// DS4 Metal routed-MoE matvec kernels.

#ifndef QK_K
#define QK_K 256
#endif
#define N_R0_Q2_K 4
#define N_R0_GLM_Q2_PAIR2_K 1
#define N_R0_Q4_K 2
#define N_R0_Q8_K 2
#define N_R0_GLM_Q4_PAIR2_K 1
#define N_R0_GLM_Q4_PAIR_K 4
#define N_R0_Q5_PAIR_K 4
#define N_R0_Q5_K 4
#define N_R0_Q6_K 2
#define N_R0_IQ2_XXS 4

static constant uchar ds4_metal_kmask_iq2xs[8] = {
    1, 2, 4, 8, 16, 32, 64, 128
};

static constant uchar ds4_metal_ksigns_iq2xs[128] = {
      0, 129, 130,   3, 132,   5,   6, 135, 136,   9,  10, 139,  12, 141, 142,  15,
    144,  17,  18, 147,  20, 149, 150,  23,  24, 153, 154,  27, 156,  29,  30, 159,
    160,  33,  34, 163,  36, 165, 166,  39,  40, 169, 170,  43, 172,  45,  46, 175,
     48, 177, 178,  51, 180,  53,  54, 183, 184,  57,  58, 187,  60, 189, 190,  63,
    192,  65,  66, 195,  68, 197, 198,  71,  72, 201, 202,  75, 204,  77,  78, 207,
     80, 209, 210,  83, 212,  85,  86, 215, 216,  89,  90, 219,  92, 221, 222,  95,
     96, 225, 226,  99, 228, 101, 102, 231, 232, 105, 106, 235, 108, 237, 238, 111,
    240, 113, 114, 243, 116, 245, 246, 119, 120, 249, 250, 123, 252, 125, 126, 255,
};

static constant ulong ds4_metal_iq2xxs_grid[256] = {
    0x0808080808080808, 0x080808080808082b, 0x0808080808081919, 0x0808080808082b08,
    0x0808080808082b2b, 0x0808080808190819, 0x0808080808191908, 0x08080808082b0808,
    0x08080808082b082b, 0x08080808082b2b08, 0x08080808082b2b2b, 0x0808080819080819,
    0x0808080819081908, 0x0808080819190808, 0x0808080819192b08, 0x08080808192b0819,
    0x08080808192b1908, 0x080808082b080808, 0x080808082b08082b, 0x080808082b082b2b,
    0x080808082b2b082b, 0x0808081908080819, 0x0808081908081908, 0x0808081908190808,
    0x0808081908191919, 0x0808081919080808, 0x080808192b081908, 0x080808192b192b08,
    0x0808082b08080808, 0x0808082b0808082b, 0x0808082b082b082b, 0x0808082b2b08082b,
    0x0808190808080819, 0x0808190808081908, 0x0808190808190808, 0x08081908082b0819,
    0x08081908082b1908, 0x0808190819080808, 0x080819081908082b, 0x0808190819082b08,
    0x08081908192b0808, 0x080819082b080819, 0x080819082b081908, 0x080819082b190808,
    0x080819082b2b1908, 0x0808191908080808, 0x080819190808082b, 0x0808191908082b08,
    0x08081919082b0808, 0x080819191908192b, 0x08081919192b2b19, 0x080819192b080808,
    0x080819192b190819, 0x0808192b08082b19, 0x0808192b08190808, 0x0808192b19080808,
    0x0808192b2b081908, 0x0808192b2b2b1908, 0x08082b0808080808, 0x08082b0808081919,
    0x08082b0808082b08, 0x08082b0808191908, 0x08082b08082b2b08, 0x08082b0819080819,
    0x08082b0819081908, 0x08082b0819190808, 0x08082b081919082b, 0x08082b082b082b08,
    0x08082b1908081908, 0x08082b1919080808, 0x08082b2b0808082b, 0x08082b2b08191908,
    0x0819080808080819, 0x0819080808081908, 0x0819080808190808, 0x08190808082b0819,
    0x0819080819080808, 0x08190808192b0808, 0x081908082b081908, 0x081908082b190808,
    0x081908082b191919, 0x0819081908080808, 0x0819081908082b08, 0x08190819082b0808,
    0x0819081919190808, 0x0819081919192b2b, 0x081908192b080808, 0x0819082b082b1908,
    0x0819082b19081919, 0x0819190808080808, 0x0819190808082b08, 0x08191908082b0808,
    0x08191908082b1919, 0x0819190819082b19, 0x081919082b080808, 0x0819191908192b08,
    0x08191919192b082b, 0x0819192b08080808, 0x0819192b0819192b, 0x08192b0808080819,
    0x08192b0808081908, 0x08192b0808190808, 0x08192b0819080808, 0x08192b082b080819,
    0x08192b1908080808, 0x08192b1908081919, 0x08192b192b2b0808, 0x08192b2b19190819,
    0x082b080808080808, 0x082b08080808082b, 0x082b080808082b2b, 0x082b080819081908,
    0x082b0808192b0819, 0x082b08082b080808, 0x082b08082b08082b, 0x082b0819082b2b19,
    0x082b081919082b08, 0x082b082b08080808, 0x082b082b0808082b, 0x082b190808080819,
    0x082b190808081908, 0x082b190808190808, 0x082b190819080808, 0x082b19081919192b,
    0x082b191908080808, 0x082b191919080819, 0x082b1919192b1908, 0x082b192b2b190808,
    0x082b2b0808082b08, 0x082b2b08082b0808, 0x082b2b082b191908, 0x082b2b2b19081908,
    0x1908080808080819, 0x1908080808081908, 0x1908080808190808, 0x1908080808192b08,
    0x19080808082b0819, 0x19080808082b1908, 0x1908080819080808, 0x1908080819082b08,
    0x190808081919192b, 0x19080808192b0808, 0x190808082b080819, 0x190808082b081908,
    0x190808082b190808, 0x1908081908080808, 0x19080819082b0808, 0x19080819192b0819,
    0x190808192b080808, 0x190808192b081919, 0x1908082b08080819, 0x1908082b08190808,
    0x1908082b19082b08, 0x1908082b1919192b, 0x1908082b192b2b08, 0x1908190808080808,
    0x1908190808082b08, 0x19081908082b0808, 0x190819082b080808, 0x190819082b192b19,
    0x190819190819082b, 0x19081919082b1908, 0x1908192b08080808, 0x19082b0808080819,
    0x19082b0808081908, 0x19082b0808190808, 0x19082b0819080808, 0x19082b0819081919,
    0x19082b1908080808, 0x19082b1919192b08, 0x19082b19192b0819, 0x19082b192b08082b,
    0x19082b2b19081919, 0x19082b2b2b190808, 0x1919080808080808, 0x1919080808082b08,
    0x1919080808190819, 0x1919080808192b19, 0x19190808082b0808, 0x191908082b080808,
    0x191908082b082b08, 0x1919081908081908, 0x191908191908082b, 0x191908192b2b1908,
    0x1919082b2b190819, 0x191919082b190808, 0x191919082b19082b, 0x1919191908082b2b,
    0x1919192b08080819, 0x1919192b19191908, 0x19192b0808080808, 0x19192b0808190819,
    0x19192b0808192b19, 0x19192b08192b1908, 0x19192b1919080808, 0x19192b2b08082b08,
    0x192b080808081908, 0x192b080808190808, 0x192b080819080808, 0x192b0808192b2b08,
    0x192b081908080808, 0x192b081919191919, 0x192b082b08192b08, 0x192b082b192b0808,
    0x192b190808080808, 0x192b190808081919, 0x192b191908190808, 0x192b19190819082b,
    0x192b19192b081908, 0x192b2b081908082b, 0x2b08080808080808, 0x2b0808080808082b,
    0x2b08080808082b2b, 0x2b08080819080819, 0x2b0808082b08082b, 0x2b08081908081908,
    0x2b08081908192b08, 0x2b08081919080808, 0x2b08082b08190819, 0x2b08190808080819,
    0x2b08190808081908, 0x2b08190808190808, 0x2b08190808191919, 0x2b08190819080808,
    0x2b081908192b0808, 0x2b08191908080808, 0x2b0819191908192b, 0x2b0819192b191908,
    0x2b08192b08082b19, 0x2b08192b19080808, 0x2b08192b192b0808, 0x2b082b080808082b,
    0x2b082b1908081908, 0x2b082b2b08190819, 0x2b19080808081908, 0x2b19080808190808,
    0x2b190808082b1908, 0x2b19080819080808, 0x2b1908082b2b0819, 0x2b1908190819192b,
    0x2b1908192b080808, 0x2b19082b19081919, 0x2b19190808080808, 0x2b191908082b082b,
    0x2b19190819081908, 0x2b19191919190819, 0x2b192b082b080819, 0x2b192b19082b0808,
    0x2b2b08080808082b, 0x2b2b080819190808, 0x2b2b08082b081919, 0x2b2b081908082b19,
    0x2b2b082b08080808, 0x2b2b190808192b08, 0x2b2b2b0819190808, 0x2b2b2b1908081908,
};

#define kmask_iq2xs ds4_metal_kmask_iq2xs
#define ksigns_iq2xs ds4_metal_ksigns_iq2xs
#define iq2xxs_grid ds4_metal_iq2xxs_grid

struct block_q2_K {
    uchar scales[QK_K/16];
    uchar qs[QK_K/4];
    half d;
    half dmin;
};

struct block_q4_K {
    half d;
    half dmin;
    uchar scales[12];
    uchar qs[QK_K/2];
};

struct block_q5_K {
    half d;
    half dmin;
    uchar scales[12];
    uchar qh[QK_K/8];
    uchar qs[QK_K/2];
};

struct block_q6_K {
    uchar ql[QK_K/2];
    uchar qh[QK_K/4];
    char scales[QK_K/16];
    half d;
};

struct block_iq2_xxs {
    half d;
    ushort qs[QK_K/8];
};

struct ds4_metal_glm_routed_moe_args {
    uint32_t in_dim;
    uint32_t mid_dim;
    uint32_t out_dim;
    uint32_t n_total_expert;
    uint32_t n_expert_used;
    uint32_t n_tokens;
    uint32_t mid_token_stride;
    uint32_t down_type;
    /* Expert ownership under tensor parallelism: tp_world 0/1 = full
     * compute; otherwise each rank owns a contiguous expert range. */
    int32_t  tp_rank;
    int32_t  tp_world;
    int32_t  tp_expert_base;
    uint64_t gate_expert_bytes;
    uint64_t gate_row_bytes;
    uint64_t up_expert_bytes;
    uint64_t up_row_bytes;
    uint64_t down_expert_bytes;
    uint64_t down_row_bytes;
};


static inline bool ds4_tp_owns_expert(int expert, int n_total,
                                      int tp_rank, int tp_world) {
    if (tp_world <= 1) return true;
    const int first = tp_rank * (n_total / tp_world);
    const int last = tp_rank + 1 == tp_world ? n_total : first + n_total / tp_world;
    return expert >= first && expert < last;
}

struct ds4_metal_dsv4_moe_swiglu_weight_args {
    uint32_t width;
    uint32_t rows;
    uint64_t gate_row_stride;
    uint64_t up_row_stride;
    uint64_t mid_row_stride;
    uint64_t weight_stride;
    uint32_t write_clamped;
    float clamp_value;
};

struct ds4_metal_dsv4_moe_sum6_args {
    uint32_t width;
    uint32_t tokens;
    uint64_t src_token_stride;
    uint64_t dst_token_stride;
};

// Routed-MoE activation for the selected experts:
// clamp(gate), clamp(up), silu(gate) * up * route_weight.  Normal inference
// does not consume gate/up after this point, so the fast path avoids writing the
// clamped intermediates back.  A diagnostic env switch can restore those writes
// when comparing the old multi-kernel intermediate tensors.
kernel void kernel_dsv4_moe_swiglu_weight(
        constant ds4_metal_dsv4_moe_swiglu_weight_args &args,
        device char *gate,
        device char *up,
        device char *mid,
        device const char *weights,
        uint row [[threadgroup_position_in_grid]],
        uint tid [[thread_position_in_threadgroup]],
        uint ntg [[threads_per_threadgroup]]) {
    if (row >= args.rows) return;

    device float *gate_row = (device float *)(gate + (uint64_t)row * args.gate_row_stride);
    device float *up_row   = (device float *)(up   + (uint64_t)row * args.up_row_stride);
    device float *mid_row  = (device float *)(mid  + (uint64_t)row * args.mid_row_stride);
    device const float *w  = (device const float *)(weights + (uint64_t)row * args.weight_stride);
    const float route_weight = w[0];
    const float c = args.clamp_value;

    for (uint i = tid; i < args.width; i += ntg) {
        float g = gate_row[i];
        float u = up_row[i];
        if (c > 1.0e-6f) {
            g = min(g, c);
            u = clamp(u, -c, c);
            if (args.write_clamped != 0) {
                gate_row[i] = g;
                up_row[i] = u;
            }
        }
        const float silu = g / (1.0f + exp(-g));
        mid_row[i] = silu * u * route_weight;
    }
}

// Same routed-MoE activation as above, but stores the down-projection input in
// half precision. The grouped matmul path converts F32 activations to half
// before MMA anyway, so this cuts the large mid write/read traffic without
// changing the effective matmul input precision.
kernel void kernel_dsv4_moe_swiglu_weight_f16(
        constant ds4_metal_dsv4_moe_swiglu_weight_args &args,
        device char *gate,
        device char *up,
        device char *mid,
        device const char *weights,
        uint row [[threadgroup_position_in_grid]],
        uint tid [[thread_position_in_threadgroup]],
        uint ntg [[threads_per_threadgroup]]) {
    if (row >= args.rows) return;

    device float *gate_row = (device float *)(gate + (uint64_t)row * args.gate_row_stride);
    device float *up_row   = (device float *)(up   + (uint64_t)row * args.up_row_stride);
    device half  *mid_row  = (device half  *)(mid  + (uint64_t)row * args.mid_row_stride);
    device const float *w  = (device const float *)(weights + (uint64_t)row * args.weight_stride);
    const float route_weight = w[0];
    const float c = args.clamp_value;

    for (uint i = tid; i < args.width; i += ntg) {
        float g = gate_row[i];
        float u = up_row[i];
        if (c > 1.0e-6f) {
            g = min(g, c);
            u = clamp(u, -c, c);
            if (args.write_clamped != 0) {
                gate_row[i] = g;
                up_row[i] = u;
            }
        }
        const float silu = g / (1.0f + exp(-g));
        mid_row[i] = (half)(silu * u * route_weight);
    }
}

kernel void kernel_dsv4_moe_sum6_f32(
        constant ds4_metal_dsv4_moe_sum6_args &args,
        device const char *src,
        device       char *dst,
        uint token[[threadgroup_position_in_grid]],
        uint tid[[thread_position_in_threadgroup]],
        uint ntg[[threads_per_threadgroup]]) {
    if (token >= args.tokens) return;

    device const float *s =
        (device const float *)(src + (uint64_t)token * args.src_token_stride);
    device float *d =
        (device float *)(dst + (uint64_t)token * args.dst_token_stride);

    for (uint col = tid; col < args.width; col += ntg) {
        float v = s[col];
        v += s[args.width + col];
        v += s[2u * args.width + col];
        v += s[3u * args.width + col];
        v += s[4u * args.width + col];
        v += s[5u * args.width + col];
        d[col] = v;
    }
}

kernel void kernel_dsv4_moe_sum8_f32(
        constant ds4_metal_dsv4_moe_sum6_args &args,
        device const char *src,
        device       char *dst,
        uint token[[threadgroup_position_in_grid]],
        uint tid[[thread_position_in_threadgroup]],
        uint ntg[[threads_per_threadgroup]]) {
    if (token >= args.tokens) return;

    device const float *s =
        (device const float *)(src + (uint64_t)token * args.src_token_stride);
    device float *d =
        (device float *)(dst + (uint64_t)token * args.dst_token_stride);

    for (uint col = tid; col < args.width; col += ntg) {
        float v = s[col];
        v += s[args.width + col];
        v += s[2u * args.width + col];
        v += s[3u * args.width + col];
        v += s[4u * args.width + col];
        v += s[5u * args.width + col];
        v += s[6u * args.width + col];
        v += s[7u * args.width + col];
        d[col] = v;
    }
}

template <typename type4x4>
void dequantize_q2_K(device const block_q2_K *xb, short il, thread type4x4 & reg) {
    const float d = xb->d;
    const float min = xb->dmin;
    device const uint8_t * q = (device const uint8_t *)xb->qs;
    float dl, ml;
    uint8_t sc = xb->scales[il];

    q = q + 32*(il/8) + 16*(il&1);
    il = (il/2)%4;

    half  coef = il>1 ? (il>2 ? 1/64.h : 1/16.h) : (il>0 ? 1/4.h : 1.h);
    uchar mask = il>1 ? (il>2 ? 192    : 48)     : (il>0 ? 12    : 3);
    dl = d * (sc & 0xF) * coef, ml = min * (sc >> 4);
    for (int i = 0; i < 16; ++i) {
        reg[i/4][i%4] = dl * (q[i] & mask) - ml;
    }
}

static inline float ds4_glm_q2_K_value(device const block_q2_K *blocks, uint k) {
    const uint block = k / QK_K;
    const uint idx = k - block * QK_K;
    device const block_q2_K *xb = blocks + block;
    const uint group = idx / 16u;
    const uint l = idx - group * 16u;
    const uint q_base = 32u * (group / 8u) + 16u * (group & 1u);
    const uint shift = ((group / 2u) & 3u) * 2u;
    const uint q = ((uint)xb->qs[q_base + l] >> shift) & 0x03u;
    const uint sc = (uint)xb->scales[group];
    return (float)xb->d * (float)(sc & 0x0fu) * (float)q -
           (float)xb->dmin * (float)(sc >> 4u);
}

static inline uchar2 get_scale_min_k4_just2(int j, int k, device const uchar * q) {
    return j < 4 ? uchar2{uchar(q[j+0+k] & 63), uchar(q[j+4+k] & 63)}
                 : uchar2{uchar((q[j+4+k] & 0xF) | ((q[j-4+k] & 0xc0) >> 2)),
                          uchar((q[j+4+k] >> 4) | ((q[j-0+k] & 0xc0) >> 2))};
}

static inline float ds4_glm_q4_K_value(device const block_q4_K *blocks, uint k) {
    const uint block = k / QK_K;
    const uint idx = k - block * QK_K;
    device const block_q4_K *xb = blocks + block;
    const uint group = idx / 32u;
    const uint l = idx - group * 32u;
    const uchar2 sm = get_scale_min_k4_just2((int)group, 0, xb->scales);
    const uint byte_off = (group >> 1u) * 32u + l;
    const uint shift = (group & 1u) * 4u;
    const uint q = (xb->qs[byte_off] >> shift) & 0x0Fu;
    return (float)xb->d * (float)sm.x * (float)q -
           (float)xb->dmin * (float)sm.y;
}

static inline float ds4_glm_q5_K_value(device const block_q5_K *blocks, uint k) {
    const uint block = k / QK_K;
    const uint idx = k - block * QK_K;
    device const block_q5_K *xb = blocks + block;
    const uint group = idx / 32u;
    const uint l = idx - group * 32u;
    const uchar2 sm = get_scale_min_k4_just2((int)group, 0, xb->scales);
    const uint ql_base = (group >> 1u) * 32u + l;
    const uint shift = (group & 1u) * 4u;
    uint q = (xb->qs[ql_base] >> shift) & 0x0Fu;
    q += (xb->qh[l] & (uchar)(1u << group)) ? 16u : 0u;
    return (float)xb->d * (float)sm.x * (float)q -
           (float)xb->dmin * (float)sm.y;
}

static inline float ds4_glm_q6_K_value(device const block_q6_K *blocks, uint k) {
    const uint block = k / QK_K;
    const uint idx = k - block * QK_K;
    device const block_q6_K *xb = blocks + block;
    const uint n128 = idx >> 7u;
    const uint r = idx & 127u;
    const uint l = r & 31u;
    const uint quarter = r >> 5u;
    const uint ql_base = n128 * 64u;
    const uint qh_base = n128 * 32u;
    const uint sc_base = n128 * 8u;
    uint q;
    int sc;

    if (quarter == 0u) {
        q = (xb->ql[ql_base + l] & 0x0Fu) | (((xb->qh[qh_base + l] >> 0u) & 3u) << 4u);
        sc = (int)xb->scales[sc_base + l / 16u + 0u];
    } else if (quarter == 1u) {
        q = (xb->ql[ql_base + 32u + l] & 0x0Fu) | (((xb->qh[qh_base + l] >> 2u) & 3u) << 4u);
        sc = (int)xb->scales[sc_base + l / 16u + 2u];
    } else if (quarter == 2u) {
        q = (xb->ql[ql_base + l] >> 4u) | (((xb->qh[qh_base + l] >> 4u) & 3u) << 4u);
        sc = (int)xb->scales[sc_base + l / 16u + 4u];
    } else {
        q = (xb->ql[ql_base + 32u + l] >> 4u) | (((xb->qh[qh_base + l] >> 6u) & 3u) << 4u);
        sc = (int)xb->scales[sc_base + l / 16u + 6u];
    }

    return (float)xb->d * (float)sc * (float)((int)q - 32);
}

kernel void kernel_glm_q4_K_pair_swiglu_f32(
        constant ds4_metal_glm_routed_moe_args &args,
        device const char *gate,
        device const char *up,
        device const float *x,
        device const int32_t *selected,
        device const float *weights,
        device float *mid,
        threadgroup float *scratch [[threadgroup(0)]],
        uint3 tgpig [[threadgroup_position_in_grid]],
        uint tid [[thread_index_in_threadgroup]]) {
    const uint ntg = 256u;
    const uint row = tgpig.x;
    const uint slot = tgpig.y;
    const uint token = tgpig.z;
    if (row >= args.mid_dim || slot >= args.n_expert_used || token >= args.n_tokens) return;

    const uint64_t selected_off = (uint64_t)token * args.n_expert_used + slot;
    const uint64_t mid_off = (uint64_t)token * args.mid_token_stride +
                             (uint64_t)slot * args.mid_dim + row;
    const int expert = selected[selected_off];
    if (expert < 0 || (uint)expert >= args.n_total_expert) {
        if (tid == 0u) mid[mid_off] = 0.0f;
        return;
    }

    device const block_q4_K *gate_row =
        (device const block_q4_K *)(gate +
            (uint64_t)(uint)expert * args.gate_expert_bytes +
            (uint64_t)row * args.gate_row_bytes);
    device const block_q4_K *up_row =
        (device const block_q4_K *)(up +
            (uint64_t)(uint)expert * args.up_expert_bytes +
            (uint64_t)row * args.up_row_bytes);

    float acc_gate = 0.0f;
    float acc_up = 0.0f;
    device const float *token_x = x + (uint64_t)token * args.in_dim;
    for (uint k = tid; k < args.in_dim; k += ntg) {
        const float xv = token_x[k];
        acc_gate += ds4_glm_q4_K_value(gate_row, k) * xv;
        acc_up += ds4_glm_q4_K_value(up_row, k) * xv;
    }

    scratch[tid] = acc_gate;
    scratch[ntg + tid] = acc_up;
    threadgroup_barrier(mem_flags::mem_threadgroup);
    for (uint stride = ntg >> 1u; stride > 0u; stride >>= 1u) {
        if (tid < stride) {
            scratch[tid] += scratch[tid + stride];
            scratch[ntg + tid] += scratch[ntg + tid + stride];
        }
        threadgroup_barrier(mem_flags::mem_threadgroup);
    }

    if (tid == 0u) {
        const float g = scratch[0];
        const float u = scratch[ntg];
        const float sw = g / (1.0f + exp(-g));
        mid[mid_off] = sw * u * weights[selected_off];
    }
}

template <short N_R0>
static inline void glm_q2_K_pair_swiglu_simd_f32_impl(
        ds4_metal_glm_routed_moe_args args,
        device const char *gate,
        device const char *up,
        device const float *x,
        device const float *weights,
        device float *mid,
        threadgroup float *scratch,
        uint3 tgpig,
        uint slot,
        uint token,
        uint64_t selected_off,
        int expert,
        ushort tiisg,
        ushort sgitg) {
    const short NSG = 2;
    const uint row0 = ((uint)tgpig.x * (uint)NSG + (uint)sgitg) * N_R0;
    if (row0 >= args.mid_dim || slot >= args.n_expert_used || token >= args.n_tokens) return;

    const uint64_t mid_base = (uint64_t)token * args.mid_token_stride +
                              (uint64_t)slot * args.mid_dim;
    if (expert < 0 || (uint)expert >= args.n_total_expert) {
        if (tiisg == 0u) {
            for (short row = 0;
                 row < N_R0 && row0 + (uint)row < args.mid_dim;
                 row++) {
                mid[mid_base + row0 + (uint)row] = 0.0f;
            }
        }
        return;
    }

    const short ix = tiisg / 8;
    const short it = tiisg % 8;
    const short iq = it / 4;
    const short ir = it % 4;
    const short is = (8 * ir) / 16;
    const int nb = args.in_dim / QK_K;
    const uint64_t expert_gate = (uint64_t)(uint)expert * args.gate_expert_bytes;
    const uint64_t expert_up = (uint64_t)(uint)expert * args.up_expert_bytes;
    device const block_q2_K *xg =
        (device const block_q2_K *)(gate + expert_gate +
            (uint64_t)row0 * args.gate_row_bytes);
    device const block_q2_K *xu =
        (device const block_q2_K *)(up + expert_up +
            (uint64_t)row0 * args.up_row_bytes);
    device const float *y = x + (uint64_t)token * args.in_dim;
    device const float *y4 = y + ix * QK_K + 128 * iq + 8 * ir;

    float yl[32];
    float sumg[N_R0] = {0.f};
    float sumu[N_R0] = {0.f};

    for (int ib = ix; ib < nb; ib += 4) {
        float4 sumy = {0.f, 0.f, 0.f, 0.f};
        for (short i = 0; i < 8; ++i) {
            yl[i +  0] = y4[i +  0]; sumy[0] += yl[i +  0];
            yl[i +  8] = y4[i + 32]; sumy[1] += yl[i +  8];
            yl[i + 16] = y4[i + 64]; sumy[2] += yl[i + 16];
            yl[i + 24] = y4[i + 96]; sumy[3] += yl[i + 24];
        }

        device const uint8_t *scg = (device const uint8_t *)xg[ib].scales + 8 * iq + is;
        device const uint16_t *qsg = (device const uint16_t *)xg[ib].qs + 16 * iq + 4 * ir;
        device const half *dhg = &xg[ib].d;
        device const uint8_t *scu = (device const uint8_t *)xu[ib].scales + 8 * iq + is;
        device const uint16_t *qsu = (device const uint16_t *)xu[ib].qs + 16 * iq + 4 * ir;
        device const half *dhu = &xu[ib].d;

        for (short row = 0;
             row < N_R0 && row0 + (uint)row < args.mid_dim;
             row++) {
            float4 acc1g = {0.f, 0.f, 0.f, 0.f};
            float4 acc2g = {0.f, 0.f, 0.f, 0.f};
            float4 acc1u = {0.f, 0.f, 0.f, 0.f};
            float4 acc2u = {0.f, 0.f, 0.f, 0.f};

            for (int i = 0; i < 8; i += 2) {
                acc1g[0] += yl[i +  0] * (qsg[i / 2] & 0x0003);
                acc2g[0] += yl[i +  1] * (qsg[i / 2] & 0x0300);
                acc1g[1] += yl[i +  8] * (qsg[i / 2] & 0x000c);
                acc2g[1] += yl[i +  9] * (qsg[i / 2] & 0x0c00);
                acc1g[2] += yl[i + 16] * (qsg[i / 2] & 0x0030);
                acc2g[2] += yl[i + 17] * (qsg[i / 2] & 0x3000);
                acc1g[3] += yl[i + 24] * (qsg[i / 2] & 0x00c0);
                acc2g[3] += yl[i + 25] * (qsg[i / 2] & 0xc000);

                acc1u[0] += yl[i +  0] * (qsu[i / 2] & 0x0003);
                acc2u[0] += yl[i +  1] * (qsu[i / 2] & 0x0300);
                acc1u[1] += yl[i +  8] * (qsu[i / 2] & 0x000c);
                acc2u[1] += yl[i +  9] * (qsu[i / 2] & 0x0c00);
                acc1u[2] += yl[i + 16] * (qsu[i / 2] & 0x0030);
                acc2u[2] += yl[i + 17] * (qsu[i / 2] & 0x3000);
                acc1u[3] += yl[i + 24] * (qsu[i / 2] & 0x00c0);
                acc2u[3] += yl[i + 25] * (qsu[i / 2] & 0xc000);
            }

            const float dg = dhg[0];
            const float mg = dhg[1] * 1.f / 16.f;
            sumg[row] += dg * ((acc1g[0] + 1.f / 256.f * acc2g[0]) * (scg[0] & 0xF) * 1.f /  1.f +
                               (acc1g[1] + 1.f / 256.f * acc2g[1]) * (scg[2] & 0xF) * 1.f /  4.f +
                               (acc1g[2] + 1.f / 256.f * acc2g[2]) * (scg[4] & 0xF) * 1.f / 16.f +
                               (acc1g[3] + 1.f / 256.f * acc2g[3]) * (scg[6] & 0xF) * 1.f / 64.f) -
                         mg * (sumy[0] * (scg[0] & 0xF0) + sumy[1] * (scg[2] & 0xF0) +
                               sumy[2] * (scg[4] & 0xF0) + sumy[3] * (scg[6] & 0xF0));

            const float du = dhu[0];
            const float mu = dhu[1] * 1.f / 16.f;
            sumu[row] += du * ((acc1u[0] + 1.f / 256.f * acc2u[0]) * (scu[0] & 0xF) * 1.f /  1.f +
                               (acc1u[1] + 1.f / 256.f * acc2u[1]) * (scu[2] & 0xF) * 1.f /  4.f +
                               (acc1u[2] + 1.f / 256.f * acc2u[2]) * (scu[4] & 0xF) * 1.f / 16.f +
                               (acc1u[3] + 1.f / 256.f * acc2u[3]) * (scu[6] & 0xF) * 1.f / 64.f) -
                         mu * (sumy[0] * (scu[0] & 0xF0) + sumy[1] * (scu[2] & 0xF0) +
                               sumy[2] * (scu[4] & 0xF0) + sumy[3] * (scu[6] & 0xF0));

            qsg += args.gate_row_bytes / 2;
            scg += args.gate_row_bytes;
            dhg += args.gate_row_bytes / 2;
            qsu += args.up_row_bytes / 2;
            scu += args.up_row_bytes;
            dhu += args.up_row_bytes / 2;
        }

        y4 += 4 * QK_K;
    }

    for (short row = 0;
         row < N_R0 && row0 + (uint)row < args.mid_dim;
         row++) {
        const float g = simd_sum(sumg[row]);
        const float u = simd_sum(sumu[row]);
        if (tiisg == 0u) {
            const float sw = g / (1.0f + exp(-g));
            mid[mid_base + row0 + (uint)row] = sw * u * weights[selected_off];
        }
    }

