RuntimeError: CUDA error:

[Usaywook]

When I tried to implement this code with the below command, I got this error.

• command

python Evaluation.py ./config/demo.py --model-name demo

• error

gpu number:1
model loaded from ./models/demo.pt
Successfully Loaded model: ./models/demo.pt
Finished Initialization in 15.365s!!!
  0%|                                                                                                                                                                                                                                                            | 0/1234 [00:00<?, ?it/s]
Traceback (most recent call last):
  File "Evaluation.py", line 77, in <module>
    out = model(data)
  File "/home/usaywook/anaconda3/envs/mmTrans/lib/python3.7/site-packages/torch/nn/modules/module.py", line 889, in _call_impl
    result = self.forward(*input, **kwargs)
  File "/media/usaywook/Samsung_T5/tmp/mmTransformer/lib/models/mmTransformer.py", line 150, in forward
    social_mask, lane_enc, lane_mask)
  File "/home/usaywook/anaconda3/envs/mmTrans/lib/python3.7/site-packages/torch/nn/modules/module.py", line 889, in _call_impl
    result = self.forward(*input, **kwargs)
  File "/media/usaywook/Samsung_T5/tmp/mmTransformer/lib/models/TF_version/stacked_transformer.py", line 128, in forward
    lane_mem = self.lane_enc(self.lane_emb(lane_enc), lane_mask) # (batch size, max_lane_num, 128)
  File "/home/usaywook/anaconda3/envs/mmTrans/lib/python3.7/site-packages/torch/nn/modules/module.py", line 889, in _call_impl
    result = self.forward(*input, **kwargs)
  File "/media/usaywook/Samsung_T5/tmp/mmTransformer/lib/models/TF_utils.py", line 49, in forward
    x = layer(x, x_mask)
  File "/home/usaywook/anaconda3/envs/mmTrans/lib/python3.7/site-packages/torch/nn/modules/module.py", line 889, in _call_impl
    result = self.forward(*input, **kwargs)
  File "/media/usaywook/Samsung_T5/tmp/mmTransformer/lib/models/TF_utils.py", line 69, in forward
    x = self.sublayer[0](x, lambda x: self.self_attn(x, x, x, mask))
  File "/home/usaywook/anaconda3/envs/mmTrans/lib/python3.7/site-packages/torch/nn/modules/module.py", line 889, in _call_impl
    result = self.forward(*input, **kwargs)
  File "/media/usaywook/Samsung_T5/tmp/mmTransformer/lib/models/TF_utils.py", line 208, in forward
    return x + self.dropout(sublayer(self.norm(x)))
  File "/media/usaywook/Samsung_T5/tmp/mmTransformer/lib/models/TF_utils.py", line 69, in <lambda>
    x = self.sublayer[0](x, lambda x: self.self_attn(x, x, x, mask))
  File "/home/usaywook/anaconda3/envs/mmTrans/lib/python3.7/site-packages/torch/nn/modules/module.py", line 889, in _call_impl
    result = self.forward(*input, **kwargs)
  File "/media/usaywook/Samsung_T5/tmp/mmTransformer/lib/models/TF_utils.py", line 170, in forward
    query, key, value, mask=mask, dropout=self.dropout)
  File "/media/usaywook/Samsung_T5/tmp/mmTransformer/lib/models/TF_utils.py", line 227, in attention
    scores = torch.matmul(query, key.transpose(-2, -1)) / math.sqrt(d_k)
RuntimeError: CUDA error: CUBLAS_STATUS_EXECUTION_FAILED when calling `cublasSgemmStridedBatched( handle, opa, opb, m, n, k, &alpha, a, lda, stridea, b, ldb, strideb, &beta, c, ldc, stridec, num_batches)`

If I had removed .cuda in line 61 and line 75 from this code, I could resolve the error. However, I cannot use the GPU to implement this code.

Moreover, in this repository, I cannot find the loss function to consider multimodal trajectories. Could you share the code for the loss function used in the original paper?