Upload 64 files

.gitattributes

@@ -33,3 +33,29 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+datasets-cli.exe filter=lfs diff=lfs merge=lfs -text
+f2py.exe filter=lfs diff=lfs merge=lfs -text
+huggingface-cli.exe filter=lfs diff=lfs merge=lfs -text
+import_pb_to_tensorboard.exe filter=lfs diff=lfs merge=lfs -text
+isympy.exe filter=lfs diff=lfs merge=lfs -text
+markdown_py.exe filter=lfs diff=lfs merge=lfs -text
+markdown-it.exe filter=lfs diff=lfs merge=lfs -text
+normalizer.exe filter=lfs diff=lfs merge=lfs -text
+numpy-config.exe filter=lfs diff=lfs merge=lfs -text
+pip.exe filter=lfs diff=lfs merge=lfs -text
+pip3.12.exe filter=lfs diff=lfs merge=lfs -text
+pip3.exe filter=lfs diff=lfs merge=lfs -text
+pygmentize.exe filter=lfs diff=lfs merge=lfs -text
+python.exe filter=lfs diff=lfs merge=lfs -text
+pythonw.exe filter=lfs diff=lfs merge=lfs -text
+saved_model_cli.exe filter=lfs diff=lfs merge=lfs -text
+tensorboard.exe filter=lfs diff=lfs merge=lfs -text
+tf_upgrade_v2.exe filter=lfs diff=lfs merge=lfs -text
+tflite_convert.exe filter=lfs diff=lfs merge=lfs -text
+toco_from_protos.exe filter=lfs diff=lfs merge=lfs -text
+toco.exe filter=lfs diff=lfs merge=lfs -text
+torchfrtrace.exe filter=lfs diff=lfs merge=lfs -text
+torchrun.exe filter=lfs diff=lfs merge=lfs -text
+tqdm.exe filter=lfs diff=lfs merge=lfs -text
+transformers-cli.exe filter=lfs diff=lfs merge=lfs -text
+wheel.exe filter=lfs diff=lfs merge=lfs -text

Activate.ps1

@@ -0,0 +1,502 @@
+<#
+.Synopsis
+Activate a Python virtual environment for the current PowerShell session.
+
+.Description
+Pushes the python executable for a virtual environment to the front of the
+$Env:PATH environment variable and sets the prompt to signify that you are
+in a Python virtual environment. Makes use of the command line switches as
+well as the `pyvenv.cfg` file values present in the virtual environment.
+
+.Parameter VenvDir
+Path to the directory that contains the virtual environment to activate. The
+default value for this is the parent of the directory that the Activate.ps1
+script is located within.
+
+.Parameter Prompt
+The prompt prefix to display when this virtual environment is activated. By
+default, this prompt is the name of the virtual environment folder (VenvDir)
+surrounded by parentheses and followed by a single space (ie. '(.venv) ').
+
+.Example
+Activate.ps1
+Activates the Python virtual environment that contains the Activate.ps1 script.
+
+.Example
+Activate.ps1 -Verbose
+Activates the Python virtual environment that contains the Activate.ps1 script,
+and shows extra information about the activation as it executes.
+
+.Example
+Activate.ps1 -VenvDir C:\Users\MyUser\Common\.venv
+Activates the Python virtual environment located in the specified location.
+
+.Example
+Activate.ps1 -Prompt "MyPython"
+Activates the Python virtual environment that contains the Activate.ps1 script,
+and prefixes the current prompt with the specified string (surrounded in
+parentheses) while the virtual environment is active.
+
+.Notes
+On Windows, it may be required to enable this Activate.ps1 script by setting the
+execution policy for the user. You can do this by issuing the following PowerShell
+command:
+
+PS C:\> Set-ExecutionPolicy -ExecutionPolicy RemoteSigned -Scope CurrentUser
+
+For more information on Execution Policies: 
+https://go.microsoft.com/fwlink/?LinkID=135170
+
+#>
+Param(
+    [Parameter(Mandatory = $false)]
+    [String]
+    $VenvDir,
+    [Parameter(Mandatory = $false)]
+    [String]
+    $Prompt
+)
+
+<# Function declarations --------------------------------------------------- #>
+
+<#
+.Synopsis
+Remove all shell session elements added by the Activate script, including the
+addition of the virtual environment's Python executable from the beginning of
+the PATH variable.
+
+.Parameter NonDestructive
+If present, do not remove this function from the global namespace for the
+session.
+
+#>
+function global:deactivate ([switch]$NonDestructive) {
+    # Revert to original values
+
+    # The prior prompt:
+    if (Test-Path -Path Function:_OLD_VIRTUAL_PROMPT) {
+        Copy-Item -Path Function:_OLD_VIRTUAL_PROMPT -Destination Function:prompt
+        Remove-Item -Path Function:_OLD_VIRTUAL_PROMPT
+    }
+
+    # The prior PYTHONHOME:
+    if (Test-Path -Path Env:_OLD_VIRTUAL_PYTHONHOME) {
+        Copy-Item -Path Env:_OLD_VIRTUAL_PYTHONHOME -Destination Env:PYTHONHOME
+        Remove-Item -Path Env:_OLD_VIRTUAL_PYTHONHOME
+    }
+
+    # The prior PATH:
+    if (Test-Path -Path Env:_OLD_VIRTUAL_PATH) {
+        Copy-Item -Path Env:_OLD_VIRTUAL_PATH -Destination Env:PATH
+        Remove-Item -Path Env:_OLD_VIRTUAL_PATH
+    }
+
+    # Just remove the VIRTUAL_ENV altogether:
+    if (Test-Path -Path Env:VIRTUAL_ENV) {
+        Remove-Item -Path env:VIRTUAL_ENV
+    }
+
+    # Just remove VIRTUAL_ENV_PROMPT altogether.
+    if (Test-Path -Path Env:VIRTUAL_ENV_PROMPT) {
+        Remove-Item -Path env:VIRTUAL_ENV_PROMPT
+    }
+
+    # Just remove the _PYTHON_VENV_PROMPT_PREFIX altogether:
+    if (Get-Variable -Name "_PYTHON_VENV_PROMPT_PREFIX" -ErrorAction SilentlyContinue) {
+        Remove-Variable -Name _PYTHON_VENV_PROMPT_PREFIX -Scope Global -Force
+    }
+
+    # Leave deactivate function in the global namespace if requested:
+    if (-not $NonDestructive) {
+        Remove-Item -Path function:deactivate
+    }
+}
+
+<#
+.Description
+Get-PyVenvConfig parses the values from the pyvenv.cfg file located in the
+given folder, and returns them in a map.
+
+For each line in the pyvenv.cfg file, if that line can be parsed into exactly
+two strings separated by `=` (with any amount of whitespace surrounding the =)
+then it is considered a `key = value` line. The left hand string is the key,
+the right hand is the value.
+
+If the value starts with a `'` or a `"` then the first and last character is
+stripped from the value before being captured.
+
+.Parameter ConfigDir
+Path to the directory that contains the `pyvenv.cfg` file.
+#>
+function Get-PyVenvConfig(
+    [String]
+    $ConfigDir
+) {
+    Write-Verbose "Given ConfigDir=$ConfigDir, obtain values in pyvenv.cfg"
+
+    # Ensure the file exists, and issue a warning if it doesn't (but still allow the function to continue).
+    $pyvenvConfigPath = Join-Path -Resolve -Path $ConfigDir -ChildPath 'pyvenv.cfg' -ErrorAction Continue
+
+    # An empty map will be returned if no config file is found.
+    $pyvenvConfig = @{ }
+
+    if ($pyvenvConfigPath) {
+
+        Write-Verbose "File exists, parse `key = value` lines"
+        $pyvenvConfigContent = Get-Content -Path $pyvenvConfigPath
+
+        $pyvenvConfigContent | ForEach-Object {
+            $keyval = $PSItem -split "\s*=\s*", 2
+            if ($keyval[0] -and $keyval[1]) {
+                $val = $keyval[1]
+
+                # Remove extraneous quotations around a string value.
+                if ("'""".Contains($val.Substring(0, 1))) {
+                    $val = $val.Substring(1, $val.Length - 2)
+                }
+
+                $pyvenvConfig[$keyval[0]] = $val
+                Write-Verbose "Adding Key: '$($keyval[0])'='$val'"
+            }
+        }
+    }
+    return $pyvenvConfig
+}
+
+
+<# Begin Activate script --------------------------------------------------- #>
+
+# Determine the containing directory of this script
+$VenvExecPath = Split-Path -Parent $MyInvocation.MyCommand.Definition
+$VenvExecDir = Get-Item -Path $VenvExecPath
+
+Write-Verbose "Activation script is located in path: '$VenvExecPath'"
+Write-Verbose "VenvExecDir Fullname: '$($VenvExecDir.FullName)"
+Write-Verbose "VenvExecDir Name: '$($VenvExecDir.Name)"
+
+# Set values required in priority: CmdLine, ConfigFile, Default
+# First, get the location of the virtual environment, it might not be
+# VenvExecDir if specified on the command line.
+if ($VenvDir) {
+    Write-Verbose "VenvDir given as parameter, using '$VenvDir' to determine values"
+}
+else {
+    Write-Verbose "VenvDir not given as a parameter, using parent directory name as VenvDir."
+    $VenvDir = $VenvExecDir.Parent.FullName.TrimEnd("\\/")
+    Write-Verbose "VenvDir=$VenvDir"
+}
+
+# Next, read the `pyvenv.cfg` file to determine any required value such
+# as `prompt`.
+$pyvenvCfg = Get-PyVenvConfig -ConfigDir $VenvDir
+
+# Next, set the prompt from the command line, or the config file, or
+# just use the name of the virtual environment folder.
+if ($Prompt) {
+    Write-Verbose "Prompt specified as argument, using '$Prompt'"
+}
+else {
+    Write-Verbose "Prompt not specified as argument to script, checking pyvenv.cfg value"
+    if ($pyvenvCfg -and $pyvenvCfg['prompt']) {
+        Write-Verbose "  Setting based on value in pyvenv.cfg='$($pyvenvCfg['prompt'])'"
+        $Prompt = $pyvenvCfg['prompt'];
+    }
+    else {
+        Write-Verbose "  Setting prompt based on parent's directory's name. (Is the directory name passed to venv module when creating the virtual environment)"
+        Write-Verbose "  Got leaf-name of $VenvDir='$(Split-Path -Path $venvDir -Leaf)'"
+        $Prompt = Split-Path -Path $venvDir -Leaf
+    }
+}
+
+Write-Verbose "Prompt = '$Prompt'"
+Write-Verbose "VenvDir='$VenvDir'"
+
+# Deactivate any currently active virtual environment, but leave the
+# deactivate function in place.
+deactivate -nondestructive
+
+# Now set the environment variable VIRTUAL_ENV, used by many tools to determine
+# that there is an activated venv.
+$env:VIRTUAL_ENV = $VenvDir
+
+if (-not $Env:VIRTUAL_ENV_DISABLE_PROMPT) {
+
+    Write-Verbose "Setting prompt to '$Prompt'"
+
+    # Set the prompt to include the env name
+    # Make sure _OLD_VIRTUAL_PROMPT is global
+    function global:_OLD_VIRTUAL_PROMPT { "" }
+    Copy-Item -Path function:prompt -Destination function:_OLD_VIRTUAL_PROMPT
+    New-Variable -Name _PYTHON_VENV_PROMPT_PREFIX -Description "Python virtual environment prompt prefix" -Scope Global -Option ReadOnly -Visibility Public -Value $Prompt
+
+    function global:prompt {
+        Write-Host -NoNewline -ForegroundColor Green "($_PYTHON_VENV_PROMPT_PREFIX) "
+        _OLD_VIRTUAL_PROMPT
+    }
+    $env:VIRTUAL_ENV_PROMPT = $Prompt
+}
+
+# Clear PYTHONHOME
+if (Test-Path -Path Env:PYTHONHOME) {
+    Copy-Item -Path Env:PYTHONHOME -Destination Env:_OLD_VIRTUAL_PYTHONHOME
+    Remove-Item -Path Env:PYTHONHOME
+}
+
+# Add the venv to the PATH
+Copy-Item -Path Env:PATH -Destination Env:_OLD_VIRTUAL_PATH
+$Env:PATH = "$VenvExecDir$([System.IO.Path]::PathSeparator)$Env:PATH"
+
+# SIG # Begin signature block
+# MIIvIwYJKoZIhvcNAQcCoIIvFDCCLxACAQExDzANBglghkgBZQMEAgEFADB5Bgor
+# BgEEAYI3AgEEoGswaTA0BgorBgEEAYI3AgEeMCYCAwEAAAQQH8w7YFlLCE63JNLG
+# KX7zUQIBAAIBAAIBAAIBAAIBADAxMA0GCWCGSAFlAwQCAQUABCBnL745ElCYk8vk
+# dBtMuQhLeWJ3ZGfzKW4DHCYzAn+QB6CCE8MwggWQMIIDeKADAgECAhAFmxtXno4h
+# MuI5B72nd3VcMA0GCSqGSIb3DQEBDAUAMGIxCzAJBgNVBAYTAlVTMRUwEwYDVQQK
+# EwxEaWdpQ2VydCBJbmMxGTAXBgNVBAsTEHd3dy5kaWdpY2VydC5jb20xITAfBgNV
+# BAMTGERpZ2lDZXJ0IFRydXN0ZWQgUm9vdCBHNDAeFw0xMzA4MDExMjAwMDBaFw0z
+# ODAxMTUxMjAwMDBaMGIxCzAJBgNVBAYTAlVTMRUwEwYDVQQKEwxEaWdpQ2VydCBJ
+# bmMxGTAXBgNVBAsTEHd3dy5kaWdpY2VydC5jb20xITAfBgNVBAMTGERpZ2lDZXJ0
+# IFRydXN0ZWQgUm9vdCBHNDCCAiIwDQYJKoZIhvcNAQEBBQADggIPADCCAgoCggIB
+# AL/mkHNo3rvkXUo8MCIwaTPswqclLskhPfKK2FnC4SmnPVirdprNrnsbhA3EMB/z
+# G6Q4FutWxpdtHauyefLKEdLkX9YFPFIPUh/GnhWlfr6fqVcWWVVyr2iTcMKyunWZ
+# anMylNEQRBAu34LzB4TmdDttceItDBvuINXJIB1jKS3O7F5OyJP4IWGbNOsFxl7s
+# Wxq868nPzaw0QF+xembud8hIqGZXV59UWI4MK7dPpzDZVu7Ke13jrclPXuU15zHL
+# 2pNe3I6PgNq2kZhAkHnDeMe2scS1ahg4AxCN2NQ3pC4FfYj1gj4QkXCrVYJBMtfb
+# BHMqbpEBfCFM1LyuGwN1XXhm2ToxRJozQL8I11pJpMLmqaBn3aQnvKFPObURWBf3
+# JFxGj2T3wWmIdph2PVldQnaHiZdpekjw4KISG2aadMreSx7nDmOu5tTvkpI6nj3c
+# AORFJYm2mkQZK37AlLTSYW3rM9nF30sEAMx9HJXDj/chsrIRt7t/8tWMcCxBYKqx
+# YxhElRp2Yn72gLD76GSmM9GJB+G9t+ZDpBi4pncB4Q+UDCEdslQpJYls5Q5SUUd0
+# viastkF13nqsX40/ybzTQRESW+UQUOsxxcpyFiIJ33xMdT9j7CFfxCBRa2+xq4aL
+# T8LWRV+dIPyhHsXAj6KxfgommfXkaS+YHS312amyHeUbAgMBAAGjQjBAMA8GA1Ud
+# EwEB/wQFMAMBAf8wDgYDVR0PAQH/BAQDAgGGMB0GA1UdDgQWBBTs1+OC0nFdZEzf
+# Lmc/57qYrhwPTzANBgkqhkiG9w0BAQwFAAOCAgEAu2HZfalsvhfEkRvDoaIAjeNk
+# aA9Wz3eucPn9mkqZucl4XAwMX+TmFClWCzZJXURj4K2clhhmGyMNPXnpbWvWVPjS
+# PMFDQK4dUPVS/JA7u5iZaWvHwaeoaKQn3J35J64whbn2Z006Po9ZOSJTROvIXQPK
+# 7VB6fWIhCoDIc2bRoAVgX+iltKevqPdtNZx8WorWojiZ83iL9E3SIAveBO6Mm0eB
+# cg3AFDLvMFkuruBx8lbkapdvklBtlo1oepqyNhR6BvIkuQkRUNcIsbiJeoQjYUIp
+# 5aPNoiBB19GcZNnqJqGLFNdMGbJQQXE9P01wI4YMStyB0swylIQNCAmXHE/A7msg
+# dDDS4Dk0EIUhFQEI6FUy3nFJ2SgXUE3mvk3RdazQyvtBuEOlqtPDBURPLDab4vri
+# RbgjU2wGb2dVf0a1TD9uKFp5JtKkqGKX0h7i7UqLvBv9R0oN32dmfrJbQdA75PQ7
+# 9ARj6e/CVABRoIoqyc54zNXqhwQYs86vSYiv85KZtrPmYQ/ShQDnUBrkG5WdGaG5
+# nLGbsQAe79APT0JsyQq87kP6OnGlyE0mpTX9iV28hWIdMtKgK1TtmlfB2/oQzxm3
+# i0objwG2J5VT6LaJbVu8aNQj6ItRolb58KaAoNYes7wPD1N1KarqE3fk3oyBIa0H
+# EEcRrYc9B9F1vM/zZn4wggawMIIEmKADAgECAhAIrUCyYNKcTJ9ezam9k67ZMA0G
+# CSqGSIb3DQEBDAUAMGIxCzAJBgNVBAYTAlVTMRUwEwYDVQQKEwxEaWdpQ2VydCBJ
+# bmMxGTAXBgNVBAsTEHd3dy5kaWdpY2VydC5jb20xITAfBgNVBAMTGERpZ2lDZXJ0
+# IFRydXN0ZWQgUm9vdCBHNDAeFw0yMTA0MjkwMDAwMDBaFw0zNjA0MjgyMzU5NTla
+# MGkxCzAJBgNVBAYTAlVTMRcwFQYDVQQKEw5EaWdpQ2VydCwgSW5jLjFBMD8GA1UE
+# AxM4RGlnaUNlcnQgVHJ1c3RlZCBHNCBDb2RlIFNpZ25pbmcgUlNBNDA5NiBTSEEz
+# ODQgMjAyMSBDQTEwggIiMA0GCSqGSIb3DQEBAQUAA4ICDwAwggIKAoICAQDVtC9C
+# 0CiteLdd1TlZG7GIQvUzjOs9gZdwxbvEhSYwn6SOaNhc9es0JAfhS0/TeEP0F9ce
+# 2vnS1WcaUk8OoVf8iJnBkcyBAz5NcCRks43iCH00fUyAVxJrQ5qZ8sU7H/Lvy0da
+# E6ZMswEgJfMQ04uy+wjwiuCdCcBlp/qYgEk1hz1RGeiQIXhFLqGfLOEYwhrMxe6T
+# SXBCMo/7xuoc82VokaJNTIIRSFJo3hC9FFdd6BgTZcV/sk+FLEikVoQ11vkunKoA
+# FdE3/hoGlMJ8yOobMubKwvSnowMOdKWvObarYBLj6Na59zHh3K3kGKDYwSNHR7Oh
+# D26jq22YBoMbt2pnLdK9RBqSEIGPsDsJ18ebMlrC/2pgVItJwZPt4bRc4G/rJvmM
+# 1bL5OBDm6s6R9b7T+2+TYTRcvJNFKIM2KmYoX7BzzosmJQayg9Rc9hUZTO1i4F4z
+# 8ujo7AqnsAMrkbI2eb73rQgedaZlzLvjSFDzd5Ea/ttQokbIYViY9XwCFjyDKK05
+# huzUtw1T0PhH5nUwjewwk3YUpltLXXRhTT8SkXbev1jLchApQfDVxW0mdmgRQRNY
+# mtwmKwH0iU1Z23jPgUo+QEdfyYFQc4UQIyFZYIpkVMHMIRroOBl8ZhzNeDhFMJlP
+# /2NPTLuqDQhTQXxYPUez+rbsjDIJAsxsPAxWEQIDAQABo4IBWTCCAVUwEgYDVR0T
+# AQH/BAgwBgEB/wIBADAdBgNVHQ4EFgQUaDfg67Y7+F8Rhvv+YXsIiGX0TkIwHwYD
+# VR0jBBgwFoAU7NfjgtJxXWRM3y5nP+e6mK4cD08wDgYDVR0PAQH/BAQDAgGGMBMG
+# A1UdJQQMMAoGCCsGAQUFBwMDMHcGCCsGAQUFBwEBBGswaTAkBggrBgEFBQcwAYYY
+# aHR0cDovL29jc3AuZGlnaWNlcnQuY29tMEEGCCsGAQUFBzAChjVodHRwOi8vY2Fj
+# ZXJ0cy5kaWdpY2VydC5jb20vRGlnaUNlcnRUcnVzdGVkUm9vdEc0LmNydDBDBgNV
+# HR8EPDA6MDigNqA0hjJodHRwOi8vY3JsMy5kaWdpY2VydC5jb20vRGlnaUNlcnRU
+# cnVzdGVkUm9vdEc0LmNybDAcBgNVHSAEFTATMAcGBWeBDAEDMAgGBmeBDAEEATAN
+# BgkqhkiG9w0BAQwFAAOCAgEAOiNEPY0Idu6PvDqZ01bgAhql+Eg08yy25nRm95Ry
+# sQDKr2wwJxMSnpBEn0v9nqN8JtU3vDpdSG2V1T9J9Ce7FoFFUP2cvbaF4HZ+N3HL
+# IvdaqpDP9ZNq4+sg0dVQeYiaiorBtr2hSBh+3NiAGhEZGM1hmYFW9snjdufE5Btf
+# Q/g+lP92OT2e1JnPSt0o618moZVYSNUa/tcnP/2Q0XaG3RywYFzzDaju4ImhvTnh
+# OE7abrs2nfvlIVNaw8rpavGiPttDuDPITzgUkpn13c5UbdldAhQfQDN8A+KVssIh
+# dXNSy0bYxDQcoqVLjc1vdjcshT8azibpGL6QB7BDf5WIIIJw8MzK7/0pNVwfiThV
+# 9zeKiwmhywvpMRr/LhlcOXHhvpynCgbWJme3kuZOX956rEnPLqR0kq3bPKSchh/j
+# wVYbKyP/j7XqiHtwa+aguv06P0WmxOgWkVKLQcBIhEuWTatEQOON8BUozu3xGFYH
+# Ki8QxAwIZDwzj64ojDzLj4gLDb879M4ee47vtevLt/B3E+bnKD+sEq6lLyJsQfmC
+# XBVmzGwOysWGw/YmMwwHS6DTBwJqakAwSEs0qFEgu60bhQjiWQ1tygVQK+pKHJ6l
+# /aCnHwZ05/LWUpD9r4VIIflXO7ScA+2GRfS0YW6/aOImYIbqyK+p/pQd52MbOoZW
+# eE4wggd3MIIFX6ADAgECAhAHHxQbizANJfMU6yMM0NHdMA0GCSqGSIb3DQEBCwUA
+# MGkxCzAJBgNVBAYTAlVTMRcwFQYDVQQKEw5EaWdpQ2VydCwgSW5jLjFBMD8GA1UE
+# AxM4RGlnaUNlcnQgVHJ1c3RlZCBHNCBDb2RlIFNpZ25pbmcgUlNBNDA5NiBTSEEz
+# ODQgMjAyMSBDQTEwHhcNMjIwMTE3MDAwMDAwWhcNMjUwMTE1MjM1OTU5WjB8MQsw
+# CQYDVQQGEwJVUzEPMA0GA1UECBMGT3JlZ29uMRIwEAYDVQQHEwlCZWF2ZXJ0b24x
+# IzAhBgNVBAoTGlB5dGhvbiBTb2Z0d2FyZSBGb3VuZGF0aW9uMSMwIQYDVQQDExpQ
+# eXRob24gU29mdHdhcmUgRm91bmRhdGlvbjCCAiIwDQYJKoZIhvcNAQEBBQADggIP
+# ADCCAgoCggIBAKgc0BTT+iKbtK6f2mr9pNMUTcAJxKdsuOiSYgDFfwhjQy89koM7
+# uP+QV/gwx8MzEt3c9tLJvDccVWQ8H7mVsk/K+X+IufBLCgUi0GGAZUegEAeRlSXx
+# xhYScr818ma8EvGIZdiSOhqjYc4KnfgfIS4RLtZSrDFG2tN16yS8skFa3IHyvWdb
+# D9PvZ4iYNAS4pjYDRjT/9uzPZ4Pan+53xZIcDgjiTwOh8VGuppxcia6a7xCyKoOA
+# GjvCyQsj5223v1/Ig7Dp9mGI+nh1E3IwmyTIIuVHyK6Lqu352diDY+iCMpk9Zanm
+# SjmB+GMVs+H/gOiofjjtf6oz0ki3rb7sQ8fTnonIL9dyGTJ0ZFYKeb6BLA66d2GA
+# LwxZhLe5WH4Np9HcyXHACkppsE6ynYjTOd7+jN1PRJahN1oERzTzEiV6nCO1M3U1
+# HbPTGyq52IMFSBM2/07WTJSbOeXjvYR7aUxK9/ZkJiacl2iZI7IWe7JKhHohqKuc
+# eQNyOzxTakLcRkzynvIrk33R9YVqtB4L6wtFxhUjvDnQg16xot2KVPdfyPAWd81w
+# tZADmrUtsZ9qG79x1hBdyOl4vUtVPECuyhCxaw+faVjumapPUnwo8ygflJJ74J+B
+# Yxf6UuD7m8yzsfXWkdv52DjL74TxzuFTLHPyARWCSCAbzn3ZIly+qIqDAgMBAAGj
+# ggIGMIICAjAfBgNVHSMEGDAWgBRoN+Drtjv4XxGG+/5hewiIZfROQjAdBgNVHQ4E
+# FgQUt/1Teh2XDuUj2WW3siYWJgkZHA8wDgYDVR0PAQH/BAQDAgeAMBMGA1UdJQQM
+# MAoGCCsGAQUFBwMDMIG1BgNVHR8Ega0wgaowU6BRoE+GTWh0dHA6Ly9jcmwzLmRp
+# Z2ljZXJ0LmNvbS9EaWdpQ2VydFRydXN0ZWRHNENvZGVTaWduaW5nUlNBNDA5NlNI
+# QTM4NDIwMjFDQTEuY3JsMFOgUaBPhk1odHRwOi8vY3JsNC5kaWdpY2VydC5jb20v
+# RGlnaUNlcnRUcnVzdGVkRzRDb2RlU2lnbmluZ1JTQTQwOTZTSEEzODQyMDIxQ0Ex
+# LmNybDA+BgNVHSAENzA1MDMGBmeBDAEEATApMCcGCCsGAQUFBwIBFhtodHRwOi8v
+# d3d3LmRpZ2ljZXJ0LmNvbS9DUFMwgZQGCCsGAQUFBwEBBIGHMIGEMCQGCCsGAQUF
+# BzABhhhodHRwOi8vb2NzcC5kaWdpY2VydC5jb20wXAYIKwYBBQUHMAKGUGh0dHA6
+# Ly9jYWNlcnRzLmRpZ2ljZXJ0LmNvbS9EaWdpQ2VydFRydXN0ZWRHNENvZGVTaWdu
+# aW5nUlNBNDA5NlNIQTM4NDIwMjFDQTEuY3J0MAwGA1UdEwEB/wQCMAAwDQYJKoZI
+# hvcNAQELBQADggIBABxv4AeV/5ltkELHSC63fXAFYS5tadcWTiNc2rskrNLrfH1N
+# s0vgSZFoQxYBFKI159E8oQQ1SKbTEubZ/B9kmHPhprHya08+VVzxC88pOEvz68nA
+# 82oEM09584aILqYmj8Pj7h/kmZNzuEL7WiwFa/U1hX+XiWfLIJQsAHBla0i7QRF2
+# de8/VSF0XXFa2kBQ6aiTsiLyKPNbaNtbcucaUdn6vVUS5izWOXM95BSkFSKdE45O
+# q3FForNJXjBvSCpwcP36WklaHL+aHu1upIhCTUkzTHMh8b86WmjRUqbrnvdyR2yd
+# I5l1OqcMBjkpPpIV6wcc+KY/RH2xvVuuoHjlUjwq2bHiNoX+W1scCpnA8YTs2d50
+# jDHUgwUo+ciwpffH0Riq132NFmrH3r67VaN3TuBxjI8SIZM58WEDkbeoriDk3hxU
+# 8ZWV7b8AW6oyVBGfM06UgkfMb58h+tJPrFx8VI/WLq1dTqMfZOm5cuclMnUHs2uq
+# rRNtnV8UfidPBL4ZHkTcClQbCoz0UbLhkiDvIS00Dn+BBcxw/TKqVL4Oaz3bkMSs
+# M46LciTeucHY9ExRVt3zy7i149sd+F4QozPqn7FrSVHXmem3r7bjyHTxOgqxRCVa
+# 18Vtx7P/8bYSBeS+WHCKcliFCecspusCDSlnRUjZwyPdP0VHxaZg2unjHY3rMYIa
+# tjCCGrICAQEwfTBpMQswCQYDVQQGEwJVUzEXMBUGA1UEChMORGlnaUNlcnQsIElu
+# Yy4xQTA/BgNVBAMTOERpZ2lDZXJ0IFRydXN0ZWQgRzQgQ29kZSBTaWduaW5nIFJT
+# QTQwOTYgU0hBMzg0IDIwMjEgQ0ExAhAHHxQbizANJfMU6yMM0NHdMA0GCWCGSAFl
+# AwQCAQUAoIHIMBkGCSqGSIb3DQEJAzEMBgorBgEEAYI3AgEEMBwGCisGAQQBgjcC
+# AQsxDjAMBgorBgEEAYI3AgEVMC8GCSqGSIb3DQEJBDEiBCBnAZ6P7YvTwq0fbF62
+# o7E75R0LxsW5OtyYiFESQckLhjBcBgorBgEEAYI3AgEMMU4wTKBGgEQAQgB1AGkA
+# bAB0ADoAIABSAGUAbABlAGEAcwBlAF8AdgAzAC4AMQAyAC4ANgBfADIAMAAyADQA
+# MAA5ADAANgAuADAAMqECgAAwDQYJKoZIhvcNAQEBBQAEggIAhen5GN03SF9I96DT
+# rrWEsN7FAyx8BHoRf9WFBqoBXpFkBwlE6OWj/rxohuwB/b+3vcBGWaP497ACku4l
+# lgrWCrmYOVMKTjeHtDDkvgmygvGAtWB5drf56553na9RYjTxRqxto5LBMsHtPZy6
+# 1D+touyLSHx+QXzqXO4ssUq7oHtsmjDCKMLdcTuoqNGtpxaIwwlOAK+0DaLLUpkX
+# VRUUzMWBb+2FlmJ2wWtXXs6OtlACm4By2hHmKhd6OYwnHPe6fDVdrhGa0BcDAIIO
+# +elm895ddmfX2KqHWrKpgZ/0DM46pbEiYX4GVwY+kmrK9p8XF7c50c331vPPuImL
+# URRShtCM9F/5e522nQm0NxQ0Pz+thMD+qGBA8WuSoD+RRG+JKOXgM8sMX46goR8P
+# 1IJLeUnEKSOgMNcP0EUeWthrqXRjVgNcazIDgPFpPGMyo4Pp0D8SPvp/RzP3CPVo
+# uVj6r0OnhyoDuDEX4KCyo/+TCSm+2T+hv+cPWQaukovXF1TmahWb/8j1+K1RkCVd
+# UQ5v07AHYoHmJ2gxEgtM9qaVDx4woVVCpUrOhiAP/K1WSRw710oTqECG+4y+g67D
+# P2UuOxxaxhPk0pITFj9pZQcVsrCk5QbW3Yj/I3fISZgjVfYK1IDKzaWQQuBhOuim
+# j2/Tfcg+cLDbY4XEs5vpbKSYsCWhghc/MIIXOwYKKwYBBAGCNwMDATGCFyswghcn
+# BgkqhkiG9w0BBwKgghcYMIIXFAIBAzEPMA0GCWCGSAFlAwQCAQUAMHcGCyqGSIb3
+# DQEJEAEEoGgEZjBkAgEBBglghkgBhv1sBwEwMTANBglghkgBZQMEAgEFAAQgS2eq
+# 9RcYET/J2twNl3zStqvYDUBOrSdHvMcFbSu+C2sCEGHEWhqgAhMA1D+QZOB9TC4Y
+# DzIwMjQwOTA2MjAyNzExWqCCEwkwggbCMIIEqqADAgECAhAFRK/zlJ0IOaa/2z9f
+# 5WEWMA0GCSqGSIb3DQEBCwUAMGMxCzAJBgNVBAYTAlVTMRcwFQYDVQQKEw5EaWdp
+# Q2VydCwgSW5jLjE7MDkGA1UEAxMyRGlnaUNlcnQgVHJ1c3RlZCBHNCBSU0E0MDk2
+# IFNIQTI1NiBUaW1lU3RhbXBpbmcgQ0EwHhcNMjMwNzE0MDAwMDAwWhcNMzQxMDEz
+# MjM1OTU5WjBIMQswCQYDVQQGEwJVUzEXMBUGA1UEChMORGlnaUNlcnQsIEluYy4x
+# IDAeBgNVBAMTF0RpZ2lDZXJ0IFRpbWVzdGFtcCAyMDIzMIICIjANBgkqhkiG9w0B
+# AQEFAAOCAg8AMIICCgKCAgEAo1NFhx2DjlusPlSzI+DPn9fl0uddoQ4J3C9Io5d6
+# OyqcZ9xiFVjBqZMRp82qsmrdECmKHmJjadNYnDVxvzqX65RQjxwg6seaOy+WZuNp
+# 52n+W8PWKyAcwZeUtKVQgfLPywemMGjKg0La/H8JJJSkghraarrYO8pd3hkYhftF
+# 6g1hbJ3+cV7EBpo88MUueQ8bZlLjyNY+X9pD04T10Mf2SC1eRXWWdf7dEKEbg8G4
+# 5lKVtUfXeCk5a+B4WZfjRCtK1ZXO7wgX6oJkTf8j48qG7rSkIWRw69XloNpjsy7p
+# Be6q9iT1HbybHLK3X9/w7nZ9MZllR1WdSiQvrCuXvp/k/XtzPjLuUjT71Lvr1KAs
+# NJvj3m5kGQc3AZEPHLVRzapMZoOIaGK7vEEbeBlt5NkP4FhB+9ixLOFRr7StFQYU
+# 6mIIE9NpHnxkTZ0P387RXoyqq1AVybPKvNfEO2hEo6U7Qv1zfe7dCv95NBB+plwK
+# WEwAPoVpdceDZNZ1zY8SdlalJPrXxGshuugfNJgvOuprAbD3+yqG7HtSOKmYCaFx
+# smxxrz64b5bV4RAT/mFHCoz+8LbH1cfebCTwv0KCyqBxPZySkwS0aXAnDU+3tTbR
+# yV8IpHCj7ArxES5k4MsiK8rxKBMhSVF+BmbTO77665E42FEHypS34lCh8zrTioPL
+# QHsCAwEAAaOCAYswggGHMA4GA1UdDwEB/wQEAwIHgDAMBgNVHRMBAf8EAjAAMBYG
+# A1UdJQEB/wQMMAoGCCsGAQUFBwMIMCAGA1UdIAQZMBcwCAYGZ4EMAQQCMAsGCWCG
+# SAGG/WwHATAfBgNVHSMEGDAWgBS6FtltTYUvcyl2mi91jGogj57IbzAdBgNVHQ4E
+# FgQUpbbvE+fvzdBkodVWqWUxo97V40kwWgYDVR0fBFMwUTBPoE2gS4ZJaHR0cDov
+# L2NybDMuZGlnaWNlcnQuY29tL0RpZ2lDZXJ0VHJ1c3RlZEc0UlNBNDA5NlNIQTI1
+# NlRpbWVTdGFtcGluZ0NBLmNybDCBkAYIKwYBBQUHAQEEgYMwgYAwJAYIKwYBBQUH
+# MAGGGGh0dHA6Ly9vY3NwLmRpZ2ljZXJ0LmNvbTBYBggrBgEFBQcwAoZMaHR0cDov
+# L2NhY2VydHMuZGlnaWNlcnQuY29tL0RpZ2lDZXJ0VHJ1c3RlZEc0UlNBNDA5NlNI
+# QTI1NlRpbWVTdGFtcGluZ0NBLmNydDANBgkqhkiG9w0BAQsFAAOCAgEAgRrW3qCp
+# tZgXvHCNT4o8aJzYJf/LLOTN6l0ikuyMIgKpuM+AqNnn48XtJoKKcS8Y3U623mzX
+# 4WCcK+3tPUiOuGu6fF29wmE3aEl3o+uQqhLXJ4Xzjh6S2sJAOJ9dyKAuJXglnSoF
+# eoQpmLZXeY/bJlYrsPOnvTcM2Jh2T1a5UsK2nTipgedtQVyMadG5K8TGe8+c+nji
+# kxp2oml101DkRBK+IA2eqUTQ+OVJdwhaIcW0z5iVGlS6ubzBaRm6zxbygzc0brBB
+# Jt3eWpdPM43UjXd9dUWhpVgmagNF3tlQtVCMr1a9TMXhRsUo063nQwBw3syYnhmJ
+# A+rUkTfvTVLzyWAhxFZH7doRS4wyw4jmWOK22z75X7BC1o/jF5HRqsBV44a/rCcs
+# QdCaM0qoNtS5cpZ+l3k4SF/Kwtw9Mt911jZnWon49qfH5U81PAC9vpwqbHkB3NpE
+# 5jreODsHXjlY9HxzMVWggBHLFAx+rrz+pOt5Zapo1iLKO+uagjVXKBbLafIymrLS
+# 2Dq4sUaGa7oX/cR3bBVsrquvczroSUa31X/MtjjA2Owc9bahuEMs305MfR5ocMB3
+# CtQC4Fxguyj/OOVSWtasFyIjTvTs0xf7UGv/B3cfcZdEQcm4RtNsMnxYL2dHZeUb
+# c7aZ+WssBkbvQR7w8F/g29mtkIBEr4AQQYowggauMIIElqADAgECAhAHNje3JFR8
+# 2Ees/ShmKl5bMA0GCSqGSIb3DQEBCwUAMGIxCzAJBgNVBAYTAlVTMRUwEwYDVQQK
+# EwxEaWdpQ2VydCBJbmMxGTAXBgNVBAsTEHd3dy5kaWdpY2VydC5jb20xITAfBgNV
+# BAMTGERpZ2lDZXJ0IFRydXN0ZWQgUm9vdCBHNDAeFw0yMjAzMjMwMDAwMDBaFw0z
+# NzAzMjIyMzU5NTlaMGMxCzAJBgNVBAYTAlVTMRcwFQYDVQQKEw5EaWdpQ2VydCwg
+# SW5jLjE7MDkGA1UEAxMyRGlnaUNlcnQgVHJ1c3RlZCBHNCBSU0E0MDk2IFNIQTI1
+# NiBUaW1lU3RhbXBpbmcgQ0EwggIiMA0GCSqGSIb3DQEBAQUAA4ICDwAwggIKAoIC
+# AQDGhjUGSbPBPXJJUVXHJQPE8pE3qZdRodbSg9GeTKJtoLDMg/la9hGhRBVCX6SI
+# 82j6ffOciQt/nR+eDzMfUBMLJnOWbfhXqAJ9/UO0hNoR8XOxs+4rgISKIhjf69o9
+# xBd/qxkrPkLcZ47qUT3w1lbU5ygt69OxtXXnHwZljZQp09nsad/ZkIdGAHvbREGJ
+# 3HxqV3rwN3mfXazL6IRktFLydkf3YYMZ3V+0VAshaG43IbtArF+y3kp9zvU5Emfv
+# DqVjbOSmxR3NNg1c1eYbqMFkdECnwHLFuk4fsbVYTXn+149zk6wsOeKlSNbwsDET
+# qVcplicu9Yemj052FVUmcJgmf6AaRyBD40NjgHt1biclkJg6OBGz9vae5jtb7IHe
+# IhTZgirHkr+g3uM+onP65x9abJTyUpURK1h0QCirc0PO30qhHGs4xSnzyqqWc0Jo
+# n7ZGs506o9UD4L/wojzKQtwYSH8UNM/STKvvmz3+DrhkKvp1KCRB7UK/BZxmSVJQ
+# 9FHzNklNiyDSLFc1eSuo80VgvCONWPfcYd6T/jnA+bIwpUzX6ZhKWD7TA4j+s4/T
+# Xkt2ElGTyYwMO1uKIqjBJgj5FBASA31fI7tk42PgpuE+9sJ0sj8eCXbsq11GdeJg
+# o1gJASgADoRU7s7pXcheMBK9Rp6103a50g5rmQzSM7TNsQIDAQABo4IBXTCCAVkw
+# EgYDVR0TAQH/BAgwBgEB/wIBADAdBgNVHQ4EFgQUuhbZbU2FL3MpdpovdYxqII+e
+# yG8wHwYDVR0jBBgwFoAU7NfjgtJxXWRM3y5nP+e6mK4cD08wDgYDVR0PAQH/BAQD
+# AgGGMBMGA1UdJQQMMAoGCCsGAQUFBwMIMHcGCCsGAQUFBwEBBGswaTAkBggrBgEF
+# BQcwAYYYaHR0cDovL29jc3AuZGlnaWNlcnQuY29tMEEGCCsGAQUFBzAChjVodHRw
+# Oi8vY2FjZXJ0cy5kaWdpY2VydC5jb20vRGlnaUNlcnRUcnVzdGVkUm9vdEc0LmNy
+# dDBDBgNVHR8EPDA6MDigNqA0hjJodHRwOi8vY3JsMy5kaWdpY2VydC5jb20vRGln
+# aUNlcnRUcnVzdGVkUm9vdEc0LmNybDAgBgNVHSAEGTAXMAgGBmeBDAEEAjALBglg
+# hkgBhv1sBwEwDQYJKoZIhvcNAQELBQADggIBAH1ZjsCTtm+YqUQiAX5m1tghQuGw
+# GC4QTRPPMFPOvxj7x1Bd4ksp+3CKDaopafxpwc8dB+k+YMjYC+VcW9dth/qEICU0
+# MWfNthKWb8RQTGIdDAiCqBa9qVbPFXONASIlzpVpP0d3+3J0FNf/q0+KLHqrhc1D
+# X+1gtqpPkWaeLJ7giqzl/Yy8ZCaHbJK9nXzQcAp876i8dU+6WvepELJd6f8oVInw
+# 1YpxdmXazPByoyP6wCeCRK6ZJxurJB4mwbfeKuv2nrF5mYGjVoarCkXJ38SNoOeY
+# +/umnXKvxMfBwWpx2cYTgAnEtp/Nh4cku0+jSbl3ZpHxcpzpSwJSpzd+k1OsOx0I
+# SQ+UzTl63f8lY5knLD0/a6fxZsNBzU+2QJshIUDQtxMkzdwdeDrknq3lNHGS1yZr
+# 5Dhzq6YBT70/O3itTK37xJV77QpfMzmHQXh6OOmc4d0j/R0o08f56PGYX/sr2H7y
+# Rp11LB4nLCbbbxV7HhmLNriT1ObyF5lZynDwN7+YAN8gFk8n+2BnFqFmut1VwDop
+# hrCYoCvtlUG3OtUVmDG0YgkPCr2B2RP+v6TR81fZvAT6gt4y3wSJ8ADNXcL50CN/
+# AAvkdgIm2fBldkKmKYcJRyvmfxqkhQ/8mJb2VVQrH4D6wPIOK+XW+6kvRBVK5xMO
+# Hds3OBqhK/bt1nz8MIIFjTCCBHWgAwIBAgIQDpsYjvnQLefv21DiCEAYWjANBgkq
+# hkiG9w0BAQwFADBlMQswCQYDVQQGEwJVUzEVMBMGA1UEChMMRGlnaUNlcnQgSW5j
+# MRkwFwYDVQQLExB3d3cuZGlnaWNlcnQuY29tMSQwIgYDVQQDExtEaWdpQ2VydCBB
+# c3N1cmVkIElEIFJvb3QgQ0EwHhcNMjIwODAxMDAwMDAwWhcNMzExMTA5MjM1OTU5
+# WjBiMQswCQYDVQQGEwJVUzEVMBMGA1UEChMMRGlnaUNlcnQgSW5jMRkwFwYDVQQL
+# ExB3d3cuZGlnaWNlcnQuY29tMSEwHwYDVQQDExhEaWdpQ2VydCBUcnVzdGVkIFJv
+# b3QgRzQwggIiMA0GCSqGSIb3DQEBAQUAA4ICDwAwggIKAoICAQC/5pBzaN675F1K
+# PDAiMGkz7MKnJS7JIT3yithZwuEppz1Yq3aaza57G4QNxDAf8xukOBbrVsaXbR2r
+# snnyyhHS5F/WBTxSD1Ifxp4VpX6+n6lXFllVcq9ok3DCsrp1mWpzMpTREEQQLt+C
+# 8weE5nQ7bXHiLQwb7iDVySAdYyktzuxeTsiT+CFhmzTrBcZe7FsavOvJz82sNEBf
+# sXpm7nfISKhmV1efVFiODCu3T6cw2Vbuyntd463JT17lNecxy9qTXtyOj4DatpGY
+# QJB5w3jHtrHEtWoYOAMQjdjUN6QuBX2I9YI+EJFwq1WCQTLX2wRzKm6RAXwhTNS8
+# rhsDdV14Ztk6MUSaM0C/CNdaSaTC5qmgZ92kJ7yhTzm1EVgX9yRcRo9k98FpiHaY
+# dj1ZXUJ2h4mXaXpI8OCiEhtmmnTK3kse5w5jrubU75KSOp493ADkRSWJtppEGSt+
+# wJS00mFt6zPZxd9LBADMfRyVw4/3IbKyEbe7f/LVjHAsQWCqsWMYRJUadmJ+9oCw
+# ++hkpjPRiQfhvbfmQ6QYuKZ3AeEPlAwhHbJUKSWJbOUOUlFHdL4mrLZBdd56rF+N
+# P8m800ERElvlEFDrMcXKchYiCd98THU/Y+whX8QgUWtvsauGi0/C1kVfnSD8oR7F
+# wI+isX4KJpn15GkvmB0t9dmpsh3lGwIDAQABo4IBOjCCATYwDwYDVR0TAQH/BAUw
+# AwEB/zAdBgNVHQ4EFgQU7NfjgtJxXWRM3y5nP+e6mK4cD08wHwYDVR0jBBgwFoAU
+# Reuir/SSy4IxLVGLp6chnfNtyA8wDgYDVR0PAQH/BAQDAgGGMHkGCCsGAQUFBwEB
+# BG0wazAkBggrBgEFBQcwAYYYaHR0cDovL29jc3AuZGlnaWNlcnQuY29tMEMGCCsG
+# AQUFBzAChjdodHRwOi8vY2FjZXJ0cy5kaWdpY2VydC5jb20vRGlnaUNlcnRBc3N1
+# cmVkSURSb290Q0EuY3J0MEUGA1UdHwQ+MDwwOqA4oDaGNGh0dHA6Ly9jcmwzLmRp
+# Z2ljZXJ0LmNvbS9EaWdpQ2VydEFzc3VyZWRJRFJvb3RDQS5jcmwwEQYDVR0gBAow
+# CDAGBgRVHSAAMA0GCSqGSIb3DQEBDAUAA4IBAQBwoL9DXFXnOF+go3QbPbYW1/e/
+# Vwe9mqyhhyzshV6pGrsi+IcaaVQi7aSId229GhT0E0p6Ly23OO/0/4C5+KH38nLe
+# JLxSA8hO0Cre+i1Wz/n096wwepqLsl7Uz9FDRJtDIeuWcqFItJnLnU+nBgMTdydE
+# 1Od/6Fmo8L8vC6bp8jQ87PcDx4eo0kxAGTVGamlUsLihVo7spNU96LHc/RzY9Hda
+# XFSMb++hUD38dglohJ9vytsgjTVgHAIDyyCwrFigDkBjxZgiwbJZ9VVrzyerbHbO
+# byMt9H5xaiNrIv8SuFQtJ37YOtnwtoeW/VvRXKwYw02fc7cBqZ9Xql4o4rmUMYID
+# djCCA3ICAQEwdzBjMQswCQYDVQQGEwJVUzEXMBUGA1UEChMORGlnaUNlcnQsIElu
+# Yy4xOzA5BgNVBAMTMkRpZ2lDZXJ0IFRydXN0ZWQgRzQgUlNBNDA5NiBTSEEyNTYg
+# VGltZVN0YW1waW5nIENBAhAFRK/zlJ0IOaa/2z9f5WEWMA0GCWCGSAFlAwQCAQUA
+# oIHRMBoGCSqGSIb3DQEJAzENBgsqhkiG9w0BCRABBDAcBgkqhkiG9w0BCQUxDxcN
+# MjQwOTA2MjAyNzExWjArBgsqhkiG9w0BCRACDDEcMBowGDAWBBRm8CsywsLJD4Jd
+# zqqKycZPGZzPQDAvBgkqhkiG9w0BCQQxIgQgXSdFKsIxhS4gvdZFC5i8csELx4EN
+# gje4K7DDRX8dz3AwNwYLKoZIhvcNAQkQAi8xKDAmMCQwIgQg0vbkbe10IszR1EBX
+# aEE2b4KK2lWarjMWr00amtQMeCgwDQYJKoZIhvcNAQEBBQAEggIAYX9aC647tWiS
+# rGwjsC+5s2CosHEwRzUG9YNI58OJgUfCwsfgMkgKWkSi/K7mumf5RHkU+P+HCwoy
+# kvIOZ7viK9fcAkK9zS3eKPUA6mGQS11yEnEhRUZbrrsG1uHQO+gSO2SgyLs8+3vX
+# /8+YEl1IkGbw4/oeLavq79jULQqZ6/00n0E0nFDmbprjFK4wUX4CoIqt8AAWCt4F
+# Az8XwvYxa63A2JQmeDzDAWR4lfNbREQaC3MdnqbnvQIBQUspJsn3t7zxU+ubzCez
+# kCkk+7Tt5FFCP9OJvc/BEv3HcXrTAoZ4VFfAwL9K1DQ4A3hbsvKlwV0OxZlhouMd
+# fGq+R8IGMsy7mGxeHx67nzKIr6Rjd426YsGskp5D3gE9shvH8i3GOTBi2Y9JUnaU
+# /KX+IMzKbvR0Y9echgTb17v3D/+fYzDD/kSGJcuQEIbJEyYsCDBF53xoKd6K0Pgz
+# 2drucT9otwOLUgGfR1N6lRwDtkMHYB25OMIKLYtcfHjQZn+Howq/TVUbp9ohhW1N
+# jim3nJfNvmRe2zN5476SOn86GzzrqxfAMCTtbZeim2ltOHxlnPUE8EJLdRFesKMK
+# 6izgaxptlT+MO0R8jx1VoOn+qbQPbNn2GCOUvh/yFkjwDLtFb/rNdoWMNrSMZDhV
+# mRCM17SwjW6qRmsrC7VSaSAgPsokYM0=
+# SIG # End signature block

