File size: 10,793 Bytes
a26206c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import streamlit as st
from transformer_lens import HookedTransformer, utils
from io import StringIO
import sys
import torch
from functools import partial
import plotly.offline as pyo
import plotly.graph_objs as go
import numpy as np
import plotly.express as px
import circuitsvis as cv

# Little bit of front end for model selector

# Radio buttons
model_name = st.sidebar.radio("Model (only use patching for\nsmall (<4L) models due to memory limits)", [
    "gelu-1l",
    "gelu-2l",
    "gelu-3l",
    "gelu-4l",
    "attn-only-1l",
    "attn-only-2l",
    "attn-only-3l",
    "attn-only-4l",
    "solu-1l",
    "solu-2l",
    "solu-3l",
    "solu-4l",
    "solu-6l",
    "solu-8l",
    "solu-10l",
    "solu-12l",
    "gpt2-small",
    "gpt2-medium",
    #"gpt2-large",
    #"gpt2-xl",
    ], index=1)


# Backend code

model = HookedTransformer.from_pretrained(model_name)

def predict_next_token(prompt):
    logits = model(prompt)[0,-1]
    answer_index = logits.argmax()
    answer = model.tokenizer.decode(answer_index)
    answer = f"<b>|{answer}|</b> (answer by {model.cfg.model_name})"
    return answer

def test_prompt(prompt, answer):
    output = StringIO()
    sys.stdout = output
    utils.test_prompt(prompt, answer, model)
    output = output.getvalue()
    return output

def compute_residual_stream_patch(clean_prompt=None, answer=None, corrupt_prompt=None, corrupt_answer=None, layers=None):
    model.reset_hooks()
    clean_answer_index = model.tokenizer.encode(answer)[0]
    corrupt_answer_index = model.tokenizer.encode(corrupt_answer)[0]
    clean_tokens = model.to_str_tokens(clean_prompt)
    _, corrupt_cache = model.run_with_cache(corrupt_prompt)
    # Patching function
    def patch_residual_stream(activations, hook, layer="blocks.6.hook_resid_post", pos=5):
        activations[:, pos, :] = corrupt_cache[layer][:, pos, :]
        return activations
    # Compute logit diffs
    n_layers = len(layers)
    n_pos = len(clean_tokens)
    patching_effect = torch.zeros(n_layers, n_pos)
    for l, layer in enumerate(layers):
        for pos in range(n_pos):
            fwd_hooks = [(layer, partial(patch_residual_stream, layer=layer, pos=pos))]
            prediction_logits = model.run_with_hooks(clean_prompt, fwd_hooks=fwd_hooks)[0, -1]
            patching_effect[l, pos] = prediction_logits[clean_answer_index] - prediction_logits[corrupt_answer_index]
    return patching_effect

def compute_attn_patch(clean_prompt=None, answer=None, corrupt_prompt=None, corrupt_answer=None):
    use_attn_result_prev = model.cfg.use_attn_result
    model.cfg.use_attn_result = True
    clean_answer_index = model.tokenizer.encode(answer)[0]
    corrupt_answer_index = model.tokenizer.encode(corrupt_answer)[0]
    clean_tokens = model.to_str_tokens(clean_prompt)
    _, corrupt_cache = model.run_with_cache(corrupt_prompt)
    # Patching function
    def patch_head_result(activations, hook, head=None, pos=None):
        activations[:, pos, head, :] = corrupt_cache[hook.name][:, pos, head, :]
        return activations

    n_layers = model.cfg.n_layers
    n_heads = model.cfg.n_heads
    n_pos = len(clean_tokens)
    patching_effect = torch.zeros(n_layers*n_heads, n_pos)
    for layer in range(n_layers):
        for head in range(n_heads):
          for pos in range(n_pos):
              fwd_hooks = [(f"blocks.{layer}.attn.hook_result", partial(patch_head_result, head=head, pos=pos))]
              prediction_logits = model.run_with_hooks(clean_prompt, fwd_hooks=fwd_hooks)[0, -1]
              patching_effect[n_heads*layer+head, pos] = prediction_logits[clean_answer_index] - prediction_logits[corrupt_answer_index]
    model.cfg.use_attn_result = use_attn_result_prev
    return patching_effect

