Add kernels as an extra hub-kernels

Also add it to `testing`, so that the kernel replacement gets tested
when using CUDA in CI.

Author: danieldk

setup.py

@@ -302,8 +302,9 @@ def run(self):
 extras["optuna"] = deps_list("optuna")
 extras["ray"] = deps_list("ray[tune]")
 extras["sigopt"] = deps_list("sigopt")
+extras["hub-kernels"] = deps_list("kernels")
 
-extras["integrations"] = extras["optuna"] + extras["ray"] + extras["sigopt"]
+extras["integrations"] = extras["hub-kernels"] + extras["optuna"] + extras["ray"] + extras["sigopt"]
 
 extras["serving"] = deps_list("pydantic", "uvicorn", "fastapi", "starlette")
 extras["audio"] = deps_list("librosa", "pyctcdecode", "phonemizer", "kenlm")
@@ -412,7 +413,6 @@ def run(self):
     "filelock",
     "huggingface-hub",
     "importlib_metadata",
-    "kernels",
     "numpy",
     "packaging",
     "protobuf",
@@ -434,7 +434,6 @@ def run(self):
 install_requires = [
     deps["filelock"],  # filesystem locks, e.g., to prevent parallel downloads
     deps["huggingface-hub"],
-    deps["kernels"],  # download kernels from the Hub
     deps["numpy"],
     deps["packaging"],  # utilities from PyPA to e.g., compare versions
     deps["pyyaml"],  # used for the model cards metadata

src/transformers/dependency_versions_check.py

@@ -29,7 +29,6 @@
     "requests",
     "packaging",
     "filelock",
-    "kernels",
     "numpy",
     "tokenizers",
     "huggingface-hub",

src/transformers/integrations/hub_kernels.py

@@ -13,28 +13,60 @@
 # limitations under the License.
 from typing import Dict, Union
 
-from kernels import (
-    Device,
-    LayerRepository,
-    register_kernel_mapping,
-    replace_kernel_forward_from_hub,
-    use_kernel_forward_from_hub,
-)
-
-
-_KERNEL_MAPPING: Dict[str, Dict[Union[Device, str], LayerRepository]] = {
-    "MultiScaleDeformableAttention": {
-        "cuda": LayerRepository(
-            repo_id="kernels-community/deformable-detr",
-            layer_name="MultiScaleDeformableAttention",
-        )
+
+try:
+    from kernels import (
+        Device,
+        LayerRepository,
+        register_kernel_mapping,
+        replace_kernel_forward_from_hub,
+        use_kernel_forward_from_hub,
+    )
+
+    _hub_kernels_available = True
+
+    _KERNEL_MAPPING: Dict[str, Dict[Union[Device, str], LayerRepository]] = {
+        "MultiScaleDeformableAttention": {
+            "cuda": LayerRepository(
+                repo_id="kernels-community/deformable-detr",
+                layer_name="MultiScaleDeformableAttention",
+            )
+        }
     }
-}
 
-register_kernel_mapping(_KERNEL_MAPPING)
+    register_kernel_mapping(_KERNEL_MAPPING)
+
+except ImportError:
+    # Stub to make decorators int transformers work when `kernels`
+    # is not installed.
+    def use_kernel_forward_from_hub(*args, **kwargs):
+        def decorator(cls):
+            return cls
+
+        return decorator
+
+    class LayerRepository:
+        def __init__(self, *args, **kwargs):
+            raise RuntimeError("LayerRepository requires `kernels` to be installed. Run `pip install kernels`.")
+
+    def replace_kernel_forward_from_hub(*args, **kwargs):
+        raise RuntimeError(
+            "replace_kernel_forward_from_hub requires `kernels` to be installed. Run `pip install kernels`."
+        )
+
+    def register_kernel_mapping(*args, **kwargs):
+        raise RuntimeError("register_kernel_mapping requires `kernels` to be installed. Run `pip install kernels`.")
+
+    _hub_kernels_available = False
+
+
+def is_hub_kernels_available():
+    return _hub_kernels_available
+
 
 __all__ = [
     "LayerRepository",
+    "is_hub_kernels_available",
     "use_kernel_forward_from_hub",
     "register_kernel_mapping",
     "replace_kernel_forward_from_hub",