INSTALLER

@@ -0,0 +1 @@
+pip

METADATA

@@ -0,0 +1,549 @@
+Metadata-Version: 2.1
+Name: torch
+Version: 2.6.0
+Summary: Tensors and Dynamic neural networks in Python with strong GPU acceleration
+Home-page: https://pytorch.org/
+Download-URL: https://github.com/pytorch/pytorch/tags
+Author: PyTorch Team
+Author-email: [email protected]
+License: BSD-3-Clause
+Keywords: pytorch,machine learning
+Classifier: Development Status :: 5 - Production/Stable
+Classifier: Intended Audience :: Developers
+Classifier: Intended Audience :: Education
+Classifier: Intended Audience :: Science/Research
+Classifier: License :: OSI Approved :: BSD License
+Classifier: Topic :: Scientific/Engineering
+Classifier: Topic :: Scientific/Engineering :: Mathematics
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Classifier: Topic :: Software Development
+Classifier: Topic :: Software Development :: Libraries
+Classifier: Topic :: Software Development :: Libraries :: Python Modules
+Classifier: Programming Language :: C++
+Classifier: Programming Language :: Python :: 3
+Classifier: Programming Language :: Python :: 3.9
+Classifier: Programming Language :: Python :: 3.10
+Classifier: Programming Language :: Python :: 3.11
+Classifier: Programming Language :: Python :: 3.12
+Classifier: Programming Language :: Python :: 3.13
+Requires-Python: >=3.9.0
+Description-Content-Type: text/markdown
+License-File: LICENSE
+License-File: NOTICE
+Requires-Dist: filelock
+Requires-Dist: typing-extensions>=4.10.0
+Requires-Dist: networkx
+Requires-Dist: jinja2
+Requires-Dist: fsspec
+Requires-Dist: nvidia-cuda-nvrtc-cu12==12.4.127; platform_system == "Linux" and platform_machine == "x86_64"
+Requires-Dist: nvidia-cuda-runtime-cu12==12.4.127; platform_system == "Linux" and platform_machine == "x86_64"
+Requires-Dist: nvidia-cuda-cupti-cu12==12.4.127; platform_system == "Linux" and platform_machine == "x86_64"
+Requires-Dist: nvidia-cudnn-cu12==9.1.0.70; platform_system == "Linux" and platform_machine == "x86_64"
+Requires-Dist: nvidia-cublas-cu12==12.4.5.8; platform_system == "Linux" and platform_machine == "x86_64"
+Requires-Dist: nvidia-cufft-cu12==11.2.1.3; platform_system == "Linux" and platform_machine == "x86_64"
+Requires-Dist: nvidia-curand-cu12==10.3.5.147; platform_system == "Linux" and platform_machine == "x86_64"
+Requires-Dist: nvidia-cusolver-cu12==11.6.1.9; platform_system == "Linux" and platform_machine == "x86_64"
+Requires-Dist: nvidia-cusparse-cu12==12.3.1.170; platform_system == "Linux" and platform_machine == "x86_64"
+Requires-Dist: nvidia-cusparselt-cu12==0.6.2; platform_system == "Linux" and platform_machine == "x86_64"
+Requires-Dist: nvidia-nccl-cu12==2.21.5; platform_system == "Linux" and platform_machine == "x86_64"
+Requires-Dist: nvidia-nvtx-cu12==12.4.127; platform_system == "Linux" and platform_machine == "x86_64"
+Requires-Dist: nvidia-nvjitlink-cu12==12.4.127; platform_system == "Linux" and platform_machine == "x86_64"
+Requires-Dist: triton==3.2.0; platform_system == "Linux" and platform_machine == "x86_64"
+Requires-Dist: setuptools; python_version >= "3.12"
+Requires-Dist: sympy==1.13.1; python_version >= "3.9"
+Provides-Extra: opt-einsum
+Requires-Dist: opt-einsum>=3.3; extra == "opt-einsum"
+Provides-Extra: optree
+Requires-Dist: optree>=0.13.0; extra == "optree"
+
+![PyTorch Logo](https://github.com/pytorch/pytorch/raw/main/docs/source/_static/img/pytorch-logo-dark.png)
+
+--------------------------------------------------------------------------------
+
+PyTorch is a Python package that provides two high-level features:
+- Tensor computation (like NumPy) with strong GPU acceleration
+- Deep neural networks built on a tape-based autograd system
+
+You can reuse your favorite Python packages such as NumPy, SciPy, and Cython to extend PyTorch when needed.
+
+Our trunk health (Continuous Integration signals) can be found at [hud.pytorch.org](https://hud.pytorch.org/ci/pytorch/pytorch/main).
+
+<!-- toc -->
+
+- [More About PyTorch](#more-about-pytorch)
+  - [A GPU-Ready Tensor Library](#a-gpu-ready-tensor-library)
+  - [Dynamic Neural Networks: Tape-Based Autograd](#dynamic-neural-networks-tape-based-autograd)
+  - [Python First](#python-first)
+  - [Imperative Experiences](#imperative-experiences)
+  - [Fast and Lean](#fast-and-lean)
+  - [Extensions Without Pain](#extensions-without-pain)
+- [Installation](#installation)
+  - [Binaries](#binaries)
+    - [NVIDIA Jetson Platforms](#nvidia-jetson-platforms)
+  - [From Source](#from-source)
+    - [Prerequisites](#prerequisites)
+      - [NVIDIA CUDA Support](#nvidia-cuda-support)
+      - [AMD ROCm Support](#amd-rocm-support)
+      - [Intel GPU Support](#intel-gpu-support)
+    - [Get the PyTorch Source](#get-the-pytorch-source)
+    - [Install Dependencies](#install-dependencies)
+    - [Install PyTorch](#install-pytorch)
+      - [Adjust Build Options (Optional)](#adjust-build-options-optional)
+  - [Docker Image](#docker-image)
+    - [Using pre-built images](#using-pre-built-images)
+    - [Building the image yourself](#building-the-image-yourself)
+  - [Building the Documentation](#building-the-documentation)
+  - [Previous Versions](#previous-versions)
+- [Getting Started](#getting-started)
+- [Resources](#resources)
+- [Communication](#communication)
+- [Releases and Contributing](#releases-and-contributing)
+- [The Team](#the-team)
+- [License](#license)
+
+<!-- tocstop -->
+
+## More About PyTorch
+
+[Learn the basics of PyTorch](https://pytorch.org/tutorials/beginner/basics/intro.html)
+
+At a granular level, PyTorch is a library that consists of the following components:
+
+| Component | Description |
+| ---- | --- |
+| [**torch**](https://pytorch.org/docs/stable/torch.html) | A Tensor library like NumPy, with strong GPU support |
+| [**torch.autograd**](https://pytorch.org/docs/stable/autograd.html) | A tape-based automatic differentiation library that supports all differentiable Tensor operations in torch |
+| [**torch.jit**](https://pytorch.org/docs/stable/jit.html) | A compilation stack (TorchScript) to create serializable and optimizable models from PyTorch code  |
+| [**torch.nn**](https://pytorch.org/docs/stable/nn.html) | A neural networks library deeply integrated with autograd designed for maximum flexibility |
+| [**torch.multiprocessing**](https://pytorch.org/docs/stable/multiprocessing.html) | Python multiprocessing, but with magical memory sharing of torch Tensors across processes. Useful for data loading and Hogwild training |
+| [**torch.utils**](https://pytorch.org/docs/stable/data.html) | DataLoader and other utility functions for convenience |
+
+Usually, PyTorch is used either as:
+
+- A replacement for NumPy to use the power of GPUs.
+- A deep learning research platform that provides maximum flexibility and speed.
+
+Elaborating Further:
+
+### A GPU-Ready Tensor Library
+
+If you use NumPy, then you have used Tensors (a.k.a. ndarray).
+
+![Tensor illustration](./docs/source/_static/img/tensor_illustration.png)
+
+PyTorch provides Tensors that can live either on the CPU or the GPU and accelerates the
+computation by a huge amount.
+
+We provide a wide variety of tensor routines to accelerate and fit your scientific computation needs
+such as slicing, indexing, mathematical operations, linear algebra, reductions.
+And they are fast!
+
+### Dynamic Neural Networks: Tape-Based Autograd
+
+PyTorch has a unique way of building neural networks: using and replaying a tape recorder.
+
+Most frameworks such as TensorFlow, Theano, Caffe, and CNTK have a static view of the world.
+One has to build a neural network and reuse the same structure again and again.
+Changing the way the network behaves means that one has to start from scratch.
+
+With PyTorch, we use a technique called reverse-mode auto-differentiation, which allows you to
+change the way your network behaves arbitrarily with zero lag or overhead. Our inspiration comes
+from several research papers on this topic, as well as current and past work such as
+[torch-autograd](https://github.com/twitter/torch-autograd),
+[autograd](https://github.com/HIPS/autograd),
+[Chainer](https://chainer.org), etc.
+
+While this technique is not unique to PyTorch, it's one of the fastest implementations of it to date.
+You get the best of speed and flexibility for your crazy research.
+
+![Dynamic graph](https://github.com/pytorch/pytorch/raw/main/docs/source/_static/img/dynamic_graph.gif)
+
+### Python First
+
+PyTorch is not a Python binding into a monolithic C++ framework.
+It is built to be deeply integrated into Python.
+You can use it naturally like you would use [NumPy](https://www.numpy.org/) / [SciPy](https://www.scipy.org/) / [scikit-learn](https://scikit-learn.org) etc.
+You can write your new neural network layers in Python itself, using your favorite libraries
+and use packages such as [Cython](https://cython.org/) and [Numba](http://numba.pydata.org/).
+Our goal is to not reinvent the wheel where appropriate.
+
+### Imperative Experiences
+
+PyTorch is designed to be intuitive, linear in thought, and easy to use.
+When you execute a line of code, it gets executed. There isn't an asynchronous view of the world.
+When you drop into a debugger or receive error messages and stack traces, understanding them is straightforward.
+The stack trace points to exactly where your code was defined.
+We hope you never spend hours debugging your code because of bad stack traces or asynchronous and opaque execution engines.
+
+### Fast and Lean
+
+PyTorch has minimal framework overhead. We integrate acceleration libraries
+such as [Intel MKL](https://software.intel.com/mkl) and NVIDIA ([cuDNN](https://developer.nvidia.com/cudnn), [NCCL](https://developer.nvidia.com/nccl)) to maximize speed.
+At the core, its CPU and GPU Tensor and neural network backends
+are mature and have been tested for years.
+
+Hence, PyTorch is quite fast — whether you run small or large neural networks.
+
+The memory usage in PyTorch is extremely efficient compared to Torch or some of the alternatives.
+We've written custom memory allocators for the GPU to make sure that
+your deep learning models are maximally memory efficient.
+This enables you to train bigger deep learning models than before.
+
+### Extensions Without Pain
+
+Writing new neural network modules, or interfacing with PyTorch's Tensor API was designed to be straightforward
+and with minimal abstractions.
+
+You can write new neural network layers in Python using the torch API
+[or your favorite NumPy-based libraries such as SciPy](https://pytorch.org/tutorials/advanced/numpy_extensions_tutorial.html).
+
+If you want to write your layers in C/C++, we provide a convenient extension API that is efficient and with minimal boilerplate.
+No wrapper code needs to be written. You can see [a tutorial here](https://pytorch.org/tutorials/advanced/cpp_extension.html) and [an example here](https://github.com/pytorch/extension-cpp).
+
+
+## Installation
+
+### Binaries
+Commands to install binaries via Conda or pip wheels are on our website: [https://pytorch.org/get-started/locally/](https://pytorch.org/get-started/locally/)
+
+
+#### NVIDIA Jetson Platforms
+
+Python wheels for NVIDIA's Jetson Nano, Jetson TX1/TX2, Jetson Xavier NX/AGX, and Jetson AGX Orin are provided [here](https://forums.developer.nvidia.com/t/pytorch-for-jetson-version-1-10-now-available/72048) and the L4T container is published [here](https://catalog.ngc.nvidia.com/orgs/nvidia/containers/l4t-pytorch)
+
+They require JetPack 4.2 and above, and [@dusty-nv](https://github.com/dusty-nv) and [@ptrblck](https://github.com/ptrblck) are maintaining them.
+
+
+### From Source
+
+#### Prerequisites
+If you are installing from source, you will need:
+- Python 3.9 or later
+- A compiler that fully supports C++17, such as clang or gcc (gcc 9.4.0 or newer is required, on Linux)
+- Visual Studio or Visual Studio Build Tool (Windows only)
+
+\* PyTorch CI uses Visual C++ BuildTools, which come with Visual Studio Enterprise,
+Professional, or Community Editions. You can also install the build tools from
+https://visualstudio.microsoft.com/visual-cpp-build-tools/. The build tools *do not*
+come with Visual Studio Code by default.
+
+\* We highly recommend installing an [Anaconda](https://www.anaconda.com/download) environment. You will get a high-quality BLAS library (MKL) and you get controlled dependency versions regardless of your Linux distro.
+
+An example of environment setup is shown below:
+
+* Linux:
+
+```bash
+$ source <CONDA_INSTALL_DIR>/bin/activate
+$ conda create -y -n <CONDA_NAME>
+$ conda activate <CONDA_NAME>
+```
+
+* Windows:
+
+```bash
+$ source <CONDA_INSTALL_DIR>\Scripts\activate.bat
+$ conda create -y -n <CONDA_NAME>
+$ conda activate <CONDA_NAME>
+$ call "C:\Program Files\Microsoft Visual Studio\<VERSION>\Community\VC\Auxiliary\Build\vcvarsall.bat" x64
+```
+
+##### NVIDIA CUDA Support
+If you want to compile with CUDA support, [select a supported version of CUDA from our support matrix](https://pytorch.org/get-started/locally/), then install the following:
+- [NVIDIA CUDA](https://developer.nvidia.com/cuda-downloads)
+- [NVIDIA cuDNN](https://developer.nvidia.com/cudnn) v8.5 or above
+- [Compiler](https://gist.github.com/ax3l/9489132) compatible with CUDA
+
+Note: You could refer to the [cuDNN Support Matrix](https://docs.nvidia.com/deeplearning/cudnn/reference/support-matrix.html) for cuDNN versions with the various supported CUDA, CUDA driver and NVIDIA hardware
+
+If you want to disable CUDA support, export the environment variable `USE_CUDA=0`.
+Other potentially useful environment variables may be found in `setup.py`.
+
+If you are building for NVIDIA's Jetson platforms (Jetson Nano, TX1, TX2, AGX Xavier), Instructions to install PyTorch for Jetson Nano are [available here](https://devtalk.nvidia.com/default/topic/1049071/jetson-nano/pytorch-for-jetson-nano/)
+
+##### AMD ROCm Support
+If you want to compile with ROCm support, install
+- [AMD ROCm](https://rocm.docs.amd.com/en/latest/deploy/linux/quick_start.html) 4.0 and above installation
+- ROCm is currently supported only for Linux systems.
+
+By default the build system expects ROCm to be installed in `/opt/rocm`. If ROCm is installed in a different directory, the `ROCM_PATH` environment variable must be set to the ROCm installation directory. The build system automatically detects the AMD GPU architecture. Optionally, the AMD GPU architecture can be explicitly set with the `PYTORCH_ROCM_ARCH` environment variable [AMD GPU architecture](https://rocm.docs.amd.com/projects/install-on-linux/en/latest/reference/system-requirements.html#supported-gpus)
+
+If you want to disable ROCm support, export the environment variable `USE_ROCM=0`.
+Other potentially useful environment variables may be found in `setup.py`.
+
+##### Intel GPU Support
+If you want to compile with Intel GPU support, follow these
+- [PyTorch Prerequisites for Intel GPUs](https://www.intel.com/content/www/us/en/developer/articles/tool/pytorch-prerequisites-for-intel-gpus.html) instructions.
+- Intel GPU is supported for Linux and Windows.
+
+If you want to disable Intel GPU support, export the environment variable `USE_XPU=0`.
+Other potentially useful environment variables may be found in `setup.py`.
+
+#### Get the PyTorch Source
+```bash
+git clone --recursive https://github.com/pytorch/pytorch
+cd pytorch
+# if you are updating an existing checkout
+git submodule sync
+git submodule update --init --recursive
+```
+
+#### Install Dependencies
+
+**Common**
+
+```bash
+conda install cmake ninja
+# Run this command from the PyTorch directory after cloning the source code using the “Get the PyTorch Source“ section below
+pip install -r requirements.txt
+```
+
+**On Linux**
+
+```bash
+pip install mkl-static mkl-include
+# CUDA only: Add LAPACK support for the GPU if needed
+conda install -c pytorch magma-cuda121  # or the magma-cuda* that matches your CUDA version from https://anaconda.org/pytorch/repo
+
+# (optional) If using torch.compile with inductor/triton, install the matching version of triton
+# Run from the pytorch directory after cloning
+# For Intel GPU support, please explicitly `export USE_XPU=1` before running command.
+make triton
+```
+
+**On MacOS**
+
+```bash
+# Add this package on intel x86 processor machines only
+pip install mkl-static mkl-include
+# Add these packages if torch.distributed is needed
+conda install pkg-config libuv
+```
+
+**On Windows**
+
+```bash
+pip install mkl-static mkl-include
+# Add these packages if torch.distributed is needed.
+# Distributed package support on Windows is a prototype feature and is subject to changes.
+conda install -c conda-forge libuv=1.39
+```
+
+#### Install PyTorch
+**On Linux**
+
+If you would like to compile PyTorch with [new C++ ABI](https://gcc.gnu.org/onlinedocs/libstdc++/manual/using_dual_abi.html) enabled, then first run this command:
+```bash
+export _GLIBCXX_USE_CXX11_ABI=1
+```
+
+Please **note** that starting from PyTorch 2.5, the PyTorch build with XPU supports both new and old C++ ABIs. Previously, XPU only supported the new C++ ABI. If you want to compile with Intel GPU support, please follow [Intel GPU Support](#intel-gpu-support).
+
+If you're compiling for AMD ROCm then first run this command:
+```bash
+# Only run this if you're compiling for ROCm
+python tools/amd_build/build_amd.py
+```
+
+Install PyTorch
+```bash
+export CMAKE_PREFIX_PATH="${CONDA_PREFIX:-'$(dirname $(which conda))/../'}:${CMAKE_PREFIX_PATH}"
+python setup.py develop
+```
+
+**On macOS**
+
+```bash
+python3 setup.py develop
+```
+
+**On Windows**
+
+If you want to build legacy python code, please refer to [Building on legacy code and CUDA](https://github.com/pytorch/pytorch/blob/main/CONTRIBUTING.md#building-on-legacy-code-and-cuda)
+
+**CPU-only builds**
+
+In this mode PyTorch computations will run on your CPU, not your GPU
+
+```cmd
+python setup.py develop
+```
+
+Note on OpenMP: The desired OpenMP implementation is Intel OpenMP (iomp). In order to link against iomp, you'll need to manually download the library and set up the building environment by tweaking `CMAKE_INCLUDE_PATH` and `LIB`. The instruction [here](https://github.com/pytorch/pytorch/blob/main/docs/source/notes/windows.rst#building-from-source) is an example for setting up both MKL and Intel OpenMP. Without these configurations for CMake, Microsoft Visual C OpenMP runtime (vcomp) will be used.
+
+**CUDA based build**
+
+In this mode PyTorch computations will leverage your GPU via CUDA for faster number crunching
+
+[NVTX](https://docs.nvidia.com/gameworks/content/gameworkslibrary/nvtx/nvidia_tools_extension_library_nvtx.htm) is needed to build Pytorch with CUDA.
+NVTX is a part of CUDA distributive, where it is called "Nsight Compute". To install it onto an already installed CUDA run CUDA installation once again and check the corresponding checkbox.
+Make sure that CUDA with Nsight Compute is installed after Visual Studio.
+
+Currently, VS 2017 / 2019, and Ninja are supported as the generator of CMake. If `ninja.exe` is detected in `PATH`, then Ninja will be used as the default generator, otherwise, it will use VS 2017 / 2019.
+<br/> If Ninja is selected as the generator, the latest MSVC will get selected as the underlying toolchain.
+
+Additional libraries such as
+[Magma](https://developer.nvidia.com/magma), [oneDNN, a.k.a. MKLDNN or DNNL](https://github.com/oneapi-src/oneDNN), and [Sccache](https://github.com/mozilla/sccache) are often needed. Please refer to the [installation-helper](https://github.com/pytorch/pytorch/tree/main/.ci/pytorch/win-test-helpers/installation-helpers) to install them.
+
+You can refer to the [build_pytorch.bat](https://github.com/pytorch/pytorch/blob/main/.ci/pytorch/win-test-helpers/build_pytorch.bat) script for some other environment variables configurations
+
+
+```cmd
+cmd
+
+:: Set the environment variables after you have downloaded and unzipped the mkl package,
+:: else CMake would throw an error as `Could NOT find OpenMP`.
+set CMAKE_INCLUDE_PATH={Your directory}\mkl\include
+set LIB={Your directory}\mkl\lib;%LIB%
+
+:: Read the content in the previous section carefully before you proceed.
+:: [Optional] If you want to override the underlying toolset used by Ninja and Visual Studio with CUDA, please run the following script block.
+:: "Visual Studio 2019 Developer Command Prompt" will be run automatically.
+:: Make sure you have CMake >= 3.12 before you do this when you use the Visual Studio generator.
+set CMAKE_GENERATOR_TOOLSET_VERSION=14.27
+set DISTUTILS_USE_SDK=1
+for /f "usebackq tokens=*" %i in (`"%ProgramFiles(x86)%\Microsoft Visual Studio\Installer\vswhere.exe" -version [15^,17^) -products * -latest -property installationPath`) do call "%i\VC\Auxiliary\Build\vcvarsall.bat" x64 -vcvars_ver=%CMAKE_GENERATOR_TOOLSET_VERSION%
+
+:: [Optional] If you want to override the CUDA host compiler
+set CUDAHOSTCXX=C:\Program Files (x86)\Microsoft Visual Studio\2019\Community\VC\Tools\MSVC\14.27.29110\bin\HostX64\x64\cl.exe
+
+python setup.py develop
+
+```
+
+##### Adjust Build Options (Optional)
+
+You can adjust the configuration of cmake variables optionally (without building first), by doing
+the following. For example, adjusting the pre-detected directories for CuDNN or BLAS can be done
+with such a step.
+
+On Linux
+```bash
+export CMAKE_PREFIX_PATH="${CONDA_PREFIX:-'$(dirname $(which conda))/../'}:${CMAKE_PREFIX_PATH}"
+python setup.py build --cmake-only
+ccmake build  # or cmake-gui build
+```
+
+On macOS
+```bash
+export CMAKE_PREFIX_PATH="${CONDA_PREFIX:-'$(dirname $(which conda))/../'}:${CMAKE_PREFIX_PATH}"
+MACOSX_DEPLOYMENT_TARGET=10.9 CC=clang CXX=clang++ python setup.py build --cmake-only
+ccmake build  # or cmake-gui build
+```
+
+### Docker Image
+
+#### Using pre-built images
+
+You can also pull a pre-built docker image from Docker Hub and run with docker v19.03+
+
+```bash
+docker run --gpus all --rm -ti --ipc=host pytorch/pytorch:latest
+```
+
+Please note that PyTorch uses shared memory to share data between processes, so if torch multiprocessing is used (e.g.
+for multithreaded data loaders) the default shared memory segment size that container runs with is not enough, and you
+should increase shared memory size either with `--ipc=host` or `--shm-size` command line options to `nvidia-docker run`.
+
+#### Building the image yourself
+
+**NOTE:** Must be built with a docker version > 18.06
+
+The `Dockerfile` is supplied to build images with CUDA 11.1 support and cuDNN v8.
+You can pass `PYTHON_VERSION=x.y` make variable to specify which Python version is to be used by Miniconda, or leave it
+unset to use the default.
+
+```bash
+make -f docker.Makefile
+# images are tagged as docker.io/${your_docker_username}/pytorch
+```
+
+You can also pass the `CMAKE_VARS="..."` environment variable to specify additional CMake variables to be passed to CMake during the build.
+See [setup.py](./setup.py) for the list of available variables.
+
+```bash
+make -f docker.Makefile
+```
+
+### Building the Documentation
+
+To build documentation in various formats, you will need [Sphinx](http://www.sphinx-doc.org) and the
+readthedocs theme.
+
+```bash
+cd docs/
+pip install -r requirements.txt
+make html
+make serve
+```
+
+Run `make` to get a list of all available output formats.
+
+If you get a katex error run `npm install katex`.  If it persists, try
+`npm install -g katex`
+
+> Note: if you installed `nodejs` with a different package manager (e.g.,
+`conda`) then `npm` will probably install a version of `katex` that is not
+compatible with your version of `nodejs` and doc builds will fail.
+A combination of versions that is known to work is `[email protected]` and
+`[email protected]`. To install the latter with `npm` you can run
+```npm install -g [email protected]```
+
+### Previous Versions
+
+Installation instructions and binaries for previous PyTorch versions may be found
+on [our website](https://pytorch.org/previous-versions).
+
+
+## Getting Started
+
+Three-pointers to get you started:
+- [Tutorials: get you started with understanding and using PyTorch](https://pytorch.org/tutorials/)
+- [Examples: easy to understand PyTorch code across all domains](https://github.com/pytorch/examples)
+- [The API Reference](https://pytorch.org/docs/)
+- [Glossary](https://github.com/pytorch/pytorch/blob/main/GLOSSARY.md)
+
+## Resources
+
+* [PyTorch.org](https://pytorch.org/)
+* [PyTorch Tutorials](https://pytorch.org/tutorials/)
+* [PyTorch Examples](https://github.com/pytorch/examples)
+* [PyTorch Models](https://pytorch.org/hub/)
+* [Intro to Deep Learning with PyTorch from Udacity](https://www.udacity.com/course/deep-learning-pytorch--ud188)
+* [Intro to Machine Learning with PyTorch from Udacity](https://www.udacity.com/course/intro-to-machine-learning-nanodegree--nd229)
+* [Deep Neural Networks with PyTorch from Coursera](https://www.coursera.org/learn/deep-neural-networks-with-pytorch)
+* [PyTorch Twitter](https://twitter.com/PyTorch)
+* [PyTorch Blog](https://pytorch.org/blog/)
+* [PyTorch YouTube](https://www.youtube.com/channel/UCWXI5YeOsh03QvJ59PMaXFw)
+
+## Communication
+* Forums: Discuss implementations, research, etc. https://discuss.pytorch.org
+* GitHub Issues: Bug reports, feature requests, install issues, RFCs, thoughts, etc.
+* Slack: The [PyTorch Slack](https://pytorch.slack.com/) hosts a primary audience of moderate to experienced PyTorch users and developers for general chat, online discussions, collaboration, etc. If you are a beginner looking for help, the primary medium is [PyTorch Forums](https://discuss.pytorch.org). If you need a slack invite, please fill this form: https://goo.gl/forms/PP1AGvNHpSaJP8to1
+* Newsletter: No-noise, a one-way email newsletter with important announcements about PyTorch. You can sign-up here: https://eepurl.com/cbG0rv
+* Facebook Page: Important announcements about PyTorch. https://www.facebook.com/pytorch
+* For brand guidelines, please visit our website at [pytorch.org](https://pytorch.org/)
+
+## Releases and Contributing
+
+Typically, PyTorch has three minor releases a year. Please let us know if you encounter a bug by [filing an issue](https://github.com/pytorch/pytorch/issues).
+
+We appreciate all contributions. If you are planning to contribute back bug-fixes, please do so without any further discussion.
+
+If you plan to contribute new features, utility functions, or extensions to the core, please first open an issue and discuss the feature with us.
+Sending a PR without discussion might end up resulting in a rejected PR because we might be taking the core in a different direction than you might be aware of.
+
+To learn more about making a contribution to Pytorch, please see our [Contribution page](CONTRIBUTING.md). For more information about PyTorch releases, see [Release page](RELEASE.md).
+
+## The Team
+
+PyTorch is a community-driven project with several skillful engineers and researchers contributing to it.
+
+PyTorch is currently maintained by [Soumith Chintala](http://soumith.ch), [Gregory Chanan](https://github.com/gchanan), [Dmytro Dzhulgakov](https://github.com/dzhulgakov), [Edward Yang](https://github.com/ezyang), and [Nikita Shulga](https://github.com/malfet) with major contributions coming from hundreds of talented individuals in various forms and means.
+A non-exhaustive but growing list needs to mention: [Trevor Killeen](https://github.com/killeent), [Sasank Chilamkurthy](https://github.com/chsasank), [Sergey Zagoruyko](https://github.com/szagoruyko), [Adam Lerer](https://github.com/adamlerer), [Francisco Massa](https://github.com/fmassa), [Alykhan Tejani](https://github.com/alykhantejani), [Luca Antiga](https://github.com/lantiga), [Alban Desmaison](https://github.com/albanD), [Andreas Koepf](https://github.com/andreaskoepf), [James Bradbury](https://github.com/jamesb93), [Zeming Lin](https://github.com/ebetica), [Yuandong Tian](https://github.com/yuandong-tian), [Guillaume Lample](https://github.com/glample), [Marat Dukhan](https://github.com/Maratyszcza), [Natalia Gimelshein](https://github.com/ngimel), [Christian Sarofeen](https://github.com/csarofeen), [Martin Raison](https://github.com/martinraison), [Edward Yang](https://github.com/ezyang), [Zachary Devito](https://github.com/zdevito).
+
+Note: This project is unrelated to [hughperkins/pytorch](https://github.com/hughperkins/pytorch) with the same name. Hugh is a valuable contributor to the Torch community and has helped with many things Torch and PyTorch.
+
+## License
+
+PyTorch has a BSD-style license, as found in the [LICENSE](LICENSE) file.

NOTICE

@@ -0,0 +1,456 @@
+=======================================================================
+Software under third_party
+=======================================================================
+Software libraries under third_party are provided as github submodule
+links, and their content is not part of the Caffe2 codebase. Their
+licences can be found under the respective software repositories.
+
+=======================================================================
+Earlier BSD License
+=======================================================================
+Early development of Caffe2 in 2015 and early 2016 is licensed under the
+BSD license. The license is attached below:
+
+All contributions by Facebook:
+Copyright (c) 2016 Facebook Inc.
+
+All contributions by Google:
+Copyright (c) 2015 Google Inc.
+All rights reserved.
+
+All contributions by Yangqing Jia:
+Copyright (c) 2015 Yangqing Jia
+All rights reserved.
+
+All contributions by Kakao Brain:
+Copyright 2019-2020 Kakao Brain
+
+All other contributions:
+Copyright(c) 2015, 2016 the respective contributors
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are met:
+
+1. Redistributions of source code must retain the above copyright notice, this
+   list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright notice,
+   this list of conditions and the following disclaimer in the documentation
+   and/or other materials provided with the distribution.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
+ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR
+ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+(INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
+LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
+ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+
+=======================================================================
+Caffe's BSD License
+=======================================================================
+Some parts of the caffe2 code is derived from the original Caffe code, which is
+created by Yangqing Jia and is now a BSD-licensed open-source project. The Caffe
+license is as follows:
+
+COPYRIGHT
+
+All contributions by the University of California:
+Copyright (c) 2014, The Regents of the University of California (Regents)
+All rights reserved.
+
+All other contributions:
+Copyright (c) 2014, the respective contributors
+All rights reserved.
+
+Caffe uses a shared copyright model: each contributor holds copyright over
+their contributions to Caffe. The project versioning records all such
+contribution and copyright details. If a contributor wants to further mark
+their specific copyright on a particular contribution, they should indicate
+their copyright solely in the commit message of the change when it is
+committed.
+
+LICENSE
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are met:
+
+1. Redistributions of source code must retain the above copyright notice, this
+   list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright notice,
+   this list of conditions and the following disclaimer in the documentation
+   and/or other materials provided with the distribution.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
+ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR
+ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+(INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
+LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
+ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+CONTRIBUTION AGREEMENT
+
+By contributing to the BVLC/caffe repository through pull-request, comment,
+or otherwise, the contributor releases their content to the
+license and copyright terms herein.
+
+=======================================================================
+Caffe2's Apache License
+=======================================================================
+
+This repo contains Caffe2 code, which was previously licensed under
+Apache License Version 2.0:
+
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+=======================================================================
+Cephes's 3-Clause BSD License
+=======================================================================
+
+Code derived from implementations in the Cephes Math Library should mention
+its derivation and reference the following license:
+
+   3-Clause BSD License for the Cephes Math Library
+   Copyright (c) 2018, Steven Moshier
+   All rights reserved.
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions are met:
+
+   * Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+   * Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+   * Neither the name of the nor the
+   names of its contributors may be used to endorse or promote products
+   derived from this software without specific prior written permission.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
+   ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+   WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+   DISCLAIMED. IN NO EVENT SHALL Steven Moshier BE LIABLE FOR ANY
+   DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+   (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
+   LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
+   ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+
+=======================================================================
+SciPy's 3-Clause BSD License
+=======================================================================
+
+Code derived from implementations in SciPy should mention its derivation
+and reference the following license:
+
+   Copyright (c) 2001-2002 Enthought, Inc.  2003-2019, SciPy Developers.
+   All rights reserved.
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   1. Redistributions of source code must retain the above copyright
+     notice, this list of conditions and the following disclaimer.
+
+   2. Redistributions in binary form must reproduce the above
+     copyright notice, this list of conditions and the following
+     disclaimer in the documentation and/or other materials provided
+     with the distribution.
+
+   3. Neither the name of the copyright holder nor the names of its
+     contributors may be used to endorse or promote products derived
+     from this software without specific prior written permission.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+=======================================================================
+Boost's 1.0 Software License
+=======================================================================
+
+Code derived from implementations in Boost 1.0 should mention its
+derivation and reference the following license:
+
+   Boost Software License - Version 1.0 - August 17th, 2003
+
+   Permission is hereby granted, free of charge, to any person or organization
+   obtaining a copy of the software and accompanying documentation covered by
+   this license (the "Software") to use, reproduce, display, distribute,
+   execute, and transmit the Software, and to prepare derivative works of the
+   Software, and to permit third-parties to whom the Software is furnished to
+   do so, all subject to the following:
+
+   The copyright notices in the Software and this entire statement, including
+   the above license grant, this restriction and the following disclaimer,
+   must be included in all copies of the Software, in whole or in part, and
+   all derivative works of the Software, unless such copies or derivative
+   works are solely in the form of machine-executable object code generated by
+   a source language processor.
+
+   THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+   IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+   FITNESS FOR A PARTICULAR PURPOSE, TITLE AND NON-INFRINGEMENT. IN NO EVENT
+   SHALL THE COPYRIGHT HOLDERS OR ANYONE DISTRIBUTING THE SOFTWARE BE LIABLE
+   FOR ANY DAMAGES OR OTHER LIABILITY, WHETHER IN CONTRACT, TORT OR OTHERWISE,
+   ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
+   DEALINGS IN THE SOFTWARE.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.
+
+=======================================================================
+PILLOW-SIMD Software License
+=======================================================================
+
+Code derived from implementations in PILLOW-SIMD should mention its derivation
+and reference the following license:
+
+    The Python Imaging Library (PIL) is
+
+        Copyright © 1997-2011 by Secret Labs AB
+        Copyright © 1995-2011 by Fredrik Lundh
+
+    Pillow is the friendly PIL fork. It is
+
+        Copyright © 2010-2022 by Alex Clark and contributors
+
+    Like PIL, Pillow is licensed under the open source HPND License:
+
+    By obtaining, using, and/or copying this software and/or its associated
+    documentation, you agree that you have read, understood, and will comply
+    with the following terms and conditions:
+
+    Permission to use, copy, modify, and distribute this software and its
+    associated documentation for any purpose and without fee is hereby granted,
+    provided that the above copyright notice appears in all copies, and that
+    both that copyright notice and this permission notice appear in supporting
+    documentation, and that the name of Secret Labs AB or the author not be
+    used in advertising or publicity pertaining to distribution of the software
+    without specific, written prior permission.
+
+    SECRET LABS AB AND THE AUTHOR DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS
+    SOFTWARE, INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS.
+    IN NO EVENT SHALL SECRET LABS AB OR THE AUTHOR BE LIABLE FOR ANY SPECIAL,
+    INDIRECT OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM
+    LOSS OF USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE
+    OR OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
+    PERFORMANCE OF THIS SOFTWARE.

