Use Tensor.numpy() instead np.array(Tensor) (#1153)

Author: eaidova

optimum/intel/openvino/modeling.py

@@ -174,9 +174,9 @@ def forward(
 
         np_inputs = isinstance(input_ids, np.ndarray)
         if not np_inputs:
-            input_ids = np.array(input_ids)
-            attention_mask = np.array(attention_mask)
-            token_type_ids = np.array(token_type_ids) if token_type_ids is not None else token_type_ids
+            input_ids = input_ids.cpu().numpy()
+            attention_mask = attention_mask.cpu().numpy()
+            token_type_ids = token_type_ids.cpu().numpy() if token_type_ids is not None else token_type_ids
 
         inputs = {
             "input_ids": input_ids,
@@ -239,9 +239,9 @@ def forward(
 
         np_inputs = isinstance(input_ids, np.ndarray)
         if not np_inputs:
-            input_ids = np.array(input_ids)
-            attention_mask = np.array(attention_mask)
-            token_type_ids = np.array(token_type_ids) if token_type_ids is not None else token_type_ids
+            input_ids = input_ids.cpu().numpy()
+            attention_mask = attention_mask.cpu().numpy()
+            token_type_ids = token_type_ids.cpu().numpy() if token_type_ids is not None else token_type_ids
 
         inputs = {
             "input_ids": input_ids,
@@ -308,9 +308,9 @@ def forward(
 
         np_inputs = isinstance(input_ids, np.ndarray)
         if not np_inputs:
-            input_ids = np.array(input_ids)
-            attention_mask = np.array(attention_mask)
-            token_type_ids = np.array(token_type_ids) if token_type_ids is not None else token_type_ids
+            input_ids = input_ids.cpu().numpy()
+            attention_mask = attention_mask.cpu().numpy()
+            token_type_ids = token_type_ids.cpu().numpy() if token_type_ids is not None else token_type_ids
 
         inputs = {
             "input_ids": input_ids,
@@ -379,9 +379,9 @@ def forward(
 
         np_inputs = isinstance(input_ids, np.ndarray)
         if not np_inputs:
-            input_ids = np.array(input_ids)
-            attention_mask = np.array(attention_mask)
-            token_type_ids = np.array(token_type_ids) if token_type_ids is not None else token_type_ids
+            input_ids = input_ids.cpu().numpy()
+            attention_mask = attention_mask.cpu().numpy()
+            token_type_ids = token_type_ids.cpu().numpy() if token_type_ids is not None else token_type_ids
 
         inputs = {
             "input_ids": input_ids,
@@ -448,9 +448,9 @@ def forward(
 
         np_inputs = isinstance(input_ids, np.ndarray)
         if not np_inputs:
-            input_ids = np.array(input_ids)
-            attention_mask = np.array(attention_mask)
-            token_type_ids = np.array(token_type_ids) if token_type_ids is not None else token_type_ids
+            input_ids = input_ids.cpu().numpy()
+            attention_mask = attention_mask.cpu().numpy()
+            token_type_ids = token_type_ids.cpu().numpy() if token_type_ids is not None else token_type_ids
 
         inputs = {
             "input_ids": input_ids,
@@ -581,7 +581,7 @@ def forward(
 
         np_inputs = isinstance(pixel_values, np.ndarray)
         if not np_inputs:
-            pixel_values = np.array(pixel_values)
+            pixel_values = pixel_values.cpu().numpy()
 
         inputs = {
             "pixel_values": pixel_values,
@@ -640,8 +640,8 @@ def forward(
 
         np_inputs = isinstance(input_values, np.ndarray)
         if not np_inputs:
-            input_values = np.array(input_values)
-            attention_mask = np.array(attention_mask) if attention_mask is not None else attention_mask
+            input_values = input_values.cpu().numpy()
+            attention_mask = attention_mask.cpu().numpy() if attention_mask is not None else attention_mask
 
         inputs = {
             "input_values": input_values,
@@ -711,8 +711,8 @@ def forward(
     ):
         np_inputs = isinstance(input_values, np.ndarray)
         if not np_inputs:
-            input_values = np.array(input_values)
-            attention_mask = np.array(attention_mask) if attention_mask is not None else attention_mask
+            input_values = input_values.cpu().numpy()
+            attention_mask = attention_mask.cpu().numpy() if attention_mask is not None else attention_mask
 
