Exclude sensative layers from quantization on wan models
This commit is contained in:
1 parent
883d60eb71
commit
e7951ad29e
3 files changed
+14
No files matched your search
@@ -88,6 +88,9 @@ class Wan22Pipeline(WanPipeline):
|
||||
text_encoder_device = self.text_encoder.device
|
||||
device = self._exec_device
|
||||
|
||||
if vae_device == torch.device("cpu"):
|
||||
vae_device = self.reset_device_map
|
||||
|
||||
if self._aggressive_offload:
|
||||
print("Unloading vae")
|
||||
self.vae.to("cpu")
|
||||
|
||||
Reference in new issue
Block a user