"""
CNN-BiGRU — Modelo Multimodal Cooperativo Autoaprendível (v2.0)

Implementação completa de um LLM multimodal com arquitetura CNN-BiGRU cooperativa:
  - Núcleo dual-stream CNN-BiGRU com 3 níveis de pontes (cross-attention, gated cells, fusão)
  - Encoders multimodais (imagem + áudio) com fusão gated
  - Gerador + Verificador + Camada anti-alucinação (lógica fuzzy de Łukasiewicz)
  - Auto-aprendizado: ajuste dinâmico de LR, spectral norm, orthogonal init
  - Hipóteses ativadas por punições (Gumbel-Softmax diferenciável)
  - Synergy search (N tentativas de combinação de hiperparâmetros)
  - Múltiplas funções de perda: L_G, L_V, L_AH, L_reg, L_EWC

NOVO v2.0:
  - EWC (Elastic Weight Consolidation) para aprendizado contínuo
  - Context Window com cache KV (sliding window + sink)
  - RoPE (Rotary Position Embeddings)
  - TransformerBlock (CausalSelfAttention + FFN)
  - Self-attention final layer verdadeira (CLS-token + multi-head)
  - Weight tying opcional
  - Inferência autoregressiva REAL via GeneratorCNNBiGRU
"""
from __future__ import annotations

__version__ = "2.0.0"
__author__ = "CNN-BiGRU Project"

# Versão e metadados
__all__ = [
    "__version__",
    "__author__",
    # Tokenizer
    "BBPETokenizer",
    # Utils
    "MemoryOptimizer",
    "SemanticEmbedder",
    "EWCConfig",
    "EWCState",
    # Runtime
    "optimize_xeon_environment",
    "get_runtime_info",
    # Data
    "MultimodalStreamingDataset",
    "collate_multimodal",
    # Models
    "CooperativeCNNBiGRU",
    "MultimodalCNNBiGRU",
    "ImageEncoder",
    "AudioEncoder",
    "MultimodalFusion",
    "GeneratorCNNBiGRU",
    "VerifierCNNBiGRU",
    "AntiHallucinationLayer",
    "RotaryPositionEmbedding",
    "CausalSelfAttention",
    "TransformerBlock",
    "TransformerDecoderStack",
    "KVCache",
    "ContextWindowManager",
    "ContextWindowConfig",
    # Losses
    "LossConfig",
    "MultiLoss",
    # Training
    "TrainerConfig",
    "CooperativeTrainer",
    "AutoLearnConfig",
    "AutoLearner",
    "HypothesisConfig",
    "HypothesisController",
    "SynergySearcher",
    # Inference
    "generate_with_sampling",
    "evaluate_perplexity",
]


# Lazy imports — apenas quando acessado
def __getattr__(name: str):
    if name in ("BBPETokenizer",):
        from .tokenizer.bbpe_tokenizer import BBPETokenizer
        return BBPETokenizer
    if name in ("MemoryOptimizer",):
        from .utils.memory_optimizer import MemoryOptimizer
        return MemoryOptimizer
    if name in ("SemanticEmbedder",):
        from .utils.semantic_embeddings import SemanticEmbedder
        return SemanticEmbedder
    if name in ("EWCConfig", "EWCState"):
        from .utils.ewc import EWCConfig, EWCState
        if name == "EWCConfig":
            return EWCConfig
        return EWCState
    if name in ("optimize_xeon_environment", "get_runtime_info"):
        from .utils import xeon_runtime
        if name == "optimize_xeon_environment":
            return xeon_runtime.optimize_xeon_environment
        return xeon_runtime.get_runtime_info
    if name in ("MultimodalStreamingDataset", "collate_multimodal"):
        from .data.streaming_dataset import MultimodalStreamingDataset, collate_multimodal
        if name == "MultimodalStreamingDataset":
            return MultimodalStreamingDataset
        return collate_multimodal
    if name == "CooperativeCNNBiGRU":
        from .models.cooperative_bigru import CooperativeCNNBiGRU
        return CooperativeCNNBiGRU
    if name == "MultimodalCNNBiGRU":
        from .models.multimodal_model import MultimodalCNNBiGRU
        return MultimodalCNNBiGRU
    if name in ("ImageEncoder", "AudioEncoder", "MultimodalFusion"):
        from .models.multimodal_encoders import ImageEncoder, AudioEncoder, MultimodalFusion
        if name == "ImageEncoder":
            return ImageEncoder
        if name == "AudioEncoder":
            return AudioEncoder
        return MultimodalFusion
    if name in ("GeneratorCNNBiGRU", "VerifierCNNBiGRU", "AntiHallucinationLayer"):
        from .models.generator_verifier import (
            GeneratorCNNBiGRU,
            VerifierCNNBiGRU,
            AntiHallucinationLayer,
        )
        if name == "GeneratorCNNBiGRU":
            return GeneratorCNNBiGRU
        if name == "VerifierCNNBiGRU":
            return VerifierCNNBiGRU
        return AntiHallucinationLayer
    if name == "RotaryPositionEmbedding":
        from .models.rope import RotaryPositionEmbedding
        return RotaryPositionEmbedding
    if name in ("CausalSelfAttention", "TransformerBlock", "TransformerDecoderStack"):
        from .models.transformer_block import (
            CausalSelfAttention,
            TransformerBlock,
            TransformerDecoderStack,
        )
        if name == "CausalSelfAttention":
            return CausalSelfAttention
        if name == "TransformerBlock":
            return TransformerBlock
        return TransformerDecoderStack
    if name in ("KVCache", "ContextWindowManager", "ContextWindowConfig"):
        from .models.context_window import KVCache, ContextWindowManager, ContextWindowConfig
        if name == "KVCache":
            return KVCache
        if name == "ContextWindowManager":
            return ContextWindowManager
        return ContextWindowConfig
    if name in ("LossConfig", "MultiLoss"):
        from .losses.losses import LossConfig, MultiLoss
        if name == "LossConfig":
            return LossConfig
        return MultiLoss
    if name in ("TrainerConfig", "CooperativeTrainer"):
        from .training.trainer import TrainerConfig, CooperativeTrainer
        if name == "TrainerConfig":
            return TrainerConfig
        return CooperativeTrainer
    if name in ("AutoLearnConfig", "AutoLearner"):
        from .training.auto_learner import AutoLearnConfig, AutoLearner
        if name == "AutoLearnConfig":
            return AutoLearnConfig
        return AutoLearner
    if name in ("HypothesisConfig", "HypothesisController", "SynergySearcher"):
        from .training.hypothesis_controller import (
            HypothesisConfig,
            HypothesisController,
            SynergySearcher,
        )
        if name == "HypothesisConfig":
            return HypothesisConfig
        if name == "HypothesisController":
            return HypothesisController
        return SynergySearcher
    if name in ("generate_with_sampling", "evaluate_perplexity"):
        from .inference.inference import generate_with_sampling, evaluate_perplexity
        if name == "generate_with_sampling":
            return generate_with_sampling
        return evaluate_perplexity
    raise AttributeError(f"module {__name__!r} has no attribute {name!r}")