Tiled attention: preserving results without a full score matrix

01 / QUESTION

Derive online softmax from a failing block-average baseline and separate smaller temporaries from measured speed.

METHOD

Freeze protocol and source before execution, save raw records, then independently recompute metrics and verify hashes.

RESULT

Nine array groups, maximum discrepancy 3.33e-16; tiled execution remains slower in 558 timings.

OBSERVATIONS

Nine array groups, maximum discrepancy 3.33e-16; tiled execution remains slower in 558 timings.

CONCLUSION

Conclusions are limited to the stated variables and acceptance criteria, retaining failures and unverified scope.

LIMITATIONS

Teaching-scale fixed conditions, not online model quality, production throughput or hardware safety. The complete assistant and VL01 remain unfinished.

REPRODUCE

The article provides a pinned revision, environment, runner and independent audit commands.

1. Mechanisms

   Tiled attention: preserving results without a full score matrix

   Derive online softmax from a failing block-average baseline and separate smaller temporaries from measured speed.

   2026-10-079 MINPublished
2. PROJECT / 002

   Java LLM Practice

   Reproducible Java conversation, retry and persistence boundaries, alongside attention, cache-cost and quantization-error experiments.

   Building
3. Mechanisms

   LLM mechanics II: what does KV caching save, and how do we verify it?

   Compare prefix recomputation, token decoding and chunk caching with output errors, projected-row counts and actual array payloads.

   2026-10-038 MINPublished