Remove CogVideoX mentions from single file docs; Test updates

docs/source/en/api/loaders/single_file.md

@@ -22,9 +22,6 @@ The [`~loaders.FromSingleFileMixin.from_single_file`] method allows you to load:
 
 ## Supported pipelines
 
-- [`CogVideoXPipeline`]
-- [`CogVideoXImageToVideoPipeline`]
-- [`CogVideoXVideoToVideoPipeline`]
 - [`StableDiffusionPipeline`]
 - [`StableDiffusionImg2ImgPipeline`]
 - [`StableDiffusionInpaintPipeline`]
@@ -52,7 +49,6 @@ The [`~loaders.FromSingleFileMixin.from_single_file`] method allows you to load:
 - [`UNet2DConditionModel`]
 - [`StableCascadeUNet`]
 - [`AutoencoderKL`]
-- [`AutoencoderKLCogVideoX`]
 - [`ControlNetModel`]
 - [`SD3Transformer2DModel`]
 - [`FluxTransformer2DModel`]

tests/pipelines/cogvideo/test_cogvideox.py

@@ -57,6 +57,7 @@ class CogVideoXPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
             "callback_on_step_end_tensor_inputs",
         ]
     )
+    test_xformers_attention = False
 
     def get_dummy_components(self):
         torch.manual_seed(0)
@@ -71,8 +72,8 @@ def get_dummy_components(self):
             time_embed_dim=2,
             text_embed_dim=32,  # Must match with tiny-random-t5
             num_layers=1,
-            sample_width=16,  # latent width: 2 -> final width: 16
-            sample_height=16,  # latent height: 2 -> final height: 16
+            sample_width=2,  # latent width: 2 -> final width: 16
+            sample_height=2,  # latent height: 2 -> final height: 16
             sample_frames=9,  # latent frames: (9 - 1) / 4 + 1 = 3 -> final frames: 9
             patch_size=2,
             temporal_compression_ratio=4,
@@ -254,6 +255,14 @@ def test_vae_tiling(self, expected_diff_max: float = 0.2):
         generator_device = "cpu"
         components = self.get_dummy_components()
 
+        # Unlike the ImageToVideo test, this is not needed here because positional embeds can be generated on-the-fly
+        # due to them being static
+        # components["transformer"] = CogVideoXTransformer3DModel.from_config(
+        #     components["transformer"].config,
+        #     sample_height=16,
+        #     sample_width=16,
+        # )
+
         pipe = self.pipeline_class(**components)
         pipe.to("cpu")
         pipe.set_progress_bar_config(disable=None)
@@ -280,10 +289,6 @@ def test_vae_tiling(self, expected_diff_max: float = 0.2):
             "VAE tiling should not affect the inference results",
         )
 
-    @unittest.skip("xformers attention processor does not exist for CogVideoX")
-    def test_xformers_attention_forwardGenerator_pass(self):
-        pass
-
     def test_fused_qkv_projections(self):
         device = "cpu"  # ensure determinism for the device-dependent torch.Generator
         components = self.get_dummy_components()

tests/pipelines/cogvideo/test_cogvideox_image2video.py

@@ -269,8 +269,9 @@ def test_vae_tiling(self, expected_diff_max: float = 0.3):
         generator_device = "cpu"
         components = self.get_dummy_components()
 
-        # The reason to modify it this way is because I2V Transformer limits the generation to resolutions.
-        # See the if-statement on "self.use_learned_positional_embeddings"
+        # The reason to modify it this way is because I2V Transformer limits the generation to resolutions used during initalization.
+        # This limitation comes from using learned positional embeddings which cannot be generated on-the-fly like sincos or RoPE embeddings.
+        # See the if-statement on "self.use_learned_positional_embeddings" in diffusers/models/embeddings.py
         components["transformer"] = CogVideoXTransformer3DModel.from_config(
             components["transformer"].config,
             sample_height=16,

tests/pipelines/cogvideo/test_cogvideox_video2video.py

@@ -51,6 +51,7 @@ class CogVideoXVideoToVideoPipelineFastTests(PipelineTesterMixin, unittest.TestC
             "callback_on_step_end_tensor_inputs",
         ]
     )
+    test_xformers_attention = False
 
     def get_dummy_components(self):
         torch.manual_seed(0)
@@ -65,8 +66,8 @@ def get_dummy_components(self):
             time_embed_dim=2,
             text_embed_dim=32,  # Must match with tiny-random-t5
             num_layers=1,
-            sample_width=16,  # latent width: 2 -> final width: 16
-            sample_height=16,  # latent height: 2 -> final height: 16
+            sample_width=2,  # latent width: 2 -> final width: 16
+            sample_height=2,  # latent height: 2 -> final height: 16
             sample_frames=9,  # latent frames: (9 - 1) / 4 + 1 = 3 -> final frames: 9
             patch_size=2,
             temporal_compression_ratio=4,
@@ -259,6 +260,14 @@ def test_vae_tiling(self, expected_diff_max: float = 0.2):
         generator_device = "cpu"
         components = self.get_dummy_components()
 
+        # Unlike the ImageToVideo test, this is not needed here because positional embeds can be generated on-the-fly
+        # due to them being static
+        # components["transformer"] = CogVideoXTransformer3DModel.from_config(
+        #     components["transformer"].config,
+        #     sample_height=16,
+        #     sample_width=16,
+        # )
+
         pipe = self.pipeline_class(**components)
         pipe.to("cpu")
         pipe.set_progress_bar_config(disable=None)
@@ -285,10 +294,6 @@ def test_vae_tiling(self, expected_diff_max: float = 0.2):
             "VAE tiling should not affect the inference results",
         )
 
-    @unittest.skip("xformers attention processor does not exist for CogVideoX")
-    def test_xformers_attention_forwardGenerator_pass(self):
-        pass
-
     def test_fused_qkv_projections(self):
         device = "cpu"  # ensure determinism for the device-dependent torch.Generator
         components = self.get_dummy_components()