Skip to content

Cast the Wan VAE weights to vae_weights_dtype instead of weights_dtype #517

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
boshjerns wants to merge 1 commit into
base: main
Choose a base branch
from
Open

Cast the Wan VAE weights to vae_weights_dtype instead of weights_dtype #517

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion src/maxdiffusion/pipelines/wan/wan_pipeline.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -631,7 +631,7 @@ def create_model(rngs: nnx.Rngs, config: HyperParameters):
# This helps with loading sharded weights directly into the accelerators without fist copying them
# all to one device and then distributing them, thus using low HBM memory.
params = load_wan_vae(config.pretrained_model_name_or_path, params, "cpu")
params = jax.tree_util.tree_map(lambda x: x.astype(config.weights_dtype), params)
params = jax.tree_util.tree_map(lambda x: x.astype(config.vae_weights_dtype), params)
for path, val in flax.traverse_util.flatten_dict(params).items():
sharding = logical_state_sharding[path].value
if config.replicate_vae:
Expand Down
21 changes: 21 additions & 0 deletions src/maxdiffusion/tests/wan/wan_vae_test.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -46,6 +46,7 @@
AutoencoderKLWanCache,
)
from maxdiffusion.models.wan.wan_utils import load_wan_vae
from maxdiffusion.pipelines.wan.wan_pipeline import WanPipeline
from maxdiffusion.utils import load_video
from maxdiffusion.video_processor import VideoProcessor
import flax
Expand Down Expand Up @@ -578,6 +579,26 @@ def vae_encode(video, wan_vae, vae_cache, key):
ssim_compare = ssim(video[0], original_video[0], multichannel=True, channel_axis=-1, data_range=255)
assert ssim_compare >= 0.9999

def test_load_vae_uses_vae_weights_dtype(self):
for weights_dtype, vae_weights_dtype in (("bfloat16", "float32"), ("float32", "bfloat16")):
with self.subTest(weights_dtype=weights_dtype, vae_weights_dtype=vae_weights_dtype):
pyconfig.initialize(
[
None,
os.path.join(THIS_DIR, "..", "..", "configs", "base_wan_14b.yml"),
f"weights_dtype={weights_dtype}",
f"vae_weights_dtype={vae_weights_dtype}",
],
unittest=True,
)
config = pyconfig.config
components = WanPipeline._create_common_components( # pylint: disable=protected-access
config, load_vae=True, load_text_encoder=False, load_scheduler=False
)
params = nnx.state(components["vae"], nnx.Param)
param_dtypes = {x.dtype for x in jax.tree_util.tree_leaves(params)}
self.assertEqual(param_dtypes, {jnp.dtype(vae_weights_dtype)})


if __name__ == "__main__":
absltest.main()
Loading