Add SDPA fallback for Siglip2Navit attention (#5)
Browse files- add sdpa fallback (8d086dc4f92f19f736480202f96ab109e74add20)
- clean (ad6d85fb24806f7fa8f0f51c20dfbae4c18f3bb8)
Co-authored-by: Isotr0py <[email protected]>
- modeling_ovis2_5.py +61 -6
modeling_ovis2_5.py
CHANGED
@@ -4,8 +4,6 @@ from typing import Dict, List, Optional, Tuple, Union
|
|
4 |
import PIL.Image
|
5 |
import numpy as np
|
6 |
import torch
|
7 |
-
from flash_attn import flash_attn_varlen_func
|
8 |
-
from flash_attn.layers.rotary import apply_rotary_emb
|
9 |
from torch import Tensor, nn
|
10 |
from torch.nn import functional as F
|
11 |
from transformers import (
|
@@ -19,9 +17,16 @@ from transformers.activations import ACT2FN
|
|
19 |
from transformers.generation.utils import GenerateOutput
|
20 |
from transformers.modeling_outputs import BaseModelOutputWithNoAttention
|
21 |
from transformers.modeling_utils import PreTrainedModel
|
|
|
22 |
|
23 |
from .configuration_ovis2_5 import Siglip2NavitConfig, Ovis2_5_Config
|
24 |
|
|
|
|
|
|
|
|
|
|
|
|
|
25 |
IMAGE_PLACEHOLDER = "<image>"
|
26 |
IMAGE_PLACEHOLDER_ID = -200
|
27 |
VIDEO_PLACEHOLDER = "<video>"
|
@@ -30,6 +35,7 @@ VIDEO_PLACEHOLDER_ID = -201
|
|
30 |
VISUAL_ATOM_ID = -300
|
31 |
INDICATOR_IDS = [-301, -302, -303, -304]
|
32 |
|
|
|
33 |
# copied from qwen2.5-vl
|
34 |
class VisionRotaryEmbedding(nn.Module):
|
35 |
def __init__(self, dim: int, theta: float = 10000.0) -> None:
|
@@ -193,6 +199,28 @@ def apply_rotary_pos_emb_flashatt(
|
|
193 |
return q_embed, k_embed
|
194 |
|
195 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
196 |
class Siglip2Attention(nn.Module):
|
197 |
"""Multi-headed attention from 'Attention Is All You Need' paper"""
|
198 |
|
@@ -238,14 +266,41 @@ class Siglip2Attention(nn.Module):
|
|
238 |
|
239 |
if self.use_rope:
|
240 |
cos, sin = position_embeddings
|
241 |
-
|
|
|
|
|
|
|
242 |
queries = queries.squeeze(0)
|
243 |
keys = keys.squeeze(0)
|
244 |
|
245 |
max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max().item()
|
246 |
-
|
247 |
-
|
248 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
249 |
attn_output = self.out_proj(attn_output)
|
250 |
return attn_output
|
251 |
|
|
|
4 |
import PIL.Image
|
5 |
import numpy as np
|
6 |
import torch
|
|
|
|
|
7 |
from torch import Tensor, nn
|
8 |
from torch.nn import functional as F
|
9 |
from transformers import (
|
|
|
17 |
from transformers.generation.utils import GenerateOutput
|
18 |
from transformers.modeling_outputs import BaseModelOutputWithNoAttention
|
19 |
from transformers.modeling_utils import PreTrainedModel
|
20 |
+
from transformers.utils import is_flash_attn_2_available
|
21 |
|
22 |
from .configuration_ovis2_5 import Siglip2NavitConfig, Ovis2_5_Config
|
23 |
|
24 |
+
|
25 |
+
if is_flash_attn_2_available():
|
26 |
+
from flash_attn import flash_attn_varlen_func
|
27 |
+
from flash_attn.layers.rotary import apply_rotary_emb
|
28 |
+
|
29 |
+
|
30 |
IMAGE_PLACEHOLDER = "<image>"
|
31 |
IMAGE_PLACEHOLDER_ID = -200
|
32 |
VIDEO_PLACEHOLDER = "<video>"
|
|
|
35 |
VISUAL_ATOM_ID = -300
|
36 |
INDICATOR_IDS = [-301, -302, -303, -304]
|
37 |
|
38 |
+
|
39 |
# copied from qwen2.5-vl
|
40 |
class VisionRotaryEmbedding(nn.Module):
|
41 |
def __init__(self, dim: int, theta: float = 10000.0) -> None:
|
|
|
199 |
return q_embed, k_embed
|
200 |
|
201 |
|
202 |
+
# Copied from transformers.models.llama.modeling_llama.rotate_half
|
203 |
+
def rotate_half(x):
|
204 |
+
"""Rotates half the hidden dims of the input."""
