Spaces:
Running
on
A100
Running
on
A100
daniel shalem
commited on
Commit
•
6a9d9a1
1
Parent(s):
91602f9
Feature: Add full bfloat16 support.
Browse files
xora/examples/image_to_video.py
CHANGED
@@ -142,6 +142,12 @@ def main():
|
|
142 |
help="Mixed precision in float32 and bfloat16",
|
143 |
)
|
144 |
|
|
|
|
|
|
|
|
|
|
|
|
|
145 |
# Prompts
|
146 |
parser.add_argument(
|
147 |
"--prompt",
|
@@ -176,6 +182,9 @@ def main():
|
|
176 |
"PixArt-alpha/PixArt-XL-2-1024-MS", subfolder="tokenizer"
|
177 |
)
|
178 |
|
|
|
|
|
|
|
179 |
# Use submodels for the pipeline
|
180 |
submodel_dict = {
|
181 |
"transformer": unet,
|
|
|
142 |
help="Mixed precision in float32 and bfloat16",
|
143 |
)
|
144 |
|
145 |
+
parser.add_argument(
|
146 |
+
"--bfloat16",
|
147 |
+
action="store_true",
|
148 |
+
help="Denoise in bfloat16",
|
149 |
+
)
|
150 |
+
|
151 |
# Prompts
|
152 |
parser.add_argument(
|
153 |
"--prompt",
|
|
|
182 |
"PixArt-alpha/PixArt-XL-2-1024-MS", subfolder="tokenizer"
|
183 |
)
|
184 |
|
185 |
+
if args.bfloat16 and unet.dtype != torch.bfloat16:
|
186 |
+
unet = unet.to(torch.bfloat16)
|
187 |
+
|
188 |
# Use submodels for the pipeline
|
189 |
submodel_dict = {
|
190 |
"transformer": unet,
|
xora/examples/text_to_video.py
CHANGED
@@ -49,6 +49,16 @@ def main():
|
|
49 |
required=True,
|
50 |
help="Path to the directory containing unet, vae, and scheduler subdirectories",
|
51 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
52 |
args = parser.parse_args()
|
53 |
|
54 |
# Paths for the separate mode directories
|
@@ -72,6 +82,9 @@ def main():
|
|
72 |
"PixArt-alpha/PixArt-XL-2-1024-MS", subfolder="tokenizer"
|
73 |
)
|
74 |
|
|
|
|
|
|
|
75 |
# Use submodels for the pipeline
|
76 |
submodel_dict = {
|
77 |
"transformer": unet, # using unet for transformer
|
@@ -115,6 +128,7 @@ def main():
|
|
115 |
**sample,
|
116 |
is_video=True,
|
117 |
vae_per_channel_normalize=True,
|
|
|
118 |
).images
|
119 |
|
120 |
print("Generated images (video frames).")
|
|
|
49 |
required=True,
|
50 |
help="Path to the directory containing unet, vae, and scheduler subdirectories",
|
51 |
)
|
52 |
+
parser.add_argument(
|
53 |
+
"--mixed_precision",
|
54 |
+
action="store_true",
|
55 |
+
help="Mixed precision in float32 and bfloat16",
|
56 |
+
)
|
57 |
+
parser.add_argument(
|
58 |
+
"--bfloat16",
|
59 |
+
action="store_true",
|
60 |
+
help="Denoise in bfloat16",
|
61 |
+
)
|
62 |
args = parser.parse_args()
|
63 |
|
64 |
# Paths for the separate mode directories
|
|
|
82 |
"PixArt-alpha/PixArt-XL-2-1024-MS", subfolder="tokenizer"
|
83 |
)
|
84 |
|
85 |
+
if args.bfloat16 and unet.dtype != torch.bfloat16:
|
86 |
+
unet = unet.to(torch.bfloat16)
|
87 |
+
|
88 |
# Use submodels for the pipeline
|
89 |
submodel_dict = {
|
90 |
"transformer": unet, # using unet for transformer
|
|
|
128 |
**sample,
|
129 |
is_video=True,
|
130 |
vae_per_channel_normalize=True,
|
131 |
+
mixed_precision=args.mixed_precision,
|
132 |
).images
|
133 |
|
134 |
print("Generated images (video frames).")
|
xora/models/transformers/transformer3d.py
CHANGED
@@ -253,7 +253,7 @@ class Transformer3DModel(ModelMixin, ConfigMixin):
|
|
253 |
return fractional_positions
|
254 |
|
255 |
def precompute_freqs_cis(self, indices_grid, spacing="exp"):
|
256 |
-
dtype =
|
257 |
dim = self.inner_dim
|
258 |
theta = self.positional_embedding_theta
|
259 |
|
@@ -305,7 +305,7 @@ class Transformer3DModel(ModelMixin, ConfigMixin):
|
|
305 |
sin_padding = torch.zeros_like(cos_freq[:, :, : dim % 6])
|
306 |
cos_freq = torch.cat([cos_padding, cos_freq], dim=-1)
|
307 |
sin_freq = torch.cat([sin_padding, sin_freq], dim=-1)
|
308 |
-
return cos_freq.to(dtype), sin_freq.to(dtype)
|
309 |
|
310 |
def forward(
|
311 |
self,
|
|
|
253 |
return fractional_positions
|
254 |
|
255 |
def precompute_freqs_cis(self, indices_grid, spacing="exp"):
|
256 |
+
dtype = torch.float32 # We need full precision in the freqs_cis computation.
|
257 |
dim = self.inner_dim
|
258 |
theta = self.positional_embedding_theta
|
259 |
|
|
|
305 |
sin_padding = torch.zeros_like(cos_freq[:, :, : dim % 6])
|
306 |
cos_freq = torch.cat([cos_padding, cos_freq], dim=-1)
|
307 |
sin_freq = torch.cat([sin_padding, sin_freq], dim=-1)
|
308 |
+
return cos_freq.to(self.dtype), sin_freq.to(self.dtype)
|
309 |
|
310 |
def forward(
|
311 |
self,
|