[OpenVINO backend] Adding support for OpenVINO backend & support inference for Mistral & Gemma & GPT2

.github/workflows/actions.yml

@@ -16,22 +16,25 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        backend: [tensorflow, jax, torch]
+        backend: [tensorflow, jax, torch, openvino]
         version: [keras-stable]
         include:
           - backend: jax
             version: keras-3.5
           - backend: jax
             version: keras-nightly
+          - backend: openvino
+            version: keras-stable
+            python-version: '3.10'
     runs-on: ubuntu-latest
     env:
       KERAS_BACKEND: ${{ matrix.backend }}
     steps:
     - uses: actions/checkout@v4
-    - name: Set up Python 3.9
+    - name: Set up Python
       uses: actions/setup-python@v5
       with:
-        python-version: 3.9
+        python-version: ${{ matrix.python-version || '3.9' }}
     - name: Get pip cache dir
       id: pip-cache
       run: |
@@ -48,6 +51,10 @@ jobs:
       run: |
           pip install -r requirements.txt --progress-bar off
           pip install --no-deps -e "." --progress-bar off
+          if [[ "${{ matrix.backend }}" == "openvino" ]]; then
+            pip uninstall -y keras
+            pip install git+https://github.com/Mohamed-Ashraf273/keras.git@gsoc2025 --upgrade --force-reinstall --progress-bar off
+          fi
     - name: Pin Keras 3.5
       if: ${{ matrix.version == 'keras-3.5'}}
       run: |
@@ -60,7 +67,13 @@ jobs:
         pip install keras-nightly --progress-bar off
     - name: Test with pytest
       run: |
-        pytest keras_hub/
+        if [ "${{ matrix.backend }}" == "openvino" ]; then
+          IGNORE_FILE="openvino_excluded_tests.txt"
+          IGNORE_ARGS=$(awk '{print "--ignore=" $0}' "$IGNORE_FILE")
+        else
+          IGNORE_ARGS=""
+        fi
+        pytest keras_hub/ $IGNORE_ARGS
     - name: Run integration tests
       run: |
         python pip_build.py --install

conftest.py

@@ -70,6 +70,10 @@ def pytest_configure(config):
         "markers",
         "kaggle_key_required: mark test needing a kaggle key",
     )
+    config.addinivalue_line(
+        "markers",
+        "requires_trainable_backend: mark test for trainable backend only",
+    )
 
 
 def pytest_collection_modifyitems(config, items):
@@ -110,6 +114,44 @@ def pytest_collection_modifyitems(config, items):
         if "kaggle_key_required" in item.keywords:
             item.add_marker(kaggle_key_required)
 
+    openvino_skipped_tests = []
+    if keras.config.backend() == "openvino":
+        from pathlib import Path
+
+        workspace_root = Path(__file__).resolve().parents[0]
+        file_path = workspace_root / "openvino_excluded_concrete_tests.txt"
+        with open(file_path, "r") as file:
+            openvino_skipped_tests = [
+                line.strip() for line in file if line.strip()
+            ]
+
+    requires_trainable_backend = pytest.mark.skipif(
+        keras.config.backend() in ["openvino"],
+        reason="fit not implemented for OpenVINO backend.",
+    )
+
+    for item in items:
+        if "requires_trainable_backend" in item.keywords:
+            item.add_marker(requires_trainable_backend)
+        # also, skip concrete tests for openvino, listed in the special file
+        # this is more granular mechanism to exclude tests rather
+        # than using --ignore option
+        for skipped_test in openvino_skipped_tests:
+            if skipped_test in item.nodeid:
+                item.add_marker(
+                    skip_if_backend(
+                        "openvino",
+                        "Not supported operation by openvino backend",
+                    )
+                )
+                break
+
+
+def skip_if_backend(given_backend, reason):
+    return pytest.mark.skipif(
+        keras.config.backend() == given_backend, reason=reason
+    )
+
 
 # Disable traceback filtering for quicker debugging of tests failures.
 keras.config.disable_traceback_filtering()

integration_tests/basic_usage_test.py

@@ -6,6 +6,10 @@
 import keras_hub
 
 
+@unittest.skipIf(
+    keras.backend.backend() == "openvino",
+    "Skip for non-trainable backends like OpenVINO",
+)
 class BasicUsageTest(unittest.TestCase):
     def test_transformer(self):
         # Tokenize some inputs with a binary label.

