move VAE decoder to fp32 execution precision on GPU (#432)

huggingface · Sep 18, 2023 · 681b946 · 681b946
1 parent a7782ae
commit 681b946
Showing 1 changed file with 10 additions and 0 deletions.
diff --git a/optimum/intel/openvino/modeling_diffusion.py b/optimum/intel/openvino/modeling_diffusion.py
@@ -606,6 +606,11 @@ def __call__(self, latent_sample: np.ndarray):
         outputs = self.request(inputs, shared_memory=True)
         return list(outputs.values())
 
+    def _compile(self):
+        if "GPU" in self.device:
+            self.ov_config.update({"INFERENCE_PRECISION_HINT": "f32"})
+        super()._compile()
+
 
 class OVModelVaeEncoder(OVModelPart):
     def __init__(
@@ -622,6 +627,11 @@ def __call__(self, sample: np.ndarray):
         outputs = self.request(inputs, shared_memory=True)
         return list(outputs.values())
 
+    def _compile(self):
+        if "GPU" in self.device:
+            self.ov_config.update({"INFERENCE_PRECISION_HINT": "f32"})
+        super()._compile()
+
 
 class OVStableDiffusionPipeline(OVStableDiffusionPipelineBase, StableDiffusionPipelineMixin):
     def __call__(