Set version to 0.7.0

* Support registering inference/training-specific layers

This change makes it possible to register kernels specialized for
inference, training, and/or `torch.compile`. To do so, the mapping
notation is extended to support registering specialized kernels
for a specific 'mode'. For instance, the following mapping,

```python
kernel_layer_mapping = {
    "SiluAndMul": {
        "cuda": {
          Mode.DEFAULT: LayerRepository(
              repo_id="kernels-community/activation",
              layer_name="SiluAndMul",
          ),
          Mode.TRAINING | Mode.TORCH_COMPILE: LayerRepository(
              repo_id="kernels-community/activation-training-optimized",
              layer_name="SiluAndMul",
          ),
      }
    }
}
```

uses `kernels-community/activation` by default, but will switch to
using `kernels-community/activation-training-optimized` if a model
is kernelized for training and `torch.compile`.

To make it easier to add more modes in the future and to unify the
`register_kernel_mapping` and `kernelize` signatures, the `training`
and `needs_torch_compile` arguments of `kernelize` are replaced by
a single `mode` argument:

```python
model = MyModel(...)
model = kernelize(model, mode=Mode.TRAINING | Mode.TORCH_COMPILE)
```

* Documentation fixes

Co-authored-by: Mohamed Mekkouri <93391238+MekkCyber@users.noreply.github.com>

* Add note on when the fallback is used

* Tighten up some Mode checks

* Fix ruff check

* Attempt to fix mypy errors

* More typing fixes

* Ignore Python < 3.11 type check SNAFU

---------

Co-authored-by: Mohamed Mekkouri <93391238+MekkCyber@users.noreply.github.com>

.github/workflows/publish.yml

@@ -0,0 +1,120 @@
+name: Publish Python 🐍 distribution 📦 to PyPI and TestPyPI
+
+on: push
+
+jobs:
+  build:
+    name: Build distribution 📦
+    runs-on: ubuntu-latest
+
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          persist-credentials: false
+      - name: Set up Python
+        uses: actions/setup-python@v5
+        with:
+          python-version: "3.9"
+      - name: Install pypa/build
+        run: >-
+          python3 -m
+          pip install
+          build
+          --user
+      - name: Build a binary wheel and a source tarball
+        run: python3 -m build
+      - name: Store the distribution packages
+        uses: actions/upload-artifact@v4
+        with:
+          name: python-package-distributions
+          path: dist/
+
+  publish-to-pypi:
+    name: >-
+      Publish Python 🐍 distribution 📦 to PyPI
+    if: startsWith(github.ref, 'refs/tags/') # only publish to PyPI on tag pushes
+    needs:
+      - build
+    runs-on: ubuntu-latest
+    environment:
+      name: pypi
+      url: https://pypi.org/p/kernels
+    permissions:
+      id-token: write # IMPORTANT: mandatory for trusted publishing
+
+    steps:
+      - name: Download all the dists
+        uses: actions/download-artifact@v4
+        with:
+          name: python-package-distributions
+          path: dist/
+      - name: Publish distribution 📦 to PyPI
+        uses: pypa/gh-action-pypi-publish@release/v1
+
+  github-release:
+    name: >-
+      Sign the Python 🐍 distribution 📦 with Sigstore
+      and upload them to GitHub Release
+    needs:
+      - publish-to-pypi
+    runs-on: ubuntu-latest
+
+    permissions:
+      contents: write # IMPORTANT: mandatory for making GitHub Releases
+      id-token: write # IMPORTANT: mandatory for sigstore
+
+    steps:
+      - name: Download all the dists
+        uses: actions/download-artifact@v4
+        with:
+          name: python-package-distributions
+          path: dist/
+      - name: Sign the dists with Sigstore
+        uses: sigstore/gh-action-sigstore-python@v3.0.0
+        with:
+          inputs: >-
+            ./dist/*.tar.gz
+            ./dist/*.whl
+      - name: Create GitHub Release
+        env:
+          GITHUB_TOKEN: ${{ github.token }}
+        run: >-
+          gh release create
+          "$GITHUB_REF_NAME"
+          --repo "$GITHUB_REPOSITORY"
+          --notes ""
+      - name: Upload artifact signatures to GitHub Release
+        env:
+          GITHUB_TOKEN: ${{ github.token }}
+        # Upload to GitHub Release using the `gh` CLI.
+        # `dist/` contains the built packages, and the
+        # sigstore-produced signatures and certificates.
+        run: >-
+          gh release upload
+          "$GITHUB_REF_NAME" dist/**
+          --repo "$GITHUB_REPOSITORY"
+
+  publish-to-testpypi:
+    name: Publish Python 🐍 distribution 📦 to TestPyPI
+    needs:
+      - build
+    runs-on: ubuntu-latest
+
+    environment:
+      name: testpypi
+      url: https://test.pypi.org/p/kernels
+
+    permissions:
+      id-token: write # IMPORTANT: mandatory for trusted publishing
+
+    steps:
+      - name: Download all the dists
+        uses: actions/download-artifact@v4
+        with:
+          name: python-package-distributions
+          path: dist/
+      - name: Publish distribution 📦 to TestPyPI
+        uses: pypa/gh-action-pypi-publish@release/v1
+        with:
+          repository-url: https://test.pypi.org/legacy/
+          skip-existing: true # Only upload when the version is unique.

.github/workflows/test.yml

@@ -24,7 +24,7 @@ jobs:
       max-parallel: 4
       matrix:
         python-version: ["3.10", "3.12"]
-        torch-version: ["2.5.1", "2.6.0"]
+        torch-version: ["2.6.0", "2.7.0"]
 
     env:
       UV_PYTHON_PREFERENCE: only-managed
@@ -53,6 +53,18 @@ jobs:
       - name: Run tests
         run: uv run pytest tests
 
+      - name: Check kernel conversion
+        run: |
+          uv pip install wheel
+          uv run kernels to-wheel kernels-community/triton-layer-norm 0.0.1
+          uv pip install triton_layer_norm-0.0.1*.whl
+          uv run python -c "import triton_layer_norm"
+
+      - name: Check README generation
+        # For now, just checks that generation doesn't fail.
+        run: |
+          uv run kernels generate-readme kernels-community/triton-layer-norm
+
       - name: Import check without torch
         run: |
           uv pip uninstall torch

README.md

@@ -61,4 +61,5 @@ the Hub.
 - [Environment variables](docs/env.md)
 - [Using kernels in a Docker container](docs/docker.md)
 - [Kernel requirements](docs/kernel-requirements.md)
+- [Frequently Asked Questions](docs/faq.md)
 - [Writing kernels](https://github.com/huggingface/kernel-builder/blob/main/docs/writing-kernels.md) using [kernel-builder](https://github.com/huggingface/kernel-builder/)

docs/faq.md

@@ -0,0 +1,13 @@
+# FAQ
+
+## Why is the kernelization step needed?
+
+In earlier versions of `kernels`, a layer's `forward` was replaced by
+`use_kernel_forward_from_hub` and `replace_kernel_forward_from_hub`. The
+new `forward` would dispatch to a kernel based on the device type,
+whether a model was training, etc. However, this approach was
+fundamentally incompatible with `torch.compile` since it relied
+on data-dependent branching.
+
+To avoid branching, we have to make dispatch decisions ahead of time,
+which is what the `kernelize` function does.

docs/kernel-requirements.md

@@ -37,8 +37,14 @@ to resolve the version constraints.
 ## Native Python module
 
 Kernels will typically contain a native Python module with precompiled
-compute kernels and bindings. This module must fulfill the following
-requirements:
+compute kernels and bindings. This module must fulfill the requirements
+outlined in this section. For all operating systems, a kernel must not
+have dynamic library dependencies outside:
+
+- Torch;
+- CUDA/ROCm libraries installed as dependencies of Torch.
+
+### Linux
 
 - Use [ABI3/Limited API](https://docs.python.org/3/c-api/stable.html#stable-application-binary-interface)
   for compatibility with Python 3.9 and later.
@@ -50,12 +56,18 @@ requirements:
   - CXXABI 1.3.11
   - GCC 7.0.0
 
-  These requirement can be checked with the ABI checker (see below).
+These requirement can be checked with the ABI checker (see below).
 
-- No dynamic library dependencies outside:
+### macOS
 
-  - Torch;
-  - CUDA/ROCm libraries installed as dependencies of Torch.
+- Use [ABI3/Limited API](https://docs.python.org/3/c-api/stable.html#stable-application-binary-interface)
+  for compatibility with Python 3.9 and later.
+- macOS deployment target 15.0.
+- Metal 3.0 (`-std=metal3.0`).
+
+The ABI3 requirement can be checked with the ABI checker (see below).
+
+### ABI checker
 
 The manylinux_2_28 and Python ABI 3.9 version requirements can be checked with
 [`kernel-abi-check`](https://crates.io/crates/kernel-abi-check):
@@ -109,9 +121,12 @@ requirements:
 - The `forward` method has a signature that is compatible with the
   `forward` method that it is extending.
 
-The only exception to the _no class variables rule_ is addition of a
-`has_backward` class variable. This variable is used to indicate whether
-the layer has a backward pass implemented (`True` when absent).
+There are two exceptions to the _no class variables rule_:
+
+1. The `has_backward` variable can be used to indicate whether the layer has
+   a backward pass implemented (`True` when absent).
+2. The `can_torch_compile` variable can be used to indicate whether the layer
+   supports `torch.compile` (`False` when absent).
 
 This is an example of a pure layer:
 

docs/layers.md

@@ -23,33 +23,93 @@ class SiluAndMul(nn.Module):
         return F.silu(input[..., :d]) * input[..., d:]
 ```
 
-The decorator changes the layer, so that other implementations of the `forward`
-method can be registered using the name `SiluAndMul`.
+The decorator does not change the behavior of the class -- it annotates
+the class with the given name (here `SiluAndMul`). The `kernelize` function
+described below uses this name to look up kernels for the layer.
 
 ### External layers
 
 An existing layer that does not (yet) have the `use_kernel_forward_from_hub`
-decorator can be made extensible by by monkeypatching it using the `replace_kernel_forward_from_hub` function.
+decorator can be made extensible using the `replace_kernel_forward_from_hub`
+function:
 
 ```python
 from somelibrary import SiluAndMul
 
 replace_kernel_forward_from_hub(SiluAndMul, "SiluAndMul")
-register_kernel_mapping(kernel_layer_mapping)
 ```
 
-The `register_kernel_mapping` call maps the name `SiluAndMul` to actual
-hub kernels. See the [Registering a hub kernel for a layer](#registering-a-hub-kernel-for-a-layer)
-section for more information.
-
 **Warning:** we strongly recommend using layers with a decorator, since
 it signifies that the maintainer intends to keep the `forward` signature
 compatible with layers from the hub.
 
+## Kernelizing a model
+
+A model will not use Hub kernels by default, even if it contains extensible
+layers. To enable the use of Hub kernels in the model, it needs to be
+'kernelized' using the `kernelize` function. This function traverses the
+model graph and replaces the `forward` methods of extensible layers for which
+Hub kernels are registered. Kernelize can be used as follows:
+
+```python
+model = MyModel(...)
+model = kernelize(model, mode=Mode.INFERENCE)
+```
+
+The `mode` specifies that the model will be used in inference. Similarly,
+you can ask `kernelize` to prepare the model for training:
+
+```python
+model = MyModel(...)
+model = kernelize(model, mode=Mode.TRAINING)
+```
+
+**Note:** the `kernelize` function modifies the model in-place, the model
+itself is returned as a convenience.
+
+### Kernel device
+
+Kernels can be registered per device type. For instance, separate `cuda` and
+`metal` kernels could be registered for the name `SiluAndMul`. By default,
+`kernelize` will try to infer the device type from the model's parameters.
+You can pass the device type to `kernelize` if the device type cannot be
+inferred (e.g. because the model has no parameters):
+
+```python
+model = MyModel(...)
+model = kernelize(model, device="cuda", mode=Mode.INFERENCE)
+```
+
+### `torch.compile`
+
+Not all Hub kernels support `torch.compile`. If you want to compile a model
+after kernelizing it, you need to add this to the mode. You can use the
+set union (`|`) operator to add `TORCH_COMPILE` to the mode:
+
+```python
+model = MyModel(...)
+model = kernelize(model, mode=Mode.INFERENCE | Mode.TORCH_COMPILE)
+```
+
+### Fallback `forward`
+
+If the `TRAINING` and/or `TORCH_COMPILE` modes are used, but a registered
+kernel does not support backward passes or `torch.compile` respectively,
+`kernenize` will fall back to the original, non-kernelized, layer. You
+can let `kernelize` raise an exception instead by using `use_fallback=False`:
+
+```python
+model = MyModel(...)
+model = kernelize(model, mode=Mode.INFERENCE | Mode.TORCH_COMPILE, use_fallback=False)
+```
+
+This can be useful if you want to guarantee that Hub kernels are used.
+
 ## Registering a hub kernel for a layer
 
-Once a layer is made extensible, users can register hub kernels for it
-by name using the `register_kernel_mapping` function. For example:
+`kernelize` relies on kernel mappings to find Hub kernels for layers.
+Kernel mappings map a kernel name such as `SiluAndMul` to a kernel on
+the Hub. For example:
 
 ```python
 kernel_layer_mapping = {
@@ -57,11 +117,14 @@ kernel_layer_mapping = {
         "cuda": LayerRepository(
             repo_id="kernels-community/activation",
             layer_name="SiluAndMul",
-            revision="layers",
         )
     }
 }
+```
 
