[Qubitium]

Mirrors the small-batch work from dphnAI/nf12#3.

What changed

m <= 15 runs the fused kernel, which decodes each 16×64 super-atom once and
dots it against every token in flight. It took m as a runtime argument, so
acc[16][2] was indexed dynamically — which keeps the array out of VGPRs — and
the warp-partial plane was always sized for 16 tokens even when one was in
flight. The token count is now a template parameter: both are compile-time
sized and the token loop fully unrolls.

At m=1 that raises occupancy enough to hurt: the kernel then runs eight CTAs
per CU and 256 concurrent packed streams, which thrashes the 64 MiB LLC.
Over-allocating the partial plane to 32 KiB throttles it to two CTAs per CU.

Same-process A/B of the old and new fused kernel (µs, lower is better)

k × n	m	before	after	gain
4096×4096	1	28.57	20.45	1.40×
4096×4096	8	32.58	24.15	1.35×
2048×8192	1	19.44	18.49	1.05×
2048×8192	8	31.95	23.87	1.34×
896×4864	1	8.36	8.08	1.03×
896×4864	8	15.40	11.57	1.33×
4096×14336	8	161.67	149.18	1.08×
25600×5120	8	364.93	323.03	1.13×

Result against dense BF16 (warm, R9700 gfx1201)

k × n	m=1 before → after	m=8 before → after
4096×4096	0.84 → 1.20	0.67 → 0.94
2048×8192	1.33 → 1.38	0.80 → 0.97
896×4864	1.63 → 1.69	0.83 → 0.97
4096×28672	1.33 → 1.34	1.22 → 1.30
25600×5120	1.33 → 1.33	1.15 → 1.31
8192×8192	1.64 → 1.71	1.09 → 1.27

Every shape is now above parity at m=1 (1.20–1.71×); m=8 is 0.94–1.36×.
m=16 still routes to the matrix-core kernel, which is 1.2–2.3× faster than the
fused one there on every measured shape.

scripts/validate_nf12_ad.py --backend hip on this branch: 4096×4096 m=1
1.32×, m=8 0.96×, m=16 0.90×; 11008×4096 m=1 3.75×.

Accuracy

Unchanged. The kernel still decodes to exact BF16 and accumulates in FP32; the
only change is codegen shape. The validator reports 0 violations of
|err| <= 2e-3 + 2e-3*|ref|, rms_rel ≈ 1.66e-3 (the BF16 rounding floor).

Tests

pytest tests/test_nf12_ad_codec.py tests/test_backend_naming.py \
       tests/test_quant_bits_validation.py tests/test_rocm_llc.py
# 235 passed, 1 skipped