Transformers 4.53 support, SmolLM3 and Fix old Transformers support (#2319)

* test

* fix seq2seq patched sdpa

* patch qwen3_moe, out_attentions, and eager_mask

* fix

* use optimum model

* editable subpackages

* Apply suggestions from code review

* smollm3 support

* deprecate tensorflow onnx export and add smollm3 to export tests

* write a more general sdpa_mask without vmap that's also vectorized and broadcastable

* better and more generic sdpa_mask_without_vmap implementation

* style and fix

* fix

* patch find_packed_sequence_indices as it's untraceable

* fix

* fix

* revert tests removal until refactor

* fix temporary hub repo import

* fix

* fix external data tests on windows

* update phi and phi3 min version

* condition modernbert optimization test

* get back old (pre 4.44) bloom modeling support and remove the need for normalized config in ort modeling

* fix test was using hardcoded architecture

* unparallelize test that uses remote code

* support older versions of mpt and phi (4.36)

* remove parallelism from slow tests

* fix vision to text pipelines test

* more specific version handling for find_packed_sequence_indices

* fix

Author: IlyasMoutawwakil

.github/workflows/test_exporters_onnx.yml

@@ -40,4 +40,4 @@ jobs:
 
       - name: Test with pytest
         run: |
-          pytest tests/exporters/onnx/test_export.py -vvvv --durations=0 -n auto
+          pytest tests/exporters/onnx/test_export.py -vvvv --durations=0

.github/workflows/test_exporters_tflite.yml

@@ -6,7 +6,6 @@ on:
     branches: [main]
   pull_request:
     branches: [main]
-    types: [opened, synchronize, reopened, labeled, unlabeled]
 
 concurrency:
   group: ${{ github.workflow }}-${{ github.head_ref || github.run_id }}

.github/workflows/test_exporters_tflite_cli.yml

@@ -6,7 +6,6 @@ on:
     branches: [main]
   pull_request:
     branches: [main]
-    types: [opened, synchronize, reopened, labeled, unlabeled]
 
 concurrency:
   group: ${{ github.workflow }}-${{ github.head_ref || github.run_id }}

.github/workflows/test_onnxruntime_slow.yml

@@ -87,7 +87,7 @@ jobs:
 
       - name: Test with pytest (in parallel)
         run: |
-          pytest tests/onnxruntime -m "not run_in_series" --durations=0 -vvvv -n auto
+          pytest tests/onnxruntime -m "not run_in_series" --durations=0 -vvvv
         env:
           HF_HUB_READ_TOKEN: ${{ secrets.HF_HUB_READ_TOKEN }}
           RUN_SLOW: 1

optimum/commands/__init__.py

@@ -11,6 +11,7 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
+__path__ = __import__("pkgutil").extend_path(__path__, __name__)
 
 from .base import BaseOptimumCLICommand, CommandInfo, RootOptimumCLICommand
 from .env import EnvironmentCommand

optimum/commands/export/__init__.py

@@ -11,7 +11,7 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-
+__path__ = __import__("pkgutil").extend_path(__path__, __name__)
 
 from .base import ExportCommand
 from .onnx import ONNXExportCommand

optimum/exporters/__init__.py

@@ -12,5 +12,7 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-from . import onnx  # noqa
+__path__ = __import__("pkgutil").extend_path(__path__, __name__)
+
 from .tasks import TasksManager  # noqa
+from .base import ExporterConfig  # noqa

optimum/exporters/base.py

@@ -69,7 +69,12 @@
 
 
 class ExportConfig(ABC):
-    pass
+    def __init_subclass__(cls, **kwargs):
+        super().__init_subclass__(**kwargs)
+        logger.warning(
+            "The `ExportConfig` class is deprecated and will be removed in a future version. "
+            "Please use `ExporterConfig` instead."
+        )
 
 
 class ExporterConfig(ABC):

optimum/exporters/onnx/__main__.py

@@ -340,6 +340,10 @@ def main_export(
         if model_type in SDPA_ARCHS_ONNX_EXPORT_NOT_SUPPORTED and is_transformers_version("<", "4.42"):
             loading_kwargs["attn_implementation"] = "eager"
 
+        # Only eager attention implementation returns attentions
+        if model_kwargs is not None and model_kwargs.get("output_attentions", False):
+            loading_kwargs["attn_implementation"] = "eager"
+
     with DisableCompileContextManager():
         model = TasksManager.get_model_from_task(
             task,

optimum/exporters/onnx/convert.py

@@ -642,6 +642,11 @@ def export_tensorflow(
         `Tuple[List[str], List[str]]`: A tuple with an ordered list of the model's inputs, and the named outputs from
         the ONNX configuration.
     """
+
+    logger.warning(
+        "The TensorFlow ONNX export is deprecated and will be removed in the next major release of Optimum."
+    )
+
     # This is needed to import onnx and tf2onnx because onnx is also the name of the current directory.
     import sys