None, :]
|
||||
|
||||
if not USE_TMA_LOAD_W:
|
||||
offs_bn = tile_n_idx * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)
|
||||
offs_bn = tl.max_contiguous(
|
||||
tl.multiple_of(offs_bn % N, BLOCK_SIZE_N), BLOCK_SIZE_N
|
||||
)
|
||||
w_ptrs = w_ptr + (n_start + offs_bn[:, None]) * K + offs_k[None, :]
|
||||
|
||||
for k_offset in range(0, K, BLOCK_SIZE_K):
|
||||
if not USE_TMA_LOAD_X:
|
||||
x = tl.load(x_ptrs, mask=row_mask)
|
||||
else:
|
||||
x = x_desc.load([m_start + off_am, k_offset])
|
||||
|
||||
if FUSE_MUL_PRE:
|
||||
# Check for correct broadcasting
|
||||
topk_weights = tl.load(
|
||||
topk_weights_ptr + topk_load_idx, mask=row_mask
|
||||
)
|
||||
x *= topk_weights.to(x.dtype)
|
||||
|
||||
if not USE_TMA_LOAD_W:
|
||||
w = tl.load(w_ptrs, mask=offs_bn[:, None] < N)
|
||||
else:
|
||||
w = w_desc.load(
|
||||
[expert_idx, tile_n_idx * BLOCK_SIZE_N, k_offset]
|
||||
)
|
||||
w = tl.reshape(w, (BLOCK_SIZE_N, BLOCK_SIZE_K))
|
||||
|
||||
accumulator += tl.dot(x, w.T)
|
||||
|
||||
if not USE_TMA_LOAD_X:
|
||||
x_ptrs += BLOCK_SIZE_K
|
||||
|
||||
if not USE_TMA_LOAD_W:
|
||||
w_ptrs += BLOCK_SIZE_K
|
||||
|
||||
y = accumulator.to(output_dtype)
|
||||
|
||||
# NOTE: order of fusing multiplication is important
|
||||
# Fusing before accumulator dtype conversion results in numerical diffs
|
||||
if FUSE_MUL_POST:
|
||||
# Check for correct broadcasting
|
||||
topk_weights = tl.load(
|
||||
topk_weights_ptr + topk_load_idx, mask=row_mask
|
||||
)
|
||||
y *= topk_weights.to(output_dtype)
|
||||
|
||||
offs_bn = tile_n_idx * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)
|
||||
store_mask = row_mask & (offs_bn[None, :] < N)
|
||||
|
||||
if USE_TMA_STORE:
|
||||
offset_m = tile_m_idx * BLOCK_SIZE_M # .to(tl.int32)
|
||||
offset_n = tile_n_idx * BLOCK_SIZE_N # .to(tl.int32)
|
||||
y_desc.store([m_start + offset_m, offset_n], y)
|
||||
else:
|
||||
tl.store(
|
||||
y_ptr + store_idx + offs_bn[None, :],
|
||||
y,
|
||||
mask=store_mask,
|
||||
)
|
||||
tidx += NUM_SMS
|
||||
|
||||
processed_tiles += num_tiles_per_expert
|
||||
|
||||
|
||||
_autotuned_grouped_gemm_forward_kernel = triton.autotune(
|
||||
configs=get_forward_configs(),
|
||||
prune_configs_by={"early_config_prune": prune_kernel_configs_fwd},
|
||||
key=[
|
||||
"NUM_EXPERTS",
|
||||
"NUM_TOKENS",
|
||||
"N",
|
||||
"K",
|
||||
"PERMUTE_X",
|
||||
"PERMUTE_Y",
|
||||
"FUSE_MUL_POST",
|
||||
],
|
||||
)(_grouped_gemm_forward_kernel)
|
||||
274
unsloth/kernels/moe/grouped_gemm/kernels/tuning.py
Normal file
274
unsloth/kernels/moe/grouped_gemm/kernels/tuning.py
Normal file
|
|
@ -0,0 +1,274 @@
|
|||
"""
|
||||
Manual tuning utils
|
||||
"""
|
||||
|
||||
from collections import OrderedDict
|
||||
from dataclasses import asdict, dataclass, fields
|
||||
from itertools import product
|
||||
from typing import Optional
|
||||
|
||||
import pandas as pd
|
||||
import torch
|
||||
import triton
|
||||
from triton.runtime.errors import OutOfResources
|
||||
|
||||
from grouped_gemm.kernels.autotuning import (
|
||||
BOOLS,
|
||||
DEFAULT_K_BLOCK_SIZES,
|
||||
DEFAULT_M_BLOCK_SIZES,
|
||||
DEFAULT_N_BLOCK_SIZES,
|
||||
DEFAULT_NUM_STAGES,
|
||||
DEFAULT_NUM_WARPS,
|
||||
)
|
||||
|
||||
|
||||
@dataclass
|
||||
class DeviceProperties:
|
||||
NUM_SM: int
|
||||
NUM_REGS: int
|
||||
SIZE_SMEM: int
|
||||
WARP_SIZE: int
|
||||
|
||||
|
||||
_DEVICE_PROPERTIES: Optional[DeviceProperties] = None
|
||||
|
||||
|
||||
def get_device_properties():
|
||||
global _DEVICE_PROPERTIES
|
||||
if _DEVICE_PROPERTIES is None:
|
||||
properties = triton.runtime.driver.active.utils.get_device_properties(
|
||||
torch.cuda.current_device()
|
||||
)
|
||||
NUM_SM = properties["multiprocessor_count"]
|
||||
NUM_REGS = properties["max_num_regs"]
|
||||
SIZE_SMEM = properties["max_shared_mem"]
|
||||
WARP_SIZE = properties["warpSize"]
|
||||
_DEVICE_PROPERTIES = DeviceProperties(NUM_SM, NUM_REGS, SIZE_SMEM, WARP_SIZE)
|
||||
return _DEVICE_PROPERTIES
|
||||
|
||||
|
||||
@dataclass
|
||||
class KernelConfig:
|
||||
BLOCK_SIZE_M: int = 32
|
||||
BLOCK_SIZE_N: int = 32
|
||||
BLOCK_SIZE_K: int = 32
|
||||
num_warps: int = 4
|
||||
num_stages: int = 2
|
||||
flatten: bool = True
|
||||
permute_x: bool = False
|
||||
permute_y: bool = False
|
||||
fuse_mul_post: bool = False
|
||||
use_tma_store: bool = False
|
||||
|
||||
def to_string(self, include_tuning_params: bool = False, include_tma: bool = False):
|
||||
s = []
|
||||
if self.permute_x:
|
||||
s.append("permute_x")
|
||||
if self.permute_y:
|
||||
s.append("permute_y")
|
||||
if include_tuning_params:
|
||||
s.append(
|
||||
f"BLOCK_SIZE_M={self.BLOCK_SIZE_M},BLOCK_SIZE_N={self.BLOCK_SIZE_N},BLOCK_SIZE_K={self.BLOCK_SIZE_K},num_warps={self.num_warps},num_stages={self.num_stages},flatten={self.flatten}"
|
||||
)
|
||||
if include_tma:
|
||||
for f in fields(self):
|
||||
if f.name.startswith("use_tma_"):
|
||||
if getattr(self, f.name):
|
||||
s.append(f.name)
|
||||
return ",".join(s)
|
||||
|
||||
|
||||
@dataclass
|
||||
class KernelConfigForward(KernelConfig):
|
||||
use_tma_load_w: bool = False
|
||||
use_tma_load_x: bool = False
|
||||
|
||||
|
||||
@dataclass
|
||||
class KernelConfigBackward_dW(KernelConfig):
|
||||
use_tma_load_dy: bool = False
|
||||
use_tma_load_x: bool = False
|
||||
|
||||
|
||||
@dataclass
|
||||
class KernelConfigBackward_dX(KernelConfig):
|
||||
use_tma_load_dy: bool = False
|
||||
use_tma_load_w: bool = False
|
||||
|
||||
|
||||
@dataclass
|
||||
class KernelResult:
|
||||
torch_time: float
|
||||
triton_time: float
|
||||
speedup: float
|
||||
kernel_config: KernelConfig
|
||||
|
||||
def to_dict(self):
|
||||
return OrderedDict(
|
||||
**asdict(self.kernel_config),
|
||||
torch_time=self.torch_time,
|
||||
triton_time=self.triton_time,
|
||||
speedup=self.speedup,
|
||||
)
|
||||
|
||||
@staticmethod
|
||||
def to_dataframe(
|
||||
results: list["KernelResult"], sort_by: str = "speedup", ascending: bool = False
|
||||
):
|
||||
df = pd.DataFrame([result.to_dict() for result in results])
|
||||
df = df.sort_values(by=sort_by, ascending=ascending)
|
||||
return df
|
||||
|
||||
@staticmethod
|
||||
def to_csv(
|
||||
results: list["KernelResult"],
|
||||
sort_by: str = "speedup",
|
||||
ascending: bool = False,
|
||||
filename: str = "results.csv",
|
||||
):
|
||||
df = KernelResult.to_dataframe(results, sort_by, ascending)
|
||||
df.to_csv(filename, index=False)
|
||||
|
||||
@staticmethod
|
||||
def print_table(
|
||||
results: list["KernelResult"],
|
||||
sort_by: str = "speedup",
|
||||
ascending: bool = False,
|
||||
num_results: int = 10,
|
||||
):
|
||||
df = KernelResult.to_dataframe(results, sort_by, ascending)
|
||||
print(df.head(num_results).to_string(index=False))
|
||||
|
||||
|
||||
def get_kernel_configs(
|
||||
BLOCK_M=DEFAULT_M_BLOCK_SIZES,
|
||||
BLOCK_N=DEFAULT_N_BLOCK_SIZES,
|
||||
BLOCK_K=DEFAULT_K_BLOCK_SIZES,
|
||||
num_warps=DEFAULT_NUM_WARPS,
|
||||
num_stages=DEFAULT_NUM_STAGES,
|
||||
use_tma_loads=BOOLS,
|
||||
fuse_permute=BOOLS,
|
||||
):
|
||||
kernel_configs_fwd = []
|
||||
kernel_configs_backward_dW = []
|
||||
kernel_configs_backward_dX = []
|
||||
for block_m, block_n, block_k, w, s, use_tma_load, permute in product(
|
||||
BLOCK_M, BLOCK_N, BLOCK_K, num_warps, num_stages, use_tma_loads, fuse_permute
|
||||
):
|
||||
kernel_configs_fwd.append(
|
||||
KernelConfigForward(
|
||||
BLOCK_SIZE_M=block_m,
|
||||
BLOCK_SIZE_N=block_n,
|
||||
BLOCK_SIZE_K=block_k,
|
||||
num_warps=w,
|
||||
num_stages=s,
|
||||
use_tma_load_x=use_tma_load,
|
||||
use_tma_load_w=use_tma_load,
|
||||
use_tma_store=False,
|
||||
permute_x=permute,
|
||||
permute_y=permute,
|
||||
)
|
||||
)
|
||||
kernel_configs_backward_dW.append(
|
||||
KernelConfigBackward_dW(
|
||||
BLOCK_SIZE_M=block_m,
|
||||
BLOCK_SIZE_N=block_n,
|
||||
BLOCK_SIZE_K=block_k,
|
||||
num_warps=w,
|
||||
num_stages=s,
|
||||
use_tma_load_dy=use_tma_load,
|
||||
use_tma_load_x=use_tma_load,
|
||||
use_tma_store=False,
|
||||
permute_x=permute,
|
||||
permute_y=permute,
|
||||
)
|
||||
)
|
||||
kernel_configs_backward_dX.append(
|
||||
KernelConfigBackward_dX(
|
||||
BLOCK_SIZE_M=block_m,
|
||||
BLOCK_SIZE_N=block_n,
|
||||
BLOCK_SIZE_K=block_k,
|
||||
num_warps=w,
|
||||
num_stages=s,
|
||||
use_tma_load_dy=use_tma_load,
|
||||
use_tma_load_w=use_tma_load,
|
||||
use_tma_store=False,
|
||||
permute_x=permute,
|
||||
permute_y=permute,
|
||||
)
|
||||
)
|
||||
|
||||
kernel_configs_fwd = prune_kernel_configs_fwd(kernel_configs_fwd)
|
||||
kernel_configs_backward_dW = prune_kernel_configs_backward_dW(
|
||||
kernel_configs_backward_dW
|
||||
)
|
||||
kernel_configs_backward_dX = prune_kernel_configs_backward_dX(
|
||||
kernel_configs_backward_dX
|
||||
)
|
||||
return kernel_configs_fwd, kernel_configs_backward_dW, kernel_configs_backward_dX
|
||||
|
||||
|
||||
def prune_kernel_configs_fwd(configs: list[KernelConfigForward]):
|
||||
pruned_configs = []
