Spaces:
Sleeping
Sleeping
s-a-malik
commited on
Commit
·
0120475
1
Parent(s):
b501b77
thread
Browse files
app.py
CHANGED
@@ -26,10 +26,10 @@ DESCRIPTION = """
|
|
26 |
"""
|
27 |
|
28 |
EXAMPLES = [
|
29 |
-
["What is the capital of France?", "
|
30 |
-
["Who landed on the moon?", "
|
31 |
-
["Who is Yarin Gal?", "
|
32 |
-
["Explain the theory of relativity in simple terms.", "
|
33 |
]
|
34 |
|
35 |
if torch.cuda.is_available():
|
@@ -93,28 +93,7 @@ class CustomStreamer(TextIteratorStreamer):
|
|
93 |
|
94 |
|
95 |
|
96 |
-
|
97 |
-
# acc_highlighted_text = ""
|
98 |
-
# for new_text in streamer:
|
99 |
-
# hidden_states = streamer.hidden_states_queue.get()
|
100 |
-
|
101 |
-
# # Semantic Uncertainty Probe
|
102 |
-
# se_token_embeddings = torch.stack([layer[0, -1, :].cpu() for layer in hidden_states])
|
103 |
-
# se_concat_layers = se_token_embeddings.numpy()[se_layer_range[0]:se_layer_range[1]].reshape(-1)
|
104 |
-
# se_probe_pred = se_probe.predict_proba(se_concat_layers.reshape(1, -1))[0][1] * 2 - 1
|
105 |
-
|
106 |
-
# # Accuracy Probe
|
107 |
-
# acc_token_embeddings = torch.stack([layer[0, -1, :].cpu() for layer in hidden_states])
|
108 |
-
# acc_concat_layers = acc_token_embeddings.numpy()[acc_layer_range[0]:acc_layer_range[1]].reshape(-1)
|
109 |
-
# acc_probe_pred = acc_probe.predict_proba(acc_concat_layers.reshape(1, -1))[0][1] * 2 - 1
|
110 |
-
|
111 |
-
# se_new_highlighted_text = highlight_text(new_text, se_probe_pred)
|
112 |
-
# acc_new_highlighted_text = highlight_text(new_text, acc_probe_pred)
|
113 |
-
|
114 |
-
# se_highlighted_text += se_new_highlighted_text
|
115 |
-
# acc_highlighted_text += acc_new_highlighted_text
|
116 |
-
|
117 |
-
# yield se_highlighted_text, acc_highlighted_text
|
118 |
|
119 |
@spaces.GPU
|
120 |
def generate(
|
@@ -137,7 +116,8 @@ def generate(
|
|
137 |
gr.Warning(f"Trimmed input from conversation as it was longer than {MAX_INPUT_TOKEN_LENGTH} tokens.")
|
138 |
input_ids = input_ids.to(model.device)
|
139 |
|
140 |
-
streamer = TextIteratorStreamer(tokenizer, timeout=10.0, skip_prompt=True, skip_special_tokens=True)
|
|
|
141 |
generation_kwargs = dict(
|
142 |
input_ids=input_ids,
|
143 |
max_new_tokens=max_new_tokens,
|
@@ -150,41 +130,84 @@ def generate(
|
|
150 |
output_hidden_states=True,
|
151 |
return_dict_in_generate=True,
|
152 |
)
|
153 |
-
|
154 |
-
|
155 |
-
|
156 |
-
outputs = model.generate(**generation_kwargs)
|
157 |
-
generated_tokens = outputs.sequences[0, input_ids.shape[1]:]
|
158 |
-
generated_text = tokenizer.decode(generated_tokens, skip_special_tokens=True)
|
159 |
-
# hidden states
|
160 |
-
hidden = outputs.hidden_states # list of tensors, one for each token, then (batch size, sequence length, hidden size)
|
161 |
-
|
162 |
-
# TODO do this loop on the fly instead of waiting for the whole generation
|
163 |
se_highlighted_text = ""
|
164 |
acc_highlighted_text = ""
|
165 |
-
for
|
166 |
-
|
167 |
# Semantic Uncertainty Probe
|
168 |
-
token_embeddings = torch.stack([generated_token[0, 0, :].cpu() for generated_token in
|
169 |
se_concat_layers = token_embeddings[se_layer_range[0]:se_layer_range[1]].reshape(-1)
|
170 |
se_probe_pred = se_probe.predict_proba(se_concat_layers.reshape(1, -1))[0][1] * 2 - 1
|
171 |
|
172 |
# Accuracy Probe
|
173 |
-
# acc_token_embeddings = torch.stack([layer[0, -1, :].cpu() for layer in hidden_states])
|
174 |
acc_concat_layers = token_embeddings[acc_layer_range[0]:acc_layer_range[1]].reshape(-1)
|
175 |
acc_probe_pred = (1 - acc_probe.predict_proba(acc_concat_layers.reshape(1, -1))[0][1]) * 2 - 1
|
176 |
|
177 |
-
|
178 |
-
output_word = tokenizer.decode(output_id)
|
179 |
-
print(output_id, output_word, se_probe_pred, acc_probe_pred)
|
180 |
|
181 |
-
se_new_highlighted_text = highlight_text(
|
182 |
-
acc_new_highlighted_text = highlight_text(
|
183 |
se_highlighted_text += f" {se_new_highlighted_text}"
|
184 |
acc_highlighted_text += f" {acc_new_highlighted_text}"
|
185 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
186 |
# yield se_highlighted_text, acc_highlighted_text
|
187 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
188 |
|
189 |
|
190 |
|
@@ -215,7 +238,7 @@ with gr.Blocks(title="Llama-2 7B Chat with Dual Probes", css="footer {visibility
|
|
215 |
|
216 |
with gr.Column():
|
217 |
max_new_tokens = gr.Slider(label="Max new tokens", minimum=1, maximum=MAX_MAX_NEW_TOKENS, step=1, value=DEFAULT_MAX_NEW_TOKENS)
|
218 |
-
temperature = gr.Slider(label="Temperature", minimum=0.
|
219 |
top_p = gr.Slider(label="Top-p (nucleus sampling)", minimum=0.05, maximum=1.0, step=0.05, value=0.9)
|
220 |
top_k = gr.Slider(label="Top-k", minimum=1, maximum=1000, step=1, value=50)
|
221 |
repetition_penalty = gr.Slider(label="Repetition penalty", minimum=1.0, maximum=2.0, step=0.05, value=1.2)
|
@@ -243,7 +266,6 @@ with gr.Blocks(title="Llama-2 7B Chat with Dual Probes", css="footer {visibility
|
|
243 |
inputs=[message, system_prompt],
|
244 |
outputs=[se_output, acc_output],
|
245 |
fn=generate,
|
246 |
-
|
247 |
)
|
248 |
|
249 |
generate_btn.click(
|
@@ -252,59 +274,6 @@ with gr.Blocks(title="Llama-2 7B Chat with Dual Probes", css="footer {visibility
|
|
252 |
outputs=[se_output, acc_output]
|
253 |
)