+You can register such a mapping using `register_kernel_mapping`:
+
+```python
 register_kernel_mapping(kernel_layer_mapping)
 ```
 
@@ -72,8 +135,63 @@ used with the `use_kernel_mapping` context manager:
 ```python
 with use_kernel_mapping(kernel_layer_mapping):
     # Use the layer for which the mapping is applied.
-    ...
+    model = kernelize(model)
 ```
 
 This ensures that the mapping is not active anymore outside the
 `with`-scope.
+
+### Registering kernels for specific modes
+
+You might want to register two different kernels for a particular layer,
+where one kernel is optimized for a specific mode. You can do so by
+registering layer repositories for specific modes. For example:
+
+```python
+kernel_layer_mapping = {
+    "SiluAndMul": {
+        "cuda": {
+          Mode.INFERENCE: LayerRepository(
+              repo_id="kernels-community/activation-inference-optimized",
+              layer_name="SiluAndMul",
+          ),
+          Mode.TRAINING | Mode.TORCH_COMPILE: LayerRepository(
+              repo_id="kernels-community/activation-training-optimized",
+              layer_name="SiluAndMul",
+          ),
+      }
+    }
+}
+```
+
+The kernels will match exactly on the mode. So, for instance in the example above, no kernel
+layer is used when the `mode` passed to `kernelize` is
+`Mode.INFERENCE | Mode.TORCH_COMPILE` or `Mode.TRAINING`. However, if you want to
+register a kernel to be used when the mode does not match any of the
+modes in the mapping, you can use the special `Mode.DEFAULT` mode to do
+so. For example:
+
+```python
+kernel_layer_mapping = {
+    "SiluAndMul": {
+        "cuda": {
+          Mode.DEFAULT: LayerRepository(
+              repo_id="kernels-community/activation",
+              layer_name="SiluAndMul",
+          ),
+          Mode.INFERENCE: LayerRepository(
+              repo_id="kernels-community/activation-inference-optimized",
+              layer_name="SiluAndMul",
+          ),
+          Mode.TRAINING | Mode.TORCH_COMPILE: LayerRepository(
+              repo_id="kernels-community/activation-training-optimized",
+              layer_name="SiluAndMul",
+          ),
+      }
+    }
+}
+```
+
+In this case, modes other than `Mode.INFERENCE` and
+`Mode.TRAINING | Mode.TORCH_COMPILE` will be kernelized using
+`kernels-community/activation`.

flake.lock

@@ -51,18 +51,38 @@
         "type": "github"
       }
     },
+    "hf-nix": {
+      "inputs": {
+        "flake-compat": "flake-compat",
+        "flake-utils": "flake-utils_2",
+        "nixpkgs": "nixpkgs"
+      },
+      "locked": {
+        "lastModified": 1750775451,
+        "narHash": "sha256-HiGqtwzIgUH7Xkh+wgpvHRZGooqrW0z663E6nauczA4=",
+        "owner": "huggingface",
+        "repo": "hf-nix",
+        "rev": "5943c3169e861618a6634bc8dbdb498e413ab9b7",
+        "type": "github"
+      },
+      "original": {
+        "owner": "huggingface",
+        "repo": "hf-nix",
+        "type": "github"
+      }
+    },
     "nixpkgs": {
       "locked": {
-        "lastModified": 1737453259,
-        "narHash": "sha256-5LaFI9SQwCZmJDasMoYMdzNouWXNk3BvjKcO19tq1Rs=",
+        "lastModified": 1747820358,
+        "narHash": "sha256-fTqsZsUX6M3yeEvgyQvXcbGmT2CaRVyVwsi8eK29Oj4=",
         "owner": "danieldk",
         "repo": "nixpkgs",
-        "rev": "e0372dbcfd19ddd783b7c3b3868f19322f83318e",
+        "rev": "d3c1681180717528068082103bf323147de6ab0b",
         "type": "github"
       },
       "original": {
         "owner": "danieldk",
-        "ref": "outlines-v0.1.4-tgi",
+        "ref": "cudatoolkit-12.9-kernel-builder",
         "repo": "nixpkgs",
         "type": "github"
       }
@@ -70,11 +90,11 @@
     "root": {
       "inputs": {
         "flake-utils": "flake-utils",
+        "hf-nix": "hf-nix",
         "nixpkgs": [
-          "tgi-nix",
+          "hf-nix",
           "nixpkgs"
-        ],
-        "tgi-nix": "tgi-nix"
+        ]
       }
     },
     "systems": {
@@ -106,27 +126,6 @@
         "repo": "default",
         "type": "github"
       }
-    },
-    "tgi-nix": {
-      "inputs": {
-        "flake-compat": "flake-compat",
-        "flake-utils": "flake-utils_2",
-        "nixpkgs": "nixpkgs"
-      },
-      "locked": {
-        "lastModified": 1741617161,
-        "narHash": "sha256-cwKYAsIVSLtoLbG48+oi3NkSrvuZRLYs8lkJmpDsTw0=",
-        "owner": "huggingface",
-        "repo": "text-generation-inference-nix",
-        "rev": "5946021ec6cb6aae18158a9dc27f893cfbab2925",
-        "type": "github"
-      },
-      "original": {
-        "owner": "huggingface",
-        "ref": "kernels-0.2.0",
-        "repo": "text-generation-inference-nix",
-        "type": "github"
-      }
     }
   },
   "root": "root",

flake.nix

@@ -1,7 +1,7 @@
 {
   inputs = {
-    tgi-nix.url = "github:huggingface/text-generation-inference-nix/kernels-0.2.0";
-    nixpkgs.follows = "tgi-nix/nixpkgs";
+    hf-nix.url = "github:huggingface/hf-nix";
+    nixpkgs.follows = "hf-nix/nixpkgs";
     flake-utils.url = "github:numtide/flake-utils";
   };
   outputs =
@@ -9,21 +9,21 @@
       self,
       nixpkgs,
       flake-utils,
-      tgi-nix,
+      hf-nix,
     }:
     flake-utils.lib.eachDefaultSystem (
       system:
       let
         pkgs = import nixpkgs {
           inherit system;
-          inherit (tgi-nix.lib) config;
+          config = hf-nix.lib.config system;
           overlays = [
-            tgi-nix.overlays.default
+            hf-nix.overlays.default
           ];
         };
       in
       {
-        formatter = pkgs.nixfmt-rfc-style;
+        formatter = pkgs.nixfmt-tree;
         devShells = with pkgs; rec {
           default = mkShell {
             buildInputs =
@@ -34,10 +34,13 @@
                 ruff
               ]
               ++ (with python3.pkgs; [
+                docutils
                 huggingface-hub
                 pytest
                 pytest-benchmark
+                pyyaml
                 torch
+                types-pyyaml
                 venvShellHook
               ]);
 

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "kernels"
-version = "0.4.4"
+version = "0.7.0"
 description = "Download compute kernels"
 authors = [
   { name = "OlivierDehaene", email = "olivier@huggingface.co" },
@@ -14,6 +14,7 @@ requires-python = ">= 3.9"
 dependencies = [
   "huggingface_hub>=0.26.0,<1.0",
   "packaging>=20.0",
+  "pyyaml>=6",
   "tomli>=2.0; python_version<'3.11'",
 ]
 
@@ -23,11 +24,12 @@ build-backend = "setuptools.build_meta"
 
 [dependency-groups]
 dev = [
-  "mypy == 1.14.1",
+  "mypy >= 1.15.0",
   "pytest >=8",
   # Whatever version is compatible with pytest.
   "pytest-benchmark",
   "torch >=2.5",
+  "types-pyyaml"
 ]
 
 [project.optional-dependencies]

pytest.ini

@@ -0,0 +1,4 @@
+[pytest]
+markers =
+    darwin_only: marks tests that should only run on macOS
+    linux_only: marks tests that should only run on Linux

src/kernels/__init__.py

@@ -1,6 +1,8 @@
 from kernels.layer import (
     Device,
     LayerRepository,
+    Mode,
+    kernelize,
     register_kernel_mapping,
     replace_kernel_forward_from_hub,
     use_kernel_forward_from_hub,
@@ -8,6 +10,7 @@ from kernels.layer import (
 )
 from kernels.utils import (
     get_kernel,
+    get_local_kernel,
     get_locked_kernel,
     has_kernel,
     install_kernel,
@@ -15,15 +18,18 @@ from kernels.utils import (
 )
 
 __all__ = [
+    "Device",
+    "LayerRepository",
+    "Mode",
     "get_kernel",
+    "get_local_kernel",
     "get_locked_kernel",
     "has_kernel",
-    "load_kernel",
     "install_kernel",
-    "use_kernel_forward_from_hub",
-    "use_kernel_mapping",
+    "kernelize",
+    "load_kernel",
     "register_kernel_mapping",
     "replace_kernel_forward_from_hub",
-    "LayerRepository",
-    "Device",
+    "use_kernel_forward_from_hub",
+    "use_kernel_mapping",
 ]

src/kernels/_vendored/convert_rst_to_mdx.py

@@ -0,0 +1,751 @@
+# coding=utf-8
+# Copyright 2021 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# 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.
+
+# Vendored from https://github.com/huggingface/doc-builder/blob/main/src/doc_builder/convert_rst_to_mdx.py
+
+import re
+
+# Re pattern to catch things inside ` ` in :obj:`thing`.
+_re_obj = re.compile(r":obj:`([^`]+)`")
+# Re pattern to catch things inside ` ` in :math:`thing`.
+_re_math = re.compile(r":math:`([^`]+)`")
+# Re pattern to catch things between single backquotes.
+_re_single_backquotes = re.compile(r"(^|[^`])`([^`]+)`([^`]|$)")
+# Re pattern to catch things between double backquotes.
+_re_double_backquotes = re.compile(r"(^|[^`])``([^`]+)``([^`]|$)")
+# Re pattern to catch things inside ` ` in :func/class/meth:`thing`.
+_re_func_class = re.compile(r":(?:func|class|meth):`([^`]+)`")
+
+
+def convert_rst_formatting(text):
+    """
+    Convert rst syntax for formatting to markdown in a given text.
+    """
+    # Remove :class:, :func: and :meth: markers. To code-links and put double backquotes
+    # (to not be caught by the italic conversion).
+    text = _re_func_class.sub(r"[``\1``]", text)
+    # Remove :obj: markers. What's after is in a single backquotes so we put in double backquotes
+    # (to not be caught by the italic conversion).
+    text = _re_obj.sub(r"``\1``", text)
+    # Remove :math: markers.
+    text = _re_math.sub(r"\\\\(\1\\\\)", text)
+    # Convert content in single backquotes to italic.
+    text = _re_single_backquotes.sub(r"\1*\2*\3", text)
+    # Convert content in double backquotes to single backquotes.
+    text = _re_double_backquotes.sub(r"\1`\2`\3", text)
+    # Remove remaining ::
+    text = re.sub(r"::\n", "", text)
+
+    # Remove new lines inside blocks in backsticks as they will be kept.
+    lines = text.split("\n")
+    in_code = False
+    text = None
+    for line in lines:
+        if in_code:
+            splits = line.split("`")
+            in_code = len(splits) > 1 and len(splits) % 2 == 1
+            if len(splits) == 1:
+                # Some forgotten lone backstick
+                text += "\n" + line
+            else:
+                text += " " + line.lstrip()
+        else:
+            if text is not None:
+                text += "\n" + line
+            else:
+                text = line
+            splits = line.split("`")
+            in_code = len(splits) % 2 == 0
+    return text
+
+
+# Re pattern to catch description and url in links of the form `description <url>`_.
+_re_links = re.compile(r"`([^`]+\S)\s+</*([^/][^>`]*)>`_+")
+# Re pattern to catch description and url in links of the form :prefix_link:`description <url>`_.
+_re_prefix_links = re.compile(r":prefix_link:`([^`]+\S)\s+</*([^/][^>`]*)>`")
+# Re pattern to catch reference in links of the form :doc:`reference`.
+_re_simple_doc = re.compile(r":doc:`([^`<]*)`")
+# Re pattern to catch description and reference in links of the form :doc:`description <reference>`.
+_re_doc_with_description = re.compile(r":doc:`([^`<]+\S)\s+</*([^/][^>`]*)>`")
+# Re pattern to catch reference in links of the form :ref:`reference`.
+_re_simple_ref = re.compile(r":ref:`([^`<]*)`")
+# Re pattern to catch description and reference in links of the form :ref:`description <reference>`.
+_re_ref_with_description = re.compile(r":ref:`([^`<]+\S)\s+<([^>]*)>`")
+
+
+def convert_rst_links(text, page_info):
+    """
+    Convert the rst links in text to markdown.
+    """
+    if "package_name" not in page_info:
+        raise ValueError("`page_info` must contain at least the package_name.")
+    package_name = page_info["package_name"]
+    version = page_info.get("version", "main")
+    language = page_info.get("language", "en")
+    no_prefix = page_info.get("no_prefix", False)
+
+    prefix = "" if no_prefix else f"/docs/{package_name}/{version}/{language}/"
+    # Links of the form :doc:`page`
+    text = _re_simple_doc.sub(rf"[\1]({prefix}\1)", text)
+    # Links of the form :doc:`text <page>`
+    text = _re_doc_with_description.sub(rf"[\1]({prefix}\2)", text)
+
+    if "page" in page_info and not no_prefix:
+        page = str(page_info["page"])
+        if page.endswith(".html"):
+            page = page[:-5]
+        prefix = f"{prefix}{page}"
+    else:
+        prefix = ""
+    # Refs of the form :ref:`page`
+    text = _re_simple_ref.sub(rf"[\1]({prefix}#\1)", text)
+    # Refs of the form :ref:`text <page>`
+    text = _re_ref_with_description.sub(rf"[\1]({prefix}#\2)", text)
+
+    # Links with a prefix
+    # TODO: when it exists, use the API to deal with prefix links properly.
+    prefix = f"https://github.com/huggingface/{package_name}/tree/main/"
+    text = _re_prefix_links.sub(rf"[\1]({prefix}\2)", text)
+    # Other links
+    text = _re_links.sub(r"[\1](\2)", text)
+    # Relative links or Transformers links need to remove the .html
+    if (
+        "(https://https://huggingface.co/" in text
+        or re.search(r"\(\.+/", text) is not None
+    ):
+        text = text.replace(".html", "")
+    return text
+
+
+# Re pattern that catches examples blocks of the form `Example::`.
+_re_example = re.compile(r"^\s*(\S.*)::\s*$")
+# Re pattern that catches rst blocks of the form `.. block_name::`.
+_re_block = re.compile(r"^\s*\.\.\s+(\S+)::")
+# Re pattern that catches what's after the :: in rst blocks of the form `.. block_name:: something`.
+_re_block_info = re.compile(r"^\s*\.\.\s+\S+::\s*(\S.*)$")
+
+
+def is_empty_line(line):
+    return len(line) == 0 or line.isspace()
+
+
+def find_indent(line):
+    """
+    Returns the number of spaces that start a line indent.
+    """
+    search = re.search(r"^(\s*)(?:\S|$)", line)
+    if search is None:
+        return 0
+    return len(search.groups()[0])
+
+
+_re_rst_option = re.compile(r"^\s*:(\S+):(.*)$")
+
+
+def convert_special_chars(text):
+    """
+    Converts { and < that have special meanings in MDX.
+    """
+    text = text.replace("{", "&amp;lcub;")
+    # We don't want to replace those by the HTML code, so we temporarily set them at LTHTML
+    text = re.sub(
+        r"<(img|br|hr|Youtube)", r"LTHTML\1", text
+    )  # html void elements with no closing counterpart
+    _re_lt_html = re.compile(r"<(\S+)([^>]*>)(((?!</\1>).)*)<(/\1>)", re.DOTALL)
+    while _re_lt_html.search(text):
+        text = _re_lt_html.sub(r"LTHTML\1\2\3LTHTML\5", text)
+    text = re.sub(r"(^|[^<])<([^<]|$)", r"\1&amp;lt;\2", text)
+    text = text.replace("LTHTML", "<")
+    return text
+
+
+def parse_options(block_content):
+    """
+    Parses the option in some rst block content.
+    """
+    block_lines = block_content.split("\n")
+    block_indent = find_indent(block_lines[0])
+    current_option = None
+    result = {}
+    for line in block_lines:
+        if _re_rst_option.search(line) is not None:
+            current_option, value = _re_rst_option.search(line).groups()
+            result[current_option] = value.lstrip()
+        elif find_indent(line) > block_indent:
+            result[current_option] += " " + line.lstrip()
+
+    return result
+
+
+def apply_min_indent(text, min_indent):
+    """
+    Make sure all lines in a text are have a minimum indentation.
+
+    Args:
+        text (`str`): The text to treat.
+        min_indent (`int`): The minimal indentation.
+
+    Returns:
+        `str`: The processed text.
+    """
+    lines = text.split("\n")
+    idx = 0
+    while idx < len(lines):
+        if is_empty_line(lines[idx]):
+            idx += 1
+            continue
+        indent = find_indent(lines[idx])
+        if indent < min_indent:
+            while idx < len(lines) and (
+                find_indent(lines[idx]) >= indent or is_empty_line(lines[idx])
+            ):
+                if not is_empty_line(lines[idx]):
+                    lines[idx] = " " * (min_indent - indent) + lines[idx]
+                idx += 1
+        else:
+            idx += 1
+
+    return "\n".join(lines)
+
+
+def convert_rst_blocks(text, page_info):
+    """
+    Converts rst special blocks (examples, notes) into MDX.
+    """
+    if "package_name" not in page_info:
+        raise ValueError("`page_info` must contain at least the package_name.")
+    package_name = page_info["package_name"]
+    version = page_info.get("version", "main")
+    language = page_info.get("language", "en")
+
+    lines = text.split("\n")
+    idx = 0
+    new_lines = []
+    while idx < len(lines):
+        block_type = None
+        block_info = None
+        if _re_block.search(lines[idx]) is not None:
+            block_type = _re_block.search(lines[idx]).groups()[0]
+            if _re_block_info.search(lines[idx]) is not None:
+                block_info = _re_block_info.search(lines[idx]).groups()[0]
+        elif _re_example.search(lines[idx]) is not None:
+            block_type = "code-block-example"
+            block_info = "python"
+            example_name = _re_example.search(lines[idx]).groups()[0]
+            new_lines.append(f"<exampletitle>{example_name}:</exampletitle>\n")
+        elif lines[idx].strip() == "..":
+            block_type = "comment"
+        elif lines[idx].strip() == "::":
+            block_type = "code-block"
+
+        if block_type is not None:
+            block_indent = find_indent(lines[idx])
+            # Find the next nonempty line
+            idx += 1
+            while idx < len(lines) and is_empty_line(lines[idx]):
+                idx += 1
+            # Grab the indent of the return line, this block will stop when we unindent under it (or has already)
+            example_indent = (
+                find_indent(lines[idx]) if idx < len(lines) else block_indent
+            )
+
+            if example_indent == block_indent:
+                block_content = ""
+            else:
+                block_lines = []
+                while idx < len(lines) and (
+                    is_empty_line(lines[idx])
+                    or find_indent(lines[idx]) >= example_indent
+                ):
+                    block_lines.append(lines[idx][example_indent:])
+                    idx += 1
+                block_content = "\n".join(block_lines)
+
+            if block_type in ["code", "code-block"]:
+                prefix = "```" if block_info is None else f"```{block_info}"
+                new_lines.append(f"{prefix}\n{block_content.strip()}\n```\n")
+            elif block_type == "code-block-example":
+                prefix = f"<example>```{block_info}"
+                new_lines.append(f"{prefix}\n{block_content.strip()}\n```\n</example>")
+            elif block_type == "note":
+                new_lines.append(
+                    apply_min_indent(
+                        f"<Tip>\n\n{block_content.strip()}\n\n</Tip>\n", block_indent
+                    )
+                )
+            elif block_type == "warning":
+                new_lines.append(
+                    apply_min_indent(
+                        "<Tip warning={true}>\n\n"
+                        + f"{block_content.strip()}\n\n</Tip>\n",
+                        block_indent,
+                    )
+                )
+            elif block_type == "raw":
+                new_lines.append(block_content.strip() + "\n")
+            elif block_type == "math":
+                new_lines.append(f"$${block_content.strip()}$$\n")
+            elif block_type == "comment":
+                new_lines.append(f"<!--{block_content.strip()}\n-->\n")
+            elif block_type == "autofunction":
+                if block_info is not None:
+                    new_lines.append(f"[[autodoc]] {block_info}\n")
+            elif block_type == "autoclass":
+                if block_info is not None:
+                    block = f"[[autodoc]] {block_info}\n"
+                    options = parse_options(block_content)
+                    if "special-members" in options:
+                        special_members = options["special-members"].split(", ")
+                        for special_member in special_members:
+                            block += f"    - {special_member}\n"
+                    if "members" in options:
+                        members = options["members"]
+                        if len(members) == 0:
+                            block += "    - all\n"
+                        else:
+                            for member in members.split(", "):
+                                block += f"    - {member}\n"
+                    new_lines.append(block)
+            elif block_type == "image":
+                options = parse_options(block_content)
+                target = options.pop("target", None)
+                if block_info is not None:
+                    options["src"] = block_info
+                else:
+                    if target is None:
+                        raise ValueError("Image source not defined.")
+                    options["src"] = target
+                # Adapt path
+                options["src"] = options["src"].replace(
+                    "/imgs/", f"/docs/{package_name}/{version}/{language}/imgs/"
+                )
+                html_code = " ".join(
+                    [f'{key}="{value}"' for key, value in options.items()]
+                )
+                new_lines.append(f"<img {html_code}/>\n")
+
+            else:
+                new_lines.append(
+                    f"{block_type},{block_info}\n{block_content.rstrip()}\n"
+                )
+
+        else:
+            new_lines.append(lines[idx])
+            idx += 1
+
+    return "\n".join(new_lines)
+
+
+# Re pattern that catches rst args blocks of the form `Parameters:`.
+_re_args = re.compile(r"^\s*(Args?|Arguments?|Attributes?|Params?|Parameters?):\s*$")
+# Re pattern that catches return blocks of the form `Return:`.
+_re_returns = re.compile(r"^\s*(Return|Yield|Raise)s?:\s*$")
+
+
+def split_return_line(line):
+    """
+    Split the return line with format `type: some doc`. Type may contain colons in the form of :obj: or :class:.
+    """
+    splits_on_colon = line.split(":")
+    idx = 1
+    while idx < len(splits_on_colon) and splits_on_colon[idx] in ["obj", "class"]:
+        idx += 2
+    if idx >= len(splits_on_colon):
+        if len(splits_on_colon) % 2 == 1 and re.search(r"`\w+`$", line.rstrip()):
+            return line, ""
+        return None, line
+    return ":".join(splits_on_colon[:idx]), ":".join(splits_on_colon[idx:])
+
+
+def split_raise_line(line):
+    """
+    Split the raise line with format `SomeError some doc`.
+    """
+    splits_on_colon = line.strip().split(" ")
+    error_type, doc = splits_on_colon[0], " ".join(splits_on_colon[1:])
+    if error_type and error_type[-1] == ":":
+        error_type = error_type[:-1]
+    return error_type, doc
+
+
+def split_arg_line(line):
+    """
+    Split the return line with format `type: some doc`. Type may contain colons in the form of :obj: or :class:.
+    """
+    splits_on_colon = line.split(":")
+    idx = 1
+    while idx < len(splits_on_colon) and splits_on_colon[idx] in ["obj", "class"]:
+        idx += 2
+    if idx >= len(splits_on_colon):
+        return line, ""
+    return ":".join(splits_on_colon[:idx]), ":".join(splits_on_colon[idx:])
+
+
+class InvalidRstDocstringError(ValueError):
+    pass
+
+
+_re_parameters = re.compile(
+    r"<parameters>(((?!<parameters>).)*)</parameters>", re.DOTALL
+)
+_re_md_link = re.compile(r"\[(.+)\]\(.+\)", re.DOTALL)
+
+
+def parse_rst_docstring(docstring):
+    """
+    Parses a docstring written in rst, in particular the list of arguments and the return type.
+    """
+    lines = docstring.split("\n")
+    idx = 0
+    while idx < len(lines):
+        # Parameters section
+        if _re_args.search(lines[idx]) is not None:
+            # Title of the section.
+            lines[idx] = "<parameters>\n"
+            # Find the next nonempty line
+            idx += 1
+            while is_empty_line(lines[idx]):
+                idx += 1
+            # Grab the indent of the list of parameters, this block will stop when we unindent under it or we see the
+            # Returns or Raises block.
+            param_indent = find_indent(lines[idx])
+            while (
+                idx < len(lines)
+                and find_indent(lines[idx]) == param_indent
+                and _re_returns.search(lines[idx]) is None
+            ):
+                intro, doc = split_arg_line(lines[idx])
+                # Line starting with a > after indent indicate a "section title" in the parameters.
+                if intro.lstrip().startswith(">"):
+                    lines[idx] = intro.lstrip()
+                else:
+                    lines[idx] = (
+                        re.sub(r"^\s*(\S+)(\s)?", r"- **\1**\2", intro) + " --" + doc
+                    )
+                idx += 1
+                while idx < len(lines) and (
+                    is_empty_line(lines[idx]) or find_indent(lines[idx]) > param_indent
+                ):
+                    idx += 1
+            lines.insert(idx, "</parameters>\n")
+            idx += 1
+
+        # Returns section
+        elif _re_returns.search(lines[idx]) is not None:
+            # tag is either `return` or `yield`
+            tag = _re_returns.match(lines[idx]).group(1).lower()
+            # Title of the section.
+            lines[idx] = f"<{tag}s>\n"
+            # Find the next nonempty line
+            idx += 1
+            while is_empty_line(lines[idx]):
+                idx += 1
+
+            # Grab the indent of the return line, this block will stop when we unindent under it.
+            return_indent = find_indent(lines[idx])
+            raised_errors = []
+            # The line may contain the return type.
+            if tag in ["return", "yield"]:
+                return_type, return_description = split_return_line(lines[idx])
+                lines[idx] = return_description
+                idx += 1
+                while idx < len(lines) and (
+                    is_empty_line(lines[idx])
+                    or find_indent(lines[idx]) >= return_indent
+                ):
+                    idx += 1
+            else:
+                while idx < len(lines) and find_indent(lines[idx]) == return_indent:
+                    return_type, return_description = split_raise_line(lines[idx])
+                    raised_error = re.sub(r"^\s*`?([\w\.]*)`?$", r"``\1``", return_type)
+                    lines[idx] = "- " + raised_error + " -- " + return_description
+                    md_link = _re_md_link.match(raised_error)
+                    if md_link:
+                        raised_error = md_link[1]
+                        raised_error = re.sub(
+                            r"^\s*`?([\w\.]*)`?$", r"``\1``", raised_error
+                        )
+                    if raised_error not in raised_errors:
+                        raised_errors.append(raised_error)
+                    idx += 1
+                    while idx < len(lines) and (
+                        is_empty_line(lines[idx])
+                        or find_indent(lines[idx]) > return_indent
+                    ):
+                        idx += 1
+
+            lines.insert(idx, f"</{tag}s>\n")
+            idx += 1
+
+            # Return block finished, we insert the return type if one was specified
+            if tag in ["return", "yield"] and return_type is not None:
+                lines[idx - 1] += f"\n<{tag}type>{return_type}</{tag}type>\n"
+            elif len(raised_errors) > 0:
+                # raised errors
+                lines[
+                    idx - 1
+                ] += f"\n<raisederrors>{' or '.join(raised_errors)}</raisederrors>\n"
+
+        else:
+            idx += 1
+
+    result = "\n".join(lines)
+
+    # combine multiple <parameters> blocks into one block
+    if result.count("<parameters>") > 1:
+        parameters_blocks = _re_parameters.findall(result)
+        parameters_blocks = [pb[0].strip() for pb in parameters_blocks]
+        parameters_str = "\n".join(parameters_blocks)
+        result = _re_parameters.sub("", result)
+        result += f"\n<parameters>{parameters_str}</parameters>\n"
+
+    return result
+
+
+_re_list = re.compile(r"^\s*(-|\*|\d+\.)\s")
+_re_autodoc = re.compile(r"^\s*\[\[autodoc\]\]\s+(\S+)\s*$")
+
+
+def remove_indent(text):
