M2 Metal: reject Q8 activation staging without an integer-dot primitive #864

Finding

On Apple M2 Pro with Apple Metal compiler 32023.883, a diagnostic MSL kernel using dot(char4, char4) fails to compile: the exposed dot overloads accept only half and float vectors. The installed metal_simdgroup_matrix headers likewise constrain multiply operands and accumulators to floating-point types; available 8x8 aliases are half, float, and bfloat. Thus a Q8_K activation staging transform cannot map its int8 products to a packed integer-dot or integer matrix primitive on this target.

Incumbent cross-check

Pinned current upstreams:

• llama.cpp master 4a08fa29705b8177e332b134306566c2c4b95902 keeps N_R0_Q4_K=2/N_SG_Q4_K=2 and its Metal Q4_K matvec consumes F32 activations directly.
• MLX main d9077d8316ad7305497a3ecf2296bd0e0e99a627 also performs quantized Metal matvec dots against floating activations rather than staging Q8 activations.

GoAI current Q4_K Metal counters are ALU-limited, so Q8 staging initially looks attractive. Without packed integer dot, however, it replaces each efficient floating multiply-accumulate with scalar/vector-lane integer multiply-add, retains the weight unpack, adds activation quantization plus scale traffic, and changes model numerics. No production proposal is justified until a target exposes a genuine packed integer-dot primitive or an isolated microbenchmark disproves this instruction-count model.

Potential detector guidance

A future perfscan GPU rule should warn when an activation-quantization proposal assumes int8 dot throughput without proving an exposed packed dot/MMA primitive for the exact backend capability tier. Treat conversion plus scalar integer multiply-add as extra work, not as a tensor-core path.