|
205 |
+
x1 = x[..., : x.shape[-1] // 2]
|
206 |
+
x2 = x[..., x.shape[-1] // 2 :]
|
207 |
+
return torch.cat((-x2, x1), dim=-1)
|
208 |
+
|
209 |
+
|
210 |
+
def apply_rotary_pos_emb_vision(
|
211 |
+
q: torch.Tensor, k: torch.Tensor, cos: torch.Tensor, sin: torch.Tensor
|
212 |
+
) -> tuple[torch.Tensor, torch.Tensor]:
|
213 |
+
orig_q_dtype = q.dtype
|
214 |
+
orig_k_dtype = k.dtype
|
215 |
+
q, k = q.float(), k.float()
|
216 |
+
cos, sin = cos.unsqueeze(-2).float(), sin.unsqueeze(-2).float()
|
217 |
+
q_embed = (q * cos) + (rotate_half(q) * sin)
|
218 |
+
k_embed = (k * cos) + (rotate_half(k) * sin)
|
219 |
+
q_embed = q_embed.to(orig_q_dtype)
|
220 |
+
k_embed = k_embed.to(orig_k_dtype)
|
221 |
+
return q_embed, k_embed
|
222 |
+
|
223 |
+
|
224 |
class Siglip2Attention(nn.Module):
|
225 |
"""Multi-headed attention from 'Attention Is All You Need' paper"""
|
226 |
|
|
|
266 |
|
267 |
if self.use_rope:
|
268 |
cos, sin = position_embeddings
|
269 |
+
if is_flash_attn_2_available():
|
270 |
+
queries, keys = apply_rotary_pos_emb_flashatt(queries.unsqueeze(0), keys.unsqueeze(0), cos, sin)
|
271 |
+
else:
|
272 |
+
queries, keys = apply_rotary_pos_emb_vision(queries.unsqueeze(0), keys.unsqueeze(0), cos, sin)
|
273 |
queries = queries.squeeze(0)
|
274 |
keys = keys.squeeze(0)
|
275 |
|
276 |
max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max().item()
|
277 |
+
if is_flash_attn_2_available():
|
278 |
+
attn_output = flash_attn_varlen_func(queries, keys, values, cu_seqlens, cu_seqlens, max_seqlen, max_seqlen).reshape(
|
279 |
+
seq_length, -1
|
280 |
+
)
|
281 |
+
else:
|
282 |
+
batch_size = cu_seqlens.shape[0] - 1
|
283 |
+
outputs = []
|
284 |
+
cu = cu_seqlens.tolist()
|
285 |
+
for i in range(batch_size):
|
286 |
+
start_idx = cu[i]
|
287 |
+
end_idx = cu[i + 1]
|
288 |
+
# Each sequence is processed independently.
|
289 |
+
q_i = queries[start_idx:end_idx].unsqueeze(0)
|
290 |
+
k_i = keys[start_idx:end_idx].unsqueeze(0)
|
291 |
+
v_i = values[start_idx:end_idx].unsqueeze(0)
|
292 |
+
# (1, seq_len, num_heads, head_dim) ->
|
293 |
+
# (1, num_heads, seq_len, head_dim)
|
294 |
+
q_i, k_i, v_i = [x.transpose(1, 2) for x in (q_i, k_i, v_i)]
|
295 |
+
output_i = F.scaled_dot_product_attention(q_i,
|
296 |
+
k_i,
|
297 |
+
v_i,
|
298 |
+
dropout_p=0.0)
|
299 |
+
# (1, num_heads, seq_len, head_dim) -> (seq_len, embed_dim)
|
300 |
+
output_i = output_i.transpose(1, 2).reshape(-1, self.embed_dim)
|
301 |
+
outputs.append(output_i)
|
302 |
+
attn_output = torch.cat(outputs, dim=0)
|
303 |
+
|
304 |
attn_output = self.out_proj(attn_output)
|
305 |
return attn_output
|
306 |
|