search

DerryHub / BEVFormer_tensorrt

BEVFormer inference on TensorRT, including INT8 Quantization and Custom TensorRT Plugins (float/half/half2/int8).
Apache License 2.0
410 stars 67 forks source link

what's the difference between mmdeploy plugins and custom plugins ? #68

Closed reverse-2020 closed 1 year ago

reverse-2020 commented 1 year ago

i find theri onnx is different , mmdeploy plugins seems not have the op MultiScaleDeformableAttnTRT ,how does it works？

DerryHub commented 1 year ago

MultiScaleDeformableAttnTRT(2) support half(half2)/int8 while the mmdeploy plugins don't support. You can refer https://github.com/DerryHub/BEVFormer_tensorrt/tree/main/TensorRT#multi-scale-deformable-attention to use it.

reverse-2020 commented 1 year ago

MultiScaleDeformableAttnTRT(2) support half(half2)/int8 while the mmdeploy plugins don't support. You can refer https://github.com/DerryHub/BEVFormer_tensorrt/tree/main/TensorRT#multi-scale-deformable-attention to use it.

thanks for reply , i have readed this md, my doubt is when i export bevformer_tiny.pth for mmdeploy plugin onnx , i don't see the plugin op of MultiScale deformable attention , which should export failed ,and the ops of trt in mmdeploy 0.10.0 dont have the plugin of MultiScale deformable attention yet , but this onnx convert to trt successful ,if i am wrong ,please tell me

reverse-2020 commented 1 year ago

MultiScaleDeformableAttnTRT(2) support half(half2)/int8 while the mmdeploy plugins don't support. You can refer https://github.com/DerryHub/BEVFormer_tensorrt/tree/main/TensorRT#multi-scale-deformable-attention to use it.

Deformable attention can be split into several grid_sample operators, is that right?

DerryHub commented 1 year ago

MultiScaleDeformableAttnTRT(2) support half(half2)/int8 while the mmdeploy plugins don't support. You can refer https://github.com/DerryHub/BEVFormer_tensorrt/tree/main/TensorRT#multi-scale-deformable-attention to use it.

Deformable attention can be split into several grid_sample operators, is that right?

Yes, you are right.

reverse-2020 commented 1 year ago

MultiScaleDeformableAttnTRT(2) support half(half2)/int8 while the mmdeploy plugins don't support. You can refer https://github.com/DerryHub/BEVFormer_tensorrt/tree/main/TensorRT#multi-scale-deformable-attention to use it.

Deformable attention can be split into several grid_sample operators, is that right?

Yes, you are right.

okay, I see

admyxs commented 6 months ago

MultiScaleDeformableAttnTRT(2) support half(half2)/int8 while the mmdeploy plugins don't support. You can refer https://github.com/DerryHub/BEVFormer_tensorrt/tree/main/TensorRT#multi-scale-deformable-attention to use it.

Deformable attention can be split into several grid_sample operators, is that right?

Yes, you are right.

thank for your great repo

I want to know how you do to make Deformable attention split into several grid_sample operators,

or MMDeploy help us to do this?