Spaces:
Sleeping
Sleeping
frankaging
commited on
Commit
·
36edf66
1
Parent(s):
7962ddb
autosteer
Browse files
app.py
CHANGED
@@ -2,12 +2,13 @@ import os, json, random
|
|
2 |
import torch
|
3 |
import gradio as gr
|
4 |
import spaces
|
5 |
-
from transformers import AutoModelForCausalLM, AutoTokenizer, TextIteratorStreamer
|
6 |
from huggingface_hub import login, hf_hub_download
|
7 |
import pyreft
|
8 |
import pyvene as pv
|
9 |
from threading import Thread
|
10 |
from typing import Iterator
|
|
|
11 |
|
12 |
HF_TOKEN = os.environ.get("HF_TOKEN")
|
13 |
login(token=HF_TOKEN)
|
@@ -16,6 +17,18 @@ MAX_MAX_NEW_TOKENS = 2048
|
|
16 |
DEFAULT_MAX_NEW_TOKENS = 256 # smaller default to save memory
|
17 |
MAX_INPUT_TOKEN_LENGTH = 4096
|
18 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
19 |
def load_jsonl(jsonl_path):
|
20 |
jsonl_data = []
|
21 |
with open(jsonl_path, 'r') as f:
|
@@ -29,19 +42,44 @@ class Steer(pv.SourcelessIntervention):
|
|
29 |
def __init__(self, **kwargs):
|
30 |
super().__init__(**kwargs, keep_last_dim=True)
|
31 |
self.proj = torch.nn.Linear(
|
32 |
-
|
33 |
-
|
34 |
def forward(self, base, source=None, subspaces=None):
|
35 |
-
if subspaces
|
36 |
return base
|
37 |
-
|
38 |
-
|
39 |
-
|
40 |
-
|
41 |
-
|
42 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
43 |
return base + steering_vec
|
44 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
45 |
# Check GPU
|
46 |
if not torch.cuda.is_available():
|
47 |
print("Warning: Running on CPU, may be slow.")
|
@@ -73,7 +111,23 @@ if torch.cuda.is_available():
|
|
73 |
concept_id_map[item["concept"]] = concept_reindex
|
74 |
concept_reindex += 1
|
75 |
|
76 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
77 |
steer.proj.weight.data = params.float()
|
78 |
|
79 |
pv_model = pv.IntervenableModel({
|
@@ -117,8 +171,10 @@ def generate(
|
|
117 |
"intervene_on_prompt": True,
|
118 |
"subspaces": [
|
119 |
{
|
120 |
-
"idx":
|
121 |
-
"mag":
|
|
|
|
|
122 |
}
|
123 |
] if subspaces_list else None,
|
124 |
"streamer": streamer,
|
@@ -133,9 +189,6 @@ def generate(
|
|
133 |
partial_text.append(token_str)
|
134 |
yield "".join(partial_text)
|
135 |
|
136 |
-
def _build_remove_choices(subspaces):
|
137 |
-
return [f"(+{x['display_mag']:.1f}*) {x['text']}" for x in subspaces]
|
138 |
-
|
139 |
def filter_concepts(search_text: str):
|
140 |
if not search_text.strip():
|
141 |
return concept_list[:500]
|
@@ -144,15 +197,21 @@ def filter_concepts(search_text: str):
|
|
144 |
|
145 |
def add_concept_to_list(selected_concept, user_slider_val, current_list):
|
146 |
if not selected_concept:
|
147 |
-
return current_list
|
148 |
|
149 |
-
|
|
|
|
|
|
|
|
|
|
|
150 |
internal_mag = user_slider_val * 50
|
151 |
new_entry = {
|
152 |
"text": selected_concept,
|
153 |
"idx": idx,
|
154 |
"display_mag": user_slider_val,
|
155 |
"internal_mag": internal_mag,
|
|
|
156 |
}
|
157 |
# Add to the beginning of the list
|
158 |
current_list = [new_entry]
|
@@ -160,16 +219,23 @@ def add_concept_to_list(selected_concept, user_slider_val, current_list):
|
|
160 |
|
161 |
def update_dropdown_choices(search_text):
|
162 |
filtered = filter_concepts(search_text)
|
163 |
-
if not filtered:
|
164 |
-
return gr.update(choices=[], value=
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
165 |
# Automatically select the first matching concept
|
166 |
return gr.update(
|
167 |
choices=filtered,
|
168 |
value=filtered[0], # Select the first match
|
169 |
-
interactive=True
|
170 |
-
)
|
171 |
|
172 |
-
with gr.Blocks(fill_height=True) as demo:
|
173 |
# Remove default subspaces
|
174 |
selected_subspaces = gr.State([])
|
175 |
|
@@ -179,7 +245,7 @@ with gr.Blocks(fill_height=True) as demo:
|
|
179 |
chat_interface = gr.ChatInterface(
|
180 |
fn=generate,
|
181 |
title="Chat with a Concept Steering Model",
|
182 |
-
description="Steer responses by selecting concepts on the right
|
183 |
type="messages",
|
184 |
additional_inputs=[selected_subspaces],
|
185 |
fill_height=True
|
@@ -188,7 +254,7 @@ with gr.Blocks(fill_height=True) as demo:
|
|
188 |
# Right side: concept management
|
189 |
with gr.Column(scale=4):
|
190 |
gr.Markdown("## Steer Model Responses")
|
191 |
-
gr.Markdown("Search and then select a concept to steer. The closest match will be automatically selected.")
