[manyblock mode] why is "act_out" in "nv_wavenet_persistent_GEMM_MxK" necessary?

[isaacleeai]

since we are storing the same values in the last layer of "accum_in", and we are actually using "accum_in" to pass values from skip_block to Zs block, Zs block to Za block, and Za to softmax, why do we even have act_out? Please tell me if I am wrong, but it seems unnecessary.

[isaacleeai]

Commented out this single line in nv_wavenet_persistent_GEMM_MxK in nv_wavenet_persistent.cuh
act_out[layer*ldc + row] = accum[0];
and ran nv_wavenet_test for manyblock with batch_size = 0, R = 64, A = 256, S = 256, num_layers = 16 ( with comparing Za and Zs commented out as well )
Except for not checking Za and Zs, everything else was correct, in manyblock mode ( but since persistent is almost like manyblock, they should both work, although I can't test persistent on the GPU I have ).
It seems like act_out is unnecessary... Let me know what you think!