Closed flbbb closed 1 year ago
Hi @flbbb ,
this is the standard behavior. Batch dimensions are, in the LazyTensor framework and by convention (which is quite common though), the first dimensions... If your data comes with a "batch last" setting, you should manually swap dimensions and make the data continuous (there can be a copy here).
NB : check the actual entry of the res
tensor in the "last batch" version of your code. My guess is that it does not give what you plan to compute...
Best,
Hi, I'm running the code below on a A5000 25Go. When the batch dimension
B
is on the last dimension (running with the script as is) it works.When running with
--swap
it fails withsegfault
after the backward call.Edit: Also tested on V100 32g and same issue.