|
254 |
|
255 |
-
# chat_interface = gr.ChatInterface(
|
256 |
-
# fn=generate,
|
257 |
-
# additional_inputs=[
|
258 |
-
# gr.Textbox(label="System prompt", lines=6),
|
259 |
-
# gr.Slider(
|
260 |
-
# label="Max new tokens",
|
261 |
-
# minimum=1,
|
262 |
-
# maximum=MAX_MAX_NEW_TOKENS,
|
263 |
-
# step=1,
|
264 |
-
# value=DEFAULT_MAX_NEW_TOKENS,
|
265 |
-
# ),
|
266 |
-
# gr.Slider(
|
267 |
-
# label="Temperature",
|
268 |
-
# minimum=0.1,
|
269 |
-
# maximum=4.0,
|
270 |
-
# step=0.1,
|
271 |
-
# value=0.6,
|
272 |
-
# ),
|
273 |
-
# gr.Slider(
|
274 |
-
# label="Top-p (nucleus sampling)",
|
275 |
-
# minimum=0.05,
|
276 |
-
# maximum=1.0,
|
277 |
-
# step=0.05,
|
278 |
-
# value=0.9,
|
279 |
-
# ),
|
280 |
-
# gr.Slider(
|
281 |
-
# label="Top-k",
|
282 |
-
# minimum=1,
|
283 |
-
# maximum=1000,
|
284 |
-
# step=1,
|
285 |
-
# value=50,
|
286 |
-
# ),
|
287 |
-
# gr.Slider(
|
288 |
-
# label="Repetition penalty",
|
289 |
-
# minimum=1.0,
|
290 |
-
# maximum=2.0,
|
291 |
-
# step=0.05,
|
292 |
-
# value=1.2,
|
293 |
-
# ),
|
294 |
-
# ],
|
295 |
-
# stop_btn=None,
|
296 |
-
# examples=[
|
297 |
-
# ["What is the capital of France?"],
|
298 |
-
# ["Who landed on the moon?"],
|
299 |
-
# ["Who is Yarin Gal?"]
|
300 |
-
# ],
|
301 |
-
# title="Llama-2 7B Chat with Streamable Semantic Uncertainty Probe",
|
302 |
-
# description=DESCRIPTION,
|
303 |
-
# )
|
304 |
-
|
305 |
-
# if __name__ == "__main__":
|
306 |
-
# chat_interface.launch()
|
307 |
-
|
308 |
|
309 |
if __name__ == "__main__":
|
310 |
demo.launch()
|
|
|
26 |
"""
|
27 |
|
28 |
EXAMPLES = [
|
29 |
+
["What is the capital of France?", ""],
|
30 |
+
["Who landed on the moon?", ""],
|
31 |
+
["Who is Yarin Gal?", ""],
|
32 |
+
["Explain the theory of relativity in simple terms.", ""],
|
33 |
]
|
34 |
|
35 |
if torch.cuda.is_available():
|
|
|
93 |
|
94 |
|
95 |
|
96 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
97 |
|
98 |
@spaces.GPU
|
99 |
def generate(
|
|
|
116 |
gr.Warning(f"Trimmed input from conversation as it was longer than {MAX_INPUT_TOKEN_LENGTH} tokens.")
|
117 |
input_ids = input_ids.to(model.device)
|
118 |
|
119 |
+
# streamer = TextIteratorStreamer(tokenizer, timeout=10.0, skip_prompt=True, skip_special_tokens=True)
|
120 |
+
streamer = CustomStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True)
|
121 |
generation_kwargs = dict(
|
122 |
input_ids=input_ids,
|
123 |
max_new_tokens=max_new_tokens,
|
|
|
130 |
output_hidden_states=True,
|
131 |
return_dict_in_generate=True,
|
132 |
)
|
133 |
+
# with threading
|
134 |
+
thread = Thread(target=model.generate, kwargs=generation_kwargs)
|
135 |
+
thread.start()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
136 |
se_highlighted_text = ""
|
137 |
acc_highlighted_text = ""
|
138 |
+
for new_text in streamer:
|
139 |
+
hidden_states = streamer.hidden_states_queue.get()
|
140 |
# Semantic Uncertainty Probe
|
141 |
+
token_embeddings = torch.stack([generated_token[0, 0, :].cpu() for generated_token in hidden_states]).numpy() # (num_layers, hidden_size)
|
142 |
se_concat_layers = token_embeddings[se_layer_range[0]:se_layer_range[1]].reshape(-1)
|
143 |
se_probe_pred = se_probe.predict_proba(se_concat_layers.reshape(1, -1))[0][1] * 2 - 1
|
144 |
|
145 |
# Accuracy Probe
|
|
|
146 |
acc_concat_layers = token_embeddings[acc_layer_range[0]:acc_layer_range[1]].reshape(-1)
|
147 |
acc_probe_pred = (1 - acc_probe.predict_proba(acc_concat_layers.reshape(1, -1))[0][1]) * 2 - 1
|
148 |
|
149 |
+
print(new_text, se_probe_pred, acc_probe_pred)
|
|
|
|
|
150 |
|
151 |
+
se_new_highlighted_text = highlight_text(new_text, se_probe_pred)
|
152 |
+
acc_new_highlighted_text = highlight_text(new_text, acc_probe_pred)
|
153 |
se_highlighted_text += f" {se_new_highlighted_text}"
|
154 |
acc_highlighted_text += f" {acc_new_highlighted_text}"
|
155 |
|
156 |
+
yield se_highlighted_text, acc_highlighted_text
|
157 |
+
|
158 |
+
# Semantic Uncertainty Probe
|
159 |
+