         inputs = {
             "input_values": input_values,
@@ -791,8 +791,8 @@ def forward(
     ):
         np_inputs = isinstance(input_values, np.ndarray)
         if not np_inputs:
-            input_values = np.array(input_values)
-            attention_mask = np.array(attention_mask) if attention_mask is not None else attention_mask
+            input_values = input_values.cpu().numpy()
+            attention_mask = attention_mask.cpu().numpy() if attention_mask is not None else attention_mask
 
         inputs = {
             "input_values": input_values,
@@ -867,8 +867,8 @@ def forward(
     ):
         np_inputs = isinstance(input_values, np.ndarray)
         if not np_inputs:
-            input_values = np.array(input_values)
-            attention_mask = np.array(attention_mask) if attention_mask is not None else attention_mask
+            input_values = input_values.cpu().numpy()
+            attention_mask = attention_mask.cpu().numpy() if attention_mask is not None else attention_mask
 
         inputs = {
             "input_values": input_values,
@@ -929,7 +929,7 @@ def forward(self, **kwargs):
         np_inputs = isinstance(next(iter(kwargs.values())), np.ndarray)
         inputs = {}
         for input_name in self.input_names:
-            inputs[input_name] = np.array(kwargs.pop(input_name)) if not np_inputs else kwargs.pop(input_name)
+            inputs[input_name] = kwargs.pop(input_name).cpu().numpy() if not np_inputs else kwargs.pop(input_name)
 
         outputs = self._inference(inputs)
         model_outputs = {}

optimum/intel/openvino/modeling_decoder.py

@@ -21,8 +21,8 @@
 import openvino
 import torch
 from huggingface_hub.constants import HUGGINGFACE_HUB_CACHE
+from openvino import Core, Tensor, Type
 from openvino.preprocess import PrePostProcessor
-from openvino.runtime import Core, Tensor, Type
 from transformers import AutoModelForCausalLM, PretrainedConfig
 from transformers.file_utils import add_start_docstrings, add_start_docstrings_to_model_forward
 from transformers.generation import GenerationMixin
@@ -492,11 +492,11 @@ def prepare_inputs(
                 self.next_beam_idx = np.arange(batch_size, dtype=int)
                 self._past_length = 0
         past_len = self._get_past_length(past_key_values)
-        inputs["input_ids"] = np.array(input_ids)
+        inputs["input_ids"] = input_ids.cpu().numpy()
         # Add the attention_mask inputs when needed
         if "attention_mask" in self.input_names or "position_ids" in self.input_names:
             if attention_mask is not None:
-                attention_mask = np.array(attention_mask)
+                attention_mask = attention_mask.cpu().numpy()
             else:
                 attention_mask = np.ones(
                     (input_ids.shape[0], input_ids.shape[1] + past_len), dtype=inputs["input_ids"].dtype
@@ -507,7 +507,7 @@ def prepare_inputs(
 
         if "position_ids" in self.input_names:
             if position_ids is not None:
-                position_ids = np.array(position_ids)
+                position_ids = position_ids.cpu().numpy()
             else:
                 position_ids = np.cumsum(attention_mask, axis=1) - 1
                 position_ids[attention_mask == 0] = 1

optimum/intel/openvino/modeling_sentence_transformers.py

@@ -43,9 +43,9 @@ def forward(self, inputs: Dict[str, torch.Tensor]):
 
         np_inputs = isinstance(input_ids, np.ndarray)
         if not np_inputs:
-            input_ids = np.array(input_ids)
-            attention_mask = np.array(attention_mask)
-            token_type_ids = np.array(token_type_ids) if token_type_ids is not None else token_type_ids
+            input_ids = input_ids.cpu().numpy()
+            attention_mask = attention_mask.cpu().numpy()
+            token_type_ids = token_type_ids.cpu().numpy() if token_type_ids is not None else token_type_ids
 
         inputs = {
             "input_ids": input_ids,

optimum/intel/openvino/modeling_visual_language.py

@@ -150,7 +150,7 @@ def prepare_inputs(
         # Add the attention_mask inputs when needed
         if "attention_mask" in self.input_names or "position_ids" in self.input_names:
             if attention_mask is not None:
-                attention_mask = np.array(attention_mask)
+                attention_mask = attention_mask.cpu().numpy()
             else:
                 attention_mask = np.ones((inputs_embeds.shape[0], inputs_embeds.shape[1] + past_len), dtype=int)
 
@@ -159,7 +159,7 @@ def prepare_inputs(
 
         if "position_ids" in self.input_names:
             if position_ids is not None:
-                position_ids = np.array(position_ids)
+                position_ids = position_ids.cpu().numpy()
             else:
                 position_ids = np.cumsum(attention_mask, axis=1) - 1
                 position_ids[attention_mask == 0] = 1