Rhythmblue
commented
3 years ago I add some lines in CMakeLists for RTX3090+CUDA11.1+pytorch1.7.1.
IF (CUDA_VERSION GREATER 8.9)
set(CUDA_NVCC_FLAGS "${CUDA_NVCC_FLAGS} -gencode arch=compute_70,code=sm_70")
ENDIF()
IF (CUDA_VERSION GREATER 9.9)
set(CUDA_NVCC_FLAGS "${CUDA_NVCC_FLAGS} -gencode arch=compute_75,code=sm_75")
ENDIF()
IF (CUDA_VERSION GREATER 10.9)
set(CUDA_NVCC_FLAGS "${CUDA_NVCC_FLAGS} -gencode arch=compute_80,code=sm_80")
ENDIF()
if (NOT APPLE)
set(CUDA_NVCC_FLAGS "${CUDA_NVCC_FLAGS} --std=c++14")
set(CUDA_NVCC_FLAGS "${CUDA_NVCC_FLAGS}")
ENDIF() It works for the following testing codes,
import torch
import warpctc_pytorch as warp_ctc
def test_simple():
probs = torch.FloatTensor([[[0.1, 0.6, 0.1, 0.1, 0.1], [0.1, 0.1, 0.6, 0.1, 0.1]]]).transpose(0, 1).contiguous()
grads = torch.zeros(probs.size())
labels = torch.IntTensor([1, 2])
label_sizes = torch.IntTensor([2])
sizes = torch.IntTensor(probs.size(1)).fill_(probs.size(0))
minibatch_size = probs.size(1)
costs = torch.zeros(minibatch_size)
warp_ctc.cpu_ctc(probs,
grads,
labels,
label_sizes,
sizes,
minibatch_size,
costs,
0)
print('CPU_cost: %f' % costs.sum())
probs = probs.clone().cuda()
grads = torch.zeros(probs.size()).cuda()
costs = torch.zeros(minibatch_size)
warp_ctc.gpu_ctc(probs,
grads,
labels,
label_sizes,
sizes,
minibatch_size,
costs,
0)
print('GPU_cost: %f' % costs.sum())
print(grads.view(grads.size(0) * grads.size(1), grads.size(2)))
if __name__ == '__main__':
test_simple()Output:
CPU_cost: 2.462858
GPU_cost: 2.462858
tensor([[ 0.1770, -0.7081, 0.1770, 0.1770, 0.1770],
[ 0.1770, 0.1770, -0.7081, 0.1770, 0.1770]], device='cuda:0')
Luna-yu
我使用 from warp_ctc import CTCLoss 之后，报错：
ImportError: ....... warp-ctc-pytorch_bindings/pytorch_binding/warpctc_pytorch/_warp_ctc.cpython-36m-x86_64-linux-gnu.so: undefined symbol: _ZN3c1011CPUTensorIdEv我之前是能用的。然后我用nvidia-smi查看cuda版本为11.1 ，请问是否不支持cuda11.1？期待解答，不胜感激！