docs(mmq): the Metal line numbers were master's, not the pin's

[glennneuber]

Summary

Follow-up to #452, from the Metal host's third review note there: the two mul_mv.metal line numbers in docs/maxusai/mmq-padding-metal-not-affected.md were three lines off at b11081. The reading was done on upstream master (dd266785c) while the file said it was done at the fork's pin, and mul_mm.metal and mul_mv.metal differ between the two.

At b11081 (161755f2) the quoted loops in kernel_mul_mv_q4_K_f32_impl are at 1569 (the unconditional src0 read) and 1607 (the guarded write), not 1572 and 1610. At the pin, 1572 is sc16[2] = … inside the read loop and 1610 is dst_f32[first_row + row] = sum_all; inside the write loop, as the review said.

The conclusion is unchanged: the specific MMQ tail-padding defect is absent from Metal. What was wrong was the tree the numbers came from.

Changes

docs/maxusai/mmq-padding-metal-not-affected.md, one file, no code:

• mul_mv.metal:1572 → :1569 and :1610 → :1607; that block's ... elision becomes [...], matching the other quoted block.
• A Which tree note at the top: read on dd266785c, re-verified at b11081, every line number in the file is the pin's.
• A third bullet in the review section recording this correction.

#452's own description still carries 1572/1610 and the "read from b11081" line. The file on main is the record.

Test plan

Every check below was run against b11081 (161755f2), not master. The tree is what was wrong last time, so each line names it.

• The two quoted mul_mv.metal lines, trailing citation comments stripped, diffed against sed -n '1569p;1607p' of the pin's file: identical, whitespace included.
• The quoted kernel_mul_mm_id block diffed against the pin's mul_mm.metal: two exact contiguous runs, 542–555 and 559–560, so the block is 542–560 verbatim with 556–558 (the il0/il lines) elided at the marker. The clamp comment is at 552.
• The review's other references: mul_mm.metal:406 is n_all += sel > 0;, and :822 is the write-back's for (short j = sgitg; j < nr1; j += 4) {.
• ggml_metal_op_mul_mat_id_extra_{tpe,ids,amax} each appear in both ggml-metal.cpp and ggml-metal-ops.cpp; MMQ_TILE_Y_K and block_q8_1_mmq appear only under ggml/src/ggml-cuda/.
• check_source_paths.py --changed-since origin/main (every referenced file resolves), name scan.
• No GPU run; not applicable to a citation fix.

ai-server/mlx-cuda

🤖 Generated with Claude Code