WHEEL

@@ -0,0 +1,5 @@
+Wheel-Version: 1.0
+Generator: setuptools (72.1.0)
+Root-Is-Purelib: false
+Tag: cp312-cp312-win_amd64
+

__init__.py

@@ -0,0 +1,10 @@
+"""torchgen
+
+This module contains codegeneration utilities for PyTorch. It is used to
+build PyTorch from source, but may also be used for out-of-tree projects
+that extend PyTorch.
+
+Note well that we provide no BC guarantees for torchgen. If you're interested
+in using torchgen and want the PyTorch team to be aware, please reach out
+on GitHub.
+"""

activate

@@ -0,0 +1,70 @@
+# This file must be used with "source bin/activate" *from bash*
+# You cannot run it directly
+
+deactivate () {
+    # reset old environment variables
+    if [ -n "${_OLD_VIRTUAL_PATH:-}" ] ; then
+        PATH="${_OLD_VIRTUAL_PATH:-}"
+        export PATH
+        unset _OLD_VIRTUAL_PATH
+    fi
+    if [ -n "${_OLD_VIRTUAL_PYTHONHOME:-}" ] ; then
+        PYTHONHOME="${_OLD_VIRTUAL_PYTHONHOME:-}"
+        export PYTHONHOME
+        unset _OLD_VIRTUAL_PYTHONHOME
+    fi
+
+    # Call hash to forget past commands. Without forgetting
+    # past commands the $PATH changes we made may not be respected
+    hash -r 2> /dev/null
+
+    if [ -n "${_OLD_VIRTUAL_PS1:-}" ] ; then
+        PS1="${_OLD_VIRTUAL_PS1:-}"
+        export PS1
+        unset _OLD_VIRTUAL_PS1
+    fi
+
+    unset VIRTUAL_ENV
+    unset VIRTUAL_ENV_PROMPT
+    if [ ! "${1:-}" = "nondestructive" ] ; then
+    # Self destruct!
+        unset -f deactivate
+    fi
+}
+
+# unset irrelevant variables
+deactivate nondestructive
+
+# on Windows, a path can contain colons and backslashes and has to be converted:
+if [ "${OSTYPE:-}" = "cygwin" ] || [ "${OSTYPE:-}" = "msys" ] ; then
+    # transform D:\path\to\venv to /d/path/to/venv on MSYS
+    # and to /cygdrive/d/path/to/venv on Cygwin
+    export VIRTUAL_ENV=$(cygpath "C:\Users\migue\Downloads\Codigo\AI\HuggingModelAI\IAservicies\myenv")
+else
+    # use the path as-is
+    export VIRTUAL_ENV="C:\Users\migue\Downloads\Codigo\AI\HuggingModelAI\IAservicies\myenv"
+fi
+
+_OLD_VIRTUAL_PATH="$PATH"
+PATH="$VIRTUAL_ENV/Scripts:$PATH"
+export PATH
+
+# unset PYTHONHOME if set
+# this will fail if PYTHONHOME is set to the empty string (which is bad anyway)
+# could use `if (set -u; : $PYTHONHOME) ;` in bash
+if [ -n "${PYTHONHOME:-}" ] ; then
+    _OLD_VIRTUAL_PYTHONHOME="${PYTHONHOME:-}"
+    unset PYTHONHOME
+fi
+
+if [ -z "${VIRTUAL_ENV_DISABLE_PROMPT:-}" ] ; then
+    _OLD_VIRTUAL_PS1="${PS1:-}"
+    PS1="(myenv) ${PS1:-}"
+    export PS1
+    VIRTUAL_ENV_PROMPT="(myenv) "
+    export VIRTUAL_ENV_PROMPT
+fi
+
+# Call hash to forget past commands. Without forgetting
+# past commands the $PATH changes we made may not be respected
+hash -r 2> /dev/null

activate.bat

@@ -0,0 +1,34 @@
+@echo off
+
+rem This file is UTF-8 encoded, so we need to update the current code page while executing it
+for /f "tokens=2 delims=:." %%a in ('"%SystemRoot%\System32\chcp.com"') do (
+    set _OLD_CODEPAGE=%%a
+)
+if defined _OLD_CODEPAGE (
+    "%SystemRoot%\System32\chcp.com" 65001 > nul
+)
+
+set VIRTUAL_ENV=C:\Users\migue\Downloads\Codigo\AI\HuggingModelAI\IAservicies\myenv
+
+if not defined PROMPT set PROMPT=$P$G
+
+if defined _OLD_VIRTUAL_PROMPT set PROMPT=%_OLD_VIRTUAL_PROMPT%
+if defined _OLD_VIRTUAL_PYTHONHOME set PYTHONHOME=%_OLD_VIRTUAL_PYTHONHOME%
+
+set _OLD_VIRTUAL_PROMPT=%PROMPT%
+set PROMPT=(myenv) %PROMPT%
+
+if defined PYTHONHOME set _OLD_VIRTUAL_PYTHONHOME=%PYTHONHOME%
+set PYTHONHOME=
+
+if defined _OLD_VIRTUAL_PATH set PATH=%_OLD_VIRTUAL_PATH%
+if not defined _OLD_VIRTUAL_PATH set _OLD_VIRTUAL_PATH=%PATH%
+
+set PATH=%VIRTUAL_ENV%\Scripts;%PATH%
+set VIRTUAL_ENV_PROMPT=(myenv) 
+
+:END
+if defined _OLD_CODEPAGE (
+    "%SystemRoot%\System32\chcp.com" %_OLD_CODEPAGE% > nul
+    set _OLD_CODEPAGE=
+)

code_template.py

@@ -0,0 +1,103 @@
+from __future__ import annotations
+
+import re
+from typing import TYPE_CHECKING
+
+
+if TYPE_CHECKING:
+    from collections.abc import Mapping, Sequence
+
+
+# match $identifier or ${identifier} and replace with value in env
+# If this identifier is at the beginning of whitespace on a line
+# and its value is a list then it is treated as
+# block substitution by indenting to that depth and putting each element
+# of the list on its own line
+# if the identifier is on a line starting with non-whitespace and a list
+# then it is comma separated ${,foo} will insert a comma before the list
+# if this list is not empty and ${foo,} will insert one after.
+
+
+class CodeTemplate:
+    substitution_str = r"(^[^\n\S]*)?\$([^\d\W]\w*|\{,?[^\d\W]\w*\,?})"
+    substitution = re.compile(substitution_str, re.MULTILINE)
+
+    pattern: str
+    filename: str
+
+    @staticmethod
+    def from_file(filename: str) -> CodeTemplate:
+        with open(filename) as f:
+            return CodeTemplate(f.read(), filename)
+
+    def __init__(self, pattern: str, filename: str = "") -> None:
+        self.pattern = pattern
+        self.filename = filename
+
+    def substitute(
+        self, env: Mapping[str, object] | None = None, **kwargs: object
+    ) -> str:
+        if env is None:
+            env = {}
+
+        def lookup(v: str) -> object:
+            assert env is not None
+            return kwargs[v] if v in kwargs else env[v]
+
+        def indent_lines(indent: str, v: Sequence[object]) -> str:
+            return "".join(
+                [indent + l + "\n" for e in v for l in str(e).splitlines()]
+            ).rstrip()
+
+        def replace(match: re.Match[str]) -> str:
+            indent = match.group(1)
+            key = match.group(2)
+            comma_before = ""
+            comma_after = ""
+            if key[0] == "{":
+                key = key[1:-1]
+                if key[0] == ",":
+                    comma_before = ", "
+                    key = key[1:]
+                if key[-1] == ",":
+                    comma_after = ", "
+                    key = key[:-1]
+            v = lookup(key)
+            if indent is not None:
+                if not isinstance(v, list):
+                    v = [v]
+                return indent_lines(indent, v)
+            elif isinstance(v, list):
+                middle = ", ".join([str(x) for x in v])
+                if len(v) == 0:
+                    return middle
+                return comma_before + middle + comma_after
+            else:
+                return str(v)
+
+        return self.substitution.sub(replace, self.pattern)
+
+
+if __name__ == "__main__":
+    c = CodeTemplate(
+        """\
+    int foo($args) {
+
+        $bar
+            $bar
+        $a+$b
+    }
+    int commatest(int a${,stuff})
+    int notest(int a${,empty,})
+    """
+    )
+    print(
+        c.substitute(
+            args=["hi", 8],
+            bar=["what", 7],
+            a=3,
+            b=4,
+            stuff=["things...", "others"],
+            empty=[],
+        )
+    )

context.py

@@ -0,0 +1,134 @@
+from __future__ import annotations
+
+import contextlib
+import functools
+from typing import Any, Callable, Optional, TYPE_CHECKING, TypeVar, Union
+
+import torchgen.local as local
+from torchgen.model import (
+    BackendIndex,
+    DispatchKey,
+    NativeFunction,
+    NativeFunctionsGroup,
+    NativeFunctionsViewGroup,
+)
+from torchgen.utils import context, S, T
+
+
+if TYPE_CHECKING:
+    from collections.abc import Iterator
+
+
+# Helper functions for defining generators on things in the model
+
+F = TypeVar(
+    "F",
+    NativeFunction,
+    NativeFunctionsGroup,
+    NativeFunctionsViewGroup,
+    Union[NativeFunction, NativeFunctionsGroup],
+    Union[NativeFunction, NativeFunctionsViewGroup],
+)
+
+F2 = TypeVar(
+    "F2",
+    NativeFunction,
+    NativeFunctionsGroup,
+    Optional[NativeFunction],
+    bool,
+    str,
+)
+
+F3 = TypeVar("F3", tuple[NativeFunction, Any], list[NativeFunction])
+
+
+@contextlib.contextmanager
+def native_function_manager(
+    g: NativeFunctionsGroup | NativeFunctionsViewGroup | NativeFunction,
+) -> Iterator[None]:
+    if isinstance(g, NativeFunctionsGroup):
+        # By default, we associate all errors with structured native functions
+        # with the out variant.  In some cases, it might be better to have
+        # a more specific place to hang things; if so, use
+        # native_function_manager again on the inside
+        f = g.out
+    elif isinstance(g, NativeFunctionsViewGroup):
+        # We associate errors with the view operator
+        f = g.view
+    else:
+        f = g
+    with context(lambda: f"in native_functions.yaml line {f.loc}:\n  {f.func}"):
+        with local.parametrize(
+            use_const_ref_for_mutable_tensors=f.use_const_ref_for_mutable_tensors,
+            use_ilistref_for_tensor_lists=f.part_of_structured_group,
+        ):
+            yield
+
+
+# Given a function that operates on NativeFunction, wrap it into a new function
+# that sets some appropriate context managers for that native function.
+# YOU MUST WRAP FUNCTIONS IN THIS for calls to api modules to be sound
+# (you will get an error if we try to access the local variables without having
+# set them).
+def with_native_function(func: Callable[[F], T]) -> Callable[[F], T]:
+    @functools.wraps(func)
+    def wrapper(f: F) -> T:
+        with native_function_manager(f):
+            return func(f)
+
+    return wrapper
+
+
+def with_native_function_and(func: Callable[[F, F2], T]) -> Callable[[F, F2], T]:
+    @functools.wraps(func)
+    def wrapper(f: F, f2: F2) -> T:
+        # The first native_function is assumed to be the one with the appropriate context.
+        with native_function_manager(f):
+            return func(f, f2)
+
+    return wrapper
+
+
+def method_with_native_function(func: Callable[[S, F], T]) -> Callable[[S, F], T]:
+    @functools.wraps(func)
+    def wrapper(slf: S, f: F) -> T:
+        with native_function_manager(f):
+            return func(slf, f)
+
+    return wrapper
+
+
+def method_with_nested_native_function(
+    func: Callable[[S, F3], T],
+) -> Callable[[S, F3], T]:
+    @functools.wraps(func)
+    def wrapper(slf: S, f: F3) -> T:
+        with native_function_manager(f[0]):
+            return func(slf, f)
+
+    return wrapper
+
+
+# Convenience decorator for functions that explicitly take in a BackendIndex,
+# instead of indirectly taking one in as a closure
+def with_native_function_and_index(
+    func: Callable[[F, BackendIndex], T],
+) -> Callable[[F, BackendIndex], T]:
+    @functools.wraps(func)
+    def wrapper(f: F, backend_index: BackendIndex) -> T:
+        with native_function_manager(f):
+            return func(f, backend_index)
+
+    return wrapper
+
+
+# Convenience decorator for functions that explicitly take in a Dict of BackendIndices
+def with_native_function_and_indices(
+    func: Callable[[F, dict[DispatchKey, BackendIndex]], T],
+) -> Callable[[F, dict[DispatchKey, BackendIndex]], T]:
+    @functools.wraps(func)
+    def wrapper(f: F, backend_indices: dict[DispatchKey, BackendIndex]) -> T:
+        with native_function_manager(f):
+            return func(f, backend_indices)
+
+    return wrapper

datasets-cli.exe

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8ecc3f2798b65a5b9ee125a1bf86f62d05a56e9210e70dddff9a6895fae506a5
+size 108451

deactivate.bat

@@ -0,0 +1,22 @@
+@echo off
+
+if defined _OLD_VIRTUAL_PROMPT (
+    set "PROMPT=%_OLD_VIRTUAL_PROMPT%"
+)
+set _OLD_VIRTUAL_PROMPT=
+
+if defined _OLD_VIRTUAL_PYTHONHOME (
+    set "PYTHONHOME=%_OLD_VIRTUAL_PYTHONHOME%"
+    set _OLD_VIRTUAL_PYTHONHOME=
+)
+
+if defined _OLD_VIRTUAL_PATH (
+    set "PATH=%_OLD_VIRTUAL_PATH%"
+)
+
+set _OLD_VIRTUAL_PATH=
+
+set VIRTUAL_ENV=
+set VIRTUAL_ENV_PROMPT=
+
+:END

distutils-precedence.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2638ce9e2500e572a5e0de7faed6661eb569d1b696fcba07b0dd223da5f5d224
+size 151

entry_points.txt

@@ -0,0 +1,6 @@
+[console_scripts]
+torchfrtrace = tools.flight_recorder.fr_trace:main
+torchrun = torch.distributed.run:main
+
+[torchrun.logs_specs]
+default = torch.distributed.elastic.multiprocessing:DefaultLogsSpecs

f2py.exe

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5463826983b536494b9bf027c1fab304632aefe10c03e89f0145a7543cb69d16
+size 108440

gen_aoti_c_shim.py

@@ -0,0 +1,508 @@
+from __future__ import annotations
+
+import textwrap
+from dataclasses import dataclass
+from typing import TYPE_CHECKING
+
+from torchgen.api.types import DispatcherSignature
+from torchgen.api.types.signatures import CppSignature, CppSignatureGroup
+from torchgen.context import method_with_native_function
+from torchgen.model import (
+    Argument,
+    BackendIndex,
+    BaseTy,
+    BaseType,
+    DispatchKey,
+    FunctionSchema,
+    ListType,
+    NativeFunction,
+    NativeFunctionsGroup,
+    OperatorName,
+    OptionalType,
+    Type,
+)
+from torchgen.utils import mapMaybe
+
+
+if TYPE_CHECKING:
+    from collections.abc import Sequence
+
+
+base_type_to_c_type = {
+    BaseTy.Tensor: "AtenTensorHandle",
+    BaseTy.bool: "int32_t",  # Use int to pass bool
+    BaseTy.int: "int64_t",
+    BaseTy.SymInt: "int64_t",  # Inductor-generated code won't see a SymInt
+    BaseTy.Scalar: "double",  # Use double to pass both integer and floating point
+    BaseTy.float: "double",  # TODO: how about other floating point types?
+    BaseTy.str: "const char*",
+    BaseTy.DeviceIndex: "int32_t",
+    BaseTy.Layout: "int32_t",  # Represent enum as int
+    BaseTy.MemoryFormat: "int32_t",  # Represent enum as int
+    BaseTy.ScalarType: "int32_t",  # Represent enum as int
+    BaseTy.Generator: "AtenGeneratorHandle",
+}
+
+base_type_to_aten_type = {
+    BaseTy.Tensor: "at::Tensor",
+    BaseTy.bool: "bool",
+    BaseTy.int: "int64_t",
+    BaseTy.SymInt: "c10::SymInt",
+    BaseTy.Scalar: "c10::Scalar",
+    BaseTy.float: "double",
+    BaseTy.str: "c10::string_view",
+    BaseTy.DeviceIndex: "c10::DeviceIndex",
+    BaseTy.Layout: "c10::Layout",
+    BaseTy.MemoryFormat: "c10::MemoryFormat",
+    BaseTy.ScalarType: "c10::ScalarType",
+    BaseTy.Generator: "at::Generator",
+}
+
+base_type_to_callsite_expr = {
+    BaseTy.Tensor: "*tensor_handle_to_tensor_pointer",
+    BaseTy.bool: "",
+    BaseTy.int: "",
+    BaseTy.SymInt: "",
+    BaseTy.Scalar: "",
+    BaseTy.float: "",
+    BaseTy.str: "",
+    BaseTy.DeviceIndex: "static_cast<c10::DeviceIndex>",
+    BaseTy.Layout: "static_cast<c10::Layout>",
+    BaseTy.MemoryFormat: "static_cast<c10::MemoryFormat>",
+    BaseTy.ScalarType: "static_cast<c10::ScalarType>",
+    BaseTy.Generator: "*generator_handle_to_generator_pointer",
+}
+
+
+# convert args to C types, names in declarations, and expressions in function bodies
+def convert_arg_type_and_name(  # type: ignore[return]
+    typ: Type,
+    name: str,
+) -> tuple[list[str], list[str], list[str], list[str]]:
+    if isinstance(typ, BaseType):
+        if typ.name in base_type_to_c_type:
+            return (
+                [base_type_to_c_type[typ.name]],
+                [name],
+                [base_type_to_aten_type[typ.name]],
+                [
+                    f"{base_type_to_callsite_expr[typ.name]}({name})"
+                    if base_type_to_callsite_expr[typ.name]
+                    else name
+                ],
+            )
+        elif typ.name == BaseTy.Device:
+            return (
+                ["int32_t", "int32_t"],
+                [name, name + "_index_"],
+                ["c10::Device"],
+                [
+                    f"c10::Device(static_cast<c10::DeviceType>({name}), static_cast<c10::DeviceIndex>({name}_index_))"
+                ],
+            )
+        else:
+            # TODO: BaseTy.Dimname, etc.
+            raise NotImplementedError(f"TODO: add support for arg type {repr(typ)}")
+    elif isinstance(typ, OptionalType):
+        c_types, names, aten_types, callsite_exprs = convert_arg_type_and_name(
+            typ.elem, name
+        )
+        j = 0  # index for names
+        new_aten_types = []
+        new_callsite_exprs = []
+        for aten_type in aten_types:
+            # Use pointer to denote optional type
+            c_types[j] = c_types[j] + "*"
+            if aten_type.startswith("c10::ArrayRef<"):
+                # ArrayRef is passed as pointer + size, but no need to add "*" to the size argument
+                new_aten_types.append(f"::std::optional<{aten_type}>")
+                base_type = aten_type[len("c10::ArrayRef<") : -1]
+                new_callsite_exprs.append(
+                    f"pointer_to_optional_list<{base_type}>({names[j]}, {names[j + 1]})"
+                )
+                j += 2
+            elif aten_type == "c10::Device":
+                # Device is passed as device_type + device_index
+                new_aten_types.append("::std::optional<c10::Device>")
+                new_callsite_exprs.append(
+                    f"pointer_to_optional_device({names[j]}, {names[j + 1]})"
+                )
+                j += 2
+            else:
+                new_aten_types.append(f"::std::optional<{aten_type}>")
+                new_callsite_exprs.append(
+                    f"pointer_to_optional<{aten_type}>({names[j]})"
+                )
+                j += 1
+
+        return (
+            c_types,
+            names,
+            new_aten_types,
+            new_callsite_exprs,
+        )
+    elif isinstance(typ, ListType):
+        # Need to explictly pass the list as pointer + length
+        c_types, names, aten_types, _ = convert_arg_type_and_name(typ.elem, name)
+        assert len(c_types) == 1, "ListType with unsupported element type " + repr(typ)
+
+        # The list content should never be modified
+        c_types[0] = f"const {c_types[0]}*"
+        c_types.append("int64_t")
+        name = names[0]
+        names.append(name + "_len_")
+
+        atype = aten_types[0]
+        callsite_exprs = []
+        if atype == "bool":
+            # no converter from std::vector<bool> to c10::ArrayRef<bool>
+            # construct std::array<bool, N> instead
+            assert typ.size is not None
+            callsite_exprs.append(f"pointer_to_list<{typ.size}>({name})")
+        elif atype == "::std::optional<at::Tensor>":
+            # convert from std::vector<::std::optional<at::Tensor>> to c10::List<::std::optional<at::Tensor>>
+            callsite_exprs.append(
+                f"c10::List<{atype}>(c10::ArrayRef<{atype}>(pointer_to_list<{atype}>({name}, {name}_len_)))"
+            )
+        else:
+            callsite_exprs.append(f"pointer_to_list<{atype}>({name}, {name}_len_)")
+
+        aten_types = [f"c10::ArrayRef<{t}>" for t in aten_types]
+        return (
+            c_types,
+            names,
+            aten_types,
+            callsite_exprs,
+        )
+
+
+def zip_type_and_name(types: list[str], names: list[str]) -> list[str]:
+    return [typ + " " + name for typ, name in zip(types, names)]
+
+
+# Generate argument declarations and callsite expressions
+def gen_arguments(flat_arguments: Sequence[Argument]) -> tuple[list[str], list[str]]:
+    types = []
+    new_names = []
+    callsite_exprs = []
+    for arg in flat_arguments:
+        new_types, names, _, new_callsite_exprs = convert_arg_type_and_name(
+            arg.type, arg.name
+        )
+        types.extend(new_types)
+        new_names.extend(names)
+        callsite_exprs.extend(new_callsite_exprs)
+    return zip_type_and_name(types, new_names), callsite_exprs
+
+
+# Return values are passed out as pointer arguments because all the C shim functions
+# are expected to return AOTITorchError.
+# Generate returns as declarations and callsite expressions
+def gen_returns(schema: FunctionSchema) -> tuple[list[str], list[str]]:
+    types = []
+    names = []
+    for idx, ret in enumerate(schema.returns):
+        names.append(f"ret{idx}")
+        if isinstance(ret.type, BaseType) and ret.type.name in base_type_to_c_type:
+            types.append(base_type_to_c_type[ret.type.name] + "*")
+        else:
+            raise NotImplementedError(
+                f"TODO: add support for return type {repr(ret.type)}"
+            )
+
+    def convert_return(typ: BaseType, val: str) -> str:
+        if typ.name == BaseTy.Tensor:
+            return f"new_tensor_handle(std::move({val}));"
+        elif typ.name == BaseTy.SymInt:
+            return f"{val}.expect_int()"
+        elif typ.name == BaseTy.Scalar:
+            return f"{val}.toDouble()"
+        else:
+            return val
+
+    ret_pointer_can_be_null = False
+    unambiguous_name = schema.name.unambiguous_name()
+    for name in [
+        "_scaled_dot_product_flash_attention",
+        "_scaled_dot_product_efficient_attention",
+        "_scaled_dot_product_cudnn_attention",
+        "convolution_backward",
+    ]:
+        if name in unambiguous_name:
+            ret_pointer_can_be_null = True
+            break
+
+    callsite_exprs: list[str] = []
+    for idx, ret in enumerate(schema.returns):
+        tmp = "tmp_result" if len(names) == 1 else f"std::get<{idx}>(tmp_result)"
+        assert isinstance(ret.type, BaseType)
+        rval = convert_return(ret.type, tmp)
+        if ret_pointer_can_be_null:
+            callsite_exprs.append(f"if ({names[idx]}) {{ *{names[idx]} = {rval}; }}")
+        else:
+            callsite_exprs.append(f"*{names[idx]} = {rval};")
+
+    return zip_type_and_name(types, names), callsite_exprs
+
+
+# gen.py generates header first and then src, so caching the result here to avoid duplicate work
+declaration_definition_cache: dict[tuple[str, str, str], tuple[str, str]] = {}
+
+
+def gen_declaration_and_definition(
+    schema: FunctionSchema, device: str, backend_call: str
+) -> tuple[str, str]:
+    func_name = schema.name.unambiguous_name()
+
+    global declaration_definition_cache
+    if (func_name, device, backend_call) in declaration_definition_cache:
+        return declaration_definition_cache[(func_name, device, backend_call)]
+
+    if schema.is_out_fn():
+        # out_variant has out arguments in the front, and it's ok to ignore return values
+        # because C shim functions only return AOTITorchError
+        args, callsite_exprs = gen_arguments(
+            [*schema.arguments.out, *schema.arguments.flat_non_out]
+        )
+        ret_assignments: list[str] = []
+    else:
+        args, callsite_exprs = gen_arguments(schema.arguments.flat_all)
+        # ignore return values for inplace ops
+        ret_declarations, ret_assignments = (
+            ([], []) if schema.name.name.inplace else gen_returns(schema)
+        )
+        args.extend(ret_declarations)
+
+    declaration = f"AOTITorchError aoti_torch_{device}_{func_name}({', '.join(args)})"
+
+    tmp_result = "auto tmp_result = " if ret_assignments else ""
+    ret_assignments_str = "\n" + "\n".join(ret_assignments) if ret_assignments else ""
+    definition = f"""
+{declaration} {{
+    AOTI_TORCH_CONVERT_EXCEPTION_TO_ERROR_CODE({{
+        {tmp_result}{backend_call}(
+{textwrap.indent(', '.join(callsite_exprs), "            ")}
+        );{textwrap.indent(ret_assignments_str, "        ")}
+    }});
+}}
+"""
+    declaration_definition_cache[(func_name, device, backend_call)] = (
+        declaration,
+        definition,
+    )
+    return declaration, definition
+
+
+def gen_static_dispatch_backend_call_signature(
+    sig: CppSignature | DispatcherSignature,
+    f: NativeFunction,
+) -> CppSignature:
+    sig = DispatcherSignature.from_schema(f.func)
+    cpp_sigs = CppSignatureGroup.from_native_function(
+        f, method=False, fallback_binding=False
+    )
+    if sig.symint and f.func.has_symint():
+        cpp_sig = cpp_sigs.symint_signature
+    else:
+        cpp_sig = cpp_sigs.signature
+    assert cpp_sig is not None
+    return cpp_sig
+
+
+def gen_static_dispatch_backend_call(
+    f: NativeFunction,
+    backend_index: BackendIndex,
+) -> str:
+    sig = DispatcherSignature.from_schema(f.func)
+    cpp_sig = gen_static_dispatch_backend_call_signature(sig, f)
+    return f"at::{backend_index.dispatch_key.lower()}::{cpp_sig.name()}"
+
+
+def get_backend_index_for_aoti(
+    func: NativeFunction,
+    func_group_mapping: dict[OperatorName, NativeFunctionsGroup],
+    dispatch_key: DispatchKey,
+    backend_indices: dict[DispatchKey, BackendIndex],
+    extend_aoti_c_shim: bool,
+) -> BackendIndex | None:
+    backend_index = None
+    if backend_indices[dispatch_key].has_kernel(func) or (
+        func.structured_delegate is not None
+        and func.structured_delegate in func_group_mapping
+        and backend_indices[dispatch_key].has_kernel(
+            func_group_mapping[func.structured_delegate]
+        )
+    ):
+        backend_index = backend_indices[dispatch_key]
+    else:
+        # for the extend out-of-tree kernels, we don't need to
+        # duplicatly create C shim wrappers for other dispatch keys
+        if extend_aoti_c_shim:
+            return backend_index
+
+        elif backend_indices[DispatchKey.CompositeExplicitAutograd].has_kernel(func):
+            # We need to create C shim wrappers for CompositeExplicitAutograd kernels
+            backend_index = backend_indices[DispatchKey.CompositeExplicitAutograd]
+        elif backend_indices[
+            DispatchKey.CompositeExplicitAutogradNonFunctional
+        ].has_kernel(func):
+            # We need to create C shim wrappers for CompositeExplicitAutogradNonFunctional kernels
+            backend_index = backend_indices[
+                DispatchKey.CompositeExplicitAutogradNonFunctional
+            ]
+        elif backend_indices[DispatchKey.CompositeImplicitAutograd].has_kernel(func):
+            backend_index = backend_indices[DispatchKey.CompositeImplicitAutograd]
+
+    return backend_index
+
+
+def get_header_for_aoti(
+    func: NativeFunction,
+    func_group_mapping: dict[OperatorName, NativeFunctionsGroup],
+    dispatch_key: DispatchKey,
+    backend_indices: dict[DispatchKey, BackendIndex],
+    extend_aoti_c_shim: bool,
+) -> str | None:
+    backend_index = get_backend_index_for_aoti(
+        func, func_group_mapping, dispatch_key, backend_indices, extend_aoti_c_shim
+    )
+    return (
+        None
+        if backend_index is None
+        else f"#include <ATen/ops/{func.root_name}_{backend_index.dispatch_key.lower()}_dispatch.h>"
+    )
+
+
+def get_fallback_op_name(func: NativeFunction) -> str:
+    return (
+        f"{func.namespace}.{func.func.name.name}.{func.func.name.overload_name}"
+        if func.func.name.overload_name
+        else f"{func.namespace}.{func.func.name.name}.default"
+    )
+
+
+def gen_c_shim(
+    func: NativeFunction,
+    func_group_mapping: dict[OperatorName, NativeFunctionsGroup],
+    dispatch_key: DispatchKey,
+    backend_indices: dict[DispatchKey, BackendIndex],
+    header: bool,
+    extend_aoti_c_shim: bool,
+) -> str | None:
+    backend_index = get_backend_index_for_aoti(
+        func, func_group_mapping, dispatch_key, backend_indices, extend_aoti_c_shim
+    )
+    if backend_index is None:
+        return None
+
+    schema = func.func
+    device = dispatch_key.lower()
+    backend_call = gen_static_dispatch_backend_call(
+        func,
+        backend_index,
+    )
+
+    try:
+        if header:
+            declaration, _ = gen_declaration_and_definition(
+                schema, device, backend_call
+            )
+            return f"AOTI_TORCH_EXPORT {declaration};"
+        else:
+            _, definition = gen_declaration_and_definition(schema, device, backend_call)
+            return definition
+
+    except NotImplementedError:
+        return None
+
+
+@dataclass(frozen=True)
+class ShimGenerator:
+    func_group_mapping: dict[OperatorName, NativeFunctionsGroup]
+    dispatch_key: DispatchKey
+    backend_indices: dict[DispatchKey, BackendIndex]
+    header: bool  # True to generate .h and False to generate .cpp
+    extend_aoti_c_shim: bool
+
+    @method_with_native_function
+    def __call__(
+        self,
+        func: NativeFunction,
+    ) -> str | None:
+        result = gen_c_shim(
+            func,
+            self.func_group_mapping,
+            self.dispatch_key,
+            self.backend_indices,
+            self.header,
+            self.extend_aoti_c_shim,
+        )
+        return result
+
+
+def gen_aoti_c_shim(
+    native_functions: Sequence[NativeFunction],
+    func_group_mapping: dict[OperatorName, NativeFunctionsGroup],
+    dispatch_key: DispatchKey,
+    backend_indices: dict[DispatchKey, BackendIndex],
+    header: bool,
+    extend_aoti_c_shim: bool,
+    includes: str = "",
+) -> str:
+    body = "\n".join(
+        list(
+            mapMaybe(
+                ShimGenerator(
+                    func_group_mapping,
+                    dispatch_key,
+                    backend_indices,
+                    header,
+                    extend_aoti_c_shim,
+                ),
+                native_functions,
+            )
+        )
+    )
+    device = dispatch_key.lower()
+    warning = """
+// WARNING: THIS FILE IS AUTOGENERATED BY torchgen. DO NOT MODIFY BY HAND.
+// See https://github.com/pytorch/pytorch/blob/7e86a7c0155295539996e0cf422883571126073e/torchgen/gen.py#L2424-L2436 for details"""
+
+    if header:
+        return f"""
+{warning}
+
+#pragma once
+
+#include <torch/csrc/inductor/aoti_torch/c/shim.h>
+
+#ifdef __cplusplus
+extern "C" {{
+#endif
+
+{body}
+
+#ifdef __cplusplus
+}} // extern "C"
+#endif
+"""
+
+    else:
+        return f"""
+{warning}
+
+#include <torch/csrc/inductor/aoti_torch/generated/{"extend/" if extend_aoti_c_shim else ""}c_shim_{device}.h>
+#include <torch/csrc/inductor/aoti_torch/utils.h>
+
+#ifndef AT_PER_OPERATOR_HEADERS
+#include <ATen/{str(dispatch_key)}Functions.h>
+#include <ATen/CompositeExplicitAutogradFunctions.h>
+#include <ATen/CompositeExplicitAutogradNonFunctionalFunctions.h>
+#include <ATen/CompositeImplicitAutogradFunctions.h>
+#else
+{includes}
+#endif
+
+using namespace torch::aot_inductor;
+
+{body}"""

