RFInversionFluxPipeline, small fix for enable_model_cpu_offload & enable_sequential_cpu_offload compatibility (huggingface#10480)

Teriks · linoytsaban · web-flow · commit 03bcf5aefef1 · 2025-01-07T15:47:28.000+01:00
RFInversionFluxPipeline.encode_image, device fix

Use self._execution_device instead of self.device when selecting
a device for the input image tensor.

This allows for compatibility with enable_model_cpu_offload &amp;
enable_sequential_cpu_offload

Co-authored-by: Teriks &lt;Teriks@users.noreply.github.com&gt;
Co-authored-by: Linoy Tsaban &lt;57615435+linoytsaban@users.noreply.github.com&gt;
diff --git a/examples/community/pipeline_flux_rf_inversion.py b/examples/community/pipeline_flux_rf_inversion.py
@@ -419,7 +419,7 @@ def encode_image(self, image, dtype=None, height=None, width=None, resize_mode="
             )
         image = image.to(dtype)
 
-        x0 = self.vae.encode(image.to(self.device)).latent_dist.sample()
+        x0 = self.vae.encode(image.to(self._execution_device)).latent_dist.sample()
         x0 = (x0 - self.vae.config.shift_factor) * self.vae.config.scaling_factor
         x0 = x0.to(dtype)
         return x0, resized

Original file line number	Diff line number	Diff line change
`@@ -419,7 +419,7 @@ def encode_image(self, image, dtype=None, height=None, width=None, resize_mode="`
`419`	`419`	`)`
`420`	`420`	`image = image.to(dtype)`
`421`	`421`
`422`		`- x0 = self.vae.encode(image.to(self.device)).latent_dist.sample()`
	`422`	`+ x0 = self.vae.encode(image.to(self._execution_device)).latent_dist.sample()`
`423`	`423`	`x0 = (x0 - self.vae.config.shift_factor) * self.vae.config.scaling_factor`
`424`	`424`	`x0 = x0.to(dtype)`
`425`	`425`	`return x0, resized`