+    """
+    Remove indents in text, except the one linked to lists (or sublists).
+    """
+    lines = text.split("\n")
+    # List of indents to remember for nested lists
+    current_indents = []
+    # List of new indents to remember for nested lists
+    new_indents = []
+    is_inside_code = False
+    code_indent = 0
+    for idx, line in enumerate(lines):
+        # Line is an item in a list.
+        if _re_list.search(line) is not None:
+            indent = find_indent(line)
+            # Is it a new list / new level of nestedness?
+            if len(current_indents) == 0 or indent > current_indents[-1]:
+                current_indents.append(indent)
+                new_indent = 0 if len(new_indents) == 0 else new_indents[-1]
+                lines[idx] = " " * new_indent + line[indent:]
+                new_indent += len(_re_list.search(line).groups()[0]) + 1
+                new_indents.append(new_indent)
+            # Otherwise it's an existing level of list (current one, or previous one)
+            else:
+                # Let's find the proper level of indentation
+                level = len(current_indents) - 1
+                while level >= 0 and current_indents[level] != indent:
+                    level -= 1
+                current_indents = current_indents[: level + 1]
+                new_indents = new_indents[:level]
+                new_indent = 0 if len(new_indents) == 0 else new_indents[-1]
+                lines[idx] = " " * new_indent + line[indent:]
+                new_indent += len(_re_list.search(line).groups()[0]) + 1
+                new_indents.append(new_indent)
+
+        # Line is an autodoc, we keep the indent for the list just after if there is one.
+        elif _re_autodoc.search(line) is not None:
+            indent = find_indent(line)
+            current_indents = [indent]
+            new_indents = [4]
+            lines[idx] = line.strip()
+
+        # Deal with empty lines separately
+        elif is_empty_line(line):
+            lines[idx] = ""
+
+        # Code blocks
+        elif line.lstrip().startswith("```"):
+            is_inside_code = not is_inside_code
+            if is_inside_code:
+                code_indent = find_indent(line)
+            lines[idx] = line[code_indent:]
+        elif is_inside_code:
+            lines[idx] = line[code_indent:]
+
+        else:
+            indent = find_indent(line)
+            if len(current_indents) > 0 and indent > current_indents[-1]:
+                lines[idx] = " " * new_indents[-1] + line[indent:]
+            elif len(current_indents) > 0:
+                # Let's find the proper level of indentation
+                level = len(current_indents) - 1
+                while level >= 0 and current_indents[level] > indent:
+                    level -= 1
+                current_indents = current_indents[: level + 1]
+                if level >= 0:
+                    if current_indents[level] < indent:
+                        new_indents = new_indents[: level + 1]
+                    else:
+                        new_indents = new_indents[:level]
+                    new_indent = 0 if len(new_indents) == 0 else new_indents[-1]
+                    lines[idx] = " " * new_indent + line[indent:]
+                    new_indents.append(new_indent)
+                else:
+                    new_indents = []
+                    lines[idx] = line[indent:]
+            else:
+                lines[idx] = line[indent:]
+
+    return "\n".join(lines)
+
+
+def base_rst_to_mdx(text, page_info, unindent=True):
+    """
+    Convert a text from rst to mdx, with the base operations necessary for both docstrings and rst docs.
+    """
+    text = convert_rst_links(text, page_info)
+    text = convert_special_chars(text)
+    text = convert_rst_blocks(text, page_info)
+    # Convert * in lists to - to avoid the formatting conversion treat them as bold.
+    text = re.sub(r"^(\s*)\*(\s)", r"\1-\2", text, flags=re.MULTILINE)
+    text = convert_rst_formatting(text)
+    return remove_indent(text) if unindent else text
+
+
+def convert_rst_docstring_to_mdx(docstring, page_info):
+    """
+    Convert a docstring written in rst to mdx.
+    """
+    text = parse_rst_docstring(docstring)
+    return base_rst_to_mdx(text, page_info)
+
+
+def process_titles(lines):
+    """Converts rst titles to markdown titles."""
+    title_chars = """= - ` : ' " ~ ^ _ * + # < >""".split(" ")
+    title_levels = {}
+    new_lines = []
+    for line in lines:
+        if (
+            len(new_lines) > 0
+            and len(line) >= len(new_lines[-1])
+            and len(set(line)) == 1
+            and line[0] in title_chars
+            and line != "::"
+        ):
+            char = line[0]
+            level = title_levels.get(char, len(title_levels) + 1)
+            if level not in title_levels:
+                title_levels[char] = level
+            new_lines[-1] = f"{'#' * level} {new_lines[-1]}"
+        else:
+            new_lines.append(line)
+    return new_lines
+
+
+# Matches lines with a pattern of a table new line in rst.
+_re_ignore_line_table = re.compile(r"^(\+[\-\s]+)+\+\s*$")
+# Matches lines with a pattern of a table new line in rst, with a first column empty.
+_re_ignore_line_table1 = re.compile(r"^\|\s+(\+[\-\s]+)+\+\s*$")
+# Matches lines with a pattern of a first table line in rst.
+_re_sep_line_table = re.compile(r"^(\+[=\s]+)+\+\s*$")
+# Re pattern that catches anchors of the type .. reference:
+_re_anchor_section = re.compile(r"^\.\.\s+_(\S+):")
+
+
+def split_pt_tf_code_blocks(text):
+    """
+    Split PyTorch and TensorFlow specific block codes.
+    """
+    lines = text.split("\n")
+    new_lines = []
+    idx = 0
+    while idx < len(lines):
+        if lines[idx].startswith("```"):
+            code_lines = {"common": [lines[idx]], "pytorch": [], "tensorflow": []}
+            is_pytorch = False
+            is_tensorflow = False
+            idx += 1
+            while idx < len(lines) and lines[idx].strip() != "```":
+                if "## PYTORCH CODE" in lines[idx]:
+                    is_pytorch = True
+                    is_tensorflow = False
+                elif "## TENSORFLOW CODE" in lines[idx]:
+                    is_tensorflow = True
+                    is_pytorch = False
+                elif is_pytorch:
+                    code_lines["pytorch"].append(lines[idx])
+                elif is_tensorflow:
+                    code_lines["tensorflow"].append(lines[idx])
+                else:
+                    code_lines["common"].append(lines[idx])
+                idx += 1
+            if len(code_lines["pytorch"]) > 0 or len(code_lines["tensorflow"]) > 0:
+                block_lines = ["<frameworkcontent>", "<pt>"]
+                block_lines.extend(code_lines["common"].copy() + code_lines["pytorch"])
+                block_lines.extend(["```", "</pt>", "<tf>"])
+                block_lines.extend(
+                    code_lines["common"].copy() + code_lines["tensorflow"]
+                )
+                block_lines.extend(["```", "</tf>", "</frameworkcontent>"])
+                new_lines.extend(block_lines)
+            else:
+                block_lines = code_lines["common"] + ["```"]
+                new_lines.extend(block_lines)
+            idx += 1
+        else:
+            new_lines.append(lines[idx])
+            idx += 1
+    return "\n".join(new_lines)
+
+
+def convert_rst_to_mdx(rst_text, page_info, add_imports=True):
+    """
+    Convert a document written in rst to mdx.
+    """
+    lines = rst_text.split("\n")
+    lines = process_titles(lines)
+    if add_imports:
+        new_lines = [
+            '<script lang="ts">',
+            '	import Tip from "$lib/Tip.svelte";',
+            '	import Youtube from "$lib/Youtube.svelte";',
+            '	import Docstring from "$lib/Docstring.svelte";',
+            '	import CodeBlock from "$lib/CodeBlock.svelte";',
+            '	import CodeBlockFw from "$lib/CodeBlockFw.svelte";',
+            '	import DocNotebookDropdown from "$lib/DocNotebookDropdown.svelte";',
+            '	import CourseFloatingBanner from "$lib/CourseFloatingBanner.svelte";',
+            '	import IconCopyLink from "$lib/IconCopyLink.svelte";',
+            '	import FrameworkContent from "$lib/FrameworkContent.svelte";',
+            '	import Markdown from "$lib/Markdown.svelte";',
+            '	import ExampleCodeBlock from "$lib/ExampleCodeBlock.svelte";',
+            '	import Added from "$lib/Added.svelte";',
+            '	import Changed from "$lib/Changed.svelte";',
+            '	import Deprecated from "$lib/Deprecated.svelte";',
+            '	import PipelineIcon from "$lib/PipelineIcon.svelte";',
+            '	import PipelineTag from "$lib/PipelineTag.svelte";',
+            "	",
+            '	export let fw: "pt" | "tf"',
+            "</script>",
+            "<svelte:head>",
+            '<meta name="hf:doc:metadata" content={JSON.stringify(metadata)} >',
+            "</svelte:head>",
+            "",
+        ]
+    else:
+        new_lines = []
+    for line in lines:
+        if _re_ignore_line_table.search(line) is not None:
+            continue
+        elif _re_ignore_line_table1.search(line) is not None:
+            continue
+        elif _re_sep_line_table.search(line) is not None:
+            line = line.replace("=", "-").replace("+", "|")
+        elif _re_anchor_section.search(line) is not None:
+            anchor_name = _re_anchor_section.search(line).groups()[0]
+            line = f"<a id='{anchor_name}'></a>"
+        new_lines.append(line)
+    text = "\n".join(new_lines)
+
+    return split_pt_tf_code_blocks(base_rst_to_mdx(text, page_info))

src/kernels/cli.py

@@ -8,6 +8,9 @@ from kernels.compat import tomllib
 from kernels.lockfile import KernelLock, get_kernel_locks
 from kernels.utils import install_kernel, install_kernel_all_variants
 
+from .doc import generate_readme_for_kernel
+from .wheel import build_variant_to_wheel
+
 
 def main():
     parser = argparse.ArgumentParser(
@@ -36,6 +39,47 @@ def main():
     )
     lock_parser.set_defaults(func=lock_kernels)
 
+    to_wheel_parser = subparsers.add_parser(
+        "to-wheel", help="Convert a kernel to a wheel file"
+    )
+    to_wheel_parser.add_argument("repo_id", type=str, help="The kernel repo ID")
+    to_wheel_parser.add_argument("version", type=str, help="The kernel version")
+    to_wheel_parser.add_argument(
+        "--python-version",
+        type=str,
+        default="3.9",
+        help="The minimum Python version. Must match the Python version that the kernel was compiled for.",
+    )
+    to_wheel_parser.add_argument(
+        "--manylinux-version",
+        type=str,
+        default="2.28",
+        help="The manylinux version. Must match the manylinux version that the kernel was compiled for.",
+    )
+    to_wheel_parser.set_defaults(func=kernels_to_wheel)
+
+    # Add generate-readme subcommand parser
+    generate_readme_parser = subparsers.add_parser(
+        "generate-readme",
+        help="Generate README snippets for a kernel's public functions",
+    )
+    generate_readme_parser.add_argument(
+        "repo_id",
+        type=str,
+        help="The kernel repo ID (e.g., kernels-community/activation)",
+    )
+    generate_readme_parser.add_argument(
+        "--revision",
+        type=str,
+        default="main",
+        help="The kernel revision (branch, tag, or commit SHA, defaults to 'main')",
+    )
+    generate_readme_parser.set_defaults(
+        func=lambda args: generate_readme_for_kernel(
+            repo_id=args.repo_id, revision=args.revision
+        )
+    )
+
     args = parser.parse_args()
     args.func(args)
 
@@ -77,6 +121,24 @@ def download_kernels(args):
         sys.exit(1)
 
 
+def kernels_to_wheel(args):
+    variants_path = install_kernel_all_variants(
+        repo_id=args.repo_id, revision=f"v{args.version}"
+    )
+    for variant_path in variants_path.iterdir():
+        if not variant_path.is_dir():
+            continue
+        wheel_path = build_variant_to_wheel(
+            manylinux_version=args.manylinux_version,
+            python_version=args.python_version,
+            repo_id=args.repo_id,
+            version=args.version,
+            variant_path=variant_path,
+            wheel_dir=Path("."),
+        )
+        print(f"☸️ {wheel_path.name}", file=sys.stderr)
+
+
 def lock_kernels(args):
     with open(args.project_dir / "pyproject.toml", "rb") as f:
         data = tomllib.load(f)

src/kernels/doc.py

@@ -0,0 +1,242 @@
+import inspect
+import re
+import sys
+from types import ModuleType
+
+import yaml
+
+from ._vendored.convert_rst_to_mdx import convert_rst_docstring_to_mdx
+from .utils import get_kernel
+
+_RE_PARAMETERS = re.compile(
+    r"<parameters>(((?!<parameters>).)*)</parameters>", re.DOTALL
+)
+_RE_RETURNS = re.compile(r"<returns>(((?!<returns>).)*)</returns>", re.DOTALL)
+_RE_RETURNTYPE = re.compile(
+    r"<returntype>(((?!<returntype>).)*)</returntype>", re.DOTALL
+)
+
+
+def _extract_description_before_tags(docstring_mdx: str) -> str:
+    """Extract the description part of a docstring before any tags."""
+    params_pos = docstring_mdx.find("<parameters>")
+    returns_pos = docstring_mdx.find("<returns>")
+    returntype_pos = docstring_mdx.find("<returntype>")
+    positions = [pos for pos in [params_pos, returns_pos, returntype_pos] if pos != -1]
+
+    if positions:
+        first_tag_pos = min(positions)
+        return docstring_mdx[:first_tag_pos].strip()
+    else:
+        return docstring_mdx.strip()
+
+
+def _print_parameters_section(docstring_mdx: str, *, header_level: int) -> None:
+    """Print the parameters section from a docstring."""
+    matches = _RE_PARAMETERS.findall(docstring_mdx)
+    if matches:
+        header = "#" * header_level
+        print(f"\n{header} Parameters")
+        for match in matches:
+            print(f"\n{match[0].strip()}")
+
+
+def _print_returns_section(
+    docstring_mdx: str, *, context_name: str, header_level: int
+) -> None:
