Spaces:
Running
Running
File size: 11,431 Bytes
f2ae1f5 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 | """Qwen-Scope SAE feature reading + steering for transformers.
End-to-end demo:
1. Loads a base Qwen3 model and a matching Qwen-Scope TopK SAE checkpoint.
2. Captures the residual-stream output of a chosen decoder layer.
3. Encodes it through the SAE -> top-K firing features.
4. Generates a baseline completion.
5. Re-generates with feature steering: residual h <- h + alpha * W_dec[:, feat]
applied via register_forward_hook on every forward pass.
Verified against:
* Qwen/SAE-Res-Qwen3-1.7B-Base-W32K-L0_50 (W_enc 32768x2048, W_dec 2048x32768,
b_enc 32768, b_dec 2048, all float32, K=50)
* Qwen/Qwen3-1.7B-Base (28 Qwen3DecoderLayer, hidden_size=2048, layer forward
returns bare torch.Tensor under transformers >= 5).
"""
from __future__ import annotations
import argparse
import contextlib
from dataclasses import dataclass
from pathlib import Path
import torch
import torch.nn.functional as F
from huggingface_hub import hf_hub_download
from transformers import AutoModelForCausalLM, AutoTokenizer
# ---------------------------------------------------------------------------
# SAE
# ---------------------------------------------------------------------------
@dataclass
class SAE:
W_enc: torch.Tensor # (n_features, d_model)
W_dec: torch.Tensor # (d_model, n_features)
b_enc: torch.Tensor # (n_features,)
b_dec: torch.Tensor # (d_model,)
k: int # TopK
layer: int # layer index this SAE belongs to
@classmethod
def from_repo(cls, repo: str, layer: int, k: int, device: str = "cpu",
dtype: torch.dtype = torch.float32) -> "SAE":
path = hf_hub_download(repo, f"layer{layer}.sae.pt")
return cls.from_path(path, layer=layer, k=k, device=device, dtype=dtype)
@classmethod
def from_path(cls, path: str | Path, layer: int, k: int,
device: str = "cpu", dtype: torch.dtype = torch.float32) -> "SAE":
sd = torch.load(str(path), map_location=device, weights_only=True)
for key in ("W_enc", "W_dec", "b_enc", "b_dec"):
if key not in sd:
raise KeyError(f"SAE checkpoint at {path} missing key {key!r}; "
f"got {list(sd.keys())}")
return cls(
W_enc=sd["W_enc"].to(device=device, dtype=dtype),
W_dec=sd["W_dec"].to(device=device, dtype=dtype),
b_enc=sd["b_enc"].to(device=device, dtype=dtype),
b_dec=sd["b_dec"].to(device=device, dtype=dtype),
k=k, layer=layer,
)
@property
def n_features(self) -> int:
return self.W_enc.shape[0]
@property
def d_model(self) -> int:
return self.W_enc.shape[1]
def encode(self, x: torch.Tensor) -> torch.Tensor:
"""Encode residual stream activations -> sparse feature codes (TopK)."""
x = x.to(device=self.W_enc.device, dtype=self.W_enc.dtype)
pre = F.linear(x, self.W_enc, self.b_enc) # (..., n_features)
topk_vals, topk_idx = pre.topk(self.k, dim=-1)
z = torch.zeros_like(pre)
z.scatter_(-1, topk_idx, topk_vals)
return z
def decode(self, z: torch.Tensor) -> torch.Tensor:
z = z.to(device=self.W_dec.device, dtype=self.W_dec.dtype)
return F.linear(z, self.W_dec, self.b_dec)
def steering_vector(self, feature_id: int) -> torch.Tensor:
return self.W_dec[:, feature_id].clone()
# ---------------------------------------------------------------------------
# Hook helpers
# ---------------------------------------------------------------------------
def _layer_output_to_tensor(out):
"""Qwen3DecoderLayer returns torch.Tensor in transformers >= 5,
a tuple (hidden_states, ...) in transformers < 5. Handle both."""
if isinstance(out, tuple):
return out[0], out
return out, None
def _rebuild_layer_output(new_h: torch.Tensor, original_out):
if original_out is None:
return new_h
return (new_h, *original_out[1:])
@contextlib.contextmanager
def capture_residual(model, layer_idx: int):
"""Capture the residual-stream output of model.model.layers[layer_idx]."""
bucket: dict = {}
layer = model.model.layers[layer_idx]
def hook(_module, _inp, out):
h, _ = _layer_output_to_tensor(out)
bucket["h"] = h.detach()
return out
handle = layer.register_forward_hook(hook)
try:
yield bucket
finally:
handle.remove()
@contextlib.contextmanager
def steer(model, layer_idx: int, direction: torch.Tensor, alpha: float):
"""Add `alpha * direction` to the residual stream output of layer_idx
on every forward pass while the context is active."""
