Upload attn_mask_utils.py with huggingface_hub
Browse files- attn_mask_utils.py +292 -0
attn_mask_utils.py
ADDED
@@ -0,0 +1,292 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import copy
|
3 |
+
|
4 |
+
def find_prefix_seq_length_by_pe(
|
5 |
+
pe: torch.Tensor
|
6 |
+
) -> torch.Tensor:
|
7 |
+
"""
|
8 |
+
Find the sequence length where position encoding drops (indicating prefix boundary).
|
9 |
+
Args:
|
10 |
+
pe: Position encoding tensor of shape [Batch size, Sequence length ]
|
11 |
+
Contains position indices for each token in the sequence.
|
12 |
+
Returns:
|
13 |
+
torch.Tensor: A tensor of shape [B] containing:
|
14 |
+
- The index where position encoding drops for each sequence
|
15 |
+
- -1 if no drop occurs in the sequence
|
16 |
+
"""
|
17 |
+
batch_size, seq_len = pe.shape
|
18 |
+
prev = pe[:, :-1]
|
19 |
+
curr = pe[:, 1:]
|
20 |
+
drop_mask = curr < prev # [batch_size, seq_len-1]
|
21 |
+
|
22 |
+
seq_len = torch.full((batch_size,), -1, dtype=torch.long)
|
23 |
+
|
24 |
+
for b in range(batch_size):
|
25 |
+
drop_pos = torch.nonzero(drop_mask[b], as_tuple=False)
|
26 |
+
if drop_pos.numel() > 0:
|
27 |
+
i = drop_pos[0].item() + 1 # Take first drop position (+1 because we compared shifted sequences)
|
28 |
+
seq_len[b] = i
|
29 |
+
|
30 |
+
return seq_len
|
31 |
+
|
32 |
+
|
33 |
+
|
34 |
+
def update_causal_mask_with_pad_non_visible_2d(
|
35 |
+
input_ids: torch.Tensor,
|
36 |
+
attn_mask_2d: torch.Tensor,
|
37 |
+
text_mask_token_id: int = 151666,
|
38 |
+
block_size: int = 4,
|
39 |
+
causal_attn: bool = False
|
40 |
+
) -> torch.Tensor:
|
41 |
+
"""
|
42 |
+
Updates a 2D attention mask for hole sequence through input_ids and text_mask_token_id
|
43 |
+
|
44 |
+
Args:
|
45 |
+
input_ids: Input token IDs (unused in current implementation)
|
46 |
+
attn_mask_2d: 2D attention mask matrix of shape [seq_len, seq_len] where:
|
47 |
+
- 0.0 indicates allowed attention
|
48 |
+
- -inf indicates masked attention
|
49 |
+
text_mask_token_id: ID representing masked tokens
|
50 |
+
block_size: Size of the diffusion window
|
51 |
+
causal_attn: If True, maintains strict causal masking throughout
|
52 |
+
|
53 |
+
Returns:
|
54 |
+
Modified attention mask with updated visibility patterns
|
55 |
+
"""
|
56 |
+
seq_len = input_ids.shape[0]
|
57 |
+
device = input_ids.device
|
58 |
+
|
59 |
+
# Identify masked tokens and their preceding positions
|
60 |
+
input_mask = input_ids.eq(text_mask_token_id)
|
61 |
+
input_before_mask = torch.zeros_like(input_mask)
|
62 |
+
input_before_mask[:-1] = input_mask[1:]
|
63 |
+
mask_cols = (input_mask | input_before_mask)
|
64 |
+
non_mask = ~mask_cols
|
65 |
+
|
66 |
+
rows = torch.arange(seq_len, device=device)[:, None] # (seq_len, 1)
|
67 |
+
cols = torch.arange(seq_len, device=device) # (seq_len,)
|
68 |
+
|
69 |
+
|
70 |
+
indices = torch.arange(seq_len, device=device)
|
71 |
+
prev_non_mask = (indices * non_mask).cummax(dim=0).values
|
72 |
+
|
73 |
+
max_value = torch.iinfo(indices.dtype).max
|
74 |
+
mask_indices = torch.where(non_mask, indices, torch.full_like(indices, max_value))
|
75 |
+
reversed_mask_indices = torch.flip(mask_indices, dims=[0])
|
76 |
+
reversed_cummin = reversed_mask_indices.cummin(dim=0).values
|
77 |
+
next_non_mask = torch.flip(reversed_cummin, dims=[0])
