Upload spa.py with huggingface_hub
Browse files
spa.py
ADDED
@@ -0,0 +1,530 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import time
|
2 |
+
import re
|
3 |
+
import torch
|
4 |
+
import torch.nn.functional as F
|
5 |
+
from transformers import (
|
6 |
+
AutoModelForCausalLM,
|
7 |
+
AutoTokenizer,
|
8 |
+
BitsAndBytesConfig,
|
9 |
+
LogitsProcessor,
|
10 |
+
GenerationConfig,
|
11 |
+
TextIteratorStreamer,
|
12 |
+
)
|
13 |
+
|
14 |
+
# --- Helper Function for Input Preparation ---
|
15 |
+
|
16 |
+
def create_masked_attention(input_ids, target_strings, tokenizer):
|
17 |
+
"""
|
18 |
+
Creates an attention mask where tokens corresponding to any of the target strings have 0 attention.
|
19 |
+
"""
|
20 |
+
# Ensure input_ids is 2D
|
21 |
+
if len(input_ids.shape) == 1:
|
22 |
+
input_ids = input_ids.unsqueeze(0)
|
23 |
+
|
24 |
+
# Create default attention mask (all 1s)
|
25 |
+
attention_mask = torch.ones_like(input_ids)
|
26 |
+
|
27 |
+
# Convert single string to list for uniform processing
|
28 |
+
if isinstance(target_strings, str):
|
29 |
+
target_strings = [target_strings]
|
30 |
+
|
31 |
+
# Get the input IDs as a list
|
32 |
+
input_ids_list = input_ids[0].tolist()
|
33 |
+
|
34 |
+
# Decode each token individually for comparison
|
35 |
+
token_texts = []
|
36 |
+
for token_id in input_ids_list:
|
37 |
+
token_texts.append(tokenizer.decode([token_id]))
|
38 |
+
|
39 |
+
|
40 |
+
|
41 |
+
masked_indices = []
|
42 |
+
|
43 |
+
# Try tokenizing each target string to find its exact token representation
|
44 |
+
for target_string in target_strings:
|
45 |
+
if not target_string:
|
46 |
+
continue
|
47 |
+
|
48 |
+
# Tokenize the target string to get its expected token IDs
|
49 |
+
target_ids = tokenizer.encode(target_string, add_special_tokens=False)
|
50 |
+
target_tokens = [tokenizer.decode([id]) for id in target_ids]
|
51 |
+
|
52 |
+
|
53 |
+
# First approach: Direct token sequence matching
|
54 |
+
# Look for the sequence of tokens in the input
|
55 |
+
for i in range(len(token_texts) - len(target_tokens) + 1):
|
56 |
+
# Check if this position starts a matching sequence
|
57 |
+
all_match = True
|
58 |
+
for j, target_token in enumerate(target_tokens):
|
59 |
+
if i+j >= len(token_texts) or target_token != token_texts[i+j]:
|
60 |
+
all_match = False
|
61 |
+
break
|
62 |
+
|
63 |
+
if all_match:
|
64 |
+
for j in range(len(target_tokens)):
|
65 |
+
attention_mask[0, i+j] = 0
|
66 |
+
masked_indices.append(i+j)
|
67 |
+
|
68 |
+
# Second approach: Look for individual tokens that make up the target
|
69 |
+
for i, token_text in enumerate(token_texts):
|
70 |
+
if token_text.strip() in target_tokens:
|
71 |
+
attention_mask[0, i] = 0
|
72 |
+
masked_indices.append(i)
|
73 |
+
|
74 |
+
# Third approach: If the target is split between tokens, try to detect it
|
75 |
+
# For example 'MASKTOKEN' might be split as ' MASK' and 'TOKEN'
|
76 |
+
if len(target_tokens) == 1 and len(target_tokens[0]) > 2: # Only for substantial single tokens
|
77 |
+
# Look for token pairs that might contain the target
|
78 |
+
for i in range(len(token_texts) - 1):
|
79 |
+
pair = token_texts[i].strip() + token_texts[i+1].strip()
|
80 |
+
if target_string in pair:
|
81 |
+
attention_mask[0, i] = 0
|
82 |
+
attention_mask[0, i+1] = 0
|
83 |
+
masked_indices.extend([i, i+1])
|
84 |
+
|
85 |
+
# Check for triplet if possible
|
86 |
+
if i < len(token_texts) - 2:
|
87 |
+
triplet = token_texts[i].strip() + token_texts[i+1].strip() + token_texts[i+2].strip()
|
88 |
+
if target_string in triplet:
|
89 |
+
attention_mask[0, i] = 0
|
90 |
+
attention_mask[0, i+1] = 0
|
91 |
+
attention_mask[0, i+2] = 0
|
92 |
+
masked_indices.extend([i, i+1, i+2])
|
93 |
+
|
94 |
+
|
95 |
+
# Print the final mask
|
96 |
+
mask_positions = list(set(masked_indices)) # Remove duplicates
|
97 |
+
mask_positions.sort()
|
98 |
+
|
99 |
+
if mask_positions:
|
100 |
+
masked_text = [token_texts[idx] for idx in mask_positions]
|
101 |
+
else:
|
102 |
+
print("WARNING: No tokens were masked!")
