Summary: 21 instances, 21 unique

Text	Count
# TODO: Add the QK dotprod to get scores	1
seed = SEEDS + col_id  # FIXME index the seed properly	1
# TODO: check this	1
# TODO add bias here	1
# TODO: flag use zeros for garbage	1
# TODO: tune these	1
# TODO: pass in key and value independently.	1
# TODO: check if need to return this or not	1
# TODO: this is here for BC	1
int64_t grain_size = 128; // TODO: tune this	1
# TODO: should be able to not have to pass in num_heads	1
# TODO: Build LUT	1
# TODO: update defaults for use_razavi_pinverse and inv_iterations	1
// TODO: compute grad only if they require grad	1
// TODO investigate misaligned address errors in values ptr	1
# TODO assume we have (N, S, hs) instead of (B, nh, S, hs), with N = B x nh	1
# TODO this is not always true, but is a fast approximation for now	1
# FIXME: Should we still accept a mask in that case ?	1
# TODO: flag use zeros for garbage	1
# TODO: link to a drawing of what these tensors are	1
# FIXME: @lefaudeux tensor shape changes are not well handled, see shape3	1