File size: 9,727 Bytes
00dd625 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 | """AlignX Model — full two-stage alignment framework.
Wraps a HuggingFace causal LM and surgically replaces the last transformer
FFN layer with the MoCaE AlignX layer (Stage 2 injection).
"""
import os
import torch
import torch.nn as nn
from typing import Optional, Dict, List, Union
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
from .mocae import MoCaE, AlignXLayer, ExpertFFN
from .task_feature_matrix import TaskFeatureMatrix
# ---------------------------------------------------------------------------
# Model-family helpers
# ---------------------------------------------------------------------------
def _get_layers(model) -> nn.ModuleList:
"""Return the transformer layer list for supported model families."""
if hasattr(model, "model") and hasattr(model.model, "layers"):
return model.model.layers
if hasattr(model, "transformer") and hasattr(model.transformer, "h"):
return model.transformer.h
raise ValueError(f"Unsupported model architecture: {type(model)}")
def _get_mlp(layer) -> nn.Module:
"""Return the FFN/MLP sub-module from a transformer layer."""
for attr in ("mlp", "feed_forward", "ffn", "ff"):
if hasattr(layer, attr):
return getattr(layer, attr)
raise ValueError(f"Cannot locate MLP in layer {type(layer)}")
def _set_mlp(layer, new_module: nn.Module):
"""Replace the FFN/MLP sub-module in a transformer layer."""
for attr in ("mlp", "feed_forward", "ffn", "ff"):
if hasattr(layer, attr):
setattr(layer, attr, new_module)
return
raise ValueError(f"Cannot locate MLP in layer {type(layer)}")
def _get_intermediate_dim(mlp_module) -> int:
"""Infer the FFN intermediate dimension from the gate/up projection."""
for attr in ("gate_proj", "w1", "fc1", "dense_h_to_4h"):
if hasattr(mlp_module, attr):
proj = getattr(mlp_module, attr)
return proj.out_features
for attr in ("up_proj", "w2", "fc2"):
if hasattr(mlp_module, attr):
return getattr(mlp_module, attr).out_features
raise ValueError(f"Cannot infer intermediate_dim from {type(mlp_module)}")
# ---------------------------------------------------------------------------
# AlignX model class
# ---------------------------------------------------------------------------
class AlignXModel(nn.Module):
"""AlignX-wrapped causal LM with MoCaE layer injected at the last FFN.
Usage:
model = AlignXModel(base_lm, hidden_dim=4096, intermediate_dim=11008)
model.register_task_matrices(T_h, T_ha, T_ho)
# Train only MoCaE params (base_lm is frozen)
logits = model(input_ids, attention_mask)
"""
def __init__(
self,
base_lm: nn.Module,
hidden_dim: int,
intermediate_dim: int,
k: int = 256,
lambda1: float = 0.6,
lambda2: float = 0.4,
epsilon: float = 0.05,
n_clusters: int = 8,
layer_idx: int = -1,
freeze_base: bool = True,
):
super().__init__()
self.base_lm = base_lm
self.hidden_dim = hidden_dim
self.layer_idx = layer_idx
self.config = base_lm.config
# Build MoCaE
self.mocae = MoCaE(
hidden_dim=hidden_dim,
intermediate_dim=intermediate_dim,
k=k,
lambda1=lambda1,
lambda2=lambda2,
epsilon=epsilon,
n_clusters=n_clusters,
)
self.alignx_layer = AlignXLayer(self.mocae, hidden_dim)
# Inject AlignX layer into the last transformer FFN
layers = _get_layers(self.base_lm)
target_layer = layers[layer_idx]
original_mlp = _get_mlp(target_layer)
# Initialise all experts from the original FFN weights (helpful default)
self.mocae.init_experts_from_ffn(original_mlp)
# Replace the FFN
_set_mlp(target_layer, self.alignx_layer)
