abliteration

final_refusal_dirs.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1c00500a2055f8669d27bc4a6c26296eb484396d9bd8868f77e31fd8b77196c8
+size 210282

modeling_qwen3_moe.py

@@ -1,174 +1,124 @@
-from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig, TextStreamer
 import torch
-import torch.nn as nn
+from torch import nn
+import torch.nn.functional as F
+import types
 import os
-import signal
-from typing import Optional, Tuple
-import einops
-import jaxtyping
-
-cpu_count = os.cpu_count()
-print(f"Number of CPU cores in the system: {cpu_count}")
-half_cpu_count = cpu_count // 2
-os.environ["MKL_NUM_THREADS"] = str(half_cpu_count)
-os.environ["OMP_NUM_THREADS"] = str(half_cpu_count)
-torch.set_num_threads(half_cpu_count)
-
-print(f"PyTorch threads: {torch.get_num_threads()}")
-print(f"MKL threads: {os.getenv('MKL_NUM_THREADS')}")
-print(f"OMP threads: {os.getenv('OMP_NUM_THREADS')}")
-
-# Load the model and tokenizer
-MODEL_ID = "Qwen/Qwen3-30B-A3B"
-print(f"Load Model {MODEL_ID} ... ")
-quant_config_4 = BitsAndBytesConfig(
-    load_in_4bit=True,
-    bnb_4bit_compute_dtype=torch.bfloat16,
-    bnb_4bit_use_double_quant=True,
-    llm_int8_enable_fp32_cpu_offload=True,
-)
-
-model = AutoModelForCausalLM.from_pretrained(
-    MODEL_ID,
-    device_map="auto",
-    trust_remote_code=True,
-    quantization_config=quant_config_4,
-    torch_dtype=torch.bfloat16
-)
-
-tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, trust_remote_code=True)
-if tokenizer.pad_token is None:
-    tokenizer.pad_token = tokenizer.eos_token
-tokenizer.pad_token_id = tokenizer.eos_token_id
-
-messages = []
-enable_thinking = True
-skip_prompt=True
-skip_special_tokens=True
-
-def direction_ablation_hook(activation: jaxtyping.Float[torch.Tensor, "... d_act"],
-                            direction: jaxtyping.Float[torch.Tensor, "d_act"]):
-    proj = einops.einsum(activation, direction.view(-1, 1), '... d_act, d_act single -> ... single') * direction
-    return activation - proj
-
-class AblationDecoderLayer(nn.Module):
-    def __init__(self, original_layer, refusal_dir):
-        super(AblationDecoderLayer, self).__init__()
-        self.original_layer = original_layer
-        self.refusal_dir = refusal_dir
+
+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):
-        hidden_states = args[0]
-        ablated = direction_ablation_hook(hidden_states, self.refusal_dir.to(hidden_states.device)).to(hidden_states.device)
-        args = (ablated,) + args[1:]
-        return self.original_layer.forward(*args, **kwargs)
-
-class CustomTextStreamer(TextStreamer):
-    def __init__(self, tokenizer, skip_prompt=True, skip_special_tokens=True):
-        super().__init__(tokenizer, skip_prompt=skip_prompt, skip_special_tokens=skip_special_tokens)
-        self.generated_text = ""
-        self.stop_flag = False
-
-    def on_finalized_text(self, text: str, stream_end: bool = False):
-        self.generated_text += text
-        print(text, end="", flush=True)
-        if self.stop_flag:
-            raise StopIteration
-
-    def stop_generation(self):
-        self.stop_flag = True
-
-def generate_stream(model, tokenizer, messages, enable_thinking, skip_prompt, skip_special_tokens, max_new_tokens):
-    input_ids = tokenizer.apply_chat_template(
-        messages,
-        tokenize=True,
-        enable_thinking = enable_thinking,
-        add_generation_prompt=True,
-        return_tensors="pt"
-    )
-    attention_mask = torch.ones_like(input_ids, dtype=torch.long)
-    tokens = input_ids.to(model.device)
-    attention_mask = attention_mask.to(model.device)
-
-    streamer = CustomTextStreamer(tokenizer, skip_prompt=skip_prompt, skip_special_tokens=skip_special_tokens)
-
-    def signal_handler(sig, frame):
-        streamer.stop_generation()
-        print("\n[Generation stopped by user with Ctrl+C]")
-
-    signal.signal(signal.SIGINT, signal_handler)
-
-    print("Response: ", end="", flush=True)
-    try:
-        generated_ids = model.generate(
-            tokens,
-            attention_mask=attention_mask,
-            use_cache=False,
-            max_new_tokens=max_new_tokens,
-            do_sample=True,
-            pad_token_id=tokenizer.pad_token_id,
-            streamer=streamer
-        )
-        del generated_ids
-    except StopIteration:
-        print("\n[Stopped by user]")
-
-    del input_ids, attention_mask
-    torch.cuda.empty_cache()
-    signal.signal(signal.SIGINT, signal.SIG_DFL)
-
-    return streamer.generated_text, streamer.stop_flag
-
-
-
-final_refusal_dirs= torch.load(MODEL_ID + "/final_refusal_dirs.pt", map_location='cpu', weights_only=True)
-# candidate layer, 16, 21 ...
-candidate_layer = 16
-
-refusal_dir = final_refusal_dirs[candidate_layer]
-
-for idx in range(len(model.model.layers)):
-    model.model.layers[idx] = AblationDecoderLayer(model.model.layers[idx], refusal_dir)
-
-while True:
-    user_input = input("User: ").strip()
-    if user_input.lower() == "/exit":
-        print("Exiting chat.")
-        break
-    if user_input.lower() == "/clear":
-        messages = []
-        print("Chat history cleared. Starting a new conversation.")
-        continue
-    if user_input.lower() == "/no_think":
-        if enable_thinking:
-            enable_thinking = False
-            print("Thinking = False.")
-        else:
-            enable_thinking = True
-            print("Thinking = True.")
-        continue
-    if user_input.lower() == "/skip_prompt":
-        if skip_prompt:
-            skip_prompt = False
-            print("skip_prompt = False.")
-        else:
-            skip_prompt = True
-            print("skip_prompt = True.")
-        continue
-    if user_input.lower() == "/skip_special_tokens":
-        if skip_special_tokens:
-            skip_special_tokens = False
-            print("skip_special_tokens = False.")
-        else:
-            skip_special_tokens = True
-            print("skip_special_tokens = True.")
-        continue
-    if not user_input:
-        print("Input cannot be empty. Please enter something.")
-        continue
-    messages.append({"role": "user", "content": user_input})
-    response, stop_flag = generate_stream(model, tokenizer, messages, enable_thinking, skip_prompt, skip_special_tokens, 8192)
-    print("", flush=True)
-    if stop_flag:
-        continue
-    messages.append({"role": "assistant", "content": response})
+        # 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)