gen_backend_stubs.py

@@ -0,0 +1,615 @@
+from __future__ import annotations
+
+import argparse
+import os
+import re
+from collections import Counter, defaultdict, namedtuple
+from pathlib import Path
+from typing import TYPE_CHECKING
+
+import yaml
+
+import torchgen.api.dispatcher as dispatcher
+import torchgen.dest as dest
+from torchgen.api.types import DispatcherSignature
+from torchgen.code_template import CodeTemplate
+from torchgen.context import native_function_manager
+from torchgen.gen import get_grouped_native_functions, parse_native_yaml
+from torchgen.model import (
+    BackendIndex,
+    BackendMetadata,
+    DispatchKey,
+    NativeFunction,
+    NativeFunctionsGroup,
+    OperatorName,
+)
+from torchgen.selective_build.selector import SelectiveBuilder
+from torchgen.utils import concatMap, context, FileManager, NamespaceHelper, Target
+from torchgen.yaml_utils import YamlLoader
+
+
+if TYPE_CHECKING:
+    from collections.abc import Sequence
+
+
+# Parses the external backend's yaml, and adds a new BackendIndex for the backend's dispatch key.
+# Returns a Tuple of (backend_key, autograd_key, cpp_namespace, updated BackendIndex mapping)
+ParsedExternalYaml = namedtuple(
+    "ParsedExternalYaml",
+    ["backend_key", "autograd_key", "class_name", "cpp_namespace", "backend_indices"],
+)
+
+
+def parse_backend_yaml(
+    backend_yaml_path: str,
+    grouped_native_functions: Sequence[NativeFunction | NativeFunctionsGroup],
+    backend_indices: dict[DispatchKey, BackendIndex],
+) -> ParsedExternalYaml:
+    native_functions_map: dict[OperatorName, NativeFunction] = {
+        f.func.name: f
+        for f in concatMap(
+            lambda f: [f] if isinstance(f, NativeFunction) else list(f.functions()),
+            grouped_native_functions,
+        )
+    }
+
+    with open(backend_yaml_path) as f:
+        yaml_values = yaml.load(f, Loader=YamlLoader)
+    assert isinstance(yaml_values, dict)
+
+    valid_keys = [
+        "backend",
+        "class_name",
+        "cpp_namespace",
+        "extra_headers",
+        "supported",
+        "autograd",
+        "full_codegen",
+        "non_native",
+        "ir_gen",
+        "symint",
+    ]
+
+    backend = yaml_values.pop("backend", None)
+    assert backend is not None, 'You must provide a value for "backend"'
+
+    class_name = yaml_values.pop("class_name", None)
+
+    cpp_namespace = yaml_values.pop("cpp_namespace", None)
+    assert cpp_namespace is not None, 'You must provide a value for "cpp_namespace"'
+
+    # Mostly just defaulting to false to stick with LazyTensor convention.
+    use_out_as_primary = yaml_values.pop("use_out_as_primary", False)
+    assert isinstance(
+        use_out_as_primary, bool
+    ), f"You must provide either True or False for use_out_as_primary. Provided: {use_out_as_primary}"
+
+    use_device_guard = yaml_values.pop("device_guard", False)
+    assert isinstance(
+        use_device_guard, bool
+    ), f"You must provide either True or False for device_guard. Provided: {use_device_guard}"
+
+    supported = yaml_values.pop("supported", [])
+    if supported is None:
+        supported = []  # Allow an empty list of supported ops
+    assert isinstance(
+        supported, list
+    ), f'expected "supported" to be a list, but got: {supported} (of type {type(supported)})'
+
+    symint = yaml_values.pop("symint", [])
+    if symint is None:
+        symint = []  # Allow an empty list of symint ops
+    assert isinstance(
+        symint, list
+    ), f'expected "symint" to be a list, but got: {supported} (of type {type(supported)})'
+    symint_set = set(symint)
+
+    supported_autograd = yaml_values.pop("autograd", [])
+    assert isinstance(
+        supported_autograd, list
+    ), f'expected "autograd" to be a list, but got: {supported_autograd}'
+
+    # full_codegen is ignored by parse_backend_yaml, and re-parsed in gen_lazy_tensor.py
+    full_codegen = yaml_values.pop("full_codegen", [])
+    supported.extend(full_codegen)
+
+    # non_native is ignored by parse_backend_yaml, and re-parsed in gen_lazy_tensor.py
+    yaml_values.pop("non_native", {})
+
+    # ir_gen is ignored by parse_backend_yaml, and re-parsed in gen_lazy_tensor.py
+    yaml_values.pop("ir_gen", {})
+
+    assert (
+        len(yaml_values.keys()) == 0
+    ), f'{backend_yaml_path} contains unexpected keys: {", ".join(yaml_values.keys())}. \
+Only the following keys are supported: {", ".join(valid_keys)}'
+
+    def create_backend_index(
+        backend_ops: list[str],
+        symint_ops: set[str],
+        dispatch_key: DispatchKey,
+        *,
+        use_out_as_primary: bool,
+        use_device_guard: bool,
+    ) -> BackendIndex:
+        metadata: dict[OperatorName, BackendMetadata] = {}
+        for op in backend_ops:
+            op_name = OperatorName.parse(op)
+            assert (
+                op_name in native_functions_map
+            ), f"Found an invalid operator name: {op_name}"
+            # See Note [External Backends Follow Dispatcher API]
+            kernel_name = dispatcher.name(native_functions_map[op_name].func)
+            if op in symint_ops:
+                kernel_name += "_symint"
+            # TODO: allow structured external backends later.
+            m = BackendMetadata(
+                kernel=kernel_name, structured=False, cpp_namespace=cpp_namespace
+            )
+            metadata[op_name] = m
+        return BackendIndex(
+            dispatch_key=dispatch_key,
+            use_out_as_primary=use_out_as_primary,
+            external=True,
+            device_guard=use_device_guard,
+            index=metadata,
+        )
+
+    backend_key: DispatchKey | None = None
+    if len(supported) > 0:
+        with context(
+            lambda: f'The provided value for "backend" must be a valid DispatchKey, but got {backend}.'
+        ):
+            backend_key = DispatchKey.parse(backend)
+
+        backend_idx = create_backend_index(
+            supported,
+            symint_set,
+            backend_key,
+            use_out_as_primary=use_out_as_primary,
+            use_device_guard=use_device_guard,
+        )
+        assert backend_key not in backend_indices
+        backend_indices[backend_key] = backend_idx
+
+    autograd_key: DispatchKey | None = None
+    if len(supported_autograd) > 0:
+        with context(
+            lambda: f'The "autograd" key was specified, which indicates that you would like to override \
+the behavior of autograd for some operators on your backend. However "Autograd{backend}" is not a valid DispatchKey.'
+        ):
+            autograd_key = DispatchKey.parse(f"Autograd{backend}")
+
+        autograd_idx = create_backend_index(
+            supported_autograd,
+            symint_set,
+            autograd_key,
+            use_out_as_primary=use_out_as_primary,
+            use_device_guard=use_device_guard,
+        )
+        assert autograd_key not in backend_indices
+        backend_indices[autograd_key] = autograd_idx
+
+    for g in grouped_native_functions:
+        if isinstance(g, NativeFunction):
+            forward_kernels = (
+                []
+                if backend_key is None
+                else [
+                    m
+                    for m in [backend_indices[backend_key].get_kernel(g)]
+                    if m is not None
+                ]
+            )
+            backward_kernels = (
+                []
+                if autograd_key is None
+                else [
+                    m
+                    for m in [backend_indices[autograd_key].get_kernel(g)]
+                    if m is not None
+                ]
+            )
+        else:
+            forward_kernels = (
+                []
+                if backend_key is None
+                else [
+                    m
+                    for m in [
+                        backend_indices[backend_key].get_kernel(f)
+                        for f in g.functions()
+                    ]
+                    if m is not None
+                ]
+            )
+            backward_kernels = (
+                []
+                if autograd_key is None
+                else [
+                    m
+                    for m in [
+                        backend_indices[autograd_key].get_kernel(f)
+                        for f in g.functions()
+                    ]
+                    if m is not None
+                ]
+            )
+
+        forward_kernels = [f for f in forward_kernels if f is not None]
+        backward_kernels = [f for f in backward_kernels if f is not None]
+        assert (
+            len(forward_kernels) == 0 or len(backward_kernels) == 0
+        ), f'Currently, all variants of an op must either be registered to a backend key, or to a backend\'s \
+autograd key. They cannot be mix and matched. If this is something you need, feel free to create an issue! \
+{forward_kernels[0].kernel} is listed under "supported", but {backward_kernels[0].kernel} is listed under "autograd".'
+
+    return ParsedExternalYaml(
+        backend_key, autograd_key, class_name, cpp_namespace, backend_indices
+    )
+
+
+def error_on_missing_kernels(
+    native_functions: Sequence[NativeFunction],
+    backend_indices: dict[DispatchKey, BackendIndex],
+    backend_key: DispatchKey,
+    autograd_key: DispatchKey | None,
+    class_name: str,
+    kernel_defn_file_path: str,
+    full_codegen: list[OperatorName] | None = None,
+) -> None:
+    try:
+        with open(kernel_defn_file_path) as f:
+            backend_defns = f.read()
+    except OSError as e:
+        raise AssertionError(
+            f"Unable to read from the specified impl_path file: {kernel_defn_file_path}"
+        ) from e
+
+    if full_codegen is None:
+        full_codegen = []
+
+    indices = [backend_indices[backend_key].index] + (
+        [] if autograd_key is None else [backend_indices[autograd_key].index]
+    )
+    # Quick mapping from each OperatorName used by the external backend
+    # to its backend kernel name
+    expected_backend_op_names: dict[OperatorName, str] = dict(
+        list(
+            concatMap(
+                lambda index: [
+                    (op_name, metadata.kernel) for op_name, metadata in index.items()
+                ],
+                indices,
+            )
+        )
+    )
+    expected_backend_native_funcs: list[NativeFunction] = [
+        f
+        for f in native_functions
+        if f.func.name in expected_backend_op_names.keys()
+        and f.func.name not in full_codegen
+    ]
+    expected_backend_kernel_name_counts: dict[str, list[NativeFunction]] = defaultdict(
+        list
+    )
+    for native_f in expected_backend_native_funcs:
+        expected_backend_kernel_name_counts[
+            expected_backend_op_names[native_f.func.name]
+        ].append(native_f)
+
+    # This just looks for lines containing "foo(", and assumes that the kernel foo has been implemented.
+    # It might cause false negatives (we won't catch all cases), but that's ok - if we catch a missing kernel
+    # here, then we get a nicer error message. If we miss it, you get a linker error.
+    kernel_defn_regex = rf"(.*){class_name}::\s*([\w\d]*)\("
+    actual_backend_kernel_name_counts = Counter(
+        # A bit unwieldy (this could probably be moved into regex),
+        # but we don't want to include kernel names that come from function calls,
+        # like "return torch_xla::XLANativeFunctions::empty_strided_symint(...)".
+        # Easy check is to ignore any lines with colons before the class name.
+        [
+            y
+            for (x, y) in re.findall(kernel_defn_regex, backend_defns)
+            if not x.endswith(":")
+        ]
+    )
+
+    missing_kernels_err_msg = ""
+    for expected_name, funcs in expected_backend_kernel_name_counts.items():
+        expected_overload_count = len(funcs)
+        actual_overload_count = actual_backend_kernel_name_counts[expected_name]
+        if expected_overload_count != actual_overload_count:
+
+            def create_decl(f: NativeFunction) -> str:
+                with native_function_manager(f):
+                    return DispatcherSignature.from_schema(f.func).decl()
+
+            expected_schemas_str = "\n".join([create_decl(f) for f in funcs])
+            missing_kernels_err_msg += f"""
+{class_name} is missing a kernel definition for {expected_name}. We found {actual_overload_count} kernel(s) with that name,
+but expected {expected_overload_count} kernel(s). The expected function schemas for the missing operator are:
+{expected_schemas_str}
+
+"""
+    assert missing_kernels_err_msg == "", missing_kernels_err_msg
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(description="Generate backend stub files")
+    parser.add_argument(
+        "-s",
+        "--source-yaml",
+        "--source_yaml",
+        help="path to source yaml file containing operator external definitions",
+    )
+    parser.add_argument("-o", "--output-dir", "--output_dir", help="output directory")
+    parser.add_argument(
+        "--dry-run", "--dry_run", type=bool, default=False, help="output directory"
+    )
+    parser.add_argument(
+        "--impl-path",
+        "--impl_path",
+        type=str,
+        default=None,
+        help="path to the source C++ file containing kernel definitions",
+    )
+    options = parser.parse_args()
+
+    run(options.source_yaml, options.output_dir, options.dry_run, options.impl_path)
+
+
+def gen_dispatchkey_nativefunc_headers(
+    fm: FileManager,
+    class_name: str,
+    cpp_namespace: str,
+    backend_indices: dict[DispatchKey, BackendIndex],
+    grouped_native_functions: Sequence[NativeFunction | NativeFunctionsGroup],
+    backend_dispatch_key: DispatchKey,
+    autograd_dispatch_key: DispatchKey | None,
+    backend_name: str = "",
+) -> None:
+    assert class_name is not None
+    generated_comment = (
+        "Autogenerated file by gen_backend_stubs.py. Do not edit directly!"
+    )
+
+    # Convert to a set first to remove duplicate kernel names.
+    # Backends are allowed to repeat kernel names; only generate the declaration once!
+    # Sort for deterministic output.
+    backend_declarations = sorted(
+        set(
+            concatMap(
+                lambda f: dest.compute_native_function_declaration(
+                    f, backend_indices[backend_dispatch_key]
+                ),
+                grouped_native_functions,
+            )
+        )
+    )
+    autograd_declarations = sorted(
+        set(
+            concatMap(
+                lambda f: []
+                if autograd_dispatch_key is None
+                else dest.compute_native_function_declaration(
+                    f, backend_indices[autograd_dispatch_key]
+                ),
+                grouped_native_functions,
+            )
+        )
+    )
+
+    ns_helper = NamespaceHelper(cpp_namespace)
+    fm.write_with_template(
+        f"{backend_dispatch_key}NativeFunctions.h",
+        "DispatchKeyNativeFunctions.h",
+        lambda: {
+            "generated_comment": generated_comment,
+            "namespace_prologue": ns_helper.prologue,
+            "class_name": class_name,
+            "namespace_epilogue": ns_helper.epilogue,
+            "dispatch_declarations": backend_declarations + autograd_declarations,
+            "BackendName": backend_name,
+            "DispatchKey": backend_dispatch_key,
+        },
+    )
+
+
+def gen_dispatcher_registrations(
+    fm: FileManager,
+    output_dir: str,
+    class_name: str,
+    backend_indices: dict[DispatchKey, BackendIndex],
+    grouped_native_functions: Sequence[NativeFunction | NativeFunctionsGroup],
+    backend_dispatch_key: DispatchKey,
+    dispatch_key: DispatchKey,
+    selector: SelectiveBuilder,
+    # build_in_tree is true for lazy TS backend and affects include paths, not used for external backends
+    build_in_tree: bool = False,
+    per_operator_headers: bool = False,
+    backend_name: str = "",
+    eager_registration: bool = True,
+) -> None:
+    headers = [
+        f"{output_dir}/{backend_dispatch_key}NativeFunctions.h",
+    ]
+    if build_in_tree:
+        external_backend_headers_str = "\n".join(f"#include <{h}>" for h in headers)
+    else:
+        external_backend_headers_str = "\n".join(f'#include "{h}"' for h in headers)
+
+    assert class_name is not None
+    backend_index = backend_indices[dispatch_key]
+
+    dispatch_registrations_body = list(
+        concatMap(
+            dest.RegisterDispatchKey(
+                backend_index,
+                Target.REGISTRATION,
+                selector,
+                rocm=False,
+                symint=True,
+                class_method_name=f"{class_name}",
+                skip_dispatcher_op_registration=False,
+            ),
+            grouped_native_functions,
+        )
+    )
+    newline = "\n"
+    ns_helper = NamespaceHelper(namespace_str="at")
+    deferred_dispatch_registrations = ""
+    static_init_dispatch_registrations = ""
+    if eager_registration:
+        static_template = CodeTemplate(
+            """\
+TORCH_LIBRARY_IMPL(aten, $dispatch_key, m) {
+    $dispatch_registrations_body
+}"""
+        )
+        static_init_dispatch_registrations = static_template.substitute(
+            dispatch_key=dispatch_key,
+            dispatch_registrations_body=dispatch_registrations_body,
+        )
+    else:
+        deferred_template = CodeTemplate(
+            """\
+TORCH_API void Register${backend_name}${dispatch_key}NativeFunctions();
+TORCH_API void Register${backend_name}${dispatch_key}NativeFunctions() {
+    static auto m = MAKE_TORCH_LIBRARY_IMPL(aten, $dispatch_key);
+    $dispatch_registrations_body
+}"""
+        )
+        deferred_dispatch_registrations = deferred_template.substitute(
+            backend_name=backend_name,
+            dispatch_key=dispatch_key,
+            dispatch_registrations_body=dispatch_registrations_body,
+        )
+
+    fm.write_with_template(
+        f"Register{dispatch_key}.cpp",
+        "RegisterDispatchKey.cpp",
+        lambda: {
+            "extra_cuda_headers": "",
+            "external_backend_headers": external_backend_headers_str,
+            "ops_headers": "#include <ATen/Functions.h>"
+            if not per_operator_headers
+            else "",
+            "DispatchKey": dispatch_key,
+            "dispatch_namespace": dispatch_key.lower(),
+            "dispatch_headers": dest.gen_registration_headers(
+                backend_index, per_operator_headers=per_operator_headers, rocm=False
+            ),
+            "dispatch_definitions": fm.substitute_with_template(
+                "RegisterDispatchDefinitions.ini",
+                lambda: {
+                    "ns_prologue": ns_helper.prologue,
+                    "ns_epilogue": ns_helper.epilogue,
+                    "static_init_dispatch_registrations": static_init_dispatch_registrations,
+                    "deferred_dispatch_registrations": deferred_dispatch_registrations,
+                    "dispatch_helpers": dest.gen_registration_helpers(backend_index),
+                    "dispatch_namespace": dispatch_key.lower(),
+                    "dispatch_namespaced_definitions": "",
+                    "dispatch_anonymous_definitions": list(
+                        concatMap(
+                            dest.RegisterDispatchKey(
+                                backend_index,
+                                Target.ANONYMOUS_DEFINITION,
+                                selector,
+                                rocm=False,
+                                symint=True,
+                                class_method_name=f"{class_name}",
+                                skip_dispatcher_op_registration=False,
+                            ),
+                            grouped_native_functions,
+                        )
+                    ),
+                },
+            ).split(newline),
+        },
+    )
+
+
+def run(
+    source_yaml: str, output_dir: str, dry_run: bool, impl_path: str | None = None
+) -> None:
+    # Assumes that this file lives at PYTORCH_ROOT/torchgen/gen_backend_stubs.py
+    pytorch_root = Path(__file__).parent.parent.absolute()
+    template_dir = os.path.join(pytorch_root, "aten/src/ATen/templates")
+
+    def make_file_manager(install_dir: str) -> FileManager:
+        return FileManager(
+            install_dir=install_dir, template_dir=template_dir, dry_run=dry_run
+        )
+
+    fm = make_file_manager(output_dir)
+
+    native_yaml_path = os.path.join(
+        pytorch_root, "aten/src/ATen/native/native_functions.yaml"
+    )
+    tags_yaml_path = os.path.join(pytorch_root, "aten/src/ATen/native/tags.yaml")
+    parsed_yaml = parse_native_yaml(native_yaml_path, tags_yaml_path)
+    native_functions, backend_indices = (
+        parsed_yaml.native_functions,
+        parsed_yaml.backend_indices,
+    )
+    grouped_native_functions = get_grouped_native_functions(native_functions)
+    parsed_backend_yaml = parse_backend_yaml(
+        source_yaml, grouped_native_functions, backend_indices
+    )
+    backend_key = parsed_backend_yaml.backend_key
+    autograd_key = parsed_backend_yaml.autograd_key
+    cpp_namespace = parsed_backend_yaml.cpp_namespace
+    class_name = parsed_backend_yaml.class_name
+    backend_indices = parsed_backend_yaml.backend_indices
+
+    selector = SelectiveBuilder.get_nop_selector()
+
+    if backend_key is None:
+        # This could be useful if a backend wants to quickly set up a noop yaml file but doesn't have any kernels ready yet.
+        return
+
+    if class_name is None:
+        # class_name is an optional argument to backend yaml file.
+        # if specified it allows an external backend to override
+        # the name of the class that all generated kernel definitions live under.
+        # if not specified, its value is given as native_function_class_name.
+        class_name = backend_indices[backend_key].native_function_class_name()
+    assert class_name is not None
+
+    if impl_path is not None:
+        error_on_missing_kernels(
+            native_functions,
+            backend_indices,
+            backend_key,
+            autograd_key,
+            class_name,
+            impl_path,
+        )
+
+    gen_dispatchkey_nativefunc_headers(
+        fm,
+        class_name,
+        cpp_namespace,
+        backend_indices,
+        grouped_native_functions,
+        backend_key,
+        autograd_key,
+    )
+
+    for dispatch_key in (
+        [backend_key] if autograd_key is None else [backend_key, autograd_key]
+    ):
+        gen_dispatcher_registrations(
+            fm,
+            output_dir,
+            class_name,
+            backend_indices,
+            grouped_native_functions,
+            backend_key,
+            dispatch_key,
+            selector,
+        )
+
+
+if __name__ == "__main__":
+    main()

gen_executorch.py

@@ -0,0 +1,1000 @@
+from __future__ import annotations
+
+import argparse
+import os
+from collections import defaultdict
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Any, Callable, TextIO, TYPE_CHECKING
+
+import yaml
+
+# Parse native_functions.yaml into a sequence of NativeFunctions and Backend Indices.
+from torchgen import dest
+from torchgen.api import cpp as aten_cpp
+from torchgen.api.types import CppSignature, CppSignatureGroup, CType, NamedCType
+from torchgen.context import (
+    method_with_native_function,
+    method_with_nested_native_function,
+    with_native_function_and_index,
+)
+from torchgen.executorch.api import et_cpp
+from torchgen.executorch.api.custom_ops import (
+    ComputeNativeFunctionStub,
+    gen_custom_ops_registration,
+)
+from torchgen.executorch.api.types import contextArg, ExecutorchCppSignature
+from torchgen.executorch.api.unboxing import Unboxing
+from torchgen.executorch.model import ETKernelIndex, ETKernelKey, ETParsedYaml
+from torchgen.executorch.parse import ET_FIELDS, parse_et_yaml, parse_et_yaml_struct
+from torchgen.gen import (
+    get_custom_build_selector,
+    get_native_function_declarations,
+    get_native_function_declarations_from_ns_grouped_kernels,
+    get_native_function_schema_registrations,
+    LineLoader,
+    parse_native_yaml,
+)
+from torchgen.model import (
+    BackendIndex,
+    BackendMetadata,
+    DEFAULT_KERNEL_NAMESPACE,
+    DispatchKey,
+    FunctionSchema,
+    Location,
+    NativeFunction,
+    NativeFunctionsGroup,
+    OperatorName,
+    Variant,
+)
+from torchgen.utils import (
+    context,
+    FileManager,
+    make_file_manager,
+    mapMaybe,
+    NamespaceHelper,
+)
+
+
+if TYPE_CHECKING:
+    from collections.abc import Sequence
+
+    from torchgen.selective_build.selector import SelectiveBuilder
+
+
+def _sig_decl_wrapper(sig: CppSignature | ExecutorchCppSignature) -> str:
+    """
+    A wrapper function to basically get `sig.decl(include_context=True)`.
+    For ATen kernel, the codegen has no idea about ET contextArg, so we
+    use this wrapper to add it.
+    """
+    if isinstance(sig, ExecutorchCppSignature):
+        return sig.decl()
+
+    returns_type = aten_cpp.returns_type(sig.func.returns).cpp_type()
+    cpp_args = [a.decl() for a in sig.arguments()]
+    cpp_args_str = ", ".join([contextArg.decl()] + cpp_args)
+    sig_decl = f"{returns_type} {sig.name()}({cpp_args_str})"
+    return sig_decl
+
+
+def static_dispatch(
+    sig: CppSignature | ExecutorchCppSignature,
+    f: NativeFunction,
+    backend_indices: list[BackendIndex],
+) -> str:
+    """
+    For a given `NativeFunction`, find out the corresponding native function and dispatch to it. If zero or more than one
+    native function exists, error out. A simplified version of register_dispatch_key.py
+    Arguments:
+        sig: A CppSignature for this native function we want to use.
+        f: NativeFunction to generate static dispatch.
+        backend_indices: All available backends.
+    Return:
+        C++ code to call backend-specific functions, e.g., "return at::native::add(self, other, scale);"
+    """
+    if len(backend_indices) == 0 or f.manual_kernel_registration:
+        return ""
+
+    backends = [b for b in backend_indices if b.has_kernel(f)]
+    static_block = None
+    if len(backends) == 1:
+        backend_metadata = backends[0].get_kernel(f)
+        if backend_metadata:
+            args = ", ".join(a.name for a in sig.arguments())
+            # Here we are assuming there's no difference between CppSignature and NativeSignature for Executorch.
+            static_block = f"return ::{backend_metadata.cpp_namespace}::{backend_metadata.kernel}({args});"
+    else:
+        static_block = f"""
+ET_ASSERT_UNREACHABLE_MSG("The number of native function(s) binding to {f.func.name} is {len(backends)}.");
+    """
+    return f"""
+// {f.namespace}::{f.func}
+TORCH_API inline {_sig_decl_wrapper(sig)} {{
+    {static_block}
+}}
+"""
+
+
+# Generates Functions.h, which provides the functional public C++ API,
+# and the scaffolding to call into the dispatcher from these functions.
+@dataclass(frozen=True)
+class ComputeFunction:
+    static_dispatch_backend_indices: list[BackendIndex]
+
+    selector: SelectiveBuilder
+
+    use_aten_lib: bool
+
+    is_custom_op: Callable[[NativeFunction], bool]
+
+    @method_with_native_function
+    def __call__(self, f: NativeFunction) -> str | None:
+        is_method_variant = False
+        if not self.selector.is_root_operator(f"{f.namespace}::{f.func.name}"):
+            return None
+
+        if Variant.function not in f.variants and Variant.method in f.variants:
+            is_method_variant = True
+
+        # only valid remaining case is only function is in f.variants
+        elif not (Variant.function in f.variants and Variant.method not in f.variants):
+            raise Exception(  # noqa: TRY002
+                f"Can't handle native function {f.func} with the following variant specification {f.variants}."
+            )
+
+        sig: CppSignature | ExecutorchCppSignature = (
+            CppSignatureGroup.from_native_function(
+                f, method=False, fallback_binding=f.manual_cpp_binding
+            ).most_faithful_signature()
+            if self.use_aten_lib
+            else ExecutorchCppSignature.from_native_function(f)
+        )
+        if self.use_aten_lib and not self.is_custom_op(f):
+            comma = ", "
+
+            if is_method_variant:
+                return f"""
+// {f.namespace}::{f.func}
+TORCH_API inline {_sig_decl_wrapper(sig)} {{
+    return {sig.arguments()[0].name}.{sig.name()}({comma.join(e.name for e in sig.arguments()[1:])});
+}}
+"""
+            else:
+                return f"""
+// {f.namespace}::{f.func}
+TORCH_API inline {_sig_decl_wrapper(sig)} {{
+    return at::{sig.name()}({comma.join(e.name for e in sig.arguments())});
+}}
+"""
+
+        else:
+            return static_dispatch(
+                sig,
+                f,
+                backend_indices=self.static_dispatch_backend_indices,
+            )
+
+
+# Generates RegisterCodegenUnboxedKernels.cpp.
+@dataclass(frozen=True)
+class ComputeCodegenUnboxedKernels:
+    selector: SelectiveBuilder
+
+    use_aten_lib: bool
+
+    @method_with_nested_native_function
+    def __call__(
+        self,
+        unbox_kernel_entry: tuple[NativeFunction, tuple[ETKernelKey, BackendMetadata]],
+    ) -> str:
+        f: NativeFunction = unbox_kernel_entry[0]
+        kernel_key: ETKernelKey | list[ETKernelKey] = unbox_kernel_entry[1][0]
+        kernel_meta: BackendMetadata = unbox_kernel_entry[1][1]
+
+        op_name = f"{f.namespace}::{f.func.name}"
+        if not self.selector.is_root_operator(op_name):
+            return ""
+
+        if not isinstance(kernel_key, list):
+            kernel_key = [kernel_key]
+        used_kernel_keys = self.selector.et_get_selected_kernels(
+            op_name, [k.to_native_string() for k in kernel_key]
+        )
+        if not used_kernel_keys:
+            return ""
+        sig: CppSignature | ExecutorchCppSignature
+        argument_type_gen: Callable[..., NamedCType]
+        return_type_gen: Callable[..., CType]
+        if self.use_aten_lib:
+            sig = CppSignatureGroup.from_native_function(
+                f, method=False, fallback_binding=f.manual_cpp_binding
+            ).most_faithful_signature()
+            argument_type_gen = aten_cpp.argumenttype_type
+            return_type_gen = aten_cpp.returns_type
+            arguments = sig.arguments()
+            kernel_call = f"torch::executor::{f.namespace}::{sig.name()}"
+        else:
+            sig = ExecutorchCppSignature.from_native_function(f)
+            argument_type_gen = et_cpp.argumenttype_type
+            return_type_gen = et_cpp.returns_type
+            arguments = sig.arguments(include_context=False)
+            kernel_call = f"{kernel_meta.cpp_namespace}::{kernel_meta.kernel}"
+        # parse arguments into C++ code
+        binding_list, code_list = Unboxing(
+            argument_type_gen=argument_type_gen
+        ).convert_arguments(arguments)
+
+        # for each C++ argument, generate the conversion code
+        code_connector = "\n\t"
+        arg_connector = ", "
+
+        args_str = f"{arg_connector.join(e.name for e in binding_list)}"
+        event_tracer_output_logging = ""
+        output_ids = []
+
+        if len(f.func.returns) == 0:
+            if len(f.func.arguments.out) == 0:
+                raise Exception(  # noqa: TRY002
+                    f"Can't handle native function {f.func} with no returns and no out yet."
+                )
+            out = f.func.arguments.out[0]
+            return_assignment = f"""stack[{len(binding_list)}] = &{out.name};"""
+            ret_prefix = ""
+            output_ids = [len(binding_list)]
+        else:
+            if len(f.func.arguments.out) == 0:
+                return_assignment = (
+                    f"""*stack[{len(binding_list)}] = EValue(result_);"""
+                )
+                ret_prefix = return_type_gen(f.func.returns).cpp_type() + " result_ = "
+                output_ids = [len(binding_list)]
+            else:
+                return_assignment = ""
+                ret_prefix = ""
+                output_ids = [
+                    len(binding_list) - (i + 1)
+                    for i in reversed(range(len(f.func.arguments.out)))
+                ]
+
+        for output_id in output_ids:
+            event_tracer_output_logging += (
+                f"internal::event_tracer_log_evalue("
+                f"context.internal_event_tracer(), "
+                f"*stack[{output_id}]);\n"
+            )
+
+        newline = "\n    "
+        return "\n".join(
+            [
+                f"""
+Kernel(
+    "{f.namespace}::{f.func.name}",{newline + '"' + (k + '",') if k != 'default' else ''}
+    []({contextArg.defn()}, EValue** stack) {{
+        {code_connector.join(code_list)}
+
+        internal::EventTracerProfileOpScope event_tracer_op_scope(context.internal_event_tracer(), "native_call_{f.func.name}");
+        EXECUTORCH_SCOPE_PROF("native_call_{f.func.name}");
+        {ret_prefix}{kernel_call}(context, {args_str});
+        {event_tracer_output_logging}
+        {return_assignment}
+    }}
+),
+"""
+                for k in used_kernel_keys
+            ]
+        )
+
+
+def gen_unboxing(
+    *,
+    native_functions: Sequence[NativeFunction],
+    cpu_fm: FileManager,
+    selector: SelectiveBuilder,
+    use_aten_lib: bool,
+    kernel_index: ETKernelIndex,
+    manual_registration: bool,
+) -> None:
+    # Iterable type for write_sharded is a Tuple of (native_function, (kernel_key, metadata))
+    def key_func(
+        item: tuple[NativeFunction, tuple[ETKernelKey, BackendMetadata]],
+    ) -> str:
+        return item[0].root_name + ":" + item[1][0].to_native_string()
+
+    items: list[tuple[NativeFunction, tuple[ETKernelKey, BackendMetadata]]] = [
+        (native_function, (kernel_key, metadata))
+        for native_function in native_functions
+        for kernel_key, metadata in kernel_index.get_kernels(native_function).items()
+    ]
+
+    header = ["Functions.h" if use_aten_lib else "NativeFunctions.h"]
+    filename = (
+        "RegisterKernels.cpp"
+        if manual_registration
+        else "RegisterCodegenUnboxedKernels.cpp"
+    )
+    cpu_fm.write_sharded(
+        filename,
+        items,
+        key_fn=key_func,
+        env_callable=lambda unbox_kernel_entry: {
+            "unboxed_kernels": [
+                ComputeCodegenUnboxedKernels(selector, use_aten_lib)(unbox_kernel_entry)
+            ],
+            "fn_header": header
+            if unbox_kernel_entry == items[0]
+            else [],  # Only write header once
+        },
+        num_shards=1,
+        sharded_keys={"unboxed_kernels", "fn_header"},
+    )
+
+
+@with_native_function_and_index  # type: ignore[arg-type]
+def compute_native_function_declaration(
+    g: NativeFunctionsGroup | NativeFunction, kernel_index: ETKernelIndex
+) -> list[str]:
+    assert isinstance(g, NativeFunction)
+    sig = ExecutorchCppSignature.from_native_function(f=g)
+    metadata_list = kernel_index.get_kernels(g).values()
+    if metadata_list is None:
+        return []
+
+    # for kernels in lean mode, we declare two versions, one with context and one without.
+    # In the end we will cleanup the unused one.
+    def gen_decl(metadata: BackendMetadata, include_context: bool) -> str:
+        return f"{sig.decl(name=metadata.kernel, include_context=include_context)};"
+
+    return [
+        gen_decl(metadata, include_context)
+        for include_context in [False, True]
+        for metadata in metadata_list
+    ]
+
+
+def gen_functions_declarations(
+    *,
+    native_functions: Sequence[NativeFunction],
+    kernel_index: ETKernelIndex,
+    selector: SelectiveBuilder,
+    use_aten_lib: bool,
+    custom_ops_native_functions: Sequence[NativeFunction] | None = None,
+) -> str:
+    """
+    Generates namespace separated C++ function API inline declaration/definitions.
+    Native functions are grouped by namespaces and the generated code is wrapped inside
+    namespace blocks.
+
+    E.g., for `custom_1::foo.out` in yaml file we will generate a C++ API as a symbol
+    in `torch::executor::custom_1::foo_out`. This way we avoid symbol conflict when
+    the other `custom_2::foo.out` is available.
+    """
+
+    # convert kernel index to BackendIndex. This is because we can't handle ETKernelIndex yet.
+    # TODO larryliu: evaluate if this code is still needed. If yes let it handle ETKernelIndex.
+
+    backend_index = kernel_index._to_backend_index()
+
+    ns_grouped_functions = defaultdict(list)
+    for native_function in native_functions:
+        ns_grouped_functions[native_function.namespace].append(native_function)
+    functions_declarations = ""
+    newline = "\n"
+    for namespace in ns_grouped_functions:
+        ns_helper = NamespaceHelper(
+            namespace_str=namespace,
+            entity_name="",
+            max_level=3,
+        )
+        declarations = list(
+            mapMaybe(
+                ComputeFunction(
+                    static_dispatch_backend_indices=[backend_index],
+                    selector=selector,
+                    use_aten_lib=use_aten_lib,
+                    is_custom_op=lambda f: custom_ops_native_functions is not None
+                    and f in custom_ops_native_functions,
+                ),
+                ns_grouped_functions[namespace],
+            )
+        )
+        functions_declarations += f"""
+{ns_helper.prologue}
+{newline.join(declarations)}
+{ns_helper.epilogue}
+        """
+    return functions_declarations
+
+
+def get_ns_grouped_kernels(
+    *,
+    native_functions: Sequence[NativeFunction],
+    kernel_index: ETKernelIndex,
+    native_function_decl_gen: Callable[
+        [
+            NativeFunctionsGroup | NativeFunction,
+            ETKernelIndex,
+        ],
+        list[str],
+    ],
+) -> dict[str, list[str]]:
+    ns_grouped_kernels: dict[str, list[str]] = defaultdict(list)
+    for f in native_functions:
+        native_function_namespaces = set()
+        op_kernels = kernel_index.get_kernels(f)
+        for backend_metadata in op_kernels.values():
+            if backend_metadata:
+                namespace = backend_metadata.cpp_namespace
+                native_function_namespaces.add(namespace)
+            else:
+                namespace = DEFAULT_KERNEL_NAMESPACE
+            assert (
+                len(native_function_namespaces) <= 1
+            ), f"Codegen only supports one namespace per operator, got {native_function_namespaces}"
+            ns_grouped_kernels[namespace].extend(
+                native_function_decl_gen(f, kernel_index)
+            )
+    return ns_grouped_kernels
+
+
+def gen_headers(
+    *,
+    native_functions: Sequence[NativeFunction],
+    gen_custom_ops_header: bool,
+    custom_ops_native_functions: Sequence[NativeFunction],
+    selector: SelectiveBuilder,
+    kernel_index: ETKernelIndex,
+    cpu_fm: FileManager,
+    use_aten_lib: bool,
+) -> None:
+    """Generate headers.
+
+    Args:
+        native_functions (Sequence[NativeFunction]): a collection of NativeFunction for ATen ops.
+        gen_custom_ops_header (bool): whether we should generate CustomOpsNativeFunctions.h
+        custom_ops_native_functions (Sequence[NativeFunction]): a collection of NativeFunction for custom ops.
+        kernel_index (ETKernelIndex): kernel collection
+        cpu_fm (FileManager): file manager manages output stream
+        use_aten_lib (bool): whether we are generating for PyTorch types or Executorch types.
+    """
+    aten_headers = ["#include <ATen/Functions.h>"]
+    backend_indices = {DispatchKey.CPU: kernel_index._to_backend_index()}
+    if gen_custom_ops_header:
+        cpu_fm.write_with_template(
+            "CustomOpsNativeFunctions.h",
+            "NativeFunctions.h",
+            lambda: {
+                "nativeFunctions_declarations": get_native_function_declarations(
+                    grouped_native_functions=custom_ops_native_functions,
+                    backend_indices=backend_indices,
+                    native_function_decl_gen=dest.compute_native_function_declaration,
+                ),
+                "headers": [
+                    "#include <ATen/ATen.h>",
+                    "#include <torch/torch.h>",
+                ],
+            },
+        )
+        aten_headers.append('#include "CustomOpsNativeFunctions.h"')
+    cpu_fm.write(
+        "Functions.h",
+        lambda: {
+            "static_dispatch_extra_headers": aten_headers
+            if use_aten_lib
+            else ['#include "NativeFunctions.h"'],
+            "Functions_declarations": gen_functions_declarations(
+                native_functions=native_functions,
+                kernel_index=kernel_index,
+                selector=selector,
+                use_aten_lib=use_aten_lib,
+                custom_ops_native_functions=custom_ops_native_functions,
+            ),
+        },
+    )
+    cpu_fm.write(
+        "RegisterKernels.h",
+        lambda: {
+            "generated_comment": "@" + "generated by torchgen/gen_executorch.py",
+        },
+    )
+    headers = {
+        "headers": [
+            "#include <executorch/runtime/core/exec_aten/exec_aten.h> // at::Tensor etc.",
+            "#include <executorch/runtime/kernel/kernel_runtime_context.h>",
+        ],
+    }
+    if use_aten_lib:
+        headers["headers"].append("#include <executorch/codegen/macros.h> // TORCH_API")
+        cpu_fm.write(
+            "NativeFunctions.h",
+            lambda: dict(
+                {
+                    "nativeFunctions_declarations": get_native_function_declarations(
+                        grouped_native_functions=native_functions,
+                        backend_indices=backend_indices,
+                        native_function_decl_gen=dest.compute_native_function_declaration,
+                    ),
+                },
+                **headers,
+            ),
+        )
+    else:
+        ns_grouped_kernels = get_ns_grouped_kernels(
+            native_functions=native_functions,
+            kernel_index=kernel_index,
+            native_function_decl_gen=compute_native_function_declaration,  # type: ignore[arg-type]
+        )
+        cpu_fm.write(
+            "NativeFunctions.h",
+            lambda: dict(
+                {
+                    "nativeFunctions_declarations": get_native_function_declarations_from_ns_grouped_kernels(
+                        ns_grouped_kernels=ns_grouped_kernels,
+                    ),
+                },
+                **headers,
+            ),
+        )
+
+
+def gen_custom_ops(
+    *,
+    native_functions: Sequence[NativeFunction],
+    selector: SelectiveBuilder,
+    kernel_index: ETKernelIndex,
+    cpu_fm: FileManager,
+    rocm: bool,
+) -> None:
+    dispatch_key = DispatchKey.CPU
+    (
+        anonymous_definition,
+        static_init_dispatch_registrations,
+    ) = gen_custom_ops_registration(
+        native_functions=native_functions,
+        selector=selector,
+        kernel_index=kernel_index,
+        rocm=rocm,
+    )
+    cpu_fm.write_with_template(
+        f"Register{dispatch_key}CustomOps.cpp",
+        "RegisterDispatchKeyCustomOps.cpp",
+        lambda: {
+            "ops_headers": '#include "CustomOpsNativeFunctions.h"',
+            "DispatchKey": dispatch_key,
+            "dispatch_namespace": dispatch_key.lower(),
+            "dispatch_namespaced_definitions": "",
+            "dispatch_anonymous_definitions": anonymous_definition,
+            "static_init_dispatch_registrations": static_init_dispatch_registrations,
+        },
+    )
+    cpu_fm.write_with_template(
+        f"Register{dispatch_key}Stub.cpp",
+        "RegisterDispatchKeyCustomOps.cpp",
+        lambda: {
+            "ops_headers": "",
+            "DispatchKey": dispatch_key,
+            "dispatch_namespace": dispatch_key.lower(),
+            "dispatch_namespaced_definitions": "",
+            "dispatch_anonymous_definitions": list(
+                mapMaybe(ComputeNativeFunctionStub(), native_functions)
+            ),
+            "static_init_dispatch_registrations": static_init_dispatch_registrations,
+        },
+    )
+
+    (
+        aten_schema_registrations,
+        schema_registrations,
+    ) = get_native_function_schema_registrations(
+        native_functions=native_functions,
+        schema_selector=selector,
+    )
+    cpu_fm.write(
+        "RegisterSchema.cpp",
+        lambda: {
+            "schema_registrations": schema_registrations,
+            "aten_schema_registrations": aten_schema_registrations,
+        },
+    )
+
+
+def translate_native_yaml(
+    tags_yaml_path: str,
+    aten_yaml_path: str,
+    native_yaml_path: str | None,
+    use_aten_lib: bool,
+    out_file: TextIO,
+) -> None:
+    """Translates Executorch DSL dialect to use the same syntax as
+    native_functions.yaml. The major difference is that Executorch DSL dialect
+    supports "op" key, where it refers to the operator name in native_functions.yaml.
+
+    For example, a functions.yaml may have the following entry:
+
+    - op: add.out
+      ...
+
+    It needs to be translated to the following:
+
+    - func: add.out(Tensor self, Tensor other, *, Scalar alpha=1, Tensor(a!) out) -> Tensor(a!)
+      ...
+
+    We go in aten_yaml_path and find the operator schema for "add.out" and add it
+    to the original functions.yaml. We also add required field "variants", where for
+    Executorch it will always be "function".
+
+    For ATen mode we don't have to do the translation because native_yaml_path is
+    the same as native_functions.yaml.
+
+    Args:
+        tags_yaml_path: Path to a tags.yaml file to satisfy codegen parsing.
+            It is not optional.
+        aten_yaml_path: Path to ATen operator yaml file native_functions.yaml.
+        native_yaml_path: Path to a functions.yaml file to parse.
+            If the path does not exist in the filesystem, it is treated as an
+            empty file. If `custom_ops_yaml_path` exists, the contents of that
+            file are appended to the yaml input to be parsed.
+        use_aten_lib: We use this flag to determine if we want to generate native
+            functions. In ATen mode we should generate out= variants.
+        out_file: The IO object that we are writing into.
+    Returns:
+        None
+    """
+    if use_aten_lib:
+        with open(aten_yaml_path) as aten_yaml:
+            out_file.writelines(aten_yaml.readlines())
+        return
+
+    native_functions, persisted_fields = parse_et_yaml(
+        aten_yaml_path,
+        tags_yaml_path,
+        None,
+        skip_native_fns_gen=False,
+    )
+
+    func_to_scoped_name: dict[FunctionSchema, str] = {
+        f.func: f"{f.namespace}::{f.func.name}" for f in native_functions
+    }
+    op_to_scoped_name: dict[OperatorName, str] = {
+        func.name: name for func, name in func_to_scoped_name.items()
+    }
+
+    schema_dict = {name: str(func) for func, name in func_to_scoped_name.items()}
+    kernel_persist_dict: dict[str, dict[str, Any]] = {
+        op_to_scoped_name[op]: v for op, v in persisted_fields.items()
+    }
+
+    if (
+        not native_yaml_path
+        or not os.path.exists(native_yaml_path)
+        or os.stat(native_yaml_path).st_size == 0
+    ):
+        return
+    with open(native_yaml_path) as native_yaml:
+        native_es = yaml.load(native_yaml, Loader=LineLoader)
+        if not native_es:
+            return
+        for e in native_es:
+            assert isinstance(e.get("__line__"), int), e
+            loc = Location(native_yaml_path, e.pop("__line__"))
+            with context(lambda: f"in {loc}:\n  "):
+                if "variants" not in e:
+                    e["variants"] = "function"
+                if "func" in e:
+                    continue
+                assert isinstance(e.get("op"), str), e
+                opname = e.pop("op")
+                if "::" not in opname:
+                    opname = "aten::" + opname
+                assert opname in schema_dict
+                e["func"] = schema_dict.get(opname)
+
+                # Write out persisted kernel information
+                if opname in kernel_persist_dict:
+                    for k, v in kernel_persist_dict[opname].items():
+                        e[k] = v
+
+        yaml.dump(native_es, out_file, width=1000)
+
+
+def parse_yaml(
+    path: str | None,
+    tags_yaml_path: str,
+    function_filter: Callable[[NativeFunction], bool],
+    skip_native_fns_gen: bool = False,
+) -> tuple[
+    list[NativeFunction],
+    dict[DispatchKey, dict[OperatorName, BackendMetadata]] | ETKernelIndex,
+]:
+    if path and os.path.exists(path) and os.stat(path).st_size > 0:
+        with open(path) as f:
+            es = yaml.load(f, Loader=LineLoader)
+
+        # Check for kernel index structure
+        kernel_index = (
+            parse_et_yaml_struct(es) if any("kernels" in e for e in es) else None
+        )
+
+        # Remove ET specific fields from entries for BC compatibility
+        for entry in es:
+            for field in ET_FIELDS:
+                entry.pop(field, None)
+
+        parsed_yaml = parse_native_yaml(
+            path,
+            tags_yaml_path,
+            None,
+            skip_native_fns_gen=skip_native_fns_gen,
+            loaded_yaml=es,
+        )
+        native_functions = list(filter(function_filter, parsed_yaml.native_functions))
+        op_names = [f.func.name for f in native_functions]
+
+        # (1) Return ETKernelIndex if kernel index is present
+        if kernel_index is not None:
+            filtered_index = {
+                op_name: kernel_mapping
+                for op_name, kernel_mapping in kernel_index.index.items()
+                if op_name in op_names
+            }
+            return native_functions, ETKernelIndex(index=filtered_index)
+
+        # (2) Return BackendIndices if kernel index is absent
+        def map_index(
+            m: dict[OperatorName, BackendMetadata],
+        ) -> dict[OperatorName, BackendMetadata]:
+            return {op: m[op] for op in m if op in op_names}
+
+        backend_indices = {
+            k: map_index(b.index) for (k, b) in parsed_yaml.backend_indices.items()
+        }
+
+        return native_functions, backend_indices
+    else:
+        return [], {}
+
+
+def parse_yaml_files(
+    tags_yaml_path: str,
+    aten_yaml_path: str,
+    native_yaml_path: str | None,
+    custom_ops_yaml_path: str | None,
+    selector: SelectiveBuilder,
+    use_aten_lib: bool,
+) -> tuple[ETParsedYaml, ETParsedYaml | None]:
+    """Parses functions.yaml and custom_ops.yaml files.
+
+    Args:
+        tags_yaml_path: Path to a tags.yaml file to satisfy codegen parsing.
+            It is not optional.
+        aten_yaml_path: Path to ATen operator yaml file native_functions.yaml.
+        native_yaml_path: Path to a functions.yaml file to parse.
+            If the path does not exist in the filesystem, it is treated as an
+            empty file. If `custom_ops_yaml_path` exists, the contents of that
+            file are appended to the yaml input to be parsed.
+        custom_ops_yaml_path: Path to a custom_ops.yaml file to parse. If
+            the path does not exist in the filesystem, it is ignored.
+        selector: For selective build.
+        use_aten_lib: We use this flag to determine if we want to generate native
+            functions. In ATen mode we should generate out= variants.
+    Returns:
+        A tuple with two elements:
+        [0]: The parsed results of concatenating the contents of
+             `native_yaml_path` and `custom_ops_yaml_path`.
+        [1]: The parsed results of the contents of `custom_ops_yaml_path`, if
+             present. If not present, None.
+    """
+    import tempfile
+
+    # only include selected ops, this is because we want to avoid
+    def function_filter(f: NativeFunction) -> bool:
+        return selector.is_native_function_selected(f)
+
+    with tempfile.TemporaryDirectory() as tmpdirname:
+        translated_yaml_path = os.path.join(tmpdirname, "translated.yaml")
+        with open(translated_yaml_path, "w") as translated:
+            translate_native_yaml(
+                tags_yaml_path,
+                aten_yaml_path,
+                native_yaml_path,
+                use_aten_lib,
+                translated,
+            )
+
+        translated_functions, translated_indices = parse_yaml(
+            translated_yaml_path, tags_yaml_path, function_filter, not use_aten_lib
+        )
+        custom_ops_functions, custom_ops_indices = parse_yaml(
+            custom_ops_yaml_path, tags_yaml_path, function_filter, True
+        )
+
+        # Convert BackendIndices to ETKernelIndex
+        if not isinstance(translated_indices, ETKernelIndex):
+            translated_indices = ETKernelIndex.from_backend_indices(translated_indices)
+        if not isinstance(custom_ops_indices, ETKernelIndex):
+            custom_ops_indices = ETKernelIndex.from_backend_indices(custom_ops_indices)
+
+        combined_functions = translated_functions + custom_ops_functions
+        combined_kernel_index = ETKernelIndex.merge_indices(
+            translated_indices, custom_ops_indices
+        )
+        combined_yaml = ETParsedYaml(combined_functions, combined_kernel_index)
+        custom_ops_parsed_yaml = ETParsedYaml(custom_ops_functions, custom_ops_indices)
+
+    return combined_yaml, custom_ops_parsed_yaml
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(description="Generate operator source files")
+    # Although we don't refer to --source-path directly, make_file_manager()
+    # expects it to point to a directory that contains a templates/ subdirectory
+    # containing the file templates.
+    parser.add_argument(
+        "-s",
+        "--source-path",
+        help="path to source directory for kernel templates",
+    )
+    parser.add_argument(
+        "--functions-yaml-path",
+        "--functions_yaml_path",
+        help="path to the functions.yaml file to use. Optional, but at least "
+        "one of --functions-yaml-path and --custom-ops-yaml-path must be "
+        "specified.",
+    )
+    parser.add_argument(
+        "--custom-ops-yaml-path",
+        "--custom_ops_yaml_path",
+        help="path to the custom_ops.yaml file to use. Optional, but at least "
+        "one of --functions-yaml-path and --custom-ops-yaml-path must be "
+        "specified.",
+    )
+    parser.add_argument(
+        "--aten-yaml-path",
+        "--aten_yaml_path",
+        help="path to native_functions.yaml file.",
+    )
+    # Note that make_file_manager() also looks at --install-dir.
+    parser.add_argument(
+        "-d",
+        "--install-dir",
+        "--install_dir",
+        help="output directory",
+        default="build/generated",
+    )
+    parser.add_argument(
+        "-o",
+        "--output-dependencies",
+        help="output a list of dependencies into the given file and exit",
+    )
+    # Although we don't refer to --dry-run directly, make_file_manager() looks
+    # for it.
+    parser.add_argument(
+        "--dry-run",
+        action="store_true",
+        help="run without writing any files (still updates outputs)",
+    )
+    parser.add_argument(
+        "--static-dispatch-backend",
+        "--static_dispatch_backend",
+        nargs="*",
+        help="generate static dispatch code for the specific backend (if set)",
+    )
+    parser.add_argument(
+        "--op-registration-whitelist",
+        "--op_registration_whitelist",
+        nargs="*",
+        help="filter op registrations by the whitelist (if set); "
+        "each item is `namespace`::`operator name` without overload name; "
+        "e.g.: aten::empty aten::conv2d ...",
+    )
+    parser.add_argument(
+        "--op-selection-yaml-path",
+        "--op_selection_yaml_path",
+        help="Provide a path to the operator selection (for custom build) YAML "
+        "that contains the information about the set of selected operators "
+        "and their categories (training, ...). Each operator is either a "
+        "full operator name with overload or just a bare operator name. "
+        "The operator names also contain the namespace prefix (e.g. aten::)",
+    )
+    parser.add_argument(
+        "--tags-path",
+        help="Path to tags.yaml. Required by yaml parsing in codegen system.",
+    )
+    parser.add_argument(
+        "--rocm",
+        action="store_true",
+        help="reinterpret CUDA as ROCm/HIP and adjust filepaths accordingly",
+    )
+    parser.add_argument(
+        "--use-aten-lib",
+        "--use_aten_lib",
+        action="store_true",
+        help="a boolean flag to indicate whether we use ATen kernels or not, in the future this flag will be per "
+        "operator",
+    )
+    parser.add_argument(
+        "--manual_registration",
+        "--manual-registration",
+        action="store_true",
+        help="a boolean flag to indicate whether we want to manually call"
+        "register_kernels() or rely on static init. ",
+    )
+    parser.add_argument(
+        "--generate",
+        type=str,
+        nargs="*",
+        choices=["headers", "sources"],
+        default=["headers", "sources"],
+        help="Generate only a subset of files",
+    )
+    options = parser.parse_args()
+    assert options.tags_path, "tags.yaml is required by codegen yaml parsing."
+
+    selector = get_custom_build_selector(
+        options.op_registration_whitelist,
+        options.op_selection_yaml_path,
+    )
+
+    parsed_yaml, custom_ops_parsed_yaml = parse_yaml_files(
+        aten_yaml_path=options.aten_yaml_path,
+        tags_yaml_path=options.tags_path,
+        native_yaml_path=options.functions_yaml_path,
+        custom_ops_yaml_path=options.custom_ops_yaml_path,
+        selector=selector,
+        use_aten_lib=options.use_aten_lib,
+    )
+    native_functions, kernel_index = (
+        parsed_yaml.native_functions,
+        parsed_yaml.kernel_index,
+    )
+    custom_ops_native_functions = (
+        custom_ops_parsed_yaml.native_functions if custom_ops_parsed_yaml else []
+    )
+
+    cpu_fm = make_file_manager(options=options)
+
+    if "headers" in options.generate:
+        # generate CustomOpsNativeFunctions.h when custom_ops.yaml is present, to match the build system.
+        gen_headers(
+            native_functions=native_functions,
+            gen_custom_ops_header=options.custom_ops_yaml_path,
+            custom_ops_native_functions=custom_ops_native_functions,
+            selector=selector,
+            kernel_index=kernel_index,
+            cpu_fm=cpu_fm,
+            use_aten_lib=options.use_aten_lib,
+        )
+
+    if "sources" in options.generate:
+        gen_unboxing(
+            native_functions=native_functions,
+            cpu_fm=cpu_fm,
+            selector=selector,
+            use_aten_lib=options.use_aten_lib,
+            kernel_index=kernel_index,
+            manual_registration=options.manual_registration,
+        )
+        if custom_ops_native_functions:
+            gen_custom_ops(
+                native_functions=custom_ops_native_functions,
+                selector=selector,
+                kernel_index=kernel_index,
+                cpu_fm=cpu_fm,
+                rocm=options.rocm,
+            )
+
+    if options.output_dependencies:
+        depfile_path = Path(options.output_dependencies).resolve()
+        depfile_name = depfile_path.name
+        depfile_stem = depfile_path.stem
+
+        for fm, prefix in [
+            (cpu_fm, ""),
+        ]:
+            varname = prefix + depfile_stem
+            path = depfile_path.parent / (prefix + depfile_name)
+            fm.write_outputs(varname, str(path))
+
+
+if __name__ == "__main__":
+    main()