|
103 |
+
# Last resort - just mask any token containing part of the target
|
104 |
+
for target_string in target_strings:
|
105 |
+
for i, token_text in enumerate(token_texts):
|
106 |
+
if (target_string in token_text) or (token_text.strip() in target_string and len(token_text.strip()) > 2):
|
107 |
+
attention_mask[0, i] = 0
|
108 |
+
masked_indices.append(i)
|
109 |
+
|
110 |
+
# Check again
|
111 |
+
mask_positions = list(set(masked_indices))
|
112 |
+
mask_positions.sort()
|
113 |
+
|
114 |
+
return attention_mask
|
115 |
+
|
116 |
+
|
117 |
+
def preprocess_anchors(anchors):
|
118 |
+
# remove duplicates in anchors
|
119 |
+
anchors = list(set(anchors))
|
120 |
+
# remove "", " " in anchors
|
121 |
+
anchors = [anchor for anchor in anchors if anchor != "" and anchor != " "]
|
122 |
+
# sort the anchors by length
|
123 |
+
anchors = sorted(anchors, key=len, reverse=True)
|
124 |
+
return anchors
|
125 |
+
|
126 |
+
|
127 |
+
# Define a wrapper function to handle different cases
|
128 |
+
# The provided anchors are viewed as global anchors
|
129 |
+
def format_spa_input(input, anchors, mask_token, whole_word_only=True):
|
130 |
+
# check if the input is a string or a list of messages
|
131 |
+
if isinstance(input, str):
|
132 |
+
# 1. Collect all anchors
|
133 |
+
current_anchors = list(anchors) # Start with global anchors
|
134 |
+
tag_anchors = []
|
135 |
+
if re.search(r"<anchor>", input):
|
136 |
+
tag_anchors = re.findall(r"<anchor>(.*?)</anchor>", input, flags=re.DOTALL)
|
137 |
+
current_anchors.extend(tag_anchors)
|
138 |
+
|
139 |
+
# 2. Clean the input string (remove tags)
|
140 |
+
cleaned_input = re.sub(r"<anchor>|</anchor>", "", input)
|
141 |
+
|
142 |
+
# 3. Preprocess all collected anchors (unique, non-empty, sorted desc)
|
143 |
+
final_anchors = preprocess_anchors(current_anchors)
|
144 |
+
|
145 |
+
# 4. Escape anchors for regex and build pattern (longest first)
|
146 |
+
masked_input = cleaned_input # Initialize with cleaned input
|
147 |
+
if final_anchors:
|
148 |
+
if whole_word_only:
|
149 |
+
# Use lookarounds to assert boundaries without consuming them (Fix 1)
|
150 |
+
escaped_anchors = [rf"(?<!\w){re.escape(a)}(?!\w)" for a in final_anchors]
|
151 |
+
else:
|
152 |
+
escaped_anchors = [re.escape(a) for a in final_anchors]
|
153 |
+
|
154 |
+
pattern = "|".join(escaped_anchors)
|
155 |
+
# 5. Perform anchor replacement in one pass
|
156 |
+
masked_input = re.sub(pattern, mask_token, cleaned_input)
|
157 |
+
|
158 |
+
# 6. Post-processing: Merge consecutive mask tokens (separated by space)
|
159 |
+
if mask_token: # Avoid processing if mask_token is empty
|
160 |
+
escaped_mask_token = re.escape(mask_token)
|
161 |
+
# Improved merging logic (Fix 2)
|
162 |
+
merge_pattern = f"{escaped_mask_token}\s+{escaped_mask_token}"
|
163 |
+
while re.search(merge_pattern, masked_input):
|
164 |
+
masked_input = re.sub(merge_pattern, mask_token, masked_input)
|
165 |
+
# Optional: merge masks without space if needed, e.g., mask_token+mask_token -> mask_token
|
166 |
+
# merge_pattern_no_space = f"{escaped_mask_token}{escaped_mask_token}"
|
167 |
+
# while re.search(merge_pattern_no_space, masked_input):
|
168 |
+
