modeling_qwen3_moe.py

import torch
from torch import nn
import torch.nn.functional as F
import types
import os

from transformers.models.qwen2_moe.modeling_qwen2_moe import Qwen2MoeDecoderLayer, Qwen2MoeForCausalLM, Qwen2MoeModel
from transformers.utils import logging


logger = logging.get_logger(__name__)

# This custom layer contains the core "abliterated" logic.
# It subtracts a "steering vector" from the hidden states.
class AbliteratedDecoderLayer(Qwen2MoeDecoderLayer):
    def __init__(self, config, layer_idx):
        super().__init__(config, layer_idx)

    def forward(self, hidden_states, refusal_directions, *args, **kwargs):
        if refusal_directions is not None and self.layer_idx in refusal_directions:
            # Move refusal directions to the correct device
            layer_refusal_directions = refusal_directions[self.layer_idx].to(hidden_states.device)
            
            # Project hidden states into the direction of the refusal vector
            projected_hidden_states = F.cosine_similarity(hidden_states, layer_refusal_directions.unsqueeze(0).unsqueeze(0), dim=-1)
            
            # Get the steering vector
            steering_vector = projected_hidden_states.unsqueeze(-1) * layer_refusal_directions
            
            # Apply the steering vector
            hidden_states = hidden_states - steering_vector

        # Call the original forward pass of the layer
        return super().forward(hidden_states, *args, **kwargs)


# This custom model class will automatically patch itself upon loading.
class AbliteratedQwen3MoeForCausalLM(Qwen2MoeForCausalLM):
    def __init__(self, config):
        super().__init__(config)
        
        self.refusal_directions = None
        try:
            # In a Hugging Face model repo, config._name_or_path is the repo path
            refusal_directions_path = os.path.join(config._name_or_path, 'final_refusal_dirs.pt')
            if os.path.exists(refusal_directions_path):
                self.refusal_directions = torch.load(refusal_directions_path, map_location="cpu")
                logger.info("Successfully loaded 'final_refusal_dirs.pt' for model abliteration.")
            else:
                 logger.warning("'final_refusal_dirs.pt' not found. Model will not be abliterated.")
                 return

        except Exception as e:
            logger.error(f"Failed to load 'final_refusal_dirs.pt'. Model will not be abliterated. Error: {e}")
            return

        # Patch the model by swapping the decoder layers
        logger.info("Patching model with AbliteratedDecoderLayer.")
        for i in range(len(self.model.layers)):
            old_layer = self.model.layers[i]
            # We need to pass the original config and layer_idx
            new_layer = AbliteratedDecoderLayer(old_layer.config, old_layer.layer_idx)
            # Copy all weights and buffers from the old layer
            new_layer.load_state_dict(old_layer.state_dict())
            self.model.layers[i] = new_layer
        logger.info("Model patching complete.")

    def forward(self, *args, **kwargs):
        # We need to correctly pass the refusal_directions to the layers.
        # The layers are called inside self.model.forward.
        # So we patch the forward method of the underlying Qwen2MoeModel instance.
        original_model_forward = self.model.forward
        
        def patched_forward(*model_args, **model_kwargs):
            # The Qwen2MoeModel's forward method does not take our custom arg.
            # We need a way to pass it down. We can temporarily attach it to the `self.model` object.
            
            # The layers' forward methods were modified to accept refusal_directions
            # But the loop in Qwen2MoeModel.forward doesn't know about it.
            # The easiest way is to modify the loop itself.
            # Let's override the `self.model.forward` method entirely
            
            # To avoid re-patching on every call, we can do it once in __init__
            # Let's move the forward patch to __init__
            pass

        # Since we replaced the layers, their `forward` methods are now different.
        # We must modify the calling code in `self.model.forward` to pass the new argument.
        # The most robust way is to monkey-patch `self.model.forward` once.
        if not hasattr(self.model, '_forward_patched'):
            original_forward = self.model.forward
            
            def new_model_forward(*f_args, **f_kwargs):
                # The original `Qwen2MoeModel.forward` iterates through `self.layers`
                # and calls each `decoder_layer(...)`.
                # We need to inject `refusal_directions` into that call.
                # Let's redefine the entire `Qwen2MoeModel.forward` logic here
                # to ensure correctness.
                
                # This is a simplified version of the original source, modified for our purpose
                hidden_states = f_kwargs.get('inputs_embeds')
                if hidden_states is None:
                    hidden_states = self.model.embed_tokens(f_kwargs.get('input_ids'))

                for decoder_layer in self.model.layers:
                    layer_outputs = decoder_layer(
                        hidden_states,
                        refusal_directions=self.refusal_directions,
                        attention_mask=f_kwargs.get('attention_mask'),
                        position_ids=f_kwargs.get('position_ids'),
                        past_key_value=f_kwargs.get('past_key_values'),
                        output_attentions=f_kwargs.get('output_attentions'),
                        output_router_logits=f_kwargs.get('output_router_logits'),
                        use_cache=f_kwargs.get('use_cache'),
                    )
                    hidden_states = layer_outputs[0]

                hidden_states = self.model.norm(hidden_states)
                return (hidden_states,) # Return in a tuple as the base class expects

            self.model.forward = types.MethodType(new_model_forward, self)
            self.model._forward_patched = True

        return super().forward(*args, **kwargs)