perf(cuda): reduce MMQ stream-k overhead

[dusterbloom]

What changed

• Backport upstream llama.cpp #22298 (9725a313be0528214c4a02fed906ddaf7b3f712e) into the GGML snapshot vendored by Lucebox Hub.
• Record the backport in server/deps/llama.cpp/VENDOR.md.
• Add a CUDA-only IQ4_XS CPU-vs-CUDA MUL_MAT oracle covering the scheduler's high-efficiency no-fixup path, partial rows, stream-k fixup path, two-wave boundary, and a deeper K dimension.

The resulting mmq.cuh Git blob is b867d39653c19cf4efd3bb7809e8b43572a79a6e, matching the upstream patch output exactly.

Why

Lucebox Hub vendors GGML directly, and its vendored mmq.cuh did not yet include upstream's MMQ stream-k overhead reduction. Qwen3.6-27B IQ4_XS prefill spends enough time in this path for the scheduler and fixup overhead to be measurable.

This is intentionally a surgical upstream backport rather than a broad GGML synchronization.

Measured impact

Fresh A-B-B-A validation was collected on Hub main at 0e0023649131a23f45d58be71f2bfc60d6cd25a0, using the same RTX 3090, Qwen3.6-27B IQ4_XS model, CUDA 12.6/GCC 11 toolchain, runtime flags, salted prompts, and disabled prefix/prefill caches. Each phase used one warmup and five measured repetitions per context.

Context	Clean pooled median	Candidate pooled median	Change
1,024	1300.68 tok/s	1365.68 tok/s	+5.00%
8,192	1312.16 tok/s	1382.97 tok/s	+5.40%

Content-addressed current-tip evidence:

• Evidence ID: evidence-sha256-be8ec967d917b0a1110e55fc5571b23756f77228739a5a374b00c6898a565964
• Evidence file SHA-256: 5805ded9110250a1b178f5952e125dd2bda6eec211074931f33017209009d143
• Clean executable SHA-256: 8ad8a9fa53e54ef640f1503527c1990455811fa19bae8b8f0fc736004d693ad0
• Candidate executable SHA-256: dddb7db95cd7c54d0a2d0c84b8ab0f3dc30cd5daf7f5153217d5b42c1669b232
• Model SHA-256: 8a3365759dc1b33b52c4e7d91d5a67d5ee1418e8408aa54196f04a98da53e5dc

The earlier full-logit, repeat-determinism, and generation-canary campaign remains separately pinned to its original Hub revision in AutoLuce's quality report. It is supporting historical evidence, not relabeled as current-tip evidence.

Validation

• dflash_server Release build, CUDA 12.6, GCC 11, sm_86, -j4: pass.
• Focused test_mmq_streamk_iq4_xs on RTX 3090: 5/5 pass; all outputs finite and all NMSE values below GGML's 5e-4 threshold.
• Candidate MMQ Git blob matches the upstream patch output.
• git diff --check: pass.
• AutoLuce harness/tests: 345 pass; Ruff: pass.
• Adversarial GLM-5.2 review: no code defects or portability blockers found.

The test is built only for the CUDA backend. The backported kernel retains upstream's existing non-CDNA HIP and pre-Volta fallback behavior; no new HIP-specific implementation is introduced.

[cubic-dev-ai]

No issues found across 4 files

_{Re-trigger cubic}