justinchuby

def solve_Ax_eq_zero(A):
    _, _, vh = np.linalg.svd(A)
    x = vh.T[:, -1]
    return x

justinchuby

Python container typing hierarchy

Including collections abstract base classes

classDiagram
    AbstractSet <-- FrozenSet
    AbstractSet <-- KeysView
    AbstractSet <-- MutableSet
 MutableSet &lt;-- Set

justinchuby

# mypy: disable-error-code="misc,arg-type,type-arg,valid-type,assignment,return-value"

justinchuby

In [3]: def foo(*a): print(a)

In [4]: class Test:
   ...:     func = foo
   ...:     def bar(self):
   ...:         self.func(42)
   ...:

In [5]: Test().bar()
(<__main__.Test object at 0x106751ab0>, 42)

justinchuby

Best practice for importing in Python

1. Import at top level only

Allow imports at the module toplevel only, unless (1) it is too expensive to load the module or (2) module may not be available.

• It is clear what a module needs when all imports are grouped in a single place. This makes refactoring easy.
• Otherwise, any import errors will be raised only when the code executes. Erroneous imports may go undetected until the code path is hit at runtime.
• Doing so reduces code duplication and improves consistency when we don't have the same import lines spread across the file.
• Pylint has rule that checks for this: https://pylint.readthedocs.io/en/latest/user_guide/messages/convention/import-outside-toplevel.html

justinchuby

# https://github.com/pytorch/pytorch/pull/106984/files

retry () {
    "$@" || (sleep 10 && "$@") || (sleep 20 && "$@") || (sleep 40 && "$@")
}

justinchuby

import torch
from torch import nn
import onnx


class LlamaMSRotaryEmbedding(nn.Module):
    def __init__(self, hidden_size, num_heads, max_sequence_length):
        super().__init__()

        self.hidden_size = hidden_size

justinchuby

import torch
import onnx


class Model(torch.nn.Module):
    def forward(self, x):
        return torch.nn.functional.hardsigmoid(x)

exported = torch.onnx.dynamo_export(Model(), torch.randn(1, 3, 224, 224))
print(onnx.printer.to_text(exported.model_proto))

justinchuby

# onnxscript/tests/function_libs/torch_lib/dynamo_export_test.py

import copy
import inspect
import itertools
import sys
import unittest

import torch
from torch.onnx import ExportOptions

justinchuby

# https://aiinfra.visualstudio.com/PublicPackages/_artifacts/feed/ORT-Nightly/PyPI/ort-nightly/overview
--index-url=https://aiinfra.pkgs.visualstudio.com/PublicPackages/_packaging/ORT-Nightly/pypi/simple/
ort-nightly==1.17.0.dev20240118001