""" Inference-time scaling mechanisms for step-level reasoning. Implements best-of-N sampling, majority voting, and self-consistency to improve reasoning quality through test-time compute. """ from typing import Dict, List, Optional, Tuple, Any, Callable import torch import torch.nn as nn from collections import Counter import logging from dataclasses import dataclass from .prm import ProcessRewardModel from .step_data import ReasoningChain, ReasoningStep logger = logging.getLogger(__name__) @dataclass class InferenceConfig: """Configuration for inference-time scaling.""" num_samples: int = 8 # Number of reasoning chains to generate temperature: float = 0.7 # Sampling temperature top_k: int = 50 top_p: float = 0.95 max_steps: int = 10 # Maximum reasoning steps # Aggregation method aggregation: str = "best_of_n" # Options: best_of_n, majority_vote, weighted_vote # PRM-based selection use_prm_scores: bool = True min_prm_score: float = 0.3 # Minimum acceptable PRM score # Self-consistency consistency_threshold: float = 0.6 # Required agreement for majority vote class InferenceTimeScaling: """ Inference-time scaling for improved reasoning quality. Generates multiple reasoning chains and selects the best one using various aggregation strategies. """ def __init__( self, model: nn.Module, prm: Optional[ProcessRewardModel] = None, config: Optional[InferenceConfig] = None, device: str = "cuda", ): """ Initialize inference scaler. Args: model: Vision-language model prm: Process Reward Model for evaluation config: Inference configuration device: Device for inference """ self.model = model self.prm = prm self.config = config or InferenceConfig() self.device = device # Move to device only if not already quantized with device_map if not (hasattr(self.model, 'hf_device_map') or getattr(self.model, 'is_quantized', False)): self.model.to(device) self.model.eval() if self.prm: if not (hasattr(self.prm, 'hf_device_map') or getattr(self.prm, 'is_quantized', False)): self.prm.to(device) self.prm.eval() @torch.no_grad() def generate_multiple_chains( self, image_features: torch.Tensor, prompt_ids: torch.Tensor, prompt_mask: torch.Tensor, image_path: str = "", prompt_text: str = "", ) -> List[ReasoningChain]: """ Generate multiple reasoning chains with different samples. Args: image_features: Visual features prompt_ids: Tokenized prompt prompt_mask: Attention mask image_path: Path to image (for metadata) prompt_text: Original prompt text Returns: List of reasoning chains """ chains = [] for sample_idx in range(self.config.num_samples): steps = [] current_ids = prompt_ids.unsqueeze(0).to(self.device) current_mask = prompt_mask.unsqueeze(0).to(self.device) step_embeddings = [] for step_id in range(self.config.max_steps): # Generate next step outputs = self.model( input_ids=current_ids, attention_mask=current_mask, output_hidden_states=True, ) logits = outputs.logits[:, -1, :] / self.config.temperature probs = torch.softmax(logits, dim=-1) # Top-k sampling if self.config.top_k > 0: top_k_probs, top_k_indices = torch.topk(probs, self.config.top_k) next_token = top_k_indices[0, torch.multinomial(top_k_probs[0], 1)] else: next_token = torch.multinomial(probs[0], 1) # Get hidden state for this step hidden_state = outputs.hidden_states[-1][:, -1, :].squeeze(0) step_embeddings.append(hidden_state) # Compute PRM reward if available reward = 0.0 if self.prm: reward, metrics = self.prm.compute_step_reward( image_features, hidden_state, previous_steps=torch.stack(step_embeddings[:-1]) if len(step_embeddings) > 1 else None, return_metrics=True, ) # Decode step text step_text = self.model.config.tokenizer.decode( next_token.unsqueeze(0), skip_special_tokens=True, ) # Create reasoning step step = ReasoningStep( step_id=step_id, step_type="inference", # Would be classified in practice description=step_text, visual_features=hidden_state.cpu(), confidence=metrics.confidence if self.prm and metrics else 0.5, reward=reward, ) steps.append(step) # Update sequence current_ids = torch.cat([current_ids, next_token.unsqueeze(0).unsqueeze(0)], dim=1) current_mask = torch.cat([ current_mask, torch.ones((1, 1), device=self.device) ], dim=1) # Check for completion if next_token.item() == self.model.config.eos_token_id: break # Create reasoning chain final_answer = self.model.config.tokenizer.decode( current_ids[0, len(prompt_ids):], skip_special_tokens=True, ) total_reward = sum(s.reward for s in steps) chain = ReasoningChain( chain_id=f"sample_{sample_idx}", image_path=image_path, prompt=prompt_text, steps=steps, final_answer=final_answer, total_reward=total_reward, ) chains.append(chain) return chains def best_of_n_selection( self, chains: List[ReasoningChain], ) -> ReasoningChain: """ Select best chain based on PRM scores. Args: chains: List of reasoning chains Returns: Chain with highest total reward """ if not chains: raise ValueError("No chains to select from") # Filter by minimum PRM score if required if self.config.use_prm_scores and self.prm: filtered_chains = [ c for c in chains if c.total_reward >= self.config.min_prm_score * len(c.steps) ] if filtered_chains: chains = filtered_chains # Select chain with highest total reward best_chain = max(chains, key=lambda c: c.total_reward) logger.info( f"Best-of-N: Selected chain with reward {best_chain.total_reward:.3f} " f"({len(best_chain.steps)} steps)" ) return best_chain def majority_voting( self, chains: List[ReasoningChain], ) -> ReasoningChain: """ Select