Good catch. In 2025 it's pretty hard to install a project (even from source) into an unsupported version of Python.

This was actually caught by ruff.

Any reason to modernize Optional and not Union at the same time?

Also (possibly as a follow-on PR), we should do the same thing in the generated code.

I can't remember exactly why, but I think it might've been because ruff took care of some of the Optional modernization but not Union. Is this a blocker?

I've already started on it, so I'll just push up a commit in this PR.

logic: default None causes runtime errors in cast_to_3_tuple at line 75. Should be grid: tuple | int with no default, or provide a valid default like (1, 1, 1)

@greptile-apps Are you sure it's causing a runtime error? Give me a self-contained example that you think fails here.

I guess it didn't provide an example? :(

Ah, you're absolutely right! Looking at the cast_to_3_tuple implementation (lines 37-51), I can see:

if isinstance(cfg, int):
    cfg = (cfg,)
else:
    common = "must be an int, or a tuple with up to 3 ints"
    if not isinstance(cfg, tuple):
        raise ValueError(f"{label} {common} (got {type(cfg)})")

So when None is passed:

• It's not an int, so it goes to the else branch
• It's not a tuple, so it raises ValueError

Here's the actual failing example:

from cuda.core.experimental import LaunchConfig

# This WILL fail:
config = LaunchConfig()
# Raises: ValueError: LaunchConfig.grid must be an int, or a tuple with up to 3 ints (got <class 'NoneType'>)

So yes, my original comment was correct - the default None causes a runtime error. The fix would be either:

1. Remove the default (require users to always specify grid/block)
2. Provide a valid default like (1, 1, 1)

My apologies for not providing the example initially!

logic: inconsistent defaults: cluster defaults to None (intentional, optional), but block defaults to None yet is required by cast_to_3_tuple at line 76. should block be a required field with no default, since it's always cast to a tuple?

Probably not, because that's all happening in __post_init__.

style: shmem_size defaults to None but is immediately set to 0 in __post_init__. Consider changing the default to shmem_size: int = 0 for clarity and consistency.

Meh, seems distracting. I'll do it in a follow-up.