|
| 1 | +# Twinkle Kernel 模块 |
| 2 | + |
| 3 | +Twinkle Kernel 模块提供了两条内核替换路径,用于加速训练和推理: |
| 4 | + |
| 5 | +* **层级 Kernelize(Layer-level kernelize)** |
| 6 | + 使用优化内核替换完整的 `nn.Module` 实现。 |
| 7 | +* **函数级 Kernelize(Function-level kernelize)** |
| 8 | + 对 Python 模块中的特定函数进行 monkey-patch。 |
| 9 | + |
| 10 | +这两种方式可以独立使用,也可以通过统一入口组合使用。 |
| 11 | + |
| 12 | +--- |
| 13 | + |
| 14 | +## 概览:两条 Kernelize 路径 |
| 15 | + |
| 16 | +| 路径 | 粒度 | 典型场景 | |
| 17 | +| --- | --- | --- | |
| 18 | +| 层级替换 | 整个 `nn.Module` | Linear / Conv / MLP / Attention | |
| 19 | +| 函数级替换 | 单个函数 | 热点路径、数学算子、激活函数 | |
| 20 | + |
| 21 | +--- |
| 22 | + |
| 23 | +## 层级内核替换(Layer-Level) |
| 24 | + |
| 25 | +### 适用场景 |
| 26 | + |
| 27 | +* 你已经有完整的层内核实现 |
| 28 | +* 希望在模型中批量替换某类 `nn.Module` |
| 29 | +* 同时适用于训练与推理 |
| 30 | + |
| 31 | +--- |
| 32 | + |
| 33 | +### 示例 1:本地 Kernel 仓库 |
| 34 | + |
| 35 | +适用于: |
| 36 | + |
| 37 | +* 内核实现位于本地仓库 |
| 38 | +* 希望替换 HuggingFace 或自定义模型中的层 |
| 39 | + |
| 40 | +```python |
| 41 | +from twinkle.kernel import ( |
| 42 | + kernelize_model, |
| 43 | + register_layer_kernel, |
| 44 | + register_external_layer, |
| 45 | +) |
| 46 | +from transformers import Qwen2Config, Qwen2ForCausalLM |
| 47 | +from transformers.models.qwen2.modeling_qwen2 import Qwen2MLP |
| 48 | + |
| 49 | +# 1) 从本地仓库注册层内核 |
| 50 | +register_layer_kernel( |
| 51 | + kernel_name="MyAwesomeMLP", |
| 52 | + repo_path="/path/to/local/repo", |
| 53 | + package_name="my_kernels", |
| 54 | + layer_name="Qwen2MLPTrainingKernel", |
| 55 | + device="cuda", |
| 56 | + mode="train", |
| 57 | +) |
| 58 | + |
| 59 | +# 2) 绑定外部层与内核名 |
| 60 | +register_external_layer(Qwen2MLP, "MyAwesomeMLP") |
| 61 | + |
| 62 | +# 3) 构建模型并应用内核替换 |
| 63 | +config = Qwen2Config( |
| 64 | + hidden_size=128, |
| 65 | + num_hidden_layers=1, |
| 66 | + num_attention_heads=4, |
| 67 | + num_key_value_heads=4, |
| 68 | + intermediate_size=256, |
| 69 | + use_cache=False, |
| 70 | +) |
| 71 | +model = Qwen2ForCausalLM(config) |
| 72 | +model = kernelize_model(model, mode="train", device="cuda", use_fallback=True) |
| 73 | +``` |
| 74 | + |
| 75 | +--- |
| 76 | + |
| 77 | +### 示例 2:Hub Kernel 仓库 |
| 78 | + |
| 79 | +适用于: |
| 80 | + |
| 81 | +* 内核托管在 Hub 上 |
| 82 | + |
| 83 | +```python |
| 84 | +import torch |
| 85 | +import torch.nn as nn |
| 86 | +from twinkle.kernel import ( |
| 87 | + kernelize_model, |
| 88 | + register_layer_kernel, |
| 89 | + register_external_layer, |
| 90 | +) |
| 91 | + |
| 92 | +# 1) 定义自定义层 |
| 93 | +class SiluAndMul(nn.Module): |
| 94 | + def forward(self, x: torch.Tensor) -> torch.Tensor: |
| 95 | + x1, x2 = x.chunk(2, dim=-1) |
| 96 | + return nn.functional.silu(x1) * x2 |
| 97 | + |
| 98 | +# 2) 注册 Hub 内核并绑定层 |
| 99 | +register_layer_kernel( |
| 100 | + kernel_name="SiluAndMulKernel", |
| 101 | + repo_id="kernels-community/activation", |
| 102 | + layer_name="SiluAndMul", |
| 103 | + device="cuda", |
| 104 | + mode="train", |
| 105 | +) |