def imshow(tensor, xlabel="X", ylabel="Y", zlabel=None, xticks=None, yticks=None, c_midpoint=0.0, c_scale="RdBu", **kwargs):
    tensor = utils.to_numpy(tensor)
    xticks = [str(x) for x in xticks]
    yticks = [str(y) for y in yticks]
    labels = {"x": xlabel, "y": ylabel}
    if zlabel is not None:
        labels["color"] = zlabel
    fig = px.imshow(tensor, x=xticks, y=yticks, labels=labels, color_continuous_midpoint=c_midpoint,
                    color_continuous_scale=c_scale, **kwargs)
    return fig

def plot_residual_stream_patch(clean_prompt=None, answer=None, corrupt_prompt=None, corrupt_answer=None):
    layers = ["blocks.0.hook_resid_pre", *[f"blocks.{i}.hook_resid_post" for i in range(model.cfg.n_layers)]]
    clean_tokens = model.to_str_tokens(clean_prompt)
    token_labels = [f"(pos {i:2}) {t}" for i, t in enumerate(clean_tokens)]
    patching_effect = compute_residual_stream_patch(clean_prompt=clean_prompt, answer=answer, corrupt_prompt=corrupt_prompt, corrupt_answer=corrupt_answer, layers=layers)
    fig = imshow(patching_effect, xticks=token_labels, yticks=layers, xlabel="Position", ylabel="Layer",
       zlabel="Logit Difference", title="Patching residual stream at specific layer and position")
    return fig

def plot_attn_patch(clean_prompt=None, answer=None, corrupt_prompt=None, corrupt_answer=None):
    clean_tokens = model.to_str_tokens(clean_prompt)
    n_layers = model.cfg.n_layers
    n_heads = model.cfg.n_heads
    layerhead_labels = [f"{l}.{h}" for l in range(n_layers) for h in range(n_heads)]
    token_labels = [f"(pos {i:2}) {t}" for i, t in enumerate(clean_tokens)]
    patching_effect = compute_attn_patch(clean_prompt=clean_prompt, answer=answer, corrupt_prompt=corrupt_prompt, corrupt_answer=corrupt_answer)
    return imshow(patching_effect, xticks=token_labels, yticks=layerhead_labels, xlabel="Position", ylabel="Layer.Head",
           zlabel="Logit Difference", title=f"Patching attention outputs for specific layer, head, and position", width=600, height=300+200*n_layers)


# Frontend code
st.title("Simple Trafo Mech Int")
st.subheader("Transformer Mechanistic Interpretability")
st.markdown("Powered by [TransformerLens](https://github.com/neelnanda-io/TransformerLens/)")
st.markdown("For _what_ these plots are, and _why_, see this [tutorial](https://docs.google.com/document/d/1e6cs8d9QNretWvOLsv_KaMp6kSPWpJEW0GWc0nwjqxo/).")

# Predict next token
st.header("Predict the next token")
st.markdown("Just a simple test UI, enter a prompt and the model will predict the next token")
prompt_simple = st.text_input("Prompt:", "Today, the weather is", key="prompt_simple")

if "prompt_simple_output" not in st.session_state:
    st.session_state.prompt_simple_output = None

if st.button("Run model", key="key_button_prompt_simple"):
    res = predict_next_token(prompt_simple)
    st.session_state.prompt_simple_output = res

if st.session_state.prompt_simple_output:
    st.markdown(st.session_state.prompt_simple_output, unsafe_allow_html=True)