|
78 |
+
|
79 |
+
# ================= Part 1: Make positions after masks invisible =================
|
80 |
+
infra_mask = (
|
81 |
+
(cols > prev_non_mask) &
|
82 |
+
(rows >= next_non_mask[None, :]) &
|
83 |
+
mask_cols[None, :]
|
84 |
+
)
|
85 |
+
attn_mask_2d.masked_fill_(infra_mask, -float('inf'))
|
86 |
+
|
87 |
+
# ================= Part 2: Allow visibility to previous positions (if not causal) =================
|
88 |
+
if not causal_attn:
|
89 |
+
visible_mask = (
|
90 |
+
(rows > prev_non_mask[None, :]) &
|
91 |
+
(rows < cols) &
|
92 |
+
mask_cols[None, :]
|
93 |
+
)
|
94 |
+
attn_mask_2d.masked_fill_(visible_mask, 0.0)
|
95 |
+
|
96 |
+
return attn_mask_2d
|
97 |
+
|
98 |
+
|
99 |
+
def update_causal_mask_for_one_gen_window_2d(
|
100 |
+
input_ids: torch.Tensor,
|
101 |
+
attn_mask_2d: torch.Tensor,
|
102 |
+
block_size: int = 4,
|
103 |
+
use_cache: bool = True,
|
104 |
+
causal_attn: bool = False
|
105 |
+
) -> torch.Tensor:
|
106 |
+
"""
|
107 |
+
Updates a 2D attention mask for a diffusion window in transformer inference.
|
108 |
+
|
109 |
+
Args:
|
110 |
+
input_ids: Input token IDs (unused in current implementation)
|
111 |
+
attn_mask_2d: 2D attention mask matrix of shape [seq_len, seq_len] where:
|
112 |
+
- 0.0 indicates allowed attention
|
113 |
+
- -inf indicates masked attention
|
114 |
+
block_size: Size of the diffusion window
|
115 |
+
use_cache: Whether key-value cache is being used
|
116 |
+
causal_attn: If True, maintains strict causal masking throughout
|
117 |
+
|
118 |
+
Returns:
|
119 |
+
Modified attention mask with updated visibility patterns
|
120 |
+
"""
|
121 |
+
|
122 |
+
if not causal_attn:
|
123 |
+
# Make the diffusion window (last block_size tokens) fully visible to itself
|
124 |
+
# This allows bidirectional attention within the diffusion window
|
125 |
+
attn_mask_2d[-block_size:, -block_size:] = 0.0
|
126 |
+
if use_cache:
|
127 |
+
# Mask the last token from previous round to prevent recomputation and maintain generation consistency.
|
128 |
+
attn_mask_2d[-block_size:, -block_size-1] = -float('inf')
|
129 |
+
|
130 |
+
return attn_mask_2d
|
131 |
+
|
132 |
+
|
133 |
+
def create_block_diff_mask_by_pe_1d(
|
134 |
+
b: int,
|
135 |
+
h: int,
|
136 |
+
q_idx: torch.Tensor,
|
137 |
+
kv_idx: torch.Tensor,
|
138 |
+
block_size: int,
|
139 |
+
x0_len_list: torch.Tensor,
|
140 |
+
position_ids_list: torch.Tensor,
|
141 |
+
causal_attn: bool = False,
|
142 |
+
) -> torch.Tensor:
|
143 |
+
"""Computes attention mask for a single query-key position in Flex Attention.
|
144 |
+
|
145 |
+
Args:
|
146 |
+
b (int): Batch index (0 <= b < batch_size).
|
147 |
+
h (int): Head index (unused in current implementation, reserved for future multi-head support).
|
148 |
+
q_idx (torch.Tensor): Query position index (scalar or 0D tensor).
|
149 |
+
kv_idx (torch.Tensor): Key/Value position index (scalar or 0D tensor).
|
150 |
+
block_size (int): Size of processing blocks for non-`x0` tokens.
|
151 |
+
x0_len_list (torch.Tensor): Tensor of shape [batch_size] with `x0` segment lengths.
|
152 |
+
position_ids_list (torch.Tensor): Tensor of shape [batch_size, seq_len] with position IDs.
|
153 |
+
causal_attn (bool, optional): Enforces causal masking in mutual blocks if True. Defaults to False.