# se_token_embeddings = torch.stack([layer[0, -1, :].cpu() for layer in hidden_states])
|
160 |
+
# se_concat_layers = se_token_embeddings.numpy()[se_layer_range[0]:se_layer_range[1]].reshape(-1)
|
161 |
+
# se_probe_pred = se_probe.predict_proba(se_concat_layers.reshape(1, -1))[0][1] * 2 - 1
|
162 |
+
|
163 |
+
# # Accuracy Probe
|
164 |
+
# acc_token_embeddings = torch.stack([layer[0, -1, :].cpu() for layer in hidden_states])
|
165 |
+
# acc_concat_layers = acc_token_embeddings.numpy()[acc_layer_range[0]:acc_layer_range[1]].reshape(-1)
|
166 |
+
# acc_probe_pred = acc_probe.predict_proba(acc_concat_layers.reshape(1, -1))[0][1] * 2 - 1
|
167 |
+
|
168 |
+
# se_new_highlighted_text = highlight_text(new_text, se_probe_pred)
|
169 |
+
# acc_new_highlighted_text = highlight_text(new_text, acc_probe_pred)
|
170 |
+
|
171 |
+
# se_highlighted_text += se_new_highlighted_text
|
172 |
+
# acc_highlighted_text += acc_new_highlighted_text
|
173 |
+
|
174 |
# yield se_highlighted_text, acc_highlighted_text
|
175 |
+
|
176 |
+
# Generate without threading
|
177 |
+
# with torch.no_grad():
|
178 |
+
# outputs = model.generate(**generation_kwargs)
|
179 |
+
# generated_tokens = outputs.sequences[0, input_ids.shape[1]:]
|
180 |
+
# generated_text = tokenizer.decode(generated_tokens, skip_special_tokens=True)
|
181 |
+
# # hidden states
|
182 |
+
# hidden = outputs.hidden_states # list of tensors, one for each token, then (batch size, sequence length, hidden size)
|
183 |
+
|
184 |
+
# # TODO do this loop on the fly instead of waiting for the whole generation
|
185 |
+
# se_highlighted_text = ""
|
186 |
+
# acc_highlighted_text = ""
|
187 |
+
|
188 |
+
# for i in range(1, len(hidden)):
|
189 |
+
|
190 |
+
# # Semantic Uncertainty Probe
|
191 |
+
# token_embeddings = torch.stack([generated_token[0, 0, :].cpu() for generated_token in hidden[i]]).numpy() # (num_layers, hidden_size)
|
192 |
+
# se_concat_layers = token_embeddings[se_layer_range[0]:se_layer_range[1]].reshape(-1)
|
193 |
+
# se_probe_pred = se_probe.predict_proba(se_concat_layers.reshape(1, -1))[0][1] * 2 - 1
|
194 |
+
|
195 |
+
# # Accuracy Probe
|
196 |
+
# # acc_token_embeddings = torch.stack([layer[0, -1, :].cpu() for layer in hidden_states])
|
197 |
+
# acc_concat_layers = token_embeddings[acc_layer_range[0]:acc_layer_range[1]].reshape(-1)
|
198 |
+
# acc_probe_pred = (1 - acc_probe.predict_proba(acc_concat_layers.reshape(1, -1))[0][1]) * 2 - 1
|
199 |
+
|
200 |
+
# output_id = outputs.sequences[0, input_ids.shape[1]+i]
|
201 |
+
# output_word = tokenizer.decode(output_id)
|
202 |
+
# print(output_id, output_word, se_probe_pred, acc_probe_pred)
|
203 |
+
|
204 |
+
# se_new_highlighted_text = highlight_text(output_word, se_probe_pred)
|
205 |
+
# acc_new_highlighted_text = highlight_text(output_word, acc_probe_pred)
|
206 |
+
# se_highlighted_text += f" {se_new_highlighted_text}"
|
207 |
+
# acc_highlighted_text += f" {acc_new_highlighted_text}"
|
208 |
+
|
209 |
+
# # yield se_highlighted_text, acc_highlighted_text
|
210 |
+
# return se_highlighted_text, acc_highlighted_text
|
211 |
|
212 |
|
213 |
|
|
|
238 |
|
239 |
with gr.Column():
|
240 |
max_new_tokens = gr.Slider(label="Max new tokens", minimum=1, maximum=MAX_MAX_NEW_TOKENS, step=1, value=DEFAULT_MAX_NEW_TOKENS)
|
241 |
+
temperature = gr.Slider(label="Temperature", minimum=0.01, maximum=2.0, step=0.1, value=0.01)
|
242 |
top_p = gr.Slider(label="Top-p (nucleus sampling)", minimum=0.05, maximum=1.0, step=0.05, value=0.9)
|
243 |
top_k = gr.Slider(label="Top-k", minimum=1, maximum=1000, step=1, value=50)
|
244 |
repetition_penalty = gr.Slider(label="Repetition penalty", minimum=1.0, maximum=2.0, step=0.05, value=1.2)
|
|
|
266 |
inputs=[message, system_prompt],
|
267 |
outputs=[se_output, acc_output],
|
268 |
fn=generate,
|
|
|
269 |
)
|
270 |
|
271 |
generate_btn.click(
|
|
|
274 |
outputs=[se_output, acc_output]
|
275 |
)
|
276 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
277 |
|
278 |
if __name__ == "__main__":
|
279 |
demo.launch()
|