Support/AnyLayerTangentVector.swift (6 lines):
	- line 416: // TODO: Re-enable this once we have settled on a single toolchain.
	- line 421: // TODO: Re-enable this once we have settled on a single toolchain.
	- line 427: // TODO: Re-enable this once we have settled on a single toolchain.
	- line 432: // TODO: Re-enable this once we have settled on a single toolchain.
	- line 627: // TODO: Re-enable this once we have settled on a single toolchain.
	- line 633: // TODO: Re-enable this once we have settled on a single toolchain.


Models/Text/BERT/BERT.swift (6 lines):
	- line 63: // TODO: [AD] Avoid using token type embeddings for RoBERTa once optionals are supported in AD.
	- line 64: // TODO: [AD] Similarly for the embedding projection used in ALBERT.
	- line 78: // TODO: Convert to a generic constraint once TF-427 is resolved.
	- line 116: /// TODO: [DOC] Add a documentation string and fix the parameter descriptions.
	- line 212: // TODO: Make dropout generic over the probability type.
	- line 220: // TODO: [AD] Change to optional once supported.


Support/Image.swift (5 lines):
	- line 71: // TODO: Add BGR byte reordering.
	- line 75: // TODO: Proper error propagation for this.
	- line 83: // TODO: Proper error propagation for this.
	- line 120: // TODO: Proper error propagation for this.
	- line 127: // TODO: Proper error propagation for this.


pix2pix/CheckpointReader.swift (4 lines):
	- line 57: // TODO: Come up with better names for these protocols
	- line 80: // TODO: read/write activation, strides, padding, and dialations from checkpoint file
	- line 109: // TODO: read/write activation, strides, and padding from checkpoint file
	- line 164: // TODO: Convert this to suitable protocol with Discriminator configuration


Support/STBImage/include/stb_image.h (3 lines):
	- line 1180: // @TODO: move stbi__convert_format to here
	- line 1204: // @TODO: move stbi__convert_format16 to here
	- line 1205: // @TODO: special case RGB-to-Y (and RGBA-to-YA) for 8-bit-to-16-bit case to keep more precision


Models/Text/BERT/TransformerBERT.swift (3 lines):
	- line 48: // TODO: Convert to a generic constraint once TF-427 is resolved.
	- line 179: // TODO: Convert to a generic constraint once TF-427 is resolved.
	- line 278: // TODO: Make dropout generic over the probability type.


Checkpoints/CheckpointIndexWriter.swift (3 lines):
	- line 19: // TODO: Extend handling to different tensor types.
	- line 35: // TODO: Expand beyond using a single binary shard.
	- line 143: // TODO: Find a more efficient way of calculating this without casting to bytes twice.


Models/Text/WordSeg/Model.swift (3 lines):
	- line 28: /// TODO: Consider promoting this into a public protocol in swift-apis?
	- line 265: // TODO: avoid copies?
	- line 342: // TODO: Remove this when differentiation uses per-instance zeros


Support/GIF.swift (3 lines):
	- line 71: // FIXME: Unless the palette length is 256xRGB the GIF is invalid for some reason...
	- line 127: // TODO: Optimize bounding box to only provided updated part.
	- line 215: // TODO: Specify number of loops, rather than looping forever.


Checkpoints/CheckpointWriter.swift (3 lines):
	- line 19: // TODO: Extend handling to different tensor types.
	- line 54: // TODO: Handle splitting into multiple shards.
	- line 68: // TODO: Write this directly to disk, rather than accumulating it in memory.


Support/FoundationFileSystem.swift (2 lines):
	- line 49: // TODO: Incorporate file offset.
	- line 58: // TODO: Incorporate file offset.


Models/Recommendation/DLRM.swift (2 lines):
	- line 94: // TODO: loss threshold clipping
	- line 154: // TODO: remove computeEmbeddingsVJP once inout differentiation is supported!


Support/Text/BytePairEncoder.swift (2 lines):
	- line 23: // TODO: Find a nice way to support caching.
	- line 154: // TODO: Add documentation.


Models/ImageClassification/ResNet.swift (2 lines):
	- line 57: // TODO: Replace the following, so as to not waste memory for non-projection cases.
	- line 89: // TODO: Find a way for this to be checked only at initialization, not during training or


Models/Text/BERT/Attention.swift (2 lines):
	- line 77: // TODO: Convert to a generic constraint once TF-427 is resolved.
	- line 158: // TODO: Make dropout generic over the probability type.


Models/Text/WordSeg/Lattice.swift (2 lines):
	- line 136: // TODO: Reduceinto and +=
	- line 332: // TODO: figure out why the string equality is being ignored


pix2pix/Generator.swift (1 line):
	- line 22: // TODO: need to persist a sample generator and set this URL


Support/Text/TextBatch.swift (1 line):
	- line 41: // TODO: Use parallelism to grab the samples in parallel.


SwiftModelsBenchmarksCore/BenchmarkSettings.swift (1 line):
	- line 106: // TODO: Remove the following tensor workaround when above is fixed.


Datasets/CoLA/CoLA.swift (1 line):
	- line 134: // FIXME: Need to generalize `DatasetUtilities.downloadResource` to accept


pix2pix/Layers.swift (1 line):
	- line 219: // TODO: need to persist the activation function.  Until then, manually set it to tanh.


SwiftModelsBenchmarksCore/Models/ImageClassificationTraining.swift (1 line):
	- line 37: // TODO: Split out the optimizer as a separate specification.


Support/BijectiveDictionary.swift (1 line):
	- line 19: // TODO: specialize where value is Int32 such that keyForValue is actually an


Support/AnimatedImage.swift (1 line):
	- line 63: // TODO: Adapt the following to the colors in the input image.


Models/Text/GPT2/GPT2.swift (1 line):
	- line 102: // TODO: Add argument that controls this.


Support/AnyLayer.swift (1 line):
	- line 15: // TODO: Re-enable this for the stock toolchain when it can be realigned with VectorProtocol.


Models/Recommendation/NeuMF.swift (1 line):
	- line 76: // TODO: regularization