[to #42322933]兼容新增clip huge模型

Link: https://code.alibaba-inc.com/Ali-MaaS/MaaS-lib/codereview/10585552

    * compatiable with vit huge, and set clip base default mm-ebed pipeline

modelscope/models/multi_modal/clip/model.py

@@ -349,11 +349,13 @@ class CLIP(nn.Module):
         text_num_hidden_layers: int,
         text_type_vocab_size: int,
         tokenizer: FullTokenizer,
+        # vision_head_width, added this param for ViT-H
+        vision_head_width: int = 64,
     ):
         super().__init__()
 
         if isinstance(vision_layers, (tuple, list)):
-            vision_heads = vision_width * 32 // 64
+            vision_heads = vision_width * 32 // vision_head_width
             self.visual = ModifiedResNet(
                 layers=vision_layers,
                 output_dim=embed_dim,
@@ -361,7 +363,7 @@ class CLIP(nn.Module):
                 input_resolution=image_resolution,
                 width=vision_width)
         else:
-            vision_heads = vision_width // 64
+            vision_heads = vision_width // vision_head_width
             self.visual = VisualTransformer(
                 input_resolution=image_resolution,
                 patch_size=vision_patch_size,

modelscope/pipelines/builder.py

@@ -93,9 +93,8 @@ DEFAULT_MODEL_FOR_PIPELINE = {
                           'damo/cv_resnet50_live-category'),
     Tasks.video_category: (Pipelines.video_category,
                            'damo/cv_resnet50_video-category'),
-    Tasks.multi_modal_embedding:
-    (Pipelines.multi_modal_embedding,
-     'damo/multi-modal_clip-vit-large-patch14_zh'),
+    Tasks.multi_modal_embedding: (Pipelines.multi_modal_embedding,
+                                  'damo/multi-modal_clip-vit-base-patch16_zh'),
     Tasks.generative_multi_modal_embedding:
     (Pipelines.generative_multi_modal_embedding,
      'damo/multi-modal_gemm-vit-large-patch14_generative-multi-modal-embedding'