Function weights_streaming_layer_experts_uniform

ds4.c:3344–3354

* Mixed-precision ("boosted") GGUFs upcast a few layers' routed experts to a * bigger quant (e.g. Q4_K among IQ2 layers). The streaming expert cache is a * single-size-class slab allocator sized from the FIRST routed layer, so those * layers can never be served from it: they must read expert weights through the * mapped-model views instead. A layer is "uniform" iff its per-expert bytes * matc

        float amax = 0.0f;
        for (int j = 0; j < QK_K; j++) {
            const float ax = fabsf(x[j]);
            if (ax > amax) {
                amax = ax;
                max = x[j];
            }
        }

        if (amax == 0.0f) {
            y[b].d = 0.0f;
            memset(y[b].qs, 0, sizeof(y[b].qs));
            memset(y[b].bsums, 0, sizeof(y[b].bsums));
            x += QK_K;
            continue;
        }