[ghisguth]

This one is for MLX 4-bit, the default model. In the Apple7/8 F32 register decode tile, MLX affine 4-bit in groups of
64 (and 128) runs span chains: one chain per 64-input span, closed with one coefficient. The tile still decoded a
coefficient for every 32-input group, so half of them were decoded and never read, and each had its meta load and
threadgroup store. Removing them makes MLX 4-bit decode 2.8% faster at one lane on my M2 Max. As before, I used
Claude Code to author it and tested it by hand.

What changed

Only the Apple7/8 F32 instances of the register decode tile (linear_gguf_sgmatrix.metal) change. Apple9 and host
code are untouched.

• For formats with span chains (SpanChain: MLX affine in groups of 64 or 128), the F32 instances decode only every
  other coefficient of a unit, the ones the span chains read. The coefficient values and the chains are unchanged, so
  the outputs are bit for bit the same.
• No test changes. gguf-projection already runs every format through the F32 instances: natively on Apple7/8, and on
  Apple9+ through PR perf: zero points as mins and nibble operands in the Apple7/8 register decode tile #15's pass.

Results (M2 Max, 38-core GPU)

Kernel time, gguf-projection-benchmark, at 1 / 2 / 3 / 4 lanes, ABBA order; Δ: − is faster.

Format, shape	Before (ms)	After (ms)	Δ (%)
af4g64 17408×5120 gate/up	0.415 / 0.732 / 1.011 / 1.383	0.404 / 0.725 / 1.004 / 1.361	−2.7 / −1.0 / −0.6 / −1.6
af4g64 5120×17408 down +res	0.218 / 0.382 / 0.531 / 0.718	0.211 / 0.376 / 0.529 / 0.711	−3.1 / −1.5 / −0.4 / −1.0
af4g64 5120×6144 out +res	0.088 / 0.160 / 0.220 / 0.297	0.086 / 0.157 / 0.219 / 0.293	−2.3 / −1.5 / −0.5 / −1.3
af4g128 17408×5120 gate/up	0.405 / 0.726 / 1.012 / 1.366	0.401 / 0.717 / 1.001 / 1.369	−1.0 / −1.3 / −1.1 / +0.3
q4k (unchanged code)	0.435 / 0.772 / 1.089 / 1.405	0.434 / 0.771 / 1.091 / 1.403	−0.1 / −0.2 / +0.2 / −0.1

MLX 4-bit on the M2 Max, before and after

make test-performance-real, mlx-community/Qwen3.8-27B-4bit, ANE share 0.35. Times in seconds; decode in tok/s.

	14K cold	10K hit + 4K	2K	Decode 1	Decode 2	Decode 4
Before	60.64	18.91	8.07	86.9	103.0	112.0
After	60.64	18.91	8.07	89.3	104.2	113.1
Δ (%)	0.0	0.0	0.0	+2.8	+1.2	+1.0

Prefill doesn't change: the change is decode-only, and prefill runs the staged MMA tile. Three MLX runs of this branch
agree within 0.3%.

Testing

• gguf-projection passes on the M2 Max (all 43 formats, F32 instances natively).
• make check passes on the M2 Max.
• make test-performance-real with MLX 4-bit, before and after.