    (void)scratch;
}

kernel void kernel_glm_q2_K_pair_swiglu_f32(
        constant ds4_metal_glm_routed_moe_args &args,
        device const char *gate,
        device const char *up,
        device const float *x,
        device const int32_t *selected,
        device const float *weights,
        device float *mid,
        threadgroup float *scratch [[threadgroup(0)]],
        uint3 tgpig [[threadgroup_position_in_grid]],
        ushort tiisg [[thread_index_in_simdgroup]],
        ushort sgitg [[simdgroup_index_in_threadgroup]]) {
    const uint slot = tgpig.y;
    const uint token = tgpig.z;
    if (slot >= args.n_expert_used || token >= args.n_tokens) return;
    const uint64_t selected_off = (uint64_t)token * args.n_expert_used + slot;
    const int expert = selected[selected_off];
    if (!ds4_tp_owns_expert(expert, args.n_total_expert,
                            args.tp_rank, args.tp_world)) return;
    glm_q2_K_pair_swiglu_simd_f32_impl<N_R0_Q2_K>(
        args, gate, up, x, weights, mid, scratch,
        tgpig, slot, token, selected_off, expert - args.tp_expert_base,
        tiisg, sgitg);
}

kernel void kernel_glm_q2_K_addr_pair_swiglu2_f32(
        constant ds4_metal_glm_routed_moe_args &args,
        device const uint64_t *gate_addrs,
        device const uint64_t *up_addrs,
        device const float *x,
        device const int32_t *selected,
        device const float *weights,
        device float *mid,
        threadgroup float *scratch [[threadgroup(0)]],
        uint3 tgpig [[threadgroup_position_in_grid]],
        ushort tiisg [[thread_index_in_simdgroup]],
        ushort sgitg [[simdgroup_index_in_threadgroup]]) {
    const uint slot = tgpig.y;
    const uint token = tgpig.z;
    if (slot >= args.n_expert_used || token >= args.n_tokens) return;

    const uint64_t selected_off = (uint64_t)token * args.n_expert_used + slot;
    const int expert = selected[selected_off];
    const short NSG = 2;
    const uint row0 = ((uint)tgpig.x * (uint)NSG + (uint)sgitg) * N_R0_GLM_Q2_PAIR2_K;
    const uint64_t mid_base = (uint64_t)token * args.mid_token_stride +
                              (uint64_t)slot * args.mid_dim;
    if (row0 >= args.mid_dim) return;

    if (expert < 0 || (uint)expert >= args.n_total_expert) {
        if (tiisg == 0u) {
            for (short row = 0;
                 row < N_R0_GLM_Q2_PAIR2_K && row0 + (uint)row < args.mid_dim;
                 row++) {
                mid[mid_base + row0 + (uint)row] = 0.0f;
            }
        }
        return;
    }
    const uint64_t gate_addr = gate_addrs[(uint)expert];
    const uint64_t up_addr = up_addrs[(uint)expert];
    if (gate_addr == 0 || up_addr == 0) {
        if (tiisg == 0u) {
            for (short row = 0;
                 row < N_R0_GLM_Q2_PAIR2_K && row0 + (uint)row < args.mid_dim;
                 row++) {
                mid[mid_base + row0 + (uint)row] = 0.0f;
            }
        }
        return;
    }

    ds4_metal_glm_routed_moe_args local = args;
    local.n_total_expert = 1;
    local.gate_expert_bytes = 0;
    local.up_expert_bytes = 0;
    glm_q2_K_pair_swiglu_simd_f32_impl<N_R0_GLM_Q2_PAIR2_K>(
        local,
        reinterpret_cast<device const char *>(gate_addr),
        reinterpret_cast<device const char *>(up_addr),
        x, weights, mid, scratch,
        tgpig, slot, token, selected_off, 0, tiisg, sgitg);
}

kernel void kernel_glm_q2_K_addr_pair_swiglu2_f32_masked(
        constant ds4_metal_glm_routed_moe_args &args,
        constant uint32_t &active_mask,
        device const uint64_t *gate_addrs,
        device const uint64_t *up_addrs,
        device const float *x,
        device const int32_t *selected,
        device const float *weights,
        device float *mid,
        threadgroup float *scratch [[threadgroup(0)]],
        uint3 tgpig [[threadgroup_position_in_grid]],
        ushort tiisg [[thread_index_in_simdgroup]],
        ushort sgitg [[simdgroup_index_in_threadgroup]]) {
    const uint slot = tgpig.y;
    const uint token = tgpig.z;
    if (slot >= args.n_expert_used || token >= args.n_tokens) return;
    if ((active_mask & (1u << slot)) == 0u) return;

    const uint64_t selected_off = (uint64_t)token * args.n_expert_used + slot;
    const int expert = selected[selected_off];
    const short NSG = 2;
    const uint row0 = ((uint)tgpig.x * (uint)NSG + (uint)sgitg) * N_R0_GLM_Q2_PAIR2_K;
    const uint64_t mid_base = (uint64_t)token * args.mid_token_stride +
                              (uint64_t)slot * args.mid_dim;
    if (row0 >= args.mid_dim) return;

    if (expert < 0 || (uint)expert >= args.n_total_expert) {
        if (tiisg == 0u) {
            for (short row = 0;
                 row < N_R0_GLM_Q2_PAIR2_K && row0 + (uint)row < args.mid_dim;
                 row++) {
                mid[mid_base + row0 + (uint)row] = 0.0f;
            }
        }
        return;
    }
    const uint64_t gate_addr = gate_addrs[(uint)expert];
    const uint64_t up_addr = up_addrs[(uint)expert];
    if (gate_addr == 0 || up_addr == 0) {
        if (tiisg == 0u) {
            for (short row = 0;
                 row < N_R0_GLM_Q2_PAIR2_K && row0 + (uint)row < args.mid_dim;
                 row++) {
                mid[mid_base + row0 + (uint)row] = 0.0f;
            }
        }
        return;
    }

    ds4_metal_glm_routed_moe_args local = args;
    local.n_total_expert = 1;
    local.gate_expert_bytes = 0;
    local.up_expert_bytes = 0;
    glm_q2_K_pair_swiglu_simd_f32_impl<N_R0_GLM_Q2_PAIR2_K>(
        local,
        reinterpret_cast<device const char *>(gate_addr),
        reinterpret_cast<device const char *>(up_addr),
        x, weights, mid, scratch,
        tgpig, slot, token, selected_off, 0, tiisg, sgitg);
}

template <short N_R0>
static inline void glm_q4_K_pair_swiglu_simd_f32_impl(
        ds4_metal_glm_routed_moe_args args,
        device const char *gate,
        device const char *up,
        device const float *x,
        device const float *weights,
        device float *mid,
        threadgroup float *scratch,
        uint3 tgpig,
        uint slot,
        uint token,
        uint64_t selected_off,
        int expert,
        ushort tiisg,
        ushort sgitg) {
    const short NSG = 2;
    constexpr uint16_t kmask1 = 0x3f3f;
    constexpr uint16_t kmask2 = 0x0f0f;
    constexpr uint16_t kmask3 = 0xc0c0;

    const uint row0 = ((uint)tgpig.x * (uint)NSG + (uint)sgitg) * N_R0;
    if (row0 >= args.mid_dim || slot >= args.n_expert_used || token >= args.n_tokens) return;

    const uint64_t mid_base = (uint64_t)token * args.mid_token_stride +
                              (uint64_t)slot * args.mid_dim;
    if (expert < 0 || (uint)expert >= args.n_total_expert) {
        if (tiisg == 0u) {
            for (short row = 0;
                 row < N_R0 && row0 + (uint)row < args.mid_dim;
                 row++) {
                mid[mid_base + row0 + (uint)row] = 0.0f;
            }
        }
        return;
    }

    const short ix = tiisg / 8;
    const short it = tiisg % 8;
    const short iq = it / 4;
    const short ir = it % 4;
    const int nb = args.in_dim / QK_K;
    const uint64_t expert_gate = (uint64_t)(uint)expert * args.gate_expert_bytes;
    const uint64_t expert_up = (uint64_t)(uint)expert * args.up_expert_bytes;
    device const block_q4_K *xg =
        (device const block_q4_K *)(gate + expert_gate +
            (uint64_t)row0 * args.gate_row_bytes);
    device const block_q4_K *xu =
        (device const block_q4_K *)(up + expert_up +
            (uint64_t)row0 * args.up_row_bytes);
    device const float *y = x + (uint64_t)token * args.in_dim;
    device const float *y4 = y + ix * QK_K + 64 * iq + 8 * ir;

    float sumg[N_R0] = {0.f};
    float sumu[N_R0] = {0.f};
    uint16_t sc16[4];
    thread const uint8_t *sc8 = (thread const uint8_t *)sc16;

    for (int ib = ix; ib < nb; ib += 4) {
        float yl[16];
        float yh[16];
        float4 sumy = {0.f, 0.f, 0.f, 0.f};

        for (short i = 0; i < 8; ++i) {
            yl[i + 0] = y4[i +   0]; sumy[0] += yl[i + 0];
            yl[i + 8] = y4[i +  32]; sumy[1] += yl[i + 8];
            yh[i + 0] = y4[i + 128]; sumy[2] += yh[i + 0];
            yh[i + 8] = y4[i + 160]; sumy[3] += yh[i + 8];
        }

        device const uint16_t *scg = (device const uint16_t *)xg[ib].scales + iq;
        device const uint16_t *qg1 = (device const uint16_t *)xg[ib].qs + 16 * iq + 4 * ir;
        device const half *dhg = &xg[ib].d;
        device const uint16_t *scu = (device const uint16_t *)xu[ib].scales + iq;
        device const uint16_t *qu1 = (device const uint16_t *)xu[ib].qs + 16 * iq + 4 * ir;
        device const half *dhu = &xu[ib].d;

        for (short row = 0;
             row < N_R0 && row0 + (uint)row < args.mid_dim;
             row++) {
            sc16[0] = scg[0] & kmask1;
            sc16[1] = scg[2] & kmask1;
            sc16[2] = ((scg[4] >> 0) & kmask2) | ((scg[0] & kmask3) >> 2);
            sc16[3] = ((scg[4] >> 4) & kmask2) | ((scg[2] & kmask3) >> 2);

            device const uint16_t *qg2 = qg1 + 32;
            float4 acc1g = {0.f, 0.f, 0.f, 0.f};
            float4 acc2g = {0.f, 0.f, 0.f, 0.f};

            FOR_UNROLL (short i = 0; i < 4; ++i) {
                acc1g[0] += yl[2 * i + 0] * (qg1[i] & 0x000F);
                acc1g[1] += yl[2 * i + 1] * (qg1[i] & 0x0F00);
                acc1g[2] += yl[2 * i + 8] * (qg1[i] & 0x00F0);
                acc1g[3] += yl[2 * i + 9] * (qg1[i] & 0xF000);
                acc2g[0] += yh[2 * i + 0] * (qg2[i] & 0x000F);
                acc2g[1] += yh[2 * i + 1] * (qg2[i] & 0x0F00);
                acc2g[2] += yh[2 * i + 8] * (qg2[i] & 0x00F0);
                acc2g[3] += yh[2 * i + 9] * (qg2[i] & 0xF000);
            }

            sumg[row] += dhg[0] * ((acc1g[0] + 1.f / 256.f * acc1g[1]) * sc8[0] +
                                   (acc1g[2] + 1.f / 256.f * acc1g[3]) * sc8[1] * 1.f / 16.f +
                                   (acc2g[0] + 1.f / 256.f * acc2g[1]) * sc8[4] +
                                   (acc2g[2] + 1.f / 256.f * acc2g[3]) * sc8[5] * 1.f / 16.f) -
                         dhg[1] * (sumy[0] * sc8[2] + sumy[1] * sc8[3] +
                                   sumy[2] * sc8[6] + sumy[3] * sc8[7]);

            sc16[0] = scu[0] & kmask1;
            sc16[1] = scu[2] & kmask1;
            sc16[2] = ((scu[4] >> 0) & kmask2) | ((scu[0] & kmask3) >> 2);
            sc16[3] = ((scu[4] >> 4) & kmask2) | ((scu[2] & kmask3) >> 2);

            device const uint16_t *qu2 = qu1 + 32;
            float4 acc1u = {0.f, 0.f, 0.f, 0.f};
            float4 acc2u = {0.f, 0.f, 0.f, 0.f};

            FOR_UNROLL (short i = 0; i < 4; ++i) {
                acc1u[0] += yl[2 * i + 0] * (qu1[i] & 0x000F);
                acc1u[1] += yl[2 * i + 1] * (qu1[i] & 0x0F00);
                acc1u[2] += yl[2 * i + 8] * (qu1[i] & 0x00F0);
                acc1u[3] += yl[2 * i + 9] * (qu1[i] & 0xF000);
                acc2u[0] += yh[2 * i + 0] * (qu2[i] & 0x000F);
                acc2u[1] += yh[2 * i + 1] * (qu2[i] & 0x0F00);
                acc2u[2] += yh[2 * i + 8] * (qu2[i] & 0x00F0);
                acc2u[3] += yh[2 * i + 9] * (qu2[i] & 0xF000);
            }

            sumu[row] += dhu[0] * ((acc1u[0] + 1.f / 256.f * acc1u[1]) * sc8[0] +
                                   (acc1u[2] + 1.f / 256.f * acc1u[3]) * sc8[1] * 1.f / 16.f +
                                   (acc2u[0] + 1.f / 256.f * acc2u[1]) * sc8[4] +
                                   (acc2u[2] + 1.f / 256.f * acc2u[3]) * sc8[5] * 1.f / 16.f) -
                         dhu[1] * (sumy[0] * sc8[2] + sumy[1] * sc8[3] +
                                   sumy[2] * sc8[6] + sumy[3] * sc8[7]);

            qg1 += args.gate_row_bytes / 2;
            scg += args.gate_row_bytes / 2;
            dhg += args.gate_row_bytes / 2;
            qu1 += args.up_row_bytes / 2;
            scu += args.up_row_bytes / 2;
            dhu += args.up_row_bytes / 2;
        }

        y4 += 4 * QK_K;
    }

    for (int row = 0;
         row < N_R0 && row0 + (uint)row < args.mid_dim;
         ++row) {
        const float g = simd_sum(sumg[row]);
        const float u = simd_sum(sumu[row]);
        if (tiisg == 0u) {
            const float sw = g / (1.0f + exp(-g));
            mid[mid_base + row0 + (uint)row] = sw * u * weights[selected_off];
        }
    }

    (void)scratch;
}

kernel void kernel_glm_q4_K_pair_swiglu2_f32(
        constant ds4_metal_glm_routed_moe_args &args,
        device const char *gate,
        device const char *up,
        device const float *x,
        device const int32_t *selected,
        device const float *weights,
        device float *mid,
        threadgroup float *scratch [[threadgroup(0)]],
        uint3 tgpig [[threadgroup_position_in_grid]],
        ushort tiisg [[thread_index_in_simdgroup]],
        ushort sgitg [[simdgroup_index_in_threadgroup]]) {
    const uint slot = tgpig.y;
    const uint token = tgpig.z;
    if (slot >= args.n_expert_used || token >= args.n_tokens) return;
    const uint64_t selected_off = (uint64_t)token * args.n_expert_used + slot;
    const int expert = selected[selected_off];
    glm_q4_K_pair_swiglu_simd_f32_impl<N_R0_GLM_Q4_PAIR2_K>(
        args, gate, up, x, weights, mid, scratch,
        tgpig, slot, token, selected_off, expert, tiisg, sgitg);
}

kernel void kernel_glm_q4_K_addr_pair_swiglu_f32(
        constant ds4_metal_glm_routed_moe_args &args,
        device const uint64_t *gate_addrs,
        device const uint64_t *up_addrs,
        device const float *x,
        device const int32_t *selected,
        device const float *weights,
        device float *mid,
        threadgroup float *scratch [[threadgroup(0)]],
        uint3 tgpig [[threadgroup_position_in_grid]],
        ushort tiisg [[thread_index_in_simdgroup]],
        ushort sgitg [[simdgroup_index_in_threadgroup]]) {
    const uint slot = tgpig.y;
    const uint token = tgpig.z;
    if (slot >= args.n_expert_used || token >= args.n_tokens) return;

    const uint64_t selected_off = (uint64_t)token * args.n_expert_used + slot;
    const int expert = selected[selected_off];
    const short NSG = 2;
    const uint row0 = ((uint)tgpig.x * (uint)NSG + (uint)sgitg) * N_R0_Q4_K;
    const uint64_t mid_base = (uint64_t)token * args.mid_token_stride +
                              (uint64_t)slot * args.mid_dim;
    if (row0 >= args.mid_dim) return;

    if (expert < 0 || (uint)expert >= args.n_total_expert) {
        if (tiisg == 0u) {
            for (short row = 0;
                 row < N_R0_Q4_K && row0 + (uint)row < args.mid_dim;
                 row++) {
                mid[mid_base + row0 + (uint)row] = 0.0f;
            }
        }
        return;
    }
    const uint64_t gate_addr = gate_addrs[(uint)expert];
    const uint64_t up_addr = up_addrs[(uint)expert];
    if (gate_addr == 0 || up_addr == 0) {
        if (tiisg == 0u) {
            for (short row = 0;
                 row < N_R0_Q4_K && row0 + (uint)row < args.mid_dim;
                 row++) {
                mid[mid_base + row0 + (uint)row] = 0.0f;
            }
        }
        return;
    }

    ds4_metal_glm_routed_moe_args local = args;
    local.n_total_expert = 1;
    local.gate_expert_bytes = 0;
    local.up_expert_bytes = 0;
    glm_q4_K_pair_swiglu_simd_f32_impl<N_R0_Q4_K>(
        local,
        reinterpret_cast<device const char *>(gate_addr),
        reinterpret_cast<device const char *>(up_addr),
        x, weights, mid, scratch,
        tgpig, slot, token, selected_off, 0, tiisg, sgitg);
}

kernel void kernel_glm_q4_K_addr_pair_swiglu_f32_masked(
        constant ds4_metal_glm_routed_moe_args &args,
        constant uint32_t &active_mask,
        device const uint64_t *gate_addrs,
        device const uint64_t *up_addrs,
        device const float *x,
        device const int32_t *selected,
        device const float *weights,
        device float *mid,
        threadgroup float *scratch [[threadgroup(0)]],
        uint3 tgpig [[threadgroup_position_in_grid]],
        ushort tiisg [[thread_index_in_simdgroup]],
        ushort sgitg [[simdgroup_index_in_threadgroup]]) {
    const uint slot = tgpig.y;
    const uint token = tgpig.z;
    if (slot >= args.n_expert_used || token >= args.n_tokens) return;
    if ((active_mask & (1u << slot)) == 0u) return;

    const uint64_t selected_off = (uint64_t)token * args.n_expert_used + slot;
    const int expert = selected[selected_off];
    const short NSG = 2;
    const uint row0 = ((uint)tgpig.x * (uint)NSG + (uint)sgitg) * N_R0_Q4_K;
    const uint64_t mid_base = (uint64_t)token * args.mid_token_stride +
                              (uint64_t)slot * args.mid_dim;
    if (row0 >= args.mid_dim) return;

    if (expert < 0 || (uint)expert >= args.n_total_expert) {
        if (tiisg == 0u) {
            for (short row = 0;
                 row < N_R0_Q4_K && row0 + (uint)row < args.mid_dim;
                 row++) {
                mid[mid_base + row0 + (uint)row] = 0.0f;
            }
        }
        return;
    }
    const uint64_t gate_addr = gate_addrs[(uint)expert];
    const uint64_t up_addr = up_addrs[(uint)expert];
    if (gate_addr == 0 || up_addr == 0) {
        if (tiisg == 0u) {
            for (short row = 0;
                 row < N_R0_Q4_K && row0 + (uint)row < args.mid_dim;
                 row++) {
                mid[mid_base + row0 + (uint)row] = 0.0f;
            }
        }
        return;
    }

    ds4_metal_glm_routed_moe_args local = args;
    local.n_total_expert = 1;
    local.gate_expert_bytes = 0;
    local.up_expert_bytes = 0;
    glm_q4_K_pair_swiglu_simd_f32_impl<N_R0_Q4_K>(
        local,
        reinterpret_cast<device const char *>(gate_addr),
        reinterpret_cast<device const char *>(up_addr),
        x, weights, mid, scratch,
        tgpig, slot, token, selected_off, 0, tiisg, sgitg);
}

kernel void kernel_glm_q4_K_slots6_pair_swiglu_f32(
        constant ds4_metal_glm_routed_moe_args &args,
        device const char *gate0,
        device const char *gate1,
        device const char *gate2,
        device const char *gate3,
        device const char *gate4,
        device const char *gate5,
        device const char *up0,
        device const char *up1,
        device const char *up2,
        device const char *up3,
        device const char *up4,
        device const char *up5,
        device const float *x,
        device const float *weights,
        device float *mid,
        threadgroup float *scratch [[threadgroup(0)]],
        uint3 tgpig [[threadgroup_position_in_grid]],
        ushort tiisg [[thread_index_in_simdgroup]],
        ushort sgitg [[simdgroup_index_in_threadgroup]]) {
    const uint slot = tgpig.y;
    const uint token = tgpig.z;
    if (slot >= args.n_expert_used || token >= args.n_tokens) return;

    device const char *gate_cur = gate0;
    device const char *up_cur = up0;
    switch (slot) {
    case 1: gate_cur = gate1; up_cur = up1; break;
    case 2: gate_cur = gate2; up_cur = up2; break;
    case 3: gate_cur = gate3; up_cur = up3; break;
    case 4: gate_cur = gate4; up_cur = up4; break;
    case 5: gate_cur = gate5; up_cur = up5; break;
    default: break;
    }

    const uint64_t selected_off = (uint64_t)token * args.n_expert_used + slot;
    glm_q4_K_pair_swiglu_simd_f32_impl<N_R0_Q4_K>(
        args, gate_cur, up_cur, x, weights, mid, scratch,
        tgpig, slot, token, selected_off, 0, tiisg, sgitg);
}

kernel void kernel_glm_q4_K_slots8_pair_swiglu_f32(
        constant ds4_metal_glm_routed_moe_args &args,
        device const char *gate0,
        device const char *gate1,
        device const char *gate2,
        device const char *gate3,
        device const char *gate4,
        device const char *gate5,
        device const char *gate6,
        device const char *gate7,
        device const char *up0,
        device const char *up1,
        device const char *up2,
        device const char *up3,
        device const char *up4,
        device const char *up5,
        device const char *up6,
        device const char *up7,
        device const float *x,
        device const float *weights,
        device float *mid,
        threadgroup float *scratch [[threadgroup(0)]],
        uint3 tgpig [[threadgroup_position_in_grid]],
        ushort tiisg [[thread_index_in_simdgroup]],
        ushort sgitg [[simdgroup_index_in_threadgroup]]) {
    const uint slot = tgpig.y;
    const uint token = tgpig.z;
    if (slot >= args.n_expert_used || token >= args.n_tokens) return;