|
||||
for config in configs:
|
||||
if config.use_tma_load_x and config.permute_x:
|
||||
continue
|
||||
if config.permute_x and config.permute_y:
|
||||
continue
|
||||
if config.use_tma_store and config.permute_y:
|
||||
continue
|
||||
pruned_configs.append(config)
|
||||
return pruned_configs
|
||||
|
||||
|
||||
def prune_kernel_configs_backward_dX(configs: list[KernelConfigBackward_dX]):
|
||||
pruned_configs = []
|
||||
for config in configs:
|
||||
if config.use_tma_load_dy and config.permute_y:
|
||||
continue
|
||||
if config.permute_x and config.permute_y:
|
||||
continue
|
||||
if config.use_tma_store and config.permute_x:
|
||||
continue
|
||||
pruned_configs.append(config)
|
||||
return pruned_configs
|
||||
|
||||
|
||||
def prune_kernel_configs_backward_dW(configs: list[KernelConfigBackward_dW]):
|
||||
pruned_configs = []
|
||||
for config in configs:
|
||||
if config.use_tma_load_dy and config.permute_y:
|
||||
continue
|
||||
if config.use_tma_load_x and config.permute_x:
|
||||
continue
|
||||
if config.permute_x and config.permute_y:
|
||||
continue
|
||||
pruned_configs.append(config)
|
||||
return pruned_configs
|
||||
|
||||
|
||||
class TritonTuningContext:
|
||||
def __init__(self, kernel_config: KernelConfig):
|
||||
self.kernel_config = kernel_config
|
||||
self.success = True
|
||||
|
||||
def __enter__(self):
|
||||
# Setup code can be added here if needed
|
||||
return self
|
||||
|
||||
def __exit__(self, exc_type, exc_value, traceback):
|
||||
if exc_type is OutOfResources:
|
||||
name = exc_value.name
|
||||
required = exc_value.required
|
||||
limit = exc_value.limit
|
||||
print(
|
||||
f"Kernel config {self.kernel_config} failed: {name}, required: {required}, limit: {limit}"
|
||||
)
|
||||
self.success = False
|
||||
elif exc_type is not None:
|
||||
print(
|
||||
f"Error running Triton grouped GEMM for kernel config: {self.kernel_config}: {exc_value}"
|
||||
)
|
||||
self.success = False
|
||||
# Return False to propagate exceptions, True to suppress them
|
||||
return True
|
||||
0
unsloth/kernels/moe/grouped_gemm/reference/__init__.py
Normal file
0
unsloth/kernels/moe/grouped_gemm/reference/__init__.py
Normal file
158
unsloth/kernels/moe/grouped_gemm/reference/moe_block.py
Normal file
158
unsloth/kernels/moe/grouped_gemm/reference/moe_block.py
Normal file
|
|
@ -0,0 +1,158 @@
|
|||
import torch
|
||||
from transformers.models.qwen3_moe.configuration_qwen3_moe import Qwen3MoeConfig
|
||||
from transformers.models.qwen3_moe.modeling_qwen3_moe import Qwen3MoeSparseMoeBlock
|
||||
|
||||
from grouped_gemm.interface import grouped_gemm
|
||||
from grouped_gemm.kernels.tuning import (
|
||||
KernelConfigBackward_dW,
|
||||
KernelConfigBackward_dX,
|
||||
KernelConfigForward,
|
||||
)
|
||||
from grouped_gemm.reference.moe_ops import (
|
||||
Qwen3MoeGroupedGEMMBlock,
|
||||
permute,
|
||||
unpermute,
|
||||
)
|
||||
|
||||
"""
|
||||
Reference implementation of MoE block using grouped gemm.
|
||||
|
||||
This is the same as the Qwen3MoeGroupedGEMMBlock but with triton grouped gemm in place of torch-native grouped gemm implementation.
|
||||
|
||||
NOTE: This is NOT to be used for production as it contains many extra checks and saves all intermediate results for debugging.
|
||||
"""
|
||||
|
||||
|
||||
class Qwen3MoeFusedGroupedGEMMBlock(Qwen3MoeGroupedGEMMBlock):
|
||||
def __init__(
|
||||
self,
|
||||
config: Qwen3MoeConfig,
|
||||
gate: torch.Tensor,
|
||||
gate_up_proj: torch.Tensor,
|
||||
down_proj: torch.Tensor,
|
||||
permute_x: bool = True,
|
||||
permute_y: bool = True,
|
||||
autotune: bool = True,
|
||||
kernel_config_fwd: KernelConfigForward = None,
|
||||
kernel_config_bwd_dW: KernelConfigBackward_dW = None,
|
||||
kernel_config_bwd_dX: KernelConfigBackward_dX = None,
|
||||
dW_only: bool = False,
|
||||
dX_only: bool = False,
|
||||
):
|
||||
super().__init__(config, gate, gate_up_proj, down_proj)
|
||||
self.permute_x = permute_x
|
||||
self.permute_y = permute_y
|
||||
self.autotune = autotune
|
||||
if not autotune:
|
||||
assert (
|
||||
kernel_config_fwd is not None
|
||||
and kernel_config_bwd_dW is not None
|
||||
and kernel_config_bwd_dX is not None
|
||||
), "Kernel configs must be provided if autotune is False"
|
||||
self.kernel_config_fwd = kernel_config_fwd
|
||||
self.kernel_config_bwd_dW = kernel_config_bwd_dW
|
||||
self.kernel_config_bwd_dX = kernel_config_bwd_dX
|
||||
self.dW_only = dW_only
|
||||
self.dX_only = dX_only
|
||||
|
||||
@classmethod
|
||||
def from_hf(
|
||||
cls,
|
||||
moe_block: Qwen3MoeSparseMoeBlock,
|
||||
permute_x: bool = True,
|
||||
permute_y: bool = True,
|
||||
autotune: bool = True,
|
||||
kernel_config_fwd: KernelConfigForward = None,
|
||||
kernel_config_bwd_dW: KernelConfigBackward_dW = None,
|
||||
kernel_config_bwd_dX: KernelConfigBackward_dX = None,
|
||||
dW_only: bool = False,
|
||||
dX_only: bool = False,
|
||||
):
|
||||
config: Qwen3MoeConfig = moe_block.experts[0].config
|
||||
gate, gate_up_proj, down_proj = Qwen3MoeGroupedGEMMBlock.extract_hf_weights(
|
||||
moe_block
|
||||
)
|
||||
return cls(
|
||||
config,
|
||||
gate,
|
||||
gate_up_proj,
|
||||
down_proj,
|
||||
permute_x=permute_x,
|
||||
permute_y=permute_y,
|
||||
autotune=autotune,
|
||||
kernel_config_fwd=kernel_config_fwd,
|
||||
kernel_config_bwd_dW=kernel_config_bwd_dW,
|
||||
kernel_config_bwd_dX=kernel_config_bwd_dX,
|
||||
dW_only=dW_only,
|
||||
dX_only=dX_only,
|
||||
)
|
||||
|
||||
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
|
||||
batch_size, sequence_length, hidden_dim = hidden_states.shape
|
||||
num_tokens = batch_size * sequence_length
|
||||
total_tokens = num_tokens * self.top_k
|
||||
|
||||
hidden_states = hidden_states.view(-1, hidden_dim)
|
||||
|
||||
router_logits, routing_weights, selected_experts = self.run_router(
|
||||
hidden_states
|
||||
)
|
||||
# Pre-processing
|
||||
# 1. Compute tokens per expert and indices for gathering tokes from token order to expert order
|
||||
# NOTE: these are auxiliary data structs which don't need to be recorded in autograd graph
|
||||
token_counts_by_expert, gather_indices = (
|
||||
self.get_token_counts_and_gather_indices(selected_experts)
|
||||
)
|
||||
|
||||
# 2. permute_x -> permutation will be fused in prologue of first grouped gemm
|
||||
if not self.permute_x:
|
||||
hidden_states = permute(hidden_states, gather_indices, self.top_k)
|
||||
# Start expert computation
|
||||
hidden_states = grouped_gemm(
|
||||
X=hidden_states,
|
||||
W=self.gate_up_proj,
|
||||
m_sizes=token_counts_by_expert,
|
||||
gather_indices=gather_indices,
|
||||
topk=self.top_k,
|
||||
permute_x=self.permute_x,
|
||||
permute_y=False, # output of first grouped gemm should never be permuted
|
||||
autotune=self.autotune,
|
||||
kernel_config_fwd=self.kernel_config_fwd,
|
||||
kernel_config_bwd_dW=self.kernel_config_bwd_dW,
|
||||
kernel_config_bwd_dX=self.kernel_config_bwd_dX,
|
||||
is_first_gemm=True,
|
||||
dW_only=self.dW_only,
|
||||
dX_only=self.dX_only,
|
||||
)
|
||||
hidden_states = self.act_and_mul(hidden_states)
|
||||
hidden_states = grouped_gemm(
|
||||
X=hidden_states,
|
||||
W=self.down_proj,
|
||||
m_sizes=token_counts_by_expert,
|
||||
gather_indices=gather_indices,
|
||||
topk=self.top_k,
|
||||
permute_x=False,
|
||||
permute_y=self.permute_y,
|
||||
autotune=self.autotune,
|
||||
kernel_config_fwd=self.kernel_config_fwd,
|
||||
kernel_config_bwd_dW=self.kernel_config_bwd_dW,
|
||||
kernel_config_bwd_dX=self.kernel_config_bwd_dX,
|
||||
is_first_gemm=False,
|
||||
dW_only=self.dW_only,
|
||||
dX_only=self.dX_only,
|
||||
)
|
||||
|
||||
# Post-processing
|
||||
# 1. Unpermute from expert order to token order
|
||||
if not self.permute_y:
|
||||
hidden_states = unpermute(hidden_states, gather_indices)
|
||||
|
||||
# 2. Merge topk weights
|
||||
hidden_states = (
|
||||
hidden_states.view(num_tokens, self.top_k, hidden_dim)
|
||||
* routing_weights[..., None]
|
||||
)
|
||||
hidden_states = hidden_states.sum(dim=1)
|
||||
|
||||
hidden_states = hidden_states.view(batch_size, sequence_length, hidden_dim)
|
||||
return hidden_states, router_logits
|
||||
335
unsloth/kernels/moe/grouped_gemm/reference/moe_ops.py
Normal file
335
unsloth/kernels/moe/grouped_gemm/reference/moe_ops.py
Normal file
|
|
@ -0,0 +1,335 @@
|
|||
from dataclasses import dataclass
|
||||
from typing import Tuple
|
||||
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
import torch.nn.functional as F
|
||||
from transformers.models.qwen3_moe import Qwen3MoeConfig
|
||||
from transformers.models.qwen3_moe.modeling_qwen3_moe import (
|
||||
ACT2FN,
|
||||
Qwen3MoeSparseMoeBlock,
|
||||
)
|
||||
|
||||
|
||||
def permute(X: torch.Tensor, gather_indices: torch.Tensor, topk: int):
|
||||
"""
|
||||
Scatters X to a new tensor with shape [total_tokens, hidden_dim] where total_tokens is num_tokens * topk,
|
||||
permuting the tokens according to sorted_token_idx.
