Fast-Reformer
所属分类:推荐系统
开发工具:Cuda
文件大小:0KB
下载次数:0
上传日期:2022-08-22 05:53:35
上 传 者:
sh-1993
说明: 具有cuda实现的重整器(),
(Reformer () with cuda implementation,)
文件列表:
CMakeLists.txt (921, 2022-08-21)
LICENSE (11357, 2022-08-21)
make.sh (86, 2022-08-21)
pybind11/ (0, 2022-08-21)
reformer/ (0, 2022-08-21)
reformer/inference/ (0, 2022-08-21)
reformer/inference/CMakeLists.txt (159, 2022-08-21)
reformer/inference/kernels.cu (37868, 2022-08-21)
reformer/inference/kernels.cuh (3313, 2022-08-21)
reformer/inference/models.cu (42448, 2022-08-21)
reformer/inference/utils.cuh (1675, 2022-08-21)
test.sh (33, 2022-08-21)
tests/ (0, 2022-08-21)
tests/__init__.py (0, 2022-08-21)
tests/config.json (1322, 2022-08-21)
tests/cpp/ (0, 2022-08-21)
tests/cpp/CMakeLists.txt (251, 2022-08-21)
tests/cpp/test_kernels.cu (7329, 2022-08-21)
tests/cpp/test_models.cu (7616, 2022-08-21)
tests/profile_fp16.py (4219, 2022-08-21)
tests/profile_fp32.py (4198, 2022-08-21)
tests/test_fp16.py (8991, 2022-08-21)
tests/test_fp32.py (8765, 2022-08-21)
# Fast-Reformer
[Reformer](https://arxiv.org/abs/2001.04451) encoder (inference only) with cuda implementation
for my cuda practice
The test target is [huggingface reformer](https://github.com/huggingface/transformers/blob/main/src/transformers/models/reformer/modeling_reformer.py)
```bash
pip install transformers
git clone --recursive https://github.com/hutao965/Fast-Reformer.git
cd Fast-Reformer
sh make.sh
sh test.sh
```
# Goal
- [x] reformer model
- [x] axial embedding
- [x] chunk ffn
- [x] local atten
- [x] lsh atten
- [x] block radix sort
- [x] unit test
- [x] pybind
- [x] profiling
- [x] fp32 speed up 38.63%
- [x] fp16 speed up 40.64%
# Ref
- [lightseq](https://github.com/bytedance/lightseq)
- [block radix sort in cub](https://github.com/NVIDIA/cub/blob/main/cub/block/block_radix_sort.cuh)
近期下载者:
相关文件:
收藏者: