fix: update model configuration parameters in test files for consistency

Author: jp1924

test/convergence/bf16/test_mini_models.py

@@ -552,10 +552,10 @@
             vision_config=CLIPVisionConfig(
                 hidden_size=1024,
                 image_size=336,
-                intermediate_size=2048,  #4096
+                intermediate_size=2048,  # 4096
                 model_type="clip_vision_model",
-                num_attention_heads=4, #16
-                num_hidden_layers=4,  #24
+                num_attention_heads=4,  # 16
+                num_hidden_layers=4,  # 24
                 patch_size=14,
                 projection_dim=768,
                 vocab_size=32000,

test/convergence/bf16/test_mini_models_multimodal.py

@@ -357,10 +357,10 @@
             vision_config=CLIPVisionConfig(
                 hidden_size=1024,
                 image_size=336,
-                intermediate_size=4096,
+                intermediate_size=2048,  # 4096
                 model_type="clip_vision_model",
-                num_attention_heads=16,
-                num_hidden_layers=24,
+                num_attention_heads=4,  # 16
+                num_hidden_layers=4,  # 24
                 patch_size=14,
                 projection_dim=768,
                 vocab_size=32000,

test/convergence/bf16/test_mini_models_with_logits.py

@@ -553,10 +553,10 @@
             vision_config=CLIPVisionConfig(
                 hidden_size=1024,
                 image_size=336,
-                intermediate_size=4096,
+                intermediate_size=2048,  # 4096
                 model_type="clip_vision_model",
-                num_attention_heads=16,
-                num_hidden_layers=24,
+                num_attention_heads=4,  # 16
+                num_hidden_layers=4,  # 24
                 patch_size=14,
                 projection_dim=768,
                 vocab_size=32000,

test/convergence/fp32/test_mini_models.py

@@ -580,10 +580,10 @@
             vision_config=CLIPVisionConfig(
                 hidden_size=1024,
                 image_size=336,
-                intermediate_size=4096,
+                intermediate_size=2048,  # 4096
                 model_type="clip_vision_model",
-                num_attention_heads=16,
-                num_hidden_layers=24,
+                num_attention_heads=4,  # 16
+                num_hidden_layers=4,  # 24
                 patch_size=14,
                 projection_dim=768,
                 vocab_size=32000,

test/convergence/fp32/test_mini_models_multimodal.py

@@ -357,10 +357,10 @@
             vision_config=CLIPVisionConfig(
                 hidden_size=1024,
                 image_size=336,
-                intermediate_size=4096,
+                intermediate_size=2048,  # 4096
                 model_type="clip_vision_model",
-                num_attention_heads=16,
-                num_hidden_layers=24,
+                num_attention_heads=4,  # 16
+                num_hidden_layers=4,  # 24
                 patch_size=14,
                 projection_dim=768,
                 vocab_size=32000,

test/convergence/fp32/test_mini_models_with_logits.py

@@ -553,10 +553,10 @@
             vision_config=CLIPVisionConfig(
                 hidden_size=1024,
                 image_size=336,
-                intermediate_size=4096,
+                intermediate_size=2048,  # 4096
                 model_type="clip_vision_model",
-                num_attention_heads=16,
-                num_hidden_layers=24,
+                num_attention_heads=4,  # 16
+                num_hidden_layers=4,  # 24
                 patch_size=14,
                 projection_dim=768,
                 vocab_size=32000,