# Move AlignXLayer (MoCaE) to the same device as the rest of the target layer.
# input_layernorm is a plain fp32/bf16 LayerNorm — its device is reliable even
# when the base model is 4-bit quantized.
if hasattr(target_layer, "input_layernorm"):
_target_device = target_layer.input_layernorm.weight.device
elif hasattr(target_layer, "ln_1"):
_target_device = target_layer.ln_1.weight.device
else:
_target_device = torch.device(
f"cuda:{torch.cuda.device_count() - 1}" if torch.cuda.is_available() else "cpu"
)
_target_dtype = target_layer.input_layernorm.weight.dtype if hasattr(target_layer, "input_layernorm") else torch.float16
self.alignx_layer.to(_target_device).to(_target_dtype)
print(f"[AlignX] AlignXLayer (MoCaE) placed on {_target_device} dtype={_target_dtype}")
# Freeze base model parameters (only MoCaE is trainable)
if freeze_base:
for name, param in self.base_lm.named_parameters():
param.requires_grad_(False)
for param in self.mocae.parameters():
param.requires_grad_(True)
total_base = sum(p.numel() for p in self.base_lm.parameters())
total_mocae = sum(p.numel() for p in self.mocae.parameters())
print(f"[AlignX] Base LM params: {total_base:,} | MoCaE params: {total_mocae:,}")
# ------------------------------------------------------------------
def register_task_matrices(self, T_helpful, T_harmless, T_honest):
self.mocae.register_task_matrices(T_helpful, T_harmless, T_honest)
def init_expert_from_finetuned(self, expert_idx: int, finetuned_lm: nn.Module):
"""Initialise expert `expert_idx` from the last FFN of a fine-tuned model."""
layers = _get_layers(finetuned_lm)
ffn = _get_mlp(layers[self.layer_idx])
self.mocae.init_expert_from_ffn(expert_idx, ffn)
# ------------------------------------------------------------------
def forward(self, input_ids, attention_mask=None, labels=None, **kwargs):
return self.base_lm(
input_ids=input_ids,
attention_mask=attention_mask,
labels=labels,
**kwargs,
)
@torch.no_grad()
def generate(self, input_ids, attention_mask=None, **kwargs):
return self.base_lm.generate(
input_ids=input_ids,
attention_mask=attention_mask,
**kwargs,
)
def save_mocae(self, path: str):
os.makedirs(path, exist_ok=True)
torch.save(self.mocae.state_dict(), os.path.join(path, "mocae.pt"))
print(f"[AlignX] Saved MoCaE weights to {path}/mocae.pt")
def load_mocae(self, path: str):
state = torch.load(os.path.join(path, "mocae.pt"), map_location="cpu")
self.mocae.load_state_dict(state, strict=False)
print(f"[AlignX] Loaded MoCaE weights from {path}/mocae.pt")
# ---------------------------------------------------------------------------
# Builder helper
# ---------------------------------------------------------------------------
def build_alignx_model(
base_model_name_or_path: str,
finetuned_paths: Optional[Dict[str, str]] = None,
task_matrix_paths: Optional[Dict[str, str]] = None,
load_in_4bit: bool = True,
device_map: str = "auto",
k: int = 256,
lambda1: float = 0.6,
lambda2: float = 0.4,
freeze_base: bool = True,
layer_idx: int = -1,
) -> AlignXModel:
"""Build a full AlignX model from a base checkpoint.
Optionally initialises each expert from a per-axis fine-tuned checkpoint
and loads precomputed task-feature matrices.
"""
bnb_config = None
if load_in_4bit:
bnb_config = BitsAndBytesConfig(
load_in_4bit=True,
bnb_4bit_compute_dtype=torch.bfloat16,
bnb_4bit_use_double_quant=True,
bnb_4bit_quant_type="nf4",
)
base_lm = AutoModelForCausalLM.from_pretrained(
base_model_name_or_path,
quantization_config=bnb_config,
device_map=device_map,
torch_dtype=torch.bfloat16,
trust_remote_code=True,
)
hidden_dim = base_lm.config.hidden_size
layers = _get_layers(base_lm)
mlp = _get_mlp(layers[layer_idx])
intermediate_dim = _get_intermediate_dim(mlp)
alignx = AlignXModel(
base_lm=base_lm,
hidden_dim=hidden_dim,
intermediate_dim=intermediate_dim,
k=k,
lambda1=lambda1,
lambda2=lambda2,
layer_idx=layer_idx,
freeze_base=freeze_base,
)
# Initialise each expert from axis-specific fine-tuned model
if finetuned_paths:
axis_to_idx = {"helpful": 0, "harmless": 1, "honest": 2}
for axis, ckpt_path in finetuned_paths.items():
if axis in axis_to_idx and os.path.exists(ckpt_path):
print(f"[AlignX] Loading expert {axis} from {ckpt_path}")
ft_lm = AutoModelForCausalLM.from_pretrained(
ckpt_path,
torch_dtype=torch.float16,
device_map="cpu",
trust_remote_code=True,
)
alignx.init_expert_from_finetuned(axis_to_idx[axis], ft_lm)
del ft_lm
# Load task-feature matrices
if task_matrix_paths:
T = {}
for axis in ("helpful", "harmless", "honest"):
path = task_matrix_paths.get(axis, "")
if path and os.path.exists(path):
T[axis] = torch.load(path, map_location="cpu")
if len(T) == 3:
alignx.register_task_matrices(T["helpful"], T["harmless"], T["honest"])
return alignx
|