RuntimeError: stack expects a non-empty TensorList

[hjx999222]

When running your code, I encounter some issues, and sometimes it results in the following bug. Could you please help me resolve it? Thank you!

Traceback (most recent call last):
File "DRAGIN_simplified/src/main.py", line 90, in
main()
File "DRAGIN_simplified/src/main.py", line 76, in main
pred = model.inference(entry["question"], entry["demo"], entry["case"])
File "DRAGIN_simplified/src/generate.py", line 433, in inference
outputs = self.generator.generate(
File "DRAGIN_simplified/src/generate.py", line 240, in generate
atten = torch.stack([atten[l:r, :].mean(dim=-2) for l, r in range_], dim=-2) # 同组的注意者(token)，对同一被注意者(word)的attention取平均
RuntimeError: stack expects a non-empty TensorList

[Bocchi7]

It seems that the list range_ is empty, which means the new generated tokens are empty (after deleting the last eos_token). Check if the program enters the branch in line 214 and the list tokens become empty.

        if ended:
            tokens = tokens[:-1]
            # 注：如果tokens中有"</s>"（"<s>"也一样），转换成字符串后仍然会保留。因此在这里删掉。

When the situation occurs, you may return the GeneratorOutput instance immediately (see the format as line 267). Remember to set ended as True.