+    """Print the returns section from a docstring."""
+    return_matches = _RE_RETURNS.findall(docstring_mdx)
+    returntype_matches = _RE_RETURNTYPE.findall(docstring_mdx)
+
+    if return_matches or returntype_matches:
+        header = "#" * header_level
+        print(f"\n{header} Returns")
+
+        if returntype_matches:
+            if len(returntype_matches) > 1:
+                raise ValueError(
+                    f"More than one <returntype> tag found in docstring for {context_name}"
+                )
+            print(f"\n**Type**: {returntype_matches[0][0].strip()}")
+
+        if return_matches:
+            for match in return_matches:
+                print(f"\n{match[0].strip()}")
+
+
+def _get_docstring(obj, use_dict_check: bool = False) -> str:
+    """Get docstring from an object, with fallback to default message."""
+    # Check whether the class/method itself has docs and not just
+    # the superclass.
+    if use_dict_check:
+        has_doc = obj.__dict__.get("__doc__", None) is not None
+    else:
+        has_doc = getattr(obj, "__doc__", None) is not None
+
+    # We use inspect.getdoc because it does normalization.
+    doc = inspect.getdoc(obj)
+
+    return doc if has_doc and doc is not None else "No documentation available."
+
+
+def _process_and_print_docstring(
+    docstring: str, *, kernel_name: str, context_name: str, header_level: int
+) -> None:
+    """Convert docstring to MDX and print description, parameters, and returns sections."""
+    docstring_mdx = convert_rst_docstring_to_mdx(
+        docstring, page_info={"package_name": kernel_name}
+    )
+
+    # Print the description
+    description = _extract_description_before_tags(docstring_mdx)
+    print(f"\n{description}")
+
+    # Print parameters and returns sections
+    _print_parameters_section(docstring_mdx, header_level=header_level)
+    _print_returns_section(
+        docstring_mdx, context_name=context_name, header_level=header_level
+    )
+
+
+def generate_readme_for_kernel(repo_id: str, *, revision: str = "main") -> None:
+    kernel_module = get_kernel(repo_id=repo_id, revision=revision)
+    kernel_name = repo_id.split("/")[-1].replace("-", "_")
+
+    generate_metadata(kernel_module)
+    generate_kernel_doc(kernel_module, kernel_name)
+    generate_function_doc(kernel_module, kernel_name)
+    generate_layers_doc(kernel_module, kernel_name)
+
+
+def generate_metadata(module: ModuleType) -> None:
+    metadata = getattr(module, "__kernel_metadata__", {})
+    if "tags" not in metadata:
+        metadata["tags"] = ["kernel"]
+    else:
+        if "kernel" not in metadata["tags"]:
+            metadata["tags"].append("kernel")
+
+    print("---")
+    print(yaml.dump(metadata), end="")
+    print("---")
+
+
+def generate_kernel_doc(module: ModuleType, kernel_name: str) -> None:
+    docstring = module.__doc__.strip() if module.__doc__ is not None else None
+    if docstring:
+        title, rest = docstring.split("\n", 1)
+        print(f"# {title.strip()}")
+        print(
+            f"\n{convert_rst_docstring_to_mdx(rest.strip(), page_info={'package_name': kernel_name})}"
+        )
+
+
+def generate_function_doc(kernel_module: ModuleType, kernel_name: str) -> None:
+    print("\n## Functions")
+
+    # Track if we found any functions
+    found_functions = False
+
+    for name, func in inspect.getmembers(kernel_module, inspect.isfunction):
+        # Do not include imported functions.
+        if func.__module__ != kernel_module.__name__:
+            continue
+
+        # Exclude private functions.
+        if name.startswith("_"):
+            continue
+
+        found_functions = True
+
+        try:
+            sig = inspect.signature(func)
+            docstring = _get_docstring(func)
+        except ValueError:
+            print(
+                f"Warning: Could not retrieve signature for {name} in {kernel_module.__name__}",
+                file=sys.stderr,
+            )
+            continue
+
+        print(f"\n### Function `{name}`")
+        print(f"\n`{sig}`")
+
+        _process_and_print_docstring(
+            docstring, kernel_name=kernel_name, context_name=name, header_level=3
+        )
+
+    if not found_functions:
+        print("\nNo public top-level functions.")
+
+
+def generate_layers_doc(kernel_module: ModuleType, kernel_name: str) -> None:
+    # Check if layers module is available
+    layers_module = getattr(kernel_module, "layers", None)
+    if layers_module is None:
+        return
+
+    print("\n## Layers")
+
+    # Track if we found any classes
+    found_classes = False
+
+    for class_name, cls in inspect.getmembers(layers_module, inspect.isclass):
+        # Exclude classes that were imported.
+        if cls.__module__ != layers_module.__name__:
+            continue
+
+        found_classes = True
+
+        try:
+            # Get docstring, but not from superclasses.
+            class_docstring = _get_docstring(cls, use_dict_check=True)
+        except Exception:
+            print(
+                f"Warning: Could not retrieve documentation for class {class_name} in {layers_module.__name__}",
+                file=sys.stderr,
+            )
+            continue
+
+        print(f"\n### Class `{class_name}`")
+
+        # Always print class description (helper handles conversion and formatting)
+        class_docstring_mdx = convert_rst_docstring_to_mdx(
+            class_docstring, page_info={"package_name": kernel_name}
+        )
+        description = _extract_description_before_tags(class_docstring_mdx)
+        print(f"\n{description}")
+
+        # Document methods
+        print("\n#### Methods")
+
+        for method_name, method in inspect.getmembers(cls, inspect.isfunction):
+            # Note: also skip __init__, since extension layers cannot have a constructor.
+            if method_name.startswith("_"):
+                continue
+
+            # Skip methods from superclasses.
+            if method_name not in cls.__dict__:
+                continue
+
+            try:
+                sig = inspect.signature(method)
+                method_docstring = _get_docstring(method)
+            except ValueError:
+                print(
+                    f"Warning: Could not retrieve signature for {method_name} in {class_name}",
+                    file=sys.stderr,
+                )
+                continue
+
+            print(f"\n##### Method `{method_name}`")
+            print(f"\n`{sig}`")
+
+            _process_and_print_docstring(
+                method_docstring,
+                kernel_name=kernel_name,
+                context_name=method_name,
+                header_level=6,
+            )
+
+    if not found_classes:
+        print("\nNo layers defined.")

src/kernels/layer.py

@@ -1,19 +1,67 @@
+from __future__ import annotations
+
 import inspect
 import os
 import warnings
 from contextvars import ContextVar
 from copy import deepcopy
 from dataclasses import dataclass, field
-from typing import TYPE_CHECKING, Dict, Union
+from enum import Flag, auto
+from types import MethodType
+from typing import (
+    TYPE_CHECKING,
+    Dict,
+    Optional,
+    Tuple,
+    Type,
+    Union,
+)
 
 from .utils import get_kernel
 
 if TYPE_CHECKING:
+    import torch
     from torch import nn
 
+
 _DISABLE_KERNEL_MAPPING: bool = bool(int(os.environ.get("DISABLE_KERNEL_MAPPING", "0")))
 
 
+class Mode(Flag):
+    """
+    Kernelize mode
+
+    The `Mode` flag is used by `kernelize` to select kernels for the given
+    mode. Mappings can be registered for specific modes.
+
+    * `INFERENCE`: The kernel is used for inference.
+    * `TRAINING`: The kernel is used for training.
+    * `TORCH_COMPILE`: The kernel is used with `torch.compile`.
+    * `DEFAULT`: In a kernel mapping, this kernel is used when no other mode
+       matches.
+
+    Different modes can be combined. For instance, `INFERENCE | TORCH_COMPILE`
+    should be used for layers that are used for inference *with* `torch.compile`.
+    """
+
+    _NONE = 0
+    DEFAULT = auto()
+    TRAINING = auto()
+    INFERENCE = auto()
+    TORCH_COMPILE = auto()
+
+    def __or__(self, other: Mode) -> Mode:
+        union = super().__or__(other)
+
+        if Mode.INFERENCE in union and Mode.TRAINING in union:
+            raise ValueError("Mode.INFERENCE and Mode.TRAINING are mutually exclusive.")
+
+        if Mode.DEFAULT in union and union != Mode.DEFAULT:
+            raise ValueError("Mode.DEFAULT cannot be combined with other modes.")
+
+        return union
+
+
 @dataclass(frozen=True)
 class Device:
     type: str
@@ -53,13 +101,19 @@ class LayerRepository:
         return hash((self.layer_name, self.repo_id, self.revision))
 
 
-_KERNEL_MAPPING: ContextVar[Dict[str, Dict[Device, LayerRepository]]] = ContextVar(
-    "_KERNEL_MAPPING", default={}
+_CACHED_LAYER: Dict[LayerRepository, Type["nn.Module"]] = {}
+
+
+_KERNEL_MAPPING: ContextVar[Dict[str, Dict[Device, Dict[Mode, LayerRepository]]]] = (
+    ContextVar("_KERNEL_MAPPING", default={})
 )
 
 
 def use_kernel_mapping(
-    mapping: Dict[str, Dict[Union[Device, str], LayerRepository]],
+    mapping: Dict[
+        str,
+        Dict[Union[Device, str], Union[LayerRepository, Dict[Mode, LayerRepository]]],
+    ],
     *,
     inherit_mapping: bool = True,
 ):
@@ -87,12 +141,17 @@ def use_kernel_mapping(
 
 
 def register_kernel_mapping(
-    mapping: Dict[str, Dict[Union[Device, str], LayerRepository]]
+    mapping: Dict[
+        str,
+        Dict[Union[Device, str], Union[LayerRepository, Dict[Mode, LayerRepository]]],
+    ],
 ):
     """
-    Allows one to register a mapping between a layer name the corresponding kernel to use, depending on the device.
-    This should be use in conjunction with `use_kernel_hub_forward` decorator on the classname.
-    Exemple usage:
+    Allows one to register a mapping between a layer name and the corresponding
+    kernel(s) to use, depending on the device. This should be used in conjunction
+    with `kernelize`.
+
+    Example usage:
 
     ```python
     from kernels import LayerRepository, register_kernel_mapping
@@ -113,32 +172,106 @@ def register_kernel_mapping(
     for new_kernel, new_device_repos in mapping.items():
         device_repo = _KERNEL_MAPPING.get().setdefault(new_kernel, {})
         for new_device, new_repo in new_device_repos.items():
-            if isinstance(new_device, str):
-                device_repo[Device(type=new_device)] = new_repo
+            device = (
+                Device(type=new_device) if isinstance(new_device, str) else new_device
+            )
+
+            if isinstance(new_repo, LayerRepository):
+                kernel_options = {Mode.DEFAULT: new_repo}
             else:
-                device_repo[new_device] = new_repo
+                kernel_options = new_repo
+
+            device_repo[device] = kernel_options
 
 
-def replace_kernel_forward_from_hub(cls, layer_name: str, *, use_fallback: bool = True):
+def replace_kernel_forward_from_hub(
+    cls,
+    layer_name: str,
+):
     """
-    Replace the forward function of a layer using a layer from the kernel hub.
-    This function monkeypatches a layer, replacing the `forward` method
-    of the layer with that of a layer from the hub. The replacement is done
-    when a layer matching `layer_name` and device type is registered through
-    `register_layer_mapping`. The device type is inferred from the first
-    argument to `forward`.
+    Decorator that prepares a layer class to use a kernel from the Hugging Face Hub.
+
+    This decorator stores the layer name and original forward method, which will be used
+    by the kernelize function to replace the forward implementation with the appropriate
+    kernel from the hub.
+
+    Args:
+        cls: The layer class to decorate
+        layer_name: The name of the layer to use for kernel lookup
     """
+    cls.kernel_layer_name = layer_name
 
-    fallback_forward = cls.forward
 
-    cached_layer: Dict[LayerRepository, nn.Module] = {}
+def _select_repository(
+    repositories: Dict[Mode, LayerRepository],
+    *,
+    mode: Mode,
+) -> Optional[Tuple[LayerRepository, Mode]]:
+    if mode in repositories:
+        return (repositories[mode], mode)
+    elif Mode.DEFAULT in repositories:
+        return (repositories[Mode.DEFAULT], Mode.DEFAULT)
+    else:
+        return None
+
+
+def kernelize(
+    model: "nn.Module",
+    *,
+    mode: Mode,
+    device: Optional[Union[str, "torch.device"]] = None,
+    use_fallback: bool = True,
+):
+    """
+    Iterate over all modules in the model and replace the `forward` method of
+    extensible layers for which kernels are registered using `register_kernel_mapping`
+    or `use_kernel_mapping`.
+
+    Args:
+        model: The PyTorch model to kernelize
+        mode: the mode that the kernel is going to be used in (e.g.
+            `Mode.TRAINING | Mode.TORCH_COMPILE` kernelizes the model for training
+            and `torch.compile`).
+        device: The device type to load kernels for. The device type will be inferred
+            from the parameters of the model when not provided.
+        use_fallback: Whether to use the original forward method of modules when no
+            compatible kernel could be found. If set to `False`, an exception will
+            be raised in such cases.
+
+    Returns:
+        The kernelized model
+    """
+    import torch
+
+    if mode == Mode.DEFAULT:
+        raise ValueError("Mode.DEFAULT can only be used to register kernel mappings.")
+
+    # Type check ignored because this causes a false negative on Python < 3.11.
+    # Looks similar to: https://github.com/python/mypy/issues/9642
+    # Remove once we start doing typing checks on >= 3.11.
+    if Mode.INFERENCE not in mode and Mode.TRAINING not in mode:  # type: ignore[operator]
+        raise ValueError("kernelize mode must contain Mode.INFERENCE or Mode.TRAINING.")
+
+    if device is None:
+        device_type = _find_device(model)
+    elif isinstance(device, str):
+        device_type = Device(type=torch.device(device).type)