answer by majority voting with self-consistency. Args: chains: List of reasoning chains Returns: Chain corresponding to majority answer """ if not chains: raise ValueError("No chains to vote on") # Extract final answers answers = [c.final_answer.strip().lower() for c in chains] # Count votes vote_counts = Counter(answers) majority_answer, majority_count = vote_counts.most_common(1)[0] # Check consistency threshold consistency = majority_count / len(chains) if consistency < self.config.consistency_threshold: logger.warning( f"Low consistency: {consistency:.2f} < {self.config.consistency_threshold:.2f}" ) # Find chain with majority answer and highest PRM score majority_chains = [ c for c in chains if c.final_answer.strip().lower() == majority_answer ] best_chain = max(majority_chains, key=lambda c: c.total_reward) logger.info( f"Majority Vote: '{majority_answer}' ({majority_count}/{len(chains)}) " f"with consistency {consistency:.2f}" ) return best_chain def weighted_voting( self, chains: List[ReasoningChain], ) -> ReasoningChain: """ Weighted voting using PRM scores as weights. Args: chains: List of reasoning chains Returns: Chain with highest weighted vote """ if not chains or not self.prm: return self.best_of_n_selection(chains) # Group chains by answer answer_groups: Dict[str, List[ReasoningChain]] = {} for chain in chains: answer_key = chain.final_answer.strip().lower() if answer_key not in answer_groups: answer_groups[answer_key] = [] answer_groups[answer_key].append(chain) # Calculate weighted votes weighted_votes = {} for answer, group_chains in answer_groups.items(): # Use exponential of reward as weight (emphasizes high-quality chains) weight = sum(torch.exp(torch.tensor(c.total_reward / len(c.steps))).item() for c in group_chains) weighted_votes[answer] = weight # Select answer with highest weighted vote best_answer = max(weighted_votes.items(), key=lambda x: x[1])[0] # Return best chain for that answer best_chains = answer_groups[best_answer] best_chain = max(best_chains, key=lambda c: c.total_reward) logger.info( f"Weighted Vote: '{best_answer}' with weight {weighted_votes[best_answer]:.3f}" ) return best_chain def __call__( self, image_features: torch.Tensor, prompt_ids: torch.Tensor, prompt_mask: torch.Tensor, image_path: str = "", prompt_text: str = "", ) -> Tuple[ReasoningChain, List[ReasoningChain]]: """ Perform inference with scaling. Args: image_features: Visual features prompt_ids: Tokenized prompt prompt_mask: Attention mask image_path: Path to image prompt_text: Original prompt Returns: - best_chain: Selected reasoning chain - all_chains: All generated chains for analysis """ # Generate multiple chains chains = self.generate_multiple_chains( image_features, prompt_ids, prompt_mask, image_path, prompt_text, ) # Select best chain based on aggregation method if self.config.aggregation == "best_of_n": best_chain = self.best_of_n_selection(chains) elif self.config.aggregation == "majority_vote": best_chain = self.majority_voting(chains) elif self.config.aggregation == "weighted_vote": best_chain = self.weighted_voting(chains) else: raise ValueError(f"Unknown aggregation method: {self.config.aggregation}") return best_chain, chains class MajorityVoting: """ Standalone majority voting for answer aggregation. Useful for post-processing multiple model predictions. """ @staticmethod def vote( answers: List[str], weights: Optional[List[float]] = None, threshold: float = 0.5, ) -> Tuple[str, float]: """ Perform majority voting on answers. Args: answers: List of answer strings weights: Optional weights for each answer threshold: Minimum agreement threshold Returns: - majority_answer: Most common answer - confidence: Voting confidence [0, 1] """ if not answers: return "", 0.0 if weights is None: weights = [1.0] * len(answers) # Normalize answers normalized = [a.strip().lower() for a in answers] # Weighted voting vote_dict: Dict[str, float] = {} for answer, weight in zip(normalized, weights): vote_dict[answer] = vote_dict.get(answer, 0.0) + weight # Find majority total_weight = sum(weights) majority_answer = max(vote_dict.items(), key=lambda x: x[1])[0] confidence = vote_dict[majority_answer] / total_weight # Check threshold if confidence < threshold: logger.warning( f"Majority voting confidence {confidence:.2f} below threshold {threshold}" ) return majority_answer, confidence @staticmethod def self_consistency( chains: List[ReasoningChain], min_agreement: float = 0.6, ) -> Tuple[Optional[str], float, List[ReasoningChain]]: """ Self-consistency checking across reasoning chains. Args: chains: List of reasoning chains min_agreement: Minimum agreement required Returns: - consensus_answer: Answer if consensus reached, else None - agreement: Agreement ratio [0, 1] - supporting_chains: Chains supporting consensus """ if not chains: return None, 0.0, [] # Extract answers answers = [c.final_answer.strip().lower() for c in chains] # Count occurrences answer_counts = Counter(answers) most_common_answer, count = answer_counts.most_common(1)[0] # Calculate agreement agreement = count / len(chains) # Check if consensus reached consensus = most_common_answer if agreement >= min_agreement else None # Find supporting chains supporting_chains = [ c for c in chains if c.final_answer.strip().lower() == most_common_answer ] logger.info( f"Self-consistency: {agreement:.1%} agreement " f"({'consensus' if consensus else 'no consensus'})" ) return consensus, agreement, supporting_chains