    device const char *gate_cur = gate0;
    device const char *up_cur = up0;
    switch (slot) {
    case 1: gate_cur = gate1; up_cur = up1; break;
    case 2: gate_cur = gate2; up_cur = up2; break;
    case 3: gate_cur = gate3; up_cur = up3; break;
    case 4: gate_cur = gate4; up_cur = up4; break;
    case 5: gate_cur = gate5; up_cur = up5; break;
    case 6: gate_cur = gate6; up_cur = up6; break;
    case 7: gate_cur = gate7; up_cur = up7; break;
    default: break;
    }

    const uint64_t selected_off = (uint64_t)token * args.n_expert_used + slot;
    glm_q4_K_pair_swiglu_simd_f32_impl<N_R0_Q4_K>(
        args, gate_cur, up_cur, x, weights, mid, scratch,
        tgpig, slot, token, selected_off, 0, tiisg, sgitg);
}

kernel void kernel_glm_q4_K_slots6_pair_swiglu4_f32(
        constant ds4_metal_glm_routed_moe_args &args,
        device const char *gate0,
        device const char *gate1,
        device const char *gate2,
        device const char *gate3,
        device const char *gate4,
        device const char *gate5,
        device const char *up0,
        device const char *up1,
        device const char *up2,
        device const char *up3,
        device const char *up4,
        device const char *up5,
        device const float *x,
        device const float *weights,
        device float *mid,
        threadgroup float *scratch [[threadgroup(0)]],
        uint3 tgpig [[threadgroup_position_in_grid]],
        ushort tiisg [[thread_index_in_simdgroup]],
        ushort sgitg [[simdgroup_index_in_threadgroup]]) {
    const uint slot = tgpig.y;
    const uint token = tgpig.z;
    if (slot >= args.n_expert_used || token >= args.n_tokens) return;

    device const char *gate_cur = gate0;
    device const char *up_cur = up0;
    switch (slot) {
    case 1: gate_cur = gate1; up_cur = up1; break;
    case 2: gate_cur = gate2; up_cur = up2; break;
    case 3: gate_cur = gate3; up_cur = up3; break;
    case 4: gate_cur = gate4; up_cur = up4; break;
    case 5: gate_cur = gate5; up_cur = up5; break;
    default: break;
    }

    const uint64_t selected_off = (uint64_t)token * args.n_expert_used + slot;
    glm_q4_K_pair_swiglu_simd_f32_impl<N_R0_GLM_Q4_PAIR_K>(
        args, gate_cur, up_cur, x, weights, mid, scratch,
        tgpig, slot, token, selected_off, 0, tiisg, sgitg);
}

kernel void kernel_glm_q4_K_slots8_pair_swiglu4_f32(
        constant ds4_metal_glm_routed_moe_args &args,
        device const char *gate0,
        device const char *gate1,
        device const char *gate2,
        device const char *gate3,
        device const char *gate4,
        device const char *gate5,
        device const char *gate6,
        device const char *gate7,
        device const char *up0,
        device const char *up1,
        device const char *up2,
        device const char *up3,
        device const char *up4,
        device const char *up5,
        device const char *up6,
        device const char *up7,
        device const float *x,
        device const float *weights,
        device float *mid,
        threadgroup float *scratch [[threadgroup(0)]],
        uint3 tgpig [[threadgroup_position_in_grid]],
        ushort tiisg [[thread_index_in_simdgroup]],
        ushort sgitg [[simdgroup_index_in_threadgroup]]) {
    const uint slot = tgpig.y;
    const uint token = tgpig.z;
    if (slot >= args.n_expert_used || token >= args.n_tokens) return;

    device const char *gate_cur = gate0;
    device const char *up_cur = up0;
    switch (slot) {
    case 1: gate_cur = gate1; up_cur = up1; break;
    case 2: gate_cur = gate2; up_cur = up2; break;
    case 3: gate_cur = gate3; up_cur = up3; break;
    case 4: gate_cur = gate4; up_cur = up4; break;
    case 5: gate_cur = gate5; up_cur = up5; break;
    case 6: gate_cur = gate6; up_cur = up6; break;
    case 7: gate_cur = gate7; up_cur = up7; break;
    default: break;
    }

    const uint64_t selected_off = (uint64_t)token * args.n_expert_used + slot;
    glm_q4_K_pair_swiglu_simd_f32_impl<N_R0_GLM_Q4_PAIR_K>(
        args, gate_cur, up_cur, x, weights, mid, scratch,
        tgpig, slot, token, selected_off, 0, tiisg, sgitg);
}

kernel void kernel_glm_q4_K_pair_swiglu4_f32(
        constant ds4_metal_glm_routed_moe_args &args,
        device const char *gate,
        device const char *up,
        device const float *x,
        device const int32_t *selected,
        device const float *weights,
        device float *mid,
        threadgroup float *scratch [[threadgroup(0)]],
        uint3 tgpig [[threadgroup_position_in_grid]],
        ushort tiisg [[thread_index_in_simdgroup]],
        ushort sgitg [[simdgroup_index_in_threadgroup]]) {
    const uint slot = tgpig.y;
    const uint token = tgpig.z;
    if (slot >= args.n_expert_used || token >= args.n_tokens) return;
    const uint64_t selected_off = (uint64_t)token * args.n_expert_used + slot;
    const int expert = selected[selected_off];
    glm_q4_K_pair_swiglu_simd_f32_impl<N_R0_GLM_Q4_PAIR_K>(
        args, gate, up, x, weights, mid, scratch,
        tgpig, slot, token, selected_off, expert, tiisg, sgitg);
}

kernel void kernel_glm_q4_K_pair_swiglu2_mapped_f32(
        constant ds4_metal_glm_routed_moe_args &args,
        device const char *gate,
        device const char *up,
        device const float *x,
        device const uint32_t *htpe,
        device const int32_t *hids,
        device const float *weights,
        device float *mid,
        threadgroup float *scratch [[threadgroup(0)]],
        uint3 tgpig [[threadgroup_position_in_grid]],
        ushort tiisg [[thread_index_in_simdgroup]],
        ushort sgitg [[simdgroup_index_in_threadgroup]]) {
    const uint expert = tgpig.z;
    if (expert >= args.n_total_expert) return;
    const uint count = htpe[expert];
    const uint map_base = tgpig.y * 32u;
    for (uint i = 0; i < 32u; i++) {
        const uint map_row = map_base + i;
        if (map_row >= count) break;
        const int id = hids[(uint64_t)expert * args.n_tokens + map_row];
        if (id < 0) continue;
        const uint token = (uint)id / args.n_expert_used;
        const uint slot = (uint)id - token * args.n_expert_used;
        if (slot >= args.n_expert_used || token >= args.n_tokens) continue;
        const uint64_t selected_off = (uint64_t)token * args.n_expert_used + slot;
        glm_q4_K_pair_swiglu_simd_f32_impl<N_R0_Q4_K>(
            args, gate, up, x, weights, mid, scratch,
            tgpig, slot, token, selected_off, (int)expert, tiisg, sgitg);
    }
}

kernel void kernel_glm_q4_K_pair_swiglu2_mapped_row_f32(
        constant ds4_metal_glm_routed_moe_args &args,
        device const char *gate,
        device const char *up,
        device const float *x,
        device const uint32_t *htpe,
        device const int32_t *hids,
        device const float *weights,
        device float *mid,
        threadgroup float *scratch [[threadgroup(0)]],
        uint3 tgpig [[threadgroup_position_in_grid]],
        ushort tiisg [[thread_index_in_simdgroup]],
        ushort sgitg [[simdgroup_index_in_threadgroup]]) {
    const uint expert = tgpig.z;
    const uint map_row = tgpig.y;
    if (expert >= args.n_total_expert || map_row >= htpe[expert]) return;
    const int id = hids[(uint64_t)expert * args.n_tokens + map_row];
    if (id < 0) return;
    const uint token = (uint)id / args.n_expert_used;
    const uint slot = (uint)id - token * args.n_expert_used;
    if (slot >= args.n_expert_used || token >= args.n_tokens) return;
    const uint64_t selected_off = (uint64_t)token * args.n_expert_used + slot;
    glm_q4_K_pair_swiglu_simd_f32_impl<N_R0_Q4_K>(
        args, gate, up, x, weights, mid, scratch,
        tgpig, slot, token, selected_off, (int)expert, tiisg, sgitg);
}

static inline void glm_q5_K_pair_swiglu_f32_impl(
        constant ds4_metal_glm_routed_moe_args &args,
        device const char *gate,
        device const char *up,
        device const float *x,
        device const float *weights,
        device float *mid,
        threadgroup float *scratch,
        uint3 tgpig,
        uint slot,
        uint token,
        uint64_t selected_off,
        int expert,
        ushort tiisg,
        ushort sgitg) {
    const short NSG = 2;
    constexpr uint16_t kmask1 = 0x3f3f;
    constexpr uint16_t kmask2 = 0x0f0f;
    constexpr uint16_t kmask3 = 0xc0c0;

    const uint row0 = ((uint)tgpig.x * (uint)NSG + (uint)sgitg) * N_R0_Q5_PAIR_K;
    if (row0 >= args.mid_dim || slot >= args.n_expert_used || token >= args.n_tokens) return;

    const uint64_t mid_base = (uint64_t)token * args.mid_token_stride +
                              (uint64_t)slot * args.mid_dim;
    if (expert < 0 || (uint)expert >= args.n_total_expert) {
        if (tiisg == 0u) {
            for (short row = 0; row < N_R0_Q5_PAIR_K && row0 + (uint)row < args.mid_dim; row++) {
                mid[mid_base + row0 + (uint)row] = 0.0f;
            }
        }
        return;
    }

    const short ix = tiisg / 8;
    const short it = tiisg % 8;
    const short iq = it / 4;
    const short ir = it % 4;
    const uint bit0 = 2u * (uint)iq;
    const uint bit1 = bit0 + 1u;
    const uint bit2 = bit0 + 4u;
    const uint bit3 = bit0 + 5u;
    const int nb = args.in_dim / QK_K;

    const uint64_t expert_gate = (uint64_t)(uint)expert * args.gate_expert_bytes;
    const uint64_t expert_up = (uint64_t)(uint)expert * args.up_expert_bytes;
    device const block_q5_K *gate_rows =
        (device const block_q5_K *)(gate +
            expert_gate + (uint64_t)row0 * args.gate_row_bytes);
    device const block_q5_K *up_rows =
        (device const block_q5_K *)(up +
            expert_up + (uint64_t)row0 * args.up_row_bytes);
    device const float *y = x + (uint64_t)token * args.in_dim;
    device const float *y4 = y + ix * QK_K + 64 * iq + 8 * ir;

    float sumg[N_R0_Q5_PAIR_K] = {0.f};
    float sumu[N_R0_Q5_PAIR_K] = {0.f};
    uint16_t sc16[4];
    thread const uint8_t *sc8 = (thread const uint8_t *)sc16;

    for (int ib = ix; ib < nb; ib += 4) {
        float yl[16];
        float yh[16];
        float4 sumy = {0.f, 0.f, 0.f, 0.f};

        for (short i = 0; i < 8; ++i) {
            yl[i + 0] = y4[i +   0]; sumy[0] += yl[i + 0];
            yl[i + 8] = y4[i +  32]; sumy[1] += yl[i + 8];
            yh[i + 0] = y4[i + 128]; sumy[2] += yh[i + 0];
            yh[i + 8] = y4[i + 160]; sumy[3] += yh[i + 8];
        }

        device const uint16_t *scg = (device const uint16_t *)gate_rows[ib].scales + iq;
        device const uint16_t *qg1 = (device const uint16_t *)gate_rows[ib].qs + 16 * iq + 4 * ir;
        device const uint16_t *qhg = (device const uint16_t *)gate_rows[ib].qh + 4 * ir;
        device const half *dhg = &gate_rows[ib].d;
        device const uint16_t *scu = (device const uint16_t *)up_rows[ib].scales + iq;
        device const uint16_t *qu1 = (device const uint16_t *)up_rows[ib].qs + 16 * iq + 4 * ir;
        device const uint16_t *qhu = (device const uint16_t *)up_rows[ib].qh + 4 * ir;
        device const half *dhu = &up_rows[ib].d;

        for (short row = 0; row < N_R0_Q5_PAIR_K && row0 + (uint)row < args.mid_dim; row++) {
            sc16[0] = scg[0] & kmask1;
            sc16[1] = scg[2] & kmask1;
            sc16[2] = ((scg[4] >> 0) & kmask2) | ((scg[0] & kmask3) >> 2);
            sc16[3] = ((scg[4] >> 4) & kmask2) | ((scg[2] & kmask3) >> 2);

            device const uint16_t *qg2 = qg1 + 32;
            float4 accg = {0.f, 0.f, 0.f, 0.f};

            FOR_UNROLL (short i = 0; i < 4; ++i) {
                const uint ql1 = (uint)qg1[i];
                const uint ql2 = (uint)qg2[i];
                const uint hb = (uint)qhg[i];
                accg[0] += yl[2 * i + 0] *
                               (float)((ql1 & 0x000Fu) + (((hb >> bit0) & 1u) << 4u)) +
                           yl[2 * i + 1] *
                               (float)(((ql1 >> 8u) & 0x000Fu) + (((hb >> (bit0 + 8u)) & 1u) << 4u));
                accg[1] += yl[2 * i + 8] *
                               (float)(((ql1 >> 4u) & 0x000Fu) + (((hb >> bit1) & 1u) << 4u)) +
                           yl[2 * i + 9] *
                               (float)(((ql1 >> 12u) & 0x000Fu) + (((hb >> (bit1 + 8u)) & 1u) << 4u));
                accg[2] += yh[2 * i + 0] *
                               (float)((ql2 & 0x000Fu) + (((hb >> bit2) & 1u) << 4u)) +
                           yh[2 * i + 1] *
                               (float)(((ql2 >> 8u) & 0x000Fu) + (((hb >> (bit2 + 8u)) & 1u) << 4u));
                accg[3] += yh[2 * i + 8] *
                               (float)(((ql2 >> 4u) & 0x000Fu) + (((hb >> bit3) & 1u) << 4u)) +
                           yh[2 * i + 9] *
                               (float)(((ql2 >> 12u) & 0x000Fu) + (((hb >> (bit3 + 8u)) & 1u) << 4u));
            }

            sumg[row] += dhg[0] * (accg[0] * sc8[0] + accg[1] * sc8[1] +
                                   accg[2] * sc8[4] + accg[3] * sc8[5]) -
                         dhg[1] * (sumy[0] * sc8[2] + sumy[1] * sc8[3] +
                                   sumy[2] * sc8[6] + sumy[3] * sc8[7]);

            sc16[0] = scu[0] & kmask1;
            sc16[1] = scu[2] & kmask1;
            sc16[2] = ((scu[4] >> 0) & kmask2) | ((scu[0] & kmask3) >> 2);
            sc16[3] = ((scu[4] >> 4) & kmask2) | ((scu[2] & kmask3) >> 2);

            device const uint16_t *qu2 = qu1 + 32;
            float4 accu = {0.f, 0.f, 0.f, 0.f};

            FOR_UNROLL (short i = 0; i < 4; ++i) {
                const uint ql1 = (uint)qu1[i];
                const uint ql2 = (uint)qu2[i];
                const uint hb = (uint)qhu[i];
                accu[0] += yl[2 * i + 0] *
                               (float)((ql1 & 0x000Fu) + (((hb >> bit0) & 1u) << 4u)) +
                           yl[2 * i + 1] *
                               (float)(((ql1 >> 8u) & 0x000Fu) + (((hb >> (bit0 + 8u)) & 1u) << 4u));
                accu[1] += yl[2 * i + 8] *
                               (float)(((ql1 >> 4u) & 0x000Fu) + (((hb >> bit1) & 1u) << 4u)) +
                           yl[2 * i + 9] *
                               (float)(((ql1 >> 12u) & 0x000Fu) + (((hb >> (bit1 + 8u)) & 1u) << 4u));
                accu[2] += yh[2 * i + 0] *
                               (float)((ql2 & 0x000Fu) + (((hb >> bit2) & 1u) << 4u)) +
                           yh[2 * i + 1] *
                               (float)(((ql2 >> 8u) & 0x000Fu) + (((hb >> (bit2 + 8u)) & 1u) << 4u));
                accu[3] += yh[2 * i + 8] *
                               (float)(((ql2 >> 4u) & 0x000Fu) + (((hb >> bit3) & 1u) << 4u)) +
                           yh[2 * i + 9] *
                               (float)(((ql2 >> 12u) & 0x000Fu) + (((hb >> (bit3 + 8u)) & 1u) << 4u));
            }

            sumu[row] += dhu[0] * (accu[0] * sc8[0] + accu[1] * sc8[1] +
                                   accu[2] * sc8[4] + accu[3] * sc8[5]) -
                         dhu[1] * (sumy[0] * sc8[2] + sumy[1] * sc8[3] +
                                   sumy[2] * sc8[6] + sumy[3] * sc8[7]);

            qg1 += args.gate_row_bytes / 2;
            qhg += args.gate_row_bytes / 2;
            scg += args.gate_row_bytes / 2;
            dhg += args.gate_row_bytes / 2;
            qu1 += args.up_row_bytes / 2;
            qhu += args.up_row_bytes / 2;
            scu += args.up_row_bytes / 2;
            dhu += args.up_row_bytes / 2;
        }

        y4 += 4 * QK_K;
    }

    for (short row = 0; row < N_R0_Q5_PAIR_K && row0 + (uint)row < args.mid_dim; row++) {
        const float g = simd_sum(sumg[row]);
        const float u = simd_sum(sumu[row]);
        if (tiisg == 0u) {
            const float sw = g / (1.0f + exp(-g));
            mid[mid_base + row0 + (uint)row] = sw * u * weights[selected_off];
        }
    }

    (void)scratch;
}

kernel void kernel_glm_q5_K_pair_swiglu_f32(
        constant ds4_metal_glm_routed_moe_args &args,
        device const char *gate,
        device const char *up,
        device const float *x,
        device const int32_t *selected,
        device const float *weights,
        device float *mid,
        threadgroup float *scratch [[threadgroup(0)]],
        uint3 tgpig [[threadgroup_position_in_grid]],
        ushort tiisg [[thread_index_in_simdgroup]],
        ushort sgitg [[simdgroup_index_in_threadgroup]]) {
    const uint slot = tgpig.y;
    const uint token = tgpig.z;
    if (slot >= args.n_expert_used || token >= args.n_tokens) return;
    const uint64_t selected_off = (uint64_t)token * args.n_expert_used + slot;
    const int expert = selected[selected_off];
    glm_q5_K_pair_swiglu_f32_impl(
        args, gate, up, x, weights, mid, scratch,
        tgpig, slot, token, selected_off, expert, tiisg, sgitg);
}

kernel void kernel_glm_q5_K_slots6_pair_swiglu_f32(
        constant ds4_metal_glm_routed_moe_args &args,
        device const char *gate0,
        device const char *gate1,
        device const char *gate2,
        device const char *gate3,
        device const char *gate4,
        device const char *gate5,
        device const char *up0,
        device const char *up1,
        device const char *up2,
        device const char *up3,
        device const char *up4,
        device const char *up5,
        device const float *x,
        device const float *weights,
        device float *mid,
        threadgroup float *scratch [[threadgroup(0)]],
        uint3 tgpig [[threadgroup_position_in_grid]],
        ushort tiisg [[thread_index_in_simdgroup]],
        ushort sgitg [[simdgroup_index_in_threadgroup]]) {
    const uint slot = tgpig.y;
    const uint token = tgpig.z;
    if (slot >= args.n_expert_used || token >= args.n_tokens) return;

    device const char *gate_cur = gate0;
    device const char *up_cur = up0;
    switch (slot) {
    case 1: gate_cur = gate1; up_cur = up1; break;
    case 2: gate_cur = gate2; up_cur = up2; break;
    case 3: gate_cur = gate3; up_cur = up3; break;
    case 4: gate_cur = gate4; up_cur = up4; break;
    case 5: gate_cur = gate5; up_cur = up5; break;
    default: break;
    }

    const uint64_t selected_off = (uint64_t)token * args.n_expert_used + slot;
    glm_q5_K_pair_swiglu_f32_impl(
        args, gate_cur, up_cur, x, weights, mid, scratch,
        tgpig, slot, token, selected_off, 0, tiisg, sgitg);
}

kernel void kernel_glm_q5_K_slots8_pair_swiglu_f32(
        constant ds4_metal_glm_routed_moe_args &args,
        device const char *gate0,
        device const char *gate1,
        device const char *gate2,
        device const char *gate3,
        device const char *gate4,
        device const char *gate5,
        device const char *gate6,
        device const char *gate7,
        device const char *up0,
        device const char *up1,
        device const char *up2,
        device const char *up3,
        device const char *up4,
        device const char *up5,
        device const char *up6,
        device const char *up7,
        device const float *x,
        device const float *weights,
        device float *mid,
        threadgroup float *scratch [[threadgroup(0)]],
        uint3 tgpig [[threadgroup_position_in_grid]],
        ushort tiisg [[thread_index_in_simdgroup]],
        ushort sgitg [[simdgroup_index_in_threadgroup]]) {
    const uint slot = tgpig.y;
    const uint token = tgpig.z;
    if (slot >= args.n_expert_used || token >= args.n_tokens) return;

    device const char *gate_cur = gate0;
    device const char *up_cur = up0;
    switch (slot) {
    case 1: gate_cur = gate1; up_cur = up1; break;
    case 2: gate_cur = gate2; up_cur = up2; break;
    case 3: gate_cur = gate3; up_cur = up3; break;
    case 4: gate_cur = gate4; up_cur = up4; break;
    case 5: gate_cur = gate5; up_cur = up5; break;
    case 6: gate_cur = gate6; up_cur = up6; break;
    case 7: gate_cur = gate7; up_cur = up7; break;
    default: break;
    }

    const uint64_t selected_off = (uint64_t)token * args.n_expert_used + slot;
    glm_q5_K_pair_swiglu_f32_impl(
        args, gate_cur, up_cur, x, weights, mid, scratch,
        tgpig, slot, token, selected_off, 0, tiisg, sgitg);
}

kernel void kernel_glm_q5_K_pair_swiglu_mapped_f32(
        constant ds4_metal_glm_routed_moe_args &args,
        device const char *gate,
        device const char *up,
        device const float *x,
        device const uint32_t *htpe,
        device const int32_t *hids,
        device const float *weights,
        device float *mid,
        threadgroup float *scratch [[threadgroup(0)]],
        uint3 tgpig [[threadgroup_position_in_grid]],
        ushort tiisg [[thread_index_in_simdgroup]],
        ushort sgitg [[simdgroup_index_in_threadgroup]]) {
    const uint expert = tgpig.z;
    if (expert >= args.n_total_expert) return;
    const uint count = htpe[expert];
    const uint map_base = tgpig.y * 32u;
    for (uint i = 0; i < 32u; i++) {
        const uint map_row = map_base + i;
        if (map_row >= count) break;
        const int id = hids[(uint64_t)expert * args.n_tokens + map_row];
        if (id < 0) continue;
        const uint token = (uint)id / args.n_expert_used;
        const uint slot = (uint)id - token * args.n_expert_used;
        if (slot >= args.n_expert_used || token >= args.n_tokens) continue;
        const uint64_t selected_off = (uint64_t)token * args.n_expert_used + slot;
        glm_q5_K_pair_swiglu_f32_impl(
            args, gate, up, x, weights, mid, scratch,
            tgpig, slot, token, selected_off, (int)expert, tiisg, sgitg);
    }
}

kernel void kernel_glm_q5_K_pair_swiglu_mapped_row_f32(
        constant ds4_metal_glm_routed_moe_args &args,
        device const char *gate,
        device const char *up,
        device const float *x,
        device const uint32_t *htpe,
        device const int32_t *hids,
        device const float *weights,
        device float *mid,
        threadgroup float *scratch [[threadgroup(0)]],
        uint3 tgpig [[threadgroup_position_in_grid]],
        ushort tiisg [[thread_index_in_simdgroup]],
        ushort sgitg [[simdgroup_index_in_threadgroup]]) {
    const uint expert = tgpig.z;
    const uint map_row = tgpig.y;
    if (expert >= args.n_total_expert || map_row >= htpe[expert]) return;
    const int id = hids[(uint64_t)expert * args.n_tokens + map_row];
    if (id < 0) return;
    const uint token = (uint)id / args.n_expert_used;
    const uint slot = (uint)id - token * args.n_expert_used;
    if (slot >= args.n_expert_used || token >= args.n_tokens) return;
    const uint64_t selected_off = (uint64_t)token * args.n_expert_used + slot;
    glm_q5_K_pair_swiglu_f32_impl(
        args, gate, up, x, weights, mid, scratch,
        tgpig, slot, token, selected_off, (int)expert, tiisg, sgitg);
}

kernel void kernel_glm_q5_K_down_f32(
        constant ds4_metal_glm_routed_moe_args &args,
        device const char *down,
        device const int32_t *selected,
        device const float *mid,
        device float *out,
        uint3 tgpig [[threadgroup_position_in_grid]],
        ushort tiisg [[thread_index_in_simdgroup]],
        ushort sgitg [[simdgroup_index_in_threadgroup]]) {
    const short NSG = 2;
    constexpr uint16_t kmask1 = 0x3f3f;
    constexpr uint16_t kmask2 = 0x0f0f;
    constexpr uint16_t kmask3 = 0xc0c0;

    const uint row0 = ((uint)tgpig.x * (uint)NSG + (uint)sgitg) * N_R0_Q5_K;
    const uint token = tgpig.y;
    if (row0 >= args.out_dim || token >= args.n_tokens) return;

    const short ix = tiisg / 8;
    const short it = tiisg % 8;
    const short iq = it / 4;
    const short ir = it % 4;
    const uint bit0 = 2u * (uint)iq;
    const uint bit1 = bit0 + 1u;
    const uint bit2 = bit0 + 4u;
    const uint bit3 = bit0 + 5u;
    const int nb = args.mid_dim / QK_K;

    float sumf[N_R0_Q5_K] = {0.f};
    uint16_t sc16[4];
    thread const uint8_t *sc8 = (thread const uint8_t *)sc16;
    const uint64_t selected_base = (uint64_t)token * args.n_expert_used;
    const uint64_t mid_base = (uint64_t)token * args.mid_token_stride;
    for (uint slot = 0; slot < args.n_expert_used; slot++) {
        const int expert = selected[selected_base + slot];
        if (expert < 0 || (uint)expert >= args.n_total_expert) continue;
        device const block_q5_K *x =
            (device const block_q5_K *)(down +
                (uint64_t)(uint)expert * args.down_expert_bytes +
                (uint64_t)row0 * args.down_row_bytes);
        device const float *y = mid + mid_base + (uint64_t)slot * args.mid_dim;
        device const float *y4 = y + ix * QK_K + 64 * iq + 8 * ir;

        for (int ib = ix; ib < nb; ib += 4) {
            float yl[16];
            float yh[16];
            float4 sumy = {0.f, 0.f, 0.f, 0.f};

            for (short i = 0; i < 8; ++i) {
                yl[i + 0] = y4[i +   0]; sumy[0] += yl[i + 0];
                yl[i + 8] = y4[i +  32]; sumy[1] += yl[i + 8];
                yh[i + 0] = y4[i + 128]; sumy[2] += yh[i + 0];
                yh[i + 8] = y4[i + 160]; sumy[3] += yh[i + 8];
            }

            device const uint16_t *sc = (device const uint16_t *)x[ib].scales + iq;
            device const uint16_t *q1 = (device const uint16_t *)x[ib].qs + 16 * iq + 4 * ir;
            device const uint16_t *qh = (device const uint16_t *)x[ib].qh + 4 * ir;
            device const half *dh = &x[ib].d;

            for (short row = 0; row < N_R0_Q5_K && row0 + (uint)row < args.out_dim; row++) {
                sc16[0] = sc[0] & kmask1;
                sc16[1] = sc[2] & kmask1;
                sc16[2] = ((sc[4] >> 0) & kmask2) | ((sc[0] & kmask3) >> 2);
                sc16[3] = ((sc[4] >> 4) & kmask2) | ((sc[2] & kmask3) >> 2);

                device const uint16_t *q2 = q1 + 32;
                float4 acc = {0.f, 0.f, 0.f, 0.f};

