Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Add a design topic page on lazy vs. eager implementations #708

Merged
merged 3 commits into from
Nov 27, 2023
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
1 change: 1 addition & 0 deletions spec/draft/design_topics/index.rst
Original file line number Diff line number Diff line change
Expand Up @@ -7,6 +7,7 @@ Design topics & constraints

copies_views_and_mutation
data_dependent_output_shapes
lazy_eager
data_interchange
device_support
static_typing
Expand Down
43 changes: 43 additions & 0 deletions spec/draft/design_topics/lazy_eager.rst
Original file line number Diff line number Diff line change
@@ -0,0 +1,43 @@
.. _lazy-eager:

Lazy vs. eager execution
========================

While the execution model for implementations is out of scope of this standard,
there are a few aspects of lazy (or graph-based) execution as contrasted to
eager execution that may have an impact on the prescribed semantics of
individual APIs, and will therefore show up in the API specification.

One important difference is data-dependent or value-dependent behavior, as
described in :ref:`data-dependent-output-shapes`. Because such behavior is hard
to implement, implementers may choose to omit such APIs from their library.

Another difference is when the Python language itself prescribes that a
specific type *must* be returned. For those cases, it is not possible to return
a lazy/delayed kind of object to avoid computing a value. This is the case for
five dunder methods: `__bool__`, `__int__`, `__float__`, `__complex__` and
`__index__`. Each implementation has only two choices when one of these methods
is called:

1. Compute a value of the required type (a Python scalar of type `bool`, `int`,
`float` or `complex`), or
2. Raise an exception.

When an implementation is 100% lazy, for example when it serializes a
computation graph, computing the value is not possible and hence such an
implementation has no choice but to raise an exception. For a "mostly lazy"
implementation, it may make sense to trigger execution instead - but it is not
required to, both choices are valid.

A common code construct where this happens is conditional logic, e.g.::

vals = compute_something()
if all(vals):
# The if-statement will make Python call the __bool__ method
# on the result of `all(vals)`.
do_something_else()

Note that the API does not contain control flow constructs, as of now, that
would allow avoiding the implicit `__bool__` call in the example above. The
only control flow-like function is `where`, but there's no function like `cond`
to replace an `if`-statement.