kernel/config_arc.py

"""
GLADIUS v2.1 — ARC-Ready Kernel Configuration

Extends KernelConfig with ARC competition settings.
"""

import torch
from dataclasses import dataclass, field


@dataclass
class ArcKernelConfig:
    """
    Configuration for GLADIUS kernel in ARC competition mode.
    
    Inherits all Wyrm settings and adds:
    - Expanded vocab (grid tokens)
    - Grid tool activation
    - Specialist residual scaling
    - Context window options
    """
    # === Base Wyrm Settings (from checkpoint) ===
    vocab_size: int = 16022          # 16000 BPE + 22 grid tokens
    hidden_dim: int = 640
    num_layers: int = 14
    num_heads: int = 20
    head_dim: int = 32
    ffn_dim: int = 2560
    max_seq_len: int = 1024          # Start at 1024, can scale to 2048/4096
    
    # Memory
    hot_memory_slots: int = 512
    hot_importance_threshold: float = 0.5
    warm_rank: int = 32
    warm_condition_threshold: float = 10.0
    warm_balance_frequency: int = 100
    warm_novelty_threshold: float = 0.1
    warm_checkpoint_interval: int = 300
    cold_embedding_dim: int = 640
    cold_top_k: int = 4
    
    # Time
    time_dim: int = 64
    time_num_frequencies: int = 16
    time_max_events: int = 64
    clock_mode: str = 'continuous'
    lattice_size: int = 256
    lattice_scales: int = 4
    
    # Cognition
    cognition_state_dim: int = 128
    cognition_modes: int = 4
    cognition_prompt_types: int = 5
    
    # Modulator
    register_dim: int = 4
    intent_dim: int = 4
    silence_threshold: float = 0.7
    
    # Tools
    max_tools: int = 32
    tool_activation_threshold: float = 0.6
    
    # Router + Specialists
    num_specialists: int = 4
    router_top_k: int = 2
    
    # Attention
    attention_sparse_budget: int = 64
    attention_alpha_init: float = 0.5
    
    # Training
    learning_rate: float = 3e-4
    weight_decay: float = 0.01
    warmup_steps: int = 500
    max_grad_norm: float = 1.0
    batch_size: int = 4
    accumulation_steps: int = 8
    
    # Device
    device: str = 'cpu'
    dtype: torch.dtype = torch.float32
    
    # Paths
    checkpoint_dir: str = 'checkpoints'
    seed: int = 42
    
    # Special tokens
    pad_token_id: int = 0
    bos_token_id: int = 1
    eos_token_id: int = 2
    unk_token_id: int = 3
    
    # === NEW: ARC-Specific Settings ===
    grid_tools_enabled: bool = True
    specialist_residual_scale: float = 0.1    # Start small, increase as specialists learn
    
    # Multi-task training LR groups
    backbone_lr: float = 3e-5       # Low — preserve backbone
    specialist_lr: float = 3e-4     # Medium — learn fast
    router_lr: float = 1e-3         # High — learn routing quickly
    tool_lr: float = 3e-4           # Medium — learn tool activation
    
    # Task mix ratios
    english_ratio: float = 0.6
    grid_ratio: float = 0.2
    program_ratio: float = 0.1
    tool_ratio: float = 0.1


def wyrm_to_arc_config(wyrm_config) -> ArcKernelConfig:
    """
    Convert a Wyrm KernelConfig to ARC-ready config.
    Copies all matching fields, adds ARC-specific ones.
    """
    arc = ArcKernelConfig()
    
    if hasattr(wyrm_config, '__dict__'):
        for key, value in vars(wyrm_config).items():
            if hasattr(arc, key) and key != 'vocab_size':  # Don't override expanded vocab
                setattr(arc, key, value)
    
    # Force ARC settings
    arc.vocab_size = 16022
    arc.grid_tools_enabled = True
    arc.specialist_residual_scale = 0.1
    
    return arc