                FOR_UNROLL (short i = 0; i < 4; ++i) {
                    const uint ql1 = (uint)q1[i];
                    const uint ql2 = (uint)q2[i];
                    const uint hb = (uint)qh[i];
                    acc[0] += yl[2 * i + 0] *
                                  (float)((ql1 & 0x000Fu) + (((hb >> bit0) & 1u) << 4u)) +
                              yl[2 * i + 1] *
                                  (float)(((ql1 >> 8u) & 0x000Fu) + (((hb >> (bit0 + 8u)) & 1u) << 4u));
                    acc[1] += yl[2 * i + 8] *
                                  (float)(((ql1 >> 4u) & 0x000Fu) + (((hb >> bit1) & 1u) << 4u)) +
                              yl[2 * i + 9] *
                                  (float)(((ql1 >> 12u) & 0x000Fu) + (((hb >> (bit1 + 8u)) & 1u) << 4u));
                    acc[2] += yh[2 * i + 0] *
                                  (float)((ql2 & 0x000Fu) + (((hb >> bit2) & 1u) << 4u)) +
                              yh[2 * i + 1] *
                                  (float)(((ql2 >> 8u) & 0x000Fu) + (((hb >> (bit2 + 8u)) & 1u) << 4u));
                    acc[3] += yh[2 * i + 8] *
                                  (float)(((ql2 >> 4u) & 0x000Fu) + (((hb >> bit3) & 1u) << 4u)) +
                              yh[2 * i + 9] *
                                  (float)(((ql2 >> 12u) & 0x000Fu) + (((hb >> (bit3 + 8u)) & 1u) << 4u));
                }

                sumf[row] += dh[0] * (acc[0] * sc8[0] + acc[1] * sc8[1] +
                                      acc[2] * sc8[4] + acc[3] * sc8[5]) -
                             dh[1] * (sumy[0] * sc8[2] + sumy[1] * sc8[3] +
                                      sumy[2] * sc8[6] + sumy[3] * sc8[7]);

                q1 += args.down_row_bytes / 2;
                qh += args.down_row_bytes / 2;
                sc += args.down_row_bytes / 2;
                dh += args.down_row_bytes / 2;
            }

            y4 += 4 * QK_K;
        }
    }

    for (short row = 0; row < N_R0_Q5_K && row0 + (uint)row < args.out_dim; row++) {
        const float sum_all = simd_sum(sumf[row]);
        if (tiisg == 0u) {
            out[(uint64_t)token * args.out_dim + row0 + (uint)row] = sum_all;
        }
    }
}

kernel void kernel_glm_q2_K_down_f32(
        constant ds4_metal_glm_routed_moe_args &args,
        device const char *down,
        device const int32_t *selected,
        device const float *mid,
        device float *out,
        threadgroup float *scratch [[threadgroup(0)]],
        uint3 tgpig [[threadgroup_position_in_grid]],
        ushort tiisg [[thread_index_in_simdgroup]],
        ushort sgitg [[simdgroup_index_in_threadgroup]]) {
    const short NSG = 2;
    const uint row0 = ((uint)tgpig.x * (uint)NSG + (uint)sgitg) * N_R0_Q2_K;
    const uint token = tgpig.y;
    if (row0 >= args.out_dim || token >= args.n_tokens) return;

    const short ix = tiisg / 8;
    const short it = tiisg % 8;
    const short iq = it / 4;
    const short ir = it % 4;
    const short is = (8 * ir) / 16;
    const int nb = args.mid_dim / QK_K;
    float sumf[N_R0_Q2_K] = {0.f};
    const uint64_t selected_base = (uint64_t)token * args.n_expert_used;
    const uint64_t mid_base = (uint64_t)token * args.mid_token_stride;
    for (uint slot = 0; slot < args.n_expert_used; slot++) {
        const int expert = selected[selected_base + slot];
        if (expert < 0 || (uint)expert >= args.n_total_expert) continue;
        if (!ds4_tp_owns_expert(expert, args.n_total_expert,
                                args.tp_rank, args.tp_world)) continue;
        device const block_q2_K *x =
            (device const block_q2_K *)(down +
                (uint64_t)(uint)(expert - args.tp_expert_base) * args.down_expert_bytes +
                (uint64_t)row0 * args.down_row_bytes);
        device const float *y = mid + mid_base + (uint64_t)slot * args.mid_dim;
        device const float *y4 = y + ix * QK_K + 128 * iq + 8 * ir;

        for (int ib = ix; ib < nb; ib += 4) {
            float yl[32];
            float4 sumy = {0.f, 0.f, 0.f, 0.f};
            for (short i = 0; i < 8; ++i) {
                yl[i +  0] = y4[i +  0]; sumy[0] += yl[i +  0];
                yl[i +  8] = y4[i + 32]; sumy[1] += yl[i +  8];
                yl[i + 16] = y4[i + 64]; sumy[2] += yl[i + 16];
                yl[i + 24] = y4[i + 96]; sumy[3] += yl[i + 24];
            }

            device const uint8_t *sc = (device const uint8_t *)x[ib].scales + 8 * iq + is;
            device const uint16_t *qs = (device const uint16_t *)x[ib].qs + 16 * iq + 4 * ir;
            device const half *dh = &x[ib].d;
            for (short row = 0; row < N_R0_Q2_K && row0 + (uint)row < args.out_dim; row++) {
                float4 acc1 = {0.f, 0.f, 0.f, 0.f};
                float4 acc2 = {0.f, 0.f, 0.f, 0.f};

                for (int i = 0; i < 8; i += 2) {
                    acc1[0] += yl[i +  0] * (qs[i / 2] & 0x0003);
                    acc2[0] += yl[i +  1] * (qs[i / 2] & 0x0300);
                    acc1[1] += yl[i +  8] * (qs[i / 2] & 0x000c);
                    acc2[1] += yl[i +  9] * (qs[i / 2] & 0x0c00);
                    acc1[2] += yl[i + 16] * (qs[i / 2] & 0x0030);
                    acc2[2] += yl[i + 17] * (qs[i / 2] & 0x3000);
                    acc1[3] += yl[i + 24] * (qs[i / 2] & 0x00c0);
                    acc2[3] += yl[i + 25] * (qs[i / 2] & 0xc000);
                }

                const float d = dh[0];
                const float m = dh[1] * 1.f / 16.f;
                sumf[row] += d * ((acc1[0] + 1.f / 256.f * acc2[0]) * (sc[0] & 0xF) * 1.f /  1.f +
                                  (acc1[1] + 1.f / 256.f * acc2[1]) * (sc[2] & 0xF) * 1.f /  4.f +
                                  (acc1[2] + 1.f / 256.f * acc2[2]) * (sc[4] & 0xF) * 1.f / 16.f +
                                  (acc1[3] + 1.f / 256.f * acc2[3]) * (sc[6] & 0xF) * 1.f / 64.f) -
                             m * (sumy[0] * (sc[0] & 0xF0) + sumy[1] * (sc[2] & 0xF0) +
                                  sumy[2] * (sc[4] & 0xF0) + sumy[3] * (sc[6] & 0xF0));

                qs += args.down_row_bytes / 2;
                sc += args.down_row_bytes;
                dh += args.down_row_bytes / 2;
            }

            y4 += 4 * QK_K;
        }
    }

    for (short row = 0; row < N_R0_Q2_K && row0 + (uint)row < args.out_dim; row++) {
        const float sum_all = simd_sum(sumf[row]);
        if (tiisg == 0u) {
            out[(uint64_t)token * args.out_dim + row0 + (uint)row] = sum_all;
        }
    }

    (void)scratch;
}

kernel void kernel_glm_q2_K_addr_down_f32(
        constant ds4_metal_glm_routed_moe_args &args,
        device const uint64_t *down_addrs,
        device const int32_t *selected,
        device const float *mid,
        device float *out,
        threadgroup float *scratch [[threadgroup(0)]],
        uint3 tgpig [[threadgroup_position_in_grid]],
        ushort tiisg [[thread_index_in_simdgroup]],
        ushort sgitg [[simdgroup_index_in_threadgroup]]) {
    const short NSG = 2;
    const uint row0 = ((uint)tgpig.x * (uint)NSG + (uint)sgitg) * N_R0_Q2_K;
    const uint token = tgpig.y;
    if (row0 >= args.out_dim || token >= args.n_tokens) return;

    const short ix = tiisg / 8;
    const short it = tiisg % 8;
    const short iq = it / 4;
    const short ir = it % 4;
    const short is = (8 * ir) / 16;
    const int nb = args.mid_dim / QK_K;
    float sumf[N_R0_Q2_K] = {0.f};
    const uint64_t selected_base = (uint64_t)token * args.n_expert_used;
    const uint64_t mid_base = (uint64_t)token * args.mid_token_stride;
    for (uint slot = 0; slot < args.n_expert_used; slot++) {
        const int expert = selected[selected_base + slot];
        if (expert < 0 || (uint)expert >= args.n_total_expert) continue;
        const uint64_t down_addr = down_addrs[(uint)expert];
        if (down_addr == 0) continue;
        device const block_q2_K *x =
            (device const block_q2_K *)(reinterpret_cast<device const char *>(down_addr) +
                (uint64_t)row0 * args.down_row_bytes);
        device const float *y = mid + mid_base + (uint64_t)slot * args.mid_dim;
        device const float *y4 = y + ix * QK_K + 128 * iq + 8 * ir;

        for (int ib = ix; ib < nb; ib += 4) {
            float yl[32];
            float4 sumy = {0.f, 0.f, 0.f, 0.f};
            for (short i = 0; i < 8; ++i) {
                yl[i +  0] = y4[i +  0]; sumy[0] += yl[i +  0];
                yl[i +  8] = y4[i + 32]; sumy[1] += yl[i +  8];
                yl[i + 16] = y4[i + 64]; sumy[2] += yl[i + 16];
                yl[i + 24] = y4[i + 96]; sumy[3] += yl[i + 24];
            }

            device const uint8_t *sc = (device const uint8_t *)x[ib].scales + 8 * iq + is;
            device const uint16_t *qs = (device const uint16_t *)x[ib].qs + 16 * iq + 4 * ir;
            device const half *dh = &x[ib].d;
            for (short row = 0; row < N_R0_Q2_K && row0 + (uint)row < args.out_dim; row++) {
                float4 acc1 = {0.f, 0.f, 0.f, 0.f};
                float4 acc2 = {0.f, 0.f, 0.f, 0.f};

                for (int i = 0; i < 8; i += 2) {
                    acc1[0] += yl[i +  0] * (qs[i / 2] & 0x0003);
                    acc2[0] += yl[i +  1] * (qs[i / 2] & 0x0300);
                    acc1[1] += yl[i +  8] * (qs[i / 2] & 0x000c);
                    acc2[1] += yl[i +  9] * (qs[i / 2] & 0x0c00);
                    acc1[2] += yl[i + 16] * (qs[i / 2] & 0x0030);
                    acc2[2] += yl[i + 17] * (qs[i / 2] & 0x3000);
                    acc1[3] += yl[i + 24] * (qs[i / 2] & 0x00c0);
                    acc2[3] += yl[i + 25] * (qs[i / 2] & 0xc000);
                }

                const float d = dh[0];
                const float m = dh[1] * 1.f / 16.f;
                sumf[row] += d * ((acc1[0] + 1.f / 256.f * acc2[0]) * (sc[0] & 0xF) * 1.f /  1.f +
                                  (acc1[1] + 1.f / 256.f * acc2[1]) * (sc[2] & 0xF) * 1.f /  4.f +
                                  (acc1[2] + 1.f / 256.f * acc2[2]) * (sc[4] & 0xF) * 1.f / 16.f +
                                  (acc1[3] + 1.f / 256.f * acc2[3]) * (sc[6] & 0xF) * 1.f / 64.f) -
                             m * (sumy[0] * (sc[0] & 0xF0) + sumy[1] * (sc[2] & 0xF0) +
                                  sumy[2] * (sc[4] & 0xF0) + sumy[3] * (sc[6] & 0xF0));

                qs += args.down_row_bytes / 2;
                sc += args.down_row_bytes;
                dh += args.down_row_bytes / 2;
            }

            y4 += 4 * QK_K;
        }
    }

    for (short row = 0; row < N_R0_Q2_K && row0 + (uint)row < args.out_dim; row++) {
        const float sum_all = simd_sum(sumf[row]);
        if (tiisg == 0u) {
            out[(uint64_t)token * args.out_dim + row0 + (uint)row] = sum_all;
        }
    }

    (void)scratch;
}

kernel void kernel_glm_q4_K_down_f32(
        constant ds4_metal_glm_routed_moe_args &args,
        device const char *down,
        device const int32_t *selected,
        device const float *mid,
        device float *out,
        threadgroup float *scratch [[threadgroup(0)]],
        uint3 tgpig [[threadgroup_position_in_grid]],
        uint tid [[thread_index_in_threadgroup]]) {
    const uint ntg = 256u;
    const uint row = tgpig.x;
    const uint token = tgpig.y;
    if (row >= args.out_dim || token >= args.n_tokens) return;

    float acc = 0.0f;
    const uint64_t selected_base = (uint64_t)token * args.n_expert_used;
    const uint64_t mid_base = (uint64_t)token * args.mid_token_stride;
    for (uint slot = 0; slot < args.n_expert_used; slot++) {
        const int expert = selected[selected_base + slot];
        if (expert < 0 || (uint)expert >= args.n_total_expert) continue;
        device const block_q4_K *down_row =
            (device const block_q4_K *)(down +
                (uint64_t)(uint)expert * args.down_expert_bytes +
                (uint64_t)row * args.down_row_bytes);
        device const float *slot_mid = mid + mid_base + (uint64_t)slot * args.mid_dim;
        for (uint k = tid; k < args.mid_dim; k += ntg) {
            acc += ds4_glm_q4_K_value(down_row, k) * slot_mid[k];
        }
    }

    scratch[tid] = acc;
    threadgroup_barrier(mem_flags::mem_threadgroup);
    for (uint stride = ntg >> 1u; stride > 0u; stride >>= 1u) {
        if (tid < stride) scratch[tid] += scratch[tid + stride];
        threadgroup_barrier(mem_flags::mem_threadgroup);
    }
    if (tid == 0u) {
        out[(uint64_t)token * args.out_dim + row] = scratch[0];
    }
}

kernel void kernel_glm_q4_K_addr_down_f32(
        constant ds4_metal_glm_routed_moe_args &args,
        device const uint64_t *down_addrs,
        device const int32_t *selected,
        device const float *mid,
        device float *out,
        threadgroup float *scratch [[threadgroup(0)]],
        uint3 tgpig [[threadgroup_position_in_grid]],
        uint tid [[thread_index_in_threadgroup]]) {
    const uint ntg = 256u;
    const uint row = tgpig.x;
    const uint token = tgpig.y;
    if (row >= args.out_dim || token >= args.n_tokens) return;

    float acc = 0.0f;
    const uint64_t selected_base = (uint64_t)token * args.n_expert_used;
    const uint64_t mid_base = (uint64_t)token * args.mid_token_stride;
    for (uint slot = 0; slot < args.n_expert_used; slot++) {
        const int expert = selected[selected_base + slot];
        if (expert < 0 || (uint)expert >= args.n_total_expert) continue;
        const uint64_t down_addr = down_addrs[(uint)expert];
        if (down_addr == 0) continue;
        device const block_q4_K *down_row =
            (device const block_q4_K *)(reinterpret_cast<device const char *>(down_addr) +
                (uint64_t)row * args.down_row_bytes);
        device const float *slot_mid = mid + mid_base + (uint64_t)slot * args.mid_dim;
        for (uint k = tid; k < args.mid_dim; k += ntg) {
            acc += ds4_glm_q4_K_value(down_row, k) * slot_mid[k];
        }
    }

    scratch[tid] = acc;
    threadgroup_barrier(mem_flags::mem_threadgroup);
    for (uint stride = ntg >> 1u; stride > 0u; stride >>= 1u) {
        if (tid < stride) scratch[tid] += scratch[tid + stride];
        threadgroup_barrier(mem_flags::mem_threadgroup);
    }
    if (tid == 0u) {
        out[(uint64_t)token * args.out_dim + row] = scratch[0];
    }
}

kernel void kernel_glm_q4_K_down_simd_f32(
        constant ds4_metal_glm_routed_moe_args &args,
        device const char *down,
        device const int32_t *selected,
        device const float *mid,
        device float *out,
        uint3 tgpig [[threadgroup_position_in_grid]],
        ushort tiisg [[thread_index_in_simdgroup]],
        ushort sgitg [[simdgroup_index_in_threadgroup]]) {
    const short NSG = 2;
    const short nr0 = N_R0_Q4_K;
    const int nb = args.mid_dim / QK_K;
    const uint row0 = ((uint)tgpig.x * (uint)NSG + (uint)sgitg) * (uint)nr0;
    const uint token = tgpig.y;
    if (row0 >= args.out_dim || token >= args.n_tokens) return;

    constexpr uint16_t kmask1 = 0x3f3f;
    constexpr uint16_t kmask2 = 0x0f0f;
    constexpr uint16_t kmask3 = 0xc0c0;

    const short ix = tiisg / 8;
    const short it = tiisg % 8;
    const short iq = it / 4;
    const short ir = it % 4;

    float sumf[N_R0_Q4_K] = {0.f};
    uint16_t sc16[4];
    thread const uint8_t *sc8 = (thread const uint8_t *)sc16;

    const uint64_t selected_base = (uint64_t)token * args.n_expert_used;
    const uint64_t mid_base = (uint64_t)token * args.mid_token_stride;
    for (uint slot = 0; slot < args.n_expert_used; slot++) {
        const int expert = selected[selected_base + slot];
        if (expert < 0 || (uint)expert >= args.n_total_expert) continue;

        device const block_q4_K *x =
            (device const block_q4_K *)(down +
                (uint64_t)(uint)expert * args.down_expert_bytes +
                (uint64_t)row0 * args.down_row_bytes);
        device const float *y = mid + mid_base + (uint64_t)slot * args.mid_dim;
        device const float *y4 = y + ix * QK_K + 64 * iq + 8 * ir;

        for (int ib = ix; ib < nb; ib += 4) {
            float yl[16];
            float yh[16];
            float4 sumy = {0.f, 0.f, 0.f, 0.f};

            for (short i = 0; i < 8; ++i) {
                yl[i + 0] = y4[i +   0]; sumy[0] += yl[i + 0];
                yl[i + 8] = y4[i +  32]; sumy[1] += yl[i + 8];
                yh[i + 0] = y4[i + 128]; sumy[2] += yh[i + 0];
                yh[i + 8] = y4[i + 160]; sumy[3] += yh[i + 8];
            }

            device const uint16_t *sc = (device const uint16_t *)x[ib].scales + iq;
            device const uint16_t *q1 = (device const uint16_t *)x[ib].qs + 16 * iq + 4 * ir;
            device const half *dh = &x[ib].d;

            for (short row = 0; row < nr0 && row0 + (uint)row < args.out_dim; row++) {
                sc16[0] = sc[0] & kmask1;
                sc16[1] = sc[2] & kmask1;
                sc16[2] = ((sc[4] >> 0) & kmask2) | ((sc[0] & kmask3) >> 2);
                sc16[3] = ((sc[4] >> 4) & kmask2) | ((sc[2] & kmask3) >> 2);

                device const uint16_t *q2 = q1 + 32;
                float4 acc1 = {0.f, 0.f, 0.f, 0.f};
                float4 acc2 = {0.f, 0.f, 0.f, 0.f};

                FOR_UNROLL (short i = 0; i < 4; ++i) {
                    acc1[0] += yl[2 * i + 0] * (q1[i] & 0x000F);
                    acc1[1] += yl[2 * i + 1] * (q1[i] & 0x0F00);
                    acc1[2] += yl[2 * i + 8] * (q1[i] & 0x00F0);
                    acc1[3] += yl[2 * i + 9] * (q1[i] & 0xF000);
                    acc2[0] += yh[2 * i + 0] * (q2[i] & 0x000F);
                    acc2[1] += yh[2 * i + 1] * (q2[i] & 0x0F00);
                    acc2[2] += yh[2 * i + 8] * (q2[i] & 0x00F0);
                    acc2[3] += yh[2 * i + 9] * (q2[i] & 0xF000);
                }

                sumf[row] += dh[0] * ((acc1[0] + 1.f / 256.f * acc1[1]) * sc8[0] +
                                      (acc1[2] + 1.f / 256.f * acc1[3]) * sc8[1] * 1.f / 16.f +
                                      (acc2[0] + 1.f / 256.f * acc2[1]) * sc8[4] +
                                      (acc2[2] + 1.f / 256.f * acc2[3]) * sc8[5] * 1.f / 16.f) -
                             dh[1] * (sumy[0] * sc8[2] + sumy[1] * sc8[3] +
                                      sumy[2] * sc8[6] + sumy[3] * sc8[7]);

                q1 += args.down_row_bytes / 2;
                sc += args.down_row_bytes / 2;
                dh += args.down_row_bytes / 2;
            }

            y4 += 4 * QK_K;
        }
    }

    for (short row = 0; row < nr0 && row0 + (uint)row < args.out_dim; row++) {
        const float sum_all = simd_sum(sumf[row]);
        if (tiisg == 0u) {
            out[(uint64_t)token * args.out_dim + row0 + (uint)row] = sum_all;
        }
    }
}

kernel void kernel_glm_q4_K_addr_down_simd_f32(
        constant ds4_metal_glm_routed_moe_args &args,
        device const uint64_t *down_addrs,
        device const int32_t *selected,
        device const float *mid,
        device float *out,
        uint3 tgpig [[threadgroup_position_in_grid]],
        ushort tiisg [[thread_index_in_simdgroup]],
        ushort sgitg [[simdgroup_index_in_threadgroup]]) {
    const short NSG = 2;
    const short nr0 = N_R0_Q4_K;
    const int nb = args.mid_dim / QK_K;
    const uint row0 = ((uint)tgpig.x * (uint)NSG + (uint)sgitg) * (uint)nr0;
    const uint token = tgpig.y;
    if (row0 >= args.out_dim || token >= args.n_tokens) return;

    constexpr uint16_t kmask1 = 0x3f3f;
    constexpr uint16_t kmask2 = 0x0f0f;
    constexpr uint16_t kmask3 = 0xc0c0;

    const short ix = tiisg / 8;
    const short it = tiisg % 8;
    const short iq = it / 4;
    const short ir = it % 4;

    float sumf[N_R0_Q4_K] = {0.f};
    uint16_t sc16[4];
    thread const uint8_t *sc8 = (thread const uint8_t *)sc16;

    const uint64_t selected_base = (uint64_t)token * args.n_expert_used;
    const uint64_t mid_base = (uint64_t)token * args.mid_token_stride;
    for (uint slot = 0; slot < args.n_expert_used; slot++) {
        const int expert = selected[selected_base + slot];
        if (expert < 0 || (uint)expert >= args.n_total_expert) continue;
        const uint64_t down_addr = down_addrs[(uint)expert];
        if (down_addr == 0) continue;

        device const block_q4_K *x =
            (device const block_q4_K *)(reinterpret_cast<device const char *>(down_addr) +
                (uint64_t)row0 * args.down_row_bytes);
        device const float *y = mid + mid_base + (uint64_t)slot * args.mid_dim;
        device const float *y4 = y + ix * QK_K + 64 * iq + 8 * ir;

        for (int ib = ix; ib < nb; ib += 4) {
            float yl[16];
            float yh[16];
            float4 sumy = {0.f, 0.f, 0.f, 0.f};

            for (short i = 0; i < 8; ++i) {
                yl[i + 0] = y4[i +   0]; sumy[0] += yl[i + 0];
                yl[i + 8] = y4[i +  32]; sumy[1] += yl[i + 8];
                yh[i + 0] = y4[i + 128]; sumy[2] += yh[i + 0];
                yh[i + 8] = y4[i + 160]; sumy[3] += yh[i + 8];
            }

            device const uint16_t *sc = (device const uint16_t *)x[ib].scales + iq;
            device const uint16_t *q1 = (device const uint16_t *)x[ib].qs + 16 * iq + 4 * ir;
            device const half *dh = &x[ib].d;

            for (short row = 0; row < nr0 && row0 + (uint)row < args.out_dim; row++) {
                sc16[0] = sc[0] & kmask1;
                sc16[1] = sc[2] & kmask1;
                sc16[2] = ((sc[4] >> 0) & kmask2) | ((sc[0] & kmask3) >> 2);
                sc16[3] = ((sc[4] >> 4) & kmask2) | ((sc[2] & kmask3) >> 2);

                device const uint16_t *q2 = q1 + 32;
                float4 acc1 = {0.f, 0.f, 0.f, 0.f};
                float4 acc2 = {0.f, 0.f, 0.f, 0.f};