gen_functionalization_type.py

@@ -0,0 +1,882 @@
+from __future__ import annotations
+
+from dataclasses import dataclass
+from typing import Callable, TYPE_CHECKING
+
+from torchgen.api import cpp, dispatcher
+from torchgen.api.translate import translate
+from torchgen.api.types import (
+    BaseCType,
+    Binding,
+    CType,
+    DispatcherSignature,
+    FunctionalizationLambda,
+    iTensorListRefT,
+    NativeSignature,
+    OptionalCType,
+    optionalSymIntArrayRefT,
+    symIntArrayRefT,
+    SymIntT,
+    tensorListT,
+    tensorT,
+    VectorCType,
+    ViewInverseSignature,
+)
+from torchgen.context import (
+    method_with_native_function,
+    native_function_manager,
+    with_native_function,
+    with_native_function_and,
+)
+from torchgen.model import (
+    Argument,
+    BackendIndex,
+    BaseTy,
+    BaseType,
+    FunctionSchema,
+    ListType,
+    NativeFunction,
+    NativeFunctionsGroup,
+    NativeFunctionsViewGroup,
+    Return,
+    SchemaKind,
+    SelfArgument,
+    TensorOptionsArguments,
+)
+from torchgen.native_function_generation import (
+    INPLACE_OPS_THAT_DONT_GET_GROUPED_PROPERLY,
+    MUTABLE_OPS_THAT_CANNOT_GET_AN_OUT_VARIANT,
+    OUT_OPS_THAT_DONT_GET_GROUPED_PROPERLY,
+)
+from torchgen.utils import dataclass_repr
+
+
+if TYPE_CHECKING:
+    from torchgen.selective_build.selector import SelectiveBuilder
+
+
+# Note: [Mutable Ops Not Using Functionalization]
+# Ops in this list currently do not work with functionalization and should be fixed.
+MUTABLE_OPS_NOT_USING_FUNCTIONALIZATION = (
+    OUT_OPS_THAT_DONT_GET_GROUPED_PROPERLY
+    + MUTABLE_OPS_THAT_CANNOT_GET_AN_OUT_VARIANT
+    + INPLACE_OPS_THAT_DONT_GET_GROUPED_PROPERLY
+    + [
+        # It will be BC-breaking, but we should fix their schemas.
+        # should be inplace?
+        "record_stream",
+        # See Note [resize_ in Functionalization]
+        "resize_",
+        "resize_as_",
+        # This function is used as for testing purposes only.
+        "_fill_mem_eff_dropout_mask_",
+    ]
+)
+
+# This file contains codegen that relates to the functionalization pass.
+# It includes:
+# - gen_functionalization_definition
+#     Generates dispatcher kernel definitions for the functionalization pass.
+# - gen_functionalization_registration
+#     Generates dispatcher kernel registrations for the functionalization pass.
+# - gen_functionalization_view_inverse_declaration
+#     Generates a declaration for an "inverse view", for every view op
+#     that is needed in functionalization. We manually implement their definitions.
+# - gen_composite_view_copy_kernel
+#     Generates view_copy() composite kernels for all view_copy operators.
+
+
+# Generates the body of the default composite C++ kernel for a {view}_copy NativeFunction
+# See Note [view_copy NativeFunctions]
+@dataclass(frozen=True)
+class GenCompositeViewCopyKernel:
+    backend_index: BackendIndex
+
+    @method_with_native_function
+    def __call__(self, g: NativeFunctionsViewGroup) -> str | None:
+        if g.view_copy is None:
+            return None
+        elif g.view_copy.func.name.name.base != f"{g.view.func.name.name}_copy":
+            # If the view_copy doesn't match the standard naming scheme of <op>_copy,
+            # assume it already exists and doesn't need to be generated.
+            # Example: slice_inverse() with the copy variant named slice_scatter()
+            # instead of slice_inverse_copy()
+            return None
+
+        metadata = self.backend_index.get_kernel(g.view_copy)
+        assert metadata is not None
+
+        # We can make view_copy work in more cases by using reshape()
+        # when a normal view call would ordinarily fail.
+        # This also makes LTC more efficient, because they don't need to include
+        # clone() calls in their graph (which is normally needed by reshape).
+        if str(g.view_copy.func.name) == "view_copy":
+            assert metadata.kernel == "view_copy_symint"
+            return """\
+at::Tensor view_copy_symint(const at::Tensor & self, at::SymIntArrayRef size) {
+  c10::SymDimVector shape = infer_size_dv(size, self.sym_numel());
+  if (!at::detail::computeStride(self.sym_sizes(), self.sym_strides(), shape).has_value()) {
+    return self.reshape_symint(size);
+  } else {
+    auto output = at::_ops::view::call(self, size);
+    return output.clone(/*memory_format=*/at::MemoryFormat::Contiguous);
+  }
+}
+"""
+        # view_copy is a native signature, since we're generating an at::native:: kernel
+        # Functionalization always operates on symints though
+        view_copy_sig = NativeSignature(
+            g.view_copy.func, symint=metadata.supports_symint()
+        )
+
+        # view is a dispatcher signature, since we're calling into the at::_ops API
+        view_sig = DispatcherSignature(g.view.func)
+
+        view_api_name = g.view.func.name.unambiguous_name()
+        exprs = ", ".join(
+            [e.expr for e in translate(view_copy_sig.arguments(), view_sig.arguments())]
+        )
+
+        # view ops today always return either a Tensor or a list of Tensors
+        assert len(g.view.func.returns) == 1
+        assert g.view.func.returns[0].type == BaseType(
+            BaseTy.Tensor
+        ) or g.view.func.returns[0].type == ListType(BaseType(BaseTy.Tensor), None)
+
+        if g.view.func.returns[0].type == BaseType(BaseTy.Tensor):
+            return_cloned_output = """\
+  return output.clone(/*memory_format=*/at::MemoryFormat::Contiguous);"""
+        else:
+            # If the return type is a list, we need to clone each tensor in the list.
+            return_cloned_output = f"""\
+  {view_copy_sig.returns_type().cpp_type()} out_clone;
+  for (const auto i : c10::irange(output.size())) {{
+    out_clone.push_back(output[i].clone(/*memory_format=*/at::MemoryFormat::Contiguous));
+  }}
+  return out_clone;"""
+
+        # The default generated composite kernel for {view}_copy() operators just clones
+        # the input tensor, and runs the underlying view on the clone.
+        return f"""
+{view_copy_sig.defn(name=metadata.kernel)} {{
+  auto output = at::_ops::{view_api_name}::call({exprs});
+  {return_cloned_output}
+}}
+"""
+
+
+def return_str(rets: tuple[Return, ...], names: list[str]) -> str:
+    assert len(rets) == len(names)
+    if len(rets) == 0:
+        return ""
+    elif len(rets) == 1:
+        return f"return {names[0]};"
+    else:
+        return f"return {dispatcher.returns_type(rets).cpp_type()}({', '.join(names)});"
+
+
+def modifies_arguments(f: NativeFunction) -> bool:
+    return any(
+        a.annotation is not None and a.annotation.is_write
+        for a in f.func.arguments.flat_all
+    )
+
+
+def wrapper_name(func: FunctionSchema) -> str:
+    if func.name.overload_name:
+        return f"{cpp.name(func)}_{func.name.overload_name}"
+    else:
+        return cpp.name(func)
+
+
+def is_tensor_like(a: Argument | TensorOptionsArguments | SelfArgument) -> bool:
+    return isinstance(a, SelfArgument) or (
+        isinstance(a, Argument) and a.type.is_tensor_like()
+    )
+
+
+# We need to wrap / unwrap various arguments from the op in the functionalization kernels.
+# Some op schemas include non-owning types though (like TensorList),
+# and when we unwrap them we expect to get out an owning type!.
+# We also return a lambda that tells you how to conver the non-owning type argument into the owning type.
+def get_owning_type(t: CType) -> tuple[CType, Callable[[str], str]]:
+    if t == BaseCType(tensorListT):
+        return VectorCType(BaseCType(tensorT)), lambda x: f"{x}.vec()"
+    if t == BaseCType(iTensorListRefT):
+        return VectorCType(BaseCType(tensorT)), lambda x: f"{{{x}.begin(), {x}.end()}}"
+    # There are technically other non-owning types out there (like IntArrayRef),
+    # but functionalization only actually cares about the ones involving tensors.
+    return t, lambda x: x
+
+
+# unwraps all tensor-like arguments, returning:
+# (1) a string containing all of the logic that does the unwrapping
+# (2) a context, to be used by translate(), with all of the relevant bindings.
+def unwrap_tensor_args(
+    sig: DispatcherSignature, *, is_view_op: bool
+) -> tuple[str, list[Binding]]:
+    context: list[Binding] = []
+    unwrapped_tensor_args: list[str] = []
+    for arg in sig.arguments():
+        if is_tensor_like(arg.argument):
+            # for tensor inputs, we want to unwrap them before passing them into the redispatch calls.
+            unwrapped_name = f"{arg.name}_"
+            # For most ops, the functionalization needs to sync any pending updates on the input tensors
+            # before calling the operator, since otherwise the operator will act on stale data.
+            # For view ops though, we can continue to defer syncing until the tensor is used by
+            # a non-view operator.
+            maybe_sync_input = (
+                "" if is_view_op else f"at::functionalization::impl::sync({arg.name});"
+            )
+            unwrapped_type, conversion_fn = get_owning_type(
+                arg.nctype.remove_const_ref().type
+            )
+            unwrapped_tensor_args.append(
+                f"""
+      {unwrapped_type.cpp_type()} {unwrapped_name};
+      if (at::functionalization::impl::isFunctionalTensor({arg.name})) {{
+        {maybe_sync_input}
+        {unwrapped_name} = at::functionalization::impl::from_functional_tensor({arg.name});
+      }} else {{
+        {unwrapped_name} = {conversion_fn(arg.name)};
+      }}"""
+            )
+            context.append(arg.with_name(unwrapped_name))
+        else:
+            # for non-tensor inputs, we want to pass them directly into the redispatch calls.
+            context.append(arg)
+    unwrap_tensor_args_str = "\n      ".join(unwrapped_tensor_args)
+    return unwrap_tensor_args_str, context
+
+
+# converts  all tensor-like arguments to meta tensors, which are used to compute stride info. Returns:
+# (1) a string containing all of the logic that does the conversions.
+# (2) a context, to be used by translate(), with all of the relevant bindings.
+def convert_to_meta_tensors(sig: DispatcherSignature) -> tuple[str, list[Binding]]:
+    context: list[Binding] = []
+    unwrapped_tensor_args: list[str] = []
+    for arg in sig.arguments():
+        if is_tensor_like(arg.argument):
+            # for tensor inputs, we want to unwrap them before passing them into the redispatch calls.
+            a_ = arg.name
+            unwrapped_name = f"{arg.name}_meta"
+            unwrapped_tensor_args.append(f"auto {unwrapped_name} = to_meta({a_});")
+            context.append(arg.with_name(unwrapped_name))
+        else:
+            # for non-tensor inputs, we want to pass them directly into the redispatch calls.
+            context.append(arg)
+    unwrap_tensor_args_str = "\n        ".join(unwrapped_tensor_args)
+    return unwrap_tensor_args_str, context
+
+
+# The functionalization codegen currently expects view op schemas to have this form:
+# foo(Tensor(a), ...) -> Tensor(a) (e.g. transpose)
+# foo(Tensor(a!), ...) -> Tensor(a!) (e.g. transpose_)
+def assert_view_op_properties(func: FunctionSchema) -> None:
+    def is_alias(a: Argument) -> bool:
+        return a.annotation is not None
+
+    args = func.arguments.flat_non_out
+    # The first argument is a tensor with an alias semantics (annotations)
+    assert (
+        len(args) > 0 and args[0].type == BaseType(BaseTy.Tensor)
+    ), f"""In the functionalization codegen, we expect the first argument of every view operator to be a tensor,
+but found an argument of type {str(args[0].type)} for operator: {str(func.name)}."""
+    # No other arguments have aliasing semantics
+    assert (
+        is_alias(args[0]) and not any(is_alias(a) for a in args[1:])
+    ), """In the functionalization codegen, we expect the first argument of every view operator to alias the output.
+View operators with multiple aliasing inputs aren't supported yet. Found an operator that doesn't satisfy this constraint"""
+
+
+# One-liner expression for checking if an expression expr of type type has any
+# symbolic values.
+def emit_expr_has_symbolic_values(expr: str, type: CType) -> str:
+    if type == BaseCType(SymIntT):
+        return f"{expr}.is_symbolic()"
+
+    if isinstance(type, OptionalCType):
+        innerexpr = f"(*{expr})"
+        return f"{expr}.has_value() ? {emit_expr_has_symbolic_values(innerexpr, type.elem)} : false"
+
+    if type == BaseCType(optionalSymIntArrayRefT):
+        return emit_expr_has_symbolic_values(
+            expr, OptionalCType(BaseCType(symIntArrayRefT))
+        )
+
+    if type in (BaseCType(symIntArrayRefT), VectorCType(BaseCType(SymIntT))):
+        argname = "arg"
+        lambda_check = emit_expr_has_symbolic_values(argname, BaseCType(SymIntT))
+        return (
+            "std::any_of("
+            f"{expr}.begin(), {expr}.end(), "
+            f"[=](auto& {argname}) {{ return {lambda_check}; }})"
+        )
+
+    raise ValueError(
+        "unsupported type for has_symbolic_values check. "
+        "It should be a SymInt or a collection of those. "
+        f"Got: {type.cpp_type()}"
+    )
+
+
+# Detects whether any of the SymInt arguments are, in fact, symbolic values.
+# This is used in the constructor of ViewMeta.
+def emit_has_symbolic_inputs(sig: DispatcherSignature) -> tuple[str, str]:
+    name = "has_symbolic_inputs"
+    statements = [
+        f"{name} = {name} | ({emit_expr_has_symbolic_values(binding.name, binding.nctype.type)});"
+        for binding in sig.arguments()
+        if (
+            isinstance(binding.argument, Argument)
+            and binding.argument.type.is_symint_like()
+        )
+    ]
+    body = "\n      ".join(statements)
+    return (
+        name,
+        f"""
+      bool {name} = false;
+      {body}""",
+    )
+
+
+# Generates the Functionalization kernel for:
+# - ops that create aliases (e.g. transpose())
+# - ops that are views AND mutations (e.g. transpose_())
+def emit_view_functionalization_body(
+    g: NativeFunctionsViewGroup, *, view_inplace: bool
+) -> str:
+    if view_inplace:
+        # This op is both an inplace op AND a view op.
+        # See Note [Functionalization Pass - Inplace View Ops] for details.
+        # I currently have the view meta call into the out-of-place variant of the view, to avoid
+        # having to define an extra ~20 inplace {view}_inverse_ functions.
+        # Most view ops don't have NativeFunctionGroup's both, because we don't define out= variants for view ops.
+        # I'm assuming that every inplace-view op has a corresponding out-of-place view op,
+        # with the same name but the trailing underscore removed.
+        # This is currently asserted at parse time in gen.py (see error_check_native_functions).
+        assert g.view_inplace is not None
+        f = g.view_inplace
+    else:
+        f = g.view
+
+    assert g.view_copy is not None
+    with native_function_manager(f):
+        call_sig = DispatcherSignature.from_schema(g.view_copy.func)
+
+        # the "view_copy" op name that the functionalization kernels need to call
+        api_name = g.view_copy.func.name.unambiguous_name()
+        # Sometimes the functionalization pass needs to no-op (e.g. if it was passed non-functional tensors)
+        # "no-op"ing in this context is just redispatching to the original op.
+        noop_api_name = f.func.name.unambiguous_name()
+
+        dispatcher_sig = DispatcherSignature.from_schema(f.func)
+        assert_view_op_properties(f.func)
+        view_tensor_name = dispatcher_sig.arguments()[0].name
+
+        return_type = dispatcher_sig.returns_type().remove_const_ref().cpp_type()
+
+        unwrap_tensor_args_str, unwrapped_args_ctx = unwrap_tensor_args(
+            dispatcher_sig, is_view_op=True
+        )
+        view_redispatch_args = [
+            e.expr
+            for e in translate(unwrapped_args_ctx, call_sig.arguments(), method=False)
+        ]
+
+        forward_lambda = FunctionalizationLambda.from_func(g, is_reverse=False)
+        reverse_lambda = FunctionalizationLambda.from_func(g, is_reverse=True)
+
+        # The meta API call should use the same arguments, but convert all tensors to meta tensors first.
+        meta_conversion_str, meta_call_ctx = convert_to_meta_tensors(dispatcher_sig)
+        meta_call_args = [
+            e.expr for e in translate(meta_call_ctx, call_sig.arguments(), method=False)
+        ]
+
+        (
+            symbolic_inputs_varname,
+            symbolic_inputs_check,
+        ) = emit_has_symbolic_inputs(call_sig)
+
+        if "inplace_view" in f.tags:
+            # See Note [Functionalization Pass - Inplace View Ops] for more details
+            return f"""
+    {dispatcher_sig.defn(name=wrapper_name(f.func), is_redispatching_fn=True)} {{
+      if (!at::functionalization::impl::isFunctionalTensor({view_tensor_name})) {{
+        // functionalization is re-entrant, but will no-op if it wasn't passed a FunctionalTensorWrapper.
+        {unwrap_tensor_args_str}
+        at::AutoDispatchSkipFunctionalize guard;
+        return at::_ops::{noop_api_name}::call({', '.join(view_redispatch_args)});
+      }}
+      auto reapply_views = at::functionalization::impl::getFunctionalizationReapplyViewsTLS();
+      auto inverse_return_mode = (
+          reapply_views ? at::functionalization::InverseReturnMode::ViewOrScatterInverse
+            : at::functionalization::InverseReturnMode::NeverView
+      );
+      {symbolic_inputs_check}
+      at::functionalization::ViewMeta view_meta = at::functionalization::ViewMeta(
+        {forward_lambda.decl()} {{
+          if (reapply_views) {{
+            return {forward_lambda.inner_call(reapply_views=True)}
+          }} else {{
+            return {forward_lambda.inner_call(reapply_views=False)}
+          }}
+        }},
+        {reverse_lambda.decl()} {{
+          return {reverse_lambda.inner_call()}
+        }},
+        /*has_symbolic_inputs=*/{symbolic_inputs_varname}
+      );
+      auto compute_reference_meta =
+        {view_tensor_name}.key_set().has_backend(c10::BackendComponent::XLABit) ||
+        {view_tensor_name}.key_set().has_backend(c10::BackendComponent::LazyBit);
+      {return_type} reference_tensor_output;
+      if (compute_reference_meta) {{
+        {meta_conversion_str}
+        at::AutoDispatchSkipFunctionalize func_guard;
+        c10::impl::ExcludeDispatchKeyGuard guard(exclude_keys_for_meta_dispatch);
+        reference_tensor_output = at::_ops::{noop_api_name}::call({', '.join(meta_call_args)});
+      }}
+      // This function adds the above view meta to the current tensor and replays them off the base,
+      // mutating the size/stride info of the current FunctionalTensorWrapper.
+      // Because of this, we need to make sure to run the reference shape function above,
+      // BEFORE doing this (otherwise we'll end up runnin the reference function using the wrong sizes/strides)
+      at::functionalization::impl::mutate_view_meta({view_tensor_name}, view_meta);
+      // See  Note [Propagating strides in the functionalization pass]
+      // XLA/LTC don't implement the logic to propagate strides correctly, so we need to rely
+      // on a reference implementation here (instead of relying on the output from the forward lambda
+      // having the correct stride info)
+      if (compute_reference_meta) {{
+        at::functionalization::impl::set_sizes_strides_offset({view_tensor_name}, reference_tensor_output);
+      }}
+      return {view_tensor_name};
+    }}
+"""
+
+        else:
+            is_multi_output_view = isinstance(f.func.returns[0].type, ListType)
+            return f"""
+    {dispatcher_sig.defn(name=wrapper_name(f.func), is_redispatching_fn=True)} {{
+      {unwrap_tensor_args_str}
+      if (!at::functionalization::impl::isFunctionalTensor({view_tensor_name})) {{
+        // functionalization is re-entrant, but will no-op if it wasn't passed a FunctionalTensorWrapper.
+        at::AutoDispatchSkipFunctionalize guard;
+        return at::_ops::{noop_api_name}::call({', '.join(view_redispatch_args)});
+      }}
+      auto reapply_views = at::functionalization::impl::getFunctionalizationReapplyViewsTLS();
+      auto inverse_return_mode = (
+          reapply_views ? at::functionalization::InverseReturnMode::ViewOrScatterInverse
+            : at::functionalization::InverseReturnMode::NeverView
+      );
+      auto compute_reference_meta =
+        {view_tensor_name}.key_set().has_backend(c10::BackendComponent::XLABit) ||
+        {view_tensor_name}.key_set().has_backend(c10::BackendComponent::LazyBit);
+      {return_type} reference_tensor_output;
+      if (compute_reference_meta) {{
+        {meta_conversion_str}
+        at::AutoDispatchSkipFunctionalize func_guard;
+        c10::impl::ExcludeDispatchKeyGuard guard(exclude_keys_for_meta_dispatch);
+        reference_tensor_output = at::_ops::{noop_api_name}::call({', '.join(meta_call_args)});
+      }}
+      {return_type} tmp_output;
+      {{
+        at::AutoDispatchSkipFunctionalize guard;
+        if (reapply_views) {{
+          tmp_output = at::_ops::{noop_api_name}::call({', '.join(view_redispatch_args)});
+        }} else {{
+          tmp_output = at::_ops::{api_name}::call({', '.join(view_redispatch_args)});
+        }}
+      }}
+      {symbolic_inputs_check}
+      at::functionalization::ViewMeta view_meta = at::functionalization::ViewMeta(
+        {forward_lambda.decl()} {{
+          if (reapply_views) {{
+            return {forward_lambda.inner_call(reapply_views=True)}
+          }} else {{
+            return {forward_lambda.inner_call(reapply_views=False)}
+          }}
+        }},
+        {reverse_lambda.decl()} {{
+          return {reverse_lambda.inner_call()}
+        }},
+        /*has_symbolic_inputs=*/{symbolic_inputs_varname},
+        /*is_multi_output=*/{str(is_multi_output_view).lower()},
+        /*is_as_strided=*/{str(str(f.func.name) == 'as_strided').lower()}
+      );
+      auto out = at::functionalization::impl::create_functional_tensor_with_view_meta(tmp_output, {view_tensor_name}, view_meta);
+      // See  Note [Propagating strides in the functionalization pass]
+      if (compute_reference_meta) {{
+        at::functionalization::impl::set_sizes_strides_offset(out, reference_tensor_output);
+      }}
+      return out;
+    }}
+"""
+
+
+def maybe_create_output(f: NativeFunction, var_name: str) -> str:
+    if len(f.func.returns) == 0:
+        return ""
+    return_type = dispatcher.returns_type(f.func.returns).remove_const_ref().cpp_type()
+    return f"{return_type} {var_name} = "
+
+
+# Given a NativeFunction, and a variable name corresponding to the output of redispatching on the function,
+# this returns two lists of names, consisting of:
+# - the names of returns corresponding to the original (mutable) inputs of the outer function
+# - the names of returns corresponding to the (immutable) outputs of the inner redispatched function
+def get_mutable_redispatch_return_names(
+    f: NativeFunction, inner_return_var: str
+) -> tuple[list[str], list[str]]:
+    aliased_returns = []
+    non_aliased_returns = []
+    for i, name in enumerate(f.func.aliased_return_names()):
+        if name is not None:
+            aliased_returns.append(name)
+        else:
+            non_aliased_returns.append(
+                inner_return_var
+                if len(f.func.returns) == 1
+                else f"std::get<{i}>({inner_return_var})"
+            )
+    return aliased_returns, non_aliased_returns
+
+
+# When functionalization "no-op's" and redispatches on a mutable operator, we need to take care so that:
+#  - For fresh outputs, we return the result of the redispatch (without wrapping outputs)
+#  - For outputs that were aliased to inputs, we return the inputs directly (since some of them might have been wrapped)
+def return_from_mutable_noop_redispatch(
+    f: NativeFunction, inner_return_var: str
+) -> str:
+    aliased, non_aliased = get_mutable_redispatch_return_names(f, inner_return_var)
+    # Just get all of the return names, and immediately return them
+    return return_str(f.func.returns, aliased + non_aliased)
+
+
+def wrap_propagate_mutations_and_return(
+    f: NativeFunction, functional_op: NativeFunction, inner_return_var: str
+) -> str:
+    mutable_arg_names = f.func.arguments.mutable_arg_names()
+    (
+        aliased_outer_rets,
+        non_aliased_outer_rets,
+    ) = get_mutable_redispatch_return_names(f, inner_return_var)
+    _, non_aliased_inner_rets = get_mutable_redispatch_return_names(
+        functional_op, inner_return_var
+    )
+    # The outer function may have a mix of aliased and non-aliased outputs,
+    # But the inner functional op that we're transforming to should only have non-aliased outputs
+    assert len(mutable_arg_names) + len(non_aliased_outer_rets) == len(
+        non_aliased_inner_rets
+    )
+
+    # First, take all of the newly created outputs from the inner call and wrap them into functional tensors
+    updates = []
+    non_aliased_wrapped_ret_names = []
+    for i, inner_ret in enumerate(
+        non_aliased_inner_rets[: len(non_aliased_outer_rets)]
+    ):
+        ret_name = f"output_{i}"
+        updates.append(
+            f"""\
+  auto output_{i} = at::functionalization::impl::to_functional_tensor({inner_ret});"""
+        )
+        non_aliased_wrapped_ret_names.append(ret_name)
+
+    # Next, take all of the mutated outputs from the inner call corresponding to mutated inputs,
+    # and propagate the mutations
+    for outer_arg, inner_ret in zip(
+        mutable_arg_names, non_aliased_inner_rets[len(non_aliased_outer_rets) :]
+    ):
+        updates.append(
+            f"""\
+  auto {outer_arg}_inner = at::functionalization::impl::from_functional_tensor({outer_arg});
+  at::functionalization::impl::replace_({outer_arg}, {inner_ret});
+  at::functionalization::impl::commit_update({outer_arg});
+  at::functionalization::impl::sync({outer_arg});
+  auto {outer_arg}_inner_updated = at::functionalization::impl::from_functional_tensor({outer_arg});
+  at::functionalization::impl::propagate_xla_data_direct({outer_arg}_inner, {outer_arg}_inner_updated);"""
+        )
+
+    # Finally, we return:
+    # - Any mutable arguments that also returns
+    # - Any immutable returns that were created wrapping the output from the inner call
+    returns_str = return_str(
+        f.func.returns, aliased_outer_rets + non_aliased_wrapped_ret_names
+    )
+    updates_str = "\n".join(updates)
+    return f"""\
+{updates_str}
+    {returns_str}"""
+
+
+# Generates the Functionalization kernel for:
+# - mutation ops (inplace and out= ops)
+@with_native_function_and
+def emit_inplace_functionalization_body(
+    f: NativeFunction, g: NativeFunctionsGroup
+) -> str:
+    # mutation case
+    assert modifies_arguments(f)
+
+    dispatcher_sig = DispatcherSignature.from_schema(f.func)
+
+    unwrap_tensor_args_str, unwrapped_args_ctx = unwrap_tensor_args(
+        dispatcher_sig, is_view_op=False
+    )
+
+    mutated_names = [
+        a.name
+        for a in f.func.arguments.flat_all
+        if a.type.is_tensor_like() and a.annotation is not None
+    ]
+    non_mutated_names = [
+        a.name
+        for a in f.func.arguments.flat_all
+        if a.type.is_tensor_like() and a.annotation is None
+    ]
+    non_mutated_tensor_names = [
+        a.name
+        for a in f.func.arguments.flat_all
+        if a.type == BaseType(BaseTy.Tensor) and a.annotation is None
+    ]
+    # all mutable inputs must be functional tensors in order to participate in functionalization
+    check_all_mutated_args_are_functional = " && ".join(
+        ["true"]
+        + [
+            f"at::functionalization::impl::isFunctionalTensor({a})"
+            for a in mutated_names
+        ]
+    )
+    check_any_non_mutated_args_are_functional = " || ".join(
+        ["false"]
+        + [
+            f"at::functionalization::impl::isFunctionalTensor({a})"
+            for a in non_mutated_names
+        ]
+    )
+
+    check_any_non_mutated_tensors_are_xla = " || ".join(
+        ["false"]
+        + [
+            f"{a}.device().type() == c10::DeviceType::XLA"
+            for a in non_mutated_tensor_names
+        ]
+    )
+    # These are used in the cases where we don't functionalize and redispatch to the inplace op
+    # case 1: we hit an inplace op that doesn't have an out-of-place equivalent
+    # case 2: we hit an inplace ops but our inputs are not functional tensors (in which case our kernel just no-ops)
+    inplace_exprs = [
+        e.expr
+        for e in translate(unwrapped_args_ctx, dispatcher_sig.arguments(), method=False)
+    ]
+
+    # call the out-of-place variant of the op
+    return_type = (
+        dispatcher.returns_type(g.functional.func.returns).remove_const_ref().cpp_type()
+    )
+    functional_sig = DispatcherSignature.from_schema(g.functional.func)
+    functional_exprs = [
+        e.expr
+        for e in translate(unwrapped_args_ctx, functional_sig.arguments(), method=False)
+    ]
+
+    if f.func.is_out_fn():
+        mutable_input_post_processing = "\n".join(
+            [
+                f"""
+      at::functionalization::impl::replace_(
+        {a.name}, {'std::get<' + str(i) + '>(tmp_output)' if len(f.func.returns) > 1 else 'tmp_output'});
+      at::functionalization::impl::commit_update({a.name});"""
+                for (i, a) in enumerate(f.func.arguments.out)
+                if a.annotation and a.annotation.is_write and a.type.is_tensor_like()
+            ]
+        )
+    else:
+        mutable_input_post_processing = "\n".join(  # noqa: F841
+            [
+                f"""
+      at::functionalization::impl::replace_({a.name}, tmp_output);
+      at::functionalization::impl::commit_update({a.name});"""
+                for a in f.func.arguments.flat_all
+                if a.annotation and a.annotation.is_write and a.type.is_tensor_like()
+            ]
+        )
+
+    meta_conversion_str, meta_call_ctx = convert_to_meta_tensors(dispatcher_sig)
+    # We don't want to run the inplace meta func for ops like .set_(), because:
+    # (1) they're unnecessary: inplace meta checks are only useful for ops like add_(),
+    #     where broadcasting will work for the out-of-place case but should fail on the inplace call
+    # (2) They'll also fail without adding extra infra: we'd need to convert the input storage argument
+    #     into a meta storage
+    any_storage_args = any(
+        a.type == BaseType(BaseTy.Storage) for a in f.func.arguments.flat_all
+    )
+
+    return f"""
+    {dispatcher_sig.defn(name=wrapper_name(f.func), is_redispatching_fn=True)} {{
+      if ({str(not any_storage_args and f.func.kind() == SchemaKind.inplace).lower()}) {{
+        // Before converting the mutable op to its functional variant, run meta tensors through the original op.
+        // This will help us catch shape errors that apply to inplace ops that wouldn't apply to their functional variants.
+        // (We can only do this for inplace ops today though, because they technically all support meta tensors).
+        {meta_conversion_str}
+        at::AutoDispatchSkipFunctionalize func_guard;
+        c10::impl::ExcludeDispatchKeyGuard guard(exclude_keys_for_meta_dispatch);
+        at::_ops::{f.func.name.unambiguous_name()}::call({', '.join(a.name for a in meta_call_ctx)});
+      }}
+      {unwrap_tensor_args_str}
+      if (!({check_all_mutated_args_are_functional})) {{
+        // We want to disable this check if there are any XLA tensors.
+        // cpu_tensor.copy_(xla_tensor) is valid code.
+        if (!({check_any_non_mutated_tensors_are_xla}) && ({check_any_non_mutated_args_are_functional})) {{
+         // case 1: trying to mutate a non functional tensor with a functional tensor is an error
+         TORCH_INTERNAL_ASSERT(false,
+           "mutating a non-functional tensor with a functional tensor is not allowed.",
+           " Please ensure that all of your inputs are wrapped inside of a functionalize() call.");
+        }} else {{
+         // case 2: arguments are not functional tensors, so we no-op and redispatch.
+         at::AutoDispatchSkipFunctionalize guard;
+         {maybe_create_output(f, 'tmp_output')}at::_ops::{f.func.name.unambiguous_name()}::call({', '.join(inplace_exprs)});
+         {return_from_mutable_noop_redispatch(f, 'tmp_output')}
+        }}
+      }} else {{
+        {return_type} tmp_output;
+        {{
+          at::AutoDispatchSkipFunctionalize guard;
+          tmp_output = at::_ops::{g.functional.func.name.unambiguous_name()}::call({', '.join(functional_exprs)});
+        }}
+        {wrap_propagate_mutations_and_return(f, g.functional, 'tmp_output')}
+      }}
+    }}"""
+
+
+# The below functions generate RegisterFunctionalization.cpp
+# These files provide the kernels that run the functionalization pass, which can be opted into
+# per backend (e.g. XLA or Vulkan), or as a composable transform (functionalize() in functorch).
+
+
+# See Note [Functionalization Pass: View Inverses].
+def gen_functionalization_view_inverse_declaration(
+    selector: SelectiveBuilder, g: NativeFunctionsViewGroup
+) -> str | None:
+    # For every (non-composite) view op, we need a corresponding "inverse view" function.
+    # This generates the declarations so we get a good compiler error when someone adds a new view.
+    @with_native_function
+    def emit_decl_helper(g: NativeFunctionsViewGroup) -> str | None:
+        if g.view.has_composite_implicit_autograd_kernel:
+            return None
+        view_inverse_sig = ViewInverseSignature(g)
+        return view_inverse_sig.decl()
+
+    return emit_decl_helper(g)
+
+
+def gen_functionalization_registration(
+    selector: SelectiveBuilder,
+    g: NativeFunction | NativeFunctionsGroup | NativeFunctionsViewGroup,
+    composite_implicit_autograd_index: BackendIndex,
+) -> list[str]:
+    @with_native_function
+    def emit_registration_helper(f: NativeFunction) -> str:
+        assert not f.has_composite_implicit_autograd_kernel
+        registration_str = f"TORCH_FN(functionalization::{wrapper_name(f.func)})"
+        return f'm.impl("{f.func.name}", {registration_str});'
+
+    # Don't generate kernels in mobile build
+    if not selector.include_all_operators:
+        return []
+
+    if isinstance(g, NativeFunctionsViewGroup):
+        # functionalization needs to register kernels for view + view_inplace ops
+        # See Note [Functionalization <> torch.Tensor constructor]
+        if str(g.view.func.name) == "lift_fresh":
+            return []
+        view_str = []
+        if not g.view.has_composite_implicit_autograd_kernel:
+            view_str.append(emit_registration_helper(g.view))
+        if (
+            g.view_inplace is not None
+            and not g.view_inplace.has_composite_implicit_autograd_kernel
+        ):
+            assert g.view_inplace.is_view_op
+            view_str.append(emit_registration_helper(g.view_inplace))
+        return view_str
+
+    elif isinstance(g, NativeFunctionsGroup):
+        # Gets a hand-written functionalization kernel
+        if g.inplace is not None and str(g.inplace.func.name) == "set_.source_Tensor":
+            fns = []
+        else:
+            fns = list(g.functions())
+    else:
+        if str(g.func.name) in MUTABLE_OPS_NOT_USING_FUNCTIONALIZATION:
+            return []
+        fns = [g]
+
+    registrations = []
+    for f in fns:
+        if f.has_composite_implicit_autograd_kernel:
+            continue
+        if str(f.func.name) == "lift":
+            # See Note [Functionalization <> torch.Tensor constructor]
+            return []
+        if str(f.func.name) == "resize_":
+            # See Note [resize_ in Functionalization]
+            return []
+        if str(f.func.name.name) != "set_":
+            assert not f.is_view_op
+        # functionalization needs to generate and register kernels for inplace ops.
+        # We *also* need to directly register CompositeImplicitAUtograd kernels
+        # so that they decompose properly before functioanlization.
+        if modifies_arguments(f):
+            registrations.append(emit_registration_helper(f))
+    return registrations
+
+
+def gen_functionalization_definition(
+    selector: SelectiveBuilder,
+    # Note: Ideally this code should never have to look at NativeFunction
+    # (and instead only need to operate on grouped NativeFunctions).
+    # The only reason currently is because we need to emit direct dispatch registrations
+    # For CompositeImplicitAutograd operators, which are potentially ungrouped.
+    g: NativeFunction | NativeFunctionsGroup | NativeFunctionsViewGroup,
+) -> list[str]:
+    # Don't generate kernels in mobile build
+    if not selector.include_all_operators:
+        return []
+
+    if isinstance(g, NativeFunctionsViewGroup):
+        # Case 1: emit view -> view_copy kernels for the functionalization pass
+        view_defs = []
+        if not g.composite:
+            # invariant: NativeFunctionsViewGroup's always have a view_copy operator
+            # if the view is not composite (implicit autograd)
+            assert g.view_copy is not None, dataclass_repr(g, indent=1)
+            view_defs.append(emit_view_functionalization_body(g, view_inplace=False))
+            if g.view_inplace is not None:
+                view_defs.append(emit_view_functionalization_body(g, view_inplace=True))
+        return view_defs
+    elif isinstance(g, NativeFunction):
+        # Invariant: all mutable operators that we need to handle in functionalization
+        # should have been properly grouped up.
+        # TODO: The below ops all have "problematic" schemas that prevent them from
+        # getting functionalized. Instead of bending over backwards to get things to work,
+        # I think we should either:
+        # (1) fix their schemas (BC-breaking)
+        # (2) hand-write their functionalization kernels
+        if (
+            str(g.func.name) not in MUTABLE_OPS_NOT_USING_FUNCTIONALIZATION
+            and str(g.func.name.name) not in MUTABLE_OPS_NOT_USING_FUNCTIONALIZATION
+        ):
+            assert g.has_composite_implicit_autograd_kernel or not modifies_arguments(g)
+        return []
+    else:
+        # Case 2: emit inplace -> out-of-place kernels for the functionalization pass
+        mutation_defs = []
+        mutation_defs.append(emit_inplace_functionalization_body(g.out, g))
+        if g.inplace is not None:
+            mutation_defs.append(emit_inplace_functionalization_body(g.inplace, g))
+        if g.mutable is not None:
+            mutation_defs.append(emit_inplace_functionalization_body(g.mutable, g))
+        return mutation_defs
+    return []