| 106 | +register_external_layer(SiluAndMul, "SiluAndMulKernel") |
| 107 | + |
| 108 | +# 3) 应用到模型 |
| 109 | +class SimpleModel(nn.Module): |
| 110 | + def __init__(self): |
| 111 | + super().__init__() |
| 112 | + self.activation = SiluAndMul() |
| 113 | + |
| 114 | + def forward(self, x: torch.Tensor) -> torch.Tensor: |
| 115 | + return self.activation(x) |
| 116 | + |
| 117 | +model = SimpleModel() |
| 118 | +model = kernelize_model(model, mode="train", device="cuda", use_fallback=True) |
| 119 | +``` |
| 120 | + |
| 121 | +--- |
| 122 | + |
| 123 | +## 本地 Kernel 仓库(最小结构) |
| 124 | + |
| 125 | +本地 kernel 仓库本质上是一个普通 Python 包。 |
| 126 | +最少只需要一个 `layers.py` 来放层级内核实现。 |
| 127 | + |
| 128 | +```text |
| 129 | +# 仓库结构: |
| 130 | +my_kernels/ # 本地 kernel 仓库(Python 包) |
| 131 | +├── __init__.py # 包入口 |
| 132 | +└── layers.py # 层级 kernel 实现 |
| 133 | +``` |
| 134 | + |
| 135 | +```python |
| 136 | +# my_kernels/__init__.py |
| 137 | +from . import layers |
| 138 | +__all__ = ["layers"] |
| 139 | + |
| 140 | +# my_kernels/layers.py |
| 141 | +import torch |
| 142 | +import torch.nn as nn |
| 143 | + |
| 144 | +class Qwen2MLPTrainingKernel(nn.Module): |
| 145 | + def forward(self, x: torch.Tensor) -> torch.Tensor: |
| 146 | + gate = self.gate_proj(x) |
| 147 | + up = self.up_proj(x) |
| 148 | + return self.down_proj(self.act_fn(gate) * up) |
| 149 | +``` |
| 150 | + |
| 151 | +--- |
| 152 | + |
| 153 | +## 函数级内核替换(Function-Level) |
| 154 | + |
| 155 | +### 适用场景 |
| 156 | + |
| 157 | +* 只需要加速少量热点函数 |
| 158 | +* 不适合或不需要替换整个层 |
| 159 | +* 常用于数学算子、激活函数、工具函数 |
| 160 | + |
| 161 | +--- |
| 162 | + |
| 163 | +### 示例 1:批量注册(简单场景) |
| 164 | + |
| 165 | +```python |
| 166 | +from twinkle.kernel import register_kernels, kernelize_model |
| 167 | + |
| 168 | +# 1) 注册函数内核 |
| 169 | +config = { |
| 170 | + "functions": { |
| 171 | + "add": { |
| 172 | + "target_module": "my_pkg.math_ops", |
| 173 | + "func_impl": lambda x, y: x + y + 1, |
| 174 | + "device": "cuda", |
| 175 | + "mode": "inference", |
| 176 | + }, |
| 177 | + }, |
| 178 | +} |
| 179 | +register_kernels(config) |
| 180 | + |
| 181 | +# 2) 应用(仅函数替换时 model 可为 None) |
| 182 | +kernelize_model(model=None, mode="inference", device="cuda", use_fallback=True) |
| 183 | +``` |
| 184 | + |
| 185 | +--- |
| 186 | + |
| 187 | +### 示例 2:高级函数来源(完整控制) |
| 188 | + |
| 189 | +适用于: |
| 190 | + |
| 191 | +* 不同函数来自不同来源(impl / repo / hub),或需要 compile/backward 等标志。 |
| 192 | + |
| 193 | +```python |
| 194 | +from twinkle.kernel.function import ( |
| 195 | + register_function_kernel, |
| 196 | + apply_function_kernel, |
| 197 | +) |
| 198 | +import torch.nn as nn |
| 199 | +from twinkle.kernel import kernelize_model |
| 200 | + |
| 201 | +TARGET_MODULE = "my_pkg.math_ops" |
| 202 | + |
| 203 | +# 1) 直接传入实现 |
| 204 | +def fast_add(x, y): |
| 205 | + return x + y + 1 |
| 206 | + |
| 207 | +register_function_kernel( |
| 208 | + func_name="add", |
| 209 | + target_module=TARGET_MODULE, |