|
||||
|
||||
Helper for grouped gemm where hidden states need be ordered by expert.
|
||||
X: [num_tokens, hidden_dim]
|
||||
sorted_token_idx: [num_tokens * topk]
|
||||
topk: int
|
||||
|
||||
Returns:
|
||||
[total_tokens, hidden_dim]
|
||||
"""
|
||||
assert gather_indices.ndim == 1
|
||||
X = X.view(-1, X.shape[-1])
|
||||
# Shortcut for topk == 1
|
||||
if topk == 1:
|
||||
return X[gather_indices]
|
||||
|
||||
return X[gather_indices // topk]
|
||||
|
||||
|
||||
def unpermute(X: torch.Tensor, gather_indices: torch.Tensor):
|
||||
X = X.view(-1, X.shape[-1]) if X.ndim > 2 else X
|
||||
unpermuted = torch.empty_like(X)
|
||||
unpermuted.index_copy_(0, gather_indices, X)
|
||||
return unpermuted.view_as(X)
|
||||
|
||||
|
||||
def calculate_topk(
|
||||
gating_output: torch.Tensor,
|
||||
top_k: int,
|
||||
use_sigmoid: bool,
|
||||
renormalize: bool,
|
||||
pre_act: bool = True,
|
||||
post_act: bool = False,
|
||||
):
|
||||
"""
|
||||
If post_act is True, then activation function is run AFTER topk
|
||||
If post_act is False, then activation function is run BEFORE topk
|
||||
|
||||
This is to align with triton_bench implementation (post_act) whereas most models use pre_act (e.g. llama4, deepseek)
|
||||
"""
|
||||
assert pre_act ^ post_act, "only one of pre_act or post_act can be True"
|
||||
|
||||
def _activation(gating_output: torch.Tensor):
|
||||
if use_sigmoid:
|
||||
scores = torch.sigmoid(gating_output.to(torch.float32)).to(
|
||||
gating_output.dtype
|
||||
)
|
||||
else:
|
||||
scores = F.softmax(gating_output.to(torch.float32), dim=1).to(
|
||||
gating_output.dtype
|
||||
)
|
||||
|
||||
return scores
|
||||
|
||||
if pre_act:
|
||||
scores = _activation(gating_output)
|
||||
else:
|
||||
scores = gating_output
|
||||
|
||||
topk_weights, topk_ids = torch.topk(scores, k=top_k, dim=1)
|
||||
|
||||
if post_act:
|
||||
topk_weights = _activation(topk_weights)
|
||||
|
||||
if renormalize:
|
||||
topk_weights /= torch.sum(topk_weights, dim=-1, keepdim=True).to(
|
||||
gating_output.dtype
|
||||
)
|
||||
|
||||
return topk_weights, topk_ids
|
||||
|
||||
|
||||
@torch.no_grad()
|
||||
def get_routing_indices(
|
||||
selected_experts, num_experts, return_scatter_indices: bool = False
|
||||
):
|
||||
"""
|
||||
Returns:
|
||||
token_counts_by_expert: [num_experts]
|
||||
gather_indices: [num_tokens]
|
||||
scatter_indices [Optional] (torch.Tensor):
|
||||
Indices for unpermuting gathered inputs back to token order, shape ``(bs * seqlen * top_k,)``.
|
||||
"""
|
||||
# group tokens together by expert indices from 0 to num_experts and pass that to experts forward
|
||||
token_counts_by_expert = torch.histc(
|
||||
selected_experts.view(-1),
|
||||
bins=num_experts,
|
||||
min=0,
|
||||
max=num_experts,
|
||||
)
|
||||
# token_indices_experts_sorted shape (bs*slen*top_k,)
|
||||
gather_indices = torch.argsort(selected_experts.view(-1), stable=True)
|
||||
if return_scatter_indices:
|
||||
scatter_indices = gather_indices.argsort()
|
||||
return token_counts_by_expert, gather_indices, scatter_indices
|
||||
else:
|
||||
return token_counts_by_expert, gather_indices
|
||||
|
||||
|
||||
def torch_grouped_gemm(X, W, m_sizes, transpose=True):
|
||||
"""
|
||||
X: [M, K] if forward, else [M, N]
|
||||
W: [E, N, K]
|
||||
m_sizes: [E]
|
||||
|
||||
Returns:
|
||||
Y: [M, N] if forward, else [M, K]
|
||||
"""
|
||||
X = X.view(-1, X.shape[-1])
|
||||
M, K = X.shape
|
||||
|
||||
assert m_sizes.ndim == 1
|
||||
E = m_sizes.shape[0]
|
||||
|
||||
assert W.ndim == 3
|
||||
assert W.shape[0] == E
|
||||
|
||||
N = W.shape[1]
|
||||
|
||||
result = torch.zeros((M, N), dtype=X.dtype, device=X.device)
|
||||
|
||||
m_start = 0
|
||||
for g in range(E):
|
||||
m_size = m_sizes[g]
|
||||
if m_size > 0:
|
||||
m_end = m_start + m_size
|
||||
|
||||
# Extract group input
|
||||
# m_size x K
|
||||
X_g = X[m_start:m_end]
|
||||
# N x K
|
||||
W_g = W[g]
|
||||
|
||||
# Y_g = X_g @ W_g.T -> [m_size, N]
|
||||
W_g = W_g.T if transpose else W_g
|
||||
Y_g = X_g @ W_g
|
||||
|
||||
result[m_start:m_end] = Y_g
|
||||
|
||||
m_start = m_end
|
||||
return result
|
||||
|
||||
|
||||
@dataclass
|
||||
class GroupedGEMMResult:
|
||||
token_counts_by_expert: torch.Tensor
|
||||
gather_indices: torch.Tensor
|
||||
topk_weights: torch.Tensor
|
||||
first_gemm: torch.Tensor
|
||||
intermediate: torch.Tensor
|
||||
second_gemm: torch.Tensor
|
||||
hidden_states_unpermute: torch.Tensor
|
||||
hidden_states: torch.Tensor # final output
|
||||
|
||||
|
||||
class Qwen3MoeGroupedGEMMBlock(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
config,
|
||||
gate: torch.Tensor,
|
||||
gate_up_proj: torch.Tensor,
|
||||
down_proj: torch.Tensor,
|
||||
):
|
||||
super().__init__()
|
||||
self.num_experts = config.num_experts
|
||||
self.top_k = config.num_experts_per_tok
|
||||
self.norm_topk_prob = config.norm_topk_prob
|
||||
self.hidden_size = config.hidden_size
|
||||
self.moe_intermediate_size = config.moe_intermediate_size
|
||||
|
||||
assert gate.shape == (config.num_experts, config.hidden_size)
|
||||
assert gate_up_proj.shape == (
|
||||
config.num_experts,
|
||||
2 * config.moe_intermediate_size,
|
||||
config.hidden_size,
|
||||
)
|
||||
assert down_proj.shape == (
|
||||
config.num_experts,
|
||||
config.hidden_size,
|
||||
config.moe_intermediate_size,
|
||||
)
|
||||
|
||||
# gating
|
||||
self.gate = torch.nn.Parameter(gate)
|
||||
|
||||
# experts
|
||||
self.gate_up_proj = torch.nn.Parameter(gate_up_proj, requires_grad=True)
|
||||
self.down_proj = torch.nn.Parameter(down_proj, requires_grad=True)
|
||||
self.act_fn = ACT2FN[config.hidden_act]
|
||||
|
||||
@staticmethod
|
||||
def extract_hf_weights(moe_block: Qwen3MoeSparseMoeBlock):
|
||||
config: Qwen3MoeConfig = moe_block.experts[0].config
|
||||
num_experts = config.num_experts
|
||||
|
||||
gate = moe_block.gate.weight.data
|
||||
gate_proj = torch.stack(
|
||||
[moe_block.experts[i].gate_proj.weight.data for i in range(num_experts)],
|
||||
dim=0,
|
||||
)
|
||||
up_proj = torch.stack(
|
||||
[moe_block.experts[i].up_proj.weight.data for i in range(num_experts)],
|
||||
dim=0,
|
||||
)
|
||||
down_proj = torch.stack(
|
||||
[moe_block.experts[i].down_proj.weight.data for i in range(num_experts)],
|
||||
dim=0,
|
||||
)
|
||||
gate_up_proj = torch.cat([gate_proj, up_proj], dim=1)
|
||||
return gate, gate_up_proj, down_proj
|
||||
|
||||
@classmethod
|
||||
def from_hf(cls, moe_block: Qwen3MoeSparseMoeBlock):
|
||||
config: Qwen3MoeConfig = moe_block.experts[0].config
|
||||
gate, gate_up_proj, down_proj = cls.extract_hf_weights(moe_block)
|
||||
return cls(config, gate, gate_up_proj, down_proj)
|
||||
|
||||
def check_weights(self, moe_block: Qwen3MoeSparseMoeBlock):
|
||||
for i in range(self.num_experts):
|
||||
assert self.gate_up_proj[i].equal(
|
||||
torch.cat(
|
||||
[
|
||||
moe_block.experts[i].gate_proj.weight.data,
|
||||
moe_block.experts[i].up_proj.weight.data,
|
||||
],
|
||||
dim=0,
|
||||
)
|
||||
)
|
||||
assert self.down_proj[i].equal(moe_block.experts[i].down_proj.weight.data)
|
||||
|
||||
def act_and_mul(self, x: torch.Tensor) -> torch.Tensor:
|
||||
assert x.shape[-1] == 2 * self.moe_intermediate_size
|
||||
gate_proj = x[..., : self.moe_intermediate_size]
|
||||
up_proj = x[..., self.moe_intermediate_size :]
|
||||
return self.act_fn(gate_proj) * up_proj
|
||||
|
||||
def run_router(self, hidden_states: torch.Tensor) -> torch.Tensor:
|
||||
# router_logits: (batch * sequence_length, n_experts)
|
||||
router_logits = torch.nn.functional.linear(hidden_states, self.gate)
|
||||
|
||||
routing_weights = F.softmax(router_logits, dim=1, dtype=torch.float)
|
||||
routing_weights, selected_experts = torch.topk(
|
||||
routing_weights, self.top_k, dim=-1
|
||||
)
|
||||
if self.norm_topk_prob: # only diff with mixtral sparse moe block!
