Integrate antirez/main at 890898cf and resolve the shared Metal Q8 conflicts.
Keep Qwen's plain NR2/NR4 matvec and reduction arithmetic alongside upstream's
BF16, RoPE, HC and row variants. Preserve Qwen dispatch guards and prefill
unpack policy, with BF16 rounding disabled for Qwen and NR4 disabled for the
BF16 entry points. Keep upstream's concurrent-unpack and cluster2 shape guards.
Update the Q8 oracle source extraction for the dedicated plain implementation;
the independent frozen references remain unchanged.
Validated CPU/Metal builds; complete Qwen kernels and SSD expert tests;
resident/SSD arithmetic in four math/buffer modes; memory, MTP/device policies
and invalid-token staging; frontend, Engram and session/TP host tests.
Frozen Q8 oracles pass 3072 reduction cases and 1024 full-GEMV fixtures with
zero non-NaN bit differences.
V4.1 synthetic coverage passes up to its decode-topk dispatch-count assertion
at tests/test_deepseek41_metal.c:1618. The same assertion also fails on an
isolated build of unmodified upstream/main on this M1 Max. CUDA/ROCm, M5/M6
hardware and full-model inference were not tested.
Giorgio Oppo
9eb4a12f25