|
154 |
+
|
155 |
+
Returns:
|
156 |
+
torch.Tensor: Boolean indicating whether attention is allowed (True = allowed).
|
157 |
+
"""
|
158 |
+
x0_len = x0_len_list[b]
|
159 |
+
position_ids = position_ids_list[b]
|
160 |
+
|
161 |
+
x0_flag_q = (q_idx < x0_len)
|
162 |
+
x0_flag_kv = (kv_idx < x0_len)
|
163 |
+
|
164 |
+
# top - left causal
|
165 |
+
block_causal = (
|
166 |
+
x0_flag_q & \
|
167 |
+
x0_flag_kv & \
|
168 |
+
(q_idx >= kv_idx)
|
169 |
+
)
|
170 |
+
|
171 |
+
q_ith_block = (q_idx - x0_len) // block_size
|
172 |
+
kv_ith_block = (kv_idx - x0_len) // block_size
|
173 |
+
|
174 |
+
# bottom - right
|
175 |
+
block_mutual = (
|
176 |
+
(~x0_flag_q & ~x0_flag_kv) & \
|
177 |
+
(q_ith_block == kv_ith_block) & \
|
178 |
+
(q_idx >= kv_idx if causal_attn else 1)
|
179 |
+
)
|
180 |
+
|
181 |
+
# bottom - left
|
182 |
+
prefix_len = position_ids[x0_len + q_ith_block * block_size] # kv_idx's cosponding prefix
|
183 |
+
block_prefix = (
|
184 |
+
(~x0_flag_q & x0_flag_kv) & \
|
185 |
+
(kv_idx < prefix_len)
|
186 |
+
)
|
187 |
+
|
188 |
+
mask_val = (block_causal | block_mutual | block_prefix)
|
189 |
+
return mask_val.to(torch.bool)
|
190 |
+
|
191 |
+
|
192 |
+
def create_block_diff_mask_by_pe_4d(
|
193 |
+
block_size: int,
|
194 |
+
x0_len_list: torch.Tensor,
|
195 |
+
position_ids: torch.Tensor,
|
196 |
+
causal_attn: bool = False
|
197 |
+
) -> tuple[torch.Tensor, torch.Tensor]:
|
198 |
+
"""Generates a 4D attention mask for block-difference attention patterns.
|
199 |
+
|
200 |
+
The mask consists of three regions:
|
201 |
+
1. Causal block (top-left): Standard causal attention for `x0` tokens.
|
202 |
+
2. Mutual block (bottom-right): Non-causal attention within the same block for non-`x0` tokens.
|
203 |
+
3. Prefix block (bottom-left): Non-`x0` tokens can attend to a prefix of `x0` tokens.
|
204 |
+
|
205 |
+
Args:
|
206 |
+
block_size (int): Size of processing blocks for non-`x0` tokens.
|
207 |
+
x0_len_list (torch.Tensor): Tensor of shape [B] containing lengths of `x0` segments per batch.
|
208 |
+
position_ids (torch.Tensor): Tensor of shape [B, seq_len] containing position IDs.
|
209 |
+
causal_attn (bool, optional): If True, enforces causal masking in mutual blocks. Defaults to False.
|
210 |
+
|
211 |
+
Returns:
|
212 |
+
tuple[torch.Tensor, torch.Tensor]:
|
213 |
+
- A float mask of shape [batch_size, 1, seq_len, seq_len] with `-inf` for masked positions (non visiable).
|
214 |
+
- A boolean mask of shape [batch_size, 1, seq_len, seq_len] indicating allowed attention positions.