# masked_input = re.sub(merge_pattern_no_space, mask_token, masked_input)
|
169 |
+
|
170 |
+
return cleaned_input, masked_input
|
171 |
+
|
172 |
+
elif isinstance(input, list):
|
173 |
+
cleaned_input_list = []
|
174 |
+
masked_input_list = []
|
175 |
+
|
176 |
+
for msg in input:
|
177 |
+
msg_copy = msg.copy() # Work on a copy
|
178 |
+
content = msg_copy.get("content", "")
|
179 |
+
|
180 |
+
# 1. Collect all anchors for this message
|
181 |
+
current_anchors = list(anchors) # Start with global anchors
|
182 |
+
if "anchors" in msg_copy:
|
183 |
+
dict_anchors = msg_copy.get("anchors", [])
|
184 |
+
if isinstance(dict_anchors, list):
|
185 |
+
current_anchors.extend(dict_anchors)
|
186 |
+
tag_anchors = []
|
187 |
+
if re.search(r"<anchor>", content):
|
188 |
+
tag_anchors = re.findall(r"<anchor>(.*?)</anchor>", content, flags=re.DOTALL)
|
189 |
+
current_anchors.extend(tag_anchors)
|
190 |
+
|
191 |
+
# 2. Clean the message content (remove tags)
|
192 |
+
cleaned_content = re.sub(r"<anchor>|</anchor>", "", content)
|
193 |
+
|
194 |
+
# 3. Preprocess all collected anchors for this message
|
195 |
+
final_anchors = preprocess_anchors(current_anchors)
|
196 |
+
|
197 |
+
# 4. Escape anchors, build pattern, and replace in one pass
|
198 |
+
masked_content = cleaned_content # Initialize
|
199 |
+
if final_anchors:
|
200 |
+
if whole_word_only:
|
201 |
+
# Use lookarounds to assert boundaries without consuming them (Fix 1)
|
202 |
+
escaped_anchors = [rf"(?<!\w){re.escape(a)}(?!\w)" for a in final_anchors]
|
203 |
+
else:
|
204 |
+
escaped_anchors = [re.escape(a) for a in final_anchors]
|
205 |
+
|
206 |
+
pattern = "|".join(escaped_anchors)
|
207 |
+
masked_content = re.sub(pattern, mask_token, cleaned_content)
|
208 |
+
|
209 |
+
# 5. Post-processing: Merge consecutive mask tokens (separated by space) for this message
|
210 |
+
if mask_token:
|
211 |
+
escaped_mask_token = re.escape(mask_token)
|
212 |
+
# Improved merging logic (Fix 2)
|
213 |
+
merge_pattern = f"{escaped_mask_token}\s+{escaped_mask_token}"
|
214 |
+
while re.search(merge_pattern, masked_content):
|
215 |
+
masked_content = re.sub(merge_pattern, mask_token, masked_content)
|
216 |
+
# Optional: merge masks without space if needed
|
217 |
+
# merge_pattern_no_space = f"{escaped_mask_token}{escaped_mask_token}"
|
218 |
+
# while re.search(merge_pattern_no_space, masked_content):
|
219 |
+
# masked_content = re.sub(merge_pattern_no_space, mask_token, masked_content)
|
220 |
+
|
221 |
+
# 6. Prepare output dictionaries
|
222 |
+
final_cleaned_msg = msg_copy.copy()
|
223 |
+
final_cleaned_msg["content"] = cleaned_content
|
224 |
+
if "anchors" in final_cleaned_msg:
|
225 |
+
del final_cleaned_msg["anchors"]
|
226 |
+
|
227 |
+
final_masked_msg = msg_copy.copy()
|
228 |
+
final_masked_msg["content"] = masked_content
|
229 |
+
if "anchors" in final_masked_msg:
|
230 |
+
del final_masked_msg["anchors"]
|
231 |
+
|
232 |
+
cleaned_input_list.append(final_cleaned_msg)
|
233 |
+
masked_input_list.append(final_masked_msg)
|
234 |
+
|
235 |
+
return cleaned_input_list, masked_input_list
|
236 |
+
else:
|
237 |
+
raise ValueError("Invalid input type. Must be string or list of dictionaries.")