gen_lazy_tensor.py

@@ -0,0 +1,585 @@
+from __future__ import annotations
+
+import argparse
+import os
+from collections import namedtuple
+from pathlib import Path
+from typing import Any, Callable, TYPE_CHECKING
+
+import yaml
+
+import torchgen.dest as dest
+from torchgen.api.lazy import setValueT
+from torchgen.api.types import BaseCppType
+from torchgen.dest.lazy_ir import GenLazyIR, GenLazyNativeFuncDefinition, GenTSLazyIR
+from torchgen.gen import get_grouped_native_functions, parse_native_yaml
+from torchgen.gen_backend_stubs import (
+    error_on_missing_kernels,
+    gen_dispatcher_registrations,
+    gen_dispatchkey_nativefunc_headers,
+    parse_backend_yaml,
+)
+from torchgen.model import NativeFunction, NativeFunctionsGroup, OperatorName
+from torchgen.selective_build.selector import SelectiveBuilder
+from torchgen.utils import FileManager, NamespaceHelper
+from torchgen.yaml_utils import YamlLoader
+
+
+if TYPE_CHECKING:
+    from collections.abc import Iterable, Iterator, Sequence
+
+
+# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ #
+#
+#                        Lazy Tensor Codegen
+#
+# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ #
+# Overview
+# ~~~~~~~~
+#
+# This codegen script builds on existing data models and helpers used
+# by all ATen backends, and adds new functionality specific to lazy
+# tensor backends.
+#
+# Inputs:
+# - <backend>_native_functions.yaml: controls which operators are
+#   supported by the backend.
+#
+# Outputs:
+# (for all backends)
+# <DispatchKey>Ir.h defines Lazy IR classes to be constructed during tracing
+# - opt-in: also generate 'lowering' methods for the TorchScript backend only
+# <DispatchKey>NativeFunctions.cpp defines implementations of native functions which perform lazy tracing
+# - opt-in: 'full_codegen' section of backend yaml; 'supported' section omits these implementations
+# <DispatchKey>NativeFunctions.h declares implementations of native functions for both 'supported' and 'full_codegen'
+# ops
+#
+# Register<DispatchKey>.cpp registers all op implementations with the dispatcher
+# RegisterAutograd<DispatchKey>.cpp registers all autograd implementations with the dispatcher
+#
+# Validation Helpers:
+# - Shape Inference: errs if any ops in backend yaml require shape inference not provided by meta kernels or
+#   implementations in torch/csrc/lazy/core/shape_inference.*
+# - native function impls: errs if any 'supported' ops do not have an implementation defined in the backend
+#   (non-codegen) implementation file
+#
+#
+# About the Data Model
+# ~~~~~~~~~~~~~~~~~~~~
+#
+# Modeled after ATen codegen, the first step is to parse yaml and build a data model for the operators
+# we care about.  In this case, the <backend>_native_functions yaml defines a subset of the core operators
+# (defined in more detail in the main native_functions.yaml), which will be supported by your backend.
+# Backends can list ops in two categories:
+#  - `supported` ops require hand-implementations but still get codegenned declarations and registrations
+#  - `full_codegen` ops get implementations (and IR classes) generated too
+#
+# Each native function is modeled as an object with a schema, and each schema has objects representing their
+# arguments.  Much of the codegen is manipulation of the arguments and their types.  For example, lazy tensor
+# backends need to transform 'at::Tensor' arguments into 'lazy::Value' objects, as well as replacing reference
+# types (stringref) with actual string objects, and this is done by manipulating the data model objects.
+# - see api/lazy.py for the lazy data model
+#
+# Once the data model is set up, the rest of this script processes a number of templates for output CPP file
+# and fills in the template values using helpers in `dest/lazy_ir.py` and `dest/lazy_ts_lowering.py`.  These
+# helpers mostly iterate over functions and their arguments, outputting different c++ snippets.
+#
+# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ #
+
+
+# Parses the external backend's yaml, and adds a new BackendIndex for the backend's dispatch key.
+# Returns a Tuple of (backend_key, autograd_key, cpp_namespace, updated BackendIndex mapping, full_codegen)
+ParsedExternalYaml = namedtuple(
+    "ParsedExternalYaml",
+    ["backend_key", "autograd_key", "cpp_namespace", "backend_indices", "full_codegen"],
+)
+
+
+def parse_native_functions_keys(
+    backend_yaml_path: str,
+    grouped_native_functions: Sequence[NativeFunction | NativeFunctionsGroup],
+) -> tuple[list[OperatorName], list[Any], list[OperatorName]]:
+    with open(backend_yaml_path) as f:
+        yaml_values = yaml.load(f, Loader=YamlLoader)
+    assert isinstance(yaml_values, dict)
+
+    full_codegen = yaml_values.pop("full_codegen", [])
+    non_native = yaml_values.pop("non_native", [])
+    ir_gen = yaml_values.pop("ir_gen", [])
+    assert isinstance(full_codegen, list)
+    assert isinstance(non_native, list)
+    assert isinstance(ir_gen, list)
+    full_codegen_opnames = [OperatorName.parse(name) for name in full_codegen]
+    ir_gen_opnames = [OperatorName.parse(name) for name in ir_gen]
+    return full_codegen_opnames, non_native, ir_gen_opnames
+
+
+def validate_shape_inference_header(
+    shape_inference_hdr: str, expected_shape_infr_decls: list[str]
+) -> None:
+    try:
+        with open(shape_inference_hdr) as f:
+            shape_infr_decls = f.read()
+            shape_infr_decl_lines = set(shape_infr_decls.split("\n"))
+    except OSError as e:
+        raise AssertionError(
+            f"Unable to read from the specified shape_inference_hdr file: {shape_inference_hdr}"
+        ) from e
+
+    # TODO(whc) add a check for shape inference functions that have meta kernels implement and should be retired.
+
+    missing_decls = [
+        decl for decl in expected_shape_infr_decls if decl not in shape_infr_decl_lines
+    ]
+    if missing_decls:
+        raise Exception(  # noqa: TRY002
+            f"""Missing shape inference function.\n
+Please add declare this function in {shape_inference_hdr}:\n
+and implement it in the corresponding shape_inference.cpp file.\n
+{os.linesep.join(missing_decls)}"""
+        )
+
+
+# Some helper functions for the codegen.
+def get_ltc_helper_fns() -> str:
+    return """\
+at::Tensor to_meta(const at::Tensor& tensor) {
+  // undefined tensors can't be converted to the meta device, since they don't have sizes/strides
+  if (!tensor.defined()) return tensor;
+  auto out = at::native::empty_strided_meta_symint(tensor.sym_sizes(), tensor.sym_strides(), \
+/*dtype=*/tensor.scalar_type(), /*layout=*/tensor.layout(), \
+/*device=*/c10::Device(c10::kMeta), /*pin_memory=*/std::nullopt);
+  // needs to handle wrapped numbers, so dtype promotion works properly.
+  if (tensor.unsafeGetTensorImpl()->is_wrapped_number()) {
+    out.unsafeGetTensorImpl()->set_wrapped_number(true);
+  }
+  return out;
+}
+std::optional<at::Tensor> to_meta(const std::optional<at::Tensor>& tensor) {
+  if (tensor.has_value()) {
+    return to_meta(*tensor);
+  }
+  return std::nullopt;
+}
+
+std::vector<at::Tensor> to_meta(at::ITensorListRef t_list) {
+  std::vector<at::Tensor> outs;
+  outs.reserve(t_list.size());
+  for (const auto& tensor : t_list) {
+    outs.push_back(to_meta(tensor));
+  }
+  return outs;
+}
+"""
+
+
+class default_args:
+    node_base: str = "Node"
+    node_base_hdr: str | None = None
+    shape_inference_hdr: str = "torch/csrc/lazy/core/shape_inference.h"
+    tensor_class: str = "torch::lazy::LazyTensor"
+    tensor_class_hdr: str = "torch/csrc/lazy/core/tensor.h"
+    lazy_ir_generator: type[GenLazyIR] = GenLazyIR
+    native_func_definition_generator: type[GenLazyNativeFuncDefinition] = (
+        GenLazyNativeFuncDefinition
+    )
+    backend_name: str = "TorchScript"
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(description="Generate Lazy Tensor backend files")
+    parser.add_argument(
+        "-s",
+        "--source-yaml",
+        "--source_yaml",
+        help="path to source yaml file containing operator external definitions",
+    )
+    parser.add_argument("-o", "--output-dir", "--output_dir", help="output directory")
+    parser.add_argument(
+        "--dry-run", "--dry_run", type=bool, default=False, help="output directory"
+    )
+    parser.add_argument(
+        "--impl-path",
+        "--impl_path",
+        type=str,
+        default=None,
+        help="path to the source C++ file containing kernel definitions",
+    )
+    parser.add_argument(
+        "--gen-ts-lowerings",
+        "--gen_ts_lowerings",
+        action="store_true",
+        help="Generate TorchScript lowerings in addition to Lazy IR and NativeFunctions",
+    )
+    parser.add_argument(
+        "--node-base",
+        "--node_base",
+        type=str,
+        default=default_args.node_base,
+        help="Name of backend specific custom Lazy IR Node base class",
+    )
+    parser.add_argument(
+        "--node-base-hdr",
+        "--node_base_hdr",
+        type=str,
+        default=default_args.node_base_hdr,
+        help="Path to header file defining custom Lazy IR Node base class",
+    )
+    parser.add_argument(
+        "--shape-inference-hdr",
+        "--shape_inference_hdr",
+        type=str,
+        default=default_args.shape_inference_hdr,
+        help="Path to header file defining custom Lazy shape inference functions",
+    )
+    parser.add_argument(
+        "--tensor-class",
+        "--tensor_class",
+        type=str,
+        default=default_args.tensor_class,
+        help="Name of backend specific custom Lazy Tensor class",
+    )
+    parser.add_argument(
+        "--tensor-class-hdr",
+        "--tensor_class_hdr",
+        type=str,
+        default=default_args.tensor_class_hdr,
+        help="Path to header file defining custom Lazy Tensor class",
+    )
+    parser.add_argument(
+        "--backend-name",
+        "--backend_name",
+        type=str,
+        default=default_args.backend_name,
+        help="Name of the backend to generate",
+    )
+    options = parser.parse_args()
+
+    # Assumes that this file lives at PYTORCH_ROOT/torchgen/gen_backend_stubs.py
+    torch_root = Path(__file__).parent.parent.parent.absolute()
+    aten_path = str(torch_root / "aten" / "src" / "ATen")
+    lazy_ir_generator: type[GenLazyIR] = default_args.lazy_ir_generator
+    if options.gen_ts_lowerings:
+        lazy_ir_generator = GenTSLazyIR
+    native_func_definition_generator: type[GenLazyNativeFuncDefinition] = (
+        default_args.native_func_definition_generator
+    )
+
+    run_gen_lazy_tensor(
+        aten_path,
+        options.source_yaml,
+        options.output_dir,
+        options.dry_run,
+        options.impl_path,
+        options.node_base,
+        options.node_base_hdr,
+        options.tensor_class,
+        options.tensor_class_hdr,
+        options.shape_inference_hdr,
+        lazy_ir_generator,
+        native_func_definition_generator,
+        options.backend_name,
+    )
+
+
+def run_gen_lazy_tensor(
+    aten_path: str,
+    source_yaml: str,
+    output_dir: str,
+    dry_run: bool,
+    impl_path: str | None,
+    node_base: str = default_args.node_base,
+    node_base_hdr: str | None = default_args.node_base_hdr,
+    tensor_class: str = default_args.tensor_class,
+    tensor_class_hdr: str = default_args.tensor_class_hdr,
+    shape_inference_hdr: str = default_args.shape_inference_hdr,
+    lazy_ir_generator: type[GenLazyIR] = default_args.lazy_ir_generator,
+    native_func_definition_generator: type[
+        GenLazyNativeFuncDefinition
+    ] = default_args.native_func_definition_generator,
+    # build_in_tree is true for TS backend and affects include paths
+    build_in_tree: bool = False,
+    # per_operator_headers changes whether ATen/Functions.h or individual operator headers are used
+    # it must match how ATen was built
+    per_operator_headers: bool = False,
+    backend_name: str = default_args.backend_name,
+    gen_forced_fallback_code: bool = False,
+    use_lazy_shape: bool = True,
+    # the following arguments are temporary customization points for xla backend migration.
+    # do not rely on them otherwise, they should be removed once migration is complete
+    backend_namespace: str = "torch::lazy",
+    get_tensorlist: str = "GetTensorList",
+    get_tensor_or_wrap_number: str = "GetLtcTensorOrCreateForWrappedNumber",
+    try_get_tensor: str = "TryGetLtcTensor",
+    metrics_counter: str = 'TORCH_LAZY_FN_COUNTER("lazy::")',
+    create_tensor: str = "LazyTensor::Create",
+    create_from_first_tensor: bool = False,
+    create_aten_from_ltc_tensor: str = "torch::lazy::CreateAtenFromLtcTensor",
+    tuple_aten_from_ltc_tensors: str = "torch::lazy::TupleAtenFromLtcTensors",
+    lazy_value_class: str = "torch::lazy::Value",
+    lazy_tensor_ptr: str = "LazyTensorPtr",
+    get_device_fn: str = "torch::lazy::GetBackendDevice",
+) -> None:
+    lv_tokens = lazy_value_class.split("::")
+    lv_class = lv_tokens[-1]
+    lv_ns = "::".join(lv_tokens[:-1])
+    setValueT(BaseCppType(lv_ns, lv_class))
+    template_dir = os.path.join(aten_path, "templates")
+
+    def make_file_manager(install_dir: str) -> FileManager:
+        return FileManager(
+            install_dir=install_dir, template_dir=template_dir, dry_run=dry_run
+        )
+
+    fm = make_file_manager(output_dir)
+
+    native_yaml_path = os.path.join(aten_path, "native/native_functions.yaml")
+    tags_yaml_path = os.path.join(aten_path, "native/tags.yaml")
+    parsed_yaml = parse_native_yaml(native_yaml_path, tags_yaml_path)
+    native_functions, backend_indices = (
+        parsed_yaml.native_functions,
+        parsed_yaml.backend_indices,
+    )
+    grouped_native_functions = get_grouped_native_functions(native_functions)
+
+    def sort_native_function(f: NativeFunctionsGroup | NativeFunction) -> str:
+        """
+        We sort the native function because of the note in concat_map_codegen.
+        TODO(alanwaketan): Remove this sorting hack once all ops are grouped properly.
+        """
+        func = f.functional.func if isinstance(f, NativeFunctionsGroup) else f.func
+        return str(func.name.name)
+
+    grouped_native_functions = sorted(
+        grouped_native_functions, key=sort_native_function
+    )
+
+    parsed_backend_yaml = parse_backend_yaml(
+        source_yaml, grouped_native_functions, backend_indices
+    )
+    backend_key = parsed_backend_yaml.backend_key
+    autograd_key = parsed_backend_yaml.autograd_key
+    cpp_namespace = parsed_backend_yaml.cpp_namespace
+    backend_indices = parsed_backend_yaml.backend_indices
+    # the following 3 keys are all processed differently
+    # for full_codegen, we generate IR, kernels, etc
+    # for ir_gen, we generate only IR
+    # non_native is used to register kernels not declared in
+    # native_functions.yaml
+    full_codegen, non_native, ir_gen = parse_native_functions_keys(
+        source_yaml, grouped_native_functions
+    )
+
+    def concat_map_codegen(
+        func: Callable[[NativeFunction], Sequence[str]],
+        xs: Iterable[NativeFunctionsGroup | NativeFunction],
+        ops_list: list[OperatorName] = full_codegen,
+    ) -> Iterator[str]:
+        """
+        We code-gen for the functional variant, which is all we need for IR classes/lowerings/shape inferences, but we
+        only code-gen additional entries for the inplace variant for the native functions.
+        """
+
+        for x in xs:
+            fs = list(x.functions()) if isinstance(x, NativeFunctionsGroup) else [x]
+            for f in fs:
+                if f.func.name in ops_list:
+                    yield from func(f)
+
+    selector = SelectiveBuilder.get_nop_selector()
+
+    assert backend_key is not None
+    class_name = backend_indices[backend_key].native_function_class_name()
+
+    if impl_path is not None:
+        error_on_missing_kernels(
+            native_functions,
+            backend_indices,
+            backend_key,
+            autograd_key,
+            class_name,
+            impl_path,
+            full_codegen,
+        )
+
+    """ Validate Shape Inference Definitions
+
+    Generated lazy native functions all perform shape inference, by first using a meta:: kernel
+    if available for that op, and otherwise using a 'compute_shape_{op}' function instead.  The generator
+    knows the call signature for compute_shape_{op} because it matches the nativefunction (and meta::) signature,
+    so it just has to check whether the op is structured and generate a call for one or the other.  It's up to the dev
+    to supply the missing compute_shape_{op} function, but the codegen at least warns you about this and provides
+    the expected signature which can be copy-pasted into shape_inference.h.
+
+    compute_shape_{op} functions are handwritten and should be replaced over time as ops get ported
+    to structured kernels.
+
+    See torch/csrc/lazy/core/shape_inference.cpp #READ THIS! for more information.
+    """
+    if shape_inference_hdr is not None:
+        expected_shape_infr_decls = list(
+            concat_map_codegen(
+                dest.GenLazyShapeInferenceDefinition(
+                    backend_indices[backend_key], tensor_class
+                ),
+                grouped_native_functions,
+            )
+        )
+
+        validate_shape_inference_header(shape_inference_hdr, expected_shape_infr_decls)
+    assert class_name is not None
+
+    # Generate nativefunction declarations
+    # Note, eager registrations is set to False for the lazy TS backend as another LTC backend
+    # may want to register their own lazy kernels instead of registering the TS ones.
+    # The registration will lazily happen when init_ts_backend is called.
+    gen_dispatchkey_nativefunc_headers(
+        fm,
+        class_name,
+        cpp_namespace,
+        backend_indices,
+        grouped_native_functions,
+        backend_key,
+        autograd_key,
+        backend_name,
+    )
+
+    # Generate Dispatcher registrations which hook up the nativefunctions
+    for dispatch_key in (
+        [backend_key] if autograd_key is None else [backend_key, autograd_key]
+    ):
+        gen_dispatcher_registrations(
+            fm,
+            output_dir,
+            class_name,
+            backend_indices,
+            grouped_native_functions,
+            backend_key,
+            dispatch_key,
+            selector,
+            build_in_tree=build_in_tree,
+            per_operator_headers=per_operator_headers,
+            backend_name=backend_name,
+            eager_registration=False,
+        )
+
+    # Generate native function impls that build IR nodes
+    ns_helper = NamespaceHelper(cpp_namespace)
+    fm.write_with_template(
+        f"{backend_key}NativeFunctions.cpp",
+        "DispatchKeyNativeFunctions.cpp",
+        lambda: {
+            "includes": [
+                f"#include <{path}>"
+                for path in [
+                    tensor_class_hdr,
+                    shape_inference_hdr,
+                    "ATen/Functions.h",
+                    "ATen/native/TensorConversions.h",
+                    "ATen/NativeFunctions.h",
+                    "ATen/CompositeExplicitAutogradNonFunctionalFunctions.h",
+                    "ATen/MetaFunctions.h",
+                    "ATen/Operators.h",
+                    "ATen/native/CPUFallback.h",
+                    "torch/csrc/lazy/core/ir_builder.h",
+                    "torch/csrc/lazy/core/lazy_graph_executor.h",
+                    "torch/csrc/lazy/core/metrics.h",
+                    "torch/csrc/lazy/core/shape.h",
+                    f"{output_dir}/{backend_key}NativeFunctions.h",
+                    f"{output_dir}/LazyIr.h",
+                ]
+                + (
+                    ["torch/csrc/lazy/ts_backend/ts_eager_fallback.h"]
+                    if gen_forced_fallback_code
+                    else []
+                )
+            ],
+            "helper_fns": get_ltc_helper_fns(),
+            "native_functions_include": "",
+            "namespace_prologue": ns_helper.prologue,
+            "namespace_epilogue": ns_helper.epilogue,
+            "native_function_definitions": list(
+                concat_map_codegen(
+                    native_func_definition_generator(
+                        f"{backend_key}NativeFunctions",
+                        backend_indices[backend_key],
+                        tensor_class,
+                        gen_forced_fallback_code,
+                        backend_namespace,
+                        get_tensorlist,
+                        get_tensor_or_wrap_number,
+                        try_get_tensor,
+                        metrics_counter,
+                        create_tensor,
+                        create_from_first_tensor,
+                        create_aten_from_ltc_tensor,
+                        tuple_aten_from_ltc_tensors,
+                        lazy_tensor_ptr,
+                        get_device_fn,
+                    ),
+                    grouped_native_functions,
+                )
+            ),
+        },
+    )
+    # Generate IR node classes
+    lazy_ir_obj = lazy_ir_generator(
+        backend_indices[backend_key], backend_name, node_base, use_lazy_shape
+    )
+
+    fm.write_with_template(
+        "LazyIr.h",
+        "LazyIr.h",
+        lambda: {
+            "lazy_ir_sysinc": [
+                f"#include <{path}>"
+                for path in [
+                    "ATen/core/Formatting.h",
+                    "c10/core/ScalarType.h",
+                    "torch/csrc/lazy/core/hash.h",
+                    "torch/csrc/lazy/core/ir.h",
+                    "torch/csrc/lazy/core/shape.h",
+                    "optional",
+                    "vector",
+                ]
+            ],
+            "lazy_ir_inc": [f'#include "{node_base_hdr}"']
+            if node_base_hdr is not None
+            else [],
+            "ir_declarations": list(
+                concat_map_codegen(
+                    lazy_ir_obj, grouped_native_functions, full_codegen + ir_gen
+                )
+            ),
+            "namespace_prologue": ns_helper.prologue,
+            "namespace_epilogue": ns_helper.epilogue,
+        },
+    )
+
+    # Generate Non Native IR Node classes
+    fm.write_with_template(
+        "LazyNonNativeIr.h",
+        "LazyNonNativeIr.h",
+        lambda: {
+            "lazy_non_native_ir_inc": [
+                f"#include <{path}>"
+                for path in [
+                    "torch/csrc/lazy/core/ir.h",
+                    "torch/csrc/lazy/core/ir_builder.h",
+                    "torch/csrc/lazy/core/internal_ops/ltc_ops.h",
+                    "torch/csrc/lazy/core/shape_inference.h",
+                ]
+                + ([node_base_hdr] if node_base_hdr else [])
+                if path
+            ],
+            "non_native_ir_nodes": dest.generate_non_native_lazy_ir_nodes(
+                non_native, lazy_ir_obj
+            ),
+            "namespace_prologue": ns_helper.prologue,
+            "namespace_epilogue": ns_helper.epilogue,
+        },
+    )
+
+
+if __name__ == "__main__":
+    main()

gen_schema_utils.py

@@ -0,0 +1,97 @@
+from typing import Any, Optional, Union
+
+from torchgen.model import (
+    Annotation,
+    Argument,
+    Arguments,
+    BaseOperatorName,
+    BaseTy,
+    BaseType,
+    CustomClassType,
+    FunctionSchema,
+    ListType,
+    OperatorName,
+    Return,
+)
+
+
+# Note: These aren't actually used in torchgen, they're some utilities for generating a schema
+# from real arguments. For example, this is used to generate HigherOrderOperators' schema since
+# their schemas can vary for different instances of the same HOP.
+
+
+class TypeGen:
+    convert_to_base_ty = {
+        int: BaseTy.int,
+        float: BaseTy.float,
+        str: BaseTy.str,
+        bool: BaseTy.bool,
+    }
+
+    @staticmethod
+    def from_example(obj: Any) -> Union[BaseType, ListType, CustomClassType]:
+        import torch
+
+        if isinstance(obj, torch.fx.GraphModule):
+            return BaseType(BaseTy.GraphModule)
+        elif isinstance(obj, torch.Tensor):
+            return BaseType(BaseTy.Tensor)
+        elif isinstance(obj, torch.SymInt):
+            return BaseType(BaseTy.SymInt)
+        elif isinstance(obj, torch.SymBool):
+            return BaseType(BaseTy.SymBool)
+        elif isinstance(obj, torch.ScriptObject):
+            return CustomClassType(obj._type().name())  # type: ignore[attr-defined]
+        elif isinstance(obj, (list, tuple)):
+            assert len(obj) > 0
+            all_base_tys = [TypeGen.from_example(x) for x in obj]
+            if len(set(all_base_tys)) > 1:
+                raise RuntimeError(
+                    f"Cannot generate schema for a seqeunce of args of heterogeneous types: {all_base_tys}. "
+                    "Consider unpacking the argument and give proper names to them if possible "
+                    "instead of using *args."
+                )
+            return ListType(all_base_tys[0], len(obj))
+        tp = type(obj)
+        if tp not in TypeGen.convert_to_base_ty:
+            raise RuntimeError(f"unsupported type {tp}")
+        return BaseType(TypeGen.convert_to_base_ty[tp])
+
+
+class ReturnGen:
+    @staticmethod
+    def from_example(
+        name: Optional[str], obj: Any, annotation: Optional[Annotation]
+    ) -> Return:
+        return Return(name, TypeGen.from_example(obj), annotation)
+
+
+class ArgumentGen:
+    @staticmethod
+    def from_example(
+        name: str, obj: Any, default: Optional[str], annotation: Optional[Annotation]
+    ) -> Argument:
+        return Argument(
+            name, TypeGen.from_example(obj), default=default, annotation=annotation
+        )
+
+
+class FunctionSchemaGen:
+    @staticmethod
+    def from_example(
+        op_name: str,
+        example_inputs: tuple[tuple[str, Any], ...],
+        example_outputs: tuple[Any, ...],
+    ) -> FunctionSchema:
+        args = []
+        for name, inp in example_inputs:
+            args.append(ArgumentGen.from_example(name, inp, None, None))
+        # ignore the annotations and other attributes for now, we could add more when needed.
+        arguments = Arguments(
+            tuple(), None, tuple(args), tuple(), None, tuple(), tuple()
+        )
+        returns = tuple(
+            ReturnGen.from_example(None, out, None) for out in example_outputs
+        )
+        op_name = OperatorName(BaseOperatorName(op_name, False, False, False), "")
+        return FunctionSchema(op_name, arguments, returns)

gen_vmap_plumbing.py

@@ -0,0 +1,275 @@
+from __future__ import annotations
+
+import textwrap
+from dataclasses import dataclass
+from typing import TYPE_CHECKING
+
+from torchgen.api.translate import translate
+from torchgen.api.types import DispatcherSignature
+from torchgen.context import method_with_native_function
+from torchgen.model import (
+    Argument,
+    BaseTy,
+    BaseType,
+    FunctionSchema,
+    ListType,
+    NativeFunction,
+    OptionalType,
+    Return,
+    SchemaKind,
+    Type,
+)
+from torchgen.utils import mapMaybe
+
+
+if TYPE_CHECKING:
+    from collections.abc import Sequence
+
+
+def is_tensor(typ: Type) -> bool:
+    return isinstance(typ, BaseType) and typ.name == BaseTy.Tensor
+
+
+def is_optional_tensor(typ: Type) -> bool:
+    return isinstance(typ, OptionalType) and is_tensor(typ.elem)
+
+
+def is_tensor_list(typ: Type) -> bool:
+    return isinstance(typ, ListType) and is_tensor(typ.elem)
+
+
+def unwrap_tensor(name: str, cur_level_var: str) -> list[str]:
+    result = f"""\
+    auto [{name}_value, {name}_bdim] = unwrapTensorAtLevel({name}, {cur_level_var});"""
+    return textwrap.dedent(result).split("\n")
+
+
+def unwrap_optional_tensor(name: str, cur_level_var: str) -> list[str]:
+    result = f"""\
+    std::optional<Tensor> {name}_value;
+    std::optional<int64_t> {name}_bdim;
+    if ({name}) {{
+        std::tie({name}_value, {name}_bdim) = unwrapTensorAtLevel({name}.value(), {cur_level_var});
+    }}"""
+    return textwrap.dedent(result).split("\n")
+
+
+def gen_unwraps(
+    flat_arguments: Sequence[Argument], cur_level_var: str
+) -> tuple[str, list[str]]:
+    arg_names = [a.name for a in flat_arguments]
+    arg_types = [a.type for a in flat_arguments]
+
+    tensors = [name for typ, name in zip(arg_types, arg_names) if is_tensor(typ)]
+    optional_tensors = [
+        name for typ, name in zip(arg_types, arg_names) if is_optional_tensor(typ)
+    ]
+
+    unwraps = []
+    for tensor in tensors:
+        unwraps += unwrap_tensor(tensor, cur_level_var)
+
+    for opt_tensor in optional_tensors:
+        unwraps += unwrap_optional_tensor(opt_tensor, cur_level_var)
+    unwrap_code = "\n".join(unwraps)
+
+    unwrapped_arg_list = []
+    for arg in arg_names:
+        if arg in tensors or arg in optional_tensors:
+            unwrapped_arg_list += [f"{arg}_value", f"{arg}_bdim"]
+        else:
+            unwrapped_arg_list.append(arg)
+    return unwrap_code, unwrapped_arg_list
+
+
+def gen_case_where_all_bdims_are_none(
+    outer_sig: DispatcherSignature, schema: FunctionSchema, cur_level_var: str
+) -> str:
+    conditions = []
+    flat_args = schema.arguments.flat_all
+    for arg in flat_args:
+        if not arg.type.is_tensor_like():
+            continue
+        conditions.append(f"!isBatchedAtLevel({arg.name}, {cur_level_var})")
+
+    sig = DispatcherSignature.from_schema(schema)
+    translated_args = ", ".join(
+        e.expr for e in translate(outer_sig.arguments(), sig.arguments())
+    )
+    return f"""\
+if ({' && '.join(conditions)}) {{
+  return at::_ops::{sig.func.name.unambiguous_name()}::call({translated_args});
+}}"""
+
+
+def gen_returns(
+    returns: tuple[Return, ...], cur_level_var: str, results_var: str
+) -> str:
+    idx = 0
+    wrapped_returns = []
+    for ret in returns:
+        if is_tensor(ret.type):
+            wrapped_returns.append(
+                f"makeBatched(std::get<{idx}>({results_var}), std::get<{idx + 1}>({results_var}), {cur_level_var})"
+            )
+            idx += 2
+        elif is_tensor_list(ret.type):
+            wrapped_returns.append(
+                f"makeBatchedVector(std::get<{idx}>({results_var}), std::get<{idx + 1}>({results_var}), {cur_level_var})"
+            )
+            idx += 2
+        else:
+            wrapped_returns.append(f"std::get<{idx}>({results_var})")
+            idx += 1
+    if len(wrapped_returns) == 1:
+        result = f"return {wrapped_returns[0]};"
+    else:
+        result = f'return std::make_tuple({", ".join(wrapped_returns)});'
+    return result
+
+
+def accepts_at_least_one_tensor_input(schema: FunctionSchema) -> bool:
+    return any(a.type.is_tensor_like() for a in schema.arguments.flat_all)
+
+
+def is_mutated_arg(argument: Argument) -> bool:
+    return argument.annotation is not None and argument.annotation.is_write
+
+
+def gen_vmap_inplace_plumbing(native_function: NativeFunction) -> str | None:
+    # Assumptions:
+    # - only one argument is being modified in-place
+    # - the argument that is being modified in-place is the first argument
+    # - all returns are either Tensor, tuple of Tensor, or TensorList
+    schema = native_function.func
+    sig = DispatcherSignature.from_schema(schema)
+    returns = schema.returns
+
+    # Check assumptions. If these are invalid we return None
+    # and punt the work to handle them to the future.
+    assert schema.kind() == SchemaKind.inplace
+    if not is_mutated_arg(schema.arguments.flat_all[0]):
+        return None
+    if not len([arg for arg in schema.arguments.flat_all if is_mutated_arg(arg)]) == 1:
+        return None
+
+    # Only support cases where all returns are Tensors or vector<Tensor>
+    if len(returns) == 0:
+        return None
+    if not all(is_tensor(ret.type) or is_tensor_list(ret.type) for ret in returns):
+        return None
+    if not accepts_at_least_one_tensor_input(schema):
+        return None
+
+    cur_level_var = "cur_level"
+
+    unwraps, unwrapped_arg_list = gen_unwraps(schema.arguments.flat_all, cur_level_var)
+    bdims_all_none_case = gen_case_where_all_bdims_are_none(sig, schema, cur_level_var)
+
+    return f"""\
+template <typename batch_rule_t, batch_rule_t batch_rule>
+{sig.decl(name=schema.name.unambiguous_name() + '_generated_plumbing')} {{
+  c10::impl::ExcludeDispatchKeyGuard guard(DispatchKey::FuncTorchBatched);
+  auto maybe_layer = maybeCurrentDynamicLayer();
+  vmap_check_escaped(maybe_layer, "gen_vmap_inplace_plumbing");
+  int64_t {cur_level_var} = maybe_layer->layerId();
+{textwrap.indent(bdims_all_none_case, "  ")}
+{textwrap.indent(unwraps, "  ")}
+  batch_rule({', '.join(unwrapped_arg_list)});
+  return {schema.arguments.flat_all[0].name};
+}}"""
+
+
+def gen_vmap_plumbing_no_returns(native_function: NativeFunction) -> str:
+    schema = native_function.func
+    sig = DispatcherSignature.from_schema(schema)
+    cur_level_var = "cur_level"
+
+    unwraps, unwrapped_arg_list = gen_unwraps(schema.arguments.flat_all, cur_level_var)
+    bdims_all_none_case = gen_case_where_all_bdims_are_none(sig, schema, cur_level_var)
+
+    return f"""\
+template <typename batch_rule_t, batch_rule_t batch_rule>
+{sig.decl(name=schema.name.unambiguous_name() + '_generated_plumbing')} {{
+  c10::impl::ExcludeDispatchKeyGuard guard(DispatchKey::FuncTorchBatched);
+  auto maybe_layer = maybeCurrentDynamicLayer();
+  vmap_check_escaped(maybe_layer, "gen_vmap_plumbing_no_returns");
+  int64_t {cur_level_var} = maybe_layer->layerId();
+{textwrap.indent(bdims_all_none_case, "  ")}
+{textwrap.indent(unwraps, "  ")}
+  batch_rule({', '.join(unwrapped_arg_list)});
+}}"""
+
+
+def gen_vmap_plumbing(native_function: NativeFunction) -> str | None:
+    schema = native_function.func
+    sig = DispatcherSignature.from_schema(schema)
+    returns = schema.returns
+
+    # Only support cases where all returns are Tensors or vector<Tensor>
+    if not accepts_at_least_one_tensor_input(schema):
+        return None
+    if len(returns) == 0:
+        return gen_vmap_plumbing_no_returns(native_function)
+    return_symint_overrides = [
+        "_scaled_dot_product_flash_attention",
+        "_scaled_dot_product_cudnn_attention",
+    ]
+    if (
+        not all(ret.type.is_tensor_like() for ret in returns)
+        and schema.name.unambiguous_name() not in return_symint_overrides
+    ):
+        return None
+    # in-place views need special handling
+    if "inplace_view" in native_function.tags:
+        return None
+
+    if schema.kind() == SchemaKind.inplace:
+        return gen_vmap_inplace_plumbing(native_function)
+
+    # Don't support these (mutable, out, scratch)
+    if schema.kind() != SchemaKind.functional:
+        return None
+
+    results_var = "results"
+    cur_level_var = "cur_level"
+
+    unwraps, unwrapped_arg_list = gen_unwraps(schema.arguments.flat_all, cur_level_var)
+    bdims_all_none_case = gen_case_where_all_bdims_are_none(sig, schema, cur_level_var)
+
+    wrapped_returns = gen_returns(returns, cur_level_var, results_var)
+    return f"""\
+template <typename batch_rule_t, batch_rule_t batch_rule>
+{sig.decl(name=schema.name.unambiguous_name() + '_generated_plumbing')} {{
+  c10::impl::ExcludeDispatchKeyGuard guard(DispatchKey::FuncTorchBatched);
+  auto maybe_layer = maybeCurrentDynamicLayer();
+  vmap_check_escaped(maybe_layer, "gen_vmap_plumbing");
+  int64_t {cur_level_var} = maybe_layer->layerId();
+{textwrap.indent(bdims_all_none_case, "  ")}
+{textwrap.indent(unwraps, "  ")}
+  auto {results_var} = batch_rule({', '.join(unwrapped_arg_list)});
+  {wrapped_returns}
+}}"""
+
+
+@dataclass(frozen=True)
+class ComputeBatchRulePlumbing:
+    @method_with_native_function
+    def __call__(self, f: NativeFunction) -> str | None:
+        result = gen_vmap_plumbing(f)
+        return result
+
+
+def gen_all_vmap_plumbing(native_functions: Sequence[NativeFunction]) -> str:
+    body = "\n".join(list(mapMaybe(ComputeBatchRulePlumbing(), native_functions)))
+    return f"""
+#pragma once
+#include <ATen/Operators.h>
+#include <ATen/functorch/PlumbingHelper.h>
+
+namespace at {{ namespace functorch {{
+
+{body}
+
+}}}} // namespace at::functorch
+"""

get_gprof

@@ -0,0 +1,75 @@
+#!C:\Users\migue\Downloads\Codigo\AI\HuggingModelAI\IAservicies\myenv\Scripts\python.exe
+#
+# Author: Mike McKerns (mmckerns @caltech and @uqfoundation)
+# Copyright (c) 2008-2016 California Institute of Technology.
+# Copyright (c) 2016-2024 The Uncertainty Quantification Foundation.
+# License: 3-clause BSD.  The full license text is available at:
+#  - https://github.com/uqfoundation/dill/blob/master/LICENSE
+'''
+build profile graph for the given instance
+
+running:
+  $ get_gprof <args> <instance>
+
+executes:
+  gprof2dot -f pstats <args> <type>.prof | dot -Tpng -o <type>.call.png
+
+where:
+  <args> are arguments for gprof2dot, such as "-n 5 -e 5"
+  <instance> is code to create the instance to profile
+  <type> is the class of the instance (i.e. type(instance))
+
+For example:
+  $ get_gprof -n 5 -e 1 "import numpy; numpy.array([1,2])"
+
+will create 'ndarray.call.png' with the profile graph for numpy.array([1,2]),
+where '-n 5' eliminates nodes below 5% threshold, similarly '-e 1' eliminates
+edges below 1% threshold
+'''
+
+if __name__ == "__main__":
+    import sys
+    if len(sys.argv) < 2:
+        print ("Please provide an object instance (e.g. 'import math; math.pi')")
+        sys.exit()
+    # grab args for gprof2dot
+    args = sys.argv[1:-1]
+    args = ' '.join(args)
+    # last arg builds the object
+    obj = sys.argv[-1]
+    obj = obj.split(';')
+    # multi-line prep for generating an instance
+    for line in obj[:-1]:
+        exec(line)
+    # one-line generation of an instance
+    try:
+        obj = eval(obj[-1])
+    except Exception:
+        print ("Error processing object instance")
+        sys.exit()
+
+    # get object 'name'
+    objtype = type(obj)
+    name = getattr(objtype, '__name__', getattr(objtype, '__class__', objtype))
+
+    # profile dumping an object
+    import dill
+    import os
+    import cProfile
+    #name = os.path.splitext(os.path.basename(__file__))[0]
+    cProfile.run("dill.dumps(obj)", filename="%s.prof" % name)
+    msg = "gprof2dot -f pstats %s %s.prof | dot -Tpng -o %s.call.png" % (args, name, name)
+    try:
+        res = os.system(msg)
+    except Exception:
+        print ("Please verify install of 'gprof2dot' to view profile graphs")
+    if res:
+        print ("Please verify install of 'gprof2dot' to view profile graphs")
+
+    # get stats
+    f_prof = "%s.prof" % name
+    import pstats
+    stats = pstats.Stats(f_prof, stream=sys.stdout)
+    stats.strip_dirs().sort_stats('cumtime')
+    stats.print_stats(20) #XXX: save to file instead of print top 20?
+    os.remove(f_prof)

get_objgraph

@@ -0,0 +1,54 @@
+#!C:\Users\migue\Downloads\Codigo\AI\HuggingModelAI\IAservicies\myenv\Scripts\python.exe
+#
+# Author: Mike McKerns (mmckerns @caltech and @uqfoundation)
+# Copyright (c) 2008-2016 California Institute of Technology.
+# Copyright (c) 2016-2024 The Uncertainty Quantification Foundation.
+# License: 3-clause BSD.  The full license text is available at:
+#  - https://github.com/uqfoundation/dill/blob/master/LICENSE
+"""
+display the reference paths for objects in ``dill.types`` or a .pkl file
+
+Notes:
+    the generated image is useful in showing the pointer references in
+    objects that are or can be pickled.  Any object in ``dill.objects``
+    listed in ``dill.load_types(picklable=True, unpicklable=True)`` works.
+
+Examples::
+
+    $ get_objgraph ArrayType
+    Image generated as ArrayType.png
+"""
+
+import dill as pickle
+#pickle.debug.trace(True)
+#import pickle
+
+# get all objects for testing
+from dill import load_types
+load_types(pickleable=True,unpickleable=True)
+from dill import objects
+
+if __name__ == "__main__":
+    import sys
+    if len(sys.argv) != 2:
+        print ("Please provide exactly one file or type name (e.g. 'IntType')")
+        msg = "\n"
+        for objtype in list(objects.keys())[:40]:
+            msg += objtype + ', '
+        print (msg + "...")
+    else:
+        objtype = str(sys.argv[-1])
+        try:
+            obj = objects[objtype]
+        except KeyError:
+            obj = pickle.load(open(objtype,'rb'))
+            import os
+            objtype = os.path.splitext(objtype)[0]
+        try:
+            import objgraph
+            objgraph.show_refs(obj, filename=objtype+'.png')
+        except ImportError:
+            print ("Please install 'objgraph' to view object graphs")
+
+
+# EOF

huggingface-cli.exe

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c574193300d5f23639772059c556e84d222138dd251cf9ddee95eb5a4663f37c
+size 108461

import_pb_to_tensorboard.exe

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5c8bab1d9edcde78e2e8158df29baa25eff6ed268a881514bb24f6b6a52849fd
+size 108471