|
192 |
# Concept Search and Selection
|
193 |
with gr.Group():
|
194 |
search_box = gr.Textbox(
|
@@ -196,6 +262,7 @@ with gr.Blocks(fill_height=True) as demo:
|
|
196 |
placeholder="Find concepts to steer the model (e.g. 'time travel')",
|
197 |
lines=2,
|
198 |
)
|
|
|
199 |
concept_dropdown = gr.Dropdown(
|
200 |
label="Select a concept to steer the model (Click to see more!)",
|
201 |
interactive=True,
|
@@ -211,10 +278,10 @@ with gr.Blocks(fill_height=True) as demo:
|
|
211 |
|
212 |
# Wire up events
|
213 |
# When search box changes, update dropdown AND trigger concept selection
|
214 |
-
search_box.
|
215 |
update_dropdown_choices,
|
216 |
[search_box],
|
217 |
-
[concept_dropdown]
|
218 |
).then( # Chain the events to automatically add the concept
|
219 |
add_concept_to_list,
|
220 |
[concept_dropdown, concept_magnitude, selected_subspaces],
|
@@ -227,6 +294,12 @@ with gr.Blocks(fill_height=True) as demo:
|
|
227 |
[selected_subspaces]
|
228 |
)
|
229 |
|
|
|
|
|
|
|
|
|
|
|
|
|
230 |
concept_magnitude.input(
|
231 |
add_concept_to_list,
|
232 |
[concept_dropdown, concept_magnitude, selected_subspaces],
|
|
|
2 |
import torch
|
3 |
import gradio as gr
|
4 |
import spaces
|
5 |
+
from transformers import AutoConfig, AutoModelForCausalLM, AutoTokenizer, TextIteratorStreamer
|
6 |
from huggingface_hub import login, hf_hub_download
|
7 |
import pyreft
|
8 |
import pyvene as pv
|
9 |
from threading import Thread
|
10 |
from typing import Iterator
|
11 |
+
import torch.nn.functional as F
|
12 |
|
13 |
HF_TOKEN = os.environ.get("HF_TOKEN")
|
14 |
login(token=HF_TOKEN)
|
|
|
17 |
DEFAULT_MAX_NEW_TOKENS = 256 # smaller default to save memory
|
18 |
MAX_INPUT_TOKEN_LENGTH = 4096
|
19 |
|
20 |
+
css = """
|
21 |
+
#alert-message textarea {
|
22 |
+
background-color: #e8f4ff;
|
23 |
+
border: 1px solid #cce5ff;
|
24 |
+
color: #084298;
|
25 |
+
font-size: 1.1em;
|
26 |
+
padding: 12px;
|
27 |
+
border-radius: 4px;
|
28 |
+
font-weight: 500;
|
29 |
+
}
|
30 |
+
"""
|
31 |
+
|
32 |
def load_jsonl(jsonl_path):
|
33 |
jsonl_data = []
|
34 |
with open(jsonl_path, 'r') as f:
|
|
|
42 |
def __init__(self, **kwargs):
|
43 |
super().__init__(**kwargs, keep_last_dim=True)
|
44 |
self.proj = torch.nn.Linear(
|
45 |
+
self.embed_dim, kwargs["latent_dim"], bias=False)
|
46 |
+
self.subspace_generator = kwargs["subspace_generator"]
|
47 |
def forward(self, base, source=None, subspaces=None):
|
48 |
+
if subspaces == None:
|
49 |
return base
|
50 |
+
if subspaces["subspace_gen_inputs"] is not None:
|
51 |
+
# we call our subspace generator to generate the subspace on-the-fly.
|
52 |
+
raw_steering_vec = self.subspace_generator(
|
53 |
+
subspaces["subspace_gen_inputs"]["input_ids"],
|
54 |
+
subspaces["subspace_gen_inputs"]["attention_mask"],
|
55 |
+
)[0]
|
56 |
+
steering_vec = torch.tensor(subspaces["mag"]) * \
|
57 |
+
raw_steering_vec.unsqueeze(dim=0)
|
58 |
+
return base + steering_vec
|
59 |
+
else:
|
60 |
+
steering_vec = torch.tensor(subspaces["mag"]) * \
|
61 |
+
self.proj.weight[subspaces["idx"]].unsqueeze(dim=0)
|
62 |
return base + steering_vec
|
63 |
|
64 |
+
class RegressionWrapper(torch.nn.Module):
|
65 |
+
def __init__(self, base_model, hidden_size, output_dim):
|
66 |
+
super().__init__()
|
67 |
+
self.base_model = base_model
|
68 |
+
self.regression_head = torch.nn.Linear(hidden_size, output_dim)
|
69 |
+
|
70 |
+
def forward(self, input_ids, attention_mask):
|
71 |
+
outputs = self.base_model.model(
|
72 |
+
input_ids=input_ids,
|
73 |
+
attention_mask=attention_mask,
|
74 |
+
output_hidden_states=True,
|
75 |
+
return_dict=True
|
76 |
+
)
|
77 |
+
last_hiddens = outputs.hidden_states[-1]
|
78 |
+
last_token_representations = last_hiddens[:, -1]
|
79 |
+
preds = self.regression_head(last_token_representations)
|
80 |
+
preds = F.normalize(preds, p=2, dim=-1)
|
81 |
+
return preds
|
82 |
+
|
83 |
# Check GPU
|
84 |
if not torch.cuda.is_available():
|
85 |
print("Warning: Running on CPU, may be slow.")
|
|
|
111 |
concept_id_map[item["concept"]] = concept_reindex
|
112 |
concept_reindex += 1
|
113 |
|
114 |
+
# load subspace generator.
|
115 |
+
base_tokenizer = AutoTokenizer.from_pretrained(
|
116 |
+
f"google/gemma-2-2b", model_max_length=512)
|
117 |
+
config = AutoConfig.from_pretrained("google/gemma-2-2b")
|
118 |
+
base_model = AutoModelForCausalLM.from_config(config)
|
119 |
+
|
120 |
+
subspace_generator_weight_path = hf_hub_download(repo_id="pyvene/gemma-reft-2b-it-res-generator", filename="l20/weight.pt")
|
121 |
+
hidden_size = base_model.config.hidden_size
|
122 |
+
subspace_generator = RegressionWrapper(
|
123 |
+
base_model, hidden_size, hidden_size).bfloat16().to("cuda")
|
124 |
+
subspace_generator.load_state_dict(torch.load(subspace_generator_weight_path))
|
125 |
+
print(f"Loading model from saved file {subspace_generator_weight_path}")
|
126 |
+
_ = subspace_generator.eval()
|
127 |
+
|
128 |
+
steer = Steer(
|
129 |
+
embed_dim=params.shape[0], latent_dim=params.shape[1],
|
130 |
+
subspace_generator=subspace_generator)
|
131 |
steer.proj.weight.data = params.float()
|
132 |
|
133 |
pv_model = pv.IntervenableModel({
|
|
|
171 |
"intervene_on_prompt": True,
|
172 |
"subspaces": [
|
173 |
{
|
174 |
+
"idx": int(subspaces_list[0]["idx"]),
|
175 |
+
"mag": int(subspaces_list[0]["internal_mag"]),
|
176 |
+
"subspace_gen_inputs": base_tokenizer(subspaces_list[0]["subspace_gen_text"], return_tensors="pt").to("cuda") \
|
177 |
+
if subspaces_list[0]["subspace_gen_text"] is not None else None
|
178 |
}
|
179 |
] if subspaces_list else None,
|
180 |
"streamer": streamer,
|
|
|
189 |
partial_text.append(token_str)
|
190 |
yield "".join(partial_text)
|
191 |
|
|
|
|
|
|
|
192 |
def filter_concepts(search_text: str):
|
193 |
if not search_text.strip():
|
194 |
return concept_list[:500]
|
|
|
197 |
|
198 |
def add_concept_to_list(selected_concept, user_slider_val, current_list):
|
199 |
if not selected_concept:
|
200 |
+
return current_list
|
201 |
|
202 |
+
selected_concept_text = None
|
203 |
+
if selected_concept.startswith("[New] "):
|
204 |
+
selected_concept_text = selected_concept[6:]
|
205 |
+
idx = 0
|
206 |
+
else:
|
207 |
+
idx = concept_id_map[selected_concept]
|
208 |
internal_mag = user_slider_val * 50
|
209 |
new_entry = {
|
210 |
"text": selected_concept,
|
