[glennneuber]

Summary

Upstream merged llama.cpp#29941 on 2026-10-04 as dd266785c and closed our #27044 in its favour. The maintainer decided the same day to keep compat 903. This PR records that decision and the re-cut it implies. It changes no code.

• What upstream merged: the one-line ne11 → ne12, the same line docs(upstream): llama.cpp#27044 was reviewed; test the maintainer's variant on sm_120 #445 measured on sm_120. Between 05043961, where docs(upstream): llama.cpp#27044 was reviewed; test the maintainer's variant on sm_120 #445 measured, and dd266785c, it is the only change to ggml-cuda's mmq, mmvq, mmid and common files.
• Why 903 stays:
  • From 128 tokens up, both lines pad 128 blocks, so every reported crash is fixed upstream.
  • Below 128 tokens, ne12 can leave the last tile up to 63 blocks short. The CPU check finds 268,440 of 2,439,360 shapes short, and memcheck with the exact allocation shows errors at 65 and 100 tokens on sm_120.
  • The stock pool hid that read in every case measured.
• What it means at the next pin: nothing changes at b11081, and v0.35.1's b11232 also predates dd266785c. At the first pin that contains it, 903's context line reads ne12, so 903 is re-cut as ne12 → ne12*n_expert_used.

Changes

• docs/maxusai/retirement-register.md: the 903 row. #27044 is closed, and the row records what now retires 903, which test gates its deletion (tasks/mmq-ids-padding-gpu.sh on the new pin, plus the CPU check), and the decision.
• llama/compat/README.md: the 903 entry gains the outcome and the re-cut. "Up to a 512-row tile" is corrected to at most 128 rows, since no MMQ config is wider. The 512 was get_J_max's cap, as docs(upstream): llama.cpp#27044 was reviewed; test the maintainer's variant on sm_120 #445's review found.
• README.md: the MoE + MMQ row and the "filed and pending" paragraph. The links now point at the PR, not the issue path.
• docs/maxusai/upstream-mmq-submission-material.md: a new "Outcome (2026-10-04)" section, and the open items updated.

Nothing was posted upstream. A follow-up there on the residual would have to be written by hand.

Test plan

• check_source_paths.py --changed-since origin/main: every referenced file resolves.
• Name scan of the diff: clean.

ai-server/mlx-cuda

🤖 Generated with Claude Code

[glennneuber]

Closing: superseded by #448 and #462. This PR records the 2026-10-04 plan, which kept 903 and re-cut it as ne12 → ne12*n_expert_used. Main has moved past that plan twice.

• The outcome is already on main, from docs(mmq): the MMQ padding must cover the padded y tile — 903 re-cut; short on CDNA3, gfx1151 and RDNA4, not NVIDIA #448 (57f9fbc87):
  • upstream-mmq-submission-material.md says "Upstream merged #29941 and closed #27044."
  • The retirement register lists #27044 as closed on 2026-10-04 for #29941.
  • The compat README's History section covers it, including the 128-column correction.
• 903's design changed:
  • On 2026-10-05 it moved to the widest padded tile.
  • With compat: 903 becomes llama.cpp#29953 plus the y-tile guard, at b11351 (verified on CUDA and ROCm) #462 it became #29953, backported to b11351, plus the y-tile guard.
  • Merging this would bring back a 903 description that main no longer ships.
• It no longer applies: all four files conflict with main.

Nothing in it is missing from main. If you see something I missed, reopen it.