Soooo close to cracking deepspeed with triton now working on windows but stuck on "lscpu"

[PleezDeez]

To preface this I'm on Cuda 11.7, python 3.10.8, most recent dev build of xformers, accelerate 0.15.0, and I've gotten both to recognize that I have triton 2.0.0 installed thanks to the recent thread about somebody finally cracking it but I do not know how to get past the next hurdle because it's a command buried deep in deepspeed "lscpu" which is what linux uses to poll cpu information (cores, threads, etc) that does not have an equivalent in Windows... Need help from a wrinklier brain than mine to get this figure this out :) we're so close!

[PleezDeez]

got passed lscpu flag by installing py_cpuinfo then rain into triton flag so edited the name in all files I saw from 2.0.0 to 1.0.0 and now I'm stuck on building the async IO bc there's no libaio-dev version for windows that I can find. it's only on apt for now

[PleezDeez]

[PleezDeez]

I'm building the .whl now with triton support...

[PleezDeez]

still no aio for NVMe offload but we'll see what happens when it's done...

[PleezDeez]

was able to capture the code for async.io.py with all the things it's looking for while trying to install:

Copyright 2020 The Microsoft DeepSpeed Team
"""
import distutils.spawn
import subprocess

from .builder import OpBuilder

class AsyncIOBuilder(OpBuilder):
BUILD_VAR = "DS_BUILD_AIO"
NAME = "async_io"

def __init__(self):
    super().__init__(name=self.NAME)

def absolute_name(self):
    return f'deepspeed.ops.aio.{self.NAME}_op'

def sources(self):
    return [
        'csrc/aio/py_lib/deepspeed_py_copy.cpp',
        'csrc/aio/py_lib/py_ds_aio.cpp',
        'csrc/aio/py_lib/deepspeed_py_aio.cpp',
        'csrc/aio/py_lib/deepspeed_py_aio_handle.cpp',
        'csrc/aio/py_lib/deepspeed_aio_thread.cpp',
        'csrc/aio/common/deepspeed_aio_utils.cpp',
        'csrc/aio/common/deepspeed_aio_common.cpp',
        'csrc/aio/common/deepspeed_aio_types.cpp'
    ]

def include_paths(self):
    return ['csrc/aio/py_lib', 'csrc/aio/common']

def cxx_args(self):
    # -O0 for improved debugging, since performance is bound by I/O
    CPU_ARCH = self.cpu_arch()
    SIMD_WIDTH = self.simd_width()
    return [
        '-g',
        '-Wall',
        '-O0',
        '-std=c++14',
        '-shared',
        '-fPIC',
        '-Wno-reorder',
        CPU_ARCH,
        '-fopenmp',
        SIMD_WIDTH,
        '-laio',
    ]

def extra_ldflags(self):
    return ['-laio']

def check_for_libaio_pkg(self):
    libs = dict(
        dpkg=["-l",
              "libaio-dev",
              "apt"],
        pacman=["-Q",
                "libaio",
                "pacman"],
        rpm=["-q",
             "libaio-devel",
             "yum"],
    )

    found = False
    for pkgmgr, data in libs.items():
        flag, lib, tool = data
        path = distutils.spawn.find_executable(pkgmgr)
        if path is not None:
            cmd = f"{pkgmgr} {flag} {lib}"
            result = subprocess.Popen(cmd,
                                      stdout=subprocess.PIPE,
                                      stderr=subprocess.PIPE,
                                      shell=True)
            if result.wait() == 0:
                found = True
            else:
                self.warning(
                    f"{self.NAME}: please install the {lib} package with {tool}")
            break
    return found

def is_compatible(self, verbose=True):
    # Check for the existence of libaio by using distutils
    # to compile and link a test program that calls io_submit,
    # which is a function provided by libaio that is used in the async_io op.
    # If needed, one can define -I and -L entries in CFLAGS and LDFLAGS
    # respectively to specify the directories for libaio.h and libaio.so.
    aio_compatible = self.has_function('io_submit', ('aio', ))
    if verbose and not aio_compatible:
        self.warning(
            f"{self.NAME} requires the dev libaio .so object and headers but these were not found."
        )

        # Check for the libaio package via known package managers
        # to print suggestions on which package to install.
        self.check_for_libaio_pkg()

        self.warning(
            "If libaio is already installed (perhaps from source), try setting the CFLAGS and LDFLAGS environment variables to where it can be found."
        )
    return super().is_compatible(verbose) and aio_compatible

[PleezDeez]

[PleezDeez]

forceWrite(self.choices[index], end)
File "C:\Automatic1111\venv\lib\site-packages\accelerate\commands\menu\helpers.py", line 36, in forceWrite
sys.stdout.write(content + end)
TypeError: unsupported operand type(s) for +: 'int' and 'str'

when trying to run Accelerate Config after enabling deepspeed option

[PleezDeez]

Used WSL to generate a Stage 2 optimized cpu offload Accelerate config and copied it into my huggingface cache folder in windows where it's normally stored to get around it and was able to start training with deepspeed including sparse attention support both enabled. Still no NVMe offload bc I have no way to build the libaio-dev version it uses for it for windows but my RAM usage spiked the second it started training.

[PleezDeez]

I still oomed bc I only have an rtx 3060 ti with 8gb vram and only 16gb total system RAM but it's a start...