cnn_bigru/__init__.py

"""
CNN-BiGRU — Modelo Multimodal Cooperativo Autoaprendível (v3.0)

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

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

NOVO v3.0:
  - Cyclic Reasoning (Raciocínio Cíclico) com convergência antecipada
  - NLG (Natural Language Generation) com Transformer Decoder + Medusa MTP
  - NLP (Natural Language Processing) com 4 tarefas: SeqCls, TokCls, Span, Embed
  - W8A8 Quantization via SmoothQuant (alpha=0.5, per-channel)
  - VQ-VAE-2 Hierárquico (top + bottom codebooks, EMA update)
  - Multi-Token Prediction (MTP) com Medusa Heads
  - Multimodal Multi-Head Attention (cross-modal + modality gate)
  - Long Context Window (1M tokens) via chunked / ring attention
  - Monitor completo (treino, inferência, evolução, sistema)
"""
from __future__ import annotations

__version__ = "3.0.0"
__author__ = "CNN-BiGRU Project"

# Versão e metadados
__all__ = [
    "__version__",
    "__author__",
    # Tokenizer
    "BBPETokenizer",
    # Utils
    "MemoryOptimizer",
    "SemanticEmbedder",
    "EWCConfig",
    "EWCState",
    "W8A8Config",
    "SmoothQuantizer",
    "quantize_model_w8a8",
    "estimate_memory_savings",
    "VQVAE2Config",
    "VQVAE2",
    "Monitor",
    "get_monitor",
    # 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",
    "LongContextConfig",
    "LongContextManager",
    "make_long_context_window",
    "CyclicReasoningConfig",
    "CyclicReasoning",
    "MedusaConfig",
    "MedusaMTP",
    "MedusaHead",
    "medusa_tree_decode",
    "NLGConfig",
    "NLGModule",
    "NLPConfig",
    "NLPModule",
    "SequenceClassificationHead",
    "TokenClassificationHead",
    "SpanDetectionHead",
    "EmbeddingHead",
    "MultimodalAttentionConfig",
    "MultimodalMultiHeadAttention",
    "CrossModalAttention",
    "ModalityGate",
    # 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
    # W8A8 Quantization
    if name in ("W8A8Config", "SmoothQuantizer"):
        from .utils.quantization import W8A8Config, SmoothQuantizer
        if name == "W8A8Config":
            return W8A8Config
        return SmoothQuantizer
    if name == "quantize_model_w8a8":
        from .utils.quantization import quantize_model_w8a8
        return quantize_model_w8a8
    if name == "estimate_memory_savings":
        from .utils.quantization import estimate_memory_savings
        return estimate_memory_savings
    # VQ-VAE-2
    if name in ("VQVAE2Config", "VQVAE2"):
        from .utils.vqvae2 import VQVAE2Config, VQVAE2
        if name == "VQVAE2Config":
            return VQVAE2Config
        return VQVAE2
    # Monitor
    if name in ("Monitor", "get_monitor"):
        from .utils.monitoring import Monitor, get_monitor
        if name == "Monitor":
            return Monitor
        return get_monitor
    # Runtime
    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
    # Data
    if name in ("MultimodalStreamingDataset", "collate_multimodal"):
        from .data.streaming_dataset import MultimodalStreamingDataset, collate_multimodal
        if name == "MultimodalStreamingDataset":
            return MultimodalStreamingDataset
        return collate_multimodal
    # Models
    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
    # Context window (incl. 1M tokens)
    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 ("LongContextConfig", "LongContextManager", "make_long_context_window"):
        from .models.context_window import (
            LongContextConfig,
            LongContextManager,
            make_long_context_window,
        )
        if name == "LongContextConfig":
            return LongContextConfig
        if name == "LongContextManager":
            return LongContextManager
        return make_long_context_window
    # Cyclic Reasoning
    if name in ("CyclicReasoningConfig", "CyclicReasoning"):
        from .models.cyclic_reasoning import CyclicReasoningConfig, CyclicReasoning
        if name == "CyclicReasoningConfig":
            return CyclicReasoningConfig
        return CyclicReasoning
    # Medusa MTP
    if name in ("MedusaConfig", "MedusaMTP", "MedusaHead", "medusa_tree_decode"):
        from .models.medusa_heads import (
            MedusaConfig,
            MedusaMTP,
            MedusaHead,
            medusa_tree_decode,
        )
        if name == "MedusaConfig":
            return MedusaConfig
        if name == "MedusaMTP":
            return MedusaMTP
        if name == "MedusaHead":
            return MedusaHead
        return medusa_tree_decode
    # NLG
    if name in ("NLGConfig", "NLGModule"):
        from .models.nlg import NLGConfig, NLGModule
        if name == "NLGConfig":
            return NLGConfig
        return NLGModule
    # NLP
    if name in ("NLPConfig", "NLPModule",
                "SequenceClassificationHead", "TokenClassificationHead",
                "SpanDetectionHead", "EmbeddingHead"):
        from .models.nlp import (
            NLPConfig,
            NLPModule,
            SequenceClassificationHead,
            TokenClassificationHead,
            SpanDetectionHead,
            EmbeddingHead,
        )
        if name == "NLPConfig":
            return NLPConfig
        if name == "NLPModule":
            return NLPModule
        if name == "SequenceClassificationHead":
            return SequenceClassificationHead
        if name == "TokenClassificationHead":
            return TokenClassificationHead
        if name == "SpanDetectionHead":
            return SpanDetectionHead
        return EmbeddingHead
    # Multimodal Attention
    if name in ("MultimodalAttentionConfig", "MultimodalMultiHeadAttention",
                "CrossModalAttention", "ModalityGate"):
        from .models.multimodal_attention import (
            MultimodalAttentionConfig,
            MultimodalMultiHeadAttention,
            CrossModalAttention,
            ModalityGate,
        )
        if name == "MultimodalAttentionConfig":
            return MultimodalAttentionConfig
        if name == "MultimodalMultiHeadAttention":
            return MultimodalMultiHeadAttention
        if name == "CrossModalAttention":
            return CrossModalAttention
        return ModalityGate
    # Losses
    if name in ("LossConfig", "MultiLoss"):
        from .losses.losses import LossConfig, MultiLoss
        if name == "LossConfig":
            return LossConfig
        return MultiLoss
    # Training
    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
    # Inference
    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}")