+    else:
+        device_type = Device(device.type)
+    assert isinstance(device_type, Device)
+
+    for _, module in model.named_modules():
+        module_class = type(module)
+        if not hasattr(module_class, "kernel_layer_name"):
+            continue
+        layer_name = module_class.kernel_layer_name
 
-    def forward(self, x, *args, **kwargs):
         if _DISABLE_KERNEL_MAPPING:
-            return fallback_forward(self, x, *args, **kwargs)
+            _replace_forward(module, module_class)
+            continue
+
+        kernel = _KERNEL_MAPPING.get().get(str(layer_name))
 
-        needs_backward = self.training
-        kernel = _KERNEL_MAPPING.get().get(layer_name)
         if kernel is None:
             warnings.warn(
                 "\n"
@@ -148,68 +281,70 @@ def replace_kernel_forward_from_hub(cls, layer_name: str, *, use_fallback: bool
             )
             if not use_fallback:
                 raise ValueError(f"No layer mapping for `{layer_name}`")
-            return fallback_forward(self, x, *args, **kwargs)
+            _replace_forward(module, module_class)
+            continue
 
-        device = getattr(x, "device", None)
-        if device is None:
-            return fallback_forward(self, x, *args, **kwargs)
+        # Get kernel options for the device
+        repos = kernel.get(device_type)
 
-        repo = kernel.get(Device(type=device.type))
-        if repo is None:
+        if repos is None:
             if not use_fallback:
                 raise ValueError(
-                    f"No layer mapping for `{layer_name}` with device type `{device.type}`"
+                    f"No layer mapping for `{layer_name}` with device type `{device_type}`"
                 )
-            return fallback_forward(self, x, *args, **kwargs)
+            _replace_forward(module, module_class)
+            continue
 
-        # Short-circuit if we already loaded the layer.
-        layer = cached_layer.get(repo, None)
-        if layer is not None:
-            if needs_backward and not getattr(layer, "has_backward", True):
-                return fallback_forward(self, x, *args, **kwargs)
-            return layer.forward(self, x, *args, **kwargs)
-
-        layer = _get_kernel_layer(
-            repo_id=repo.repo_id,
-            layer_name=repo.layer_name,
-            revision=repo.revision,
+        repo_with_mode = _select_repository(
+            repos,
+            mode=mode,
         )
 
-        # We have to validate against the original signature.
-        orig_forward = cls.forward
-        try:
-            cls.forward = fallback_forward
-            _validate_layer(check_cls=cls, cls=layer)
-        finally:
-            cls.forward = orig_forward
+        if repo_with_mode is None:
+            if not use_fallback:
+                raise ValueError(
+                    f"No repository for `{layer_name}` for configuration mode={mode}"
+                )
+            _replace_forward(module, module_class)
+            continue
 
-        cached_layer[repo] = layer
+        repo, repo_mode = repo_with_mode
 
-        if needs_backward and not getattr(layer, "has_backward", True):
-            return fallback_forward(self, x, *args, **kwargs)
-        return layer.forward(self, x, *args, **kwargs)
+        layer = _get_layer_memoize(repo, module_class)
 
-    cls.forward = forward
+        # Ideally we would do validation on the mapping where we check that
+        # e.g. if a repo class is registered for TRAINING | TORCH_COMPILE,
+        # the actual layer is compatible with that. Unfortunately, this would
+        # mean that we have to pre-download everything.
+        _validate_layer_has_mode(
+            layer_name=layer_name, module=layer, repo=repo, repo_mode=repo_mode
+        )
+
+        _conditionally_replace_forward(
+            module=module,
+            layer=layer,
+            mode=mode,
+            use_fallback=use_fallback,
+        )
+
+    return model
 
 
-def use_kernel_forward_from_hub(layer_name: str, *, use_fallback: bool = True):
+def use_kernel_forward_from_hub(layer_name: str):
     """
-    Replace the forward function of a layer using a layer from the kernel hub.
-    This decorator can be applied to a layer and replaces the forward method
-    of the layer with that of a layer from the hub. The replacement is done
-    when a layer matching `layer_name` and device type is registered through
-    `register_layer_mapping`. The device type is inferred from the first
-    argument to `forward`.
+    Make a layer extensible using the name `layer_name`.
     """
 
     def decorator(cls):
-        replace_kernel_forward_from_hub(cls, layer_name, use_fallback=use_fallback)
+        replace_kernel_forward_from_hub(cls, layer_name)
         return cls
 
     return decorator
 
 
-def _get_kernel_layer(*, repo_id: str, layer_name: str, revision: str) -> "nn.Module":
+def _get_kernel_layer(
+    *, repo_id: str, layer_name: str, revision: str
+) -> Type["nn.Module"]:
     """Get a layer from a kernel."""
 
     kernel = get_kernel(repo_id, revision=revision)
@@ -226,13 +361,13 @@ def _get_kernel_layer(*, repo_id: str, layer_name: str, revision: str) -> "nn.Mo
 
 
 def _validate_layer(*, check_cls, cls):
+    import torch.nn as nn
+
     # The layer must have at least have the following properties: (1) it
     # must be stateless; (2) the forward signature should correspond to
     # the signature it is replacing; (3) forward should not call other
     # methods.
 
-    from torch import nn
-
     if not issubclass(cls, nn.Module):
         raise TypeError(f"Layer `{cls}` is not a Torch layer.")
 
@@ -245,7 +380,8 @@ def _validate_layer(*, check_cls, cls):
     torch_module_members = {name for name, _ in inspect.getmembers(nn.Module)}
     cls_members = {name for name, _ in inspect.getmembers(cls)}
     difference = cls_members - torch_module_members
-    if difference != set() and difference != {"has_backward"}:
+    # verify if : difference ⊄ {"can_torch_compile", "has_backward"}
+    if not difference <= {"can_torch_compile", "has_backward"}:
         raise TypeError("Layer must not contain additional members.")
 
     # Check whether the forward signatures are similar.
@@ -262,3 +398,92 @@ def _validate_layer(*, check_cls, cls):
             raise TypeError(
                 f"Forward signature does not match: different kind of arguments ({param} ({param.kind}) and {ref_param} ({ref_param.kind})"
             )
+
+
+def _find_device(model: "nn.Module") -> Device:
+    try:
+        param = next(model.parameters())
+    except StopIteration:
+        raise ValueError(
+            "Cannot determine model device, provide as `device` argument to `kernelize`."
+        )
+
+    return Device(type=param.device.type)
+
+
+def _conditionally_replace_forward(
+    *,
+    module: "nn.Module",
+    layer: Type["nn.Module"],
+    mode: Mode,
+    use_fallback: bool,
+):
+    module_class = type(module)
+
+    # Switch to fallback if the mode is not supported by the layer.
+    # Note that this is useful even after _validate_layer_has_mode because
+    # layers registered with the DEFAULT mode never get rejected by
+    # _validate_layer_has_mode. For such layers, we want to fall back in
+    # case the layer does not support the given mode.
+    needs_fallback = Mode.TORCH_COMPILE in mode and not getattr(
+        layer, "can_torch_compile", False
+    )
+    needs_fallback |= Mode.TRAINING in mode and not getattr(layer, "has_backward", True)
+
+    if needs_fallback:
+        if use_fallback:
+            _replace_forward(module, module_class)
+        else:
+            raise ValueError(f"Available kernel does not support mode: {mode}")
+    else:
+        _replace_forward(module, layer)
+
+
+def _replace_forward(module: "nn.Module", layer: Type["nn.Module"]):
+    module.forward = MethodType(layer.forward, module)  # type: ignore[method-assign]
+
+
+def _validate_layer_has_mode(
+    *,
+    layer_name: str,
+    module: Type["nn.Module"],
+    repo: LayerRepository,
+    repo_mode: Mode,
+):
+    """
+    Check that a repository supports the mode that it was registered for.
+    """
+
+    if Mode.TRAINING in repo_mode and not getattr(module, "has_backward", True):
+        raise ValueError(
+            f"Layer `{repo.layer_name}` ({repo.repo_id}, revision: {repo.revision}) does not support backward.\n"
+            f"Was registered for `{layer_name}` with mode `{repo_mode}`"
+        )
+
+    if Mode.TORCH_COMPILE in repo_mode and not getattr(
+        module, "can_torch_compile", False
+    ):
+        raise ValueError(
+            f"Layer `{repo.layer_name}` ({repo.repo_id}, revision: {repo.revision}) does not support torch.compile.\n"
+            f"Was registered for `{layer_name}` with mode `{repo_mode}`"
+        )
+
+    return True
+
+
+def _get_layer_memoize(
+    repo: LayerRepository, module_class: Type["nn.Module"]
+) -> Type["nn.Module"]:
+    layer = _CACHED_LAYER.get(repo, None)
+    if layer is not None:
+        return layer
+
+    layer = _get_kernel_layer(
+        repo_id=repo.repo_id,
+        layer_name=repo.layer_name,
+        revision=repo.revision,
+    )
+    _validate_layer(check_cls=module_class, cls=layer)
+    _CACHED_LAYER[repo] = layer
+
+    return layer

src/kernels/utils.py

@@ -43,14 +43,23 @@ def build_variant() -> str:
     elif torch.version.hip is not None:
         rocm_version = parse(torch.version.hip.split("-")[0])
         compute_framework = f"rocm{rocm_version.major}{rocm_version.minor}"
+    elif torch.backends.mps.is_available():
+        compute_framework = "metal"
     else:
-        raise AssertionError("Torch was not compiled with CUDA or ROCm enabled.")
+        raise AssertionError(
+            "Torch was not compiled with CUDA, Metal, or ROCm enabled."
+        )
 
     torch_version = parse(torch.__version__)
-    cxxabi = "cxx11" if torch.compiled_with_cxx11_abi() else "cxx98"
     cpu = platform.machine()
     os = platform.system().lower()
 
+    if os == "darwin":
+        cpu = "aarch64" if cpu == "arm64" else cpu
+        return f"torch{torch_version.major}{torch_version.minor}-{compute_framework}-{cpu}-{os}"
+
+    cxxabi = "cxx11" if torch.compiled_with_cxx11_abi() else "cxx98"
+
     return f"torch{torch_version.major}{torch_version.minor}-{cxxabi}-{compute_framework}-{cpu}-{os}"
 
 
@@ -101,6 +110,23 @@ def install_kernel(
         )
     )
 
+    try:
+        return _load_kernel_from_path(repo_path, package_name, variant_locks)
+    except FileNotFoundError:
+        # Redo with more specific error message.
+        raise FileNotFoundError(
+            f"Kernel `{repo_id}` at revision {revision} does not have build: {variant}"
+        )
+
+
+def _load_kernel_from_path(
+    repo_path: Path,
+    package_name: str,
+    variant_locks: Optional[Dict[str, VariantLock]] = None,
+) -> Tuple[str, Path]:
+    variant = build_variant()
+    universal_variant = universal_build_variant()
+
     variant_path = repo_path / "build" / variant
     universal_variant_path = repo_path / "build" / universal_variant
 
@@ -119,7 +145,7 @@ def install_kernel(
 
     if not os.path.exists(module_init_path):
         raise FileNotFoundError(
-            f"Kernel `{repo_id}` at revision {revision} does not have build: {variant}"
+            f"Kernel at path `{repo_path}` does not have build: {variant}"
         )
 
     return package_name, variant_path
@@ -157,10 +183,24 @@ def install_kernel_all_variants(
 
 
 def get_kernel(repo_id: str, revision: str = "main") -> ModuleType:
+    """
+    Download and import a kernel from the Hugging Face Hub.
+
+    The kernel is downloaded from the repository `repo_id` at
+    branch/commit/tag `revision`.
+    """
     package_name, package_path = install_kernel(repo_id, revision=revision)
     return import_from_path(package_name, package_path / package_name / "__init__.py")
 
 
+def get_local_kernel(repo_path: Path, package_name: str) -> ModuleType:
+    """
+    Import a kernel from a local kernel repository path.
+    """
+    package_name, package_path = _load_kernel_from_path(repo_path, package_name)
+    return import_from_path(package_name, package_path / package_name / "__init__.py")
+
+
 def has_kernel(repo_id: str, revision: str = "main") -> bool:
     """
     Check whether a kernel build exists for the current environment

src/kernels/wheel.py

@@ -0,0 +1,186 @@
+import email.policy
+import os
+from dataclasses import dataclass
+from email.message import Message
+from importlib.metadata import PackageNotFoundError, version
+from pathlib import Path
+from typing import Optional
+
+try:
+    KERNELS_VERSION = version("kernels")
+except PackageNotFoundError:
+    KERNELS_VERSION = "unknown"
+
+
+@dataclass
+class Metadata:
+    name: str
+    version: str
+    cuda_version: Optional[str]
+    cxx_abi_version: Optional[str]
+    torch_version: Optional[str]
+    os: Optional[str]
+    platform: Optional[str]
+
+    @property
+    def is_universal(self) -> bool:
+        return self.platform is None
+
+
+def build_variant_to_wheel(
+    repo_id: str,
+    *,
+    version: str,
+    variant_path: Path,
+    wheel_dir: Path,
+    manylinux_version: str = "2.28",
+    python_version: str = "3.9",
+) -> Path:
+    """
+    Create a wheel file from the variant path.
+    """
+    name = repo_id.split("/")[-1].replace("_", "-")
+    metadata = extract_metadata(name, version, variant_path)
+    return build_wheel(
+        metadata,
+        variant_path=variant_path,
+        wheel_dir=wheel_dir,
+        manylinux_version=manylinux_version,
+        python_version=python_version,
+    )
+
+
+def extract_metadata(name: str, version: str, variant_path: Path) -> Metadata:
+    """
+    Extract metadata from the variant path.
+    """
+    if variant_path.name == "torch-universal":
+        return Metadata(
+            name=name,
+            version=version,
+            cuda_version=None,
+            cxx_abi_version=None,
+            torch_version=None,
+            os=None,
+            platform=None,
+        )