|
||||
routing_weights /= routing_weights.sum(dim=-1, keepdim=True)
|
||||
# we cast back to the input dtype
|
||||
routing_weights = routing_weights.to(hidden_states.dtype)
|
||||
|
||||
return router_logits, routing_weights, selected_experts
|
||||
|
||||
def get_token_counts_and_gather_indices(
|
||||
self, selected_experts: torch.Tensor
|
||||
) -> Tuple[torch.Tensor, torch.Tensor]:
|
||||
token_counts_by_expert, gather_indices = get_routing_indices(
|
||||
selected_experts, self.num_experts
|
||||
)
|
||||
assert not token_counts_by_expert.requires_grad
|
||||
assert not gather_indices.requires_grad
|
||||
return token_counts_by_expert, gather_indices
|
||||
|
||||
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
|
||||
""" """
|
||||
batch_size, sequence_length, hidden_dim = hidden_states.shape
|
||||
num_tokens = batch_size * sequence_length
|
||||
total_tokens = num_tokens * self.top_k
|
||||
|
||||
hidden_states = hidden_states.view(-1, hidden_dim)
|
||||
|
||||
router_logits, routing_weights, selected_experts = self.run_router(
|
||||
hidden_states
|
||||
)
|
||||
|
||||
# 1. Compute tokens per expert and indices for gathering tokes from token order to expert order
|
||||
# NOTE: these are auxiliary data structs which don't need to be recorded in autograd graph
|
||||
token_counts_by_expert, gather_indices = (
|
||||
self.get_token_counts_and_gather_indices(selected_experts)
|
||||
)
|
||||
|
||||
# 2. Permute tokens from token order to expert order
|
||||
hidden_states = permute(hidden_states, gather_indices, self.top_k)
|
||||
assert hidden_states.shape == (total_tokens, hidden_dim)
|
||||
|
||||
# Start expert computation
|
||||
first_gemm = torch_grouped_gemm(
|
||||
X=hidden_states, W=self.gate_up_proj, m_sizes=token_counts_by_expert
|
||||
)
|
||||
assert first_gemm.shape == (total_tokens, 2 * self.moe_intermediate_size)
|
||||
intermediate = self.act_and_mul(first_gemm)
|
||||
assert intermediate.shape == (total_tokens, self.moe_intermediate_size)
|
||||
second_gemm = torch_grouped_gemm(
|
||||
X=intermediate, W=self.down_proj, m_sizes=token_counts_by_expert
|
||||
)
|
||||
assert second_gemm.shape == (total_tokens, hidden_dim)
|
||||
|
||||
# Post-processing
|
||||
# 1. Unpermute from expert order to token order
|
||||
hidden_states_unpermute = unpermute(second_gemm, gather_indices)
|
||||
assert hidden_states_unpermute.shape == (total_tokens, hidden_dim)
|
||||
|
||||
# 2. Merge topk weights
|
||||
hidden_states = (
|
||||
hidden_states_unpermute.view(num_tokens, self.top_k, hidden_dim)
|
||||
* routing_weights[..., None]
|
||||
)
|
||||
hidden_states = hidden_states.sum(dim=1)
|
||||
assert hidden_states.shape == (num_tokens, hidden_dim)
|
||||
|
||||
hidden_states = hidden_states.view(batch_size, sequence_length, hidden_dim)
|
||||
return GroupedGEMMResult(
|
||||
token_counts_by_expert=token_counts_by_expert,
|
||||
gather_indices=gather_indices,
|
||||
topk_weights=routing_weights,
|
||||
first_gemm=first_gemm,
|
||||
intermediate=intermediate,
|
||||
second_gemm=second_gemm,
|
||||
hidden_states_unpermute=hidden_states_unpermute,
|
||||
hidden_states=hidden_states,
|
||||
), router_logits
|
||||
5
unsloth/kernels/moe/requirements.txt
Normal file
5
unsloth/kernels/moe/requirements.txt
Normal file
|
|
@ -0,0 +1,5 @@
|
|||
torch
|
||||
git+https://github.com/huggingface/transformers.git@main
|
||||
pytest
|
||||
pandas
|
||||
ruff
|
||||
0
unsloth/kernels/moe/tests/__init__.py
Normal file
0
unsloth/kernels/moe/tests/__init__.py
Normal file
333
unsloth/kernels/moe/tests/common.py
Normal file
333
unsloth/kernels/moe/tests/common.py
Normal file
|
|
@ -0,0 +1,333 @@
|
|||
import itertools
|
||||
from contextlib import contextmanager
|
||||
from dataclasses import dataclass, field
|
||||
|
||||
import torch
|
||||
|
||||
from grouped_gemm.kernels.tuning import (
|
||||
KernelConfig,
|
||||
KernelConfigBackward_dW,
|
||||
KernelConfigBackward_dX,
|
||||
KernelConfigForward,
|
||||
prune_kernel_configs_backward_dW,
|
||||
prune_kernel_configs_backward_dX,
|
||||
prune_kernel_configs_fwd,
|
||||
)
|
||||
|
||||
|
||||
def print_delimiter(char="-", length=80):
|
||||
print(char * length)
|
||||
|
||||
|
||||
@contextmanager
|
||||
def delimiter_context():
|
||||
print_delimiter()
|
||||
yield
|
||||
print_delimiter()
|
||||
|
||||
|
||||
def make_inputs(M, N, K, E, topk, dtype, requires_grad=False):
|
||||
X1 = (
|
||||
torch.randn((M, K), device="cuda", dtype=dtype, requires_grad=requires_grad)
|
||||
/ 10
|
||||
)
|
||||
X2 = (
|
||||
torch.randn(
|
||||
(M * topk, N), device="cuda", dtype=dtype, requires_grad=requires_grad
|
||||
)
|
||||
/ 10
|
||||
)
|
||||
W1 = (
|
||||
torch.randn(
|
||||
(E, 2 * N, K), device="cuda", dtype=dtype, requires_grad=requires_grad
|
||||
)
|
||||
/ 10
|
||||
)
|
||||
W2 = (
|
||||
torch.randn((E, K, N), device="cuda", dtype=dtype, requires_grad=requires_grad)
|
||||
/ 10
|
||||
)
|
||||
score = torch.randn((M, E), device="cuda", dtype=dtype, requires_grad=requires_grad)
|
||||
if requires_grad:
|
||||
X1.retain_grad()
|
||||
X2.retain_grad()
|
||||
W1.retain_grad()
|
||||
W2.retain_grad()
|
||||
score.retain_grad()
|
||||
return X1, X2, W1, W2, score
|
||||
|
||||
|
||||
@dataclass(kw_only=True)
|
||||
class DataConfig:
|
||||
seq_len: int
|
||||
dtype: torch.dtype
|
||||
device: str = "cuda"
|
||||
bs: int = 1
|
||||
|
||||
|
||||
@dataclass(kw_only=True)
|
||||
class ModelConfig:
|
||||
hidden_size: int
|
||||
intermediate_size: int
|
||||
num_experts: int
|
||||
topk: int
|
||||
use_sigmoid: bool
|
||||
renormalize: bool
|
||||
pre_mul: bool = False
|
||||
post_mul: bool = field(init=False)
|
||||
|
||||
def __post_init__(self):
|
||||
self.post_mul = not self.pre_mul
|
||||
|
||||
|
||||
@dataclass(kw_only=True)
|
||||
class GroupedGEMMTestConfig:
|
||||
name: str = "test"
|
||||
data_config: DataConfig
|
||||
model_config: ModelConfig
|
||||
|
||||
|
||||
TOLERANCE = {
|
||||
torch.bfloat16: (1e-3, 1e-3),
|
||||
torch.float16: (1e-4, 1e-4),
|
||||
torch.float32: (1e-5, 1e-5),
|
||||
}
|
||||
|
||||
|
||||
# from https://github.com/triton-lang/triton/blob/main/bench/triton_bench/testing.py
|
||||
def assert_equal(ref, tri):
|
||||
if isinstance(ref, torch.Tensor):
|
||||
assert torch.all(ref == tri), f"tensors not equal {ref} != {tri}"
|
||||
else:
|
||||
assert ref == tri, f"ref not equal to tri {ref} != {tri}"
|
||||
|
||||
|
||||
def assert_close(ref, tri, maxtol=None, rmstol=None, description="--", verbose=True):
|
||||
if tri.dtype.itemsize == 1:
|
||||
ref_as_type = ref.to(tri.dtype)
|
||||
if ref.dtype == tri.dtype:
|
||||
assert torch.all(ref_as_type == tri)
|
||||
return
|
||||
ref = ref_as_type
|
||||
|
||||
if maxtol is None:
|
||||
maxtol = 2e-2
|
||||
if rmstol is None:
|
||||
rmstol = 4e-3
|
||||
"""
|
||||
Compare reference values against obtained values.
