changed channel parameters for UNET and VAE. Changed configs parameters of CLIPText (#5370)

* changed channel parameters for UNET and VAE. Decreased hidden layers size with increased attention heads and intermediate size

* changed the assertion check range

* clean up

---------

Co-authored-by: Dhruv Nair <dhruv.nair@gmail.com>

Author: arkajyotimitra

tests/pipelines/stable_diffusion/test_stable_diffusion.py

@@ -82,6 +82,7 @@ def _test_stable_diffusion_compile(in_queue, out_queue, timeout):
 
         assert image.shape == (1, 512, 512, 3)
         expected_slice = np.array([0.38019, 0.28647, 0.27321, 0.40377, 0.38290, 0.35446, 0.39218, 0.38165, 0.42239])
+
         assert np.abs(image_slice - expected_slice).max() < 5e-3
     except Exception:
         error = f"{traceback.format_exc()}"
@@ -103,14 +104,15 @@ class StableDiffusionPipelineFastTests(
     def get_dummy_components(self):
         torch.manual_seed(0)
         unet = UNet2DConditionModel(
-            block_out_channels=(32, 64),
-            layers_per_block=2,
+            block_out_channels=(4, 8),
+            layers_per_block=1,
             sample_size=32,
             in_channels=4,
             out_channels=4,
             down_block_types=("DownBlock2D", "CrossAttnDownBlock2D"),
             up_block_types=("CrossAttnUpBlock2D", "UpBlock2D"),
             cross_attention_dim=32,
+            norm_num_groups=2,
         )
         scheduler = DDIMScheduler(
             beta_start=0.00085,
@@ -121,22 +123,23 @@ def get_dummy_components(self):
         )
         torch.manual_seed(0)
         vae = AutoencoderKL(
-            block_out_channels=[32, 64],
+            block_out_channels=[4, 8],
             in_channels=3,
             out_channels=3,
             down_block_types=["DownEncoderBlock2D", "DownEncoderBlock2D"],
             up_block_types=["UpDecoderBlock2D", "UpDecoderBlock2D"],
             latent_channels=4,
+            norm_num_groups=2,
         )
         torch.manual_seed(0)
         text_encoder_config = CLIPTextConfig(
             bos_token_id=0,
             eos_token_id=2,
             hidden_size=32,
-            intermediate_size=37,
+            intermediate_size=64,
             layer_norm_eps=1e-05,
-            num_attention_heads=4,
-            num_hidden_layers=5,
+            num_attention_heads=8,
+            num_hidden_layers=3,
             pad_token_id=1,
             vocab_size=1000,
         )
@@ -183,7 +186,7 @@ def test_stable_diffusion_ddim(self):
         image_slice = image[0, -3:, -3:, -1]
 
         assert image.shape == (1, 64, 64, 3)
-        expected_slice = np.array([0.5756, 0.6118, 0.5005, 0.5041, 0.5471, 0.4726, 0.4976, 0.4865, 0.4864])
+        expected_slice = np.array([0.3203, 0.4555, 0.4711, 0.3505, 0.3973, 0.4650, 0.5137, 0.3392, 0.4045])
 
         assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2
 
@@ -317,7 +320,7 @@ def test_stable_diffusion_ddim_factor_8(self):
         image_slice = image[0, -3:, -3:, -1]
 
         assert image.shape == (1, 136, 136, 3)
-        expected_slice = np.array([0.5524, 0.5626, 0.6069, 0.4727, 0.386, 0.3995, 0.4613, 0.4328, 0.4269])
+        expected_slice = np.array([0.4346, 0.5621, 0.5016, 0.3926, 0.4533, 0.4134, 0.5625, 0.5632, 0.5265])
 
         assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2
 
@@ -335,7 +338,7 @@ def test_stable_diffusion_pndm(self):
         image_slice = image[0, -3:, -3:, -1]
 
