新增功能
- 全面支持 HuggingFace 模型,包括Bert、Albert等热点模型以及对应不同任务类型,共55种 TorchScript 以及40种 ONNX 模型
- 针对 Bert结构 ONNX模型,新增 Gelu pass,预测速度提升20%以上 https://github.com/PaddlePaddle/X2Paddle/pull/809
- 针对 Bert结构 ONNX模型,新增 div2scale pass,使其命中PaddleInference multihead_matmul_fuse_pass,性能提升30%以上 https://github.com/PaddlePaddle/X2Paddle/pull/813
新增模型支持
PyTorch->Paddle
- MobileNetv3
- HuggingFace T5
- HuggingFace Marian
- HuggingFace Bart
- HuggingFace wav2vec2
- HuggingFace BigBird
ONNX->Paddle
新增op
- 新增PyTorch op(7个):
aten::rsqrt
、aten::replication_pad1d
、aten::full
、aten::group_norm
、aten::argmax
、aten::copy
、aten::linear
https://github.com/PaddlePaddle/X2Paddle/pull/783 https://github.com/PaddlePaddle/X2Paddle/pull/781 https://github.com/PaddlePaddle/X2Paddle/pull/786 https://github.com/PaddlePaddle/X2Paddle/pull/791 https://github.com/PaddlePaddle/X2Paddle/pull/789 https://github.com/PaddlePaddle/X2Paddle/pull/769 - 新增ONNX op(2个):
CumSum
、LessOrEqual
https://github.com/PaddlePaddle/X2Paddle/pull/793 https://github.com/PaddlePaddle/X2Paddle/pull/794
修复op
- 修复PyTorch op(2个):
aten::masked_fill
、aten::index
https://github.com/PaddlePaddle/X2Paddle/pull/786 https://github.com/PaddlePaddle/X2Paddle/pull/787 - 修复ONNX op(12个):
ConvTranspose
、Unsqueeze
、MatMul
、Gather
、Slice
、Squeeze
、Flatten
、Constant
、ConstantOfShape
、Where
、Topk
、NonZero
https://github.com/PaddlePaddle/X2Paddle/pull/792 https://github.com/PaddlePaddle/X2Paddle/pull/794 https://github.com/PaddlePaddle/X2Paddle/pull/796 https://github.com/PaddlePaddle/X2Paddle/pull/797 https://github.com/PaddlePaddle/X2Paddle/pull/805 https://github.com/PaddlePaddle/X2Paddle/pull/808
修复功能
- 修复LN fuser pass https://github.com/PaddlePaddle/X2Paddle/pull/777
- 修复Linear fuser pass https://github.com/PaddlePaddle/X2Paddle/pull/800
其他
- 在转换完成之后添加问卷链接 https://github.com/PaddlePaddle/X2Paddle/pull/801
- 清除
paddle.fluid
接口 https://github.com/PaddlePaddle/X2Paddle/pull/806