+
+    if not variant_path.name.startswith("torch"):
+        raise ValueError("Currently only conversion of Torch kernels is supported.")
+
+    variant_parts = variant_path.name.removeprefix("torch").split("-")
+    if len(variant_parts) != 5:
+        raise ValueError(f"Invalid variant name: {variant_path.name}")
+
+    torch_version = f"{variant_parts[0][:-1]}.{variant_parts[0][-1:]}"
+    cpp_abi_version = variant_parts[1].removeprefix("cxx")
+    cuda_version = variant_parts[2].removeprefix("cu")
+    platform = variant_parts[3].replace("-", "_")
+    os = variant_parts[4]
+
+    return Metadata(
+        name=name,
+        version=version,
+        cuda_version=cuda_version,
+        cxx_abi_version=cpp_abi_version,
+        torch_version=torch_version,
+        os=os,
+        platform=platform,
+    )
+
+
+def build_wheel(
+    metadata: Metadata,
+    *,
+    variant_path: Path,
+    wheel_dir: Path,
+    manylinux_version: str = "2.28",
+    python_version: str = "3.9",
+) -> Path:
+    """
+    Build the wheel file.
+    """
+    try:
+        from wheel.wheelfile import WheelFile  # type: ignore
+    except ImportError:
+        raise ImportError(
+            "The 'wheel' package is required to build wheels. Please install it with: `pip install wheel`"
+        )
+
+    name = metadata.name.replace("-", "_")
+    python_version_flat = python_version.replace(".", "")
+
+    if metadata.is_universal:
+        python_tag = f"py{python_version_flat}"
+        abi_tag = "none"
+        platform_tag = "any"
+        wheel_filename = (
+            f"{name}-{metadata.version}-{python_tag}-{abi_tag}-{platform_tag}.whl"
+        )
+        dist_info_dir_name = f"{name}-{metadata.version}.dist-info"
+        root_is_purelib = "true"
+        requires_dist_torch = "torch"
+    else:
+        python_tag = f"cp{python_version_flat}"
+        abi_tag = "abi3"
+
+        if (
+            metadata.torch_version is None
+            or metadata.cuda_version is None
+            or metadata.cxx_abi_version is None
+            or metadata.os is None
+            or metadata.platform is None
+        ):
+            raise ValueError(
+                "Torch version, CUDA version, C++ ABI version, OS, and platform must be specified for non-universal wheels."
+            )
+
+        local_version = f"torch{metadata.torch_version.replace('.', '')}cu{metadata.cuda_version}cxx{metadata.cxx_abi_version}"
+
+        if metadata.os == "linux":
+            platform_tag = (
+                f"manylinux_{manylinux_version.replace('.', '_')}_{metadata.platform}"
+            )
+        else:
+            platform_tag = f"{metadata.os}_{metadata.platform.replace('-', '_')}"
+
+        wheel_filename = f"{name}-{metadata.version}+{local_version}-{python_tag}-{abi_tag}-{platform_tag}.whl"
+        dist_info_dir_name = f"{name}-{metadata.version}+{local_version}.dist-info"
+        root_is_purelib = "false"
+        requires_dist_torch = f"torch=={metadata.torch_version}.*"
+
+    wheel_path = wheel_dir / wheel_filename
+
+    wheel_msg = Message(email.policy.compat32)
+    wheel_msg.add_header("Wheel-Version", "1.0")
+    wheel_msg.add_header("Generator", f"kernels ({KERNELS_VERSION})")
+    wheel_msg.add_header("Root-Is-Purelib", root_is_purelib)
+    wheel_msg.add_header("Tag", f"{python_tag}-{abi_tag}-{platform_tag}")
+
+    metadata_msg = Message(email.policy.compat32)
+    metadata_msg.add_header("Metadata-Version", "2.1")
+    metadata_msg.add_header("Name", name)
+    metadata_msg.add_header("Version", metadata.version)
+    metadata_msg.add_header("Summary", f"{name} kernel")
+    metadata_msg.add_header("Requires-Python", ">=3.9")
+    metadata_msg.add_header("Requires-Dist", requires_dist_torch)
+
+    source_pkg_dir = variant_path / name
+
+    with WheelFile(wheel_path, "w") as wheel_file:
+        for root, dirnames, filenames in os.walk(source_pkg_dir):
+            for filename in filenames:
+                if filename.endswith(".pyc"):
+                    continue
+
+                abs_filepath = os.path.join(root, filename)
+                entry_name = os.path.relpath(abs_filepath, variant_path)
+                wheel_file.write(abs_filepath, entry_name)
+
+        wheel_metadata_path = os.path.join(dist_info_dir_name, "WHEEL")
+        wheel_file.writestr(wheel_metadata_path, str(wheel_msg).encode("utf-8"))
+
+        metadata_path = os.path.join(dist_info_dir_name, "METADATA")
+        wheel_file.writestr(metadata_path, str(metadata_msg).encode("utf-8"))
+
+    return wheel_path

tests/conftest.py

@@ -0,0 +1,10 @@
+import sys
+
+import pytest
+
+
+def pytest_runtest_setup(item):
+    if "linux_only" in item.keywords and not sys.platform.startswith("linux"):
+        pytest.skip("skipping Linux-only test on non-Linux platform")
+    if "darwin_only" in item.keywords and not sys.platform.startswith("darwin"):
+        pytest.skip("skipping macOS-only test on non-macOS platform")

tests/kernel_locking/kernels.lock

@@ -1,54 +1,82 @@
 [
   {
     "repo_id": "kernels-community/activation",
-    "sha": "6a030420d0dd33ffdc1281afc8ae8e94b4f4f9d0",
+    "sha": "fd6842e88f1f23f198551d78a4541b8eb07e0538",
     "variants": {
       "torch25-cxx11-cu118-x86_64-linux": {
-        "hash": "sha256-3e39de10721a6b21806834fc95c96526b9cfe2c2052829184f2d3fa48ef5849d",
+        "hash": "sha256-61e3e51b5b59b30d4a6ba943a5e6e4ef5a9c8260cc4bca40b9fb462c0777842b",
         "hash_type": "git_lfs_concat"
       },
       "torch25-cxx11-cu121-x86_64-linux": {
-        "hash": "sha256-b0dee22c65bb277fa8150f9ea3fc90e2b1c11f84b5d760bbf4ab9c7a4b102e58",
+        "hash": "sha256-baa6b872040730bd1d676c011381f6f626fb96189837b828f587c806af8994fa",
         "hash_type": "git_lfs_concat"
       },
       "torch25-cxx11-cu124-x86_64-linux": {
-        "hash": "sha256-8960cf857d641d591a7c2d4264925cc2bf7b4a6f9d738b74082b2fb0806db19a",
+        "hash": "sha256-c1ec7457847fa1f0e4ab43234dfc3cd0959977e03dc2ffe89b4f6b90970c7965",
         "hash_type": "git_lfs_concat"
       },
       "torch25-cxx98-cu118-x86_64-linux": {
-        "hash": "sha256-0496e04c2900a2dc7ab0f3b95fe8ce9da69faab6b5ca3f55ddd62c26c81268d0",
+        "hash": "sha256-412f9c841f20741e42f2c6cdb8c7da0e33ab436b219975acffe18b62b97ecd7c",
         "hash_type": "git_lfs_concat"
       },
       "torch25-cxx98-cu121-x86_64-linux": {
-        "hash": "sha256-172b793b24dfed3dcb9adc7d3487f260c05b310c598fc6ee8abb3e230c59a0a8",
+        "hash": "sha256-2fde7f97859506e000c1072b3916c0a75bc8cee750a9853ea8b68199e7b57bcd",
         "hash_type": "git_lfs_concat"
       },
       "torch25-cxx98-cu124-x86_64-linux": {
-        "hash": "sha256-12f5e66f32dc4cf4b21f43f76efad198556024da67a1ce28e88ea2d49ad8bdcc",
+        "hash": "sha256-93309986f39a64a5630378108154866f0545178fa8dfef9b8f8ccfef9a78608e",
         "hash_type": "git_lfs_concat"
       },
       "torch26-cxx11-cu118-x86_64-linux": {
-        "hash": "sha256-bb70e2f36f0b4d12868956c2ad713c756570ff0e0eb4cf7fc3a78ebde617975b",
+        "hash": "sha256-3284d3c64b76d92c1ee930bce8013aff307f16eefb16c2d5dea9f2ca70e71e1f",
         "hash_type": "git_lfs_concat"
       },
       "torch26-cxx11-cu124-x86_64-linux": {
-        "hash": "sha256-a745732eb9ec5d6a54565dbeec5b3c983cc6aa072a4a2576ab2fef9b2a600005",
+        "hash": "sha256-36a8c93773c08ddf8ef624a8a6b2866be26d1861450dfe1ecac0bed59f9ffa47",
+        "hash_type": "git_lfs_concat"
+      },
+      "torch26-cxx11-cu126-aarch64-linux": {
+        "hash": "sha256-f5afb734520f587717665659798ff738a69e5ae1e34d4bd95624edd18fb165cd",
         "hash_type": "git_lfs_concat"
       },
       "torch26-cxx11-cu126-x86_64-linux": {
-        "hash": "sha256-1160684ca09c065864f27c5c110281807a1ec31d603bf05fcb974e9e7cfe35cc",
+        "hash": "sha256-940841a7cb44f76c9a896d8b39f5bc0e0420f1c4c05ae9423da96778de4d1f2c",
         "hash_type": "git_lfs_concat"
       },
       "torch26-cxx98-cu118-x86_64-linux": {
-        "hash": "sha256-24459d068943b93e4d55e94811469bf7e850d7958785132b108f1240724b846f",
+        "hash": "sha256-8e0f907830c3acc8c6bebfc162c744012ff6973e8110d7bf8ecd74b492418204",
         "hash_type": "git_lfs_concat"
       },
       "torch26-cxx98-cu124-x86_64-linux": {
-        "hash": "sha256-5b009ba63ab6d52ac1aaf70057a2d0fa6ea5d1788a2416111be02103c6bcaaaf",
+        "hash": "sha256-0833414cbe658baec55b7ff63537cddccc973fe99e3c03008cced5e66e38b6c1",
+        "hash_type": "git_lfs_concat"
+      },
+      "torch26-cxx98-cu126-aarch64-linux": {
+        "hash": "sha256-d94fa59a13a5b623b2071aadcd1e6c8477c4d557fd06ad144f15b46b1fc71aab",
         "hash_type": "git_lfs_concat"
       },
       "torch26-cxx98-cu126-x86_64-linux": {
-        "hash": "sha256-05128889b4bdaf9ef58f3c07d93218deaa08e06f9121931b47efef8826482e4a",
+        "hash": "sha256-64784f5f2f9e232d0f2fd824fbc47eadde505e3c232f351bead5b04c429c65c2",
+        "hash_type": "git_lfs_concat"
+      },
+      "torch27-cxx11-cu118-x86_64-linux": {
+        "hash": "sha256-bcba3765f061649bac0e5a9159bea8349ced4780e24a2330aa62ce0f8d3a9d78",
+        "hash_type": "git_lfs_concat"
+      },
+      "torch27-cxx11-cu126-aarch64-linux": {
+        "hash": "sha256-e4625df5706af025c70bd824d952b928d9a2965eeaefda72fc47be0fae680c5e",
+        "hash_type": "git_lfs_concat"
+      },
+      "torch27-cxx11-cu126-x86_64-linux": {
+        "hash": "sha256-7d7d3e655f34a7b03d5603d7c1ab723ef3efc823291762421a8b3a4aa51bd405",
+        "hash_type": "git_lfs_concat"
+      },
+      "torch27-cxx11-cu128-aarch64-linux": {
+        "hash": "sha256-60e076194dcd55b32c5aca72f09816cba0fff52f340c8a063b17ff0577154d99",
+        "hash_type": "git_lfs_concat"
+      },
+      "torch27-cxx11-cu128-x86_64-linux": {
+        "hash": "sha256-f0a3802382efdcd78b40601187a9c416579a24ef2ed5a60d2296ef0951a89597",
         "hash_type": "git_lfs_concat"
       }
     }

tests/test_basic.py

@@ -1,7 +1,7 @@
 import pytest
 import torch
 
-from kernels import get_kernel, has_kernel
+from kernels import get_kernel, get_local_kernel, has_kernel, install_kernel
 
 
 @pytest.fixture
@@ -9,6 +9,19 @@ def kernel():
     return get_kernel("kernels-community/activation")
 
 
+@pytest.fixture
+def local_kernel():
+    package_name, path = install_kernel("kernels-community/activation", "main")
+    # Path is the build variant path (build/torch-<...>), so the grandparent
+    # is the kernel repository path.
+    return get_local_kernel(path.parent.parent, package_name)
+
+
+@pytest.fixture
+def metal_kernel():
+    return get_kernel("kernels-test/relu-metal")
+
+
 @pytest.fixture
 def universal_kernel():
     return get_kernel("kernels-community/triton-scaled-mm")
@@ -21,16 +34,12 @@ def device():
     return "cuda"
 
 
-@pytest.mark.parametrize("torch_compile", [False, True])
-def test_gelu_fast(kernel, device, torch_compile):
+@pytest.mark.linux_only
+def test_gelu_fast(kernel, device):
     x = torch.arange(1, 10, dtype=torch.float16, device=device).view(3, 3)
     y = torch.empty_like(x)
 
-    op = kernel.gelu_fast
-    if torch_compile:
-        op = torch.compile(op)
-
-    op(y, x)
+    kernel.gelu_fast(y, x)
 
     expected = torch.tensor(
         [[0.8408, 1.9551, 2.9961], [4.0000, 5.0000, 6.0000], [7.0000, 8.0000, 9.0000]],
@@ -41,6 +50,31 @@ def test_gelu_fast(kernel, device, torch_compile):
     assert torch.allclose(y, expected)
 
 
+@pytest.mark.linux_only
+def test_local_kernel(local_kernel, device):
+    x = torch.arange(1, 10, dtype=torch.float16, device=device).view(3, 3)
+    y = torch.empty_like(x)
+
+    local_kernel.gelu_fast(y, x)
+
+    expected = torch.tensor(
+        [[0.8408, 1.9551, 2.9961], [4.0000, 5.0000, 6.0000], [7.0000, 8.0000, 9.0000]],
+        device=device,
+        dtype=torch.float16,
+    )
+
+    assert torch.allclose(y, expected)
+
+
+@pytest.mark.darwin_only
+@pytest.mark.parametrize("dtype", [torch.float16, torch.float32])
+def test_relu_metal(metal_kernel, dtype):
+    x = torch.arange(-10, 10, dtype=dtype, device="mps")
+    y = metal_kernel.relu(x)
+    assert torch.allclose(y, torch.relu(x))
+
+
+@pytest.mark.linux_only
 @pytest.mark.parametrize(
     "kernel_exists",
     [
@@ -57,6 +91,7 @@ def test_has_kernel(kernel_exists):
     assert has_kernel(repo_id, revision=revision) == kernel
 
 
+@pytest.mark.linux_only
 def test_universal_kernel(universal_kernel):
     torch.manual_seed(0)
     A = torch.randint(-10, 10, (64, 128), dtype=torch.int8, device="cuda")

tests/test_benchmarks.py

@@ -16,18 +16,21 @@ def device():
     return "cuda"
 
 
+@pytest.mark.linux_only
 def test_gelu_small(kernel, device, benchmark):
     x = torch.randn(32, 32, dtype=torch.float16, device=device)
     y = torch.empty_like(x)
     benchmark(kernel.gelu_fast, y, x)
 
 
+@pytest.mark.linux_only
 def test_gelu_medium(kernel, device, benchmark):
     x = torch.randn(128, 128, dtype=torch.float16, device=device)
     y = torch.empty_like(x)
     benchmark(kernel.gelu_fast, y, x)
 
 
+@pytest.mark.linux_only
 def test_gelu_large(kernel, device, benchmark):
     x = torch.randn(512, 512, dtype=torch.float16, device=device)
     y = torch.empty_like(x)

tests/test_kernel_locking.py

@@ -1,6 +1,8 @@
 from dataclasses import dataclass
 from pathlib import Path
 
+import pytest
+
 from kernels import load_kernel
 from kernels.cli import download_kernels
 
@@ -17,6 +19,7 @@ def test_download_all_hash_validation():
     download_kernels(DownloadArgs(all_variants=True, project_dir=project_dir))
 
 
+@pytest.mark.linux_only
 def test_load_locked():
     project_dir = Path(__file__).parent / "kernel_locking"
     # Also validates that hashing works correctly.

tests/test_layer.py

@@ -1,3 +1,5 @@
+from contextlib import nullcontext
+
 import pytest
 import torch
 import torch.nn as nn
@@ -6,6 +8,8 @@ from torch.nn import functional as F
 from kernels import (
     Device,
     LayerRepository,
+    Mode,
+    kernelize,
     register_kernel_mapping,
     use_kernel_forward_from_hub,
 )
@@ -16,14 +20,18 @@ kernel_layer_mapping = {
         Device(type="cuda"): LayerRepository(
             repo_id="kernels-community/activation",
             layer_name="SiluAndMul",
