inference.py

#!/usr/bin/env python3
"""
ARC-8B: Adaptive Repetition Controller
=======================================
Decode-time behavioral control for language models.

This script loads the complete ARC system and runs inference with
multi-head cognitive control that detects and suppresses:
- Repetition loops (125× separation)
- Hedging phrases (1.5× separation)  
- Verbosity/filler (2.1× separation)
- Sycophancy (experimental)

Usage:
    python inference.py                          # Interactive mode
    python inference.py --prompt "Hello"         # Single prompt
    python inference.py --no-arc                 # Disable ARC (baseline)

Requirements:
    pip install torch transformers accelerate bitsandbytes

Model: LoganResearch/ARC-Base-8B (16GB, runs in ~10GB with 4-bit)
"""

import os
import sys
import argparse
import torch
import torch.nn as nn
import torch.nn.functional as F
from typing import Dict, List, Optional, Tuple
from dataclasses import dataclass


# =============================================================================
# CONFIGURATION
# =============================================================================

@dataclass
class ARCConfig:
    """ARC System Configuration"""
    # Model
    model_id: str = "LoganResearch/ARC-Base-8B"
    load_in_4bit: bool = True
    load_in_8bit: bool = False
    device_map: str = "auto"
    
    # Architecture (must match training)
    d_model: int = 4096
    n_layers: int = 32
    d_fiber: int = 16
    d_control: int = 64
    
    # Intervention thresholds (tuned empirically)
    repetition_threshold: float = 0.70
    hedging_threshold: float = 0.60
    verbosity_threshold: float = 0.65
    sycophancy_threshold: float = 0.60
    
    # Intervention penalties
    repetition_penalty: float = 5.0
    hedging_penalty: float = 3.0
    verbosity_penalty: float = 2.0
    sycophancy_penalty: float = 2.0
    
    # Generation
    max_new_tokens: int = 512
    temperature: float = 0.8
    top_p: float = 0.92
    repetition_window: int = 32


# =============================================================================
# MULTI-HEAD PREDICTOR
# =============================================================================

class MultiHeadPredictor(nn.Module):
    """
    Prediction heads that monitor hidden states and detect behavioral patterns.
    
    The system uses shared "fiber projections" that compress hidden states,
    then individual heads that predict risk scores for specific behaviors.
    
    Architecture:
        Hidden States [n_layers × d_model] 
            → Fiber Projections [n_layers × d_fiber]
            → Weighted Aggregation [d_fiber]
            → Per-Head MLP → Risk Score [0-1]
    """
    
    def __init__(self, config: ARCConfig):
        super().__init__()
        self.config = config
        
        # Shared fiber projections (learned during CF-HoT training)
        self.fiber_projs = nn.ModuleList([
            nn.Linear(config.d_model, config.d_fiber, bias=False) 
            for _ in range(config.n_layers)
        ])
        
        # Learned layer importance weights
        self.layer_weights = nn.Parameter(torch.ones(config.n_layers) / config.n_layers)
        
        # Individual prediction heads
        self.heads = nn.ModuleDict()
        self.loaded_heads: set = set()
    
    def _make_head(self) -> nn.Sequential:
        """Create a prediction head: fiber features → risk score"""
        return nn.Sequential(
            nn.Linear(self.config.d_fiber, self.config.d_control),
            nn.GELU(),
            nn.Linear(self.config.d_control, self.config.d_control),
            nn.GELU(),
            nn.Linear(self.config.d_control, 1)
        )
    
    def add_head(self, name: str) -> None:
        """Add a new prediction head"""
        self.heads[name] = self._make_head()
    
    def get_fiber_features(self, hidden_states: List[torch.Tensor]) -> torch.Tensor:
        """
        Project hidden states through fiber projections and aggregate.
        
        Args:
            hidden_states: List of [batch, seq, d_model] tensors from each layer
            
        Returns:
            Aggregated features [batch, seq, d_fiber]
        """
        device = hidden_states[0].device
        fibers = []
        for i, (proj, hidden) in enumerate(zip(self.fiber_projs, hidden_states)):
            if i < len(hidden_states):
                proj = proj.to(device)
                fibers.append(proj(hidden.float()))
        
        # Weighted sum across layers
        weights = F.softmax(self.layer_weights.to(device)[:len(fibers)], dim=0)
        aggregated = sum(w * f for w, f in zip(weights, fibers))
        return aggregated
    
    def get_risk(self, head_name: str, hidden_states: List[torch.Tensor]) -> torch.Tensor:
        """Get risk score from a specific head"""
        if head_name not in self.loaded_heads:
            return torch.zeros(1, device=hidden_states[0].device)
        
        features = self.get_fiber_features(hidden_states)
        logits = self.heads[head_name](features).squeeze(-1)
        return torch.sigmoid(logits)
    
    def get_all_risks(self, hidden_states: List[torch.Tensor]) -> Dict[str, torch.Tensor]:
        """Get risk scores from all loaded heads"""
        if not self.loaded_heads:
            return {}
        
        device = hidden_states[0].device
        features = self.get_fiber_features(hidden_states)
        risks = {}
        for name in self.loaded_heads:
            self.heads[name] = self.heads[name].to(device)
            logits = self.heads[name](features).squeeze(-1)
            risks[name] = torch.sigmoid(logits)
        return risks


# =============================================================================
# ARC SYSTEM
# =============================================================================

class ARCSystem:
    """
    Complete ARC (Adaptive Repetition Controller) System
    
    Loads model + prediction heads and provides controlled generation
    with real-time behavioral intervention.
    """
    
    # Tokens to suppress for each behavior type
    HEDGE_STARTERS = [
        "As", "I'm", "I", "It's", "While", "Although", "However",
        "That", "This", "Please", "Well", "So", "Actually"
    ]
    VERBOSE_STARTERS = [
        "Let", "Basically", "Essentially", "Simply", "Indeed",
        "Furthermore", "Moreover", "Additionally", "Firstly"
    ]
    SYCOPHANCY_STARTERS = [
        "Great", "Excellent", "Wonderful", "Absolutely", "Of",
        "Thank", "Sure", "Certainly", "Definitely"
    ]
    
    def __init__(self, config: Optional[ARCConfig] = None):
        self.config = config or ARCConfig()
        
        self.model = None
        self.tokenizer = None
        self.predictor = None
        
        # Token ID caches for suppression
        self._hedge_token_ids: set = set()
        self._verbose_token_ids: set = set()
        self._sycophancy_token_ids: set = set()
        
        # Stats
        self.total_interventions = {"repetition": 0, "hedging": 0, "verbosity": 0, "sycophancy": 0}
    
    def load(self, verbose: bool = True) -> "ARCSystem":
        """
        Load all components from HuggingFace.
        
        Downloads and initializes:
        1. Base model (Hermes-3-Llama-3.1-8B based)
        2. Tokenizer
        3. Prediction heads (repetition, hedging, verbosity, sycophancy)
        
        Returns:
            self (for chaining)
        """
        from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
        from huggingface_hub import hf_hub_download
        
        if verbose:
            print("=" * 60)
            print("  ARC-8B: Adaptive Repetition Controller")
            print("  Decode-time behavioral control system")
            print("=" * 60)
        
        # === 1. Tokenizer ===
        if verbose:
            print("\n[1/4] Loading tokenizer...")
        self.tokenizer = AutoTokenizer.from_pretrained(
            self.config.model_id,
            trust_remote_code=True
        )
        if self.tokenizer.pad_token is None:
            self.tokenizer.pad_token = self.tokenizer.eos_token
        
        # === 2. Model ===
        if verbose:
            print("[2/4] Loading model...")
            if self.config.load_in_4bit:
                print("       (4-bit quantization enabled)")
        
        quantization_config = None
        if self.config.load_in_4bit:
            quantization_config = BitsAndBytesConfig(
                load_in_4bit=True,
                bnb_4bit_compute_dtype=torch.float16,
                bnb_4bit_use_double_quant=True,
                bnb_4bit_quant_type="nf4"
            )
        elif self.config.load_in_8bit:
            quantization_config = BitsAndBytesConfig(load_in_8bit=True)
        
        self.model = AutoModelForCausalLM.from_pretrained(
            self.config.model_id,
            quantization_config=quantization_config,
            device_map=self.config.device_map,
            torch_dtype=torch.float16,
            trust_remote_code=True
        )
        self.model.eval()
        