                FOR_UNROLL (short i = 0; i < 4; ++i) {
                    acc1[0] += yl[2 * i + 0] * (q1[i] & 0x000F);
                    acc1[1] += yl[2 * i + 1] * (q1[i] & 0x0F00);
                    acc1[2] += yl[2 * i + 8] * (q1[i] & 0x00F0);
                    acc1[3] += yl[2 * i + 9] * (q1[i] & 0xF000);
                    acc2[0] += yh[2 * i + 0] * (q2[i] & 0x000F);
                    acc2[1] += yh[2 * i + 1] * (q2[i] & 0x0F00);
                    acc2[2] += yh[2 * i + 8] * (q2[i] & 0x00F0);
                    acc2[3] += yh[2 * i + 9] * (q2[i] & 0xF000);
                }

                sumf[row] += dh[0] * ((acc1[0] + 1.f / 256.f * acc1[1]) * sc8[0] +
                                      (acc1[2] + 1.f / 256.f * acc1[3]) * sc8[1] * 1.f / 16.f +
                                      (acc2[0] + 1.f / 256.f * acc2[1]) * sc8[4] +
                                      (acc2[2] + 1.f / 256.f * acc2[3]) * sc8[5] * 1.f / 16.f) -
                             dh[1] * (sumy[0] * sc8[2] + sumy[1] * sc8[3] +
                                      sumy[2] * sc8[6] + sumy[3] * sc8[7]);

                q1 += args.down_row_bytes / 2;
                sc += args.down_row_bytes / 2;
                dh += args.down_row_bytes / 2;
            }

            y4 += 4 * QK_K;
        }
    }

    for (short row = 0; row < nr0 && row0 + (uint)row < args.out_dim; row++) {
        const float sum_all = simd_sum(sumf[row]);
        if (tiisg == 0u) {
            out[(uint64_t)token * args.out_dim + row0 + (uint)row] = sum_all;
        }
    }
}

kernel void kernel_glm_q6_K_down_f32(
        constant ds4_metal_glm_routed_moe_args &args,
        device const char *down,
        device const int32_t *selected,
        device const float *mid,
        device float *out,
        uint3 tgpig [[threadgroup_position_in_grid]],
        ushort tiisg [[thread_index_in_simdgroup]],
        ushort sgitg [[simdgroup_index_in_threadgroup]]) {
    const short NSG = 2;
    constexpr uint kmask1 = 0x03u;
    constexpr uint kmask2 = 0x0Cu;
    constexpr uint kmask3 = 0x30u;
    constexpr uint kmask4 = 0xC0u;

    const uint row0 = ((uint)tgpig.x * (uint)NSG + (uint)sgitg) * N_R0_Q6_K;
    const uint token = tgpig.y;
    if (row0 >= args.out_dim || token >= args.n_tokens) return;

    const int nb = args.mid_dim / QK_K;
    float sumf[N_R0_Q6_K] = {0.f};
    float yl[16];
    const short tid = tiisg / 2;
    const short ix = tiisg % 2;
    const short ip = tid / 8;
    const short il = tid % 8;
    const short l0 = 4 * il;
    const short is = 8 * ip + l0 / 16;
    const short y_offset = 128 * ip + l0;
    const short q_offset_l = 64 * ip + l0;
    const short q_offset_h = 32 * ip + l0;

    const uint64_t selected_base = (uint64_t)token * args.n_expert_used;
    const uint64_t mid_base = (uint64_t)token * args.mid_token_stride;
    for (uint slot = 0; slot < args.n_expert_used; slot++) {
        const int expert = selected[selected_base + slot];
        if (expert < 0 || (uint)expert >= args.n_total_expert) continue;
        device const block_q6_K *x =
            (device const block_q6_K *)(down +
                (uint64_t)(uint)expert * args.down_expert_bytes +
                (uint64_t)row0 * args.down_row_bytes);
        device const float *yy = mid + mid_base + (uint64_t)slot * args.mid_dim;

        for (int ib = ix; ib < nb; ib += 2) {
            device const uchar *q1 = x[ib].ql + q_offset_l;
            device const uchar *q2 = q1 + 32;
            device const uchar *qh = x[ib].qh + q_offset_h;
            device const char *sc = x[ib].scales + is;
            device const half *dh = &x[ib].d;
            device const float *y = yy + (uint64_t)ib * QK_K + y_offset;

            for (short l = 0; l < 4; ++l) {
                yl[4 * l + 0] = y[l + 0];
                yl[4 * l + 1] = y[l + 32];
                yl[4 * l + 2] = y[l + 64];
                yl[4 * l + 3] = y[l + 96];
            }

            for (short row = 0; row < N_R0_Q6_K && row0 + (uint)row < args.out_dim; row++) {
                float4 sums = {0.f, 0.f, 0.f, 0.f};

                FOR_UNROLL (short l = 0; l < 4; ++l) {
                    const uint h = (uint)qh[l];
                    sums[0] += yl[4 * l + 0] *
                        (float)((int)((q1[l] & 0x0Fu) | ((h & kmask1) << 4u)) - 32);
                    sums[1] += yl[4 * l + 1] *
                        (float)((int)((q2[l] & 0x0Fu) | ((h & kmask2) << 2u)) - 32);
                    sums[2] += yl[4 * l + 2] *
                        (float)((int)((q1[l] >> 4u) | (h & kmask3)) - 32);
                    sums[3] += yl[4 * l + 3] *
                        (float)((int)((q2[l] >> 4u) | ((h & kmask4) >> 2u)) - 32);
                }

                sumf[row] += dh[0] * (sums[0] * sc[0] + sums[1] * sc[2] +
                                      sums[2] * sc[4] + sums[3] * sc[6]);

                q1 += args.down_row_bytes;
                q2 += args.down_row_bytes;
                qh += args.down_row_bytes;
                sc += args.down_row_bytes;
                dh += args.down_row_bytes / 2;
            }
        }
    }

    for (short row = 0; row < N_R0_Q6_K && row0 + (uint)row < args.out_dim; row++) {
        const float sum_all = simd_sum(sumf[row]);
        if (tiisg == 0u) {
            out[(uint64_t)token * args.out_dim + row0 + (uint)row] = sum_all;
        }
    }
}

kernel void kernel_glm_q5_K_slots6_down_f32(
        constant ds4_metal_glm_routed_moe_args &args,
        device const char *down0,
        device const char *down1,
        device const char *down2,
        device const char *down3,
        device const char *down4,
        device const char *down5,
        device const float *mid,
        device float *out,
        uint3 tgpig [[threadgroup_position_in_grid]],
        ushort tiisg [[thread_index_in_simdgroup]],
        ushort sgitg [[simdgroup_index_in_threadgroup]]) {
    const short NSG = 2;
    constexpr uint16_t kmask1 = 0x3f3f;
    constexpr uint16_t kmask2 = 0x0f0f;
    constexpr uint16_t kmask3 = 0xc0c0;

    const uint row0 = ((uint)tgpig.x * (uint)NSG + (uint)sgitg) * N_R0_Q5_K;
    const uint token = tgpig.y;
    if (row0 >= args.out_dim || token >= args.n_tokens) return;

    const short ix = tiisg / 8;
    const short it = tiisg % 8;
    const short iq = it / 4;
    const short ir = it % 4;
    const uint bit0 = 2u * (uint)iq;
    const uint bit1 = bit0 + 1u;
    const uint bit2 = bit0 + 4u;
    const uint bit3 = bit0 + 5u;
    const int nb = args.mid_dim / QK_K;

    float sumf[N_R0_Q5_K] = {0.f};
    uint16_t sc16[4];
    thread const uint8_t *sc8 = (thread const uint8_t *)sc16;
    const uint64_t mid_base = (uint64_t)token * args.mid_token_stride;
    for (uint slot = 0; slot < args.n_expert_used; slot++) {
        device const char *down_cur = down0;
        switch (slot) {
        case 1: down_cur = down1; break;
        case 2: down_cur = down2; break;
        case 3: down_cur = down3; break;
        case 4: down_cur = down4; break;
        case 5: down_cur = down5; break;
        default: break;
        }

        device const block_q5_K *x =
            (device const block_q5_K *)(down_cur +
                (uint64_t)row0 * args.down_row_bytes);
        device const float *y = mid + mid_base + (uint64_t)slot * args.mid_dim;
        device const float *y4 = y + ix * QK_K + 64 * iq + 8 * ir;

        for (int ib = ix; ib < nb; ib += 4) {
            float yl[16];
            float yh[16];
            float4 sumy = {0.f, 0.f, 0.f, 0.f};

            for (short i = 0; i < 8; ++i) {
                yl[i + 0] = y4[i +   0]; sumy[0] += yl[i + 0];
                yl[i + 8] = y4[i +  32]; sumy[1] += yl[i + 8];
                yh[i + 0] = y4[i + 128]; sumy[2] += yh[i + 0];
                yh[i + 8] = y4[i + 160]; sumy[3] += yh[i + 8];
            }

            device const uint16_t *sc = (device const uint16_t *)x[ib].scales + iq;
            device const uint16_t *q1 = (device const uint16_t *)x[ib].qs + 16 * iq + 4 * ir;
            device const uint16_t *qh = (device const uint16_t *)x[ib].qh + 4 * ir;
            device const half *dh = &x[ib].d;

            for (short row = 0; row < N_R0_Q5_K && row0 + (uint)row < args.out_dim; row++) {
                sc16[0] = sc[0] & kmask1;
                sc16[1] = sc[2] & kmask1;
                sc16[2] = ((sc[4] >> 0) & kmask2) | ((sc[0] & kmask3) >> 2);
                sc16[3] = ((sc[4] >> 4) & kmask2) | ((sc[2] & kmask3) >> 2);

                device const uint16_t *q2 = q1 + 32;
                float4 acc = {0.f, 0.f, 0.f, 0.f};

                FOR_UNROLL (short i = 0; i < 4; ++i) {
                    const uint ql1 = (uint)q1[i];
                    const uint ql2 = (uint)q2[i];
                    const uint hb = (uint)qh[i];
                    acc[0] += yl[2 * i + 0] *
                                  (float)((ql1 & 0x000Fu) + (((hb >> bit0) & 1u) << 4u)) +
                              yl[2 * i + 1] *
                                  (float)(((ql1 >> 8u) & 0x000Fu) + (((hb >> (bit0 + 8u)) & 1u) << 4u));
                    acc[1] += yl[2 * i + 8] *
                                  (float)(((ql1 >> 4u) & 0x000Fu) + (((hb >> bit1) & 1u) << 4u)) +
                              yl[2 * i + 9] *
                                  (float)(((ql1 >> 12u) & 0x000Fu) + (((hb >> (bit1 + 8u)) & 1u) << 4u));
                    acc[2] += yh[2 * i + 0] *
                                  (float)((ql2 & 0x000Fu) + (((hb >> bit2) & 1u) << 4u)) +
                              yh[2 * i + 1] *
                                  (float)(((ql2 >> 8u) & 0x000Fu) + (((hb >> (bit2 + 8u)) & 1u) << 4u));
                    acc[3] += yh[2 * i + 8] *
                                  (float)(((ql2 >> 4u) & 0x000Fu) + (((hb >> bit3) & 1u) << 4u)) +
                              yh[2 * i + 9] *
                                  (float)(((ql2 >> 12u) & 0x000Fu) + (((hb >> (bit3 + 8u)) & 1u) << 4u));
                }

                sumf[row] += dh[0] * (acc[0] * sc8[0] + acc[1] * sc8[1] +
                                      acc[2] * sc8[4] + acc[3] * sc8[5]) -
                             dh[1] * (sumy[0] * sc8[2] + sumy[1] * sc8[3] +
                                      sumy[2] * sc8[6] + sumy[3] * sc8[7]);

                q1 += args.down_row_bytes / 2;
                qh += args.down_row_bytes / 2;
                sc += args.down_row_bytes / 2;
                dh += args.down_row_bytes / 2;
            }

            y4 += 4 * QK_K;
        }
    }

    for (short row = 0; row < N_R0_Q5_K && row0 + (uint)row < args.out_dim; row++) {
        const float sum_all = simd_sum(sumf[row]);
        if (tiisg == 0u) {
            out[(uint64_t)token * args.out_dim + row0 + (uint)row] = sum_all;
        }
    }
}

kernel void kernel_glm_q5_K_slots8_down_f32(
        constant ds4_metal_glm_routed_moe_args &args,
        device const char *down0,
        device const char *down1,
        device const char *down2,
        device const char *down3,
        device const char *down4,
        device const char *down5,
        device const char *down6,
        device const char *down7,
        device const float *mid,
        device float *out,
        uint3 tgpig [[threadgroup_position_in_grid]],
        ushort tiisg [[thread_index_in_simdgroup]],
        ushort sgitg [[simdgroup_index_in_threadgroup]]) {
    const short NSG = 2;
    constexpr uint16_t kmask1 = 0x3f3f;
    constexpr uint16_t kmask2 = 0x0f0f;
    constexpr uint16_t kmask3 = 0xc0c0;

    const uint row0 = ((uint)tgpig.x * (uint)NSG + (uint)sgitg) * N_R0_Q5_K;
    const uint token = tgpig.y;
    if (row0 >= args.out_dim || token >= args.n_tokens) return;

    const short ix = tiisg / 8;
    const short it = tiisg % 8;
    const short iq = it / 4;
    const short ir = it % 4;
    const uint bit0 = 2u * (uint)iq;
    const uint bit1 = bit0 + 1u;
    const uint bit2 = bit0 + 4u;
    const uint bit3 = bit0 + 5u;
    const int nb = args.mid_dim / QK_K;

    float sumf[N_R0_Q5_K] = {0.f};
    uint16_t sc16[4];
    thread const uint8_t *sc8 = (thread const uint8_t *)sc16;
    const uint64_t mid_base = (uint64_t)token * args.mid_token_stride;
    for (uint slot = 0; slot < args.n_expert_used; slot++) {
        device const char *down_cur = down0;
        switch (slot) {
        case 1: down_cur = down1; break;
        case 2: down_cur = down2; break;
        case 3: down_cur = down3; break;
        case 4: down_cur = down4; break;
        case 5: down_cur = down5; break;
        case 6: down_cur = down6; break;
        case 7: down_cur = down7; break;
        default: break;
        }

        device const block_q5_K *x =
            (device const block_q5_K *)(down_cur +
                (uint64_t)row0 * args.down_row_bytes);
        device const float *y = mid + mid_base + (uint64_t)slot * args.mid_dim;
        device const float *y4 = y + ix * QK_K + 64 * iq + 8 * ir;

        for (int ib = ix; ib < nb; ib += 4) {
            float yl[16];
            float yh[16];
            float4 sumy = {0.f, 0.f, 0.f, 0.f};

            for (short i = 0; i < 8; ++i) {
                yl[i + 0] = y4[i +   0]; sumy[0] += yl[i + 0];
                yl[i + 8] = y4[i +  32]; sumy[1] += yl[i + 8];
                yh[i + 0] = y4[i + 128]; sumy[2] += yh[i + 0];
                yh[i + 8] = y4[i + 160]; sumy[3] += yh[i + 8];
            }

            device const uint16_t *sc = (device const uint16_t *)x[ib].scales + iq;
            device const uint16_t *q1 = (device const uint16_t *)x[ib].qs + 16 * iq + 4 * ir;
            device const uint16_t *qh = (device const uint16_t *)x[ib].qh + 4 * ir;
            device const half *dh = &x[ib].d;

            for (short row = 0; row < N_R0_Q5_K && row0 + (uint)row < args.out_dim; row++) {
                sc16[0] = sc[0] & kmask1;
                sc16[1] = sc[2] & kmask1;
                sc16[2] = ((sc[4] >> 0) & kmask2) | ((sc[0] & kmask3) >> 2);
                sc16[3] = ((sc[4] >> 4) & kmask2) | ((sc[2] & kmask3) >> 2);

                device const uint16_t *q2 = q1 + 32;
                float4 acc = {0.f, 0.f, 0.f, 0.f};

                FOR_UNROLL (short i = 0; i < 4; ++i) {
                    const uint ql1 = (uint)q1[i];
                    const uint ql2 = (uint)q2[i];
                    const uint hb = (uint)qh[i];
                    acc[0] += yl[2 * i + 0] *
                                  (float)((ql1 & 0x000Fu) + (((hb >> bit0) & 1u) << 4u)) +
                              yl[2 * i + 1] *
                                  (float)(((ql1 >> 8u) & 0x000Fu) + (((hb >> (bit0 + 8u)) & 1u) << 4u));
                    acc[1] += yl[2 * i + 8] *
                                  (float)(((ql1 >> 4u) & 0x000Fu) + (((hb >> bit1) & 1u) << 4u)) +
                              yl[2 * i + 9] *
                                  (float)(((ql1 >> 12u) & 0x000Fu) + (((hb >> (bit1 + 8u)) & 1u) << 4u));
                    acc[2] += yh[2 * i + 0] *
                                  (float)((ql2 & 0x000Fu) + (((hb >> bit2) & 1u) << 4u)) +
                              yh[2 * i + 1] *
                                  (float)(((ql2 >> 8u) & 0x000Fu) + (((hb >> (bit2 + 8u)) & 1u) << 4u));
                    acc[3] += yh[2 * i + 8] *
                                  (float)(((ql2 >> 4u) & 0x000Fu) + (((hb >> bit3) & 1u) << 4u)) +
                              yh[2 * i + 9] *
                                  (float)(((ql2 >> 12u) & 0x000Fu) + (((hb >> (bit3 + 8u)) & 1u) << 4u));
                }

                sumf[row] += dh[0] * (acc[0] * sc8[0] + acc[1] * sc8[1] +
                                      acc[2] * sc8[4] + acc[3] * sc8[5]) -
                             dh[1] * (sumy[0] * sc8[2] + sumy[1] * sc8[3] +
                                      sumy[2] * sc8[6] + sumy[3] * sc8[7]);

                q1 += args.down_row_bytes / 2;
                qh += args.down_row_bytes / 2;
                sc += args.down_row_bytes / 2;
                dh += args.down_row_bytes / 2;
            }

            y4 += 4 * QK_K;
        }
    }

    for (short row = 0; row < N_R0_Q5_K && row0 + (uint)row < args.out_dim; row++) {
        const float sum_all = simd_sum(sumf[row]);
        if (tiisg == 0u) {
            out[(uint64_t)token * args.out_dim + row0 + (uint)row] = sum_all;
        }
    }
}

kernel void kernel_glm_q6_K_slots6_down_f32(
        constant ds4_metal_glm_routed_moe_args &args,
        device const char *down0,
        device const char *down1,
        device const char *down2,
        device const char *down3,
        device const char *down4,
        device const char *down5,
        device const float *mid,
        device float *out,
        uint3 tgpig [[threadgroup_position_in_grid]],
        ushort tiisg [[thread_index_in_simdgroup]],
        ushort sgitg [[simdgroup_index_in_threadgroup]]) {
    const short NSG = 2;
    constexpr uint kmask1 = 0x03u;
    constexpr uint kmask2 = 0x0Cu;
    constexpr uint kmask3 = 0x30u;
    constexpr uint kmask4 = 0xC0u;

    const uint row0 = ((uint)tgpig.x * (uint)NSG + (uint)sgitg) * N_R0_Q6_K;
    const uint token = tgpig.y;
    if (row0 >= args.out_dim || token >= args.n_tokens) return;

    const int nb = args.mid_dim / QK_K;
    float sumf[N_R0_Q6_K] = {0.f};
    float yl[16];
    const short tid = tiisg / 2;
    const short ix = tiisg % 2;
    const short ip = tid / 8;
    const short il = tid % 8;
    const short l0 = 4 * il;
    const short is = 8 * ip + l0 / 16;
    const short y_offset = 128 * ip + l0;
    const short q_offset_l = 64 * ip + l0;
    const short q_offset_h = 32 * ip + l0;

    const uint64_t mid_base = (uint64_t)token * args.mid_token_stride;
    for (uint slot = 0; slot < args.n_expert_used; slot++) {
        device const char *down_cur = down0;
        switch (slot) {
        case 1: down_cur = down1; break;
        case 2: down_cur = down2; break;
        case 3: down_cur = down3; break;
        case 4: down_cur = down4; break;
        case 5: down_cur = down5; break;
        default: break;
        }

        device const block_q6_K *x =
            (device const block_q6_K *)(down_cur +
                (uint64_t)row0 * args.down_row_bytes);
        device const float *yy = mid + mid_base + (uint64_t)slot * args.mid_dim;

        for (int ib = ix; ib < nb; ib += 2) {
            device const uchar *q1 = x[ib].ql + q_offset_l;
            device const uchar *q2 = q1 + 32;
            device const uchar *qh = x[ib].qh + q_offset_h;
            device const char *sc = x[ib].scales + is;
            device const half *dh = &x[ib].d;
            device const float *y = yy + (uint64_t)ib * QK_K + y_offset;

            for (short l = 0; l < 4; ++l) {
                yl[4 * l + 0] = y[l + 0];
                yl[4 * l + 1] = y[l + 32];
                yl[4 * l + 2] = y[l + 64];
                yl[4 * l + 3] = y[l + 96];
            }

            for (short row = 0; row < N_R0_Q6_K && row0 + (uint)row < args.out_dim; row++) {
                float4 sums = {0.f, 0.f, 0.f, 0.f};

                FOR_UNROLL (short l = 0; l < 4; ++l) {
                    const uint h = (uint)qh[l];
                    sums[0] += yl[4 * l + 0] *
                        (float)((int)((q1[l] & 0x0Fu) | ((h & kmask1) << 4u)) - 32);
                    sums[1] += yl[4 * l + 1] *
                        (float)((int)((q2[l] & 0x0Fu) | ((h & kmask2) << 2u)) - 32);
                    sums[2] += yl[4 * l + 2] *
                        (float)((int)((q1[l] >> 4u) | (h & kmask3)) - 32);
                    sums[3] += yl[4 * l + 3] *
                        (float)((int)((q2[l] >> 4u) | ((h & kmask4) >> 2u)) - 32);
                }

                sumf[row] += dh[0] * (sums[0] * sc[0] + sums[1] * sc[2] +
                                      sums[2] * sc[4] + sums[3] * sc[6]);

                q1 += args.down_row_bytes;
                q2 += args.down_row_bytes;
                qh += args.down_row_bytes;
                sc += args.down_row_bytes;
                dh += args.down_row_bytes / 2;
            }
        }
    }

    for (short row = 0; row < N_R0_Q6_K && row0 + (uint)row < args.out_dim; row++) {
        const float sum_all = simd_sum(sumf[row]);
        if (tiisg == 0u) {
            out[(uint64_t)token * args.out_dim + row0 + (uint)row] = sum_all;
        }
    }
}

kernel void kernel_glm_q6_K_slots8_down_f32(
        constant ds4_metal_glm_routed_moe_args &args,
        device const char *down0,
        device const char *down1,
        device const char *down2,
        device const char *down3,
        device const char *down4,
        device const char *down5,
        device const char *down6,
        device const char *down7,
        device const float *mid,
        device float *out,
        uint3 tgpig [[threadgroup_position_in_grid]],
        ushort tiisg [[thread_index_in_simdgroup]],
        ushort sgitg [[simdgroup_index_in_threadgroup]]) {
    const short NSG = 2;
    constexpr uint kmask1 = 0x03u;
    constexpr uint kmask2 = 0x0Cu;
    constexpr uint kmask3 = 0x30u;
    constexpr uint kmask4 = 0xC0u;

    const uint row0 = ((uint)tgpig.x * (uint)NSG + (uint)sgitg) * N_R0_Q6_K;
    const uint token = tgpig.y;
    if (row0 >= args.out_dim || token >= args.n_tokens) return;

    const int nb = args.mid_dim / QK_K;
    float sumf[N_R0_Q6_K] = {0.f};
    float yl[16];
    const short tid = tiisg / 2;
    const short ix = tiisg % 2;
    const short ip = tid / 8;
    const short il = tid % 8;
    const short l0 = 4 * il;
    const short is = 8 * ip + l0 / 16;
    const short y_offset = 128 * ip + l0;
    const short q_offset_l = 64 * ip + l0;
    const short q_offset_h = 32 * ip + l0;

    const uint64_t mid_base = (uint64_t)token * args.mid_token_stride;
    for (uint slot = 0; slot < args.n_expert_used; slot++) {
        device const char *down_cur = down0;
        switch (slot) {
        case 1: down_cur = down1; break;
        case 2: down_cur = down2; break;
        case 3: down_cur = down3; break;
        case 4: down_cur = down4; break;
        case 5: down_cur = down5; break;
        case 6: down_cur = down6; break;
        case 7: down_cur = down7; break;
        default: break;
        }

        device const block_q6_K *x =
            (device const block_q6_K *)(down_cur +
                (uint64_t)row0 * args.down_row_bytes);
        device const float *yy = mid + mid_base + (uint64_t)slot * args.mid_dim;

        for (int ib = ix; ib < nb; ib += 2) {
            device const uchar *q1 = x[ib].ql + q_offset_l;
            device const uchar *q2 = q1 + 32;
            device const uchar *qh = x[ib].qh + q_offset_h;
            device const char *sc = x[ib].scales + is;
            device const half *dh = &x[ib].d;
            device const float *y = yy + (uint64_t)ib * QK_K + y_offset;

            for (short l = 0; l < 4; ++l) {
                yl[4 * l + 0] = y[l + 0];
                yl[4 * l + 1] = y[l + 32];
                yl[4 * l + 2] = y[l + 64];
                yl[4 * l + 3] = y[l + 96];
            }

            for (short row = 0; row < N_R0_Q6_K && row0 + (uint)row < args.out_dim; row++) {
                float4 sums = {0.f, 0.f, 0.f, 0.f};

