docs(mmq): the test cases' comments, against #29953's head

[glennneuber]

Summary

Comment-only fix to docs/maxusai/tasks/mmq-tests.patch, the test-backend-ops cases behind the llama.cpp#29953 work. The cases are unchanged. Three comments were wrong, and the rest described the rules from before #29953.

• dense321 said master's get_J_max(321) = 80. It is 128: the launch picks J = 80 only because J = 96..128 need more than sm_75's 64 KiB of shared memory, a filter get_J_max() does not apply. That is why master passed this case and the published #29953, padding exactly 80 blocks, faulted. Computed on the CPU from dd266785c's config tables.
• ids16/ids64 said every rule gives src1 enough room, so only ids_dst can be read past. ids16 is the case where src1 is short, 21 blocks needed against 16 given, and get_J_max(ne11) = get_J_max(ne12) = J does not hold for ids64 (64, 96 and J = 112).
• The MUL_MAT_ID header said the padding must cover the widest tile the launch can pick: the fork's compat 903 rule, not what #29953's head does.

Every comment now describes its case against #29953's head (3070d927f), names dd266785c where it means master's old get_J_max(), and says the cases fault only with an allocation that ends at its data.

Test plan

• Comment-only: tests/test-backend-ops.cpp with the old and with the new patch applied, comments stripped (cpp -fpreprocessed), is byte-identical (496,625 bytes), and the same comparison catches a one-character change to a test case.
• The patch applies cleanly to dd266785c and to 3070d927f, and the harness marker get_J_max(113) = 64 is still found after applying it.
• No changed comment line over 120 columns; check_source_paths.py; name scan.

🤖 Generated with Claude Code

---

From ai-server/mlx-cuda via /collab · 2026-10-05 · build llama.cpp dd266785c / 3070d927f · signed ai-server