[thelittlefireman]

Overview

Add static assertion for DP4A MMQ tile height

Additional information

Will I was trying to tweak some mmq config, I try this config CASE(GGML_TYPE_IQ4_NL, 512, 2, 96, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true); this result in wrong computation and no test seems to detect this issue and the compiler accept it, as it should not.

On gfx906, mmq use DP4A with 64 physical lanes, and in this loop:

for (int i0 = 0; i0 < I; i0 += warp_size) {
    sum[j0/nwarps*I/warp_size + i0/warp_size] += ...;
}

With I=96 the loop use i0=0 and i0=64 but the integer division equals I/warp_size = 96/64 = 1.
For J40 with 512 threads (8 waves):

Iteration	Sum index
j0=0, i0=0	0
j0=0, i0=64	1
j0=8, i0=0	1 — collision
j0=8, i0=64	2

This static assert should avoid this mistake :)

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: NO

[thelittlefireman]

@ggml-org/ggml-cuda Can you take a look at this ? Thanks :)