                FOR_UNROLL (short l = 0; l < 4; ++l) {
                    const uint h = (uint)qh[l];
                    sums[0] += yl[4 * l + 0] *
                        (float)((int)((q1[l] & 0x0Fu) | ((h & kmask1) << 4u)) - 32);
                    sums[1] += yl[4 * l + 1] *
                        (float)((int)((q2[l] & 0x0Fu) | ((h & kmask2) << 2u)) - 32);
                    sums[2] += yl[4 * l + 2] *
                        (float)((int)((q1[l] >> 4u) | (h & kmask3)) - 32);
                    sums[3] += yl[4 * l + 3] *
                        (float)((int)((q2[l] >> 4u) | ((h & kmask4) >> 2u)) - 32);
                }

                sumf[row] += dh[0] * (sums[0] * sc[0] + sums[1] * sc[2] +
                                      sums[2] * sc[4] + sums[3] * sc[6]);

                q1 += args.down_row_bytes;
                q2 += args.down_row_bytes;
                qh += args.down_row_bytes;
                sc += args.down_row_bytes;
                dh += args.down_row_bytes / 2;
            }
        }
    }

    for (short row = 0; row < N_R0_Q6_K && row0 + (uint)row < args.out_dim; row++) {
        const float sum_all = simd_sum(sumf[row]);
        if (tiisg == 0u) {
            out[(uint64_t)token * args.out_dim + row0 + (uint)row] = sum_all;
        }
    }
}

template <typename type4x4>
void dequantize_q4_K(device const block_q4_K *xb, short il, thread type4x4 &reg) {
    device const uchar *q = xb->qs;

    short is = (il / 4) * 2;
    q = q + (il / 4) * 32 + 16 * (il & 1);
    il = il & 3;
    const uchar2 sc = get_scale_min_k4_just2(is, il / 2, xb->scales);
    const float d = il < 2 ?
        (float)xb->d :
        (float)xb->d * (1.0f / 16.0f);
    const float min = (float)xb->dmin;
    const float dl = d * sc[0];
    const float ml = min * sc[1];

    const ushort mask = il < 2 ? 0x0F : 0xF0;
    for (int i = 0; i < 16; ++i) {
        reg[i / 4][i % 4] = dl * (q[i] & mask) - ml;
    }
}

template <typename type4x4>
void dequantize_q5_K(device const block_q5_K *xb, short il, thread type4x4 &reg) {
    const short group = il / 2;
    const short l0 = (il & 1) * 16;
    const uchar2 sm = get_scale_min_k4_just2(group, 0, xb->scales);
    const float dl = (float)xb->d * (float)sm.x;
    const float ml = (float)xb->dmin * (float)sm.y;
    device const uchar *q = xb->qs + (group / 2) * 32 + l0;
    device const uchar *qh = xb->qh + l0;
    const uint shift = (uint)(group & 1) * 4u;
    const uchar hmask = (uchar)(1u << (uint)group);

    for (int i = 0; i < 16; ++i) {
        uint v = ((uint)q[i] >> shift) & 0x0Fu;
        v += (qh[i] & hmask) ? 16u : 0u;
        reg[i / 4][i % 4] = dl * (float)v - ml;
    }
}

template <typename type4x4>
void dequantize_q6_K(device const block_q6_K *xb, short il, thread type4x4 &reg) {
    const short n128 = il / 8;
    const short il128 = il & 7;
    const short quarter = il128 / 2;
    const short l0 = (il128 & 1) * 16;
    const uint ql_base = (uint)n128 * 64u;
    const uint qh_base = (uint)n128 * 32u;
    const uint sc_base = (uint)n128 * 8u + (uint)quarter * 2u + (uint)(il128 & 1);
    const float d = (float)xb->d * (float)((int)xb->scales[sc_base]);

    for (int i = 0; i < 16; ++i) {
        const uint l = (uint)l0 + (uint)i;
        uint v;
        if (quarter == 0) {
            v = ((uint)xb->ql[ql_base + l] & 0x0Fu) |
                ((((uint)xb->qh[qh_base + l] >> 0u) & 3u) << 4u);
        } else if (quarter == 1) {
            v = ((uint)xb->ql[ql_base + 32u + l] & 0x0Fu) |
                ((((uint)xb->qh[qh_base + l] >> 2u) & 3u) << 4u);
        } else if (quarter == 2) {
            v = ((uint)xb->ql[ql_base + l] >> 4u) |
                ((((uint)xb->qh[qh_base + l] >> 4u) & 3u) << 4u);
        } else {
            v = ((uint)xb->ql[ql_base + 32u + l] >> 4u) |
                ((((uint)xb->qh[qh_base + l] >> 6u) & 3u) << 4u);
        }
        reg[i / 4][i % 4] = d * (float)((int)v - 32);
    }
}

template <typename type4x4>
void dequantize_iq2_xxs(device const block_iq2_xxs * xb, short il, thread type4x4 & reg) {
    const float d = xb->d;
    const int ib32 = il/2;
    il = il%2;
    device const uint16_t * q2 = xb->qs + 4*ib32;
    const uint32_t aux32_g = q2[0] | (q2[1] << 16);
    const uint32_t aux32_s = q2[2] | (q2[3] << 16);
    thread const uint8_t * aux8 = (thread const uint8_t *)&aux32_g;
    const float dl = d * (0.5f + (aux32_s >> 28)) * 0.25f;
    constant uint8_t * grid = (constant uint8_t *)(iq2xxs_grid + aux8[2*il+0]);
    uint8_t signs = ksigns_iq2xs[(aux32_s >> 14*il) & 127];
    for (int i = 0; i < 8; ++i) {
        reg[i/4][i%4] = dl * grid[i] * (signs & kmask_iq2xs[i] ? -1.f : 1.f);
    }
    grid = (constant uint8_t *)(iq2xxs_grid + aux8[2*il+1]);
    signs = ksigns_iq2xs[(aux32_s >> (14*il+7)) & 127];
    for (int i = 0; i < 8; ++i) {
        reg[2+i/4][i%4] = dl * grid[i] * (signs & kmask_iq2xs[i] ? -1.f : 1.f);
    }
}

struct ds4_metal_args_mul_mv_id {
    int32_t  nei0;
    int32_t  nei1;
    uint64_t nbi1;
    int32_t  ne00;
    int32_t  ne01;
    int32_t  ne02;
    uint64_t nb00;
    uint64_t nb01;
    uint64_t nb02;
    int32_t  ne10;
    int32_t  ne11;
    int32_t  ne12;
    int32_t  ne13;
    uint64_t nb10;
    uint64_t nb11;
    uint64_t nb12;
    int32_t  ne0;
    int32_t  ne1;
    uint64_t nb1;
    int32_t  nr0;
    /* Tensor-parallel expert ownership: when tp_world > 1, routed-MoE
     * kernels skip experts outside this rank's contiguous id range so this
     * rank never touches their weights. Zero-initialized (tp_world 0) means
     * no split; only the routed id kernels read these. */
    int32_t  tp_rank;
    int32_t  tp_world;
    /* When set, the sum6 kernels add this per-row vector (the shared-expert
     * partial) into their accumulation, folding the TP local-add. */
    int32_t  tp_addend;
    /* First expert id present at the bound blob base (TP mappings bind only
     * the owned range; kernels rebase before pointer math). */
    int32_t  tp_expert_base;
};

struct ds4_metal_moe_expert_group_args {
    uint32_t expert_base;
    uint32_t expert_count;
    uint32_t accumulate;
    uint32_t pad0;
};

struct ds4_metal_q4_gather_slots6_args {
    uint64_t expert_bytes;
    uint32_t group_size;
    uint32_t n_slots;
};

struct ds4_metal_q4_expert_table {
    array<device const char *, 384> experts [[id(0)]];
};

struct ds4_metal_expert_address_table {
    device const uint64_t *addrs;
};

struct ds4_metal_stream_expert_validate_args {
    uint32_t n_total_expert;
    uint32_t n_expert;
};

struct ds4_metal_stream_expert_split_args {
    uint32_t active_mask;
    uint32_t accumulate;
};

struct ds4_metal_args_mul_mm_id_map0 {
    int32_t  ne02;
    int32_t  ne10;
    int32_t  ne11;
    uint64_t nb11;
    uint64_t nb12;
    int32_t  ne21;
    int32_t  ne20;
    uint64_t nb21;
};

struct ds4_metal_args_mul_mm_id {
    int32_t  ne00;
    int32_t  ne02;
    uint64_t nb01;
    uint64_t nb02;
    uint64_t nb03;
    int32_t  ne11;
    uint64_t nb10;
    uint64_t nb11;
    uint64_t nb12;
    uint64_t nb13;
    int32_t  ne20;
    int32_t  ne21;
    int32_t  ne0;
    int32_t  ne1;
    int16_t  r2;
    int16_t  r3;
    /* Tensor-parallel expert ownership; tp_world<=1 disables the split.
     * tp_expert_base rebases expert ids when the bind covers only the
     * owned contiguous range of the expert blob. */
    int32_t  tp_rank;
    int32_t  tp_world;
    int32_t  tp_expert_base;
};

template<int nr0, typename args_t>
void kernel_mul_mv_q2_K_f32_impl(
        args_t args,
        device const char * src0,
        device const char * src1,
        device       char * dst,
        threadgroup  char * shmem,
        uint3  tgpig,
        ushort tiisg,
        ushort sgitg) {
    const short NSG = FC_mul_mv_nsg;

    const int nb = args.ne00/QK_K;

    const int r0 = tgpig.x;
    const int r1 = tgpig.y;
    const int im = tgpig.z;

    const int first_row = (r0 * NSG + sgitg) * nr0;

    const uint i12 = im%args.ne12;
    const uint i13 = im/args.ne12;

    const uint64_t offset0 = first_row*args.nb01 + (i12/args.r2)*args.nb02 + (i13/args.r3)*args.nb03;
    const uint64_t offset1 =        r1*args.nb11 + (i12        )*args.nb12 + (i13        )*args.nb13;

    device const block_q2_K * x = (device const block_q2_K *) (src0 + offset0);
    device const float      * y = (device const float      *) (src1 + offset1);

    float yl[32];
    float sumf[nr0]={0.f};

    const short ix = tiisg/8;  // 0...3
    const short it = tiisg%8;  // 0...7
    const short iq = it/4;     // 0 or 1
    const short ir = it%4;     // 0...3
    const short is = (8*ir)/16;// 0 or 1

    device const float * y4 = y + ix * QK_K + 128 * iq + 8 * ir;

    for (int ib = ix; ib < nb; ib += 4) {
        float4 sumy = {0.f, 0.f, 0.f, 0.f};
        for (short i = 0; i < 8; ++i) {
            yl[i+ 0] = y4[i+ 0]; sumy[0] += yl[i+ 0];
            yl[i+ 8] = y4[i+32]; sumy[1] += yl[i+ 8];
            yl[i+16] = y4[i+64]; sumy[2] += yl[i+16];
            yl[i+24] = y4[i+96]; sumy[3] += yl[i+24];
        }

        device const uint8_t  * sc = (device const uint8_t  *)x[ib].scales + 8*iq + is;
        device const uint16_t * qs = (device const uint16_t *)x[ib].qs + 16 * iq + 4 * ir;
        device const half     * dh = &x[ib].d;

        for (short row = 0; row < nr0; row++) {
            float4 acc1 = {0.f, 0.f, 0.f, 0.f};
            float4 acc2 = {0.f, 0.f, 0.f, 0.f};
            for (int i = 0; i < 8; i += 2) {
                acc1[0] += yl[i+ 0] * (qs[i/2] & 0x0003);
                acc2[0] += yl[i+ 1] * (qs[i/2] & 0x0300);
                acc1[1] += yl[i+ 8] * (qs[i/2] & 0x000c);
                acc2[1] += yl[i+ 9] * (qs[i/2] & 0x0c00);
                acc1[2] += yl[i+16] * (qs[i/2] & 0x0030);
                acc2[2] += yl[i+17] * (qs[i/2] & 0x3000);
                acc1[3] += yl[i+24] * (qs[i/2] & 0x00c0);
                acc2[3] += yl[i+25] * (qs[i/2] & 0xc000);
            }
            float dall = dh[0];
            float dmin = dh[1] * 1.f/16.f;
            sumf[row] += dall * ((acc1[0] + 1.f/256.f * acc2[0]) * (sc[0] & 0xF) * 1.f/ 1.f +
                                 (acc1[1] + 1.f/256.f * acc2[1]) * (sc[2] & 0xF) * 1.f/ 4.f +
                                 (acc1[2] + 1.f/256.f * acc2[2]) * (sc[4] & 0xF) * 1.f/16.f +
                                 (acc1[3] + 1.f/256.f * acc2[3]) * (sc[6] & 0xF) * 1.f/64.f) -
                         dmin * (sumy[0] * (sc[0] & 0xF0) + sumy[1] * (sc[2] & 0xF0) + sumy[2] * (sc[4] & 0xF0) + sumy[3] * (sc[6] & 0xF0));

            qs += args.nb01/2;
            sc += args.nb01;
            dh += args.nb01/2;
        }

        y4 += 4 * QK_K;
    }

    device float * dst_f32 = (device float *) dst + (uint64_t)im*args.ne0*args.ne1 + (uint64_t)r1*args.ne0;

    for (int row = 0; row < nr0 && first_row + row < args.ne0; ++row) {
        float sum_all = simd_sum(sumf[row]);
        if (tiisg == 0) {
            dst_f32[first_row + row] = sum_all;
        }
    }
}

template<int nr0, typename args_t>
void kernel_mul_mv_q4_K_f32_impl(
        args_t args,
        device const char *src0,
        device const char *src1,
        device       char *dst,
        threadgroup  char *shmem,
        uint3  tgpig,
        ushort tiisg,
        ushort sgitg) {
    const short NSG = FC_mul_mv_nsg;

    constexpr uint16_t kmask1 = 0x3f3f;
    constexpr uint16_t kmask2 = 0x0f0f;
    constexpr uint16_t kmask3 = 0xc0c0;

    const short ix = tiisg / 8;
    const short it = tiisg % 8;
    const short iq = it / 4;
    const short ir = it % 4;

    const int nb = args.ne00 / QK_K;

    const int r0 = tgpig.x;
    const int r1 = tgpig.y;
    const int im = tgpig.z;

    const int first_row = (r0 * NSG + sgitg) * nr0;

    const uint i12 = im % args.ne12;
    const uint i13 = im / args.ne12;

    const uint64_t offset0 = first_row * args.nb01 + (i12 / args.r2) * args.nb02 + (i13 / args.r3) * args.nb03;
    const uint64_t offset1 = r1 * args.nb11 + i12 * args.nb12 + i13 * args.nb13;

    device const block_q4_K *x = (device const block_q4_K *)(src0 + offset0);
    device const float *y = (device const float *)(src1 + offset1);

    float yl[16];
    float yh[16];
    float sumf[nr0] = {0.f};

    device const float *y4 = y + ix * QK_K + 64 * iq + 8 * ir;

    uint16_t sc16[4];
    thread const uint8_t *sc8 = (thread const uint8_t *)sc16;

    for (int ib = ix; ib < nb; ib += 4) {
        float4 sumy = {0.f, 0.f, 0.f, 0.f};

        for (short i = 0; i < 8; ++i) {
            yl[i + 0] = y4[i +   0]; sumy[0] += yl[i + 0];
            yl[i + 8] = y4[i +  32]; sumy[1] += yl[i + 8];
            yh[i + 0] = y4[i + 128]; sumy[2] += yh[i + 0];
            yh[i + 8] = y4[i + 160]; sumy[3] += yh[i + 8];
        }

        device const uint16_t *sc = (device const uint16_t *)x[ib].scales + iq;
        device const uint16_t *q1 = (device const uint16_t *)x[ib].qs + 16 * iq + 4 * ir;
        device const half *dh = &x[ib].d;

        for (short row = 0; row < nr0; row++) {
            sc16[0] = sc[0] & kmask1;
            sc16[1] = sc[2] & kmask1;
            sc16[2] = ((sc[4] >> 0) & kmask2) | ((sc[0] & kmask3) >> 2);
            sc16[3] = ((sc[4] >> 4) & kmask2) | ((sc[2] & kmask3) >> 2);

            device const uint16_t *q2 = q1 + 32;

            float4 acc1 = {0.f, 0.f, 0.f, 0.f};
            float4 acc2 = {0.f, 0.f, 0.f, 0.f};

            FOR_UNROLL (short i = 0; i < 4; ++i) {
                acc1[0] += yl[2 * i + 0] * (q1[i] & 0x000F);
                acc1[1] += yl[2 * i + 1] * (q1[i] & 0x0F00);
                acc1[2] += yl[2 * i + 8] * (q1[i] & 0x00F0);
                acc1[3] += yl[2 * i + 9] * (q1[i] & 0xF000);
                acc2[0] += yh[2 * i + 0] * (q2[i] & 0x000F);
                acc2[1] += yh[2 * i + 1] * (q2[i] & 0x0F00);
                acc2[2] += yh[2 * i + 8] * (q2[i] & 0x00F0);
                acc2[3] += yh[2 * i + 9] * (q2[i] & 0xF000);
            }

            sumf[row] += dh[0] * ((acc1[0] + 1.f / 256.f * acc1[1]) * sc8[0] +
                                  (acc1[2] + 1.f / 256.f * acc1[3]) * sc8[1] * 1.f / 16.f +
                                  (acc2[0] + 1.f / 256.f * acc2[1]) * sc8[4] +
                                  (acc2[2] + 1.f / 256.f * acc2[3]) * sc8[5] * 1.f / 16.f) -
                         dh[1] * (sumy[0] * sc8[2] + sumy[1] * sc8[3] + sumy[2] * sc8[6] + sumy[3] * sc8[7]);

            q1 += args.nb01 / 2;
            sc += args.nb01 / 2;
            dh += args.nb01 / 2;
        }

        y4 += 4 * QK_K;
    }

    device float *dst_f32 = (device float *)dst + (uint64_t)im * args.ne0 * args.ne1 + (uint64_t)r1 * args.ne0;

    for (int row = 0; row < nr0 && first_row + row < args.ne0; ++row) {
        float sum_all = simd_sum(sumf[row]);
        if (tiisg == 0) {
            dst_f32[first_row + row] = sum_all;
        }
    }

    (void)shmem;
}

template<int nr0, typename args_t>
void kernel_mul_mv_q8_K_f32_impl(
        args_t args,
        device const char *src0,
        device const char *src1,
        device       char *dst,
        threadgroup  char *shmem,
        uint3  tgpig,
        ushort tiisg,
        ushort sgitg) {
    const short NSG = FC_mul_mv_nsg;

    const int nb = args.ne00 / QK_K;

    const int r0 = tgpig.x;
    const int r1 = tgpig.y;
    const int im = tgpig.z;

    const int first_row = (r0 * NSG + sgitg) * nr0;

    const uint i12 = im % args.ne12;
    const uint i13 = im / args.ne12;

    const uint64_t offset0 = first_row * args.nb01 + (i12 / args.r2) * args.nb02 + (i13 / args.r3) * args.nb03;
    const uint64_t offset1 = r1 * args.nb11 + i12 * args.nb12 + i13 * args.nb13;

    device const block_q8_K *x = (device const block_q8_K *)(src0 + offset0);
    device const float *y = (device const float *)(src1 + offset1);

    float sumf[nr0] = {0.f};

    const short ix = tiisg / 8;
    const short il = tiisg % 8;
    const int ib0 = ix;

    device const float *yb = y + ib0 * QK_K + il * 32;

    for (int ib = ib0; ib < nb; ib += 4) {
        float yl[32];
        for (short i = 0; i < 32; ++i) {
            yl[i] = yb[i];
        }

        for (short row = 0; row < nr0; row++) {
            device const block_q8_K *xr =
                (device const block_q8_K *)((device const char *)x + (uint64_t)row * args.nb01);
            device const int8_t *qs = xr[ib].qs + il * 32;

            float sumq = 0.f;
            for (short i = 0; i < 32; ++i) {
                sumq += (float)qs[i] * yl[i];
            }

            sumf[row] += sumq * xr[ib].d;
        }

        yb += 4 * QK_K;
    }

    device float *dst_f32 =
        (device float *)dst + (uint64_t)im * args.ne0 * args.ne1 + (uint64_t)r1 * args.ne0;

    for (int row = 0; row < nr0 && first_row + row < args.ne0; ++row) {
        float sum_all = simd_sum(sumf[row]);
        if (tiisg == 0) {
            dst_f32[first_row + row] = sum_all;
        }
    }

    (void)shmem;
}

template<int nr0, typename args_t>
void kernel_mul_mv_iq2_xxs_f32_impl(
        args_t args,
        device const char * src0,
        device const char * src1,
        device       char * dst,
        threadgroup  char * shmem,
        uint3  tgpig,
        ushort tiisg,
        ushort sgitg) {
    const short NSG = FC_mul_mv_nsg;

    const int nb = args.ne00/QK_K;

    const int r0 = tgpig.x;
    const int r1 = tgpig.y;
    const int im = tgpig.z;

    const int first_row = (r0 * NSG + sgitg) * nr0;

    const uint i12 = im%args.ne12;
    const uint i13 = im/args.ne12;

    const uint64_t offset0 = first_row*args.nb01 + (i12/args.r2)*args.nb02 + (i13/args.r3)*args.nb03;
    const uint64_t offset1 =        r1*args.nb11 + (i12        )*args.nb12 + (i13        )*args.nb13;

    device const block_iq2_xxs * x = (device const block_iq2_xxs *) (src0 + offset0);
    device const float         * y = (device const float         *) (src1 + offset1);

    float yl[32];
    float sumf[nr0]={0.f};

    const int nb32 = nb * (QK_K / 32);

    threadgroup uint64_t * svalues = (threadgroup uint64_t *)(shmem);
    threadgroup uint8_t  * ssigns  = (threadgroup uint8_t  *)(svalues + 256);
    {
        int nval = 4;
        int pos  = (32*sgitg + tiisg)*nval;
        for (int i = 0; i < nval; ++i) svalues[pos + i] = ds4_metal_iq2xxs_grid[pos + i];
        nval = 2;
        pos  = (32*sgitg + tiisg)*nval;
        for (int i = 0; i < nval; ++i) ssigns[pos+i] = ds4_metal_ksigns_iq2xs[pos+i];
        threadgroup_barrier(mem_flags::mem_threadgroup);
    }

    const int ix = tiisg;

    device const float * y4 = y + 32 * ix;

    for (int ib32 = ix; ib32 < nb32; ib32 += 32) {
        for (short i = 0; i < 32; ++i) {
            yl[i] = y4[i];
        }

        const int ibl = ib32 / (QK_K / 32);
        const int ib  = ib32 % (QK_K / 32);

        device const block_iq2_xxs * xr = x + ibl;
        device const uint16_t * q2 = xr->qs + 4 * ib;
        device const half * dh = &xr->d;

        for (short row = 0; row < nr0; row++) {
            const float db = dh[0];
            device const uint8_t * aux8 = (device const uint8_t *)q2;
            const uint32_t aux32 = q2[2] | (q2[3] << 16);
            const float d = db * (0.5f + (aux32 >> 28));

            float sum = 0;
            for (short l = 0; l < 4; ++l) {
                const threadgroup uint8_t * grid = (const threadgroup uint8_t *)(svalues + aux8[l]);
                const uint8_t signs = ssigns[(aux32 >> 7*l) & 127];
                for (short j = 0; j < 8; ++j) {
                    sum += yl[8*l + j] * grid[j] * (signs & ds4_metal_kmask_iq2xs[j] ? -1.f : 1.f);
                }
            }
            sumf[row] += d * sum;

            dh += args.nb01/2;
            q2 += args.nb01/2;
        }

        y4 += 32 * 32;
    }

    device float * dst_f32 = (device float *) dst + (uint64_t)im*args.ne0*args.ne1 + (uint64_t)r1*args.ne0;

    for (int row = 0; row < nr0 && first_row + row < args.ne0; ++row) {
        float sum_all = simd_sum(sumf[row]);
        if (tiisg == 0) {
            dst_f32[first_row + row] = sum_all * 0.25f;
        }
    }
}

template<int nr0>
void kernel_mul_mv_iq2_xxs_pair_f32_impl(
        ds4_metal_args_mul_mv args,
        device const char * src0_gate,
        device const char * src0_up,
        device const char * src1,
        device       char * dst_gate,
        device       char * dst_up,
        threadgroup  char * shmem,
        uint3  tgpig,
        ushort tiisg,
        ushort sgitg) {
    const short NSG = FC_mul_mv_nsg;

    const int nb = args.ne00/QK_K;

    const int r0 = tgpig.x;
    const int r1 = tgpig.y;
    const int im = tgpig.z;

    const int first_row = (r0 * NSG + sgitg) * nr0;

    const uint i12 = im%args.ne12;
    const uint i13 = im/args.ne12;

    const uint64_t offset0 = first_row*args.nb01 + (i12/args.r2)*args.nb02 + (i13/args.r3)*args.nb03;
    const uint64_t offset1 =        r1*args.nb11 + (i12        )*args.nb12 + (i13        )*args.nb13;

    device const block_iq2_xxs * xg = (device const block_iq2_xxs *) (src0_gate + offset0);
    device const block_iq2_xxs * xu = (device const block_iq2_xxs *) (src0_up   + offset0);
    device const float         * y  = (device const float         *) (src1      + offset1);

    float yl[32];
    float sumg[nr0]={0.f};
    float sumu[nr0]={0.f};

    const int nb32 = nb * (QK_K / 32);

    threadgroup uint64_t * svalues = (threadgroup uint64_t *)(shmem);
    threadgroup uint8_t  * ssigns  = (threadgroup uint8_t  *)(svalues + 256);
    {
        int nval = 4;
        int pos  = (32*sgitg + tiisg)*nval;
        for (int i = 0; i < nval; ++i) svalues[pos + i] = ds4_metal_iq2xxs_grid[pos + i];
        nval = 2;
        pos  = (32*sgitg + tiisg)*nval;
        for (int i = 0; i < nval; ++i) ssigns[pos+i] = ds4_metal_ksigns_iq2xs[pos+i];
        threadgroup_barrier(mem_flags::mem_threadgroup);
    }

    const int ix = tiisg;
    device const float * y4 = y + 32 * ix;

    for (int ib32 = ix; ib32 < nb32; ib32 += 32) {
        for (short i = 0; i < 32; ++i) {
            yl[i] = y4[i];
        }

        const int ibl = ib32 / (QK_K / 32);
        const int ib  = ib32 % (QK_K / 32);

        device const block_iq2_xxs * xgr = xg + ibl;
        device const block_iq2_xxs * xur = xu + ibl;
        device const uint16_t * qg = xgr->qs + 4 * ib;
        device const uint16_t * qu = xur->qs + 4 * ib;
        device const half * dhg = &xgr->d;
        device const half * dhu = &xur->d;

        for (short row = 0; row < nr0; row++) {
            device const uint8_t * aux8g = (device const uint8_t *)qg;
            device const uint8_t * aux8u = (device const uint8_t *)qu;
            const uint32_t aux32g = qg[2] | (qg[3] << 16);
            const uint32_t aux32u = qu[2] | (qu[3] << 16);
            const float dg = (float)dhg[0] * (0.5f + (aux32g >> 28));
            const float du = (float)dhu[0] * (0.5f + (aux32u >> 28));

            float sg = 0;
            float su = 0;
            for (short l = 0; l < 4; ++l) {
                const threadgroup uint8_t * gridg = (const threadgroup uint8_t *)(svalues + aux8g[l]);
                const threadgroup uint8_t * gridu = (const threadgroup uint8_t *)(svalues + aux8u[l]);
                const uint8_t signg = ssigns[(aux32g >> 7*l) & 127];
                const uint8_t signu = ssigns[(aux32u >> 7*l) & 127];
                for (short j = 0; j < 8; ++j) {
                    const float v = yl[8*l + j];
                    sg += v * gridg[j] * (signg & ds4_metal_kmask_iq2xs[j] ? -1.f : 1.f);
                    su += v * gridu[j] * (signu & ds4_metal_kmask_iq2xs[j] ? -1.f : 1.f);
                }
            }
            sumg[row] += dg * sg;
            sumu[row] += du * su;

            dhg += args.nb01/2;
            dhu += args.nb01/2;
            qg  += args.nb01/2;
            qu  += args.nb01/2;
        }

        y4 += 32 * 32;
    }

    device float * dst_gate_f32 = (device float *) dst_gate + (uint64_t)im*args.ne0*args.ne1 + (uint64_t)r1*args.ne0;
    device float * dst_up_f32   = (device float *) dst_up   + (uint64_t)im*args.ne0*args.ne1 + (uint64_t)r1*args.ne0;

    for (int row = 0; row < nr0 && first_row + row < args.ne0; ++row) {
        const float sum_gate = simd_sum(sumg[row]);
        const float sum_up   = simd_sum(sumu[row]);
        if (tiisg == 0) {
            dst_gate_f32[first_row + row] = sum_gate * 0.25f;
            dst_up_f32[first_row + row]   = sum_up   * 0.25f;
        }
    }
}

typedef void (kernel_mul_mv2_disp_t)(
        ds4_metal_args_mul_mv args,
        device const char * src0,
        device const char * src1,
        device       char * dst,
        threadgroup  char * shmem,
        uint3  tgpig,
        ushort tiisg,
        ushort sgitg);

template<kernel_mul_mv2_disp_t disp_fn>
void mmv_fn(
        ds4_metal_args_mul_mv args,
        device const char * src0,
        device const char * src1,
        device       char * dst,
        threadgroup  char * shmem,
        uint3  tgpig,
        ushort tiitg,
        ushort tiisg,
        ushort sgitg) {
    disp_fn(args, src0, src1, dst, shmem, tgpig, tiisg, sgitg);
}

typedef decltype(mmv_fn<kernel_mul_mv_q2_K_f32_impl<N_R0_Q2_K>>) mul_mv_id_disp_fn_t;

