src/bigru_t/__init__.py

"""BiGRU_T_version — V6.5 (reestruturado).

V6.5 cleanup: pre-V6.4 modules (unified_model, u8cell_T, BiGRU4,
TransformerUnit, OrqCell, TrainT, ModuleSelector) foram REMOVIDOS.
Apenas KohonenLearningSystem (V6.4, movido de kohonen_refactored/) e
HypT (V6.4) permanecem como módulos centrais.

V6.5 novidades:
- KohonenLearningSystem com VQ-VAE-2 compressor + reasoning_engine integrados
- VQ-VAE-2 ativo no pipeline de compressão (HierarchicalVQVAE2)
- ReasoningEngine com streaming <think>/<plan>/<answer> (compat Ollama)
- 6 datasets PT-BR para esgotar (incl. Dexavator/English-PTBR)

Lemas históricos (preservados nos módulos restantes):
- Lema 3: Cancelamento de ruído W8A8 (HypT + QuantizedLinear)
- Lema 4: Auto-configuração (MetaConfigurator)
"""
__version__ = "6.5.0"

from .model.kohonen_learning_system import (
    SimpleBBPETokenizer,
    positional_encoding,
    text_to_4d_vector,
    KohonenSOM4D,
    HypothesisClassifier,
    KohonenLearningSystem,
)
from .model.hyp_t import HypT

from .quantization.quantized_linear import QuantizedLinear, quantize_tensor, apply_w8a8

from .training.gradient_surgery import apply_gradient_surgery, orthogonalize_gradient
from .training.meta_configurator import MetaConfigurator
from .training.kill_switch import KillSwitch, KillSwitchState
from .training.dpo import dpo_loss, compute_dynamic_beta, compute_sequence_logps, dpo_step
from .training.hypothesis_monitor import HypothesisMonitor, HypothesisMonitorReport

from .reasoning.circular_reasoning_wasserstein import CircularReasoningWasserstein

from .inference.generator import BiGRUTGenerator

from .optim.hamiltonian_wasserstein import HamiltonianWassersteinOptimizer

from .utils.memory_cleanup import (
    aggressive_cleanup, production_cleanup, TimeBudget, StepTimer, get_rss_mb,
)

__all__ = [
    # V6.5 core
    "SimpleBBPETokenizer", "positional_encoding", "text_to_4d_vector",
    "KohonenSOM4D", "HypothesisClassifier", "KohonenLearningSystem",
    "HypT",
    # Quantization
    "QuantizedLinear", "quantize_tensor", "apply_w8a8",
    # Training
    "apply_gradient_surgery", "orthogonalize_gradient",
    "MetaConfigurator",
    "KillSwitch", "KillSwitchState",
    "dpo_loss", "compute_dynamic_beta", "compute_sequence_logps", "dpo_step",
    "HypothesisMonitor", "HypothesisMonitorReport",
    "CircularReasoningWasserstein",
    "BiGRUTGenerator",
    "HamiltonianWassersteinOptimizer",
    "aggressive_cleanup", "production_cleanup", "TimeBudget", "StepTimer", "get_rss_mb",
]