|
238 |
+
|
239 |
+
|
240 |
+
def get_mask_messages(messages, mask_token):
|
241 |
+
mask_msg = messages.copy() # get a copy of the messages
|
242 |
+
|
243 |
+
# Debug anchor count
|
244 |
+
for msg in mask_msg:
|
245 |
+
if "anchors" in msg:
|
246 |
+
# Debug pre-replacement content
|
247 |
+
original_content = msg["content"]
|
248 |
+
|
249 |
+
# Sort anchors by length (descending) to replace longest matches first
|
250 |
+
anchors = sorted(msg["anchors"], key=len, reverse=True)
|
251 |
+
|
252 |
+
for anchor in anchors:
|
253 |
+
if anchor in msg["content"]:
|
254 |
+
# Replace the anchor with mask token
|
255 |
+
msg["content"] = msg["content"].replace(anchor, mask_token)
|
256 |
+
|
257 |
+
# Debug post-replacement content
|
258 |
+
if original_content == msg["content"]:
|
259 |
+
print(f"WARNING: No anchors were replaced in message: {original_content[:50]}...")
|
260 |
+
print(f"Anchors: {anchors}")
|
261 |
+
|
262 |
+
return mask_msg
|
263 |
+
|
264 |
+
|
265 |
+
def convert_to_tensor_format(inputs, device=None):
|
266 |
+
# Case 1: Already a tensor in correct format
|
267 |
+
if isinstance(inputs, torch.Tensor) and len(inputs.shape) == 2:
|
268 |
+
if device is not None:
|
269 |
+
inputs = inputs.to(device)
|
270 |
+
return inputs
|
271 |
+
|
272 |
+
# Case 2: Object with input_ids attribute
|
273 |
+
if hasattr(inputs, 'input_ids'):
|
274 |
+
inputs = inputs.input_ids
|
275 |
+
|
276 |
+
# Case 3: Dictionary with input_ids key
|
277 |
+
elif isinstance(inputs, dict) and 'input_ids' in inputs:
|
278 |
+
inputs = inputs['input_ids']
|
279 |
+
|
280 |
+
# Case 4: List of token IDs
|
281 |
+
elif isinstance(inputs, list):
|
282 |
+
inputs = torch.tensor([inputs], device=device)
|
283 |
+
|
284 |
+
# Case 5: Single tensor but needs reshaping
|
285 |
+
elif isinstance(inputs, torch.Tensor):
|
286 |
+
if len(inputs.shape) == 1:
|
287 |
+
inputs = inputs.unsqueeze(0)
|
288 |
+
|
289 |
+
# Ensure it's on the correct device
|
290 |
+
if isinstance(inputs, torch.Tensor) and device is not None:
|
291 |
+
inputs = inputs.to(device)
|
292 |
+
|
293 |
+
return inputs
|
294 |
+
|
295 |
+
def create_default_attention_mask(input_ids, device=None):
|
296 |
+
"""
|
297 |
+
Creates a default attention mask (all 1s) for the given input_ids tensor.