// Decode-time expert matvec. The ids tensor selects the routed expert for each
// slot, then this wrapper invokes the quantized row kernel for Q8_0, Q2_K, or
// IQ2_XXS weights without materializing per-expert dispatches on the CPU.
template<mul_mv_id_disp_fn_t disp_fn>
kernel void kernel_mul_mv_id(
        constant ds4_metal_args_mul_mv_id & args,
        device const char * src0s,
        device const char * src1,
        device       char * dst,
        device const char * ids,
        threadgroup  char * shmem [[threadgroup(0)]],
        uint3  tgpig[[threadgroup_position_in_grid]],
        ushort tiitg[[thread_index_in_threadgroup]],
        ushort tiisg[[thread_index_in_simdgroup]],
        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
    (void)tiitg;

    const int iid1 = tgpig.z/args.nei0;
    const int idx  = tgpig.z%args.nei0;

    tgpig.z = 0;

    const int32_t i02 = ((device const int32_t *) (ids + iid1*args.nbi1))[idx];

    const int64_t i11 = idx % args.ne11;
    const int64_t i12 = iid1;

    const int64_t i1 = idx;
    const int64_t i2 = i12;

    device char * dst_cur = dst + (i1*args.ne0 + i2*args.ne1*args.ne0)*sizeof(float);

    if (!ds4_tp_owns_expert(i02, args.ne02, args.tp_rank, args.tp_world)) {
        /* Unowned expert under the TP split: zero this threadgroup's output
         * rows so the downstream expert-sum stages stay unchanged. */
        const short NSG = FC_mul_mv_nsg;
        const int row0 = (tgpig.x * NSG + sgitg) * args.nr0;
        device float *dst_f32 = (device float *)dst_cur;
        for (int r = 0; r < args.nr0 && row0 + r < args.ne0; r++) {
            if (tiisg == 0) dst_f32[row0 + r] = 0.0f;
        }
        return;
    }

    device const char * src0_cur =
        src0s + (int64_t)(i02 - args.tp_expert_base)*args.nb02;
    device const char * src1_cur = src1  + i11*args.nb11 + i12*args.nb12;

    ds4_metal_args_mul_mv args0 = {
        /*.ne00 =*/ args.ne00,
        /*.ne01 =*/ args.ne01,
        /*.ne02 =*/ 1,
        /*.nb00 =*/ args.nb00,
        /*.nb01 =*/ args.nb01,
        /*.nb02 =*/ args.nb02,
        /*.nb03 =*/ args.nb02,
        /*.ne10 =*/ args.ne10,
        /*.ne11 =*/ 1,
        /*.ne12 =*/ 1,
        /*.nb10 =*/ args.nb10,
        /*.nb11 =*/ args.nb11,
        /*.nb12 =*/ args.nb12,
        /*.nb13 =*/ args.nb12,
        /*.ne0  =*/ args.ne0,
        /*.ne1  =*/ 1,
        /*.nr0  =*/ args.nr0,
        /*.r2   =*/ 1,
        /*.r3   =*/ 1,
    };

    disp_fn(
        args0,
        /* src0 */ src0_cur,
        /* src1 */ src1_cur,
        /* dst  */ dst_cur,
        shmem,
        tgpig,
        tiitg,
        tiisg,
        sgitg);
}

typedef decltype(kernel_mul_mv_id<mmv_fn<kernel_mul_mv_q2_K_f32_impl<N_R0_Q2_K>>>) kernel_mul_mv_id_q_t;
typedef decltype(kernel_mul_mv_id<mmv_fn<kernel_mul_mv_q8_0_f32_impl<N_R0_Q8_0>>>) kernel_mul_mv_id_q8_0_t;

// Host-visible decode MoE matvec variants for the DS4 quant formats.
template [[host_name("kernel_mul_mv_id_q8_0_f32")]]    kernel kernel_mul_mv_id_q8_0_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_q8_0_f32_impl<N_R0_Q8_0>>>;
template [[host_name("kernel_mul_mv_id_q2_K_f32")]]    kernel kernel_mul_mv_id_q_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_q2_K_f32_impl<N_R0_Q2_K>>>;
template [[host_name("kernel_mul_mv_id_q4_K_f32")]]    kernel kernel_mul_mv_id_q_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_q4_K_f32_impl<N_R0_Q4_K>>>;
template [[host_name("kernel_mul_mv_id_q8_K_f32")]]    kernel kernel_mul_mv_id_q_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_q8_K_f32_impl<N_R0_Q8_K>>>;
template [[host_name("kernel_mul_mv_id_iq2_xxs_f32")]] kernel kernel_mul_mv_id_q_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_iq2_xxs_f32_impl<N_R0_IQ2_XXS>>>;

// Plain dense Q4_K matvec on the classic impl. The mul_mv_ext family used
// by the generic dense path tops out around 220 GB/s on M5 for the GLM
// DenseQ4 decode shapes; this impl streams the same rows at 530-650 GB/s.
kernel void kernel_mul_mv_q4_K_dense_f32(
        constant ds4_metal_args_mul_mv & args,
        device const char * src0,
        device const char * src1,
        device       char * dst,
        threadgroup  char * shmem [[threadgroup(0)]],
        uint3  tgpig[[threadgroup_position_in_grid]],
        ushort tiisg[[thread_index_in_simdgroup]],
        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
    kernel_mul_mv_q4_K_f32_impl<N_R0_Q4_K>(args, src0, src1, dst, shmem, tgpig, tiisg, sgitg);
}

// DS4 attention output low projection, specialized for the fixed block
// diagonal mapping used by the model:
//
//     low[token, group, rank] = heads[token, group, :] * Woa[group, rank, :]
//
// The generic GGML-style id matvec supports arbitrary routed expert ids.  Here
// the id is always equal to the group number, so this wrapper keeps the exact
// Q8_0 dot kernel but removes the id-buffer load and the CPU-side id table.
kernel void kernel_dsv4_attn_out_low_q8_0_f32(
        constant ds4_metal_args_mul_mv_id & args,
        device const char * src0s,
        device const char * src1,
        device       char * dst,
        threadgroup  char * shmem [[threadgroup(0)]],
        uint3  tgpig[[threadgroup_position_in_grid]],
        ushort tiitg[[thread_index_in_threadgroup]],
        ushort tiisg[[thread_index_in_simdgroup]],
        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
    const int iid1 = tgpig.z/args.nei0;
    const int idx  = tgpig.z%args.nei0;

    tgpig.z = 0;

    const int64_t i11 = idx % args.ne11;
    const int64_t i12 = iid1;

    device const char * src0_cur = src0s + idx*args.nb02;
    device const char * src1_cur = src1  + i11*args.nb11 + i12*args.nb12;
    device       char * dst_cur  = dst   + (idx*args.ne0 + i12*args.ne1*args.ne0)*sizeof(float);

    ds4_metal_args_mul_mv args0 = {
        /*.ne00 =*/ args.ne00,
        /*.ne01 =*/ args.ne01,
        /*.ne02 =*/ 1,
        /*.nb00 =*/ args.nb00,
        /*.nb01 =*/ args.nb01,
        /*.nb02 =*/ args.nb02,
        /*.nb03 =*/ args.nb02,
        /*.ne10 =*/ args.ne10,
        /*.ne11 =*/ 1,
        /*.ne12 =*/ 1,
        /*.nb10 =*/ args.nb10,
        /*.nb11 =*/ args.nb11,
        /*.nb12 =*/ args.nb12,
        /*.nb13 =*/ args.nb12,
        /*.ne0  =*/ args.ne0,
        /*.ne1  =*/ 1,
        /*.nr0  =*/ args.nr0,
        /*.r2   =*/ 1,
        /*.r3   =*/ 1,
    };

    kernel_mul_mv_q8_0_f32_impl<N_R0_Q8_0, thread ds4_metal_args_mul_mv &>(
        args0,
        src0_cur,
        src1_cur,
        dst_cur,
        shmem,
        tgpig,
        tiisg,
        sgitg);
}

kernel void kernel_dsv4_attn_out_low_q4_K_f32(
        constant ds4_metal_args_mul_mv_id & args,
        device const char * src0s,
        device const char * src1,
        device       char * dst,
        threadgroup  char * shmem [[threadgroup(0)]],
        uint3  tgpig[[threadgroup_position_in_grid]],
        ushort tiitg[[thread_index_in_threadgroup]],
        ushort tiisg[[thread_index_in_simdgroup]],
        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
    (void)tiitg;

    const int iid1 = tgpig.z / args.nei0;
    const int idx  = tgpig.z % args.nei0;

    tgpig.z = 0;

    const int64_t i11 = idx % args.ne11;
    const int64_t i12 = iid1;

    device const char * src0_cur = src0s + idx * args.nb02;
    device const char * src1_cur = src1  + i11 * args.nb11 + i12 * args.nb12;
    device       char * dst_cur  = dst   + (idx * args.ne0 + i12 * args.ne1 * args.ne0) * sizeof(float);

    ds4_metal_args_mul_mv args0 = {
        /*.ne00 =*/ args.ne00,
        /*.ne01 =*/ args.ne01,
        /*.ne02 =*/ 1,
        /*.nb00 =*/ args.nb00,
        /*.nb01 =*/ args.nb01,
        /*.nb02 =*/ args.nb02,
        /*.nb03 =*/ args.nb02,
        /*.ne10 =*/ args.ne10,
        /*.ne11 =*/ 1,
        /*.ne12 =*/ 1,
        /*.nb10 =*/ args.nb10,
        /*.nb11 =*/ args.nb11,
        /*.nb12 =*/ args.nb12,
        /*.nb13 =*/ args.nb12,
        /*.ne0  =*/ args.ne0,
        /*.ne1  =*/ 1,
        /*.nr0  =*/ args.nr0,
        /*.r2   =*/ 1,
        /*.r3   =*/ 1,
    };

    kernel_mul_mv_q4_K_f32_impl<N_R0_Q4_K>(
        args0,
        src0_cur,
        src1_cur,
        dst_cur,
        shmem,
        tgpig,
        tiisg,
        sgitg);
}

kernel void kernel_mul_mv_id_iq2_xxs_pair_f32(
        constant ds4_metal_args_mul_mv_id & args,
        device const char * src0_gate,
        device const char * src0_up,
        device const char * src1,
        device       char * dst_gate,
        device       char * dst_up,
        device const char * ids,
        threadgroup  char * shmem [[threadgroup(0)]],
        uint3  tgpig[[threadgroup_position_in_grid]],
        ushort tiitg[[thread_index_in_threadgroup]],
        ushort tiisg[[thread_index_in_simdgroup]],
        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
    const int iid1 = tgpig.z/args.nei0;
    const int idx  = tgpig.z%args.nei0;

    tgpig.z = 0;

    const int32_t i02 = ((device const int32_t *) (ids + iid1*args.nbi1))[idx];

    const int64_t i11 = idx % args.ne11;
    const int64_t i12 = iid1;

    device char * dst_gate_cur = dst_gate + (idx*args.ne0 + i12*args.ne1*args.ne0)*sizeof(float);
    device char * dst_up_cur   = dst_up   + (idx*args.ne0 + i12*args.ne1*args.ne0)*sizeof(float);

    if (!ds4_tp_owns_expert(i02, args.ne02, args.tp_rank, args.tp_world)) {
        /* Unowned expert: zero this threadgroup's gate/up rows so the
         * separate swiglu-weight stage yields zero mid rows. */
        const short NSG_z = FC_mul_mv_nsg;
        const int row0_z = (tgpig.x * NSG_z + sgitg) * args.nr0;
        device float *zg = (device float *)dst_gate_cur;
        device float *zu = (device float *)dst_up_cur;
        for (int r = 0; r < args.nr0 && row0_z + r < args.ne0; r++) {
            if (tiisg == 0) {
                zg[row0_z + r] = 0.0f;
                zu[row0_z + r] = 0.0f;
            }
        }
        return;
    }

    device const char * src0_gate_cur = src0_gate + (int64_t)(i02 - args.tp_expert_base)*args.nb02;
    device const char * src0_up_cur   = src0_up   + (int64_t)(i02 - args.tp_expert_base)*args.nb02;
    device const char * src1_cur      = src1      + i11*args.nb11 + i12*args.nb12;

    ds4_metal_args_mul_mv args0 = {
        args.ne00, args.ne01, 1,
        args.nb00, args.nb01, args.nb02, args.nb02,
        args.ne10, 1, 1,
        args.nb10, args.nb11, args.nb12, args.nb12,
        args.ne0, 1, args.nr0, 1, 1,
    };

    (void)tiitg;
    kernel_mul_mv_iq2_xxs_pair_f32_impl<N_R0_IQ2_XXS>(
        args0,
        src0_gate_cur,
        src0_up_cur,
        src1_cur,
        dst_gate_cur,
        dst_up_cur,
        shmem,
        tgpig,
        tiisg,
        sgitg);
}

// Decode-only routed expert gate/up projection fused with the DS4 activation:
//
//     mid = silu(clamp(gate)) * clamp(up) * route_weight
//
// The quantized dot products are intentionally the same IQ2_XXS paired path as
// `kernel_mul_mv_id_iq2_xxs_pair_f32`.  The only extra work is done by lane 0
// after each exact reduced row has been produced.  This removes the separate
// routed activation dispatch and avoids rereading the gate/up rows before the
// down projection.  The host uses this only for the normal release path where
// diagnostics do not request clamped gate/up intermediates.
kernel void kernel_mul_mv_id_iq2_xxs_pair_swiglu_f32(
        constant ds4_metal_args_mul_mv_id & args,
        constant ds4_metal_dsv4_moe_swiglu_weight_args & act,
        device const char * src0_gate,
        device const char * src0_up,
        device const char * src1,
        device       char * dst_gate,
        device       char * dst_up,
        device       char * dst_mid,
        device const char * ids,
        device const char * weights,
        threadgroup  char * shmem [[threadgroup(0)]],
        uint3  tgpig[[threadgroup_position_in_grid]],
        ushort tiitg[[thread_index_in_threadgroup]],
        ushort tiisg[[thread_index_in_simdgroup]],
        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
    const short NSG = FC_mul_mv_nsg;
    const int iid1 = tgpig.z / args.nei0;
    const int idx  = tgpig.z % args.nei0;

    tgpig.z = 0;

    const int32_t i02 = ((device const int32_t *) (ids + iid1 * args.nbi1))[idx];
    if (!ds4_tp_owns_expert(i02, args.ne02, args.tp_rank, args.tp_world)) return;
    const int i02b = i02 - args.tp_expert_base;
    const int64_t i11 = idx % args.ne11;
    const int64_t i12 = iid1;

    const int nb = args.ne00 / QK_K;
    const int first_row = (tgpig.x * NSG + sgitg) * N_R0_IQ2_XXS;
    const int nb32 = nb * (QK_K / 32);

    device const block_iq2_xxs *xg =
        (device const block_iq2_xxs *)(src0_gate + (int64_t)i02b * args.nb02 + (uint64_t)first_row * args.nb01);
    device const block_iq2_xxs *xu =
        (device const block_iq2_xxs *)(src0_up + (int64_t)i02b * args.nb02 + (uint64_t)first_row * args.nb01);
    device const float *y =
        (device const float *)(src1 + i11 * args.nb11 + i12 * args.nb12);

    float yl[32];
    float sumg[N_R0_IQ2_XXS] = {0.f};
    float sumu[N_R0_IQ2_XXS] = {0.f};

    threadgroup uint64_t *svalues = (threadgroup uint64_t *)(shmem);
    threadgroup uint8_t  *ssigns  = (threadgroup uint8_t *)(svalues + 256);
    {
        int nval = 4;
        int pos = (32 * sgitg + tiisg) * nval;
        for (int i = 0; i < nval; ++i) svalues[pos + i] = ds4_metal_iq2xxs_grid[pos + i];
        nval = 2;
        pos = (32 * sgitg + tiisg) * nval;
        for (int i = 0; i < nval; ++i) ssigns[pos + i] = ds4_metal_ksigns_iq2xs[pos + i];
        threadgroup_barrier(mem_flags::mem_threadgroup);
    }

    const int ix = tiisg;
    device const float *y4 = y + 32 * ix;

    for (int ib32 = ix; ib32 < nb32; ib32 += 32) {
        for (short i = 0; i < 32; ++i) {
            yl[i] = y4[i];
        }

        const int ibl = ib32 / (QK_K / 32);
        const int ib  = ib32 % (QK_K / 32);

        device const block_iq2_xxs *xgr = xg + ibl;
        device const block_iq2_xxs *xur = xu + ibl;
        device const uint16_t *qg = xgr->qs + 4 * ib;
        device const uint16_t *qu = xur->qs + 4 * ib;
        device const half *dhg = &xgr->d;
        device const half *dhu = &xur->d;

        for (short row = 0; row < N_R0_IQ2_XXS; row++) {
            device const uint8_t *aux8g = (device const uint8_t *)qg;
            device const uint8_t *aux8u = (device const uint8_t *)qu;
            const uint32_t aux32g = qg[2] | (qg[3] << 16);
            const uint32_t aux32u = qu[2] | (qu[3] << 16);
            const float dg = (float)dhg[0] * (0.5f + (aux32g >> 28));
            const float du = (float)dhu[0] * (0.5f + (aux32u >> 28));

            float sg = 0;
            float su = 0;
            for (short l = 0; l < 4; ++l) {
                const threadgroup uint8_t *gridg = (const threadgroup uint8_t *)(svalues + aux8g[l]);
                const threadgroup uint8_t *gridu = (const threadgroup uint8_t *)(svalues + aux8u[l]);
                const uint8_t signg = ssigns[(aux32g >> 7 * l) & 127];
                const uint8_t signu = ssigns[(aux32u >> 7 * l) & 127];
                for (short j = 0; j < 8; ++j) {
                    const float v = yl[8 * l + j];
                    sg += v * gridg[j] * (signg & ds4_metal_kmask_iq2xs[j] ? -1.f : 1.f);
                    su += v * gridu[j] * (signu & ds4_metal_kmask_iq2xs[j] ? -1.f : 1.f);
                }
            }
            sumg[row] += dg * sg;
            sumu[row] += du * su;

            dhg += args.nb01 / 2;
            dhu += args.nb01 / 2;
            qg  += args.nb01 / 2;
            qu  += args.nb01 / 2;
        }

        y4 += 32 * 32;
    }

    device float *dst_gate_f32 =
        (device float *)dst_gate + (uint64_t)i12 * args.ne0 * args.ne1 + (uint64_t)i11 * args.ne0;
    device float *dst_up_f32 =
        (device float *)dst_up + (uint64_t)i12 * args.ne0 * args.ne1 + (uint64_t)i11 * args.ne0;
    const uint64_t pair_row = (uint64_t)i12 * (uint64_t)args.nei0 + (uint64_t)idx;
    device float *dst_mid_f32 =
        (device float *)(dst_mid + pair_row * act.mid_row_stride);
    device const float *route_w =
        (device const float *)(weights + pair_row * act.weight_stride);

    const float c = act.clamp_value;
    const float route_weight = route_w[0];
    for (int row = 0; row < N_R0_IQ2_XXS && first_row + row < args.ne0; ++row) {
        const float sum_gate = simd_sum(sumg[row]);
        const float sum_up   = simd_sum(sumu[row]);
        if (tiisg == 0) {
            const uint out_row = first_row + row;
            const float gate = sum_gate * 0.25f;
            const float up = sum_up * 0.25f;
            float g = gate;
            float u = up;
            if (c > 1.0e-6f) {
                g = min(g, c);
                u = clamp(u, -c, c);
            }
            dst_gate_f32[out_row] = gate;
            dst_up_f32[out_row] = up;
            const float silu = g / (1.0f + exp(-g));
            dst_mid_f32[out_row] = silu * u * route_weight;
        }
    }

    (void)tiitg;
}

kernel void kernel_mul_mv_slots6_iq2_xxs_pair_swiglu_f32(
        constant ds4_metal_args_mul_mv_id & args,
        constant ds4_metal_dsv4_moe_swiglu_weight_args & act,
        device const char * src0_gate0,
        device const char * src0_gate1,
        device const char * src0_gate2,
        device const char * src0_gate3,
        device const char * src0_gate4,
        device const char * src0_gate5,
        device const char * src0_up0,
        device const char * src0_up1,
        device const char * src0_up2,
        device const char * src0_up3,
        device const char * src0_up4,
        device const char * src0_up5,
        device const char * src1,
        device       char * dst_gate,
        device       char * dst_up,
        device       char * dst_mid,
        device const char * weights,
        threadgroup  char * shmem [[threadgroup(0)]],
        uint3  tgpig[[threadgroup_position_in_grid]],
        ushort tiitg[[thread_index_in_threadgroup]],
        ushort tiisg[[thread_index_in_simdgroup]],
        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
    const int iid1 = tgpig.z / args.nei0;
    const int idx  = tgpig.z % args.nei0;

    tgpig.z = 0;

    const int64_t i11 = idx % args.ne11;
    const int64_t i12 = iid1;

    device const char *src0_gate_cur = src0_gate0;
    device const char *src0_up_cur = src0_up0;
    switch (idx) {
    case 1: src0_gate_cur = src0_gate1; src0_up_cur = src0_up1; break;
    case 2: src0_gate_cur = src0_gate2; src0_up_cur = src0_up2; break;
    case 3: src0_gate_cur = src0_gate3; src0_up_cur = src0_up3; break;
    case 4: src0_gate_cur = src0_gate4; src0_up_cur = src0_up4; break;
    case 5: src0_gate_cur = src0_gate5; src0_up_cur = src0_up5; break;
    default: break;
    }

    device const char *src1_cur = src1 + i11 * args.nb11 + i12 * args.nb12;

    device char *dst_gate_cur = dst_gate + (idx * args.ne0 + i12 * args.ne1 * args.ne0) * sizeof(float);
    device char *dst_up_cur   = dst_up   + (idx * args.ne0 + i12 * args.ne1 * args.ne0) * sizeof(float);

    ds4_metal_args_mul_mv args0 = {
        args.ne00, args.ne01, 1,
        args.nb00, args.nb01, args.nb02, args.nb02,
        args.ne10, 1, 1,
        args.nb10, args.nb11, args.nb12, args.nb12,
        args.ne0, 1, args.nr0, 1, 1,
    };

    kernel_mul_mv_iq2_xxs_pair_f32_impl<N_R0_IQ2_XXS>(
        args0,
        src0_gate_cur,
        src0_up_cur,
        src1_cur,
        dst_gate_cur,
        dst_up_cur,
        shmem,
        tgpig,
        tiisg,
        sgitg);

    const short NSG = FC_mul_mv_nsg;
    const int first_row = (tgpig.x * NSG + sgitg) * N_R0_IQ2_XXS;
    device float *gate_f32 = (device float *)dst_gate_cur;
    device float *up_f32 = (device float *)dst_up_cur;
    const uint64_t pair_row = (uint64_t)i12 * (uint64_t)args.nei0 + (uint64_t)idx;
    device float *mid_f32 = (device float *)(dst_mid + pair_row * act.mid_row_stride);
    device const float *route_w = (device const float *)(weights + pair_row * act.weight_stride);
    const float c = act.clamp_value;
    const float route_weight = route_w[0];

    if (tiisg == 0) {
        for (int row = 0; row < N_R0_IQ2_XXS && first_row + row < args.ne0; ++row) {
            const uint out_row = first_row + row;
            float g = gate_f32[out_row];
            float u = up_f32[out_row];
            if (c > 1.0e-6f) {
                g = min(g, c);
                u = clamp(u, -c, c);
            }
            const float silu = g / (1.0f + exp(-g));
            mid_f32[out_row] = silu * u * route_weight;
        }
    }

    (void)tiitg;
}

kernel void kernel_mul_mv_addr_iq2_xxs_pair_swiglu_f32(
        constant ds4_metal_args_mul_mv_id & args,
        constant ds4_metal_dsv4_moe_swiglu_weight_args & act,
        device const uint64_t * gate_addrs,
        device const uint64_t * up_addrs,
        device const char * src1,
        device       char * dst_gate,
        device       char * dst_up,
        device       char * dst_mid,
        device const char * ids,
        device const char * weights,
        threadgroup  char * shmem [[threadgroup(0)]],
        uint3  tgpig[[threadgroup_position_in_grid]],
        ushort tiitg[[thread_index_in_threadgroup]],
        ushort tiisg[[thread_index_in_simdgroup]],
        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
    const int iid1 = tgpig.z / args.nei0;
    const int idx  = tgpig.z % args.nei0;

    tgpig.z = 0;

    const int32_t i02 = ((device const int32_t *)(ids + iid1 * args.nbi1))[idx];
    if (i02 < 0 || i02 >= args.ne02 || i02 >= 384) {
        return;
    }
    const uint64_t gate_addr = gate_addrs[(uint)i02];
    const uint64_t up_addr = up_addrs[(uint)i02];
    if (gate_addr == 0 || up_addr == 0) {
        return;
    }

    const int64_t i11 = idx % args.ne11;
    const int64_t i12 = iid1;

    device const char *src0_gate_cur =
        reinterpret_cast<device const char *>(gate_addr);
    device const char *src0_up_cur =
        reinterpret_cast<device const char *>(up_addr);
    device const char *src1_cur = src1 + i11 * args.nb11 + i12 * args.nb12;

    device char *dst_gate_cur = dst_gate + (idx * args.ne0 + i12 * args.ne1 * args.ne0) * sizeof(float);
    device char *dst_up_cur   = dst_up   + (idx * args.ne0 + i12 * args.ne1 * args.ne0) * sizeof(float);

    ds4_metal_args_mul_mv args0 = {
        args.ne00, args.ne01, 1,
        args.nb00, args.nb01, args.nb02, args.nb02,
        args.ne10, 1, 1,
        args.nb10, args.nb11, args.nb12, args.nb12,
        args.ne0, 1, args.nr0, 1, 1,
    };

    kernel_mul_mv_iq2_xxs_pair_f32_impl<N_R0_IQ2_XXS>(
        args0,
        src0_gate_cur,
        src0_up_cur,
        src1_cur,
        dst_gate_cur,
        dst_up_cur,
        shmem,
        tgpig,
        tiisg,
        sgitg);

    const short NSG = FC_mul_mv_nsg;
    const int first_row = (tgpig.x * NSG + sgitg) * N_R0_IQ2_XXS;
    device float *gate_f32 = (device float *)dst_gate_cur;
    device float *up_f32 = (device float *)dst_up_cur;
    const uint64_t pair_row = (uint64_t)i12 * (uint64_t)args.nei0 + (uint64_t)idx;
    device float *mid_f32 = (device float *)(dst_mid + pair_row * act.mid_row_stride);
    device const float *route_w = (device const float *)(weights + pair_row * act.weight_stride);
    const float c = act.clamp_value;
    const float route_weight = route_w[0];

    if (tiisg == 0) {
        for (int row = 0; row < N_R0_IQ2_XXS && first_row + row < args.ne0; ++row) {
            const uint out_row = first_row + row;
            float g = gate_f32[out_row];
            float u = up_f32[out_row];
            if (c > 1.0e-6f) {
                g = min(g, c);
                u = clamp(u, -c, c);
            }
            const float silu = g / (1.0f + exp(-g));
            mid_f32[out_row] = silu * u * route_weight;
        }
    }