|
||||
"""
|
||||
|
||||
# cast to float32:
|
||||
ref = ref.to(torch.float32).detach()
|
||||
tri = tri.to(torch.float32).detach()
|
||||
assert ref.shape == tri.shape, (
|
||||
f"Tensors must have same size {ref.shape=} {tri.shape=}"
|
||||
)
|
||||
|
||||
# deal with infinite elements:
|
||||
inf_mask_ref = torch.isinf(ref)
|
||||
inf_mask_tri = torch.isinf(tri)
|
||||
assert torch.equal(inf_mask_ref, inf_mask_tri), (
|
||||
"Tensor must have same infinite elements"
|
||||
)
|
||||
refn = torch.where(inf_mask_ref, 0, ref)
|
||||
trin = torch.where(inf_mask_tri, 0, tri)
|
||||
|
||||
# normalise so that RMS calculation doesn't overflow:
|
||||
eps = 1.0e-30
|
||||
multiplier = 1.0 / (torch.max(torch.abs(refn)) + eps)
|
||||
refn *= multiplier
|
||||
trin *= multiplier
|
||||
|
||||
ref_rms = torch.sqrt(torch.square(refn).mean()) + eps
|
||||
|
||||
rel_err = torch.abs(refn - trin) / torch.maximum(ref_rms, torch.abs(refn))
|
||||
max_err = torch.max(rel_err).item()
|
||||
rms_err = torch.sqrt(torch.square(rel_err).mean()).item()
|
||||
|
||||
if verbose:
|
||||
print(
|
||||
"%s maximum relative error = %s (threshold = %s)"
|
||||
% (description, max_err, maxtol)
|
||||
)
|
||||
print(
|
||||
"%s RMS relative error = %s (threshold = %s)"
|
||||
% (description, rms_err, rmstol)
|
||||
)
|
||||
|
||||
if max_err > maxtol:
|
||||
bad_idxs = torch.nonzero(rel_err > maxtol)
|
||||
num_nonzero = bad_idxs.size(0)
|
||||
bad_idxs = bad_idxs[:1000]
|
||||
print(
|
||||
"%d / %d mismatched elements (shape = %s) at coords %s"
|
||||
% (num_nonzero, rel_err.numel(), tuple(rel_err.shape), bad_idxs.tolist())
|
||||
)
|
||||
|
||||
bad_idxs = bad_idxs.unbind(-1)
|
||||
print("ref values: ", ref[*bad_idxs].cpu())
|
||||
print("tri values: ", tri[*bad_idxs].cpu())
|
||||
|
||||
assert max_err <= maxtol
|
||||
assert rms_err <= rmstol
|
||||
|
||||
|
||||
def assert_indx_equal(ref, tri):
|
||||
assert_equal(ref, tri[: len(ref)])
|
||||
assert torch.all(tri[len(ref) :] == -1)
|
||||
|
||||
|
||||
def get_kernel_test_configs(
|
||||
BLOCK_SIZE_M=32,
|
||||
BLOCK_SIZE_N=32,
|
||||
BLOCK_SIZE_K=32,
|
||||
num_warps=4,
|
||||
num_stages=2,
|
||||
) -> list[KernelConfig]:
|
||||
configs_fwd = []
|
||||
configs_bwd_dX = []
|
||||
configs_bwd_dW = []
|
||||
|
||||
for permute_x in [False, True]:
|
||||
for permute_y in [False, True]:
|
||||
for use_tma_load_w in [True, False]:
|
||||
for use_tma_load_x in [True, False]:
|
||||
for use_tma_store in [True, False]:
|
||||
configs_fwd.append(
|
||||
KernelConfigForward(
|
||||
BLOCK_SIZE_M=BLOCK_SIZE_M,
|
||||
BLOCK_SIZE_N=BLOCK_SIZE_N,
|
||||
BLOCK_SIZE_K=BLOCK_SIZE_K,
|
||||
num_warps=num_warps,
|
||||
num_stages=num_stages,
|
||||
use_tma_load_w=use_tma_load_w,
|
||||
use_tma_load_x=use_tma_load_x,
|
||||
use_tma_store=use_tma_store,
|
||||
permute_x=permute_x,
|
||||
permute_y=permute_y,
|
||||
)
|
||||
)
|
||||
configs_bwd_dX.append(
|
||||
KernelConfigBackward_dX(
|
||||
BLOCK_SIZE_M=BLOCK_SIZE_M,
|
||||
BLOCK_SIZE_N=BLOCK_SIZE_N,
|
||||
BLOCK_SIZE_K=BLOCK_SIZE_K,
|
||||
num_warps=num_warps,
|
||||
num_stages=num_stages,
|
||||
use_tma_load_dy=use_tma_load_x,
|
||||
use_tma_load_w=use_tma_load_w,
|
||||
permute_x=permute_x,
|
||||
permute_y=permute_y,
|
||||
use_tma_store=use_tma_store,
|
||||
)
|
||||
)
|
||||
configs_bwd_dW.append(
|
||||
KernelConfigBackward_dW(
|
||||
BLOCK_SIZE_M=BLOCK_SIZE_M,
|
||||
BLOCK_SIZE_N=BLOCK_SIZE_N,
|
||||
BLOCK_SIZE_K=BLOCK_SIZE_K,
|
||||
num_warps=num_warps,
|
||||
num_stages=num_stages,
|
||||
use_tma_load_dy=use_tma_load_w,
|
||||
use_tma_load_x=use_tma_load_x,
|
||||
permute_x=permute_x,
|
||||
permute_y=permute_y,
|
||||
use_tma_store=use_tma_store,
|
||||
)
|
||||
)
|
||||
configs_fwd = prune_kernel_configs_fwd(configs_fwd)
|
||||
configs_bwd_dX = prune_kernel_configs_backward_dX(configs_bwd_dX)
|
||||
configs_bwd_dW = prune_kernel_configs_backward_dW(configs_bwd_dW)
|
||||
return configs_fwd, configs_bwd_dX, configs_bwd_dW
|
||||
|
||||
|
||||
def remove_feature_flags(
|
||||
kernel_configs: list[KernelConfig],
|
||||
permute_x: bool = True,
|
||||
permute_y: bool = True,
|
||||
tma_loads: bool = True,
|
||||
tma_store: bool = True,
|
||||
):
|
||||
pruned_configs = []
|
||||
for config in kernel_configs:
|
||||
# Remove permute flags first:
|
||||
if permute_x and config.permute_x:
|
||||
continue
|
||||
if permute_y and config.permute_y:
|
||||
continue
|
||||
if tma_loads:
|
||||
if isinstance(config, KernelConfigForward):
|
||||
if config.use_tma_load_w or config.use_tma_load_x:
|
||||
continue
|
||||
if isinstance(config, KernelConfigBackward_dX):
|
||||
if config.use_tma_load_dy or config.use_tma_load_w:
|
||||
continue
|
||||
if isinstance(config, KernelConfigBackward_dW):
|
||||
if config.use_tma_load_dy or config.use_tma_load_x:
|
||||
continue
|
||||
if tma_store:
|
||||
if config.use_tma_store:
|
||||
continue
|
||||
pruned_configs.append(config)
|
||||
return pruned_configs
|
||||
|
||||
|
||||
# Test Configs
|
||||
|
||||
TOPK = [1, 4]
|
||||
NUM_EXPERTS = [4, 16]
|
||||
|
||||
TEST_MODEL_SIZES = [
|
||||
(32, 32), # Debug
|
||||
(128, 128), # Small
|
||||
(512, 512), # Medium
|
||||
]
|
||||
|
||||
SMALL_MODEL_CONFIGS = [
|
||||
ModelConfig(
|
||||
topk=topk,
|
||||
num_experts=num_experts,
|
||||
hidden_size=model_size[0],
|
||||
intermediate_size=model_size[1],
|
||||
use_sigmoid=False,
|
||||
renormalize=False,
|
||||
)
|
||||
for topk, num_experts, model_size in itertools.product(
|
||||
TOPK, NUM_EXPERTS, TEST_MODEL_SIZES
|
||||
)
|
||||
]
|
||||
LLAMA_MODEL_CONFIG = ModelConfig(
|
||||
topk=1,
|
||||
num_experts=16,
|
||||
hidden_size=5120,
|
||||
intermediate_size=8192,
|
||||
use_sigmoid=True,
|
||||
renormalize=False,
|
||||
)
|
||||
QWEN_MODEL_CONFIG = ModelConfig(
|
||||
topk=8,
|
||||
num_experts=128,
|
||||
hidden_size=2048,
|
||||
intermediate_size=768,
|
||||
use_sigmoid=False,
|
||||
renormalize=False,
|
||||
)
|
||||
|
||||
SEQLENS = [128, 1024]
|
||||
DTYPE = [torch.bfloat16]
|
||||
|
||||
DATA_CONFIGS = [
|
||||
DataConfig(seq_len=seq_len, dtype=dtype)
|
||||
for seq_len, dtype in itertools.product(SEQLENS, DTYPE)
|
||||
]
|
||||
KERNEL_CONFIGS_FWD, KERNEL_CONFIGS_BWD_dX, KERNEL_CONFIGS_BWD_dW = (
|
||||
get_kernel_test_configs()
|
||||
)
|
||||
|
||||
if __name__ == "__main__":
|
||||
print(
|
||||
KERNEL_CONFIGS_BWD_dX[0].to_string(
|
||||
include_tuning_params=False, include_tma=False
|
||||
)
|
||||
)
|
||||
505
unsloth/kernels/moe/tests/moe_utils.py
Normal file
505
unsloth/kernels/moe/tests/moe_utils.py
Normal file
|
|
@ -0,0 +1,505 @@
|
|||
from dataclasses import dataclass, fields
|
||||
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
from huggingface_hub import HfApi
|
||||
from huggingface_hub.utils import _safetensors
|
||||
from transformers.models.qwen3_moe.configuration_qwen3_moe import Qwen3MoeConfig
|
||||
from transformers.models.qwen3_moe.modeling_qwen3_moe import Qwen3MoeSparseMoeBlock
|
||||
|
||||
from grouped_gemm.interface import grouped_gemm
|
||||
from grouped_gemm.kernels.tuning import (
|
||||
KernelConfigBackward_dW,
|
||||
KernelConfigBackward_dX,
|
||||
KernelConfigForward,
|
||||
)
|
||||
from grouped_gemm.reference.moe_ops import (
|
||||
GroupedGEMMResult,
|
||||
Qwen3MoeGroupedGEMMBlock,
|
||||
permute,
|
||||
unpermute,
|
||||
)
|
||||
|
||||
|
||||
def rebind_experts_to_shared_buffer(
|
||||
moe_block: Qwen3MoeSparseMoeBlock, config: Qwen3MoeConfig
|
||||
):
|
||||
num_experts = config.num_experts
|
||||
hidden_size = config.hidden_size
|
||||
interm_size = config.moe_intermediate_size
|
||||
device = moe_block.experts[0].down_proj.weight.device
|
||||
dtype = moe_block.experts[0].down_proj.weight.dtype
|
||||
|
||||
buffer_up = torch.empty(
|
||||
num_experts, interm_size, hidden_size, device=device, dtype=dtype
|
||||
)
|
||||
buffer_gate = torch.empty(
|
||||
num_experts, interm_size, hidden_size, device=device, dtype=dtype
|
||||
)
|
||||
buffer_down = torch.empty(
|
||||
num_experts, hidden_size, interm_size, device=device, dtype=dtype
|
||||
)
|
||||
|
||||
# Step 2: Copy existing expert weights into buffers
|
||||
for i, expert in enumerate(moe_block.experts):
|
||||
buffer_up[i].copy_(expert.up_proj.weight.data)
|
||||
buffer_gate[i].copy_(expert.gate_proj.weight.data)
|
||||
buffer_down[i].copy_(expert.down_proj.weight.data)
|
||||
|
||||
# Step 3: Rebind expert weights to views in shared buffer
|
||||
for i, expert in enumerate(moe_block.experts):
|
||||
expert.up_proj.weight = torch.nn.Parameter(buffer_up[i])
|
||||
expert.gate_proj.weight = torch.nn.Parameter(buffer_gate[i])
|
||||
expert.down_proj.weight = torch.nn.Parameter(buffer_down[i])
|
||||
|
||||
return buffer_up, buffer_gate, buffer_down
|
||||
|
||||
|
||||
def get_expert_metadata(model_id: str):
|
||||
api = HfApi()
|
||||
metadata: _safetensors.SafetensorsRepoMetadata = api.get_safetensors_metadata(
|
||||
model_id
|
||||
)
|
||||
return metadata.files_metadata
|
||||
|
||||
|
||||
def clone_experts(
|
||||
moe_block: Qwen3MoeSparseMoeBlock, config: Qwen3MoeConfig, copy: bool = True
|
||||
):
|
||||
down_projs = torch.empty(
|
||||
config.num_experts, config.hidden_size, config.moe_intermediate_size
|
||||
)
|
||||
up_projs = torch.empty(
|
||||
config.num_experts, config.moe_intermediate_size, config.hidden_size
|
||||
)
|
||||
gate_projs = torch.empty(
|
||||
config.num_experts, config.moe_intermediate_size, config.hidden_size
|
||||
)
|
||||
for expert_idx, expert in enumerate(moe_block.experts):
|
||||
down_projs[expert_idx].copy_(expert.down_proj.weight.data)
|
||||
up_projs[expert_idx].copy_(expert.up_proj.weight.data)
|
||||
gate_projs[expert_idx].copy_(expert.gate_proj.weight.data)
|
||||
return gate_projs, up_projs, down_projs
|
||||
|
||||
|
||||
@dataclass
|
||||
class ForwardResult:
|
||||
output: torch.Tensor
|
||||
router_logits: torch.Tensor
|