-            revision="layers",
+        )
+    },
+    "SiluAndMulNoCompile": {
+        "cuda": LayerRepository(
+            repo_id="kernels-test/op-without-fake-test",
+            layer_name="SiluAndMul",
         )
     },
     "SiluAndMulStringDevice": {
         "cuda": LayerRepository(
             repo_id="kernels-community/activation",
             layer_name="SiluAndMul",
-            revision="layers",
         )
     },
 }
@@ -43,6 +51,11 @@ class SiluAndMul(nn.Module):
         return F.silu(input[..., :d]) * input[..., d:]
 
 
+@use_kernel_forward_from_hub("SiluAndMulNoCompile")
+class SiluAndMulNoCompileKernel(SiluAndMul):
+    pass
+
+
 @use_kernel_forward_from_hub("SiluAndMul")
 class SiluAndMulWithKernel(SiluAndMul):
     pass
@@ -53,6 +66,18 @@ class SiluAndMulStringDevice(SiluAndMul):
     pass
 
 
+@use_kernel_forward_from_hub("Linear")
+class TorchLinearWithCounter(nn.Linear):
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        # Used to check that we called hub kernel.
+        self.n_calls = 0
+
+    def forward(self, input: torch.Tensor) -> torch.Tensor:
+        self.n_calls += 1
+        return super().forward(input)
+
+
 def test_arg_kinds():
     @use_kernel_forward_from_hub("ArgKind")
     class ArgKind(nn.Module):
@@ -71,19 +96,17 @@ def test_arg_kinds():
     assert arg_kind("foo", "bar", kwarg1="baz", kwarg2=5) == ("foo", "bar", "baz", 5)
 
 
+@pytest.mark.linux_only
 @pytest.mark.parametrize("cls", [SiluAndMulWithKernel, SiluAndMulStringDevice])
 @pytest.mark.parametrize("device", ["cuda", "cpu"])
-@pytest.mark.parametrize("torch_compile", [False, True])
-def test_hub_forward(cls, device, torch_compile):
+def test_hub_forward(cls, device):
     torch.random.manual_seed(0)
 
     silu_and_mul = SiluAndMul()
     X = torch.randn((32, 64), device=device)
     Y = silu_and_mul(X)
 
-    silu_and_mul_with_kernel = cls()
-    if torch_compile:
-        silu_and_mul_with_kernel = torch.compile(silu_and_mul_with_kernel)
+    silu_and_mul_with_kernel = kernelize(cls(), device=device, mode=Mode.INFERENCE)
     Y_kernel = silu_and_mul_with_kernel(X)
 
     torch.testing.assert_close(Y_kernel, Y)
@@ -101,11 +124,70 @@ def test_layer_fallback_works():
         pass
 
     # Check that we don't raise an exception for a non-existing kernel.
-    SiluAndMulWithKernelFallback()
+    silu_and_mul = SiluAndMulWithKernelFallback()
+    kernelize(silu_and_mul, device="cuda", mode=Mode.INFERENCE)
+
+
+@pytest.mark.linux_only
+@pytest.mark.parametrize("cls", [SiluAndMulWithKernel, SiluAndMulNoCompileKernel])
+@pytest.mark.parametrize("device", ["cuda"])
+def test_torch_compile_layer_without_fallback(cls, device):
+    silu_and_mul = SiluAndMul()
+
+    X = torch.randn((32, 64), dtype=torch.float32, device=device)
+    Y = silu_and_mul(X)
+
+    silu_and_mul_with_kernel = cls()
+    silu_and_mul_with_kernel.eval()
+
+    ctx = (
+        pytest.raises(ValueError, match="does not support mode")
+        if cls is SiluAndMulNoCompileKernel
+        else nullcontext()
+    )
+    with ctx:
+        silu_and_mul_with_kernel = kernelize(
+            silu_and_mul_with_kernel,
+            device=device,
+            mode=Mode.INFERENCE | Mode.TORCH_COMPILE,
+            use_fallback=False,
+        )
+    silu_and_mul_compiled = torch.compile(silu_and_mul_with_kernel, fullgraph=True)
+
+    Y_compiled = silu_and_mul_compiled(X)
+
+    torch.testing.assert_close(Y_compiled, Y)
+
+
+@pytest.mark.linux_only
+@pytest.mark.parametrize("cls", [SiluAndMulWithKernel, SiluAndMulNoCompileKernel])
+@pytest.mark.parametrize("device", ["cuda"])
+def test_torch_compile_layer_with_fallback(cls, device):
+    silu_and_mul = SiluAndMul()
+
+    X = torch.randn((32, 64), dtype=torch.float32, device=device)
+    Y = silu_and_mul(X)
+
+    silu_and_mul_with_kernel = cls()
+    silu_and_mul_with_kernel.eval()
+    silu_and_mul_with_kernel = kernelize(
+        silu_and_mul_with_kernel,
+        device=device,
+        mode=Mode.INFERENCE | Mode.TORCH_COMPILE,
+    )
+    silu_and_mul_compiled = torch.compile(silu_and_mul_with_kernel, fullgraph=True)
+
+    Y_compiled = silu_and_mul_compiled(X)
+
+    torch.testing.assert_close(Y_compiled, Y)
 
 
 def test_mapping_contexts():
-    assert set(_KERNEL_MAPPING.get().keys()) == {"SiluAndMul", "SiluAndMulStringDevice"}
+    assert set(_KERNEL_MAPPING.get().keys()) == {
+        "SiluAndMul",
+        "SiluAndMulStringDevice",
+        "SiluAndMulNoCompile",
+    }
 
     extra_mapping1 = {
         "TestKernel": {
@@ -121,6 +203,7 @@ def test_mapping_contexts():
         assert set(_KERNEL_MAPPING.get().keys()) == {
             "SiluAndMul",
             "SiluAndMulStringDevice",
+            "SiluAndMulNoCompile",
             "TestKernel",
         }
 
@@ -138,20 +221,26 @@ def test_mapping_contexts():
             assert set(_KERNEL_MAPPING.get().keys()) == {
                 "SiluAndMul",
                 "SiluAndMulStringDevice",
+                "SiluAndMulNoCompile",
                 "TestKernel",
             }
             assert (
-                _KERNEL_MAPPING.get()["SiluAndMul"][Device(type="cuda")].repo_id
+                _KERNEL_MAPPING.get()["SiluAndMul"][Device(type="cuda")][
+                    Mode.DEFAULT
+                ].repo_id
                 == "kernels-community/non-existing"
             )
 
         assert set(_KERNEL_MAPPING.get().keys()) == {
             "SiluAndMul",
             "SiluAndMulStringDevice",
+            "SiluAndMulNoCompile",
             "TestKernel",
         }
         assert (
-            _KERNEL_MAPPING.get()["SiluAndMul"][Device(type="cuda")].repo_id
+            _KERNEL_MAPPING.get()["SiluAndMul"][Device(type="cuda")][
+                Mode.DEFAULT
+            ].repo_id
             == "kernels-community/activation"
         )
 
@@ -160,23 +249,29 @@ def test_mapping_contexts():
                 "SiluAndMul",
             }
             assert (
-                _KERNEL_MAPPING.get()["SiluAndMul"][Device(type="cuda")].repo_id
+                _KERNEL_MAPPING.get()["SiluAndMul"][Device(type="cuda")][
+                    Mode.DEFAULT
+                ].repo_id
                 == "kernels-community/non-existing"
             )
 
         assert set(_KERNEL_MAPPING.get().keys()) == {
             "SiluAndMul",
             "SiluAndMulStringDevice",
+            "SiluAndMulNoCompile",
             "TestKernel",
         }
         assert (
-            _KERNEL_MAPPING.get()["SiluAndMul"][Device(type="cuda")].repo_id
+            _KERNEL_MAPPING.get()["SiluAndMul"][Device(type="cuda")][
+                Mode.DEFAULT
+            ].repo_id
             == "kernels-community/activation"
         )
 
     assert set(_KERNEL_MAPPING.get().keys()) == {
         "SiluAndMul",
         "SiluAndMulStringDevice",
+        "SiluAndMulNoCompile",
     }
 
 
@@ -208,39 +303,152 @@ def test_validate_kernel_layer():
         _validate_layer(cls=BadLayer4, check_cls=SiluAndMul)
 
 
-def test_fallback_used_when_training():
-    @use_kernel_forward_from_hub("Linear")
-    class TorchLinear(nn.Linear):
-        def __init__(self, *args, **kwargs):
-            super().__init__(*args, **kwargs)
-            # Used to check that we called hub kernel.
-            self.n_calls = 0
-
-        def forward(self, input: torch.Tensor) -> torch.Tensor:
-            self.n_calls += 1
-            return super().forward(input)
-
-    linear = TorchLinear(32, 32).to("cuda")
+def test_invalid_mode_for_mapping_rejected():
+    linear = TorchLinearWithCounter(32, 32).to("cuda")
 
     with use_kernel_mapping(
         {
             "Linear": {
-                Device(type="cuda"): LayerRepository(
+                "cuda": {
+                    Mode.TRAINING: LayerRepository(
+                        repo_id="kernels-test/backward-marker-test",
+                        layer_name="LinearNoBackward",
+                    )
+                }
+            }
+        }
+    ):
+        with pytest.raises(ValueError, match="does not support backward"):
+            kernelize(linear, mode=Mode.TRAINING)
+
+
+def test_kernel_modes():
+    linear = TorchLinearWithCounter(32, 32).to("cuda")
+
+    # Case 1: layer without further specification, becomes the
+    #         base layer.
+    with use_kernel_mapping(
+        {
+            "Linear": {
+                "cuda": LayerRepository(
                     repo_id="kernels-test/backward-marker-test",
-                    layer_name="LinearImplicitBackward",
+                    layer_name="LinearBackward",
                 )
             }
         }
     ):
-        linear.train()
+        kernelize(linear, mode=Mode.INFERENCE)
         X = torch.randn(10, 32, device="cuda")
         linear(X)
         assert linear.n_calls == 0
 
-        linear.eval()
+        kernelize(linear, mode=Mode.TRAINING)
         linear(X)
         assert linear.n_calls == 0
 
+        kernelize(linear, mode=Mode.TRAINING | Mode.TORCH_COMPILE)
+        linear(X)
+        assert linear.n_calls == 0
+
+    # Case 2: register a kernel just for training. If no base kernel
+    #         layer is registered, we fall back to the original layer.
+    with use_kernel_mapping(
+        {
+            "Linear": {
+                "cuda": {
+                    Mode.TRAINING: LayerRepository(
+                        repo_id="kernels-test/backward-marker-test",
+                        layer_name="LinearBackward",
+                    )
+                }
+            }
+        }
+    ):
+        kernelize(linear, mode=Mode.INFERENCE)
+        X = torch.randn(10, 32, device="cuda")
+        linear(X)
+        assert linear.n_calls == 1
+
+        kernelize(linear, mode=Mode.TRAINING)
+        linear(X)
+        # Training has a kernel, so fallback.
+        assert linear.n_calls == 1
+
+        kernelize(linear, mode=Mode.TRAINING | Mode.TORCH_COMPILE)
+        linear(X)
+        # No kernel for training + torch.compile, so fallback.
+        assert linear.n_calls == 2
+
+    # Case 3: register a kernel just for training and one for fallback.
+    with use_kernel_mapping(
+        {
+            "Linear": {
+                "cuda": {
+                    Mode.DEFAULT: LayerRepository(
+                        repo_id="kernels-test/backward-marker-test",
+                        layer_name="LinearBackward",
+                    ),
+                    Mode.TRAINING: LayerRepository(
+                        repo_id="kernels-test/backward-marker-test",
+                        layer_name="LinearBackward",
+                    ),
+                }
+            }
+        }
+    ):
+        kernelize(linear, mode=Mode.INFERENCE)
+        X = torch.randn(10, 32, device="cuda")
+        linear(X)
+        # Uses the base kernel.
+        assert linear.n_calls == 2
+
+        kernelize(linear, mode=Mode.TRAINING)
+        linear(X)
+        # Uses the training kernel.
+        assert linear.n_calls == 2
+
+        kernelize(linear, mode=Mode.TRAINING | Mode.TORCH_COMPILE)
+        linear(X)
+        # Uses the base kernel.
+        assert linear.n_calls == 2
+
+    # Case 4: register a kernel with two preferences.
+    with use_kernel_mapping(
+        {
+            "Linear": {
+                "cuda": {
+                    Mode.TRAINING
+                    | Mode.TORCH_COMPILE: LayerRepository(
+                        repo_id="kernels-test/backward-marker-test",
+                        layer_name="LinearBackward",
+                    )
+                }
+            }
+        }
+    ):
+        kernelize(linear, mode=Mode.INFERENCE)
+        X = torch.randn(10, 32, device="cuda")
+        linear(X)
+        # No inference kernel, so fallback.
+        assert linear.n_calls == 3
+
+        kernelize(linear, mode=Mode.TRAINING)
+        linear(X)
+        # No training kernel, so fallback.
+        assert linear.n_calls == 4
+
+        kernelize(linear, mode=Mode.TRAINING | Mode.TORCH_COMPILE)
+        linear(X)
+        # We do have a training + torch.compile kernel.
+        assert linear.n_calls == 4
+
+
+@pytest.mark.linux_only
+def test_fallback_used_when_training():
+    linear = TorchLinearWithCounter(32, 32).to("cuda")
+
+    # Case 1: kernel with explicit backward support should always
+    #         use the kernel.
     with use_kernel_mapping(
         {
             "Linear": {
@@ -252,6 +460,7 @@ def test_fallback_used_when_training():
         }
     ):
         linear.train()
+        kernelize(linear, mode=Mode.INFERENCE)
         X = torch.randn(10, 32, device="cuda")
         linear(X)
         assert linear.n_calls == 0
@@ -260,21 +469,40 @@ def test_fallback_used_when_training():
         linear(X)
         assert linear.n_calls == 0
 
+    # Case 2: kernel with implicit backward support should always
+    #         use the kernel.
     with use_kernel_mapping(
         {
             "Linear": {
                 Device(type="cuda"): LayerRepository(
                     repo_id="kernels-test/backward-marker-test",
-                    layer_name="LinearNoBackward",
+                    layer_name="LinearImplicitBackward",
                 )
             }
         }
     ):
         linear.train()
+        kernelize(linear, mode=Mode.INFERENCE)
         X = torch.randn(10, 32, device="cuda")
         linear(X)
-        assert linear.n_calls == 1
+        assert linear.n_calls == 0
 
         linear.eval()
         linear(X)
-        assert linear.n_calls == 1
+        assert linear.n_calls == 0
+
+
+def test_invalid_mode_rejected():
+    with pytest.raises(ValueError, match="mutually exclusive"):
+        _ = Mode.INFERENCE | Mode.TRAINING
+
+    with pytest.raises(ValueError, match="cannot be combined with other modes"):
+        _ = Mode.DEFAULT | Mode.TORCH_COMPILE
+
+    with pytest.raises(
+        ValueError, match="can only be used to register kernel mappings"
+    ):
+        kernelize(torch.nn.Linear(32, 32), mode=Mode.DEFAULT)
+
+    with pytest.raises(ValueError, match="mode must contain"):
+        kernelize(torch.nn.Linear(32, 32), mode=Mode.TORCH_COMPILE)