Use non-editable installs for test generators

[danceratopz]

Fixes up a regression introduced in:

• Migrate config to pyproject.toml and use uv for project management #4627

This issue describes the problem and fix:

• Editable installs + multiprocessing break vector generation #4633

This will cause the venv to flip-flop between installing eth2spec in editable and non-editable modes if you switch between non-generator/generator targets. But that should be ok; uv's good at that :)

[danceratopz]

Could a maintainer please trigger Generate Vectors to test the fix?

Repo:

danceratopz/consensus-specs

Ref:

fix/vector-gen-with-process-based-pools

[etan-status]

I'm not a maintainer but I ran it locally and it works fine now! Thanks for fixing it.

% rm -rf ../consensus-spec-tests/tests; \
    mkdir -p ../consensus-spec-tests/tests && \
    make clean && \
    make lint verbose=true && \
    caffeinate -ims make reftests && \
    find ../consensus-spec-tests/tests -depth -type d -empty -delete

Completed *::*::* in 2h 11m 35s

  Reference Test Generation Summary  
╭───────────┬──────────┬────────────╮
│ Metric    │    Count │ Percentage │
├───────────┼──────────┼────────────┤
│ Found     │   197562 │     100.0% │
│ Filtered  │        0 │       0.0% │
│ Selected  │   197562 │     100.0% │
│ Completed │    73619 │      37.3% │
│ Skipped   │   123943 │      62.7% │
│ Time      │ 7895.28s │            │
╰───────────┴──────────┴────────────╯

---

If we ever need pickling support, the C libraries can be wrapped easily, e.g., for lru-dict:

from collections.abc import Hashable
from typing import Any, Generic, TypeVar

from lru import LRU as OrigLRU

_KT = TypeVar("_KT", bound=Hashable)
_VT = TypeVar("_VT")


class LRU(Generic[_KT, _VT]):
    def __init__(self, size):
        self.cache_dict = OrigLRU(size=size)

    def __contains__(self, __o: Any) -> bool:
        return self.cache_dict.__contains__(__o)

    def __getitem__(self, item: _KT) -> _VT:
        return self.cache_dict.__getitem__(item)

    def __setitem__(self, key: _KT, value: _VT) -> None:
        self.cache_dict.__setitem__(key, value)

    def __reduce__(self):
        return (self.__class__, (self.cache_dict.get_size()))

with __reduce__ essentially describing the arguments with which __init__ should be called on the destination process. then, just import that module rather than the C library directly, caller code doesn't need other changes. BLS G1Point etc would likewise need a similar treatment.

However, #4633's long term plan suggests to move away from pickling entirely. Still leaving the comment in case it becomes an issue in the future.

[danceratopz]

Thanks for verifying, @etan-status! And thanks for your alternative suggestion - your approach seems absolutely valid 👍

I'm happy if maintainers prefer to apply your suggestion over the less intrusive "hotfix" in this PR.

These are the reasons I took the approach in this PR:

1. It maintains the status quo for vector generation from before Migrate config to pyproject.toml and use uv for project management #4627. I looked at similar solutions to yours, but wasn't sure if there could be over side effects/issues, see also next point :).
2. I won't be able to PR a more involved approach, as I'll be afk for a wee while.
3. As you mentioned, I lean towards this low effort hotfix as I believe we should put our efforts into unifying make test and make reftests to use the same test framework, i.e., to both use pytest. This would make your fix redundant, as it would work out the box in pytest via pytest-xdist (more info in Editable installs + multiprocessing break vector generation #4633).

[etan-status]

I like this PR as is, my comment was more meant as further documentation in case pickling gets revisited later on. Status quo worked well, no reason to change it on a whim.

[jtraglia]

LGTM, thanks @danceratopz! Can confirm it fixes the issue.