Spaces:
Paused
Paused
Update text2vid_torch2.py
Browse files- text2vid_torch2.py +54 -32
text2vid_torch2.py
CHANGED
@@ -224,8 +224,9 @@ class AttnProcessor2_0:
|
|
224 |
return query, key, dynamic_lambda, key1
|
225 |
'''
|
226 |
|
227 |
-
|
228 |
-
|
|
|
229 |
r"""
|
230 |
Compute the attention scores.
|
231 |
Args:
|
@@ -240,45 +241,66 @@ class AttnProcessor2_0:
|
|
240 |
dynamic_lambda = None
|
241 |
key1 = None
|
242 |
|
243 |
-
|
244 |
-
if self.
|
245 |
-
|
246 |
-
|
247 |
-
|
248 |
-
|
249 |
-
|
250 |
-
|
251 |
-
|
252 |
-
dynamic_lambda = torch.tensor([1 + self.LAMBDA * (i/50) for i in range(self.num_frames)]).to(key.dtype).cuda()
|
253 |
|
254 |
-
|
255 |
-
|
256 |
-
|
257 |
-
|
258 |
-
|
259 |
-
|
260 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
261 |
|
262 |
-
|
263 |
-
query_slice_shape = query[i*all_dim:(i*all_dim) + target_size, :query_list.shape[1], :query_list.shape[2]].shape
|
264 |
-
query_list_slice_shape = query_list[i*batch_dim:i*batch_dim + target_size].shape
|
265 |
|
266 |
-
|
267 |
-
|
268 |
-
|
269 |
-
|
|
|
|
|
|
|
|
|
270 |
|
271 |
-
|
272 |
-
|
273 |
-
|
274 |
|
275 |
-
|
276 |
-
|
277 |
-
|
|
|
|
|
|
|
278 |
|
|
|
|
|
|
|
|
|
|
|
|
|
279 |
return query, key, dynamic_lambda, key1
|
280 |
|
281 |
|
|
|
|
|
282 |
def init_attention_func(unet):
|
283 |
|
284 |
for name, module in unet.named_modules():
|
|
|
224 |
return query, key, dynamic_lambda, key1
|
225 |
'''
|
226 |
|
227 |
+
import torch
|
228 |
+
|
229 |
+
def get_qk(self, query, key):
|
230 |
r"""
|
231 |
Compute the attention scores.
|
232 |
Args:
|
|
|
241 |
dynamic_lambda = None
|
242 |
key1 = None
|
243 |
|
244 |
+
try:
|
245 |
+
if self.use_last_attn_slice:
|
246 |
+
if self.last_attn_slice is not None:
|
247 |
+
|
248 |
+
query_list = self.last_attn_slice[0]
|
249 |
+
key_list = self.last_attn_slice[1]
|
250 |
+
|
251 |
+
if query.shape[1] == self.num_frames and query.shape == key.shape:
|
252 |
+
key1 = key.clone()
|
|
|
253 |
|
254 |
+
# Ensure the batch dimension of key1 and key_list match
|
255 |
+
batch_size_key1 = key1.shape[0]
|
256 |
+
batch_size_key_list = key_list.shape[0]
|
257 |
+
|
258 |
+
if batch_size_key1 != batch_size_key_list:
|
259 |
+
# Handle mismatch: either pad or slice to match sizes
|
260 |
+
if batch_size_key1 > batch_size_key_list:
|
261 |
+
# Pad key_list if key1 batch size is larger
|
262 |
+
padding = (0, 0, 0, batch_size_key1 - batch_size_key_list) # (left, right, top, bottom)
|
263 |
+
key_list = torch.nn.functional.pad(key_list, padding, "constant", 0)
|
264 |
+
else:
|
265 |
+
# Slice key1 if key_list batch size is larger
|
266 |
+
key1 = key1[:batch_size_key_list]
|
267 |
+
|
268 |
+
# Proceed with assignment after matching batch dimensions
|
269 |
+
key1[:,:1,:key_list.shape[2]] = key_list[:,:1]
|
270 |
|
271 |
+
dynamic_lambda = torch.tensor([1 + self.LAMBDA * (i/50) for i in range(self.num_frames)]).to(key.dtype).cuda()
|
|
|
|
|
272 |
|
273 |
+
if q_old.shape == k_old.shape and q_old.shape[1] != self.num_frames:
|
274 |
+
batch_dim = query_list.shape[0] // self.bs
|
275 |
+
all_dim = query.shape[0] // self.bs
|
276 |
+
for i in range(self.bs):
|
277 |
+
target_size = min(query[i*all_dim:(i*all_dim) + batch_dim, :query_list.shape[1], :query_list.shape[2]].size(0),
|
278 |
+
query_list[i*batch_dim:(i+1)*batch_dim].size(0))
|
279 |
+
query_slice_shape = query[i*all_dim:(i*all_dim) + target_size, :query_list.shape[1], :query_list.shape[2]].shape
|
280 |
+
query_list_slice_shape = query_list[i*batch_dim:i*batch_dim + target_size].shape
|
281 |
|
282 |
+
if query_slice_shape[1] != query_list_slice_shape[1]:
|
283 |
+
print(f"Warning: Dimension mismatch. query_slice_shape: {query_slice_shape}, query_list_slice_shape: {query_list_slice_shape}. Adjusting to compatible sizes.")
|
284 |
+
target_size = min(query_slice_shape[1], query_list_slice_shape[1])
|
285 |
|
286 |
+
query[i*all_dim:(i*all_dim) + target_size, :query_list.shape[1], :query_list.shape[2]] = \
|
287 |
+
query_list[i*batch_dim:i*batch_dim + target_size]
|
288 |
+
|
289 |
+
if self.save_last_attn_slice:
|
290 |
+
self.last_attn_slice = [query, key]
|
291 |
+
self.save_last_attn_slice = False
|
292 |
|
293 |
+
except RuntimeError as e:
|
294 |
+
# If a RuntimeError happens, catch it and clean CUDA memory
|
295 |
+
print(f"RuntimeError occurred: {e}. Cleaning up CUDA memory...")
|
296 |
+
torch.cuda.empty_cache()
|
297 |
+
raise # Re-raise the error to let the caller handle it further if needed
|
298 |
+
|
299 |
return query, key, dynamic_lambda, key1
|
300 |
|
301 |
|
302 |
+
|
303 |
+
|
304 |
def init_attention_func(unet):
|
305 |
|
306 |
for name, module in unet.named_modules():
|