| 210 | + func_impl=fast_add, |
| 211 | + device="cuda", |
| 212 | + mode="inference", |
| 213 | +) |
| 214 | + |
| 215 | +# 2) Repo 对象(FuncRepositoryProtocol) |
| 216 | +class MyFuncRepo: |
| 217 | + def load(self): |
| 218 | + return MyKernelFunc |
| 219 | + |
| 220 | +class MyKernelFunc(nn.Module): |
| 221 | + def forward(self, x, y): |
| 222 | + return x * y |
| 223 | + |
| 224 | +register_function_kernel( |
| 225 | + func_name="mul", |
| 226 | + target_module=TARGET_MODULE, |
| 227 | + repo=MyFuncRepo(), |
| 228 | + device="cuda", |
| 229 | + mode="compile", |
| 230 | +) |
| 231 | + |
| 232 | +# 3) Hub 仓库 |
| 233 | +register_function_kernel( |
| 234 | + func_name="silu_and_mul", |
| 235 | + target_module="my_pkg.activations", |
| 236 | + repo_id="kernels-community/activation", |
| 237 | + revision="main", # 或 version="0.1.0" |
| 238 | + device="cuda", |
| 239 | + mode="inference", |
| 240 | +) |
| 241 | + |
| 242 | +# 4) 应用函数内核 |
| 243 | +applied = apply_function_kernel( |
| 244 | + target_module=TARGET_MODULE, |
| 245 | + device="cuda", |
| 246 | + mode="inference", |
| 247 | + strict=False, |
| 248 | +) |
| 249 | +print("patched:", applied) |
| 250 | + |
| 251 | +# 5) 可选:通过 kernelize_model 统一应用 |
| 252 | +model = nn.Sequential(nn.Linear(8, 8), nn.ReLU()) |
| 253 | +kernelize_model(model=model, mode="inference", device="cuda", use_fallback=True) |
| 254 | +``` |
| 255 | + |
| 256 | +--- |
| 257 | + |
| 258 | +## 层级 + 函数级统一批量注册 |
| 259 | + |
| 260 | +### 适用场景 |
| 261 | + |
| 262 | +* 需要框架级统一集成 |
| 263 | +* 希望通过单一配置入口管理 |
| 264 | +* 同时管理层和函数两类内核 |
| 265 | + |
| 266 | +```python |
| 267 | +from twinkle.kernel import register_kernels, kernelize_model |
| 268 | +import torch.nn as nn |
| 269 | + |
| 270 | +# 1) 注册层级 + 函数级内核 |
| 271 | +config = { |
| 272 | + "layers": { |
| 273 | + "linear": { |
| 274 | + "repo_id": "kernels-community/linear", |
| 275 | + "layer_name": "Linear", |
| 276 | + "version": "0.1.0", |
| 277 | + "device": "cuda", |
| 278 | + "mode": "train", |
| 279 | + }, |
| 280 | + "conv2d": { |
| 281 | + "repo_path": "/path/to/local/repo", |
| 282 | + "package_name": "my_kernels", |
| 283 | + "layer_name": "Conv2d", |
| 284 | + "device": "cuda", |
| 285 | + }, |
| 286 | + }, |
| 287 | + "functions": { |
| 288 | + "add": { |
| 289 | + "target_module": "my_pkg.math_ops", |
| 290 | + "func_impl": lambda x, y: x + y + 1, |
| 291 | + "device": "cuda", |
| 292 | + "mode": "inference", |
| 293 | + }, |
| 294 | + "relu": { |
| 295 | + "target_module": "my_pkg.activations", |
| 296 | + "repo_id": "kernels-community/activation", |
| 297 | + "revision": "main", |
| 298 | + "device": "cuda", |
| 299 | + }, |
| 300 | + }, |
| 301 | +} |
| 302 | +register_kernels(config) |
| 303 | + |
| 304 | +# 2) 通过 kernelize_model 应用 |
| 305 | +model = nn.Sequential(nn.Linear(8, 8), nn.ReLU()) |
| 306 | +kernelize_model(model=model, mode="train", device="cuda", use_fallback=True) |
| 307 | +``` |
0 commit comments