||||
X: torch.Tensor
|
||||
# When using grouped gemm MoE implementation to additional debugging / checking of intermediate results
|
||||
grouped_gemm_result: GroupedGEMMResult = None
|
||||
|
||||
|
||||
@dataclass
|
||||
class BackwardResult:
|
||||
X_grad: torch.Tensor
|
||||
gate_grad: torch.Tensor
|
||||
gate_proj_grad: torch.Tensor
|
||||
up_proj_grad: torch.Tensor
|
||||
down_proj_grad: torch.Tensor
|
||||
|
||||
|
||||
def check_down_proj_grad(
|
||||
moe_block: Qwen3MoeSparseMoeBlock,
|
||||
grouped_gemm_block: Qwen3MoeGroupedGEMMBlock,
|
||||
atol: float,
|
||||
rtol: float,
|
||||
):
|
||||
for i, expert in enumerate(moe_block.experts):
|
||||
ref_grad = expert.down_proj.weight.grad
|
||||
assert ref_grad is not None
|
||||
test_grad = grouped_gemm_block.down_proj.grad[i]
|
||||
assert test_grad is not None
|
||||
diff = (ref_grad - test_grad).abs().max()
|
||||
if not torch.allclose(ref_grad, test_grad, atol=atol, rtol=rtol):
|
||||
print(f"expert {i} down_proj_grad_diff: {diff.detach().cpu().item():.6f}")
|
||||
|
||||
|
||||
def check_gate_up_proj_grad(
|
||||
moe_block: Qwen3MoeSparseMoeBlock,
|
||||
grouped_gemm_block: Qwen3MoeGroupedGEMMBlock,
|
||||
atol: float,
|
||||
rtol: float,
|
||||
):
|
||||
moe_intermediate_size = grouped_gemm_block.moe_intermediate_size
|
||||
for i, expert in enumerate(moe_block.experts):
|
||||
ref_gate_proj_grad = expert.gate_proj.weight.grad
|
||||
ref_up_proj_grad = expert.up_proj.weight.grad
|
||||
assert ref_gate_proj_grad is not None
|
||||
assert ref_up_proj_grad is not None
|
||||
|
||||
# Extract gradients
|
||||
test_gate_proj_grad = grouped_gemm_block.gate_up_proj.grad[
|
||||
i, :moe_intermediate_size
|
||||
]
|
||||
test_up_proj_grad = grouped_gemm_block.gate_up_proj.grad[
|
||||
i, moe_intermediate_size:
|
||||
]
|
||||
assert test_gate_proj_grad is not None
|
||||
assert test_up_proj_grad is not None
|
||||
|
||||
# Sanity check shapes
|
||||
assert ref_gate_proj_grad.shape == test_gate_proj_grad.shape, (
|
||||
f"{ref_gate_proj_grad.shape} != {test_gate_proj_grad.shape}"
|
||||
)
|
||||
assert ref_up_proj_grad.shape == test_up_proj_grad.shape, (
|
||||
f"{ref_up_proj_grad.shape} != {test_up_proj_grad.shape}"
|
||||
)
|
||||
|
||||
# Check gradients
|
||||
diff = (ref_gate_proj_grad - test_gate_proj_grad).abs().max()
|
||||
if not torch.allclose(
|
||||
ref_gate_proj_grad, test_gate_proj_grad, atol=atol, rtol=rtol
|
||||
):
|
||||
print(f"expert {i} gate_proj_grad_diff: {diff.detach().cpu().item():.6f}")
|
||||
diff = (ref_up_proj_grad - test_up_proj_grad).abs().max()
|
||||
if not torch.allclose(
|
||||
ref_up_proj_grad, test_up_proj_grad, atol=atol, rtol=rtol
|
||||
):
|
||||
print(f"expert {i} up_proj_grad_diff: {diff.detach().cpu().item():.6f}")
|
||||
|
||||
|
||||
def check_gate_grad(
|
||||
moe_block: Qwen3MoeSparseMoeBlock,
|
||||
grouped_gemm_block: Qwen3MoeGroupedGEMMBlock,
|
||||
atol: float,
|
||||
rtol: float,
|
||||
):
|
||||
ref_grad = moe_block.gate.weight.grad
|
||||
assert ref_grad is not None
|
||||
test_grad = grouped_gemm_block.gate.grad
|
||||
assert test_grad is not None
|
||||
diff = (ref_grad - test_grad).abs().max()
|
||||
if not torch.allclose(ref_grad, test_grad, atol=atol, rtol=rtol):
|
||||
print(f"gate_grad_diff: {diff.detach().cpu().item():.6f}")
|
||||
|
||||
|
||||
def check_wgrad(
|
||||
moe_block: Qwen3MoeSparseMoeBlock,
|
||||
grouped_gemm_block: Qwen3MoeGroupedGEMMBlock,
|
||||
atol: float,
|
||||
rtol: float,
|
||||
):
|
||||
check_down_proj_grad(moe_block, grouped_gemm_block, atol, rtol)
|
||||
check_gate_up_proj_grad(moe_block, grouped_gemm_block, atol, rtol)
|
||||
check_gate_grad(moe_block, grouped_gemm_block, atol, rtol)
|
||||
|
||||
|
||||
def check_tensor_allclose(
|
||||
X_ref: torch.Tensor,
|
||||
X_test: torch.Tensor,
|
||||
atol: float,
|
||||
rtol: float,
|
||||
name: str,
|
||||
verbose: bool = False,
|
||||
):
|
||||
diff = (X_ref - X_test).abs().max()
|
||||
if verbose:
|
||||
print(f"{name} diff: {diff.detach().cpu().item():.6f}")
|
||||
assert torch.allclose(X_ref, X_test, atol=atol, rtol=rtol), (
|
||||
f"{name} diff: {diff.detach().cpu().item():.6f}"
|
||||
)
|
||||
|
||||
|
||||
def check_expert_grads(
|
||||
ref_result: BackwardResult,
|
||||
test_result: BackwardResult,
|
||||
atol: float,
|
||||
rtol: float,
|
||||
verbose: bool = False,
|
||||
):
|
||||
fields_to_check = [f.name for f in fields(BackwardResult) if "proj" in f.name]
|
||||
assert len(fields_to_check) == 3
|
||||
|
||||
for field in fields_to_check:
|
||||
ref_grads = getattr(ref_result, field)
|
||||
test_grads = getattr(test_result, field)
|
||||
assert ref_grads.shape == test_grads.shape, (
|
||||
f"{field}: {ref_grads.shape} != {test_grads.shape}"
|
||||
)
|
||||
|
||||
# Test each expert
|
||||
for i in range(ref_grads.shape[0]):
|
||||
ref_grad = ref_grads[i]
|
||||
test_grad = test_grads[i]
|
||||
diff = (ref_grad - test_grad).abs().max()
|
||||
assert torch.allclose(ref_grad, test_grad, atol=atol, rtol=rtol), (
|
||||
f"{field}[{i}] diff: {diff.detach().cpu().item():.6f}"
|
||||
)
|
||||
|
||||
# Test all experts
|
||||
diff = (ref_grads - test_grads).abs().max()
|
||||
if verbose:
|
||||
print(f"{field} diff: {diff.detach().cpu().item():.6f}")
|
||||
assert torch.allclose(ref_grads, test_grads, atol=atol, rtol=rtol), (
|
||||
f"{field} diff: {diff.detach().cpu().item():.6f}"
|
||||
)
|
||||
|
||||
|
||||
def check_grads(
|
||||
ref_result: BackwardResult,
|
||||
test_result: BackwardResult,
|
||||
atol: float,
|
||||
rtol: float,
|
||||
verbose: bool = False,
|
||||
):
|
||||
check_tensor_allclose(
|
||||
ref_result.X_grad, test_result.X_grad, atol, rtol, "X.grad", verbose
|
||||
)
|
||||
check_tensor_allclose(
|
||||
ref_result.gate_grad, test_result.gate_grad, atol, rtol, "gate.grad", verbose
|
||||
)
|
||||
check_expert_grads(ref_result, test_result, atol, rtol, verbose)
|
||||
|
||||
|
||||
def check_fwd(
|
||||
ref_result: ForwardResult,
|
||||
test_result: ForwardResult,
|
||||
atol: float,
|
||||
rtol: float,
|
||||
verbose: bool = False,
|
||||
):
|
||||
# First check hidden states (output)
|
||||
ref_output = ref_result.output
|
||||
test_output = test_result.output
|
||||
diff = (ref_output - test_output).abs().max()
|
||||
if verbose:
|
||||
print(f"output diff: {diff.detach().cpu().item():.6f}")
|
||||
assert torch.allclose(ref_output, test_output, atol=atol, rtol=rtol), (
|
||||
f"output diff: {diff.detach().cpu().item():.6f}"
|
||||
)
|
||||
|
||||
# Check router logits
|
||||
ref_router_logits = ref_result.router_logits
|
||||
test_router_logits = test_result.router_logits
|
||||
diff = (ref_router_logits - test_router_logits).abs().max()
|
||||
if verbose:
|
||||
print(f"router_logits diff: {diff.detach().cpu().item():.6f}")
|
||||
assert torch.allclose(
|
||||
ref_router_logits, test_router_logits, atol=atol, rtol=rtol
|
||||
), f"router_logits diff: {diff.detach().cpu().item():.6f}"
|
||||
|
||||
|
||||
def check_grouped_gemm_results(
|
||||
grouped_result: GroupedGEMMResult,
|
||||
fused_result: GroupedGEMMResult,
|
||||
permute_y: bool,
|
||||
atol: float,
|
||||
rtol: float,
|
||||
verbose: bool = False,
|
||||
):
|
||||
for field in fields(GroupedGEMMResult):
|
||||
ref_value = getattr(grouped_result, field.name)
|
||||
test_value = getattr(fused_result, field.name)
|
||||
diff = (ref_value - test_value).abs().max()
|
||||
|
||||
# second_gemm in torch grouped gemm is not yet unpermuted so comparing the fused unpermuted second_gemm will result in error
|
||||
# instead the hidden_states_unpermute should match since hidden_states_unpermute for the fused result is the same as second_gemm
|
||||
if field.name == "second_gemm" and permute_y:
|
||||
continue
|
||||
|
||||
if verbose:
|
||||
print(f"{field.name} diff: {diff.detach().cpu().item():.6f}")
|
||||
|
||||
assert torch.allclose(ref_value, test_value, atol=atol, rtol=rtol), (
|
||||
f"{field.name} diff: {diff.detach().cpu().item():.6f}"
|
||||
)
|
||||
|
||||
|
||||
def run_forward(model: nn.Module, X: torch.Tensor, is_grouped_gemm: bool = False):
|
||||
X = X.detach().clone().requires_grad_(True)
|
||||
output, router_logits = model(X)
|
||||
if is_grouped_gemm:
|
||||
result = ForwardResult(
|
||||
output=output.hidden_states,
|
||||
router_logits=router_logits,
|
||||
X=X,
|
||||
grouped_gemm_result=output,
|
||||
)
|
||||
else:
|
||||
result = ForwardResult(output=output, router_logits=router_logits, X=X)
|
||||
return result
|
||||
|
||||
|
||||
def run_backward(
|
||||
model: nn.Module, grad_output: torch.Tensor, output: torch.Tensor, X: torch.Tensor
|
||||
):
|
||||
output.backward(grad_output)
|
||||
assert X.grad is not None
|
||||
for name, param in model.named_parameters():
|
||||
assert param.grad is not None, f"{name} grad is None"
|
||||
if isinstance(model, Qwen3MoeSparseMoeBlock):
|
||||
gate_grad = model.gate.weight.grad
|
||||
gate_proj_grad = torch.stack(
|
||||
[expert.gate_proj.weight.grad for expert in model.experts]
|
||||
)
|
||||
up_proj_grad = torch.stack(
|
||||
[expert.up_proj.weight.grad for expert in model.experts]
|
||||
)
|
||||
down_proj_grad = torch.stack(
|
||||
[expert.down_proj.weight.grad for expert in model.experts]
|
||||
)
|
||||
elif isinstance(model, Qwen3MoeGroupedGEMMBlock):
|
||||
gate_grad = model.gate.grad
|
||||
gate_proj_grad, up_proj_grad = model.gate_up_proj.grad.chunk(2, dim=1)
|
||||
down_proj_grad = model.down_proj.grad
|
||||
else:
|
||||
raise ValueError(f"Unsupported model type: {type(model)}")
|
||||
return BackwardResult(
|
||||
X_grad=X.grad,
|
||||
gate_grad=gate_grad,
|
||||
gate_proj_grad=gate_proj_grad,
|
||||
up_proj_grad=up_proj_grad,
|
||||
down_proj_grad=down_proj_grad,
|
||||
)
|
||||
|
||||
|
||||
class Qwen3MoeFusedGroupedGEMMBlock(Qwen3MoeGroupedGEMMBlock):
|
||||
"""
|
||||
Reference implementation of MoE block using grouped gemm.
