Sources/TensorFlow/Operators/Basic.swift (8 lines):
	- line 579: // TODO: precondition(axis >= 0 && axis < rank, "'axis' is out of range.")
	- line 580: // TODO: precondition(batchDimensionCount <= axis,
	- line 911: // TODO: What about precedence? Why is this operator used for broadcasting?
	- line 945: // TODO: Simplify this once differentiating control flow is supported.
	- line 1096: // TODO: Negative indexing and strides syntax.
	- line 1107: // TODO: Precondition `lowerBounds.count == upperBounds.count`,
	- line 1109: // TODO: Differentiating control flow is not supported yet, thus the thunks.
	- line 1197: // TODO: Cannot extend non-nominal type 'UnboundedRange'.


Sources/TensorFlow/Operators/NN.swift (5 lines):
	- line 634: // TODO: Currently this is not higher order differentiable. Redefine in
	- line 694: // TODO: Currently this is not higher order differentiable. Redefine in
	- line 752: // TODO: Currently this is not higher order differentiable. Redefine in
	- line 812: // TODO: Currently this is not higher order differentiable. Redefine in
	- line 888: // TODO: Currently this is not higher order differentiable. Redefine in


Sources/TensorFlow/Operators/Math.swift (4 lines):
	- line 23: // TODO:
	- line 27: // TODO: Remove the following extension once `./` and `./=` are defined for
	- line 899: // FIXME: Scoped imports are not yet supported in parseable module interfaces, so
	- line 2727: // TODO: Consider making the return type be generic over `FloatingPoint` types


Sources/TensorFlow/Epochs/NonuniformTrainingEpochs.swift (4 lines):
	- line 46: // TODO: Figure out how to handle non-threasafe PRNGs with a parallel shuffle
	- line 94: // TODO: use a parallel shuffle like mergeshuffle
	- line 120: // TODO: fully sorting is overkill; we should use introselect here.
	- line 218: // TODO: Test the laziness of the result.


Sources/x10/xla_tensor/cross_replica_reduces.cpp (3 lines):
	- line 105: // TODO: We use pseudo-tokens ATM, which are real values. This need to be
	- line 147: // TODO: This is missing layout pinning ATM. If XLA scheduling is not exactly
	- line 162: // TODO: This is missing layout pinning ATM. If XLA scheduling is not exactly


Sources/x10/xla_tensor/tensor.cpp (2 lines):
	- line 483: // LOG(FATAL) << "TODO check device";
	- line 943: // TODO: This can be optimized via proper XRT/XLA computation.


Sources/TensorFlow/Core/MixedPrecision.swift (2 lines):
	- line 155: // TODO (SR-12968): Mark `tensor` with `@noDerivative` and remove custom vjp below.
	- line 165: // TODO (SR-12968): Remove when `tensor` can be marked `@noDerivative` in `init`.


Sources/TensorFlow/Core/Runtime.swift (2 lines):
	- line 57: // TODO: change this and subsequent properties from static to thread local.
	- line 329: // TODO: Can we interop between modes?


Sources/TensorFlow/Epochs/TrainingEpochs.swift (2 lines):
	- line 36: // TODO: Figure out how to handle non-threasafe PRNGs with a parallel shuffle
	- line 68: // TODO: use a parallel shuffle like mergeshuffle


Sources/TensorFlow/Core/LazyTensorTrace.swift (1 line):
	- line 65: // TODO: We only pick operations on which `lazyOp` depends on. Note that


Sources/TensorFlow/Bindings/generate_wrappers.py (1 line):
	- line 745: # FIXME: Re-add `@_frozen` after SR-9739 is resolved.


Sources/TensorFlow/Core/Utilities.swift (1 line):
	- line 158: // TODO: Consider revising the call sites where this is necessary to only need UnsafeMutablePointer


Sources/x10/swift_bindings/optimizers/Optimizer.swift (1 line):
	- line 102: // TODO: Experiment with efficiently fusing these...


Sources/TensorFlow/Operators/Comparison.swift (1 line):
	- line 148: // TODO: infix operator ≈: ComparisonPrecedence


Sources/TensorFlow/Operators/Dataset.swift (1 line):
	- line 29: // TODO: There's no support for TF's "global seed" yet, so we always use the default graph seed as


Sources/TensorFlow/StdlibExtensions.swift (1 line):
	- line 284: // FIXME: `one` should probably be removed from the protocol. `Array` cannot represent `one`.


Sources/TensorFlow/X10/Device.swift (1 line):
	- line 149: // TODO: Pull this from withDevice() {} mechanism?


Sources/TensorFlow/Epochs/Algorithms.swift (1 line):
	- line 193: // FIXME: this algorithm should be benchmarked on arrays against


Sources/CX10/functional_while.cc (1 line):
	- line 49: // TODO: color when building the graph as this can be n^2


Sources/TensorFlow/Epochs/Collatable.swift (1 line):
	- line 31: // TODO: derived conformance


Sources/Tensor/Random.swift (1 line):
	- line 511: // FIXME: Box-Muller can generate two values for only a little more than the


Sources/x10/swift_bindings/Device.swift (1 line):
	- line 149: // TODO: Pull this from withDevice() {} mechanism?


Sources/TensorFlow/Core/LazyTensorTraceCache.swift (1 line):
	- line 116: // TODO: we might avoid running the following check based on results of promotableConstant