Closed jylink closed 2 years ago
fixed, set_ifndef(CUDA_ARCH_SM 70)
is not enough, nvcc gencode flags are required
### alonet/torch2trt/plugins/ms_deform_im2col/CMakeLists.txt
set_ifndef(CUDA_ARCH_SM 70) # should be fine for Tesla V100
...
# Link TensorRT's nvinfer lib
target_link_libraries(ms_deform_im2col_trt PRIVATE ${NVINFER_LIB})
# NEW
target_compile_options(ms_deform_im2col_trt PRIVATE $<$<COMPILE_LANGUAGE:CUDA>:
-gencode=arch=compute_70,code=sm_70
>)
I'm trying to convert deformable detr to trt, but got these errors when I run
load_trt_plugins_for_deformable_detr()
It looks like a cuda arch problem but my gpu compute capability looks fine...
FYI