* Set `kernelize` default mode to `Mode.TRAINING | Mode.TORCH_COMPILE`
Also update docs and tests.
* Rename `Mode.DEFAULT` to `Mode.FALLBACK`
* More fine-grained fallbacks
For instance, INFERENCE can fall back to INFERENCE | TORCH_COMPILE,
TRAINING, TRAINING | TORCH_COMPILE, and FALLBACK.
* Update documtenation for mode fallback
* Mention that you can rerun `kernelize` to change the mode
* Add interval tree implementation
* Support registering layers with a range of CUDA capabilities
This change adds support for registering a layers for ranges
of CUDA capabilities. This makes it possible to use newer, faster
kernels for new GPUs, while falling back to another implementation
on older GPUs.
* Add docs for registering kernels with CUDA capabilities
* Fix typing errors
* 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>
For kernels that do not support backward, fall back to the original
implementation if `model.train(True)` is called. This removes the
need for the `needs_backward` argument of `kernelize`.
* first commit
* style
* update
* fix
* different approach
* Polish kernelize
- Process comment from the PR.
- Replacement should be on instances, not the class.
- Remove torch compile checks (not relevant during kernelize). We
might add it back in a different way in another commit: add an
option to `kernelize`.
* Fixup tests
* Fix `torch.compile` support
* Remove some unused code
* Sync the docs
* CI: update Torch versions
---------
Co-authored-by: Daniël de Kok <me@danieldk.eu>
* Add `generate-readme` subcommand for generating a README
This README includes all the top-level functions with docs (if
docstrings are available).
* CI: attempt README generation
* Add PyYAML dependencies
* Typing fixes
* kernels: add the `to-wheel` subcommand
This subcommand accepts a kernel repo and version as arguments:
kernels to-wheel kernels-community/activation 0.0.3
Wheels will then be generated for every build variant.
* CI: check kernel -> wheel conversion
* No typing for wheel.wheelfile
* Allow layers to opt in to `torch.compile`
This change allows a layer to set the `can_torch_compile` class
variable to indicate that the layer is compatible with `torch.compile`.
When enabled, the layer does not fall back to the original
implementation when `torch.compile` is used.
* Comment fixes
Co-authored-by: Mohamed Mekkouri <93391238+MekkCyber@users.noreply.github.com>
---------
Co-authored-by: Mohamed Mekkouri <93391238+MekkCyber@users.noreply.github.com>
* Add `has_kernel` function
This function checks whether a kernel build exists for the current
environment (Torch version and compute framework).
* Test kernel repo that only contains Torch 2.4
* Disable kernel mappings with `DISABLE_KERNEL_MAPPING=1`
* Rename HF_KERNELS_CACHE to KERNELS_CACHE
But still recognize the old variant for compatibility.
* Add documentation for environment variables
`inherit_mapping` is the default and extends the existing mapping
with the given mapping. If `inherit_mapping` is `False`, existing
mappings are not inherited.
