Commit
·
1d3fed2
1
Parent(s):
3e7ee7c
Refactor attention module to improve xformers integration. Renamed availability flag to HAS_XFORMERS and added safe_memory_efficient_attention function for better handling of attention operations across devices. Updated related assertions and calls to ensure compatibility with systems lacking GPU support.
Browse files
imagedream/ldm/modules/attention.py
CHANGED
@@ -12,10 +12,9 @@ from .diffusionmodules.util import checkpoint
|
|
12 |
try:
|
13 |
import xformers
|
14 |
import xformers.ops
|
15 |
-
|
16 |
-
|
17 |
-
|
18 |
-
XFORMERS_IS_AVAILBLE = False
|
19 |
|
20 |
# CrossAttn precision handling
|
21 |
import os
|
@@ -138,6 +137,20 @@ class SpatialSelfAttention(nn.Module):
|
|
138 |
return x + h_
|
139 |
|
140 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
141 |
class MemoryEfficientCrossAttention(nn.Module):
|
142 |
# https://github.com/MatthieuTPHR/diffusers/blob/d80b531ff8060ec1ea982b65a1b8df70f73aa67c/src/diffusers/models/attention.py#L223
|
143 |
def __init__(self, query_dim, context_dim=None, heads=8, dim_head=64, dropout=0.0, **kwargs):
|
@@ -195,7 +208,7 @@ class MemoryEfficientCrossAttention(nn.Module):
|
|
195 |
)
|
196 |
|
197 |
# actually compute the attention, what we cannot get enough of
|
198 |
-
out =
|
199 |
q, k, v, attn_bias=None, op=self.attention_op
|
200 |
)
|
201 |
|
@@ -209,7 +222,7 @@ class MemoryEfficientCrossAttention(nn.Module):
|
|
209 |
(k_ip, v_ip),
|
210 |
)
|
211 |
# actually compute the attention, what we cannot get enough of
|
212 |
-
out_ip =
|
213 |
q, k_ip, v_ip, attn_bias=None, op=self.attention_op
|
214 |
)
|
215 |
out = out + self.ip_weight * out_ip
|
@@ -239,7 +252,7 @@ class BasicTransformerBlock(nn.Module):
|
|
239 |
**kwargs
|
240 |
):
|
241 |
super().__init__()
|
242 |
-
assert
|
243 |
attn_cls = MemoryEfficientCrossAttention
|
244 |
self.disable_self_attn = disable_self_attn
|
245 |
self.attn1 = attn_cls(
|
|
|
12 |
try:
|
13 |
import xformers
|
14 |
import xformers.ops
|
15 |
+
HAS_XFORMERS = True
|
16 |
+
except ImportError:
|
17 |
+
HAS_XFORMERS = False
|
|
|
18 |
|
19 |
# CrossAttn precision handling
|
20 |
import os
|
|
|
137 |
return x + h_
|
138 |
|
139 |
|
140 |
+
def safe_memory_efficient_attention(q, k, v, attn_bias=None, op=None, p=0.0):
|
141 |
+
if q.device.type == "cuda" and HAS_XFORMERS:
|
142 |
+
return xformers.ops.memory_efficient_attention(q, k, v, attn_bias=attn_bias, op=op, p=p)
|
143 |
+
else:
|
144 |
+
# Standard attention for CPU
|
145 |
+
scale = 1.0 / (q.shape[-1] ** 0.5)
|
146 |
+
attn = torch.matmul(q * scale, k.transpose(-2, -1))
|
147 |
+
if attn_bias is not None:
|
148 |
+
attn = attn + attn_bias
|
149 |
+
attn = torch.softmax(attn, dim=-1)
|
150 |
+
attn = torch.nn.functional.dropout(attn, p=p)
|
151 |
+
return torch.matmul(attn, v)
|
152 |
+
|
153 |
+
|
154 |
class MemoryEfficientCrossAttention(nn.Module):
|
155 |
# https://github.com/MatthieuTPHR/diffusers/blob/d80b531ff8060ec1ea982b65a1b8df70f73aa67c/src/diffusers/models/attention.py#L223
|
156 |
def __init__(self, query_dim, context_dim=None, heads=8, dim_head=64, dropout=0.0, **kwargs):
|
|
|
208 |
)
|
209 |
|
210 |
# actually compute the attention, what we cannot get enough of
|
211 |
+
out = safe_memory_efficient_attention(
|
212 |
q, k, v, attn_bias=None, op=self.attention_op
|
213 |
)
|
214 |
|
|
|
222 |
(k_ip, v_ip),
|
223 |
)
|
224 |
# actually compute the attention, what we cannot get enough of
|
225 |
+
out_ip = safe_memory_efficient_attention(
|
226 |
q, k_ip, v_ip, attn_bias=None, op=self.attention_op
|
227 |
)
|
228 |
out = out + self.ip_weight * out_ip
|
|
|
252 |
**kwargs
|
253 |
):
|
254 |
super().__init__()
|
255 |
+
assert HAS_XFORMERS, "xformers is not available"
|
256 |
attn_cls = MemoryEfficientCrossAttention
|
257 |
self.disable_self_attn = disable_self_attn
|
258 |
self.attn1 = attn_cls(
|
imagedream/ldm/modules/diffusionmodules/model.py
CHANGED
@@ -11,10 +11,9 @@ from ..attention import MemoryEfficientCrossAttention
|
|
11 |
try:
|
12 |
import xformers
|
13 |
import xformers.ops
|
14 |
-
|
15 |
-
|
16 |
-
|
17 |
-
XFORMERS_IS_AVAILBLE = False
|
18 |
print("No module 'xformers'. Proceeding without it.")
|
19 |
|
20 |
|
@@ -238,7 +237,7 @@ class MemoryEfficientAttnBlock(nn.Module):
|
|
238 |
.contiguous(),
|
239 |
(q, k, v),
|
240 |
)
|
241 |
-
out =
|
242 |
q, k, v, attn_bias=None, op=self.attention_op
|
243 |
)
|
244 |
|
@@ -262,6 +261,20 @@ class MemoryEfficientCrossAttentionWrapper(MemoryEfficientCrossAttention):
|
|
262 |
return x + out
|
263 |
|
264 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
265 |
def make_attn(in_channels, attn_type="vanilla", attn_kwargs=None):
|
266 |
assert attn_type in [
|
267 |
"vanilla",
|
@@ -270,7 +283,7 @@ def make_attn(in_channels, attn_type="vanilla", attn_kwargs=None):
|
|
270 |
"linear",
|
271 |
"none",
|
272 |
], f"attn_type {attn_type} unknown"
|
273 |
-
if
|
274 |
attn_type = "vanilla-xformers"
|
275 |
print(f"making attention of type '{attn_type}' with {in_channels} in_channels")
|
276 |
if attn_type == "vanilla":
|
|
|
11 |
try:
|
12 |
import xformers
|
13 |
import xformers.ops
|
14 |
+
HAS_XFORMERS = True
|
15 |
+
except ImportError:
|
16 |
+
HAS_XFORMERS = False
|
|
|
17 |
print("No module 'xformers'. Proceeding without it.")
|
18 |
|
19 |
|
|
|
237 |
.contiguous(),
|
238 |
(q, k, v),
|
239 |
)
|
240 |
+
out = safe_memory_efficient_attention(
|
241 |
q, k, v, attn_bias=None, op=self.attention_op
|
242 |
)
|
243 |
|
|
|
261 |
return x + out
|
262 |
|
263 |
|
264 |
+
def safe_memory_efficient_attention(q, k, v, attn_bias=None, op=None, p=0.0):
|
265 |
+
if q.device.type == "cuda" and HAS_XFORMERS:
|
266 |
+
return xformers.ops.memory_efficient_attention(q, k, v, attn_bias=attn_bias, op=op, p=p)
|
267 |
+
else:
|
268 |
+
# Standard attention for CPU
|
269 |
+
scale = 1.0 / (q.shape[-1] ** 0.5)
|
270 |
+
attn = torch.matmul(q * scale, k.transpose(-2, -1))
|
271 |
+
if attn_bias is not None:
|
272 |
+
attn = attn + attn_bias
|
273 |
+
attn = torch.softmax(attn, dim=-1)
|
274 |
+
attn = torch.nn.functional.dropout(attn, p=p)
|
275 |
+
return torch.matmul(attn, v)
|
276 |
+
|
277 |
+
|
278 |
def make_attn(in_channels, attn_type="vanilla", attn_kwargs=None):
|
279 |
assert attn_type in [
|
280 |
"vanilla",
|
|
|
283 |
"linear",
|
284 |
"none",
|
285 |
], f"attn_type {attn_type} unknown"
|
286 |
+
if HAS_XFORMERS and attn_type == "vanilla":
|
287 |
attn_type = "vanilla-xformers"
|
288 |
print(f"making attention of type '{attn_type}' with {in_channels} in_channels")
|
289 |
if attn_type == "vanilla":
|