keras_hub/src/layers/modeling/alibi_bias_test.py

@@ -1,4 +1,5 @@
 import keras
+import pytest
 from keras import ops
 from keras import random
 
@@ -7,6 +8,7 @@
 
 
 class AlibiBiasTest(TestCase):
+    @pytest.mark.requires_trainable_backend
     def test_layer_behaviors(self):
         alibi_bias_max = 8
         batch_size = 4

keras_hub/src/layers/modeling/anchor_generator_test.py

@@ -14,6 +14,7 @@
     reason="Bbox utils are not supported before keras < 3.8.0",
 )
 class AnchorGeneratorTest(TestCase):
+    @pytest.mark.requires_trainable_backend
     def test_layer_behaviors(self):
         images_shape = (8, 128, 128, 3)
         self.run_layer_test(

keras_hub/src/layers/modeling/cached_multi_head_attention_test.py

@@ -1,3 +1,4 @@
+import pytest
 from keras import ops
 from keras import random
 
@@ -8,6 +9,7 @@
 
 
 class CachedMultiHeadAttentionTest(TestCase):
+    @pytest.mark.requires_trainable_backend
     def test_layer_behaviors(self):
         self.run_layer_test(
             cls=CachedMultiHeadAttention,
@@ -75,6 +77,7 @@ def call(outputs, cache):
         self.assertAllClose(output, no_loop_outputs)
         self.assertAllClose(output_cache, no_loop_cache)
 
+    @pytest.mark.requires_trainable_backend
     def test_training_propagation(self):
         batch_size = 2
         seq_len = 5

keras_hub/src/layers/modeling/f_net_encoder_test.py

@@ -1,3 +1,4 @@
+import pytest
 from keras import ops
 from keras import random
 
@@ -6,6 +7,7 @@
 
 
 class FNetEncoderTest(TestCase):
+    @pytest.mark.requires_trainable_backend
     def test_layer_behaviors(self):
         self.run_layer_test(
             cls=FNetEncoder,
@@ -31,6 +33,7 @@ def test_value_error_when_invalid_kernel_initializer(self):
                 kernel_initializer="Invalid",
             )
 
+    @pytest.mark.requires_trainable_backend
     def test_training_propagation(self):
         x = random.uniform(shape=(2, 4, 6))
         layer = FNetEncoder(

keras_hub/src/layers/modeling/masked_lm_head_test.py

@@ -1,3 +1,4 @@
+import pytest
 from keras import random
 
 from keras_hub.src.layers.modeling.masked_lm_head import MaskedLMHead
@@ -8,6 +9,7 @@
 
 
 class MaskedLMHeadTest(TestCase):
+    @pytest.mark.requires_trainable_backend
     def test_layer_behaviors(self):
         self.run_layer_test(
             cls=MaskedLMHead,
@@ -27,6 +29,7 @@ def test_layer_behaviors(self):
             expected_num_trainable_weights=6,
         )
 
+    @pytest.mark.requires_trainable_backend
     def test_layer_behaviors_with_embedding(self):
         embedding = ReversibleEmbedding(100, 16)
         embedding.build((4, 10))

keras_hub/src/layers/modeling/position_embedding_test.py

@@ -1,5 +1,6 @@
 import keras
 import numpy as np
+import pytest
 from keras import ops
 from keras import random
 
@@ -15,6 +16,7 @@ def custom_init(shape, dtype=None):
 
 
 class PositionEmbeddingTest(TestCase):
+    @pytest.mark.requires_trainable_backend
     def test_layer_behaviors(self):
         self.run_layer_test(
             cls=PositionEmbedding,
@@ -26,6 +28,7 @@ def test_layer_behaviors(self):
             expected_num_trainable_weights=1,
         )
 