|
||||
|
||||
This is the same as the Qwen3MoeGroupedGEMMBlock but with triton grouped gemm in place of torch-native grouped gemm implementation.
|
||||
|
||||
NOTE: This is NOT to be used for production as it contains many extra checks and saves all intermediate results for debugging.
|
||||
See grouped_gemm/reference/moe_block.py for a cleaner implementation.
|
||||
"""
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
config: Qwen3MoeConfig,
|
||||
gate: torch.Tensor,
|
||||
gate_up_proj: torch.Tensor,
|
||||
down_proj: torch.Tensor,
|
||||
permute_x: bool = False,
|
||||
permute_y: bool = False,
|
||||
autotune: bool = True,
|
||||
kernel_config_fwd: KernelConfigForward = None,
|
||||
kernel_config_bwd_dW: KernelConfigBackward_dW = None,
|
||||
kernel_config_bwd_dX: KernelConfigBackward_dX = None,
|
||||
):
|
||||
super().__init__(config, gate, gate_up_proj, down_proj)
|
||||
self.permute_x = permute_x
|
||||
self.permute_y = permute_y
|
||||
self.autotune = autotune
|
||||
if not autotune:
|
||||
assert (
|
||||
kernel_config_fwd is not None
|
||||
and kernel_config_bwd_dW is not None
|
||||
and kernel_config_bwd_dX is not None
|
||||
), "Kernel configs must be provided if autotune is False"
|
||||
self.kernel_config_fwd = kernel_config_fwd
|
||||
self.kernel_config_bwd_dW = kernel_config_bwd_dW
|
||||
self.kernel_config_bwd_dX = kernel_config_bwd_dX
|
||||
|
||||
@classmethod
|
||||
def from_hf(
|
||||
cls,
|
||||
moe_block: Qwen3MoeSparseMoeBlock,
|
||||
permute_x: bool = False,
|
||||
permute_y: bool = False,
|
||||
autotune: bool = True,
|
||||
kernel_config_fwd: KernelConfigForward = None,
|
||||
kernel_config_bwd_dW: KernelConfigBackward_dW = None,
|
||||
kernel_config_bwd_dX: KernelConfigBackward_dX = None,
|
||||
):
|
||||
config: Qwen3MoeConfig = moe_block.experts[0].config
|
||||
gate, gate_up_proj, down_proj = Qwen3MoeGroupedGEMMBlock.extract_hf_weights(
|
||||
moe_block
|
||||
)
|
||||
return cls(
|
||||
config,
|
||||
gate,
|
||||
gate_up_proj,
|
||||
down_proj,
|
||||
permute_x=permute_x,
|
||||
permute_y=permute_y,
|
||||
autotune=autotune,
|
||||
kernel_config_fwd=kernel_config_fwd,
|
||||
kernel_config_bwd_dW=kernel_config_bwd_dW,
|
||||
kernel_config_bwd_dX=kernel_config_bwd_dX,
|
||||
)
|
||||
|
||||
def forward(self, hidden_states: torch.Tensor, debug: bool = False) -> torch.Tensor:
|
||||
batch_size, sequence_length, hidden_dim = hidden_states.shape
|
||||
num_tokens = batch_size * sequence_length
|
||||
total_tokens = num_tokens * self.top_k
|
||||
|
||||
hidden_states = hidden_states.view(-1, hidden_dim)
|
||||
|
||||
router_logits, routing_weights, selected_experts = self.run_router(
|
||||
hidden_states
|
||||
)
|
||||
# Pre-processing
|
||||
# 1. Compute tokens per expert and indices for gathering tokes from token order to expert order
|
||||
# NOTE: these are auxiliary data structs which don't need to be recorded in autograd graph
|
||||
token_counts_by_expert, gather_indices = (
|
||||
self.get_token_counts_and_gather_indices(selected_experts)
|
||||
)
|
||||
|
||||
# 2. permute_x -> permutation will be fused in prologue of first grouped gemm
|
||||
if not self.permute_x:
|
||||
hidden_states = permute(hidden_states, gather_indices, self.top_k)
|
||||
assert hidden_states.shape == (total_tokens, hidden_dim)
|
||||
|
||||
# Start expert computation
|
||||
first_gemm = grouped_gemm(
|
||||
X=hidden_states,
|
||||
W=self.gate_up_proj,
|
||||
m_sizes=token_counts_by_expert,
|
||||
gather_indices=gather_indices,
|
||||
topk=self.top_k,
|
||||
permute_x=self.permute_x,
|
||||
permute_y=False, # output of first grouped gemm should never be permuted
|
||||
autotune=self.autotune,
|
||||
kernel_config_fwd=self.kernel_config_fwd,
|
||||
kernel_config_bwd_dW=self.kernel_config_bwd_dW,
|
||||
kernel_config_bwd_dX=self.kernel_config_bwd_dX,
|
||||
is_first_gemm=True,
|
||||
)
|
||||
assert first_gemm.shape == (total_tokens, 2 * self.moe_intermediate_size)
|
||||
intermediate = self.act_and_mul(first_gemm)
|
||||
assert intermediate.shape == (total_tokens, self.moe_intermediate_size)
|
||||
second_gemm = grouped_gemm(
|
||||
X=intermediate,
|
||||
W=self.down_proj,
|
||||
m_sizes=token_counts_by_expert,
|
||||
gather_indices=gather_indices,
|
||||
topk=self.top_k,
|
||||
permute_x=False,
|
||||
permute_y=self.permute_y,
|
||||
autotune=self.autotune,
|
||||
kernel_config_fwd=self.kernel_config_fwd,
|
||||
kernel_config_bwd_dW=self.kernel_config_bwd_dW,
|
||||
kernel_config_bwd_dX=self.kernel_config_bwd_dX,
|
||||
is_first_gemm=False,
|
||||
)
|
||||
assert second_gemm.shape == (total_tokens, hidden_dim)
|
||||
|
||||
# Post-processing
|
||||
# 1. Unpermute from expert order to token order
|
||||
if not self.permute_y:
|
||||
hidden_states_unpermute = unpermute(second_gemm, gather_indices)
|
||||
assert hidden_states_unpermute.shape == (total_tokens, hidden_dim)
|
||||
else:
|
||||
hidden_states_unpermute = second_gemm
|
||||
|
||||
# 2. Merge topk weights
|
||||
hidden_states = (
|
||||
hidden_states_unpermute.view(num_tokens, self.top_k, hidden_dim)
|
||||
* routing_weights[..., None]
|
||||
)
|
||||
hidden_states = hidden_states.sum(dim=1)
|
||||
assert hidden_states.shape == (num_tokens, hidden_dim)
|
||||
|
||||
hidden_states = hidden_states.view(batch_size, sequence_length, hidden_dim)
|
||||
return GroupedGEMMResult(
|
||||
token_counts_by_expert=token_counts_by_expert,
|
||||
gather_indices=gather_indices,
|
||||
topk_weights=routing_weights,
|
||||
first_gemm=first_gemm,
|
||||
intermediate=intermediate,
|
||||
second_gemm=second_gemm,
|
||||
hidden_states_unpermute=hidden_states_unpermute,
|
||||
hidden_states=hidden_states,
|
||||
), router_logits
|
||||
35
unsloth/kernels/moe/tests/run_qwen3_moe_tests.sh
Executable file
35
unsloth/kernels/moe/tests/run_qwen3_moe_tests.sh
Executable file
|
|
@ -0,0 +1,35 @@
|
|||
#!/bin/bash
|
||||
|
||||
set -euo pipefail
|
||||
|
||||
SEQLENS=(1024)
|
||||
DTYPES=(bfloat16)
|
||||
PERMUTE_X=(false true)
|
||||
PERMUTE_Y=(false true)
|
||||
AUTOTUNE=(false true)
|
||||
|
||||
for SEQLEN in "${SEQLENS[@]}"; do
|
||||
for DTYPE in "${DTYPES[@]}"; do
|
||||
for PX in "${PERMUTE_X[@]}"; do
|
||||
for PY in "${PERMUTE_Y[@]}"; do
|
||||
for AT in "${AUTOTUNE[@]}"; do
|
||||
|
||||
ARGS=()
|
||||
[[ "$PX" == "true" ]] && ARGS+=("--permute_x")
|
||||
[[ "$PY" == "true" ]] && ARGS+=("--permute_y")
|
||||
[[ "$AT" == "true" ]] && ARGS+=("--autotune")
|
||||
|
||||
ARGS+=(--seqlen "$SEQLEN" --dtype "$DTYPE")
|
||||
|
||||
echo "Running with args: ${ARGS[*]}"
|
||||
if ! python -m tests.test_qwen3_moe "${ARGS[@]}"; then
|
||||
echo "❌ Test failed with args: --permute_x=$PX --permute_y=$PY --autotune=$AT --seqlen=$SEQLEN --dtype=$DTYPE" >&2
|
||||
else
|
||||
echo "✅ Test passed with args: --permute_x=$PX --permute_y=$PY --autotune=$AT --seqlen=$SEQLEN --dtype=$DTYPE"
|
||||
fi
|
||||
|
||||
done
|
||||
done
|
||||
done
|
||||
done
|
||||
done
|
||||
1204
unsloth/kernels/moe/tests/test_grouped_gemm.py
Normal file
1204
unsloth/kernels/moe/tests/test_grouped_gemm.py
Normal file
File diff suppressed because it is too large
Load diff
267
unsloth/kernels/moe/tests/test_qwen3_moe.py
Normal file
267
unsloth/kernels/moe/tests/test_qwen3_moe.py
Normal file
|
|
@ -0,0 +1,267 @@
|
|||
import argparse
|
||||
from contextlib import contextmanager
|
||||
|
||||
import pytest
|
||||
import torch
|
||||
from transformers import AutoConfig
|
||||
from transformers.models.qwen3_moe import Qwen3MoeConfig
|
||||
from transformers.models.qwen3_moe.modeling_qwen3_moe import Qwen3MoeSparseMoeBlock
|
||||
|
||||
from grouped_gemm.kernels.tuning import (
|
||||
KernelConfigBackward_dW,
|
||||
KernelConfigBackward_dX,
|
||||
KernelConfigForward,
|
||||
)
|
||||
from grouped_gemm.reference.moe_ops import Qwen3MoeGroupedGEMMBlock
|
||||
|
||||
from .moe_utils import (
|
||||
Qwen3MoeFusedGroupedGEMMBlock,
|
||||
check_fwd,
|
||||
check_grads,
|
||||
check_grouped_gemm_results,
|
||||
run_backward,
|
||||
run_forward,
|
||||
)
|
||||
|
||||
"""
|
||||
Qwen3 MoE tests
|
||||
|
||||
NOTE: Test this as a module and NOT with pytest as running with pytest results in random numerical errors: python -m tests.test_qwen3_moe --permute_x --permute_y --autotune NOT pytest -sv tests/test_qwen3_moe.py
|
||||
More specifically, all tests pass when run individually, but some will fail randomly (even with the same seed) when the entire test is run as a parametrized test suite using pytest, likely due to how pytest interacts with triton / autotuning.