# Test prompt
st.header("Verbose test prompt")
st.markdown("Enter a prompt and the correct answer, the model will run the prompt and print the results")

prompt = st.text_input("Prompt:", "The most popular programming language is", key="prompt")
answer = st.text_input("Answer:", " Java", key="answer")

if "test_prompt_output" not in st.session_state:
    st.session_state.test_prompt_output = None

if st.button("Run model", key="key_button_test_prompt"):
    res = test_prompt(prompt, answer)
    st.session_state.test_prompt_output = res
    
if st.session_state.test_prompt_output:
    st.code(st.session_state.test_prompt_output)


# Residual stream patching

st.header("Residual stream patching")
st.markdown("Enter a clean prompt, correct answer, corrupt prompt and corrupt answer, the model will compute the patching effect")

default_clean_prompt = "Her name was Alex Hart. Tomorrow at lunch time Alex"
default_clean_answer = "Hart"
default_corrupt_prompt = "Her name was Alex Carroll. Tomorrow at lunch time Alex"
default_corrupt_answer = "Carroll"

clean_prompt   = st.text_input("Clean Prompt:",   default_clean_prompt)
clean_answer   = st.text_input("Correct Answer:", default_clean_answer)
corrupt_prompt = st.text_input("Corrupt Prompt:", default_corrupt_prompt)
corrupt_answer = st.text_input("Corrupt Answer:", default_corrupt_answer)

if "residual_stream_patch_out" not in st.session_state:
    st.session_state.residual_stream_patch_out = None

if st.button("Run model", key="key_button_residual_stream_patch"):
    fig = plot_residual_stream_patch(clean_prompt=clean_prompt, answer=clean_answer, corrupt_prompt=corrupt_prompt, corrupt_answer=corrupt_answer)
    st.session_state.residual_stream_patch_out = fig

if st.session_state.residual_stream_patch_out:
    st.plotly_chart(st.session_state.residual_stream_patch_out)


# Attention head output

st.header("Attention head output patching")
st.markdown("Enter a clean prompt, correct answer, corrupt prompt and corrupt answer, the model will compute the patching effect")

clean_prompt_attn   = st.text_input("Clean Prompt:",   default_clean_prompt, key="key2_clean_prompt_attn")
clean_answer_attn   = st.text_input("Correct Answer:", default_clean_answer, key="key2_clean_answer_attn")
corrupt_prompt_attn = st.text_input("Corrupt Prompt:", default_corrupt_prompt, key="key2_corrupt_prompt_attn")
corrupt_answer_attn = st.text_input("Corrupt Answer:", default_corrupt_answer, key="key2_corrupt_answer_attn")

if "attn_head_patch_out" not in st.session_state:
    st.session_state.attn_head_patch_out = None

if st.button("Run model", key="key_button_attn_head_patch"):
    fig = plot_attn_patch(clean_prompt=clean_prompt_attn, answer=clean_answer_attn, corrupt_prompt=corrupt_prompt_attn, corrupt_answer=corrupt_answer_attn)
    st.session_state.attn_head_patch_out = fig

if st.session_state.attn_head_patch_out:
    st.plotly_chart(st.session_state.attn_head_patch_out)


# Attention Head Visualization

st.header("Attention Pattern Visualization")
st.markdown("Powered by [CircuitsVis](https://github.com/alan-cooney/CircuitsVis)")
st.markdown("Enter a prompt, show attention patterns")

default_prompt_attn = "Her name was Alex Hart. Tomorrow at lunch time Alex"
prompt_attn   = st.text_input("Prompt:",   default_prompt_attn)

if "attn_html" not in st.session_state:
    st.session_state.attn_html = None

if st.button("Run model", key="key_button_attention_head"):
    _, cache = model.run_with_cache(prompt_attn)
    st.session_state.attn_html = []
    for layer in range(model.cfg.n_layers):
        html = cv.attention.attention_patterns(tokens=model.to_str_tokens(prompt_attn),
                                attention=cache[f'blocks.{layer}.attn.hook_pattern'][0])
        st.session_state.attn_html.append(html.show_code())

if st.session_state.attn_html:
    for layer in range(len(st.session_state.attn_html)):
        st.write(f"Attention patterns Layer {layer}:")
        st.components.v1.html(st.session_state.attn_html[layer], height=500)