isympy.1

@@ -0,0 +1,188 @@
+'\" -*- coding: us-ascii -*-
+.if \n(.g .ds T< \\FC
+.if \n(.g .ds T> \\F[\n[.fam]]
+.de URL
+\\$2 \(la\\$1\(ra\\$3
+..
+.if \n(.g .mso www.tmac
+.TH isympy 1 2007-10-8 "" ""
+.SH NAME
+isympy \- interactive shell for SymPy
+.SH SYNOPSIS
+'nh
+.fi
+.ad l
+\fBisympy\fR \kx
+.if (\nx>(\n(.l/2)) .nr x (\n(.l/5)
+'in \n(.iu+\nxu
+[\fB-c\fR | \fB--console\fR] [\fB-p\fR ENCODING | \fB--pretty\fR ENCODING] [\fB-t\fR TYPE | \fB--types\fR TYPE] [\fB-o\fR ORDER | \fB--order\fR ORDER] [\fB-q\fR | \fB--quiet\fR] [\fB-d\fR | \fB--doctest\fR] [\fB-C\fR | \fB--no-cache\fR] [\fB-a\fR | \fB--auto\fR] [\fB-D\fR | \fB--debug\fR] [
+-- | PYTHONOPTIONS]
+'in \n(.iu-\nxu
+.ad b
+'hy
+'nh
+.fi
+.ad l
+\fBisympy\fR \kx
+.if (\nx>(\n(.l/2)) .nr x (\n(.l/5)
+'in \n(.iu+\nxu
+[
+{\fB-h\fR | \fB--help\fR}
+|
+{\fB-v\fR | \fB--version\fR}
+]
+'in \n(.iu-\nxu
+.ad b
+'hy
+.SH DESCRIPTION
+isympy is a Python shell for SymPy. It is just a normal python shell
+(ipython shell if you have the ipython package installed) that executes
+the following commands so that you don't have to:
+.PP
+.nf
+\*(T<
+>>> from __future__ import division
+>>> from sympy import *
+>>> x, y, z = symbols("x,y,z")
+>>> k, m, n = symbols("k,m,n", integer=True)
+    \*(T>
+.fi
+.PP
+So starting isympy is equivalent to starting python (or ipython) and
+executing the above commands by hand. It is intended for easy and quick
+experimentation with SymPy. For more complicated programs, it is recommended
+to write a script and import things explicitly (using the "from sympy
+import sin, log, Symbol, ..." idiom).
+.SH OPTIONS
+.TP
+\*(T<\fB\-c \fR\*(T>\fISHELL\fR, \*(T<\fB\-\-console=\fR\*(T>\fISHELL\fR
+Use the specified shell (python or ipython) as
+console backend instead of the default one (ipython
+if present or python otherwise).
+
+Example: isympy -c python
+
+\fISHELL\fR could be either
+\&'ipython' or 'python'
+.TP
+\*(T<\fB\-p \fR\*(T>\fIENCODING\fR, \*(T<\fB\-\-pretty=\fR\*(T>\fIENCODING\fR
+Setup pretty printing in SymPy. By default, the most pretty, unicode
+printing is enabled (if the terminal supports it). You can use less
+pretty ASCII printing instead or no pretty printing at all.
+
+Example: isympy -p no
+
+\fIENCODING\fR must be one of 'unicode',
+\&'ascii' or 'no'.
+.TP
+\*(T<\fB\-t \fR\*(T>\fITYPE\fR, \*(T<\fB\-\-types=\fR\*(T>\fITYPE\fR
+Setup the ground types for the polys. By default, gmpy ground types
+are used if gmpy2 or gmpy is installed, otherwise it falls back to python
+ground types, which are a little bit slower. You can manually
+choose python ground types even if gmpy is installed (e.g., for testing purposes).
+
+Note that sympy ground types are not supported, and should be used
+only for experimental purposes.
+
+Note that the gmpy1 ground type is primarily intended for testing; it the
+use of gmpy even if gmpy2 is available.
+
+This is the same as setting the environment variable
+SYMPY_GROUND_TYPES to the given ground type (e.g.,
+SYMPY_GROUND_TYPES='gmpy')
+
+The ground types can be determined interactively from the variable
+sympy.polys.domains.GROUND_TYPES inside the isympy shell itself.
+
+Example: isympy -t python
+
+\fITYPE\fR must be one of 'gmpy',
+\&'gmpy1' or 'python'.
+.TP
+\*(T<\fB\-o \fR\*(T>\fIORDER\fR, \*(T<\fB\-\-order=\fR\*(T>\fIORDER\fR
+Setup the ordering of terms for printing. The default is lex, which
+orders terms lexicographically (e.g., x**2 + x + 1). You can choose
+other orderings, such as rev-lex, which will use reverse
+lexicographic ordering (e.g., 1 + x + x**2).
+
+Note that for very large expressions, ORDER='none' may speed up
+printing considerably, with the tradeoff that the order of the terms
+in the printed expression will have no canonical order
+
+Example: isympy -o rev-lax
+
+\fIORDER\fR must be one of 'lex', 'rev-lex', 'grlex',
+\&'rev-grlex', 'grevlex', 'rev-grevlex', 'old', or 'none'.
+.TP
+\*(T<\fB\-q\fR\*(T>, \*(T<\fB\-\-quiet\fR\*(T>
+Print only Python's and SymPy's versions to stdout at startup, and nothing else.
+.TP
+\*(T<\fB\-d\fR\*(T>, \*(T<\fB\-\-doctest\fR\*(T>
+Use the same format that should be used for doctests. This is
+equivalent to '\fIisympy -c python -p no\fR'.
+.TP
+\*(T<\fB\-C\fR\*(T>, \*(T<\fB\-\-no\-cache\fR\*(T>
+Disable the caching mechanism. Disabling the cache may slow certain
+operations down considerably. This is useful for testing the cache,
+or for benchmarking, as the cache can result in deceptive benchmark timings.
+
+This is the same as setting the environment variable SYMPY_USE_CACHE
+to 'no'.
+.TP
+\*(T<\fB\-a\fR\*(T>, \*(T<\fB\-\-auto\fR\*(T>
+Automatically create missing symbols. Normally, typing a name of a
+Symbol that has not been instantiated first would raise NameError,
+but with this option enabled, any undefined name will be
+automatically created as a Symbol. This only works in IPython 0.11.
+
+Note that this is intended only for interactive, calculator style
+usage. In a script that uses SymPy, Symbols should be instantiated
+at the top, so that it's clear what they are.
+
+This will not override any names that are already defined, which
+includes the single character letters represented by the mnemonic
+QCOSINE (see the "Gotchas and Pitfalls" document in the
+documentation). You can delete existing names by executing "del
+name" in the shell itself. You can see if a name is defined by typing
+"'name' in globals()".
+
+The Symbols that are created using this have default assumptions.
+If you want to place assumptions on symbols, you should create them
+using symbols() or var().
+
+Finally, this only works in the top level namespace. So, for
+example, if you define a function in isympy with an undefined
+Symbol, it will not work.
+.TP
+\*(T<\fB\-D\fR\*(T>, \*(T<\fB\-\-debug\fR\*(T>
+Enable debugging output. This is the same as setting the
+environment variable SYMPY_DEBUG to 'True'. The debug status is set
+in the variable SYMPY_DEBUG within isympy.
+.TP
+-- \fIPYTHONOPTIONS\fR
+These options will be passed on to \fIipython (1)\fR shell.
+Only supported when ipython is being used (standard python shell not supported).
+
+Two dashes (--) are required to separate \fIPYTHONOPTIONS\fR
+from the other isympy options.
+
+For example, to run iSymPy without startup banner and colors:
+
+isympy -q -c ipython -- --colors=NoColor
+.TP
+\*(T<\fB\-h\fR\*(T>, \*(T<\fB\-\-help\fR\*(T>
+Print help output and exit.
+.TP
+\*(T<\fB\-v\fR\*(T>, \*(T<\fB\-\-version\fR\*(T>
+Print isympy version information and exit.
+.SH FILES
+.TP
+\*(T<\fI${HOME}/.sympy\-history\fR\*(T>
+Saves the history of commands when using the python
+shell as backend.
+.SH BUGS
+The upstreams BTS can be found at \(lahttps://github.com/sympy/sympy/issues\(ra
+Please report all bugs that you find in there, this will help improve
+the overall quality of SymPy.
+.SH "SEE ALSO"
+\fBipython\fR(1), \fBpython\fR(1)

isympy.exe

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c5b00e29646484644b74c45d8d9469b294cbe12de1ec5a9f23483ad98cf8ee2b
+size 108427

isympy.py

@@ -0,0 +1,342 @@
+"""
+Python shell for SymPy.
+
+This is just a normal Python shell (IPython shell if you have the
+IPython package installed), that executes the following commands for
+the user:
+
+    >>> from __future__ import division
+    >>> from sympy import *
+    >>> x, y, z, t = symbols('x y z t')
+    >>> k, m, n = symbols('k m n', integer=True)
+    >>> f, g, h = symbols('f g h', cls=Function)
+    >>> init_printing()
+
+So starting 'isympy' is equivalent to starting Python (or IPython) and
+executing the above commands by hand.  It is intended for easy and quick
+experimentation with SymPy.  isympy is a good way to use SymPy as an
+interactive calculator. If you have IPython and Matplotlib installed, then
+interactive plotting is enabled by default.
+
+COMMAND LINE OPTIONS
+--------------------
+
+-c CONSOLE, --console=CONSOLE
+
+     Use the specified shell (Python or IPython) shell as the console
+     backend instead of the default one (IPython if present, Python
+     otherwise), e.g.:
+
+        $isympy -c python
+
+    CONSOLE must be one of 'ipython' or 'python'
+
+-p PRETTY, --pretty PRETTY
+
+    Setup pretty-printing in SymPy. When pretty-printing is enabled,
+    expressions can be printed with Unicode or ASCII. The default is
+    to use pretty-printing (with Unicode if the terminal supports it).
+    When this option is 'no', expressions will not be pretty-printed
+    and ASCII will be used:
+
+        $isympy -p no
+
+    PRETTY must be one of 'unicode', 'ascii', or 'no'
+
+-t TYPES, --types=TYPES
+
+    Setup the ground types for the polys.  By default, gmpy ground types
+    are used if gmpy2 or gmpy is installed, otherwise it falls back to python
+    ground types, which are a little bit slower.  You can manually
+    choose python ground types even if gmpy is installed (e.g., for
+    testing purposes):
+
+        $isympy -t python
+
+    TYPES must be one of 'gmpy', 'gmpy1' or 'python'
+
+    Note that the ground type gmpy1 is primarily intended for testing; it
+    forces the use of gmpy version 1 even if gmpy2 is available.
+
+    This is the same as setting the environment variable
+    SYMPY_GROUND_TYPES to the given ground type (e.g.,
+    SYMPY_GROUND_TYPES='gmpy')
+
+    The ground types can be determined interactively from the variable
+    sympy.polys.domains.GROUND_TYPES.
+
+-o ORDER, --order ORDER
+
+    Setup the ordering of terms for printing.  The default is lex, which
+    orders terms lexicographically (e.g., x**2 + x + 1). You can choose
+    other orderings, such as rev-lex, which will use reverse
+    lexicographic ordering (e.g., 1 + x + x**2):
+
+        $isympy -o rev-lex
+
+    ORDER must be one of 'lex', 'rev-lex', 'grlex', 'rev-grlex',
+    'grevlex', 'rev-grevlex', 'old', or 'none'.
+
+    Note that for very large expressions, ORDER='none' may speed up
+    printing considerably but the terms will have no canonical order.
+
+-q, --quiet
+
+    Print only Python's and SymPy's versions to stdout at startup.
+
+-d, --doctest
+
+    Use the same format that should be used for doctests.  This is
+    equivalent to -c python -p no.
+
+-C, --no-cache
+
+    Disable the caching mechanism.  Disabling the cache may slow certain
+    operations down considerably.  This is useful for testing the cache,
+    or for benchmarking, as the cache can result in deceptive timings.
+
+    This is equivalent to setting the environment variable
+    SYMPY_USE_CACHE to 'no'.
+
+-a, --auto-symbols (requires at least IPython 0.11)
+
+    Automatically create missing symbols.  Normally, typing a name of a
+    Symbol that has not been instantiated first would raise NameError,
+    but with this option enabled, any undefined name will be
+    automatically created as a Symbol.
+
+    Note that this is intended only for interactive, calculator style
+    usage. In a script that uses SymPy, Symbols should be instantiated
+    at the top, so that it's clear what they are.
+
+    This will not override any names that are already defined, which
+    includes the single character letters represented by the mnemonic
+    QCOSINE (see the "Gotchas and Pitfalls" document in the
+    documentation). You can delete existing names by executing "del
+    name".  If a name is defined, typing "'name' in dir()" will return True.
+
+    The Symbols that are created using this have default assumptions.
+    If you want to place assumptions on symbols, you should create them
+    using symbols() or var().
+
+    Finally, this only works in the top level namespace. So, for
+    example, if you define a function in isympy with an undefined
+    Symbol, it will not work.
+
+    See also the -i and -I options.
+
+-i, --int-to-Integer (requires at least IPython 0.11)
+
+    Automatically wrap int literals with Integer.  This makes it so that
+    things like 1/2 will come out as Rational(1, 2), rather than 0.5.  This
+    works by preprocessing the source and wrapping all int literals with
+    Integer.  Note that this will not change the behavior of int literals
+    assigned to variables, and it also won't change the behavior of functions
+    that return int literals.
+
+    If you want an int, you can wrap the literal in int(), e.g. int(3)/int(2)
+    gives 1.5 (with division imported from __future__).
+
+-I, --interactive (requires at least IPython 0.11)
+
+    This is equivalent to --auto-symbols --int-to-Integer.  Future options
+    designed for ease of interactive use may be added to this.
+
+-D, --debug
+
+    Enable debugging output.  This is the same as setting the
+    environment variable SYMPY_DEBUG to 'True'.  The debug status is set
+    in the variable SYMPY_DEBUG within isympy.
+
+-- IPython options
+
+    Additionally you can pass command line options directly to the IPython
+    interpreter (the standard Python shell is not supported).  However you
+    need to add the '--' separator between two types of options, e.g the
+    startup banner option and the colors option. You need to enter the
+    options as required by the version of IPython that you are using, too:
+
+    in IPython 0.11,
+
+        $isympy -q -- --colors=NoColor
+
+    or older versions of IPython,
+
+        $isympy -q -- -colors NoColor
+
+See also isympy --help.
+"""
+
+import os
+import sys
+
+# DO NOT IMPORT SYMPY HERE! Or the setting of the sympy environment variables
+# by the command line will break.
+
+def main() -> None:
+    from argparse import ArgumentParser, RawDescriptionHelpFormatter
+
+    VERSION = None
+    if '--version' in sys.argv:
+        # We cannot import sympy before this is run, because flags like -C and
+        # -t set environment variables that must be set before SymPy is
+        # imported. The only thing we need to import it for is to get the
+        # version, which only matters with the --version flag.
+        import sympy
+        VERSION = sympy.__version__
+
+    usage = 'isympy [options] -- [ipython options]'
+    parser = ArgumentParser(
+        usage=usage,
+        description=__doc__,
+        formatter_class=RawDescriptionHelpFormatter,
+    )
+
+    parser.add_argument('--version', action='version', version=VERSION)
+
+    parser.add_argument(
+        '-c', '--console',
+        dest='console',
+        action='store',
+        default=None,
+        choices=['ipython', 'python'],
+        metavar='CONSOLE',
+        help='select type of interactive session: ipython | python; defaults '
+        'to ipython if IPython is installed, otherwise python')
+
+    parser.add_argument(
+        '-p', '--pretty',
+        dest='pretty',
+        action='store',
+        default=None,
+        metavar='PRETTY',
+        choices=['unicode', 'ascii', 'no'],
+        help='setup pretty printing: unicode | ascii | no; defaults to '
+        'unicode printing if the terminal supports it, otherwise ascii')
+
+    parser.add_argument(
+        '-t', '--types',
+        dest='types',
+        action='store',
+        default=None,
+        metavar='TYPES',
+        choices=['gmpy', 'gmpy1', 'python'],
+        help='setup ground types: gmpy | gmpy1 | python; defaults to gmpy if gmpy2 '
+        'or gmpy is installed, otherwise python')
+
+    parser.add_argument(
+        '-o', '--order',
+        dest='order',
+        action='store',
+        default=None,
+        metavar='ORDER',
+        choices=['lex', 'grlex', 'grevlex', 'rev-lex', 'rev-grlex', 'rev-grevlex', 'old', 'none'],
+        help='setup ordering of terms: [rev-]lex | [rev-]grlex | [rev-]grevlex | old | none; defaults to lex')
+
+    parser.add_argument(
+        '-q', '--quiet',
+        dest='quiet',
+        action='store_true',
+        default=False,
+        help='print only version information at startup')
+
+    parser.add_argument(
+        '-d', '--doctest',
+        dest='doctest',
+        action='store_true',
+        default=False,
+        help='use the doctest format for output (you can just copy and paste it)')
+
+    parser.add_argument(
+        '-C', '--no-cache',
+        dest='cache',
+        action='store_false',
+        default=True,
+        help='disable caching mechanism')
+
+    parser.add_argument(
+        '-a', '--auto-symbols',
+        dest='auto_symbols',
+        action='store_true',
+        default=False,
+        help='automatically construct missing symbols')
+
+    parser.add_argument(
+        '-i', '--int-to-Integer',
+        dest='auto_int_to_Integer',
+        action='store_true',
+        default=False,
+        help="automatically wrap int literals with Integer")
+
+    parser.add_argument(
+        '-I', '--interactive',
+        dest='interactive',
+        action='store_true',
+        default=False,
+        help="equivalent to -a -i")
+
+    parser.add_argument(
+        '-D', '--debug',
+        dest='debug',
+        action='store_true',
+        default=False,
+        help='enable debugging output')
+
+    (options, ipy_args) = parser.parse_known_args()
+    if '--' in ipy_args:
+        ipy_args.remove('--')
+
+    if not options.cache:
+        os.environ['SYMPY_USE_CACHE'] = 'no'
+
+    if options.types:
+        os.environ['SYMPY_GROUND_TYPES'] = options.types
+
+    if options.debug:
+        os.environ['SYMPY_DEBUG'] = str(options.debug)
+
+    if options.doctest:
+        options.pretty = 'no'
+        options.console = 'python'
+
+    session = options.console
+
+    if session is not None:
+        ipython = session == 'ipython'
+    else:
+        try:
+            import IPython
+            ipython = True
+        except ImportError:
+            if not options.quiet:
+                from sympy.interactive.session import no_ipython
+                print(no_ipython)
+            ipython = False
+
+    args = {
+        'pretty_print': True,
+        'use_unicode':  None,
+        'use_latex':    None,
+        'order':        None,
+        'argv':         ipy_args,
+    }
+
+    if options.pretty == 'unicode':
+        args['use_unicode'] = True
+    elif options.pretty == 'ascii':
+        args['use_unicode'] = False
+    elif options.pretty == 'no':
+        args['pretty_print'] = False
+
+    if options.order is not None:
+        args['order'] = options.order
+
+    args['quiet'] = options.quiet
+    args['auto_symbols'] = options.auto_symbols or options.interactive
+    args['auto_int_to_Integer'] = options.auto_int_to_Integer or options.interactive
+
+    from sympy.interactive import init_session
+    init_session(ipython, **args)
+
+if __name__ == "__main__":
+    main()

local.py

@@ -0,0 +1,63 @@
+from __future__ import annotations
+
+import threading
+from contextlib import contextmanager
+from typing import TYPE_CHECKING
+
+
+if TYPE_CHECKING:
+    from collections.abc import Iterator
+
+
+# Simple dynamic scoping implementation.  The name "parametrize" comes
+# from Racket.
+#
+# WARNING WARNING: LOOKING TO EDIT THIS FILE?  Think carefully about
+# why you need to add a toggle to the global behavior of code
+# generation.  The parameters here should really only be used
+# for "temporary" situations, where we need to temporarily change
+# the codegen in some cases because we cannot conveniently update
+# all call sites, and are slated to be eliminated once all call
+# sites are eliminated.  If you don't have a plan for how to get there,
+# DON'T add a new entry here.
+
+
+class Locals(threading.local):
+    use_const_ref_for_mutable_tensors: bool | None = None
+    use_ilistref_for_tensor_lists: bool | None = None
+
+
+_locals = Locals()
+
+
+def use_const_ref_for_mutable_tensors() -> bool:
+    assert _locals.use_const_ref_for_mutable_tensors is not None, (
+        "need to initialize local.use_const_ref_for_mutable_tensors with "
+        "local.parametrize"
+    )
+    return _locals.use_const_ref_for_mutable_tensors
+
+
+def use_ilistref_for_tensor_lists() -> bool:
+    assert _locals.use_ilistref_for_tensor_lists is not None, (
+        "need to initialize local.use_ilistref_for_tensor_lists with "
+        "local.parametrize"
+    )
+    return _locals.use_ilistref_for_tensor_lists
+
+
+@contextmanager
+def parametrize(
+    *, use_const_ref_for_mutable_tensors: bool, use_ilistref_for_tensor_lists: bool
+) -> Iterator[None]:
+    old_use_const_ref_for_mutable_tensors = _locals.use_const_ref_for_mutable_tensors
+    old_use_ilistref_for_tensor_lists = _locals.use_ilistref_for_tensor_lists
+    try:
+        _locals.use_const_ref_for_mutable_tensors = use_const_ref_for_mutable_tensors
+        _locals.use_ilistref_for_tensor_lists = use_ilistref_for_tensor_lists
+        yield
+    finally:
+        _locals.use_const_ref_for_mutable_tensors = (
+            old_use_const_ref_for_mutable_tensors
+        )
+        _locals.use_ilistref_for_tensor_lists = old_use_ilistref_for_tensor_lists

markdown-it.exe

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:11d8735f26e48e2a2f578c86866dd60b5c4e99c0afa7e7904bbc1c3a213003f9
+size 108444

markdown_py.exe

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:eb545c7705c5f8f2fc008a81c5848028645bf057f596c6340d1dfa6f115d2eb8
+size 108438

native_function_generation.py

@@ -0,0 +1,651 @@
+from __future__ import annotations
+
+import string
+from collections import defaultdict
+from typing import TYPE_CHECKING
+
+import torchgen.api.dispatcher as dispatcher
+from torchgen.api.translate import translate
+from torchgen.api.types import Binding, DispatcherSignature, Expr
+from torchgen.context import with_native_function
+from torchgen.model import (
+    Annotation,
+    Argument,
+    BackendIndex,
+    BackendMetadata,
+    BaseOperatorName,
+    BaseTy,
+    BaseType,
+    DEFAULT_KERNEL_NAMESPACE,
+    DeviceCheckType,
+    DispatchKey,
+    FunctionSchema,
+    NativeFunction,
+    NativeFunctionsGroup,
+    OperatorName,
+    Return,
+    SchemaKind,
+    Variant,
+)
+from torchgen.utils import concatMap
+
+
+if TYPE_CHECKING:
+    from collections.abc import Sequence
+
+
+# See Note: [Out ops with functional variants that don't get grouped properly]
+OUT_OPS_THAT_DONT_GET_GROUPED_PROPERLY = [
+    # This has a functional variant, but it's currently marked private.
+    # This function should be marked private as well (*_backward ops aren't exposed to python anyway).
+    "adaptive_avg_pool3d_backward.grad_input",
+    # There's a functional variant, _slow_conv2d_backward.output_mask, that isn't grouped properly.
+    # Maybe we can kill this operator in favor of convolution_backward?
+    "_slow_conv2d_backward.grad_input",
+]
+
+
+# See Note: [Mutable ops that cannot get an out variant]
+MUTABLE_OPS_THAT_CANNOT_GET_AN_OUT_VARIANT = [
+    # should be out=?
+    "_cummax_helper",
+    # should be out=?
+    "_cummin_helper",
+]
+
+# All of these operators don't have any tensor like returns
+FUNCTIONAL_OPS_THAT_CANNOT_GET_AN_OUT_VARIANT = [
+    "_assert_async",  # no return
+    "_assert_async.msg",  # no return
+    "_assert_tensor_metadata",  # no return
+    "_cslt_sparse_mm_search",  # returns an int
+    "_assert_scalar",  # no return
+    "_dimI",  # returns an int
+    "_dimV",  # returns an int
+    "_has_same_storage_numel",  # returns a boolean
+    "_linalg_check_errors",  # no return
+    "_local_scalar_dense",  # returns a Scalar
+    "_nested_tensor_from_mask_left_aligned",  # returns a boolean
+    "_nnz",  # returns an int
+    "_use_cudnn_ctc_loss",  # returns a boolean
+    "_use_cudnn_ctc_loss.Tensor",  # returns a boolean
+    "_validate_compressed_sparse_indices",  # no return
+    "allclose",  # returns a boolean
+    "dense_dim",  # returns an int
+    "equal",  # returns a boolean
+    "is_coalesced",  # returns an boolean
+    "is_pinned",  # returns a boolean
+    "is_same_size",  # returns a boolean
+    "is_set_to",  # returns a boolean
+    "q_per_channel_axis",  # returns an int
+    "q_scale",  # returns a float
+    "q_zero_point",  # returns an int
+    "qscheme",  # returns a QScheme
+    "record_stream",  # no return
+    "sparse_dim",  # returns an int
+    "sym_constrain_range",  # no return
+    "sym_constrain_range_for_size",  # no return
+    "_nested_tensor_storage_offsets",  # returns a vector of ints
+    "_chunk_grad_outputs_efficient_attention",  # returns a bool
+    "_fused_sdp_choice",  # returns an int
+    "_print",  # no return
+    "_sink_tokens",  # no return
+    "_nested_get_ragged_idx",  # returns an int
+]
+
+INPLACE_OPS_THAT_DONT_GET_GROUPED_PROPERLY = [
+    # polygamma and polygamma.out both exist, but have a
+    # pre-self arg (while polygamma_ does not)
+    # We should either fix this schema so it can be grouped properly,
+    # or allow the codegen to generate new functional/out= NativeFunctions for this op
+    # (which would require changing its overload name to prevent overload ambiguity).
+    "polygamma_"
+]
+
+
+# Groups "similar" NativeFunctions together
+# example add.Tensor, add_.Tensor, add.out
+# "similar" NativeFunctions are all expected to have an identical `signature()`,
+# But have differing SchemaKinds.
+def pre_group_native_functions(
+    native_functions: Sequence[NativeFunction],
+) -> dict[FunctionSchema, dict[SchemaKind, NativeFunction]]:
+    pre_grouped_native_functions: dict[
+        FunctionSchema, dict[SchemaKind, NativeFunction]
+    ] = defaultdict(dict)
+    for f in native_functions:
+        d = pre_grouped_native_functions[f.func.signature()]
+        assert f.func.kind() not in d
+        d[f.func.kind()] = f
+    return pre_grouped_native_functions
+
+
+# Returns the out variant overload name given a base function overload name
+def get_expected_out_variant_overload_name(overload_name: str | None) -> str:
+    return "out" if not overload_name else f"{overload_name}_out"
+
+
+# Helper function: given an inplace FunctionSchema, generate its corresponding out= variant
+# Example before:
+#   _add_relu_.Scalar(Tensor(a!) self, Scalar other, Scalar alpha=1) -> Tensor(a!)
+# Example after:
+#   _add_relu.Scalar_out(Tensor self, Scalar other, Scalar alpha=1, *, Tensor(a!) out)
+def self_to_out_signature(func: FunctionSchema) -> FunctionSchema:
+    # Generating an out= schema from an inplace schema.
+    assert func.kind() == SchemaKind.inplace
+    assert func.arguments.self_arg is not None
+    # The new out= schema has:
+    # - a new out argument with the same type as "func" (but with a mutable annotation)
+    # - The returns (if any) now alias the out= argument instead of "func"
+    # - an "out" overload name
+    return FunctionSchema(
+        name=func.name.remove_inplace().with_overload(
+            get_expected_out_variant_overload_name(func.name.overload_name)
+        ),
+        arguments=func.arguments.remove_self_annotation().with_out_args(
+            [
+                Argument(
+                    name="out",
+                    type=func.arguments.self_arg.argument.type,
+                    default=None,
+                    annotation=func.arguments.self_arg.argument.annotation,
+                )
+            ]
+        ),
+        returns=func.returns,
+    )
+
+
+# Helper function: given a functional FunctionSchema, generate its corresponding out= variant
+# Example before:
+#   _to_copy(Tensor self, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None,
+#       bool? pin_memory=None, bool non_blocking=False, MemoryFormat? memory_format=None) -> Tensor
+# Example after:
+#   _to_copy._out(Tensor self, *, bool non_blocking=False, MemoryFormat? memory_format=None,
+#       Tensor(a!) out) -> Tensor(a!)
+def functional_to_out_signature(func: FunctionSchema) -> FunctionSchema:
+    # Generating an out= schema from a functional schema.
+    assert func.kind() == SchemaKind.functional
+
+    new_returns, new_out_args = generate_out_args_from_schema(func)
+    # The new out= schema has:
+    # - one or more new out argument(s) with the same type as returns (but with a mutable annotation)
+    # - The returns now alias the out= arguments
+    # - an "_out" overload name
+    return FunctionSchema(
+        name=func.name.with_overload(
+            get_expected_out_variant_overload_name(func.name.overload_name)
+        ),
+        arguments=func.arguments.signature().with_out_args(
+            new_out_args,
+        ),
+        returns=tuple(new_returns),
+    )
+
+
+# Helper function: given a function schema, generate corresponding out arguments, also the updated return annotations.
+def generate_out_args_from_schema(
+    func: FunctionSchema,
+) -> tuple[list[Return], list[Argument]]:
+    # More of a sanity check - our existing restrictions on schemas should enforce that
+    # mutable schema kinds never return their mutable arguments.
+    assert not any(
+        r.annotation is not None and r.annotation.is_write for r in func.returns
+    )
+
+    tensorlike_rets = [r for r in func.returns if r.type.is_tensor_like()]
+    assert len(tensorlike_rets) > 0
+
+    used_annotations = concatMap(
+        lambda a: [] if a.annotation is None else a.annotation.alias_set,
+        func.arguments.flat_all,
+    )
+    valid_annotations = [x for x in string.ascii_lowercase if x not in used_annotations]
+
+    all_rets_are_tensors = all(r.type == BaseType(BaseTy.Tensor) for r in func.returns)
+
+    new_out_args: list[Argument] = []
+    # The end result of new_returns is that:
+    # - If every return is a plain tensor, then the new returns == the old returns, but with the out= alias annotations added.
+    # - Otherwise, none of the out arguments show up in the returns (and we're only left with non-tensor-like returns, if any).
+    new_returns: list[Return] = []
+    for i, r in enumerate(func.returns):
+        if r.type.is_tensor_like():
+            new_out = Argument(
+                name="out" if len(func.returns) == 1 else f"out{i}",
+                type=r.type,
+                default=None,
+                annotation=Annotation.parse(f"{valid_annotations[i]}!"),
+            )
+            new_out_args.append(new_out)
+            if all_rets_are_tensors:
+                # The convention for out= schemas is that they only return their out arguments
+                # if the return is a plain Tensor (or if it's a tuple of plain Tensors)
+                new_ret = Return(
+                    name=None, type=new_out.type, annotation=new_out.annotation
+                )
+                new_returns.append(new_ret)
+        else:
+            new_returns.append(r)
+    return new_returns, new_out_args
+
+
+# Helper function: given a mutable FunctionSchema, generate its corresponding out= variant
+# Example before:
+#   _fused_moving_avg_obs_fq_helper(Tensor self, Tensor observer_on, Tensor fake_quant_on, Tensor(a!) running_min, Tensor(b!) running_max, Tensor(c!) scale, Tensor(d!) zero_point, float averaging_const, int quant_min, int quant_max, int ch_axis, bool per_row_fake_quant=False, bool symmetric_quant=False) -> (Tensor output, Tensor mask)  # noqa: B950
+# Example after:
+#   _fused_moving_avg_obs_fq_helper._out(Tensor self, Tensor observer_on, Tensor fake_quant_on, Tensor(a!) running_min, Tensor(b!) running_max, Tensor(c!) scale, Tensor(d!) zero_point, float averaging_const, int quant_min, int quant_max, int ch_axis, bool per_row_fake_quant=False, bool symmetric_quant=False, *, Tensor(e!) out0, Tensor(f!) out1) -> (Tensor(e!), Tensor(f!))  # noqa: B950
+def mutable_to_out_signature(func: FunctionSchema) -> FunctionSchema:
+    # Generating an out= schema from a mutable schema.
+    assert func.kind() == SchemaKind.mutable
+    # The new out= schema has:
+    # - Any non-aliased tensor-like returns are converted to mutable, aliased out= arguments
+    #   (if the argument is a tensor then we also return it for method chaining,
+    #   otherwise we return nothing)
+    # - an "out" overload name
+    #
+    # Note that:
+    # (1) This also means that we can *only* generate an out= variant from a mutable schema
+    #     if the mutable schema has at least one tensor-like non-aliasing return.
+    # (2) The generated out= variant still has mutable positional arguments,
+    #     but if necessary we could probably add another out= variant that also
+    #     functionalizes the mutable arguments (a functional_out variant)
+
+    new_returns, new_out_args = generate_out_args_from_schema(func)
+
+    return FunctionSchema(
+        name=func.name.remove_inplace().with_overload(
+            get_expected_out_variant_overload_name(func.name.overload_name)
+        ),
+        arguments=func.arguments.with_out_args(new_out_args),
+        returns=tuple(new_returns),
+    )
+
+
+# This function, given function of one SchemaKind, as well as a target SchemaKind,
+# generates a new NativeFunction with the same properties, but using the target SchemaKind.
+# We only actually generate functions for either functional or out= SchemaKinds.
+# This function returns a tuple, with:
+# - The generated NativeFunction
+# - a dictionary of `BackendIndex` objects, describing which dispatch keys
+#   we will generate kernels for, for the new NativeFunction.
+#   Details are in the function, but we only generate composite kernels (in some cases) today.
+def generate_function(
+    f: NativeFunction, k: SchemaKind
+) -> tuple[NativeFunction, dict[DispatchKey, dict[OperatorName, BackendMetadata]]]:
+    from torchgen.api import cpp
+
+    if k == SchemaKind.functional:
+        assert f.func.kind() != SchemaKind.functional
+        # The new "functional" NativeFunction has:
+        # - any mutable arguments have been converted into (immutable) returns.
+        #   (if a mutable argument was not also a return, it gets converted to one)
+        # - "_functional" appended to the base name, ONLY IF this op has a mutable variant.
+        #   See Note [Overload Ambiguity With Functional Variants]
+        # The default grouping logic in signature() actually already does this,
+        # so we can piggy-back off it (but we still want return names)
+        func = f.func.signature(keep_return_names=True).with_name(
+            OperatorName(
+                name=BaseOperatorName(
+                    base=f.func.name.name.base,
+                    inplace=False,
+                    dunder_method=f.func.name.name.dunder_method,
+                    # See Note [Overload Ambiguity With Functional Variants]
+                    functional_overload=f.func.kind() == SchemaKind.mutable,
+                ),
+                overload_name=f.func.name.overload_name,
+            )
+        )
+    elif k == SchemaKind.out:
+        # We generate out= ops mostly just so that we can pair up NativeFunctions into groups easily,
+        # but at least today, there is no good reason to actually use them.
+        # we'll generate a dispatcher entry for them, but won't actually register any kernels for them.
+        if f.func.kind() == SchemaKind.inplace:
+            func = self_to_out_signature(f.func)
+        elif f.func.kind() == SchemaKind.mutable:
+            func = mutable_to_out_signature(f.func)
+        elif f.func.kind() == SchemaKind.functional:
+            func = functional_to_out_signature(f.func)
+        else:
+            raise AssertionError(
+                "We only bother generating out= functions from either inplace or mutable or functional variants"
+            )
+    else:
+        raise AssertionError(
+            "We currently only generate either functional or out= NativeFunctions"
+        )
+
+    # Generated kernel naming convention for out: <op_name>_<overload_name>. The reason for this is to
+    # disambiguate operator with the same name but different overload name, e.g., `randn.names_out` and
+    # `randn.generator_with_names_out`.
+    kernel_name = (
+        func.name.unambiguous_name()
+        if func.kind() == SchemaKind.out
+        else cpp.name(func)
+    )
+    if f.func.has_symint():
+        kernel_name += "_symint"
+    backend_metadata = {
+        DispatchKey.CompositeExplicitAutograd: {
+            func.name: BackendMetadata(
+                kernel=kernel_name,
+                structured=False,
+                cpp_namespace=DEFAULT_KERNEL_NAMESPACE,
+            )
+        }
+    }
+    tags = {"generated"} | set(
+        f.tags & {"nondeterministic_seeded", "view_copy", "pt2_compliant_tag"}
+    )
+
+    return (
+        NativeFunction(
+            func=func,
+            use_const_ref_for_mutable_tensors=f.use_const_ref_for_mutable_tensors,
+            # These generated fn's aren't meant to be user friendly- don't generate methods.
+            variants={Variant.function},
+            structured=False,
+            structured_delegate=None,
+            structured_inherits=None,
+            precomputed=None,
+            autogen=[],
+            ufunc_inner_loop={},
+            manual_kernel_registration=False,
+            manual_cpp_binding=False,
+            python_module=None,
+            category_override=None,
+            device_guard=False,
+            device_check=DeviceCheckType.NoCheck,
+            loc=f.loc,
+            cpp_no_default_args=set(),
+            is_abstract=f.is_abstract,
+            has_composite_implicit_autograd_kernel=False,
+            has_composite_implicit_autograd_nested_tensor_kernel=False,
+            has_composite_explicit_autograd_kernel=True,
+            has_composite_explicit_autograd_non_functional_kernel=False,
+            # Every generated NativeFunction gets a "generated" tag, so it's easy to tell
+            # which NativeFunction objects did not come directly from native_functions.yaml.
+            tags=tags,
+            namespace=f.namespace,
+        ),
+        backend_metadata,
+    )
+
+
+# This function is responsible for adding generated NativeFunctions which don't appear
+# explicitly in the codegen.
+# You can inspect the full list of NativeFunctions yourself with the torchgen package, by running
+# torchgen.parse_native_yaml("aten/src/ATen/native/native_functions.yaml", "aten/src/ATen/native/tags.yaml")
+# (Maybe we should make a friendly API for this)
+#
+# Note: this function *mutates* its two inputs,
+# adding the new NativeFunctions / BackendMetadata to them
+def add_generated_native_functions(
+    rs: list[NativeFunction],
+    indices: dict[DispatchKey, dict[OperatorName, BackendMetadata]],
+) -> None:
+    # The main code for generating new NativeFunctions
+    # First we group of NativeFunctions by schema kind,
+    # then we detect which ones are missing and generate them.
+    pre_grouped_native_functions = pre_group_native_functions(rs)
+    for d in pre_grouped_native_functions.values():
+        has_functional = SchemaKind.functional in d
+        has_inplace = SchemaKind.inplace in d
+        has_mutable = SchemaKind.mutable in d
+        has_out = SchemaKind.out in d
+        is_core = any("core" in variant.tags for variant in d.values())
+
+        # We automatically generate a few native functions that don't exist in the yaml, for a few reasons:
+        # (1) If an operator has an inplace/out= variant but no functional variant, we can generate
+        #     a simple functional variant that the functionalization pass can consume.
+        # (2) If an operator has an inplace or functional but no out= variant, we generate an out=
+        #     variant, mostly so we can easily pair up functions into NativeFunctionsGroup,
+        #     while maintaining the constraint that the out= variant is "required".
+        if has_mutable or has_inplace or has_out or has_functional:
+            # Don't bother generating functions trio's for native functions that bypass the dispatcher.
+            are_manual = all(f.manual_cpp_binding for f in d.values())
+            # Don't bother generating functional + out= variants for view operators
+            # set_ is technically an inplace_view, but for now it is treated
+            # as a normal inplace op in the codegen
+            has_view_ops = any(
+                f.is_view_op and str(f.func.name.name) != "set_" for f in d.values()
+            )
+            # Don't generate the other variants for non-core CompositeImplicitAutograd operators.
+            # We could probably do this, but the main benefit of generating the function triplets
+            # is for transforms that need them, and transforms don't need to act directly
+            # on CompositeImplicitAutograd operators (since we let them decompose).
+            are_composite_implicit = all(
+                f.has_composite_implicit_autograd_kernel for f in d.values()
+            )
+            if are_manual or has_view_ops or are_composite_implicit and not is_core:
+                continue
+            if has_out and len(d.values()) == 1:
+                # Note: [Out ops with functional variants that don't get grouped properly]
+                # In theory we could validly have an out= operator in native_functions.yaml
+                # that has no other variants.
+                # But today, all of the operators where that's the case actually do have
+                # functional variants, that we are just unable to pair up properly.
+                # I think banning this all together is probably safer
+                # (you can always add a functional variant yourself if you want to add a new out= operator).
+                #
+                # We should probably fix the existing cases; this check is to prevent us from adding more over time.
+                if (
+                    str(d[SchemaKind.out].func.name)
+                    not in OUT_OPS_THAT_DONT_GET_GROUPED_PROPERLY
+                ):
+                    raise AssertionError(
+                        f"Found an out= operator that we could not find any other variants of: {str(d[SchemaKind.out].func)}"
+                    )
+                continue
+
+            # Some inplace ops that have problematic schemas (that we should fix), which prevent us
+            # from generating out= and functional variants
+            if (
+                has_inplace
+                and str(d[SchemaKind.inplace].func.name)
+                in INPLACE_OPS_THAT_DONT_GET_GROUPED_PROPERLY
+            ):
+                continue
+
+            base_fn = (
+                d[SchemaKind.mutable]
+                if has_mutable
+                else d[SchemaKind.inplace]
+                if has_inplace
+                else d[SchemaKind.out]
+                if has_out
+                else d[SchemaKind.functional]
+            )
+
+            # Note: [Mutable ops that cannot get an out variant]
+            # We can only generate an out= variant if either:
+            # - the original function has tensor-like returns (since we can convert them to out kwargs)
+            # - or it's inplace (since we can convert `self` to an out kwarg)
+            # There are only two functions that don't fit this criteria today though,
+            # and they both look like they should be fixed to be out= variants,
+            # so if feels safer to ban this schema all-together
+            base_fn_valid = base_fn.func.kind() == SchemaKind.inplace or any(
+                r.type.is_tensor_like() for r in base_fn.func.returns
+            )
+            # Note: [Loosen the assertion that all functional should have out variant]
+            # By design all functional operators should have our variants. The needs_out check
+            # is loosening this requirement, changing it to only generate out variant if there's
+            # an `autogen` block in the native function, in the long run it should be removed.
+            # FIXME: Remove this after figuring out CI job failures related to min, max, mean
+            needs_out = any("out" in str(op_name) for op_name in base_fn.autogen)
+            gets_out_variant = not has_out and base_fn_valid and needs_out
+            if not has_out and not base_fn_valid:
+                if (
+                    str(base_fn.func.name)
+                    not in MUTABLE_OPS_THAT_CANNOT_GET_AN_OUT_VARIANT
+                    and str(base_fn.func.name)
+                    not in FUNCTIONAL_OPS_THAT_CANNOT_GET_AN_OUT_VARIANT
+                ):
+                    raise AssertionError(
+                        f"""Found an operator that we could not generate an out= variant for: {str(base_fn.func)}.
+This type of operators don't have tensor-like return, making it difficult to generate a proper out= variant. If
+out= variant is not needed, please add the function name into FUNCTIONAL_OPS_THAT_CANNOT_GET_AN_OUT_VARIANT list."""
+                    )
+
+            # Generate an out= variant
+            if gets_out_variant:
+                fn, metadata = generate_function(base_fn, SchemaKind.out)
+                d[SchemaKind.out] = fn
+                BackendIndex.grow_index(indices, metadata)
+                rs.append(fn)
+
+            # Generate a functional variant, but only do it if the operator got an out= variant
+            # (Functional variants are only useful if we can group up the variants,
+            # which we can only do if they have an out= variant)
+            if not has_functional and (has_out or gets_out_variant):
+                fn, metadata = generate_function(base_fn, SchemaKind.functional)
+                d[SchemaKind.functional] = fn
+                BackendIndex.grow_index(indices, metadata)
+                rs.append(fn)
+
+
+def return_str(rets: tuple[Return, ...], names: list[str]) -> str:
+    assert len(rets) == len(names)
+    if len(rets) == 0:
+        return ""
+    elif len(rets) == 1:
+        return f"return {names[0]};"
+    else:
+        return f"return {dispatcher.returns_type(rets).cpp_type()}({', '.join(names)});"
+
+
+# Given a function, and the name of a variable corresponding to the output of that function,
+# gather up all of the individual returns that are not aliased
+def gather_nonaliased_inner_rets(func: FunctionSchema, out_var: str) -> list[str]:
+    aliased_rets = func.aliased_return_names()
+    non_aliased_names = []
+    is_out_var_a_tuple = len(func.returns) > 1
+    for i, r in enumerate(aliased_rets):
+        if r is None:
+            non_aliased_names.append(
+                f"std::get<{i}>({out_var})" if is_out_var_a_tuple else out_var
+            )
+    return non_aliased_names
+
+
+# Generates functional kernels in terms of their inplace.mutable counterparts.
+# We only do this for "generated" NativeFunctions
+@with_native_function
+def gen_composite_functional_kernel(g: NativeFunctionsGroup) -> str | None:
+    # We should only be generating these for code-generated NativeFunctions
+    if "generated" not in g.functional.tags:
+        return None
+    # And we always write the kernel for a generated op in terms of a non-generated op.
+    if g.inplace is not None and "generated" not in g.inplace.tags:
+        target_f = g.inplace
+    elif g.mutable is not None and "generated" not in g.mutable.tags:
+        target_f = g.mutable
+    else:
+        # We should be guaranteed to have a valid inplace/mutable variant to call into.
+        # See Note: [Mutable Ops Not Using Functionalization]
+        raise AssertionError(str(g.functional.func))
+
+    sig = DispatcherSignature(g.functional.func)
+    target_sig = DispatcherSignature(target_f.func)
+
+    context: list[Binding | Expr] = []
+    clone_mutable_inputs = []
+    cloned_return_names = []
+    # We can't just directly pass all of the arguments from the functional op into the mutating op.
+    # We need to check for which inputs to the mutating operator are mutable,
+    # and clone those inputs first.
+    for a_curr, a_tgt in zip(
+        dispatcher.jit_arguments(g.functional.func),
+        dispatcher.jit_arguments(target_f.func),
+    ):
+        if a_tgt.annotation is not None and a_tgt.annotation.is_write:
+            clone_mutable_inputs.append(
+                f"auto {a_curr.name}_clone = clone_arg({a_curr.name});"
+            )
+            context.append(
+                Expr(
+                    expr=f"{a_curr.name}_clone",
+                    type=dispatcher.argument_type(a_curr, binds=a_curr.name),
+                )
+            )
+            # Invariant: mutable arguments on the inner mutable op are always returns on the functional op.
+            cloned_return_names.append(f"{a_curr.name}_clone")
+        else:
+            context.append(dispatcher.argument(a_curr))
+    exprs = ", ".join([e.expr for e in translate(context, target_sig.arguments())])
+
+    out_name = "output"
+    maybe_assign = f"auto {out_name} = " if len(target_f.func.returns) > 0 else ""
+    inner_return_names = gather_nonaliased_inner_rets(target_f.func, out_name)
+    ret_str = return_str(
+        g.functional.func.returns, inner_return_names + cloned_return_names
+    )
+
+    clone_mutable_inputs_str = "\n".join(clone_mutable_inputs)
+    return f"""
+{sig.defn(name=sig.name() + ("_symint" if g.out.func.has_symint() else ""))} {{
+  {clone_mutable_inputs_str}
+  {maybe_assign}at::_ops::{target_f.func.name.unambiguous_name()}::call({exprs});
+  {ret_str}
+}}
+"""
+
+
+# Generates out= kernels in terms of their functional counterparts.
+# We only do this for "generated" NativeFunctions
+@with_native_function
+def gen_composite_out_kernel(g: NativeFunctionsGroup) -> str | None:
+    # We should only be generating these for code-generated NativeFunctions
+    if "generated" not in g.out.tags:
+        return None
+    # And we always write the kernel for the out= op in terms of the functional.
+    # Note that the functional op might have also been generated, but we don't have to
+    # worry about cycles, because the generated functional kernels are always implemented
+    # in terms of non-generated kernels (see gen_composite_functional_kernel).
+
+    sig = DispatcherSignature(g.out.func)
+    target_sig = DispatcherSignature(g.functional.func)
+
+    exprs = ", ".join(
+        [e.expr for e in translate(sig.arguments(), target_sig.arguments())]
+    )
+
+    copy_outs = []
+    out_name = "tmp_output"
+    for i, out_arg in enumerate(g.out.func.arguments.out):
+        functional_return_name = (
+            out_name
+            if len(g.functional.func.returns) == 1
+            else f"std::get<{i}>({out_name})"
+        )
+        copy_outs.append(
+            f"""\
+  resize_out_helper({out_arg.name}, {functional_return_name});
+  copy_arg({out_arg.name}, {functional_return_name});"""
+        )
+
+    rets = []
+    # For each return arg in the calling (out=) operator,
+    # If it corresponds to an aliased input, return the input.
+    # Otherwise, return the corresponding output from calling the functional operator.
+    for i, ret_name in enumerate(g.out.func.aliased_return_names()):
+        if ret_name is not None:
+            rets.append(ret_name)
+        else:
+            functional_return_name = (
+                out_name
+                if len(g.functional.func.returns) == 1
+                else f"std::get<{i}>({out_name})"
+            )
+            rets.append(functional_return_name)
+
+    copy_outs_str = "\n".join(copy_outs)
+
+    # Kernel name needs to follow the naming convention defined in `generate_function()`
+    return f"""
+{sig.defn(name=g.out.func.name.unambiguous_name() + ("_symint" if g.out.func.has_symint() else ""))} {{
+  auto {out_name} = at::_ops::{g.functional.func.name.unambiguous_name()}::call({exprs});
+  {copy_outs_str}
+  {return_str(g.out.func.returns, rets)}
+}}
+"""