        # === 3. Prediction Heads ===
        if verbose:
            print("[3/4] Loading prediction heads...")
        
        device = next(self.model.parameters()).device
        self.predictor = MultiHeadPredictor(self.config).to(device).float()
        
        # Load risk_predictor.pt (contains fiber projections + repetition head)
        try:
            risk_path = hf_hub_download(self.config.model_id, "risk_predictor.pt")
            ckpt = torch.load(risk_path, map_location=device, weights_only=False)
            
            # The checkpoint contains the full state dict
            state = ckpt.get('risk_predictor', ckpt)
            
            # Load fiber projections
            for i in range(self.config.n_layers):
                key = f'fiber_projs.{i}.weight'
                if key in state:
                    self.predictor.fiber_projs[i].weight.data = state[key].to(device).float()
            
            # Load layer weights
            if 'layer_weights' in state:
                self.predictor.layer_weights.data = state['layer_weights'].to(device).float()
            
            # Load repetition head
            self.predictor.add_head('repetition')
            self.predictor.heads['repetition'][0].weight.data = state['predictor.0.weight'].to(device).float()
            self.predictor.heads['repetition'][0].bias.data = state['predictor.0.bias'].to(device).float()
            self.predictor.heads['repetition'][2].weight.data = state['predictor.2.weight'].to(device).float()
            self.predictor.heads['repetition'][2].bias.data = state['predictor.2.bias'].to(device).float()
            self.predictor.heads['repetition'][4].weight.data = state['predictor.4.weight'].to(device).float()
            self.predictor.heads['repetition'][4].bias.data = state['predictor.4.bias'].to(device).float()
            self.predictor.loaded_heads.add('repetition')
            
            if verbose:
                print("       ✓ Repetition head (125× separation)")
        except Exception as e:
            if verbose:
                print(f"       ✗ Repetition head: {e}")
        
        # Load additional heads
        for head_name in ['hedging', 'verbosity', 'sycophancy']:
            try:
                head_path = hf_hub_download(self.config.model_id, f"{head_name}_head.pt")
                ckpt = torch.load(head_path, map_location=device, weights_only=False)
                
                self.predictor.add_head(head_name)
                head_state = ckpt.get('head_state', ckpt)
                self.predictor.heads[head_name].load_state_dict(head_state)
                self.predictor.loaded_heads.add(head_name)
                
                if verbose:
                    print(f"       ✓ {head_name.capitalize()} head")
            except Exception as e:
                if verbose:
                    print(f"       ✗ {head_name.capitalize()} head: {e}")
        
        self.predictor.eval()
        
        # === 4. Build Token Suppression Sets ===
        if verbose:
            print("[4/4] Building suppression vocabularies...")
        
        self._build_suppression_sets()
        
        if verbose:
            print("\n" + "=" * 60)
            print(f"  ✓ ARC System Ready")
            print(f"  Active heads: {list(self.predictor.loaded_heads)}")
            print("=" * 60 + "\n")
        
        return self
    
    def _build_suppression_sets(self) -> None:
        """Build token ID sets for behavioral suppression"""
        for word in self.HEDGE_STARTERS:
            tokens = self.tokenizer.encode(word, add_special_tokens=False)
            if tokens:
                self._hedge_token_ids.add(tokens[0])
        
        for word in self.VERBOSE_STARTERS:
            tokens = self.tokenizer.encode(word, add_special_tokens=False)
            if tokens:
                self._verbose_token_ids.add(tokens[0])
        
        for word in self.SYCOPHANCY_STARTERS:
            tokens = self.tokenizer.encode(word, add_special_tokens=False)
            if tokens:
                self._sycophancy_token_ids.add(tokens[0])
    
    def _apply_interventions(
        self,
        logits: torch.Tensor,
        risks: Dict[str, torch.Tensor],
        recent_tokens: List[int]
    ) -> Tuple[torch.Tensor, Dict[str, bool]]:
        """
        Apply behavioral interventions based on risk scores.
        
