feat: update nebullvm api

Update nebullvm api calls to new interface & fix dynamic shape for openvino

.pre-commit-config.yaml

@@ -1,6 +1,6 @@
 repos:
 - repo: https://github.com/ambv/black
-  rev: 20.8b1
+  rev: 22.10.0
   hooks:
   - id: black
     types: [python]

server/clip_server/model/clip_nebullvm.py

@@ -1,6 +1,7 @@
 import os
 from pathlib import Path
 
+import numpy as np
 import torch.cuda
 
 from .clip import _download, available_models
@@ -35,32 +36,47 @@ def optimize_models(
         self,
         **kwargs,
     ):
-        from nebullvm.api.frontend.onnx import optimize_onnx_model
-
-        save_dir = os.path.expanduser("~/.cache/clip/nebullvm")
-        Path(save_dir).mkdir(exist_ok=True)
-        visual_save_dir = os.path.join(save_dir, "visual")
-        Path(visual_save_dir).mkdir(exist_ok=True)
-        text_save_dir = os.path.join(save_dir, "text")
-        Path(text_save_dir).mkdir(exist_ok=True)
-        general_kwargs = {
-            "batch_size": 1,
-        }
+        from nebullvm.api.functions import optimize_model
+
+        general_kwargs = {}
         general_kwargs.update(kwargs)
-        self._visual_model = optimize_onnx_model(
+
+        dynamic_info = {
+            "inputs": [
+                {0: 'batch', 1: 'num_channels', 2: 'pixel_size', 3: 'pixel_size'}
+            ],
+            "outputs": [{0: 'batch'}],
+        }
+
+        self._visual_model = optimize_model(
             self._visual_path,
-            input_sizes=[(3, self.pixel_size, self.pixel_size)],
-            save_dir=visual_save_dir,
-            ignore_compilers=["tvm"],
+            input_data=[
+                (
+                    (
+                        np.random.randn(1, 3, self.pixel_size, self.pixel_size).astype(
+                            np.float32
+                        ),
+                    ),
+                    0,
+                )
+            ],
+            dynamic_info=dynamic_info,
             **general_kwargs,
         )
 
-        self._textual_model = optimize_onnx_model(
+        dynamic_info = {
+            "inputs": [
+                {0: 'batch', 1: 'num_tokens'},
+            ],
+            "outputs": [
+                {0: 'batch'},
+            ],
+        }
+
+        self._textual_model = optimize_model(
             self._textual_path,
-            input_sizes=[(77,)],
-            save_dir=text_save_dir,
-            input_types=["int"],
-            ignore_compilers=["tvm"],
+            input_data=[((np.random.randint(0, 100, (1, 77)),), 0)],
+            dynamic_info=dynamic_info,
             **general_kwargs,
         )
 

server/setup.py

@@ -56,7 +56,7 @@
             'onnx',
         ]
         + (['onnxruntime-gpu>=1.8.0'] if sys.platform != 'darwin' else []),
-        'nebullvm': ['nebullvm>=0.2.1'],
+        'nebullvm': ['nebullvm>=0.4.3'],
         'tensorrt': ['nvidia-tensorrt'],
     },
     classifiers=[