File size: 4,286 Bytes
db704cb | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 | # Copyright 2025 the LlamaFactory team.
#
# 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.
"""The definition of kernel interface.
Init Phase:
1. Scan all kernels.
2. Register default kernels.
3. Define kernel plugin.
"""
import importlib
from pathlib import Path
from ....utils import logging
from ....utils.plugin import BasePlugin
from ....utils.types import HFModel
from .registry import Registry
logger = logging.get_logger(__name__)
def scan_all_kernels():
"""Scan all kernels in the ``ops`` directory.
Scans the ``ops`` directory for all ``.py`` files and attempts to import them.
Importing triggers the :func:`~registry.register_kernel` decorator, which automatically registers the kernels.
Returns:
dict[str, type[BaseKernel]]: A dictionary of registered kernels.
.. note::
This function assumes that the ``ops`` directory is located in the same directory as this file.
It recursively searches for ``.py`` files and constructs the module path for import.
"""
ops_path = Path(__file__).parent / "ops"
if not ops_path.exists():
return
base_package = __package__
for file_path in ops_path.rglob("*.py"):
if file_path.name == "__init__.py":
continue
# calculate the relative path:
# file_path = .../kernels_v2/ops/mlp/npu_swiglu.py
# rel_path = ops/mlp/npu_swiglu.py
rel_path = file_path.relative_to(Path(__file__).parent)
# build module path:
module_name = ".".join(rel_path.parts)[:-3]
full_module_name = f"{base_package}.{module_name}"
try:
importlib.import_module(full_module_name)
except Exception as e:
logger.warning(f"[Kernel Registry] Failed to import {full_module_name} when loading kernels: {e}")
return Registry.get_registered_kernels()
default_kernels = scan_all_kernels()
def get_default_kernels():
"""Get a list of default registered kernel IDs.
Returns:
list[str]: List of kernel IDs.
"""
return list(default_kernels.keys())
def apply_kernel(kernel_id: str, **kwargs):
"""Applies a specific kernel to the model.
Args:
kernel_id (str): The ID of the kernel to apply.
**kwargs: Keyword arguments passed to the kernel application function.
Typically includes the model instance.
Returns:
HFModel: The model with applied kernel.
"""
kernel = default_kernels.get(kernel_id)
if kernel is None:
raise ValueError(f"Kernel {kernel_id} not found")
kernel.apply(**kwargs)
class KernelPlugin(BasePlugin):
"""Plugin for managing kernel optimizations."""
pass
@KernelPlugin("auto").register()
def apply_default_kernels(model: HFModel, include_kernels: str = None) -> HFModel:
"""Applies all default registered kernels to the model.
Args:
model (HFModel): The model instance to apply kernels to.
include_kernels (str, optional): Comma-separated list of kernel IDs to apply.
If "auto" or True, applies all default kernels.
If None or False, no kernels are applied.
Defaults to None.
Returns:
HFModel: The model with applied kernels.
"""
if not include_kernels:
return model
elif include_kernels == "auto" or include_kernels is True:
use_kernels = default_kernels.keys()
else:
use_kernels = include_kernels.split(",") # "kernel_id1,kernel_id2,kernel_id3"
for kernel in use_kernels:
if kernel not in default_kernels:
raise ValueError(f"Kernel {kernel} not found")
apply_kernel(kernel, model=model)
return model
|