|
298 |
+
|
299 |
+
Args:
|
300 |
+
input_ids (torch.Tensor): The input IDs tensor, shape (batch_size, seq_len)
|
301 |
+
device: The device to place the attention mask on
|
302 |
+
|
303 |
+
Returns:
|
304 |
+
torch.Tensor: Attention mask with the same shape as input_ids, all values set to 1
|
305 |
+
"""
|
306 |
+
# Ensure input_ids is on the right device if specified
|
307 |
+
if device is not None and input_ids.device != device:
|
308 |
+
input_ids = input_ids.to(device)
|
309 |
+
|
310 |
+
# Create attention mask filled with 1s (all tokens attend to all positions)
|
311 |
+
attention_mask = torch.ones_like(input_ids)
|
312 |
+
|
313 |
+
return attention_mask
|
314 |
+
|
315 |
+
def spa_tokenize(prompt_with_anchors, global_anchors, tokenizer, device):
|
316 |
+
|
317 |
+
# Set pad token if missing
|
318 |
+
if tokenizer.pad_token is None:
|
319 |
+
print("Setting pad token to EOS token")
|
320 |
+
tokenizer.pad_token = tokenizer.eos_token
|
321 |
+
# Remove reference to global model variable
|
322 |
+
# model.config.pad_token_id = model.config.eos_token_id
|
323 |
+
|
324 |
+
if tokenizer.mask_token:
|
325 |
+
mask_token = tokenizer.mask_token
|
326 |
+
else:
|
327 |
+
mask_token = "MASKTOKEN"
|
328 |
+
|
329 |
+
|
330 |
+
main_prompt, aux_prompt = format_spa_input(
|
331 |
+
input=prompt_with_anchors,
|
332 |
+
anchors=global_anchors,
|
333 |
+
mask_token=mask_token,
|
334 |
+
whole_word_only=False
|
335 |
+
)
|
336 |
+
|
337 |
+
|
338 |
+
# detect if tokenizer has chat_template
|
339 |
+
if isinstance(main_prompt, list):
|
340 |
+
# Expected for chat models
|
341 |
+
# print("--- Message list processed by chat template")
|
342 |
+
if hasattr(tokenizer, "chat_template") and tokenizer.chat_template:
|
343 |
+
|
344 |
+
main_inputs = tokenizer.apply_chat_template(
|
345 |
+
main_prompt,
|
346 |
+
tokenize=True,
|
347 |
+
add_generation_prompt=True,
|
348 |
+
return_tensors="pt"
|
349 |
+
).to(device)
|
350 |
+
|
351 |
+
aux_inputs = tokenizer.apply_chat_template(
|
352 |
+
aux_prompt,
|
353 |
+
tokenize=True,
|
354 |
+
add_generation_prompt=True,
|
355 |
+
return_tensors="pt"
|
356 |
+
).to(device)
|
357 |
+
|
358 |
+
else:
|
359 |
+
# non-chat models, need to convert to a string prompt
|
360 |
+
# print("--- Message list processed by flat prompt")
|
361 |
+
flat_prompt_main = ""
|
362 |
+
for msg in main_prompt:
|
363 |
+
flat_prompt_main += f"{msg['role']}: {msg['content']}\n"
|
364 |
+
flat_prompt_main += "Assistant: " # Add assistant prefix for generation
|
365 |
+
|
366 |
+
flat_prompt_aux = ""
|
367 |
+
for msg in aux_prompt:
|
368 |
+
flat_prompt_aux += f"{msg['role']}: {msg['content']}\n"
|
369 |
+
flat_prompt_aux += "Assistant: " # Add assistant prefix for generation
|
370 |
+
|
371 |
+
# Tokenize the flattened prompts
|
372 |
+
main_inputs = tokenizer(flat_prompt_main, return_tensors="pt").to(device)
|
373 |
+
aux_inputs = tokenizer(flat_prompt_aux, return_tensors="pt").to(device)
|
374 |
+
|
375 |
+
# User provides a string prompt
|
376 |
+
elif isinstance(prompt_with_anchors, str):
|
377 |
+
if hasattr(tokenizer, "chat_template") and tokenizer.chat_template:
|
378 |
+
# print("--- String prompt processed by chat template")
|
379 |
+
|
380 |
+