        Args:
            logits: [1, vocab_size] logits for next token
            risks: Dict of risk scores for each head
            recent_tokens: Recently generated token IDs
            
        Returns:
            Modified logits and dict of which interventions fired
        """
        interventions = {}
        
        # Repetition: suppress recently used tokens
        if risks.get('repetition', 0) > self.config.repetition_threshold:
            for tok in set(recent_tokens[-self.config.repetition_window:]):
                logits[0, tok] -= self.config.repetition_penalty
            interventions['repetition'] = True
            self.total_interventions['repetition'] += 1
        
        # Hedging: suppress hedge phrase starters
        if risks.get('hedging', 0) > self.config.hedging_threshold:
            for tok in self._hedge_token_ids:
                logits[0, tok] -= self.config.hedging_penalty
            interventions['hedging'] = True
            self.total_interventions['hedging'] += 1
        
        # Verbosity: suppress filler phrase starters
        if risks.get('verbosity', 0) > self.config.verbosity_threshold:
            for tok in self._verbose_token_ids:
                logits[0, tok] -= self.config.verbosity_penalty
            interventions['verbosity'] = True
            self.total_interventions['verbosity'] += 1
        
        # Sycophancy: suppress sycophantic starters
        if risks.get('sycophancy', 0) > self.config.sycophancy_threshold:
            for tok in self._sycophancy_token_ids:
                logits[0, tok] -= self.config.sycophancy_penalty
            interventions['sycophancy'] = True
            self.total_interventions['sycophancy'] += 1
        
        return logits, interventions
    
    def generate(
        self,
        prompt: str,
        system_prompt: Optional[str] = None,
        max_new_tokens: Optional[int] = None,
        temperature: Optional[float] = None,
        use_arc: bool = True,
        verbose: bool = False
    ) -> str:
        """
        Generate text with optional ARC behavioral control.
        
        Args:
            prompt: User input
            system_prompt: Optional system message
            max_new_tokens: Max tokens to generate (default: config value)
            temperature: Sampling temperature (default: config value)
            use_arc: Whether to use ARC intervention (default: True)
            verbose: Print intervention info (default: False)
            
        Returns:
            Generated text
        """
        max_new_tokens = max_new_tokens or self.config.max_new_tokens
        temperature = temperature or self.config.temperature
        
        # Build chat format
        if system_prompt is None:
            system_prompt = "You are a helpful assistant."
        
        full_prompt = f"<|im_start|>system\n{system_prompt}<|im_end|>\n"
        full_prompt += f"<|im_start|>user\n{prompt}<|im_end|>\n"
        full_prompt += "<|im_start|>assistant\n"
        
        device = next(self.model.parameters()).device
        input_ids = self.tokenizer.encode(full_prompt, return_tensors='pt').to(device)
        attention_mask = torch.ones_like(input_ids)
        
        generated_ids = input_ids.clone()
        intervention_counts = {"repetition": 0, "hedging": 0, "verbosity": 0, "sycophancy": 0}
        
        # Generation loop
        for step in range(max_new_tokens):
            with torch.no_grad():
                outputs = self.model(
                    input_ids=generated_ids,
                    attention_mask=attention_mask,
                    output_hidden_states=True,
                    return_dict=True
                )
            
            logits = outputs.logits[:, -1, :] / temperature
            
            # ARC intervention
            if use_arc and self.predictor.loaded_heads:
                hidden_states = outputs.hidden_states[1:]  # Skip embedding layer
                risks = self.predictor.get_all_risks(hidden_states)
                current_risks = {name: r[:, -1].item() for name, r in risks.items()}
                
                recent = generated_ids[0, -self.config.repetition_window:].tolist()
                logits, fired = self._apply_interventions(logits, current_risks, recent)
                
                for k, v in fired.items():
                    if v:
                        intervention_counts[k] += 1
            
            # Top-p sampling
            sorted_logits, sorted_indices = torch.sort(logits, descending=True)
            cumulative_probs = torch.cumsum(F.softmax(sorted_logits, dim=-1), dim=-1)
            sorted_indices_to_remove = cumulative_probs > self.config.top_p
            sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone()
            sorted_indices_to_remove[..., 0] = 0
            indices_to_remove = sorted_indices_to_remove.scatter(1, sorted_indices, sorted_indices_to_remove)
            logits[indices_to_remove] = float('-inf')
            
            probs = F.softmax(logits, dim=-1)
            next_token = torch.multinomial(probs, num_samples=1)
            
            generated_ids = torch.cat([generated_ids, next_token], dim=-1)
            attention_mask = torch.cat([attention_mask, torch.ones(1, 1, device=device)], dim=-1)
            
