Summary: 28 instances, 26 unique

Text	Count
// TODO: Templates for scenario and mode are overused, given the loadgen	1
# TODO: allow encoding of extra metadata	1
// TODO: Verify accuracy with the data copied here.	1
// TODO: Only set up the sample parenting once after all the queries have	1
pattern = '{folder}/*.log'.format(folder=benchmark_folder) # TODO allow other kinds of files	1
# TODO Benchmark may OOM stochastically - reduce it	1
# TODO measure at least avg latency	1
// TODO: Use PYBIND11_MAKE_OPAQUE for the following vector types.	1
// TODO: Convert this block to a simple write once we are confidient	2
// TODO: Convert this block to a simple write once we are confidient	1
# TODO: remove batch info args and migrate to record_batch_info	2
// TODO: Rate limit in accuracy mode so accuracy mode works even	1
# TODO: change name to FAMLogger	1
# TODO: fix units and percentile for latency	1
# TODO check this lr setup	1
# TODO use torchscript? jit/script this model?	1
# TODO: Don't I need to stop grad at some point now?	1
// TODO: std::this_thread::id is a class but MLPERF_GET_TID() assigned to	1
TODO was (seq_len, batch, in_features).	1
# TODO This kind of code probably exists somewhere? Fun to code but find a replacement.	1
/// TODO: Get rid of copies.	1
/// TODO: This isn't very useful anymore. Remove it.	1
# TODO We might also want to consider:	1
# TODO: Use half kwarg to generate half input data when appropriate	1
// TODO: Reserve in advance.	1
// TODO: Instead of these bools, use a class that forwards to two streams.	1