# If user only provides a string prompt, we need to convert it to a chat prompt
|
381 |
+
main_prompt = [{"role": "user", "content": main_prompt}]
|
382 |
+
aux_prompt = [{"role": "user", "content": aux_prompt}]
|
383 |
+
|
384 |
+
main_inputs = tokenizer.apply_chat_template(
|
385 |
+
main_prompt,
|
386 |
+
tokenize=True,
|
387 |
+
add_generation_prompt=True,
|
388 |
+
return_tensors="pt"
|
389 |
+
).to(device)
|
390 |
+
|
391 |
+
aux_inputs = tokenizer.apply_chat_template(
|
392 |
+
aux_prompt,
|
393 |
+
tokenize=True,
|
394 |
+
add_generation_prompt=True,
|
395 |
+
return_tensors="pt"
|
396 |
+
).to(device)
|
397 |
+
|
398 |
+
else:
|
399 |
+
# non-chat models, need to convert to a string prompt
|
400 |
+
# print("--- String prompt processed by flat prompt")
|
401 |
+
main_inputs = tokenizer(main_prompt, return_tensors="pt").to(device)
|
402 |
+
aux_inputs = tokenizer(aux_prompt, return_tensors="pt").to(device)
|
403 |
+
|
404 |
+
else:
|
405 |
+
raise ValueError("Invalid prompt format")
|
406 |
+
|
407 |
+
# Make sure the returned input_ids follow the expected format: tensor([[1, 2, 3]], device='x')
|
408 |
+
# Handle all possible tokenizer output formats
|
409 |
+
|
410 |
+
main_inputs = convert_to_tensor_format(main_inputs, device)
|
411 |
+
aux_inputs = convert_to_tensor_format(aux_inputs, device)
|
412 |
+
|
413 |
+
return main_inputs, aux_inputs, mask_token
|
414 |
+
|
415 |
+
|
416 |
+
class SPALogitsProcessor(LogitsProcessor):
|
417 |
+
"""Processor that combines logits from a main and auxiliary model."""
|
418 |
+
|
419 |
+
def __init__(self, aux_model, aux_input_ids, mask_token, strength=1.5, modulated_by_prob=True, tokenizer=None, use_attention_mask=True):
|
420 |
+
self.aux_model = aux_model # Same model, used for aux inputs
|
421 |
+
self.aux_input_ids = aux_input_ids
|
422 |
+
self.aux_past_key_values = None
|
423 |
+
self.strength = strength
|
424 |
+
self.modulated_by_prob = modulated_by_prob # Whether to modulate weight by probability
|
425 |
+
self.tokenizer = tokenizer # Optional, for debug printing
|
426 |
+
self.mask_token = mask_token # Store mask_token
|
427 |
+
# Store the device of the input_ids to use consistently
|
428 |
+
self.device = aux_input_ids.device
|
429 |
+
self.use_attention_mask = use_attention_mask
|
430 |
+
if self.use_attention_mask:
|
431 |
+
self.attention_mask = create_masked_attention(self.aux_input_ids, [mask_token], self.tokenizer)
|
432 |
+
else:
|
433 |
+
self.attention_mask = None
|
434 |
+
|
435 |
+
def __call__(self, input_ids, scores):
|
436 |
+
# Get aux model outputs for the current step
|
437 |
+
if self.aux_past_key_values is None:
|
438 |
+
# First step, run on full aux prompt
|
439 |
+
aux_outputs = self.aux_model(
|
440 |
+
input_ids=self.aux_input_ids,
|
441 |
+
use_cache=True,
|
442 |
+
return_dict=True,
|
443 |
+
attention_mask=self.attention_mask
|
444 |
+
)
|
445 |
+
self.aux_past_key_values = aux_outputs.past_key_values
|
446 |
+
aux_logits = aux_outputs.logits[:, -1, :]
|
447 |
+
else:
|
448 |
+
# Subsequent steps, run only on new token with past_key_values
|
449 |
+
last_token = input_ids[:, -1].unsqueeze(-1).to(self.device) # Ensure same device
|
450 |
+
# For subsequent tokens, we don't need to pass the attention mask
|
451 |
+
aux_outputs = self.aux_model(
