tests/test_upgrades.py

"""Tests for BPE tokenizer, HV attention, and abstraction engine."""

from __future__ import annotations

import pytest
import numpy as np

from palimseste import hv
from palimseste.learner import Encoder
from palimseste.bpe import BPETokenizer
from palimseste.attention import HVAttention, AttentionConfig
from palimseste.abstraction import AbstractionEngine, AbstractionConfig
from palimseste.memory import Memory
from palimseste.lm import PalimpsesteForCausalLM, PalimpsesteConfig


# ================================================================ BPE
class TestBPE:
    def _tok(self, D=2000):
        enc = Encoder(D=D, rng=np.random.default_rng(0))
        return BPETokenizer(encoder=enc, vocab_size=1000)

    def test_train_and_encode(self):
        tok = self._tok()
        text = "hello world hello world hello world hello world"
        tok.train(text, target_vocab_size=300)
        ids = tok.encode("hello world")
        assert len(ids) > 0
        assert len(ids) < len("hello world")  # fewer tokens than chars

    def test_encode_decode_roundtrip(self):
        tok = self._tok()
        text = "the quick brown fox jumps over the lazy dog"
        tok.train(text * 10, target_vocab_size=500)
        ids = tok.encode(text)
        decoded = tok.decode(ids)
        assert decoded == text

    def test_bpe_reduces_sequence_length(self):
        """BPE should produce fewer tokens than char-level."""
        tok = self._tok()
        text = "hello world " * 50
        tok.train(text, target_vocab_size=2000)
        bpe_ids = tok.encode("hello world hello world")
        char_count = len("hello world hello world")
        assert len(bpe_ids) < char_count, f"BPE {len(bpe_ids)} should be < chars {char_count}"

    def test_bos_eos(self):
        tok = self._tok()
        tok.train("abc abc abc", target_vocab_size=300)
        ids = tok.encode("abc", add_bos=True, add_eos=True)
        assert ids[0] == 1  # BOS
        assert ids[-1] == 2  # EOS

    def test_token_hv_stable(self):
        tok = self._tok()
        tok.train("hello hello hello", target_vocab_size=300)
        h1 = tok.token_hv(4)
        h2 = tok.token_hv(4)
        assert h1 == h2

    def test_save_load(self, tmp_path):
        tok = self._tok()
        tok.train("test test test " * 20, target_vocab_size=300)
        ids_before = tok.encode("test test")
        tok.save_vocabulary(tmp_path / "bpe.json")
        enc = Encoder(D=2000, rng=np.random.default_rng(0))
        tok2 = BPETokenizer.load_vocabulary(tmp_path / "bpe.json", encoder=enc)
        ids_after = tok2.encode("test test")
        assert ids_before == ids_after

    def test_encode_context(self):
        tok = self._tok(D=2000)
        tok.train("hello world test " * 20, target_vocab_size=500)
        ids = tok.encode("hello world test")
        ctx_hv = tok.encode_context(ids, window=32)
        assert isinstance(ctx_hv, hv.HV)
        assert ctx_hv.D == 2000


# ================================================================ Attention
class TestHVAttention:
    def _enc(self, D=2000):
        return Encoder(D=D, rng=np.random.default_rng(0))

    def test_attend_returns_hv(self):
        enc = self._enc()
        att = HVAttention(config=AttentionConfig(max_context=64, top_k=8),
                          encoder=enc)
        rng = np.random.default_rng(1)
        tokens = [hv.random_hv(2000, rng=rng) for _ in range(20)]
        query = hv.random_hv(2000, rng=rng)
        result = att.attend(tokens, query)
        assert isinstance(result, hv.HV)
        assert result.D == 2000

    def test_attention_is_selective(self):
        """The attended HV should be more similar to the query-matching
        tokens than to a random HV."""
        enc = self._enc(D=5000)
        att = HVAttention(config=AttentionConfig(max_context=64, top_k=4, temperature=2.0),
                          encoder=enc)
        rng = np.random.default_rng(2)
        # create tokens where one is very similar to the query
        query = hv.random_hv(5000, rng=rng)
        near_token = query  # identical = maximum attention
        far_tokens = [hv.random_hv(5000, rng=rng) for _ in range(20)]
        tokens = far_tokens[:10] + [near_token] + far_tokens[10:]
        result = att.attend(tokens, query)
        # result should be more similar to query than a random HV
        random_hv = hv.random_hv(5000, rng=rng)
        assert hv.similarity(result, query) > hv.similarity(random_hv, query)

    def test_top_k_limits_selection(self):
        enc = self._enc()
        att = HVAttention(config=AttentionConfig(max_context=32, top_k=4),
                          encoder=enc)
        rng = np.random.default_rng(3)
        tokens = [hv.random_hv(2000, rng=rng) for _ in range(32)]
        query = hv.random_hv(2000, rng=rng)
        result = att.attend(tokens, query)
        assert isinstance(result, hv.HV)

    def test_empty_context(self):
        enc = self._enc()
        att = HVAttention(config=AttentionConfig(), encoder=enc)
        query = hv.random_hv(2000, rng=np.random.default_rng(0))
        result = att.attend([], query)
        assert result == query