+    @pytest.mark.requires_trainable_backend
     def test_layer_behaviors_4d(self):
         self.run_layer_test(
             cls=PositionEmbedding,

keras_hub/src/layers/modeling/reversible_embedding_test.py

@@ -2,6 +2,7 @@
 
 import keras
 import numpy as np
+import pytest
 from absl.testing import parameterized
 from keras import ops
 from keras import random
@@ -17,6 +18,7 @@ class ReversibleEmbeddingTest(TestCase):
         ("tie_weights", True),
         ("untie_weights", False),
     )
+    @pytest.mark.requires_trainable_backend
     def test_layer_behaviors_tied(self, tie_weights):
         self.run_layer_test(
             cls=ReversibleEmbedding,

keras_hub/src/layers/modeling/rotary_embedding_test.py

@@ -1,5 +1,6 @@
 import keras
 import numpy as np
+import pytest
 from keras import ops
 from keras import random
 
@@ -8,6 +9,7 @@
 
 
 class RotaryEmbeddingTest(TestCase):
+    @pytest.mark.requires_trainable_backend
     def test_layer_behaviors(self):
         self.run_layer_test(
             cls=RotaryEmbedding,
@@ -21,6 +23,7 @@ def test_layer_behaviors(self):
             expected_output_shape=(2, 4, 6),
         )
 
+    @pytest.mark.requires_trainable_backend
     def test_layer_behaviors_4d(self):
         self.run_layer_test(
             cls=RotaryEmbedding,

keras_hub/src/layers/modeling/sine_position_encoding_test.py

@@ -1,4 +1,5 @@
 import keras
+import pytest
 from keras import ops
 from keras import random
 
@@ -9,6 +10,7 @@
 
 
 class SinePositionEncodingTest(TestCase):
+    @pytest.mark.requires_trainable_backend
     def test_layer_behaviors(self):
         self.run_layer_test(
             cls=SinePositionEncoding,
@@ -19,6 +21,7 @@ def test_layer_behaviors(self):
             expected_output_shape=(2, 4, 6),
         )
 
+    @pytest.mark.requires_trainable_backend
     def test_layer_behaviors_4d(self):
         self.run_layer_test(
             cls=SinePositionEncoding,

keras_hub/src/layers/modeling/token_and_position_embedding_test.py

@@ -1,4 +1,5 @@
 import numpy as np
+import pytest
 from keras import ops
 from keras import random
 
@@ -9,6 +10,7 @@
 
 
 class TokenAndPositionEmbeddingTest(TestCase):
+    @pytest.mark.requires_trainable_backend
     def test_layer_behaviors(self):
         self.run_layer_test(
             cls=TokenAndPositionEmbedding,

keras_hub/src/layers/modeling/transformer_decoder_test.py

@@ -1,3 +1,4 @@
+import pytest
 from absl.testing import parameterized
 from keras import ops
 from keras import random
@@ -11,6 +12,7 @@ class TransformerDecoderTest(TestCase):
         ("without_norm_first", False),
         ("with_norm_first", True),
     )
+    @pytest.mark.requires_trainable_backend
     def test_layer_behaviors(self, normalize_first):
         self.run_layer_test(
             cls=TransformerDecoder,
@@ -34,6 +36,7 @@ def test_layer_behaviors(self, normalize_first):
         ("without_norm_first", False),
         ("with_norm_first", True),
     )
+    @pytest.mark.requires_trainable_backend
     def test_layer_behaviors_with_cross_attention(self, normalize_first):
         self.run_layer_test(
             cls=TransformerDecoder,
@@ -89,6 +92,7 @@ def test_value_error_when_invalid_kernel_inititalizer(self):
                 kernel_initializer="Invalid",
             )
 
+    @pytest.mark.requires_trainable_backend
     def test_training_propagation(self):
         decoder = TransformerDecoder(
             intermediate_dim=4,

keras_hub/src/layers/modeling/transformer_encoder_test.py

@@ -1,4 +1,5 @@
 import keras
+import pytest
 from absl.testing import parameterized
 from keras import ops
 from keras import random
@@ -12,6 +13,7 @@ class TransformerEncoderTest(TestCase):
         ("without_norm_first", False),
         ("with_norm_first", True),
     )
+    @pytest.mark.requires_trainable_backend
     def test_layer_behaviors(self, normalize_first):
         self.run_layer_test(
             cls=TransformerEncoder,
@@ -69,6 +71,7 @@ def test_value_error_when_invalid_kernel_inititalizer(self):
                 kernel_initializer="Invalid",
             )
 
+    @pytest.mark.requires_trainable_backend
     def test_training_propagation(self):
         encoder = TransformerEncoder(
             intermediate_dim=4,