|
||||
|
||||
See tests/run_qwen3_moe_tests.sh for a script that runs all the tests
|
||||
|
||||
The tests run the following:
|
||||
Huggingface's Qwen3 MoE block (Qwen3MoeSparseMoeBlock)
|
||||
Torch-native grouped gemm version of MoE block (Qwen3MoeGroupedGEMMBlock), which is the HF block with the expert computation replaced with a torch-native grouped gemm
|
||||
Triton kernel grouped gemm version of MoE block (Qwen3MoeFusedGroupedGEMMBlock), which is the HF block with the expert computation replaced with the fused triton grouped gemm kernel
|
||||
|
||||
The tests check the following:
|
||||
- HF MoE block vs torch grouped gemm MoE block (sanity check)
|
||||
- torch grouped gemm MoE block vs fused grouped gemm MoE block -- this allows us to test each of the intermediate results for easier debugging
|
||||
- HF MoE block vs fused grouped gemm MoE block -- this is the actual test
|
||||
|
||||
Both forward and backward passes are tests:
|
||||
- forward: output of the moe block
|
||||
- backwards:
|
||||
- X: gradient of the input to the moe block
|
||||
- gate.weight: gradient of the gate weights (router weights)
|
||||
- gate_proj: gradient of concatenated gate projections
|
||||
- up_proj: gradient of the concatenated up projections
|
||||
- down_proj: gradient of the concatenated down projections
|
||||
|
||||
Additionally, for the torch grouped gemm and triton grouped gemm versions, the intermediate outputs of the forward pass are checked:
|
||||
- first_gemm: output of the first grouped gemm (X @ fused_gate_proj)
|
||||
- intermediate: output of silu_mul(first_gemm)
|
||||
- second_gemm: output of the second grouped gemm (intermediate @ down_proj)
|
||||
- hidden_states_unpermute: output of the second_gemm after unpermuting back to token order (from expert grouped order); in the case where the permutation is fused in the triton kernel, this is the same as second_gemm
|
||||
- hidden_states: output with the topk_weights applied
|
||||
"""
|
||||
|
||||
TOLERANCES = {
|
||||
torch.bfloat16: (1e-2, 1e-2),
|
||||
torch.float16: (1e-3, 1e-3),
|
||||
torch.float: (1e-5, 1e-5),
|
||||
}
|
||||
|
||||
|
||||
@pytest.fixture(scope="module")
|
||||
def model_id():
|
||||
return "Qwen/Qwen3-30B-A3B"
|
||||
|
||||
|
||||
@pytest.fixture(scope="module")
|
||||
def config(model_id: str):
|
||||
return AutoConfig.from_pretrained(model_id)
|
||||
|
||||
|
||||
@contextmanager
|
||||
def test_context(prelude, epilogue="Passed!", char="-", num_chars=80):
|
||||
print(char * num_chars)
|
||||
print(prelude)
|
||||
yield
|
||||
print(epilogue)
|
||||
print(char * num_chars)
|
||||
|
||||
|
||||
SEED = 42
|
||||
SEQ_LENS = [1024]
|
||||
DTYPES = [torch.bfloat16]
|
||||
|
||||
# Reduce the number of autotuning configs to prevent excessive runtime
|
||||
NUM_AUTOTUNE_CONFIGS = 50
|
||||
|
||||
|
||||
@pytest.mark.parametrize(
|
||||
"permute_y", [True], ids=lambda x: "permute_y" if x else "no_permute_y"
|
||||
)
|
||||
@pytest.mark.parametrize(
|
||||
"permute_x", [True], ids=lambda x: "permute_x" if x else "no_permute_x"
|
||||
)
|
||||
@pytest.mark.parametrize(
|
||||
"autotune", [True], ids=lambda x: "autotune" if x else "manual"
|
||||
)
|
||||
@pytest.mark.parametrize("seqlen", SEQ_LENS, ids=lambda x: f"seqlen={x}")
|
||||
@pytest.mark.parametrize("dtype", DTYPES, ids=str)
|
||||
def test_qwen3_moe(
|
||||
config: Qwen3MoeConfig,
|
||||
seqlen: int,
|
||||
dtype: torch.dtype,
|
||||
permute_x: bool,
|
||||
permute_y: bool,
|
||||
autotune: bool,
|
||||
atol: float,
|
||||
rtol: float,
|
||||
):
|
||||
torch.manual_seed(
|
||||
SEED
|
||||
) # Should not be needed when running using pytest -- autouse fixture in conftest.py
|
||||
device = "cuda"
|
||||
hidden_size = config.hidden_size
|
||||
bs = 1
|
||||
# Reference op -- HF
|
||||
moe_block = Qwen3MoeSparseMoeBlock(config).to(device, dtype)
|
||||
|
||||
# Torch-native grouped gemm version of MoE Block -- for sanity checking
|
||||
grouped_gemm_block = Qwen3MoeGroupedGEMMBlock.from_hf(moe_block).to(device, dtype)
|
||||
grouped_gemm_block.check_weights(moe_block)
|
||||
|
||||
if not autotune:
|
||||
kernel_config_fwd = KernelConfigForward()
|
||||
kernel_config_bwd_dW = KernelConfigBackward_dW()
|
||||
kernel_config_bwd_dX = KernelConfigBackward_dX()
|
||||
else:
|
||||
from grouped_gemm.kernels.backward import (
|
||||
_autotuned_grouped_gemm_dW_kernel,
|
||||
_autotuned_grouped_gemm_dX_kernel,
|
||||
)
|
||||
from grouped_gemm.kernels.forward import _autotuned_grouped_gemm_forward_kernel
|
||||
|
||||
# Hack to reduce number of autotuning configs
|
||||
_autotuned_grouped_gemm_forward_kernel.configs = (
|
||||
_autotuned_grouped_gemm_forward_kernel.configs[:NUM_AUTOTUNE_CONFIGS]
|
||||
)
|
||||
_autotuned_grouped_gemm_dW_kernel.configs = (
|
||||
_autotuned_grouped_gemm_dW_kernel.configs[:NUM_AUTOTUNE_CONFIGS]
|
||||
)
|
||||
_autotuned_grouped_gemm_dX_kernel.configs = (
|
||||
_autotuned_grouped_gemm_dX_kernel.configs[:NUM_AUTOTUNE_CONFIGS]
|
||||
)
|
||||
|
||||
kernel_config_fwd = None
|
||||
kernel_config_bwd_dW = None
|
||||
kernel_config_bwd_dX = None
|
||||
|
||||
# Triton kernel grouped gemm version of MoE Block -- this is what we're testing
|
||||
fused_gemm_block = Qwen3MoeFusedGroupedGEMMBlock.from_hf(
|
||||
moe_block,
|
||||
permute_x=permute_x,
|
||||
permute_y=permute_y,
|
||||
autotune=autotune,
|
||||
kernel_config_fwd=kernel_config_fwd,
|
||||
kernel_config_bwd_dW=kernel_config_bwd_dW,
|
||||
kernel_config_bwd_dX=kernel_config_bwd_dX,
|
||||
).to(device, dtype)
|
||||
fused_gemm_block.check_weights(moe_block)
|
||||
|
||||
X = torch.randn(
|
||||
bs, seqlen, hidden_size, dtype=dtype, device=device, requires_grad=True
|
||||
)
|
||||
|
||||
# Forward
|
||||
ref_result = run_forward(moe_block, X, is_grouped_gemm=False)
|
||||
grouped_result = run_forward(grouped_gemm_block, X, is_grouped_gemm=True)
|
||||
fused_result = run_forward(fused_gemm_block, X, is_grouped_gemm=True)
|
||||
|
||||
with test_context(
|
||||
"Testing forward pass",
|
||||
epilogue="Passed forward tests!",
|
||||
char="=",
|
||||
num_chars=100,
|
||||
):
|
||||
# Sanity checks
|
||||
|
||||
with test_context("Checking HF vs torch grouped gemm MoE forward outputs..."):
|
||||
check_fwd(ref_result, grouped_result, atol, rtol, verbose=False)
|
||||
|
||||
with test_context(
|
||||
"Checking torch grouped gemm MoE vs fused grouped gemm MoE forward outputs..."
|
||||
):
|
||||
# We implement a custom check for grouped gemm results to test each of the intermediate results for easier debugging
|
||||
check_grouped_gemm_results(
|
||||
grouped_result.grouped_gemm_result,
|
||||
fused_result.grouped_gemm_result,
|
||||
permute_y=permute_y,
|
||||
atol=atol,
|
||||
rtol=rtol,
|
||||
verbose=False,
|
||||
)
|
||||
# Actual test
|
||||
with test_context("Checking HF vs fused grouped gemm MoE forward outputs..."):
|
||||
check_fwd(ref_result, fused_result, atol, rtol, verbose=True)
|
||||
|
||||
# Backward
|
||||
grad_output = torch.randn_like(ref_result.output)
|
||||
ref_backward_result = run_backward(
|
||||
moe_block, grad_output, output=ref_result.output, X=ref_result.X
|
||||
)
|
||||
grouped_backward_result = run_backward(
|
||||
grouped_gemm_block,
|
||||
grad_output,
|
||||
output=grouped_result.output,
|
||||
X=grouped_result.X,
|
||||
)
|
||||
fused_backward_result = run_backward(
|
||||
fused_gemm_block, grad_output, output=fused_result.output, X=fused_result.X
|
||||
)
|
||||
|
||||
with test_context(
|
||||
"Testing backward pass",
|
||||
epilogue="Passed backward tests!",
|
||||
char="=",
|
||||
num_chars=100,
|
||||
):
|
||||
# Sanity checks
|
||||
with test_context("Checking HF vs torch grouped gemm MoE grads..."):
|
||||
check_grads(
|
||||
ref_backward_result, grouped_backward_result, atol, rtol, verbose=False
|
||||
)
|
||||
with test_context(
|
||||
"Checking torch grouped gemm MoE vs fused grouped gemm MoE grads..."
|
||||
):
|
||||
check_grads(
|
||||
grouped_backward_result,
|
||||
fused_backward_result,
|
||||
atol,
|
||||
rtol,
|
||||
verbose=False,
|
||||
)
|
||||
|
||||
# Actual test
|
||||
with test_context("Checking HF vs fused grouped gemm MoE grads..."):
|
||||
check_grads(
|
||||
ref_backward_result, fused_backward_result, atol, rtol, verbose=True
|
||||
)
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
parser = argparse.ArgumentParser()
|
||||
parser.add_argument("--seqlen", type=int, default=1024)
|
||||
parser.add_argument(
|
||||
"--dtype", type=str, choices=["bfloat16", "float16"], default="bfloat16"
|
||||
)
|
||||
parser.add_argument("--permute_x", action="store_true")
|
||||
parser.add_argument("--permute_y", action="store_true")
|
||||
parser.add_argument("--autotune", action="store_true")
|
||||
args = parser.parse_args()
|
||||
args.dtype = getattr(torch, args.dtype)
|
||||
args_dict = vars(args)
|
||||
|
||||
model_id = "Qwen/Qwen3-30B-A3B"
|
||||
config = AutoConfig.from_pretrained(model_id)
|
||||
atol, rtol = TOLERANCES[args.dtype]
|
||||
|
||||
print(
|
||||
f"Testing {model_id} with seqlen={args.seqlen}, dtype={args.dtype}, permute_x={args.permute_x}, permute_y={args.permute_y}, autotune={args.autotune}, atol={atol}, rtol={rtol}"
|
||||
)
|
||||
test_qwen3_moe(config, atol=atol, rtol=rtol, **args_dict)
|
||||
Loading…
Add table
Add a link
Reference in a new issue