"python ./lora.py" crashed on "RuntimeError: view size is not compatible with input tensor's size and stride"

[jeffDebug]

I was trying to increase the batch size in lora.py, changing the following line (36):

BATCH_SIZE = 2

Then run:

python ./lora.py

It crashed with the following backtrace:

ERROR:main:Error in batch 172: view size is not compatible with input tensor's size and stride (at least one dimension spans across two contiguous subspaces). Use .reshape(...) instead.
ERROR:main:Traceback (most recent call last):
File "/home/jeff/csm-streaming/./lora.py", line 915, in finetune
loss = single_pass_forward(model, bridging_module,
File "/home/jeff/csm-streaming/./lora.py", line 706, in single_pass_forward
loss = compute_loss_for_codebooks_single_pass(
File "/home/jeff/csm-streaming/./lora.py", line 648, in compute_loss_for_codebooks_single_pass
flat_dec = decoder_out.view(bsz * seq_len, -1)
RuntimeError: view size is not compatible with input tensor's size and stride (at least one dimension spans across two contiguous subspaces). Use .reshape(...) instead.

Any ideas/suggeations to fix this? Thanks!

[matthen]

Change decoder_out.view(bsz * seq_len, -1) to decoder_out.reshape(bsz * seq_len, -1)