torch_xla/csrc/aten_xla_type.cpp (3 lines):
	- line 1315: // TODO: for now route to native, which dispatches supported XLA operations.
	- line 2852: // TODO: implement scatter_mul
	- line 2872: // TODO: implement scatter_mul


torch_xla/csrc/cross_replica_reduces.cpp (2 lines):
	- line 91: // TODO: We use pseudo-tokens ATM, which are real values. This need to be
	- line 162: // TODO: This is missing layout pinning ATM. If XLA scheduling is not exactly


torch_xla/csrc/tensor.cpp (1 line):
	- line 1072: // TODO: This can be optimized via proper XRT/XLA computation.