         assert image.shape == (1, 64, 64, 3)
-        expected_slice = np.array([0.5122, 0.5712, 0.4825, 0.5053, 0.5646, 0.4769, 0.5179, 0.4894, 0.4994])
+        expected_slice = np.array([0.3411, 0.5032, 0.4704, 0.3135, 0.4323, 0.4740, 0.5150, 0.3498, 0.4022])
 
         assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2
 
@@ -375,7 +378,7 @@ def test_stable_diffusion_k_lms(self):
         image_slice = image[0, -3:, -3:, -1]
 
         assert image.shape == (1, 64, 64, 3)
-        expected_slice = np.array([0.4873, 0.5443, 0.4845, 0.5004, 0.5549, 0.4850, 0.5191, 0.4941, 0.5065])
+        expected_slice = np.array([0.3149, 0.5246, 0.4796, 0.3218, 0.4469, 0.4729, 0.5151, 0.3597, 0.3954])
 
         assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2
 
@@ -394,7 +397,7 @@ def test_stable_diffusion_k_euler_ancestral(self):
         image_slice = image[0, -3:, -3:, -1]
 
         assert image.shape == (1, 64, 64, 3)
-        expected_slice = np.array([0.4872, 0.5444, 0.4846, 0.5003, 0.5549, 0.4850, 0.5189, 0.4941, 0.5067])
+        expected_slice = np.array([0.3151, 0.5243, 0.4794, 0.3217, 0.4468, 0.4728, 0.5152, 0.3598, 0.3954])
 
         assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2
 
@@ -413,7 +416,7 @@ def test_stable_diffusion_k_euler(self):
         image_slice = image[0, -3:, -3:, -1]
 
         assert image.shape == (1, 64, 64, 3)
-        expected_slice = np.array([0.4873, 0.5443, 0.4845, 0.5004, 0.5549, 0.4850, 0.5191, 0.4941, 0.5065])
+        expected_slice = np.array([0.3149, 0.5246, 0.4796, 0.3218, 0.4469, 0.4729, 0.5151, 0.3597, 0.3954])
 
         assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2
 
@@ -485,7 +488,7 @@ def test_stable_diffusion_negative_prompt(self):
         image_slice = image[0, -3:, -3:, -1]
 
         assert image.shape == (1, 64, 64, 3)
-        expected_slice = np.array([0.5114, 0.5706, 0.4772, 0.5028, 0.5637, 0.4732, 0.5169, 0.4881, 0.4977])
+        expected_slice = np.array([0.3458, 0.5120, 0.4800, 0.3116, 0.4348, 0.4802, 0.5237, 0.3467, 0.3991])
 
         assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2
 
@@ -638,7 +641,7 @@ def test_stable_diffusion_1_1_pndm(self):
         image_slice = image[0, -3:, -3:, -1].flatten()
 
         assert image.shape == (1, 512, 512, 3)
-        expected_slice = np.array([0.43625, 0.43554, 0.36670, 0.40660, 0.39703, 0.38658, 0.43936, 0.43557, 0.40592])
+        expected_slice = np.array([0.3149, 0.5246, 0.4796, 0.3218, 0.4469, 0.4729, 0.5151, 0.3597, 0.3954])
         assert np.abs(image_slice - expected_slice).max() < 3e-3
 
     def test_stable_diffusion_v1_4_with_freeu(self):
@@ -665,7 +668,7 @@ def test_stable_diffusion_1_4_pndm(self):
         image_slice = image[0, -3:, -3:, -1].flatten()
 
         assert image.shape == (1, 512, 512, 3)
-        expected_slice = np.array([0.57400, 0.47841, 0.31625, 0.63583, 0.58306, 0.55056, 0.50825, 0.56306, 0.55748])
+        expected_slice = np.array([0.3458, 0.5120, 0.4800, 0.3116, 0.4348, 0.4802, 0.5237, 0.3467, 0.3991])
         assert np.abs(image_slice - expected_slice).max() < 3e-3
 
     def test_stable_diffusion_ddim(self):