Can someone explain how the memory is allocated for alpha, beta and log_probs2?

[ronggong]

warp-transducer/include/detail/cpu_rnnt.h

Line 103 in f546575

alphas = reinterpret_cast<ProbT *>(static_cast<char *>(workspace) + bytes_used);

Hi,

I am looking for an explanation of L103 to L110. Afaiu, workspace is the expected memory size to allocate for all batch, see below code:

warp-transducer/src/rnnt_entrypoint.cpp

Line 125 in f546575

*size_bytes = per_minibatch_bytes * minibatch;

and byte_used is the allocated memory size. I could not understand, why in the code, we allocate workspace + byte_used memory size for alpha, beta and log_probs2? The reason, from my understanding, workspace is already the memory size for the entire batch, so the allocation here is redundant?

Cheers,