|
452 |
+
input_ids=last_token,
|
453 |
+
past_key_values=self.aux_past_key_values,
|
454 |
+
use_cache=True,
|
455 |
+
return_dict=True
|
456 |
+
)
|
457 |
+
self.aux_past_key_values = aux_outputs.past_key_values
|
458 |
+
aux_logits = aux_outputs.logits[:, -1, :]
|
459 |
+
|
460 |
+
# Special case: strength = 1 means use only main logits
|
461 |
+
if abs(self.strength - 1.0) < 1e-4:
|
462 |
+
return scores
|
463 |
+
|
464 |
+
# if strength is 0, return the aux logits
|
465 |
+
if abs(self.strength - 0.0) < 1e-4:
|
466 |
+
return aux_logits
|
467 |
+
|
468 |
+
# Ensure scores and aux_logits are on the same device
|
469 |
+
if scores.device != aux_logits.device:
|
470 |
+
aux_logits = aux_logits.to(scores.device)
|
471 |
+
|
472 |
+
# Check for NaNs in the inputs
|
473 |
+
if torch.isnan(scores).any() or torch.isnan(aux_logits).any():
|
474 |
+
print("Warning: NaN values detected in input scores or aux_logits")
|
475 |
+
scores = torch.nan_to_num(scores, nan=0.0)
|
476 |
+
aux_logits = torch.nan_to_num(aux_logits, nan=0.0)
|
477 |
+
|
478 |
+
# Calculate the difference between main and aux logits
|
479 |
+
diff = scores - aux_logits
|
480 |
+
|
481 |
+
# Calculate the base weight
|
482 |
+
base_weight = self.strength - 1.0
|
483 |
+
|
484 |
+
# Modulate the weight by probability if enabled
|
485 |
+
# Only do this when strength > 1 (that's what can cause random behavior. If -1 < strength < 1, it is semantic dimishment, disable this for more precise control)
|
486 |
+
if self.modulated_by_prob and (self.strength > 1 or self.strength < -1):
|
487 |
+
# Convert logits to probabilities with temperature scaling for stability
|
488 |
+
temperature = 1.0
|
489 |
+
scaled_logits = scores / temperature
|
490 |
+
main_probs = F.softmax(scaled_logits, dim=-1)
|
491 |
+
|
492 |
+
# Clamp probabilities to avoid numerical issues
|
493 |
+
main_probs = torch.clamp(main_probs, min=1e-6, max=1.0)
|
494 |
+
|
495 |
+
# Each token's weight is scaled by its probability
|
496 |
+
|
497 |
+
# get the max probability
|
498 |
+
max_prob = torch.max(main_probs)
|
499 |
+
# normalize the base weight by the max probability
|
500 |
+
base_weight = base_weight / max_prob
|
501 |
+
# get different weights for each token based on their main probability
|
502 |
+
token_weights = base_weight * main_probs
|
503 |
+
|
504 |
+
# Apply the weighted adjustment
|
505 |
+
adjustment = token_weights * diff
|
506 |
+
|
507 |
+
# Clamp the adjustment to avoid extreme values
|
508 |
+
adjustment = torch.clamp(adjustment, min=-1e2, max=1e2)
|
509 |
+
|
510 |
+
# Compute final scores
|
511 |
+
final_scores = scores + adjustment
|
512 |
+
else:
|
513 |
+
# Safe computation of weighted difference
|
514 |
+
weighted_diff = base_weight * diff
|
515 |
+
# Check for and handle any NaNs that might have appeared
|
516 |
+
weighted_diff = torch.nan_to_num(weighted_diff, nan=0.0)
|
517 |
+
# Clamp to avoid extreme values
|
518 |
+
weighted_diff = torch.clamp(weighted_diff, min=-1e3, max=1e3)
|
519 |
+
final_scores = scores + weighted_diff
|
520 |
+
|
521 |
+
|
522 |
+
# Final stability check
|
523 |
+
final_scores = torch.clamp(final_scores, min=-1e3, max=1e3)
|
524 |
+
|
525 |
+
return final_scores
|
526 |
+
|
527 |
+
|
528 |
+
|
529 |
+
|
530 |
+
|