[BUG] BlockFusion bug in model training after PR#242

[xysmlx]

Test case: mnist training, lstm training

[nnfbot]

Thanks for the report @xysmlx! I will look into it ASAP! (I'm a bot).

[xysmlx]

Update: BlockFusion fuses dead gnodes, while the memory allocator does not allocate memory for dead tensors (i.e., calculating the needed output tensors does not need these dead gnodes and dead tensors). This leads to memory illegal accesses and results in such training problems.