docs(mmq): Verify the #29953 y-tile guard under hipcc for #455

[glennneuber]

Wants review before merge. This PR is documentation only: a hipcc verify script and its results for #455. It changes no shipped code. The guard variant it measures also passes on nvcc (#455).

The change

#455 asked whether mmq-29953-y-tile-guard.patch compiles and fires under hipcc. This PR adds the script that answers that question, and the results:

• docs/maxusai/tasks/mmq-29953-guard-verify-hip.sh compiles mmq.cu with hipcc for gfx1151 and scores each build the way mmq-29953-guard-verify.sh does. It runs that script's six builds, and splits build 5 three ways: rdna3_5 at 192 and 160 threads, and blackwell at 192. It uses either CMake's own compile command or a minimal hipcc command. It records the blob of the patch it applies, and checks that mmq.o is byte-identical with the guard on and off. VARIANT=instantiate applies the variant below, and does nothing to a patch that already has it.
• mmq-successor-results/rocm-gfx1151/verify-29953-guard.txt holds the runs: the patch as published (blob 729f22c2) and the variant, each with both commands, and one run against a local fold of the variant into the patch.
• The README in that folder gets a section on Help wanted (ROCm): does the llama.cpp#29953 y-tile guard compile, and fire, under hipcc? #455.

The measured effect

build	hipcc, the patch as published	hipcc, VARIANT=instantiate
1, 2, 6: the head; with the guard; with the guard compiled out	ok	ok
3, 4: helper shrunk to J blocks; helper rounded to one warp	10 guard errors and 10 knock-on errors, all ten tables named	220 guard errors, all ten tables named
5a: an rdna3_5 config at 192 threads (#455's mutation)	ok: there is nothing to catch	ok
5b: the same config at 160 threads	1 guard error and 1 knock-on error, only rdna3_5 named	1 guard error, only rdna3_5 named
5c: a blackwell config at 192 threads (the nvcc script's build 5)	1 guard error and 1 knock-on error, only blackwell named	1 guard error, only blackwell named

The knock-on error is the guard's outer static_assert, which fails after a per-type assertion fails. The variant removes that static_assert and uses an explicitly instantiated class instead. The guard adds 0.8 s to each compile (2.89 s against 2.06 s) and no code.

Verified on CUDA

The CUDA host compiled the variant with nvcc for sm_120. Its results match the hipcc results above, table for table, including rdna3_5 at 160 threads (#455). The CUDA host will fold the variant into mmq-29953-y-tile-guard.patch. After that, this script runs the folded patch without VARIANT.

    template <typename> struct ggml_cuda_mmq_y_tile_guards_##TABLE;
    template <int... ts> struct ggml_cuda_mmq_y_tile_guards_##TABLE<std::integer_sequence<int, ts...>>
        : ggml_cuda_mmq_y_tile_guard_##TABLE<ts>... {};
    template struct ggml_cuda_mmq_y_tile_guards_##TABLE<std::make_integer_sequence<int, GGML_TYPE_COUNT>>;

amd-server/rocm

🤖 Generated with Claude Code