[tests] use eval() on t5 for deterministic results.

src/diffusers/models/transformers/transformer_wan_vace.py

@@ -177,7 +177,7 @@ class WanVACETransformer3DModel(
     _supports_gradient_checkpointing = True
     _skip_layerwise_casting_patterns = ["patch_embedding", "vace_patch_embedding", "condition_embedder", "norm"]
     _no_split_modules = ["WanTransformerBlock", "WanVACETransformerBlock"]
-    _keep_in_fp32_modules = ["time_embedder", "scale_shift_table", "norm1", "norm2", "norm3"]
+    _keep_in_fp32_modules = ["rope", "time_embedder", "scale_shift_table", "norm1", "norm2", "norm3"]
     _keys_to_ignore_on_load_unexpected = ["norm_added_q"]
     _repeated_blocks = ["WanTransformerBlock", "WanVACETransformerBlock"]
 

src/diffusers/pipelines/wan/pipeline_wan_video2video.py

@@ -419,6 +419,7 @@ def prepare_latents(
         )
 
         if latents is None:
+            video = video.to(device=device, dtype=self.vae.dtype)
             init_latents = [retrieve_latents(self.vae.encode(vid.unsqueeze(0)), sample_mode="argmax") for vid in video]
 
             init_latents = torch.cat(init_latents, dim=0).to(dtype)

tests/pipelines/chroma/test_pipeline_chroma.py

@@ -42,7 +42,9 @@ def get_dummy_components(self, num_layers: int = 1, num_single_layers: int = 1):
 
         torch.manual_seed(0)
         config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
-        text_encoder = T5EncoderModel(config)
+        # `eval()` because a directly constructed model stays in training mode, which leaves T5's
+        # dropout active and makes the pipeline outputs non-deterministic across calls.
+        text_encoder = T5EncoderModel(config).eval()
 
         tokenizer = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")
 

tests/pipelines/chroma/test_pipeline_chroma_img2img.py

@@ -43,7 +43,9 @@ def get_dummy_components(self, num_layers: int = 1, num_single_layers: int = 1):
 
         torch.manual_seed(0)
         config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
-        text_encoder = T5EncoderModel(config)
+        # `eval()` because a directly constructed model stays in training mode, which leaves T5's
+        # dropout active and makes the pipeline outputs non-deterministic across calls.
+        text_encoder = T5EncoderModel(config).eval()
 
         tokenizer = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")
 

tests/pipelines/ltx/test_ltx.py

@@ -89,7 +89,9 @@ def get_dummy_components(self, num_layers: int = 1):
         torch.manual_seed(0)
         scheduler = FlowMatchEulerDiscreteScheduler()
         config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
-        text_encoder = T5EncoderModel(config)
+        # `eval()` because a directly constructed model stays in training mode, which leaves T5's
+        # dropout active and makes the pipeline outputs non-deterministic across calls.
+        text_encoder = T5EncoderModel(config).eval()
         tokenizer = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")
 
         components = {

tests/pipelines/ltx/test_ltx_condition.py

@@ -93,7 +93,9 @@ def get_dummy_components(self):
         torch.manual_seed(0)
         scheduler = FlowMatchEulerDiscreteScheduler()
         config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
-        text_encoder = T5EncoderModel(config)
+        # `eval()` because a directly constructed model stays in training mode, which leaves T5's
+        # dropout active and makes the pipeline outputs non-deterministic across calls.
+        text_encoder = T5EncoderModel(config).eval()
         tokenizer = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")
 
         components = {

tests/pipelines/ltx/test_ltx_image2video.py

@@ -92,7 +92,9 @@ def get_dummy_components(self):
         torch.manual_seed(0)
         scheduler = FlowMatchEulerDiscreteScheduler()
         config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
-        text_encoder = T5EncoderModel(config)
+        # `eval()` because a directly constructed model stays in training mode, which leaves T5's
+        # dropout active and makes the pipeline outputs non-deterministic across calls.
+        text_encoder = T5EncoderModel(config).eval()
         tokenizer = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")
 
         components = {

tests/pipelines/wan/test_wan.py

@@ -48,7 +48,9 @@ def get_dummy_components(self):
         # TODO: impl FlowDPMSolverMultistepScheduler
         scheduler = FlowMatchEulerDiscreteScheduler(shift=7.0)
         config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
-        text_encoder = T5EncoderModel(config)
+        # `eval()` because a directly constructed model stays in training mode, which leaves T5's
+        # dropout active and makes the pipeline outputs non-deterministic across calls.
+        text_encoder = T5EncoderModel(config).eval()
         tokenizer = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")
 
         torch.manual_seed(0)

tests/pipelines/wan/test_wan_22.py

@@ -47,7 +47,9 @@ def get_dummy_components(self):
         torch.manual_seed(0)
         scheduler = UniPCMultistepScheduler(prediction_type="flow_prediction", use_flow_sigmas=True, flow_shift=3.0)
         config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
-        text_encoder = T5EncoderModel(config)
+        # `eval()` because a directly constructed model stays in training mode, which leaves T5's
+        # dropout active and makes the pipeline outputs non-deterministic across calls.
+        text_encoder = T5EncoderModel(config).eval()
         tokenizer = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")
 
