Added support for multiple input images
This commit is contained in:
@@ -541,6 +541,6 @@ class CLIPModel:
|
||||
videos = self.transforms.transforms[-1](videos.mul_(0.5).add_(0.5))
|
||||
|
||||
# forward
|
||||
with torch.cuda.amp.autocast(dtype=self.dtype):
|
||||
out = self.model.visual(videos, use_31_block=True)
|
||||
with torch.amp.autocast(dtype=self.dtype, device_type="cuda"):
|
||||
out = self.model.visual(videos.to(torch.bfloat16), use_31_block=True)
|
||||
return out
|
||||
|
||||
Reference in New Issue
Block a user