[Museformer] Errors about Ninja when inference about muzic

[Alpha-Half]

I tried editing PyTorch to solve "ninja, -v" by replacing with "ninja, --version", but it followed
ImportError: /root/.cache/torch_extensions/py38_cu113/range_fill_cuda/range_fill_cuda.so: cannot open shared object file: No such file or directory
which is the same as
https://github.com/microsoft/muzic/issues/110
but I've pip installed ninja

[Alpha-Half]

@btyu Hi there
Compilation problem solved by using docker. But still cannot generate.
When inference, stuck at: (w/ no error reported and it takes forever)
2023-06-18 15:43:48 | INFO | fairseq_cli.interactive | Type the input sentence and press return:
So I run the evaluation, w/ log:
2023-06-18 15:03:30 | INFO | museformer.museformer_lm_task | loaded 2155 samples for test 2023-06-18 15:11:36 | INFO | test | | valid on 'test' subset | loss 3.337 | ppl 10.11 | wps 0 | wpb 2.14513e+07 | bsz 2155
My system is Ubuntu 20.04 LTS 64 with CUDA 11.4.3, cnDNN 8.2.4, GPU drive ver. 470.82.01
With docker env, I use default checkpoint and src code. I preprocess midi on my own (28251 midi) and rewrite the metadata to train (24,580 lines), test (2,155) and valid (1,516). Since I only compress the track to 6, so no various time signatures or instruments.
And when fairseq-interactive, GPU's not fully used so I guess it's not simply slow (wait for up to an hour QAQ

[Alpha-Half]

With respect to results, it's similar with #113. So there might be some problem in arbitrary dataset I guess...?