feat: Support for finetuning Hunyuan Video-1.5 model #80
Conversation
Signed-off-by: Pranav Prashant Thombre <[email protected]>
Signed-off-by: Pranav Prashant Thombre <[email protected]>
Signed-off-by: Pranav Prashant Thombre <[email protected]>
Signed-off-by: Pranav Prashant Thombre <[email protected]>
Signed-off-by: Pranav Prashant Thombre <[email protected]>
Signed-off-by: Pranav Prashant Thombre <[email protected]>
Signed-off-by: Pablo Garay <[email protected]>
Signed-off-by: Pablo Garay <[email protected]>
pablo-garay
force-pushed
the
pranav/hunyuan_support
branch
from
56f6f0f to
aa475ee
Compare
|
|
||
|
|
||
| @dataclass | ||
| class FlowMatchingContext: |
There was a problem hiding this comment.
is this context limited to just flow matching? if i'm not missing something EDM pipeline requires the same set of attributes.
There was a problem hiding this comment.
Not specifically. But since I've implemented only flow matching thought to call it this.
|
|
||
| return sigma, timesteps, sampling_method | ||
|
|
||
| def _sample_from_distribution(self, batch_size: int, device: torch.device) -> torch.Tensor: |
There was a problem hiding this comment.
shall we separate this out of flow matching pipeline? we can have time step scheduler classes.
There was a problem hiding this comment.
I felt like the sampling process is an intrinsic part of the flow matching logic so included it here. But I would also be okay to move it out if the goal is to develop a more scheduler class.
|
/ok to test aa475ee |
huvunvidia
left a comment
huvunvidia
left a comment
There was a problem hiding this comment.
Left a comment. Thanks for the PR.
|
|
||
|
|
||
| @dataclass | ||
| class FlowMatchingOutput: |
There was a problem hiding this comment.
This class doesn't seem to be used anywhere?
Signed-off-by: Pranav Prashant Thombre <[email protected]>
Signed-off-by: Pranav Prashant Thombre <[email protected]>
|
/ok to test 33ea6fc |
No description provided.