normalizer.exe

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:655847268f9ff668e36268319e9e202056c33038014bec44346f031db40ac9b5
+size 108450

numpy-config.exe

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:980d55316535217f8748a295df5784e4eb26100a1110bae9a110492242fe7c34
+size 108440

pip.exe

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6bf6f1c22184613cb885fc8cadd6e9407d20d1ebcafec3282fbd1a52c25210a8
+size 108445

pip3.12.exe

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6bf6f1c22184613cb885fc8cadd6e9407d20d1ebcafec3282fbd1a52c25210a8
+size 108445

pip3.exe

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6bf6f1c22184613cb885fc8cadd6e9407d20d1ebcafec3282fbd1a52c25210a8
+size 108445

pygmentize.exe

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:be176558d7b05f8849604d14f1885243e5c96fd78711fdccd01ef12014e41a1b
+size 108439

python.exe

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3470f7919170d235d7e6079691462c4b217745ec67ee612e745730e46d98f238
+size 270104

pythonw.exe

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dbb5bf39746dadaf2d2513da6746ed61a04b493289272dbed72923eabfd552f8
+size 258840

saved_model_cli.exe

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:abeeaed0c37d09ad4bb9710ee2af44fd6eb08053cd4b4528e7ca6f5c1a709ee6
+size 108462

six.py

@@ -0,0 +1,1003 @@
+# Copyright (c) 2010-2024 Benjamin Peterson
+#
+# Permission is hereby granted, free of charge, to any person obtaining a copy
+# of this software and associated documentation files (the "Software"), to deal
+# in the Software without restriction, including without limitation the rights
+# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+# copies of the Software, and to permit persons to whom the Software is
+# furnished to do so, subject to the following conditions:
+#
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+#
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+# SOFTWARE.
+
+"""Utilities for writing code that runs on Python 2 and 3"""
+
+from __future__ import absolute_import
+
+import functools
+import itertools
+import operator
+import sys
+import types
+
+__author__ = "Benjamin Peterson <[email protected]>"
+__version__ = "1.17.0"
+
+
+# Useful for very coarse version differentiation.
+PY2 = sys.version_info[0] == 2
+PY3 = sys.version_info[0] == 3
+PY34 = sys.version_info[0:2] >= (3, 4)
+
+if PY3:
+    string_types = str,
+    integer_types = int,
+    class_types = type,
+    text_type = str
+    binary_type = bytes
+
+    MAXSIZE = sys.maxsize
+else:
+    string_types = basestring,
+    integer_types = (int, long)
+    class_types = (type, types.ClassType)
+    text_type = unicode
+    binary_type = str
+
+    if sys.platform.startswith("java"):
+        # Jython always uses 32 bits.
+        MAXSIZE = int((1 << 31) - 1)
+    else:
+        # It's possible to have sizeof(long) != sizeof(Py_ssize_t).
+        class X(object):
+
+            def __len__(self):
+                return 1 << 31
+        try:
+            len(X())
+        except OverflowError:
+            # 32-bit
+            MAXSIZE = int((1 << 31) - 1)
+        else:
+            # 64-bit
+            MAXSIZE = int((1 << 63) - 1)
+        del X
+
+if PY34:
+    from importlib.util import spec_from_loader
+else:
+    spec_from_loader = None
+
+
+def _add_doc(func, doc):
+    """Add documentation to a function."""
+    func.__doc__ = doc
+
+
+def _import_module(name):
+    """Import module, returning the module after the last dot."""
+    __import__(name)
+    return sys.modules[name]
+
+
+class _LazyDescr(object):
+
+    def __init__(self, name):
+        self.name = name
+
+    def __get__(self, obj, tp):
+        result = self._resolve()
+        setattr(obj, self.name, result)  # Invokes __set__.
+        try:
+            # This is a bit ugly, but it avoids running this again by
+            # removing this descriptor.
+            delattr(obj.__class__, self.name)
+        except AttributeError:
+            pass
+        return result
+
+
+class MovedModule(_LazyDescr):
+
+    def __init__(self, name, old, new=None):
+        super(MovedModule, self).__init__(name)
+        if PY3:
+            if new is None:
+                new = name
+            self.mod = new
+        else:
+            self.mod = old
+
+    def _resolve(self):
+        return _import_module(self.mod)
+
+    def __getattr__(self, attr):
+        _module = self._resolve()
+        value = getattr(_module, attr)
+        setattr(self, attr, value)
+        return value
+
+
+class _LazyModule(types.ModuleType):
+
+    def __init__(self, name):
+        super(_LazyModule, self).__init__(name)
+        self.__doc__ = self.__class__.__doc__
+
+    def __dir__(self):
+        attrs = ["__doc__", "__name__"]
+        attrs += [attr.name for attr in self._moved_attributes]
+        return attrs
+
+    # Subclasses should override this
+    _moved_attributes = []
+
+
+class MovedAttribute(_LazyDescr):
+
+    def __init__(self, name, old_mod, new_mod, old_attr=None, new_attr=None):
+        super(MovedAttribute, self).__init__(name)
+        if PY3:
+            if new_mod is None:
+                new_mod = name
+            self.mod = new_mod
+            if new_attr is None:
+                if old_attr is None:
+                    new_attr = name
+                else:
+                    new_attr = old_attr
+            self.attr = new_attr
+        else:
+            self.mod = old_mod
+            if old_attr is None:
+                old_attr = name
+            self.attr = old_attr
+
+    def _resolve(self):
+        module = _import_module(self.mod)
+        return getattr(module, self.attr)
+
+
+class _SixMetaPathImporter(object):
+
+    """
+    A meta path importer to import six.moves and its submodules.
+
+    This class implements a PEP302 finder and loader. It should be compatible
+    with Python 2.5 and all existing versions of Python3
+    """
+
+    def __init__(self, six_module_name):
+        self.name = six_module_name
+        self.known_modules = {}
+
+    def _add_module(self, mod, *fullnames):
+        for fullname in fullnames:
+            self.known_modules[self.name + "." + fullname] = mod
+
+    def _get_module(self, fullname):
+        return self.known_modules[self.name + "." + fullname]
+
+    def find_module(self, fullname, path=None):
+        if fullname in self.known_modules:
+            return self
+        return None
+
+    def find_spec(self, fullname, path, target=None):
+        if fullname in self.known_modules:
+            return spec_from_loader(fullname, self)
+        return None
+
+    def __get_module(self, fullname):
+        try:
+            return self.known_modules[fullname]
+        except KeyError:
+            raise ImportError("This loader does not know module " + fullname)
+
+    def load_module(self, fullname):
+        try:
+            # in case of a reload
+            return sys.modules[fullname]
+        except KeyError:
+            pass
+        mod = self.__get_module(fullname)
+        if isinstance(mod, MovedModule):
+            mod = mod._resolve()
+        else:
+            mod.__loader__ = self
+        sys.modules[fullname] = mod
+        return mod
+
+    def is_package(self, fullname):
+        """
+        Return true, if the named module is a package.
+
+        We need this method to get correct spec objects with
+        Python 3.4 (see PEP451)
+        """
+        return hasattr(self.__get_module(fullname), "__path__")
+
+    def get_code(self, fullname):
+        """Return None
+
+        Required, if is_package is implemented"""
+        self.__get_module(fullname)  # eventually raises ImportError
+        return None
+    get_source = get_code  # same as get_code
+
+    def create_module(self, spec):
+        return self.load_module(spec.name)
+
+    def exec_module(self, module):
+        pass
+
+_importer = _SixMetaPathImporter(__name__)
+
+
+class _MovedItems(_LazyModule):
+
+    """Lazy loading of moved objects"""
+    __path__ = []  # mark as package
+
+
+_moved_attributes = [
+    MovedAttribute("cStringIO", "cStringIO", "io", "StringIO"),
+    MovedAttribute("filter", "itertools", "builtins", "ifilter", "filter"),
+    MovedAttribute("filterfalse", "itertools", "itertools", "ifilterfalse", "filterfalse"),
+    MovedAttribute("input", "__builtin__", "builtins", "raw_input", "input"),
+    MovedAttribute("intern", "__builtin__", "sys"),
+    MovedAttribute("map", "itertools", "builtins", "imap", "map"),
+    MovedAttribute("getcwd", "os", "os", "getcwdu", "getcwd"),
+    MovedAttribute("getcwdb", "os", "os", "getcwd", "getcwdb"),
+    MovedAttribute("getoutput", "commands", "subprocess"),
+    MovedAttribute("range", "__builtin__", "builtins", "xrange", "range"),
+    MovedAttribute("reload_module", "__builtin__", "importlib" if PY34 else "imp", "reload"),
+    MovedAttribute("reduce", "__builtin__", "functools"),
+    MovedAttribute("shlex_quote", "pipes", "shlex", "quote"),
+    MovedAttribute("StringIO", "StringIO", "io"),
+    MovedAttribute("UserDict", "UserDict", "collections", "IterableUserDict", "UserDict"),
+    MovedAttribute("UserList", "UserList", "collections"),
+    MovedAttribute("UserString", "UserString", "collections"),
+    MovedAttribute("xrange", "__builtin__", "builtins", "xrange", "range"),
+    MovedAttribute("zip", "itertools", "builtins", "izip", "zip"),
+    MovedAttribute("zip_longest", "itertools", "itertools", "izip_longest", "zip_longest"),
+    MovedModule("builtins", "__builtin__"),
+    MovedModule("configparser", "ConfigParser"),
+    MovedModule("collections_abc", "collections", "collections.abc" if sys.version_info >= (3, 3) else "collections"),
+    MovedModule("copyreg", "copy_reg"),
+    MovedModule("dbm_gnu", "gdbm", "dbm.gnu"),
+    MovedModule("dbm_ndbm", "dbm", "dbm.ndbm"),
+    MovedModule("_dummy_thread", "dummy_thread", "_dummy_thread" if sys.version_info < (3, 9) else "_thread"),
+    MovedModule("http_cookiejar", "cookielib", "http.cookiejar"),
+    MovedModule("http_cookies", "Cookie", "http.cookies"),
+    MovedModule("html_entities", "htmlentitydefs", "html.entities"),
+    MovedModule("html_parser", "HTMLParser", "html.parser"),
+    MovedModule("http_client", "httplib", "http.client"),
+    MovedModule("email_mime_base", "email.MIMEBase", "email.mime.base"),
+    MovedModule("email_mime_image", "email.MIMEImage", "email.mime.image"),
+    MovedModule("email_mime_multipart", "email.MIMEMultipart", "email.mime.multipart"),
+    MovedModule("email_mime_nonmultipart", "email.MIMENonMultipart", "email.mime.nonmultipart"),
+    MovedModule("email_mime_text", "email.MIMEText", "email.mime.text"),
+    MovedModule("BaseHTTPServer", "BaseHTTPServer", "http.server"),
+    MovedModule("CGIHTTPServer", "CGIHTTPServer", "http.server"),
+    MovedModule("SimpleHTTPServer", "SimpleHTTPServer", "http.server"),
+    MovedModule("cPickle", "cPickle", "pickle"),
+    MovedModule("queue", "Queue"),
+    MovedModule("reprlib", "repr"),
+    MovedModule("socketserver", "SocketServer"),
+    MovedModule("_thread", "thread", "_thread"),
+    MovedModule("tkinter", "Tkinter"),
+    MovedModule("tkinter_dialog", "Dialog", "tkinter.dialog"),
+    MovedModule("tkinter_filedialog", "FileDialog", "tkinter.filedialog"),
+    MovedModule("tkinter_scrolledtext", "ScrolledText", "tkinter.scrolledtext"),
+    MovedModule("tkinter_simpledialog", "SimpleDialog", "tkinter.simpledialog"),
+    MovedModule("tkinter_tix", "Tix", "tkinter.tix"),
+    MovedModule("tkinter_ttk", "ttk", "tkinter.ttk"),
+    MovedModule("tkinter_constants", "Tkconstants", "tkinter.constants"),
+    MovedModule("tkinter_dnd", "Tkdnd", "tkinter.dnd"),
+    MovedModule("tkinter_colorchooser", "tkColorChooser",
+                "tkinter.colorchooser"),
+    MovedModule("tkinter_commondialog", "tkCommonDialog",
+                "tkinter.commondialog"),
+    MovedModule("tkinter_tkfiledialog", "tkFileDialog", "tkinter.filedialog"),
+    MovedModule("tkinter_font", "tkFont", "tkinter.font"),
+    MovedModule("tkinter_messagebox", "tkMessageBox", "tkinter.messagebox"),
+    MovedModule("tkinter_tksimpledialog", "tkSimpleDialog",
+                "tkinter.simpledialog"),
+    MovedModule("urllib_parse", __name__ + ".moves.urllib_parse", "urllib.parse"),
+    MovedModule("urllib_error", __name__ + ".moves.urllib_error", "urllib.error"),
+    MovedModule("urllib", __name__ + ".moves.urllib", __name__ + ".moves.urllib"),
+    MovedModule("urllib_robotparser", "robotparser", "urllib.robotparser"),
+    MovedModule("xmlrpc_client", "xmlrpclib", "xmlrpc.client"),
+    MovedModule("xmlrpc_server", "SimpleXMLRPCServer", "xmlrpc.server"),
+]
+# Add windows specific modules.
+if sys.platform == "win32":
+    _moved_attributes += [
+        MovedModule("winreg", "_winreg"),
+    ]
+
+for attr in _moved_attributes:
+    setattr(_MovedItems, attr.name, attr)
+    if isinstance(attr, MovedModule):
+        _importer._add_module(attr, "moves." + attr.name)
+del attr
+
+_MovedItems._moved_attributes = _moved_attributes
+
+moves = _MovedItems(__name__ + ".moves")
+_importer._add_module(moves, "moves")
+
+
+class Module_six_moves_urllib_parse(_LazyModule):
+
+    """Lazy loading of moved objects in six.moves.urllib_parse"""
+
+
+_urllib_parse_moved_attributes = [
+    MovedAttribute("ParseResult", "urlparse", "urllib.parse"),
+    MovedAttribute("SplitResult", "urlparse", "urllib.parse"),
+    MovedAttribute("parse_qs", "urlparse", "urllib.parse"),
+    MovedAttribute("parse_qsl", "urlparse", "urllib.parse"),
+    MovedAttribute("urldefrag", "urlparse", "urllib.parse"),
+    MovedAttribute("urljoin", "urlparse", "urllib.parse"),
+    MovedAttribute("urlparse", "urlparse", "urllib.parse"),
+    MovedAttribute("urlsplit", "urlparse", "urllib.parse"),
+    MovedAttribute("urlunparse", "urlparse", "urllib.parse"),
+    MovedAttribute("urlunsplit", "urlparse", "urllib.parse"),
+    MovedAttribute("quote", "urllib", "urllib.parse"),
+    MovedAttribute("quote_plus", "urllib", "urllib.parse"),
+    MovedAttribute("unquote", "urllib", "urllib.parse"),
+    MovedAttribute("unquote_plus", "urllib", "urllib.parse"),
+    MovedAttribute("unquote_to_bytes", "urllib", "urllib.parse", "unquote", "unquote_to_bytes"),
+    MovedAttribute("urlencode", "urllib", "urllib.parse"),
+    MovedAttribute("splitquery", "urllib", "urllib.parse"),
+    MovedAttribute("splittag", "urllib", "urllib.parse"),
+    MovedAttribute("splituser", "urllib", "urllib.parse"),
+    MovedAttribute("splitvalue", "urllib", "urllib.parse"),
+    MovedAttribute("uses_fragment", "urlparse", "urllib.parse"),
+    MovedAttribute("uses_netloc", "urlparse", "urllib.parse"),
+    MovedAttribute("uses_params", "urlparse", "urllib.parse"),
+    MovedAttribute("uses_query", "urlparse", "urllib.parse"),
+    MovedAttribute("uses_relative", "urlparse", "urllib.parse"),
+]
+for attr in _urllib_parse_moved_attributes:
+    setattr(Module_six_moves_urllib_parse, attr.name, attr)
+del attr
+
+Module_six_moves_urllib_parse._moved_attributes = _urllib_parse_moved_attributes
+
+_importer._add_module(Module_six_moves_urllib_parse(__name__ + ".moves.urllib_parse"),
+                      "moves.urllib_parse", "moves.urllib.parse")
+
+
+class Module_six_moves_urllib_error(_LazyModule):
+
+    """Lazy loading of moved objects in six.moves.urllib_error"""
+
+
+_urllib_error_moved_attributes = [
+    MovedAttribute("URLError", "urllib2", "urllib.error"),
+    MovedAttribute("HTTPError", "urllib2", "urllib.error"),
+    MovedAttribute("ContentTooShortError", "urllib", "urllib.error"),
+]
+for attr in _urllib_error_moved_attributes:
+    setattr(Module_six_moves_urllib_error, attr.name, attr)
+del attr
+
+Module_six_moves_urllib_error._moved_attributes = _urllib_error_moved_attributes
+
+_importer._add_module(Module_six_moves_urllib_error(__name__ + ".moves.urllib.error"),
+                      "moves.urllib_error", "moves.urllib.error")
+
+
+class Module_six_moves_urllib_request(_LazyModule):
+
+    """Lazy loading of moved objects in six.moves.urllib_request"""
+
+
+_urllib_request_moved_attributes = [
+    MovedAttribute("urlopen", "urllib2", "urllib.request"),
+    MovedAttribute("install_opener", "urllib2", "urllib.request"),
+    MovedAttribute("build_opener", "urllib2", "urllib.request"),
+    MovedAttribute("pathname2url", "urllib", "urllib.request"),
+    MovedAttribute("url2pathname", "urllib", "urllib.request"),
+    MovedAttribute("getproxies", "urllib", "urllib.request"),
+    MovedAttribute("Request", "urllib2", "urllib.request"),
+    MovedAttribute("OpenerDirector", "urllib2", "urllib.request"),
+    MovedAttribute("HTTPDefaultErrorHandler", "urllib2", "urllib.request"),
+    MovedAttribute("HTTPRedirectHandler", "urllib2", "urllib.request"),
+    MovedAttribute("HTTPCookieProcessor", "urllib2", "urllib.request"),
+    MovedAttribute("ProxyHandler", "urllib2", "urllib.request"),
+    MovedAttribute("BaseHandler", "urllib2", "urllib.request"),
+    MovedAttribute("HTTPPasswordMgr", "urllib2", "urllib.request"),
+    MovedAttribute("HTTPPasswordMgrWithDefaultRealm", "urllib2", "urllib.request"),
+    MovedAttribute("AbstractBasicAuthHandler", "urllib2", "urllib.request"),
+    MovedAttribute("HTTPBasicAuthHandler", "urllib2", "urllib.request"),
+    MovedAttribute("ProxyBasicAuthHandler", "urllib2", "urllib.request"),
+    MovedAttribute("AbstractDigestAuthHandler", "urllib2", "urllib.request"),
+    MovedAttribute("HTTPDigestAuthHandler", "urllib2", "urllib.request"),
+    MovedAttribute("ProxyDigestAuthHandler", "urllib2", "urllib.request"),
+    MovedAttribute("HTTPHandler", "urllib2", "urllib.request"),
+    MovedAttribute("HTTPSHandler", "urllib2", "urllib.request"),
+    MovedAttribute("FileHandler", "urllib2", "urllib.request"),
+    MovedAttribute("FTPHandler", "urllib2", "urllib.request"),
+    MovedAttribute("CacheFTPHandler", "urllib2", "urllib.request"),
+    MovedAttribute("UnknownHandler", "urllib2", "urllib.request"),
+    MovedAttribute("HTTPErrorProcessor", "urllib2", "urllib.request"),
+    MovedAttribute("urlretrieve", "urllib", "urllib.request"),
+    MovedAttribute("urlcleanup", "urllib", "urllib.request"),
+    MovedAttribute("proxy_bypass", "urllib", "urllib.request"),
+    MovedAttribute("parse_http_list", "urllib2", "urllib.request"),
+    MovedAttribute("parse_keqv_list", "urllib2", "urllib.request"),
+]
+if sys.version_info[:2] < (3, 14):
+    _urllib_request_moved_attributes.extend(
+        [
+            MovedAttribute("URLopener", "urllib", "urllib.request"),
+            MovedAttribute("FancyURLopener", "urllib", "urllib.request"),
+        ]
+    )
+for attr in _urllib_request_moved_attributes:
+    setattr(Module_six_moves_urllib_request, attr.name, attr)
+del attr
+
+Module_six_moves_urllib_request._moved_attributes = _urllib_request_moved_attributes
+
+_importer._add_module(Module_six_moves_urllib_request(__name__ + ".moves.urllib.request"),
+                      "moves.urllib_request", "moves.urllib.request")
+
+
+class Module_six_moves_urllib_response(_LazyModule):
+
+    """Lazy loading of moved objects in six.moves.urllib_response"""
+
+
+_urllib_response_moved_attributes = [
+    MovedAttribute("addbase", "urllib", "urllib.response"),
+    MovedAttribute("addclosehook", "urllib", "urllib.response"),
+    MovedAttribute("addinfo", "urllib", "urllib.response"),
+    MovedAttribute("addinfourl", "urllib", "urllib.response"),
+]
+for attr in _urllib_response_moved_attributes:
+    setattr(Module_six_moves_urllib_response, attr.name, attr)
+del attr
+
+Module_six_moves_urllib_response._moved_attributes = _urllib_response_moved_attributes
+
+_importer._add_module(Module_six_moves_urllib_response(__name__ + ".moves.urllib.response"),
+                      "moves.urllib_response", "moves.urllib.response")
+
+
+class Module_six_moves_urllib_robotparser(_LazyModule):
+
+    """Lazy loading of moved objects in six.moves.urllib_robotparser"""
+
+
+_urllib_robotparser_moved_attributes = [
+    MovedAttribute("RobotFileParser", "robotparser", "urllib.robotparser"),
+]
+for attr in _urllib_robotparser_moved_attributes:
+    setattr(Module_six_moves_urllib_robotparser, attr.name, attr)
+del attr
+
+Module_six_moves_urllib_robotparser._moved_attributes = _urllib_robotparser_moved_attributes
+
+_importer._add_module(Module_six_moves_urllib_robotparser(__name__ + ".moves.urllib.robotparser"),
+                      "moves.urllib_robotparser", "moves.urllib.robotparser")
+
+
+class Module_six_moves_urllib(types.ModuleType):
+
+    """Create a six.moves.urllib namespace that resembles the Python 3 namespace"""
+    __path__ = []  # mark as package
+    parse = _importer._get_module("moves.urllib_parse")
+    error = _importer._get_module("moves.urllib_error")
+    request = _importer._get_module("moves.urllib_request")
+    response = _importer._get_module("moves.urllib_response")
+    robotparser = _importer._get_module("moves.urllib_robotparser")
+
+    def __dir__(self):
+        return ['parse', 'error', 'request', 'response', 'robotparser']
+
+_importer._add_module(Module_six_moves_urllib(__name__ + ".moves.urllib"),
+                      "moves.urllib")
+
+
+def add_move(move):
+    """Add an item to six.moves."""
+    setattr(_MovedItems, move.name, move)
+
+
+def remove_move(name):
+    """Remove item from six.moves."""
+    try:
+        delattr(_MovedItems, name)
+    except AttributeError:
+        try:
+            del moves.__dict__[name]
+        except KeyError:
+            raise AttributeError("no such move, %r" % (name,))
+
+
+if PY3:
+    _meth_func = "__func__"
+    _meth_self = "__self__"
+
+    _func_closure = "__closure__"
+    _func_code = "__code__"
+    _func_defaults = "__defaults__"
+    _func_globals = "__globals__"
+else:
+    _meth_func = "im_func"
+    _meth_self = "im_self"
+
+    _func_closure = "func_closure"
+    _func_code = "func_code"
+    _func_defaults = "func_defaults"
+    _func_globals = "func_globals"
+
+
+try:
+    advance_iterator = next
+except NameError:
+    def advance_iterator(it):
+        return it.next()
+next = advance_iterator
+
+
+try:
+    callable = callable
+except NameError:
+    def callable(obj):
+        return any("__call__" in klass.__dict__ for klass in type(obj).__mro__)
+
+
+if PY3:
+    def get_unbound_function(unbound):
+        return unbound
+
+    create_bound_method = types.MethodType
+
+    def create_unbound_method(func, cls):
+        return func
+
+    Iterator = object
+else:
+    def get_unbound_function(unbound):
+        return unbound.im_func
+
+    def create_bound_method(func, obj):
+        return types.MethodType(func, obj, obj.__class__)
+
+    def create_unbound_method(func, cls):
+        return types.MethodType(func, None, cls)
+
+    class Iterator(object):
+
+        def next(self):
+            return type(self).__next__(self)
+
+    callable = callable
+_add_doc(get_unbound_function,
+         """Get the function out of a possibly unbound function""")
+
+
+get_method_function = operator.attrgetter(_meth_func)
+get_method_self = operator.attrgetter(_meth_self)
+get_function_closure = operator.attrgetter(_func_closure)
+get_function_code = operator.attrgetter(_func_code)
+get_function_defaults = operator.attrgetter(_func_defaults)
+get_function_globals = operator.attrgetter(_func_globals)
+
+
+if PY3:
+    def iterkeys(d, **kw):
+        return iter(d.keys(**kw))
+
+    def itervalues(d, **kw):
+        return iter(d.values(**kw))
+
+    def iteritems(d, **kw):
+        return iter(d.items(**kw))
+
+    def iterlists(d, **kw):
+        return iter(d.lists(**kw))
+
+    viewkeys = operator.methodcaller("keys")
+
+    viewvalues = operator.methodcaller("values")
+
+    viewitems = operator.methodcaller("items")
+else:
+    def iterkeys(d, **kw):
+        return d.iterkeys(**kw)
+
+    def itervalues(d, **kw):
+        return d.itervalues(**kw)
+
+    def iteritems(d, **kw):
+        return d.iteritems(**kw)
+
+    def iterlists(d, **kw):
+        return d.iterlists(**kw)
+
+    viewkeys = operator.methodcaller("viewkeys")
+
+    viewvalues = operator.methodcaller("viewvalues")
+
+    viewitems = operator.methodcaller("viewitems")
+
+_add_doc(iterkeys, "Return an iterator over the keys of a dictionary.")
+_add_doc(itervalues, "Return an iterator over the values of a dictionary.")
+_add_doc(iteritems,
+         "Return an iterator over the (key, value) pairs of a dictionary.")
+_add_doc(iterlists,
+         "Return an iterator over the (key, [values]) pairs of a dictionary.")
+
+
+if PY3:
+    def b(s):
+        return s.encode("latin-1")
+
+    def u(s):
+        return s
+    unichr = chr
+    import struct
+    int2byte = struct.Struct(">B").pack
+    del struct
+    byte2int = operator.itemgetter(0)
+    indexbytes = operator.getitem
+    iterbytes = iter
+    import io
+    StringIO = io.StringIO
+    BytesIO = io.BytesIO
+    del io
+    _assertCountEqual = "assertCountEqual"
+    if sys.version_info[1] <= 1:
+        _assertRaisesRegex = "assertRaisesRegexp"
+        _assertRegex = "assertRegexpMatches"
+        _assertNotRegex = "assertNotRegexpMatches"
+    else:
+        _assertRaisesRegex = "assertRaisesRegex"
+        _assertRegex = "assertRegex"
+        _assertNotRegex = "assertNotRegex"
+else:
+    def b(s):
+        return s
+    # Workaround for standalone backslash
+
+    def u(s):
+        return unicode(s.replace(r'\\', r'\\\\'), "unicode_escape")
+    unichr = unichr
+    int2byte = chr
+
+    def byte2int(bs):
+        return ord(bs[0])
+
+    def indexbytes(buf, i):
+        return ord(buf[i])
+    iterbytes = functools.partial(itertools.imap, ord)
+    import StringIO
+    StringIO = BytesIO = StringIO.StringIO
+    _assertCountEqual = "assertItemsEqual"
+    _assertRaisesRegex = "assertRaisesRegexp"
+    _assertRegex = "assertRegexpMatches"
+    _assertNotRegex = "assertNotRegexpMatches"
+_add_doc(b, """Byte literal""")
+_add_doc(u, """Text literal""")
+
+
+def assertCountEqual(self, *args, **kwargs):
+    return getattr(self, _assertCountEqual)(*args, **kwargs)
+
+
+def assertRaisesRegex(self, *args, **kwargs):
+    return getattr(self, _assertRaisesRegex)(*args, **kwargs)
+
+
+def assertRegex(self, *args, **kwargs):
+    return getattr(self, _assertRegex)(*args, **kwargs)
+
+
+def assertNotRegex(self, *args, **kwargs):
+    return getattr(self, _assertNotRegex)(*args, **kwargs)
+
+
+if PY3:
+    exec_ = getattr(moves.builtins, "exec")
+
+    def reraise(tp, value, tb=None):
+        try:
+            if value is None:
+                value = tp()
+            if value.__traceback__ is not tb:
+                raise value.with_traceback(tb)
+            raise value
+        finally:
+            value = None
+            tb = None
+
+else:
+    def exec_(_code_, _globs_=None, _locs_=None):
+        """Execute code in a namespace."""
+        if _globs_ is None:
+            frame = sys._getframe(1)
+            _globs_ = frame.f_globals
+            if _locs_ is None:
+                _locs_ = frame.f_locals
+            del frame
+        elif _locs_ is None:
+            _locs_ = _globs_
+        exec("""exec _code_ in _globs_, _locs_""")
+
+    exec_("""def reraise(tp, value, tb=None):
+    try:
+        raise tp, value, tb
+    finally:
+        tb = None
+""")
+
+
+if sys.version_info[:2] > (3,):
+    exec_("""def raise_from(value, from_value):
+    try:
+        raise value from from_value
+    finally:
+        value = None
+""")
+else:
+    def raise_from(value, from_value):
+        raise value
+
+
+print_ = getattr(moves.builtins, "print", None)
+if print_ is None:
+    def print_(*args, **kwargs):
+        """The new-style print function for Python 2.4 and 2.5."""
+        fp = kwargs.pop("file", sys.stdout)
+        if fp is None:
+            return
+
+        def write(data):
+            if not isinstance(data, basestring):
+                data = str(data)
+            # If the file has an encoding, encode unicode with it.
+            if (isinstance(fp, file) and
+                    isinstance(data, unicode) and
+                    fp.encoding is not None):
+                errors = getattr(fp, "errors", None)
+                if errors is None:
+                    errors = "strict"
+                data = data.encode(fp.encoding, errors)
+            fp.write(data)
+        want_unicode = False
+        sep = kwargs.pop("sep", None)
+        if sep is not None:
+            if isinstance(sep, unicode):
+                want_unicode = True
+            elif not isinstance(sep, str):
+                raise TypeError("sep must be None or a string")
+        end = kwargs.pop("end", None)
+        if end is not None:
+            if isinstance(end, unicode):
+                want_unicode = True
+            elif not isinstance(end, str):
+                raise TypeError("end must be None or a string")
+        if kwargs:
+            raise TypeError("invalid keyword arguments to print()")
+        if not want_unicode:
+            for arg in args:
+                if isinstance(arg, unicode):
+                    want_unicode = True
+                    break
+        if want_unicode:
+            newline = unicode("\n")
+            space = unicode(" ")
+        else:
+            newline = "\n"
+            space = " "
+        if sep is None:
+            sep = space
+        if end is None:
+            end = newline
+        for i, arg in enumerate(args):
+            if i:
+                write(sep)
+            write(arg)
+        write(end)
+if sys.version_info[:2] < (3, 3):
+    _print = print_
+
+    def print_(*args, **kwargs):
+        fp = kwargs.get("file", sys.stdout)
+        flush = kwargs.pop("flush", False)
+        _print(*args, **kwargs)
+        if flush and fp is not None:
+            fp.flush()
+
+_add_doc(reraise, """Reraise an exception.""")
+
+if sys.version_info[0:2] < (3, 4):
+    # This does exactly the same what the :func:`py3:functools.update_wrapper`
+    # function does on Python versions after 3.2. It sets the ``__wrapped__``
+    # attribute on ``wrapper`` object and it doesn't raise an error if any of
+    # the attributes mentioned in ``assigned`` and ``updated`` are missing on
+    # ``wrapped`` object.
+    def _update_wrapper(wrapper, wrapped,
+                        assigned=functools.WRAPPER_ASSIGNMENTS,
+                        updated=functools.WRAPPER_UPDATES):
+        for attr in assigned:
+            try:
+                value = getattr(wrapped, attr)
+            except AttributeError:
+                continue
+            else:
+                setattr(wrapper, attr, value)
+        for attr in updated:
+            getattr(wrapper, attr).update(getattr(wrapped, attr, {}))
+        wrapper.__wrapped__ = wrapped
+        return wrapper
+    _update_wrapper.__doc__ = functools.update_wrapper.__doc__
+
+    def wraps(wrapped, assigned=functools.WRAPPER_ASSIGNMENTS,
+              updated=functools.WRAPPER_UPDATES):
+        return functools.partial(_update_wrapper, wrapped=wrapped,
+                                 assigned=assigned, updated=updated)
+    wraps.__doc__ = functools.wraps.__doc__
+
+else:
+    wraps = functools.wraps
+
+
+def with_metaclass(meta, *bases):
+    """Create a base class with a metaclass."""
+    # This requires a bit of explanation: the basic idea is to make a dummy
+    # metaclass for one level of class instantiation that replaces itself with
+    # the actual metaclass.
+    class metaclass(type):
+
+        def __new__(cls, name, this_bases, d):
+            if sys.version_info[:2] >= (3, 7):
+                # This version introduced PEP 560 that requires a bit
+                # of extra care (we mimic what is done by __build_class__).
+                resolved_bases = types.resolve_bases(bases)
+                if resolved_bases is not bases:
+                    d['__orig_bases__'] = bases
+            else:
+                resolved_bases = bases
+            return meta(name, resolved_bases, d)
+
+        @classmethod
+        def __prepare__(cls, name, this_bases):
+            return meta.__prepare__(name, bases)
+    return type.__new__(metaclass, 'temporary_class', (), {})
+
+
+def add_metaclass(metaclass):
+    """Class decorator for creating a class with a metaclass."""
+    def wrapper(cls):
+        orig_vars = cls.__dict__.copy()
+        slots = orig_vars.get('__slots__')
+        if slots is not None:
+            if isinstance(slots, str):
+                slots = [slots]
+            for slots_var in slots:
+                orig_vars.pop(slots_var)
+        orig_vars.pop('__dict__', None)
+        orig_vars.pop('__weakref__', None)
+        if hasattr(cls, '__qualname__'):
+            orig_vars['__qualname__'] = cls.__qualname__
+        return metaclass(cls.__name__, cls.__bases__, orig_vars)
+    return wrapper
+
+
+def ensure_binary(s, encoding='utf-8', errors='strict'):
+    """Coerce **s** to six.binary_type.
+
+    For Python 2:
+      - `unicode` -> encoded to `str`
+      - `str` -> `str`
+
+    For Python 3:
+      - `str` -> encoded to `bytes`
+      - `bytes` -> `bytes`
+    """
+    if isinstance(s, binary_type):
+        return s
+    if isinstance(s, text_type):
+        return s.encode(encoding, errors)
+    raise TypeError("not expecting type '%s'" % type(s))
+
+
+def ensure_str(s, encoding='utf-8', errors='strict'):
+    """Coerce *s* to `str`.
+
+    For Python 2:
+      - `unicode` -> encoded to `str`
+      - `str` -> `str`
+
+    For Python 3:
+      - `str` -> `str`
+      - `bytes` -> decoded to `str`
+    """
+    # Optimization: Fast return for the common case.
+    if type(s) is str:
+        return s
+    if PY2 and isinstance(s, text_type):
+        return s.encode(encoding, errors)
+    elif PY3 and isinstance(s, binary_type):
+        return s.decode(encoding, errors)
+    elif not isinstance(s, (text_type, binary_type)):
+        raise TypeError("not expecting type '%s'" % type(s))
+    return s
+
+
+def ensure_text(s, encoding='utf-8', errors='strict'):
+    """Coerce *s* to six.text_type.
+
+    For Python 2:
+      - `unicode` -> `unicode`
+      - `str` -> `unicode`
+
+    For Python 3:
+      - `str` -> `str`
+      - `bytes` -> decoded to `str`
+    """
+    if isinstance(s, binary_type):
+        return s.decode(encoding, errors)
+    elif isinstance(s, text_type):
+        return s
+    else:
+        raise TypeError("not expecting type '%s'" % type(s))
+
+
+def python_2_unicode_compatible(klass):
+    """
+    A class decorator that defines __unicode__ and __str__ methods under Python 2.
+    Under Python 3 it does nothing.
+
+    To support Python 2 and 3 with a single code base, define a __str__ method
+    returning text and apply this decorator to the class.
+    """
+    if PY2:
+        if '__str__' not in klass.__dict__:
+            raise ValueError("@python_2_unicode_compatible cannot be applied "
+                             "to %s because it doesn't define __str__()." %
+                             klass.__name__)
+        klass.__unicode__ = klass.__str__
+        klass.__str__ = lambda self: self.__unicode__().encode('utf-8')
+    return klass
+
+
+# Complete the moves implementation.
+# This code is at the end of this module to speed up module loading.
+# Turn this module into a package.
+__path__ = []  # required for PEP 302 and PEP 451
+__package__ = __name__  # see PEP 366 @ReservedAssignment
+if globals().get("__spec__") is not None:
+    __spec__.submodule_search_locations = []  # PEP 451 @UndefinedVariable
+# Remove other six meta path importers, since they cause problems. This can
+# happen if six is removed from sys.modules and then reloaded. (Setuptools does
+# this for some reason.)
+if sys.meta_path:
+    for i, importer in enumerate(sys.meta_path):
+        # Here's some real nastiness: Another "instance" of the six module might
+        # be floating around. Therefore, we can't use isinstance() to check for
+        # the six meta path importer, since the other six instance will have
+        # inserted an importer with different class.
+        if (type(importer).__name__ == "_SixMetaPathImporter" and
+                importer.name == __name__):
+            del sys.meta_path[i]
+            break
+    del i, importer
+# Finally, add the importer to the meta path import hook.
+sys.meta_path.append(_importer)