Function streaming_layer_routed_expert_bytes

ds4.c:3248–3289

            const float av = fabsf(x[off + i]);
            if (av > amax) amax = av;
        }

        if (amax < 1.0e-4f) amax = 1.0e-4f;
        const float scale = ldexpf(1.0f, (int)ceilf(log2f(amax / 448.0f)));
        for (uint32_t i = 0; i < 64; i++) {
            float v = x[off + i] / scale;
            if (v > 448.0f) v = 448.0f;
            if (v < -448.0f) v = -448.0f;
            x[off + i] = dsv4_e4m3fn_dequant_cpu(v) * scale;
        }
    }
}

static float dsv4_e2m1fn_value_cpu(int i) {
    static const float values[8] = {
        0.0f, 0.5f, 1.0f, 1.5f, 2.0f, 3.0f, 4.0f, 6.0f,
    };
    return values[i & 7];
}

static float dsv4_e2m1fn_dequant_cpu(float x) {
    const float sign = x < 0.0f ? -1.0f : 1.0f;
    const float ax = fminf(fabsf(x), 6.0f);
    int best = 0;
    float best_diff = fabsf(ax - dsv4_e2m1fn_value_cpu(0));
    for (int i = 1; i < 8; i++) {
        const float diff = fabsf(ax - dsv4_e2m1fn_value_cpu(i));
        if (diff < best_diff || (diff == best_diff && (i & 1) == 0 && (best & 1) != 0)) {
            best = i;
            best_diff = diff;
        }
    }
    return sign * dsv4_e2m1fn_value_cpu(best);
}

static void dsv4_hadamard128_inplace_cpu(float *x) {
    for (uint32_t stride = 1; stride < 128; stride <<= 1) {
        for (uint32_t base = 0; base < 128; base += 2u * stride) {
            for (uint32_t i = 0; i < stride; i++) {
                const float a = x[base + i];
                const float b = x[base + stride + i];
                x[base + i] = a + b;
                x[base + stride + i] = a - b;
            }
        }