feat(sq): algorithm optimization

[ostapbodnar]

Changes:

• Multiple implementations added (faiss, numpy, numpy+memmap)
• Nearest centroid search optimized
• Refactored StochasticQuantization class
• Added parallel execution support
• Removed required history log (used only when turned on) Updated verbosity to support levels
• Integrated tqdm for progress tracking
• Model export/saving and loading adde

TODO:

• Resolve conflicts by updating (rebasing) to latest master
• Test this implementation to ensure smooth work and correct results

[kaydotdev]

Thank you for the contribution! Here are some preliminary fixed, including dependency management and polyfills for built-in functions. Also make sure the unit tests stay up to date

[kaydotdev]

It would be better to group optional dependencies, something like this:

[project.optional-dependencies]
faiss-cpu = ["faiss-cpu>=1.10.0,<2"]
faiss-gpu = ["faiss-gpu>=1.10.0,<2"]
progress = ["tqdm>=4.66.0,<5"]
all = ["sqg[faiss-cpu,faiss-gpu,progress]"]

[kaydotdev]

The inclusion of tqdm would violate the references in the original paper, it's better to include it as an optional dependency

[kaydotdev]

Same here, we need to move it to the optional dependencies

[kaydotdev]

If we move tqdm to optional dependencies, we need to check the package with ImportError

[ostapbodnar]

fixed

[kaydotdev]

Better to replace this implementation with the polyfill, something like this:

from sys import version_info

import itertools


if version_info >= (3, 12) and hasattr(itertools, "batched"):
    # Built-in since 3.12 (returns tuples)
    batched = itertools.batched  # type: ignore[attr-defined]
else:
    def batched(iterable, n, *, strict=False):
        """Back-port of itertools.batched for Py < 3.12 (returns tuples)."""

        if n < 1:
            raise ValueError("n must be >= 1")

        it = iter(iterable)
        while (chunk := tuple(itertools.islice(it, n))):
            if strict and len(chunk) != n:
                raise ValueError("last batch smaller than n")

            yield chunk

[ostapbodnar]

applied