Source
Chen et al. (2018), LIT-063 — ARXIV-1802.04799.
This is the practice its source argues against
LIT-063’s premise is that hand-optimisation does not scale. Frameworks lean on vendor-specific operator libraries tuned for a narrow range of server-class GPUs, so every new target — mobile, embedded, FPGA, ASIC — costs a manual re-optimisation, and TVM exists to replace that manual work with a compiler and a search.
“Implement custom kernels for critical ops” is the manual work. Citing this paper for it credits the argument to the position it was written against.
What survives, and what it should say
Something real does survive, and it is narrower: there are operations no compiler covers well, and for those a hand-written kernel is still the answer. The fused attention kernels in SOTA-085 and SOTA-106 are exactly that case, and they are hand-written for reasons a search does not reach — the algorithm is different, not just the schedule.
So the honest form is a fallback rather than a default: automate first, measure, and hand-write the few operations where the generated code is materially short of what the hardware can do. Stated as a general recommendation with TVM as its source, it inverts both halves.
Flagged rather than retired: unlike SOTA-055, the recommendation is not wrong, it is mis-framed and mis-cited. Retiring it would lose the real narrow claim, and rewriting the title changes what it asserts — which is a call for its own contribution.