layer = model.model.layers[layer_idx]
direction = direction.detach()
def hook(_module, _inp, out):
h, original = _layer_output_to_tensor(out)
d = direction.to(device=h.device, dtype=h.dtype)
new_h = h + alpha * d
return _rebuild_layer_output(new_h, original)
handle = layer.register_forward_hook(hook)
try:
yield
finally:
handle.remove()
# ---------------------------------------------------------------------------
# Pipeline
# ---------------------------------------------------------------------------
def read_top_features(model, tokenizer, sae: SAE, prompt: str,
layer_idx: int, top_n: int = 10):
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
with torch.no_grad(), capture_residual(model, layer_idx) as bucket:
model(**inputs)
h = bucket["h"] # (1, T, d_model) on model.device
h_last = h[0, -1].unsqueeze(0) # (1, d_model) — encode() handles device/dtype
z = sae.encode(h_last)[0]
nonzero = z.nonzero(as_tuple=False).flatten()
vals = z[nonzero]
order = vals.argsort(descending=True)
top = nonzero[order][:top_n]
return [(int(f.item()), float(z[f].item())) for f in top]
def generate(model, tokenizer, prompt: str, max_new_tokens: int = 40):
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
with torch.no_grad():
out = model.generate(
**inputs,
max_new_tokens=max_new_tokens,
do_sample=False, # deterministic for A/B comparison
pad_token_id=tokenizer.eos_token_id,
)
return tokenizer.decode(out[0], skip_special_tokens=True)
# ---------------------------------------------------------------------------
# CLI
# ---------------------------------------------------------------------------
def parse_topk_from_repo(repo: str) -> int:
# e.g. "Qwen/SAE-Res-Qwen3-1.7B-Base-W32K-L0_50" -> 50
suffix = repo.rsplit("L0_", 1)
if len(suffix) == 2 and suffix[1].isdigit():
return int(suffix[1])
return 50
def main():
ap = argparse.ArgumentParser()
ap.add_argument("--model", default="Qwen/Qwen3-1.7B-Base")
ap.add_argument("--sae-repo", default="Qwen/SAE-Res-Qwen3-1.7B-Base-W32K-L0_50")
ap.add_argument("--layer", type=int, default=14)
ap.add_argument("--prompt", default="The capital of France is")
ap.add_argument("--max-new-tokens", type=int, default=40)
ap.add_argument("--alpha", type=float, default=-10.0,
help="Steering magnitude. Negative suppresses, positive amplifies.")
ap.add_argument("--suppress-rank", type=int, default=0,
help="Which top-firing feature (0 = strongest) to steer.")
ap.add_argument("--feature-id", type=int, default=None,
help="Override: steer this exact feature instead of a top-rank pick.")
ap.add_argument("--topk", type=int, default=None,
help="Override SAE TopK (auto-detected from repo name).")
ap.add_argument("--device", default=None,
help="cuda | mps | cpu (auto if omitted)")
ap.add_argument("--dtype", default="bfloat16",
choices=["bfloat16", "float16", "float32"])
args = ap.parse_args()
if args.device is None:
if torch.cuda.is_available():
args.device = "cuda"
elif torch.backends.mps.is_available():
args.device = "mps"
else:
args.device = "cpu"
dtype = {"bfloat16": torch.bfloat16, "float16": torch.float16,
"float32": torch.float32}[args.dtype]
print(f"[load] model={args.model} device={args.device} dtype={args.dtype}")
tokenizer = AutoTokenizer.from_pretrained(args.model)
model = AutoModelForCausalLM.from_pretrained(
args.model, dtype=dtype, device_map=args.device,
)
model.eval()
n_layers = len(model.model.layers)
if not (0 <= args.layer < n_layers):
raise ValueError(f"--layer {args.layer} out of range; model has {n_layers} layers")
hidden = model.config.hidden_size
print(f"[load] {type(model).__name__}: {n_layers} layers, hidden={hidden}")
k = args.topk or parse_topk_from_repo(args.sae_repo)
print(f"[load] SAE repo={args.sae_repo} layer={args.layer} K={k}")
sae = SAE.from_repo(args.sae_repo, layer=args.layer, k=k,
device=args.device, dtype=dtype)
if sae.d_model != hidden:
raise ValueError(f"SAE d_model={sae.d_model} != model hidden_size={hidden}; "
f"this SAE doesn't match this model.")
# 1. Top features for the prompt
print(f"\n[features] top firing at layer {args.layer} for prompt: {args.prompt!r}")
top = read_top_features(model, tokenizer, sae, args.prompt, args.layer, top_n=10)
for rank, (fid, act) in enumerate(top):
print(f" rank {rank:2d} feature {fid:>6d} act={act:+.4f}")
# Pick steering target
if args.feature_id is not None:
target_id = args.feature_id
else:
target_id = top[args.suppress_rank][0]
# 2. Baseline generation
print(f"\n[baseline] generating (no steering)...")
baseline = generate(model, tokenizer, args.prompt, args.max_new_tokens)
print(f" >>> {baseline!r}")
# 3. Steered generation
print(f"\n[steer] feature {target_id} at layer {args.layer} with alpha={args.alpha}")
direction = sae.steering_vector(target_id)
with steer(model, args.layer, direction, args.alpha):
steered = generate(model, tokenizer, args.prompt, args.max_new_tokens)
print(f" >>> {steered!r}")
# 4. Verify the steering actually moved the feature
inputs = tokenizer(args.prompt, return_tensors="pt").to(model.device)
with torch.no_grad(), capture_residual(model, args.layer) as bucket:
model(**inputs)
base_act = sae.encode(bucket["h"][0, -1].unsqueeze(0))[0, target_id].item()
with torch.no_grad(), steer(model, args.layer, direction, args.alpha), \
capture_residual(model, args.layer) as bucket:
model(**inputs)
steered_act = sae.encode(bucket["h"][0, -1].unsqueeze(0))[0, target_id].item()
print(f"\n[verify] feature {target_id} activation: baseline={base_act:+.4f} "
f"steered={steered_act:+.4f} delta={steered_act - base_act:+.4f}")
if args.alpha > 0 and steered_act <= base_act:
print(" WARN: alpha>0 but activation didn't go up — unexpected.")
if args.alpha < 0 and steered_act >= base_act:
print(" WARN: alpha<0 but activation didn't go down — unexpected.")
if __name__ == "__main__":
main()
|