211 |
"idx": idx,
|
212 |
"display_mag": user_slider_val,
|
213 |
"internal_mag": internal_mag,
|
214 |
+
"subspace_gen_text": selected_concept_text
|
215 |
}
|
216 |
# Add to the beginning of the list
|
217 |
current_list = [new_entry]
|
|
|
219 |
|
220 |
def update_dropdown_choices(search_text):
|
221 |
filtered = filter_concepts(search_text)
|
222 |
+
if not filtered or len(filtered) == 0:
|
223 |
+
return gr.update(choices=[f"[New] {search_text}"], value=f"[New] {search_text}", interactive=True), gr.Textbox(
|
224 |
+
label="No matching existing concepts were found!",
|
225 |
+
value="Good news! Based on the concept you provided, we will automatically generate a steering vector. Try it out by starting a chat!",
|
226 |
+
lines=3,
|
227 |
+
interactive=False,
|
228 |
+
visible=True,
|
229 |
+
elem_id="alert-message"
|
230 |
+
)
|
231 |
# Automatically select the first matching concept
|
232 |
return gr.update(
|
233 |
choices=filtered,
|
234 |
value=filtered[0], # Select the first match
|
235 |
+
interactive=True, visible=True
|
236 |
+
), gr.Textbox(visible=False)
|
237 |
|
238 |
+
with gr.Blocks(css=css, fill_height=True) as demo:
|
239 |
# Remove default subspaces
|
240 |
selected_subspaces = gr.State([])
|
241 |
|
|
|
245 |
chat_interface = gr.ChatInterface(
|
246 |
fn=generate,
|
247 |
title="Chat with a Concept Steering Model",
|
248 |
+
description="""Steer responses by selecting concepts on the right →\n\nWe are using Gemma-2-2B-it with steering vectors added to the residual stream at layer 20. Our auto-steer steering vector generated is a finetuned Gemma-2-2B model.""",
|
249 |
type="messages",
|
250 |
additional_inputs=[selected_subspaces],
|
251 |
fill_height=True
|
|
|
254 |
# Right side: concept management
|
255 |
with gr.Column(scale=4):
|
256 |
gr.Markdown("## Steer Model Responses")
|
257 |
+
gr.Markdown("Search and then select a concept to steer. The closest match will be automatically selected. If there is no match, we will use our steering vector generator to auto-steer for you!")
|
258 |
# Concept Search and Selection
|
259 |
with gr.Group():
|
260 |
search_box = gr.Textbox(
|
|
|
262 |
placeholder="Find concepts to steer the model (e.g. 'time travel')",
|
263 |
lines=2,
|
264 |
)
|
265 |
+
msg = gr.TextArea(visible=False)
|
266 |
concept_dropdown = gr.Dropdown(
|
267 |
label="Select a concept to steer the model (Click to see more!)",
|
268 |
interactive=True,
|
|
|
278 |
|
279 |
# Wire up events
|
280 |
# When search box changes, update dropdown AND trigger concept selection
|
281 |
+
search_box.input(
|
282 |
update_dropdown_choices,
|
283 |
[search_box],
|
284 |
+
[concept_dropdown, msg]
|
285 |
).then( # Chain the events to automatically add the concept
|
286 |
add_concept_to_list,
|
287 |
[concept_dropdown, concept_magnitude, selected_subspaces],
|
|
|
294 |
[selected_subspaces]
|
295 |
)
|
296 |
|
297 |
+
concept_dropdown.change(
|
298 |
+
add_concept_to_list,
|
299 |
+
[concept_dropdown, concept_magnitude, selected_subspaces],
|
300 |
+
[selected_subspaces]
|
301 |
+
)
|
302 |
+
|
303 |
concept_magnitude.input(
|
304 |
add_concept_to_list,
|
305 |
[concept_dropdown, concept_magnitude, selected_subspaces],
|
style.css
DELETED
@@ -1,19 +0,0 @@
|
|
1 |
-
#alert-message label {
|
2 |
-
font-weight: 700;
|
3 |
-
background-color: #fff3cd;
|
4 |
-
padding: 8px;
|
5 |
-
border-radius: 4px;
|
6 |
-
color: #664d03;
|
7 |
-
display: inline-block;
|
8 |
-
margin-bottom: 8px;
|
9 |
-
}
|
10 |
-
|
11 |
-
#alert-message textarea {
|
12 |
-
background-color: #e8f4ff;
|
13 |
-
border: 1px solid #cce5ff;
|
14 |
-
color: #084298;
|
15 |
-
font-size: 1.1em;
|
16 |
-
padding: 12px;
|
17 |
-
border-radius: 4px;
|
18 |
-
font-weight: 500;
|
19 |
-
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|