            # Check for EOS
            if next_token.item() == self.tokenizer.eos_token_id:
                break
            
            # Check for end of turn
            if next_token.item() == self.tokenizer.encode("<|im_end|>", add_special_tokens=False)[0]:
                break
        
        # Decode response
        full_output = self.tokenizer.decode(generated_ids[0], skip_special_tokens=False)
        
        # Extract assistant response
        if "<|im_start|>assistant\n" in full_output:
            response = full_output.split("<|im_start|>assistant\n")[-1]
            if "<|im_end|>" in response:
                response = response.split("<|im_end|>")[0]
        else:
            response = full_output
        
        if verbose:
            total = sum(intervention_counts.values())
            print(f"\n[ARC Stats] Interventions: {total} total")
            for k, v in intervention_counts.items():
                if v > 0:
                    print(f"  - {k}: {v}")
        
        return response.strip()
    
    def chat(self, system_prompt: Optional[str] = None) -> None:
        """
        Interactive chat mode.
        
        Args:
            system_prompt: Optional system message
        """
        print("\n" + "=" * 60)
        print("  ARC-8B Interactive Chat")
        print("  Commands: /quit, /stats, /arc on|off, /clear")
        print("=" * 60 + "\n")
        
        use_arc = True
        history = []
        
        while True:
            try:
                user_input = input("You: ").strip()
            except (KeyboardInterrupt, EOFError):
                print("\nGoodbye!")
                break
            
            if not user_input:
                continue
            
            # Commands
            if user_input.lower() == '/quit':
                print("Goodbye!")
                break
            elif user_input.lower() == '/stats':
                print(f"\nTotal interventions: {self.total_interventions}\n")
                continue
            elif user_input.lower() == '/arc on':
                use_arc = True
                print("ARC enabled\n")
                continue
            elif user_input.lower() == '/arc off':
                use_arc = False
                print("ARC disabled (baseline mode)\n")
                continue
            elif user_input.lower() == '/clear':
                history = []
                self.total_interventions = {k: 0 for k in self.total_interventions}
                print("History cleared\n")
                continue
            
            # Generate response
            response = self.generate(
                user_input,
                system_prompt=system_prompt,
                use_arc=use_arc,
                verbose=True
            )
            
            print(f"\nAssistant: {response}\n")
            history.append({"user": user_input, "assistant": response})


# =============================================================================
# MAIN
# =============================================================================

def main():
    parser = argparse.ArgumentParser(
        description="ARC-8B: Adaptive Repetition Controller",
        formatter_class=argparse.RawDescriptionHelpFormatter,
        epilog="""
Examples:
    python inference.py                     # Interactive chat
    python inference.py --prompt "Hello"    # Single prompt
    python inference.py --no-arc            # Disable ARC (baseline)
    python inference.py --8bit              # Use 8-bit quantization
        """
    )
    parser.add_argument("--prompt", "-p", type=str, help="Single prompt to process")
    parser.add_argument("--system", "-s", type=str, help="System prompt")
    parser.add_argument("--no-arc", action="store_true", help="Disable ARC intervention")
    parser.add_argument("--4bit", dest="load_4bit", action="store_true", default=True, help="Use 4-bit quantization (default)")
    parser.add_argument("--8bit", dest="load_8bit", action="store_true", help="Use 8-bit quantization")
    parser.add_argument("--no-quant", action="store_true", help="Disable quantization (requires ~32GB VRAM)")
    parser.add_argument("--max-tokens", type=int, default=512, help="Max tokens to generate")
    parser.add_argument("--temperature", type=float, default=0.8, help="Sampling temperature")
    
    args = parser.parse_args()
    
    # Configure
    config = ARCConfig(
        max_new_tokens=args.max_tokens,
        temperature=args.temperature
    )
    
    if args.load_8bit:
        config.load_in_4bit = False
        config.load_in_8bit = True
    elif args.no_quant:
        config.load_in_4bit = False
        config.load_in_8bit = False
    
    # Load
    arc = ARCSystem(config)
    arc.load()
    
    # Run
    if args.prompt:
        response = arc.generate(
            args.prompt,
            system_prompt=args.system,
            use_arc=not args.no_arc,
            verbose=True
        )
        print(f"\n{response}\n")
    else:
        arc.chat(system_prompt=args.system)


if __name__ == "__main__":
    main()