Issue on MacOS15.0

[erlebach]

I tried running train_mac.py on my mac running Sequoia OS. I got the error:

 python train_mac.py
training:   0%|                                                                                                                                                                  | 0/100000 [00:00<?, ?it/s]training loss: 5.670731067657471
validation loss: 5.639256954193115
%s

 %s ("ap w/ sling loop; Navy/&quot;SEF&quot; trigger group *'''MP5SD2''' â\x80\x94 integrated suppressor (''Scha", '****************************************************************************************************')
100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 412/412 [01:31<00:00,  4.50it/s]
Ø`c<U+0084><U+009C>5h® .4¨>v¯<U+0091>hÚÆÞ©<U+009F><U+0088><U+008C>tK2% ¼ âð<U+0099>><U+008D>¡0kè<U+0098>ß<U+0096>ı1®BÒ<U+0086>aÐa; <U+008C><U+009A>1[ÂKĹG³<U+0088> Ä ûn ß <U+008D>rØ\ø¦?þ  @iàU±´>À«ú <U+0<U+009C>1 ãÉR¨ÝåO;r9 p<U+008E>>¦TIV6a5<U+0096>¿<U+008D><U+009F>°å7kH¶[yseä}8<U+0083> ×>«)^r^¦2ÁS¥è<U+008C>Ø<U+008B>feU ¦{ãC ¨ñ$S\SÃQ<U+0085>I¥YÂ<U+008C>  ©ð T×Æ<U+008D>9O£Á¤ÇÒDáQE áóÞ$Æ5<U+009F>VU<U+009C>\<U+0099>{0<U+008F>}S<U+0083>É~Þ<U+009C>·ÞTs Õ¿ 3¸¿Ú <U+0091>8:à ¿t Øt b<U+008B> Ô<U+0085> ÎØ Ë¯FÀV4®Àáë ¿sJDÊà0p6ôÏDY_u:züzg<U+008B>>³a ?égdt£<U+0087>©õ Ø:d<U+0084>LrÛ¨<<U+0082>£3·f <U+009B>.  à 1^?âÚÄv<<U+0098>arüJ~ <U+0099>à   ¤ n<U+0096>Ý  aàNU<U+0096>hð<U+0080><U+009D>»<U+0088>é&¤¿Äeßu<U+0082><U+008B>.<U+008B><U+0087>IqÀÇÚü¾f<U+008B>³¤µ  j~0 \O¯ 9ŲÍV<U+0082><U+0083>D<U+008B>õ<U+0099> p-<U+009F>rmz<U<U+0091>
training:   0%|                                                                                                                                                    | 1/100000 [01:50<3064:23:04, 110.32s/it]training loss: 5.677858352661133
training:   0%|                                                                                                                                                    | 1/100000 [02:06<3521:39:51, 126.78s/it]
Traceback (most recent call last):
  File "/Users/erlebach/src/2024/titans-pytorch/train_mac.py", line 170, in <module>
    optim.step()
  File "/Users/erlebach/src/2024/titans-pytorch/.venv/lib/python3.10/site-packages/torch/optim/optimizer.py", line 493, in wrapper
    out = func(*args, **kwargs)
  File "/Users/erlebach/src/2024/titans-pytorch/.venv/lib/python3.10/site-packages/torch/utils/_contextlib.py", line 116, in decorate_context
    return func(*args, **kwargs)
  File "/Users/erlebach/src/2024/titans-pytorch/.venv/lib/python3.10/site-packages/adam_atan2_pytorch/adopt_atan2.py", line 126, in step
    p.add_(m * scale, alpha = -lr * a)
RuntimeError: unsupported operation: more than one element of the written-to tensor refers to a single memory location. Please clone() the tensor before performing the operation.

I read about the error, which is related to a tensor having overlapping memory locations. I am running on the CPU (so that could be an issue), and I disabled USE_ACCELERATED_SCAN, which requires CUDA. Note that I was able to train a single epoch.

Any help would be appreciated. Has anybody run the code on a CPU rather than with CUDA?

Gordon