Summary: 102 instances, 64 unique

Text	Count
//TODO: need to make sure opId is not an output layer of the model	1
#FIXME MaxPool2d supports ceil_mode	2
// FIXME: These 2 functions could execute independently in parallel	1
// TODO: remove this line	1
// FIXME: currently we zero init the region if not read output	1
#TODO: target shape does not support -1	2
// TODO: to be removed when we have attention layers	1
#FIXME assume there is 1 output	2
//TODO: implement	4
// FIXME: Currently assume 1st input for 1st layer = batch_size	1
#FIXME fix kernel, stride and padding	2
// TODO: why cublas/cudnn stream is needed here?	1
# TODO: input shape should not contain batch size for now	2
// TODO: only support relu and sigmoid for now	1
//TODO: change to use output instead of recomputing	1
//TODO: currently do not support splitting over the channel dimension	1
//FIXME: for now only consider i * j == NUM_PARTITIONS	1
//TODO: check data type matches	2
//FIXME: uncomment us	1
//TODO: implement measure_forward	1
# TODO: automatically call this function	2
# TODO: finish API	2
// FIXME: Legion tracing currently does not support MUST_EPOCH	1
// TODO: Use index launcher instead of task launcher	4
//TODO: need to make sure opId is not an output layer of the model	1
_remove_long_seq = sequence._remove_long_seq  # TODO: make it public?	2
// FIXME: For now, set upper limits Better: Do as follows, but memory is	3
// TODO: missing profiling here	1
// FIXME: even though it is a CPU task, we use data parallelism	2
//TODO: delete fused_op to avoid memory leakage	1
// FIXME: currently assume the final layer has exactly one output	1
// TODO: why cublas/cudnn stream is needed here	1
//FIXME: currently assume continous indices	4
// TODO: Check why cublas/cudnn stream is needed here	1
//TODO: dataloader does not support CR	1
//TODO: Currently we use a task launch, change to index launch for NCCL parameter	1
int k = out1_domain.hi()[0] - out1_domain.lo()[0] + 1; /*TODO: This prints to 5*/	1
# TODO: move check shape into another function	4
// FIXME: currently create gradients for constants since the current auto grad algorithm	1
//TODO: swich to use the Legion dim ordering	1
// TODO: perform prefetch for performance imporvement	1
// TODO add parameter synchronization time	1
# TODO: add unsqueeze	2
Tensor label_tensor_with_final_part;//FIXME: to be removed	1
assert(_input.numDim == 2); // TODO: support dims > 2	1
// TODO: assign priorities	1
//TODO: remove data loaders except single data loader	1
// TODO: add support for meta->a_seq_length_dim >= 0	1
// FIXME: Currently only support the sample dimension for operators with NCCL	1
if(l == metrics_input && metrics_input < (int)layers.size()-1) continue; // TODO: If layer serves for metrics and for further prop	1
// FIXME: Need to be atomic depending on the strategy	1
#TODO: finish API	6
// TODO: temp work, will let users to pick either NCCL or PS	1
#FIXME may be 3	2
#TODO: seperate compute_metrics from backward	2
//TODO: consider reduction dim for conv2d and linear	1
#FIXME assume it is a merge	2
# TODO: add range	2
// FIXME: it seems curand has an internal bug with volume < 4	1
# FIXME BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) args are not in FF	2
//TODO: implement broadcast op	2
# TODO: add cast	2
#TODO: this path has not tested	2
//FIXME: Not functionaly correct.	1