|
215 |
+
"""
|
216 |
+
batch_size, seq_len = position_ids.shape
|
217 |
+
device = position_ids.device
|
218 |
+
|
219 |
+
# Create position indices [batch_size, seq_len, seq_len]
|
220 |
+
q_idx = torch.arange(seq_len, device=device).view(1, seq_len, 1) # [1, seq_len, 1]
|
221 |
+
kv_idx = torch.arange(seq_len, device=device).view(1, 1, seq_len) # [1, 1, seq_len]
|
222 |
+
|
223 |
+
# Broadcast to [B, seq_len, seq_len]
|
224 |
+
x0_len = x0_len_list.view(batch_size, 1, 1) # [batch_size, 1, 1]
|
225 |
+
x0_flag_q = q_idx < x0_len # [batch_size, seq_len, seq_len]
|
226 |
+
x0_flag_kv = kv_idx < x0_len
|
227 |
+
|
228 |
+
# Block indices calculation [batch_size, seq_len, seq_len]
|
229 |
+
q_block_idx = (q_idx - x0_len) // block_size
|
230 |
+
kv_block_idx = (kv_idx - x0_len) // block_size
|
231 |
+
|
232 |
+
# causal block (top-left)
|
233 |
+
block_causal = x0_flag_q & x0_flag_kv & (q_idx >= kv_idx)
|
234 |
+
|
235 |
+
# Mutual block (bottom-right)
|
236 |
+
mutual_condition = (q_idx >= kv_idx) if causal_attn else torch.ones_like(q_idx, dtype=torch.bool)
|
237 |
+
block_mutual = (~x0_flag_q & ~x0_flag_kv &
|
238 |
+
(q_block_idx == kv_block_idx) &
|
239 |
+
mutual_condition)
|
240 |
+
|
241 |
+
# Prefix block (bottom-left)
|
242 |
+
q_blk = torch.div(q_idx - x0_len, block_size, rounding_mode='floor')
|
243 |
+
q_blk_start = (x0_len_list.view(batch_size, 1) + q_blk[:, :, 0] * block_size).clamp(min=0, max=seq_len-1) # (batch_size, L)
|
244 |
+
prefix_len = position_ids.gather(1, q_blk_start)
|
245 |
+
prefix_len = prefix_len.unsqueeze(2)
|
246 |
+
block_prefix = (~x0_flag_q & x0_flag_kv) & (kv_idx < prefix_len)
|
247 |
+
|
248 |
+
# FIXME Padding Mask
|
249 |
+
# padding_mask = (position_ids.view(batch_size, 1, seq_len) != -1) & (position_ids.view(batch_size, seq_len, -1) != -1)
|
250 |
+
|
251 |
+
# Combine masks
|
252 |
+
final_mask = (block_causal | block_mutual | block_prefix) # bool
|
253 |
+
# & padding_mask
|
254 |
+
customized_mask = torch.full_like(final_mask, float('-inf'), dtype=torch.bfloat16)
|
255 |
+
customized_mask.masked_fill_(final_mask, 0.0) # 0.0 or -inf
|
256 |
+
|
257 |
+
# Add head dimension [batch_size, 1, seq_len, seq_len]
|
258 |
+
return customized_mask.unsqueeze(1).to(device=device), final_mask.unsqueeze(1).to(device=device)
|
259 |
+
|
260 |
+
|
261 |
+
def find_pred_pos_from_input_ids(
|
262 |
+
input_ids: torch.LongTensor = None,
|
263 |
+
text_mask_token_id: int = 151666,
|
264 |
+
) -> torch.Tensor:
|
265 |
+
"""Compute the relative prediction positions for masked tokens in a sequence.
|
266 |
+
|
267 |
+
For non-masked positions, the output is 0. For masked positions, the value increments
|
268 |
+
by 1 for each consecutive mask token, indicating how many steps ahead the prediction is.
|
269 |
+
|
270 |
+
Args:
|
271 |
+
input_ids (torch.LongTensor): Input token IDs of shape [batch_size, seq_len].
|
272 |
+
text_mask_token_id (int, optional): Token ID representing masked positions. Defaults to 151666.
|
273 |
+
|
274 |
+
Returns:
|
275 |
+
torch.Tensor: A tensor of shape [batch_size, seq_len] where:
|
276 |
+
- 0 indicates a non-masked token.
|
277 |
+
- n > 0 indicates the nth consecutive masked token (e.g., 1 = first mask, 2 = second mask, etc.).
|
278 |
+
"""
|
279 |
+
batch_size, seq_len = input_ids.shape
|
280 |
+
device = input_ids.device
|
281 |
+
|
282 |
+
is_mask = (input_ids == text_mask_token_id)
|
283 |
+
|
284 |
+
base_mask = torch.zeros((batch_size, seq_len), dtype=torch.int8, device=device)
|
285 |
+
|
286 |
+
for b in range(batch_size):
|
287 |
+
for ix in range(1, seq_len):
|
288 |
+
if is_mask[b][ix] == True:
|
289 |
+
# Increment counter if current token is masked
|
290 |
+
base_mask[b][ix] = base_mask[b][ix-1] + 1
|
291 |
+
|
292 |
+
return base_mask
|