         torch.manual_seed(0)
@@ -191,7 +193,9 @@ def get_dummy_components(self):
         torch.manual_seed(0)
         scheduler = UniPCMultistepScheduler(prediction_type="flow_prediction", use_flow_sigmas=True, flow_shift=3.0)
         config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
-        text_encoder = T5EncoderModel(config)
+        # `eval()` because a directly constructed model stays in training mode, which leaves T5's
+        # dropout active and makes the pipeline outputs non-deterministic across calls.
+        text_encoder = T5EncoderModel(config).eval()
         tokenizer = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")
 
         torch.manual_seed(0)

tests/pipelines/wan/test_wan_22_image_to_video.py

@@ -49,7 +49,9 @@ def get_dummy_components(self):
         torch.manual_seed(0)
         scheduler = UniPCMultistepScheduler(prediction_type="flow_prediction", use_flow_sigmas=True, flow_shift=3.0)
         config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
-        text_encoder = T5EncoderModel(config)
+        # `eval()` because a directly constructed model stays in training mode, which leaves T5's
+        # dropout active and makes the pipeline outputs non-deterministic across calls.
+        text_encoder = T5EncoderModel(config).eval()
         tokenizer = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")
 
         torch.manual_seed(0)
@@ -203,7 +205,9 @@ def get_dummy_components(self):
         torch.manual_seed(0)
         scheduler = UniPCMultistepScheduler(prediction_type="flow_prediction", use_flow_sigmas=True, flow_shift=3.0)
         config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
-        text_encoder = T5EncoderModel(config)
+        # `eval()` because a directly constructed model stays in training mode, which leaves T5's
+        # dropout active and makes the pipeline outputs non-deterministic across calls.
+        text_encoder = T5EncoderModel(config).eval()
         tokenizer = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")
 
         torch.manual_seed(0)

tests/pipelines/wan/test_wan_animate.py

@@ -60,7 +60,9 @@ def get_dummy_components(self):
         torch.manual_seed(0)
         scheduler = FlowMatchEulerDiscreteScheduler(shift=7.0)
         config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
-        text_encoder = T5EncoderModel(config)
+        # `eval()` because a directly constructed model stays in training mode, which leaves T5's
+        # dropout active and makes the pipeline outputs non-deterministic across calls.
+        text_encoder = T5EncoderModel(config).eval()
         tokenizer = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")
 
         torch.manual_seed(0)

tests/pipelines/wan/test_wan_image_to_video.py

@@ -56,7 +56,9 @@ def get_dummy_components(self):
         # TODO: impl FlowDPMSolverMultistepScheduler
         scheduler = FlowMatchEulerDiscreteScheduler(shift=7.0)
         config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
-        text_encoder = T5EncoderModel(config)
+        # `eval()` because a directly constructed model stays in training mode, which leaves T5's
+        # dropout active and makes the pipeline outputs non-deterministic across calls.
+        text_encoder = T5EncoderModel(config).eval()
         tokenizer = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")
 
         torch.manual_seed(0)
@@ -198,7 +200,9 @@ def get_dummy_components(self):
         # TODO: impl FlowDPMSolverMultistepScheduler
         scheduler = FlowMatchEulerDiscreteScheduler(shift=7.0)
         config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
-        text_encoder = T5EncoderModel(config)
+        # `eval()` because a directly constructed model stays in training mode, which leaves T5's
+        # dropout active and makes the pipeline outputs non-deterministic across calls.
+        text_encoder = T5EncoderModel(config).eval()
         tokenizer = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")
 
         torch.manual_seed(0)

tests/pipelines/wan/test_wan_vace.py

@@ -55,7 +55,9 @@ def get_dummy_components(self):
         torch.manual_seed(0)
         scheduler = FlowMatchEulerDiscreteScheduler(shift=7.0)
         config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
-        text_encoder = T5EncoderModel(config)
+        # `eval()` because a directly constructed model stays in training mode, which leaves T5's
+        # dropout active and makes the pipeline outputs non-deterministic across calls.
+        text_encoder = T5EncoderModel(config).eval()
         tokenizer = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")
 
         torch.manual_seed(0)

tests/pipelines/wan/test_wan_video_to_video.py

@@ -49,7 +49,9 @@ def get_dummy_components(self):
         torch.manual_seed(0)
         scheduler = UniPCMultistepScheduler(flow_shift=3.0)
         config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
-        text_encoder = T5EncoderModel(config)
+        # `eval()` because a directly constructed model stays in training mode, which leaves T5's
+        # dropout active and makes the pipeline outputs non-deterministic across calls.
+        text_encoder = T5EncoderModel(config).eval()
         tokenizer = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")
 
         torch.manual_seed(0)