    def test_max_context_truncation(self):
        enc = self._enc()
        att = HVAttention(config=AttentionConfig(max_context=10, top_k=5),
                          encoder=enc)
        rng = np.random.default_rng(4)
        tokens = [hv.random_hv(2000, rng=rng) for _ in range(50)]
        query = hv.random_hv(2000, rng=rng)
        # should not crash with more tokens than max_context
        result = att.attend(tokens, query)
        assert isinstance(result, hv.HV)

    def test_config_validation(self):
        with pytest.raises(ValueError):
            AttentionConfig(max_context=0)
        with pytest.raises(ValueError):
            AttentionConfig(top_k=0)
        with pytest.raises(ValueError):
            AttentionConfig(top_k=100, max_context=50)


# ================================================================ Abstraction
class TestAbstraction:
    def _mem(self, D=2000):
        return Memory(D=D, rng=np.random.default_rng(0))

    def test_extract_concepts_from_similar_traces(self):
        """Traces with similar addresses should form a cluster."""
        mem = self._mem(D=3000)
        rng = np.random.default_rng(1)
        base = hv.random_hv(3000, rng=rng)

        # create traces with addresses near `base`
        for i in range(10):
            signs = hv.bits_to_signs(base)
            # flip a few bits to make it similar but not identical
            flip = rng.choice(3000, size=20, replace=False)
            signs[flip] = -signs[flip]
            addr = hv.signs_to_bits(signs)
            val = hv.random_hv(3000, rng=rng)
            mem.write(addr, val, tag=f"item_{i}")

        engine = AbstractionEngine(
            mem=mem,
            config=AbstractionConfig(
                n_clusters=10,
                min_cluster_size=3,
                similarity_threshold=0.2,
                sample_size=100,
            ),
            rng=rng,
        )
        concepts = engine.extract_concepts()
        assert len(concepts) > 0
        assert concepts[0].n_members >= 3

    def test_no_concepts_from_dissimilar_traces(self):
        mem = self._mem(D=3000)
        rng = np.random.default_rng(2)
        for _ in range(20):
            mem.write(hv.random_hv(3000, rng=rng), hv.random_hv(3000, rng=rng))

        engine = AbstractionEngine(
            mem=mem,
            config=AbstractionConfig(
                similarity_threshold=0.9,  # very strict
                min_cluster_size=3,
                sample_size=100,
            ),
            rng=rng,
        )
        concepts = engine.extract_concepts()
        assert len(concepts) == 0

    def test_find_concept(self):
        mem = self._mem(D=3000)
        rng = np.random.default_rng(3)
        base = hv.random_hv(3000, rng=rng)
        for i in range(5):
            signs = hv.bits_to_signs(base)
            flip = rng.choice(3000, size=10, replace=False)
            signs[flip] = -signs[flip]
            mem.write(hv.signs_to_bits(signs), hv.random_hv(3000, rng=rng))

        engine = AbstractionEngine(
            mem=mem,
            config=AbstractionConfig(similarity_threshold=0.3, min_cluster_size=2),
            rng=rng,
        )
        engine.extract_concepts()
        # query with something near the base
        concept = engine.find_concept(base)
        assert concept is not None

    def test_find_concept_returns_none_for_dissimilar(self):
        mem = self._mem(D=2000)
        rng = np.random.default_rng(4)
        for _ in range(5):
            mem.write(hv.random_hv(2000, rng=rng), hv.random_hv(2000, rng=rng))
        engine = AbstractionEngine(mem=mem, rng=rng)
        engine.extract_concepts()
        result = engine.find_concept(hv.random_hv(2000, rng=rng))
        # might be None or a weak match
        assert result is None or isinstance(result, object)

    def test_concepts_stored_in_memory(self):
        """Extracted concept centroids should be written into M."""
        mem = self._mem(D=3000)
        rng = np.random.default_rng(5)
        base = hv.random_hv(3000, rng=rng)
        for i in range(10):
            signs = hv.bits_to_signs(base)
            flip = rng.choice(3000, size=15, replace=False)
            signs[flip] = -signs[flip]
            mem.write(hv.signs_to_bits(signs), hv.random_hv(3000, rng=rng))

        n_before = len(mem)
        engine = AbstractionEngine(
            mem=mem,
            config=AbstractionConfig(similarity_threshold=0.2, min_cluster_size=3),
            rng=rng,
        )
        engine.extract_concepts()
        assert len(mem) > n_before  # concept centroids were added

    def test_empty_memory(self):
        mem = self._mem(D=2000)
        engine = AbstractionEngine(mem=mem, rng=np.random.default_rng(0))
        concepts = engine.extract_concepts()
        assert concepts == []

    def test_config_validation(self):
        with pytest.raises(ValueError):
            AbstractionConfig(n_clusters=0)
        with pytest.raises(ValueError):
            AbstractionConfig(min_cluster_size=1)
        with pytest.raises(ValueError):
            AbstractionConfig(similarity_threshold=0)
        with pytest.raises(ValueError):
            AbstractionConfig(similarity_threshold=1.5)