    (void)tiitg;
}

kernel void kernel_mul_mv_addr_iq2_xxs_f32(
        constant ds4_metal_args_mul_mv_id & args,
        device const uint64_t * addrs,
        device const char * src1,
        device       char * dst,
        device const char * ids,
        threadgroup  char * shmem [[threadgroup(0)]],
        uint3  tgpig[[threadgroup_position_in_grid]],
        ushort tiitg[[thread_index_in_threadgroup]],
        ushort tiisg[[thread_index_in_simdgroup]],
        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
    (void)tiitg;

    const int iid1 = tgpig.z / args.nei0;
    const int idx  = tgpig.z % args.nei0;

    tgpig.z = 0;

    const int32_t i02 = ((device const int32_t *)(ids + iid1 * args.nbi1))[idx];
    if (i02 < 0 || i02 >= args.ne02 || i02 >= 384) {
        return;
    }
    const uint64_t addr = addrs[(uint)i02];
    if (addr == 0) {
        return;
    }

    const int64_t i11 = idx % args.ne11;
    const int64_t i12 = iid1;

    device const char *src0_cur = reinterpret_cast<device const char *>(addr);
    device const char *src1_cur = src1 + i11 * args.nb11 + i12 * args.nb12;
    device char *dst_cur = dst + (idx * args.ne0 + i12 * args.ne1 * args.ne0) * sizeof(float);

    ds4_metal_args_mul_mv args0 = {
        args.ne00, args.ne01, 1,
        args.nb00, args.nb01, args.nb02, args.nb02,
        args.ne10, 1, 1,
        args.nb10, args.nb11, args.nb12, args.nb12,
        args.ne0, 1, args.nr0, 1, 1,
    };

    kernel_mul_mv_iq2_xxs_f32_impl<N_R0_IQ2_XXS>(
        args0,
        src0_cur,
        src1_cur,
        dst_cur,
        shmem,
        tgpig,
        tiisg,
        sgitg);
}

kernel void kernel_mul_mv_addr_iq2_xxs_pair_swiglu_masked_f32(
        constant ds4_metal_args_mul_mv_id & args,
        constant ds4_metal_dsv4_moe_swiglu_weight_args & act,
        constant ds4_metal_stream_expert_split_args & split,
        device const uint64_t * gate_addrs,
        device const uint64_t * up_addrs,
        device const char * src1,
        device       char * dst_gate,
        device       char * dst_up,
        device       char * dst_mid,
        device const char * ids,
        device const char * weights,
        threadgroup  char * shmem [[threadgroup(0)]],
        uint3  tgpig[[threadgroup_position_in_grid]],
        ushort tiitg[[thread_index_in_threadgroup]],
        ushort tiisg[[thread_index_in_simdgroup]],
        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
    const int iid1 = tgpig.z / args.nei0;
    const int idx  = tgpig.z % args.nei0;
    if ((split.active_mask & (1u << (uint)idx)) == 0) {
        return;
    }

    tgpig.z = 0;

    const int32_t i02 = ((device const int32_t *)(ids + iid1 * args.nbi1))[idx];
    if (i02 < 0 || i02 >= args.ne02 || i02 >= 384) {
        return;
    }
    const uint64_t gate_addr = gate_addrs[(uint)i02];
    const uint64_t up_addr = up_addrs[(uint)i02];
    if (gate_addr == 0 || up_addr == 0) {
        return;
    }

    const int64_t i11 = idx % args.ne11;
    const int64_t i12 = iid1;

    device const char *src0_gate_cur =
        reinterpret_cast<device const char *>(gate_addr);
    device const char *src0_up_cur =
        reinterpret_cast<device const char *>(up_addr);
    device const char *src1_cur = src1 + i11 * args.nb11 + i12 * args.nb12;

    device char *dst_gate_cur = dst_gate + (idx * args.ne0 + i12 * args.ne1 * args.ne0) * sizeof(float);
    device char *dst_up_cur   = dst_up   + (idx * args.ne0 + i12 * args.ne1 * args.ne0) * sizeof(float);

    ds4_metal_args_mul_mv args0 = {
        args.ne00, args.ne01, 1,
        args.nb00, args.nb01, args.nb02, args.nb02,
        args.ne10, 1, 1,
        args.nb10, args.nb11, args.nb12, args.nb12,
        args.ne0, 1, args.nr0, 1, 1,
    };

    kernel_mul_mv_iq2_xxs_pair_f32_impl<N_R0_IQ2_XXS>(
        args0,
        src0_gate_cur,
        src0_up_cur,
        src1_cur,
        dst_gate_cur,
        dst_up_cur,
        shmem,
        tgpig,
        tiisg,
        sgitg);

    const short NSG = FC_mul_mv_nsg;
    const int first_row = (tgpig.x * NSG + sgitg) * N_R0_IQ2_XXS;
    device float *gate_f32 = (device float *)dst_gate_cur;
    device float *up_f32 = (device float *)dst_up_cur;
    const uint64_t pair_row = (uint64_t)i12 * (uint64_t)args.nei0 + (uint64_t)idx;
    device float *mid_f32 = (device float *)(dst_mid + pair_row * act.mid_row_stride);
    device const float *route_w = (device const float *)(weights + pair_row * act.weight_stride);
    const float c = act.clamp_value;
    const float route_weight = route_w[0];

    if (tiisg == 0) {
        for (int row = 0; row < N_R0_IQ2_XXS && first_row + row < args.ne0; ++row) {
            const uint out_row = first_row + row;
            float g = gate_f32[out_row];
            float u = up_f32[out_row];
            if (c > 1.0e-6f) {
                g = min(g, c);
                u = clamp(u, -c, c);
            }
            const float silu = g / (1.0f + exp(-g));
            mid_f32[out_row] = silu * u * route_weight;
        }
    }

    (void)tiitg;
}

kernel void kernel_stream_expert_cache_validate(
        constant ds4_metal_stream_expert_validate_args & args,
        device const char * ids,
        device const uint64_t * gate_addrs,
        device const uint64_t * up_addrs,
        device const uint64_t * down_addrs,
        device uint32_t * status,
        uint tid [[thread_position_in_grid]]) {
    if (tid != 0) return;

    uint32_t miss_mask = 0;
    uint32_t invalid_mask = 0;
    const uint32_t n_expert = min(args.n_expert, (uint32_t)6);
    device const int32_t *selected = (device const int32_t *)ids;

    status[3] = n_expert;
    for (uint32_t i = 0; i < 6; i++) {
        const int32_t expert = i < n_expert ? selected[i] : -1;
        status[4 + i] = as_type<uint32_t>(expert);
        if (i >= n_expert) continue;
        if (expert < 0 ||
            (uint32_t)expert >= args.n_total_expert ||
            (uint32_t)expert >= 384) {
            invalid_mask |= (1u << i);
            continue;
        }
        const uint32_t e = (uint32_t)expert;
        if (gate_addrs[e] == 0 || up_addrs[e] == 0 || down_addrs[e] == 0) {
            miss_mask |= (1u << i);
        }
    }

    status[0] = (miss_mask == 0 && invalid_mask == 0) ? 1u : 0u;
    status[1] = miss_mask;
    status[2] = invalid_mask;
}

kernel void kernel_mul_mv_id_q4_K_pair_f32(
        constant ds4_metal_args_mul_mv_id & args,
        device const char * src0_gate,
        device const char * src0_up,
        device const char * src1,
        device       char * dst_gate,
        device       char * dst_up,
        device const char * ids,
        threadgroup  char * shmem [[threadgroup(0)]],
        uint3  tgpig[[threadgroup_position_in_grid]],
        ushort tiitg[[thread_index_in_threadgroup]],
        ushort tiisg[[thread_index_in_simdgroup]],
        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
    const int iid1 = tgpig.z / args.nei0;
    const int idx  = tgpig.z % args.nei0;

    tgpig.z = 0;

    const int32_t i02 = ((device const int32_t *)(ids + iid1 * args.nbi1))[idx];
    const int64_t i11 = idx % args.ne11;
    const int64_t i12 = iid1;

    device char *dst_gate_cur = dst_gate + (idx * args.ne0 + i12 * args.ne1 * args.ne0) * sizeof(float);
    device char *dst_up_cur   = dst_up   + (idx * args.ne0 + i12 * args.ne1 * args.ne0) * sizeof(float);

    if (!ds4_tp_owns_expert(i02, args.ne02, args.tp_rank, args.tp_world)) {
        const short NSG_z = FC_mul_mv_nsg;
        const int row0_z = (tgpig.x * NSG_z + sgitg) * args.nr0;
        device float *zg = (device float *)dst_gate_cur;
        device float *zu = (device float *)dst_up_cur;
        for (int r = 0; r < args.nr0 && row0_z + r < args.ne0; r++) {
            if (tiisg == 0) {
                zg[row0_z + r] = 0.0f;
                zu[row0_z + r] = 0.0f;
            }
        }
        return;
    }

    device const char *src0_gate_cur = src0_gate + (int64_t)(i02 - args.tp_expert_base) * args.nb02;
    device const char *src0_up_cur   = src0_up   + (int64_t)(i02 - args.tp_expert_base) * args.nb02;
    device const char *src1_cur      = src1      + i11 * args.nb11 + i12 * args.nb12;

    ds4_metal_args_mul_mv args0 = {
        args.ne00, args.ne01, 1,
        args.nb00, args.nb01, args.nb02, args.nb02,
        args.ne10, 1, 1,
        args.nb10, args.nb11, args.nb12, args.nb12,
        args.ne0, 1, args.nr0, 1, 1,
    };

    (void)tiitg;
    kernel_mul_mv_q4_K_f32_impl<N_R0_Q4_K>(
        args0,
        src0_gate_cur,
        src1_cur,
        dst_gate_cur,
        shmem,
        tgpig,
        tiisg,
        sgitg);
    kernel_mul_mv_q4_K_f32_impl<N_R0_Q4_K>(
        args0,
        src0_up_cur,
        src1_cur,
        dst_up_cur,
        shmem,
        tgpig,
        tiisg,
        sgitg);
}

// Same release-path fusion as the IQ2_XXS kernel above for the Q4_K expert
// variant.  The Q4 pair path reuses the existing exact matvec implementation
// for gate and up, then the same lane that wrote each row derives the routed
// SwiGLU input.  This keeps Q4 behavior aligned with the Q2 optimization while
// preserving the old pair projection arithmetic.
kernel void kernel_mul_mv_id_q4_K_pair_swiglu_f32(
        constant ds4_metal_args_mul_mv_id & args,
        constant ds4_metal_dsv4_moe_swiglu_weight_args & act,
        device const char * src0_gate,
        device const char * src0_up,
        device const char * src1,
        device       char * dst_gate,
        device       char * dst_up,
        device       char * dst_mid,
        device const char * ids,
        device const char * weights,
        threadgroup  char * shmem [[threadgroup(0)]],
        uint3  tgpig[[threadgroup_position_in_grid]],
        ushort tiitg[[thread_index_in_threadgroup]],
        ushort tiisg[[thread_index_in_simdgroup]],
        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
    const int iid1 = tgpig.z / args.nei0;
    const int idx  = tgpig.z % args.nei0;

    tgpig.z = 0;

    const int32_t i02 = ((device const int32_t *)(ids + iid1 * args.nbi1))[idx];
    if (!ds4_tp_owns_expert(i02, args.ne02, args.tp_rank, args.tp_world)) return;
    const int i02b = i02 - args.tp_expert_base;
    const int64_t i11 = idx % args.ne11;
    const int64_t i12 = iid1;

    device const char *src0_gate_cur = src0_gate + (int64_t)i02b * args.nb02;
    device const char *src0_up_cur   = src0_up   + (int64_t)i02b * args.nb02;
    device const char *src1_cur      = src1      + i11 * args.nb11 + i12 * args.nb12;

    device char *dst_gate_cur = dst_gate + (idx * args.ne0 + i12 * args.ne1 * args.ne0) * sizeof(float);
    device char *dst_up_cur   = dst_up   + (idx * args.ne0 + i12 * args.ne1 * args.ne0) * sizeof(float);

    const short NSG = FC_mul_mv_nsg;
    constexpr uint16_t kmask1 = 0x3f3f;
    constexpr uint16_t kmask2 = 0x0f0f;
    constexpr uint16_t kmask3 = 0xc0c0;

    const short ix = tiisg / 8;
    const short it = tiisg % 8;
    const short iq = it / 4;
    const short ir = it % 4;
    const int nb = args.ne00 / QK_K;
    const int first_row = (tgpig.x * NSG + sgitg) * N_R0_Q4_K;
    device float *gate_f32 = (device float *)dst_gate_cur;
    device float *up_f32 = (device float *)dst_up_cur;
    const uint64_t pair_row = (uint64_t)i12 * (uint64_t)args.nei0 + (uint64_t)idx;
    device float *mid_f32 = (device float *)(dst_mid + pair_row * act.mid_row_stride);
    device const float *route_w = (device const float *)(weights + pair_row * act.weight_stride);
    const float c = act.clamp_value;
    const float route_weight = route_w[0];

    device const block_q4_K *xg =
        (device const block_q4_K *)(src0_gate_cur + (uint64_t)first_row * args.nb01);
    device const block_q4_K *xu =
        (device const block_q4_K *)(src0_up_cur + (uint64_t)first_row * args.nb01);
    device const float *y = (device const float *)src1_cur;
    device const float *y4 = y + ix * QK_K + 64 * iq + 8 * ir;

    float sumg[N_R0_Q4_K] = {0.f};
    float sumu[N_R0_Q4_K] = {0.f};
    uint16_t sc16[4];
    thread const uint8_t *sc8 = (thread const uint8_t *)sc16;

    for (int ib = ix; ib < nb; ib += 4) {
        float yl[16];
        float yh[16];
        float4 sumy = {0.f, 0.f, 0.f, 0.f};

        for (short i = 0; i < 8; ++i) {
            yl[i + 0] = y4[i +   0]; sumy[0] += yl[i + 0];
            yl[i + 8] = y4[i +  32]; sumy[1] += yl[i + 8];
            yh[i + 0] = y4[i + 128]; sumy[2] += yh[i + 0];
            yh[i + 8] = y4[i + 160]; sumy[3] += yh[i + 8];
        }

        device const uint16_t *scg = (device const uint16_t *)xg[ib].scales + iq;
        device const uint16_t *qg1 = (device const uint16_t *)xg[ib].qs + 16 * iq + 4 * ir;
        device const half *dhg = &xg[ib].d;
        device const uint16_t *scu = (device const uint16_t *)xu[ib].scales + iq;
        device const uint16_t *qu1 = (device const uint16_t *)xu[ib].qs + 16 * iq + 4 * ir;
        device const half *dhu = &xu[ib].d;

        for (short row = 0; row < N_R0_Q4_K; row++) {
            sc16[0] = scg[0] & kmask1;
            sc16[1] = scg[2] & kmask1;
            sc16[2] = ((scg[4] >> 0) & kmask2) | ((scg[0] & kmask3) >> 2);
            sc16[3] = ((scg[4] >> 4) & kmask2) | ((scg[2] & kmask3) >> 2);

            device const uint16_t *qg2 = qg1 + 32;
            float4 acc1g = {0.f, 0.f, 0.f, 0.f};
            float4 acc2g = {0.f, 0.f, 0.f, 0.f};

            FOR_UNROLL (short i = 0; i < 4; ++i) {
                acc1g[0] += yl[2 * i + 0] * (qg1[i] & 0x000F);
                acc1g[1] += yl[2 * i + 1] * (qg1[i] & 0x0F00);
                acc1g[2] += yl[2 * i + 8] * (qg1[i] & 0x00F0);
                acc1g[3] += yl[2 * i + 9] * (qg1[i] & 0xF000);
                acc2g[0] += yh[2 * i + 0] * (qg2[i] & 0x000F);
                acc2g[1] += yh[2 * i + 1] * (qg2[i] & 0x0F00);
                acc2g[2] += yh[2 * i + 8] * (qg2[i] & 0x00F0);
                acc2g[3] += yh[2 * i + 9] * (qg2[i] & 0xF000);
            }

            sumg[row] += dhg[0] * ((acc1g[0] + 1.f / 256.f * acc1g[1]) * sc8[0] +
                                   (acc1g[2] + 1.f / 256.f * acc1g[3]) * sc8[1] * 1.f / 16.f +
                                   (acc2g[0] + 1.f / 256.f * acc2g[1]) * sc8[4] +
                                   (acc2g[2] + 1.f / 256.f * acc2g[3]) * sc8[5] * 1.f / 16.f) -
                         dhg[1] * (sumy[0] * sc8[2] + sumy[1] * sc8[3] +
                                   sumy[2] * sc8[6] + sumy[3] * sc8[7]);

            sc16[0] = scu[0] & kmask1;
            sc16[1] = scu[2] & kmask1;
            sc16[2] = ((scu[4] >> 0) & kmask2) | ((scu[0] & kmask3) >> 2);
            sc16[3] = ((scu[4] >> 4) & kmask2) | ((scu[2] & kmask3) >> 2);

            device const uint16_t *qu2 = qu1 + 32;
            float4 acc1u = {0.f, 0.f, 0.f, 0.f};
            float4 acc2u = {0.f, 0.f, 0.f, 0.f};

            FOR_UNROLL (short i = 0; i < 4; ++i) {
                acc1u[0] += yl[2 * i + 0] * (qu1[i] & 0x000F);
                acc1u[1] += yl[2 * i + 1] * (qu1[i] & 0x0F00);
                acc1u[2] += yl[2 * i + 8] * (qu1[i] & 0x00F0);
                acc1u[3] += yl[2 * i + 9] * (qu1[i] & 0xF000);
                acc2u[0] += yh[2 * i + 0] * (qu2[i] & 0x000F);
                acc2u[1] += yh[2 * i + 1] * (qu2[i] & 0x0F00);
                acc2u[2] += yh[2 * i + 8] * (qu2[i] & 0x00F0);
                acc2u[3] += yh[2 * i + 9] * (qu2[i] & 0xF000);
            }

            sumu[row] += dhu[0] * ((acc1u[0] + 1.f / 256.f * acc1u[1]) * sc8[0] +
                                   (acc1u[2] + 1.f / 256.f * acc1u[3]) * sc8[1] * 1.f / 16.f +
                                   (acc2u[0] + 1.f / 256.f * acc2u[1]) * sc8[4] +
                                   (acc2u[2] + 1.f / 256.f * acc2u[3]) * sc8[5] * 1.f / 16.f) -
                         dhu[1] * (sumy[0] * sc8[2] + sumy[1] * sc8[3] +
                                   sumy[2] * sc8[6] + sumy[3] * sc8[7]);

            qg1 += args.nb01 / 2;
            scg += args.nb01 / 2;
            dhg += args.nb01 / 2;
            qu1 += args.nb01 / 2;
            scu += args.nb01 / 2;
            dhu += args.nb01 / 2;
        }

        y4 += 4 * QK_K;
    }

    for (int row = 0; row < N_R0_Q4_K && first_row + row < args.ne0; ++row) {
        const float gate = simd_sum(sumg[row]);
        const float up = simd_sum(sumu[row]);
        if (tiisg == 0) {
            const uint out_row = first_row + row;
            float g = gate;
            float u = up;
            if (c > 1.0e-6f) {
                g = min(g, c);
                u = clamp(u, -c, c);
            }
            gate_f32[out_row] = gate;
            up_f32[out_row] = up;
            const float silu = g / (1.0f + exp(-g));
            mid_f32[out_row] = silu * u * route_weight;
        }
    }

    (void)tiitg;
}

kernel void kernel_mul_mv_table_q4_K_pair_swiglu_f32(
        constant ds4_metal_args_mul_mv_id & args,
        constant ds4_metal_dsv4_moe_swiglu_weight_args & act,
        device const ds4_metal_q4_expert_table & gate_table,
        device const ds4_metal_q4_expert_table & up_table,
        device const char * src1,
        device       char * dst_gate,
        device       char * dst_up,
        device       char * dst_mid,
        device const char * ids,
        device const char * weights,
        threadgroup  char * shmem [[threadgroup(0)]],
        uint3  tgpig[[threadgroup_position_in_grid]],
        ushort tiitg[[thread_index_in_threadgroup]],
        ushort tiisg[[thread_index_in_simdgroup]],
        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
    const int iid1 = tgpig.z / args.nei0;
    const int idx  = tgpig.z % args.nei0;

    tgpig.z = 0;

    const int32_t i02 = ((device const int32_t *)(ids + iid1 * args.nbi1))[idx];
    if (i02 < 0 || i02 >= args.ne02 || i02 >= 384) {
        return;
    }
    const int64_t i11 = idx % args.ne11;
    const int64_t i12 = iid1;

    device const char *src0_gate_cur = gate_table.experts[(uint)i02];
    device const char *src0_up_cur   = up_table.experts[(uint)i02];
    device const char *src1_cur      = src1 + i11 * args.nb11 + i12 * args.nb12;

    device char *dst_gate_cur = dst_gate + (idx * args.ne0 + i12 * args.ne1 * args.ne0) * sizeof(float);
    device char *dst_up_cur   = dst_up   + (idx * args.ne0 + i12 * args.ne1 * args.ne0) * sizeof(float);

    const short NSG = FC_mul_mv_nsg;
    constexpr uint16_t kmask1 = 0x3f3f;
    constexpr uint16_t kmask2 = 0x0f0f;
    constexpr uint16_t kmask3 = 0xc0c0;

    const short ix = tiisg / 8;
    const short it = tiisg % 8;
    const short iq = it / 4;
    const short ir = it % 4;
    const int nb = args.ne00 / QK_K;
    const int first_row = (tgpig.x * NSG + sgitg) * N_R0_Q4_K;
    device float *gate_f32 = (device float *)dst_gate_cur;
    device float *up_f32 = (device float *)dst_up_cur;
    const uint64_t pair_row = (uint64_t)i12 * (uint64_t)args.nei0 + (uint64_t)idx;
    device float *mid_f32 = (device float *)(dst_mid + pair_row * act.mid_row_stride);
    device const float *route_w = (device const float *)(weights + pair_row * act.weight_stride);
    const float c = act.clamp_value;
    const float route_weight = route_w[0];

    device const block_q4_K *xg =
        (device const block_q4_K *)(src0_gate_cur + (uint64_t)first_row * args.nb01);
    device const block_q4_K *xu =
        (device const block_q4_K *)(src0_up_cur + (uint64_t)first_row * args.nb01);
    device const float *y = (device const float *)src1_cur;
    device const float *y4 = y + ix * QK_K + 64 * iq + 8 * ir;

    float sumg[N_R0_Q4_K] = {0.f};
    float sumu[N_R0_Q4_K] = {0.f};
    uint16_t sc16[4];
    thread const uint8_t *sc8 = (thread const uint8_t *)sc16;

    for (int ib = ix; ib < nb; ib += 4) {
        float yl[16];
        float yh[16];
        float4 sumy = {0.f, 0.f, 0.f, 0.f};

        for (short i = 0; i < 8; ++i) {
            yl[i + 0] = y4[i +   0]; sumy[0] += yl[i + 0];
            yl[i + 8] = y4[i +  32]; sumy[1] += yl[i + 8];
            yh[i + 0] = y4[i + 128]; sumy[2] += yh[i + 0];
            yh[i + 8] = y4[i + 160]; sumy[3] += yh[i + 8];
        }

        device const uint16_t *scg = (device const uint16_t *)xg[ib].scales + iq;
        device const uint16_t *qg1 = (device const uint16_t *)xg[ib].qs + 16 * iq + 4 * ir;
        device const half *dhg = &xg[ib].d;
        device const uint16_t *scu = (device const uint16_t *)xu[ib].scales + iq;
        device const uint16_t *qu1 = (device const uint16_t *)xu[ib].qs + 16 * iq + 4 * ir;
        device const half *dhu = &xu[ib].d;

        for (short row = 0; row < N_R0_Q4_K; row++) {
            sc16[0] = scg[0] & kmask1;
            sc16[1] = scg[2] & kmask1;
            sc16[2] = ((scg[4] >> 0) & kmask2) | ((scg[0] & kmask3) >> 2);
            sc16[3] = ((scg[4] >> 4) & kmask2) | ((scg[2] & kmask3) >> 2);

            device const uint16_t *qg2 = qg1 + 32;
            float4 acc1g = {0.f, 0.f, 0.f, 0.f};
            float4 acc2g = {0.f, 0.f, 0.f, 0.f};

            FOR_UNROLL (short i = 0; i < 4; ++i) {
                acc1g[0] += yl[2 * i + 0] * (qg1[i] & 0x000F);
                acc1g[1] += yl[2 * i + 1] * (qg1[i] & 0x0F00);
                acc1g[2] += yl[2 * i + 8] * (qg1[i] & 0x00F0);
                acc1g[3] += yl[2 * i + 9] * (qg1[i] & 0xF000);
                acc2g[0] += yh[2 * i + 0] * (qg2[i] & 0x000F);
                acc2g[1] += yh[2 * i + 1] * (qg2[i] & 0x0F00);
                acc2g[2] += yh[2 * i + 8] * (qg2[i] & 0x00F0);
                acc2g[3] += yh[2 * i + 9] * (qg2[i] & 0xF000);
            }

            sumg[row] += dhg[0] * ((acc1g[0] + 1.f / 256.f * acc1g[1]) * sc8[0] +
                                   (acc1g[2] + 1.f / 256.f * acc1g[3]) * sc8[1] * 1.f / 16.f +
                                   (acc2g[0] + 1.f / 256.f * acc2g[1]) * sc8[4] +
                                   (acc2g[2] + 1.f / 256.f * acc2g[3]) * sc8[5] * 1.f / 16.f) -
                         dhg[1] * (sumy[0] * sc8[2] + sumy[1] * sc8[3] +
                                   sumy[2] * sc8[6] + sumy[3] * sc8[7]);

            sc16[0] = scu[0] & kmask1;
            sc16[1] = scu[2] & kmask1;
            sc16[2] = ((scu[4] >> 0) & kmask2) | ((scu[0] & kmask3) >> 2);
            sc16[3] = ((scu[4] >> 4) & kmask2) | ((scu[2] & kmask3) >> 2);

            device const uint16_t *qu2 = qu1 + 32;
            float4 acc1u = {0.f, 0.f, 0.f, 0.f};
            float4 acc2u = {0.f, 0.f, 0.f, 0.f};

            FOR_UNROLL (short i = 0; i < 4; ++i) {
                acc1u[0] += yl[2 * i + 0] * (qu1[i] & 0x000F);
                acc1u[1] += yl[2 * i + 1] * (qu1[i] & 0x0F00);
                acc1u[2] += yl[2 * i + 8] * (qu1[i] & 0x00F0);
                acc1u[3] += yl[2 * i + 9] * (qu1[i] & 0xF000);
                acc2u[0] += yh[2 * i + 0] * (qu2[i] & 0x000F);
                acc2u[1] += yh[2 * i + 1] * (qu2[i] & 0x0F00);
                acc2u[2] += yh[2 * i + 8] * (qu2[i] & 0x00F0);
                acc2u[3] += yh[2 * i + 9] * (qu2[i] & 0xF000);
            }

            sumu[row] += dhu[0] * ((acc1u[0] + 1.f / 256.f * acc1u[1]) * sc8[0] +
                                   (acc1u[2] + 1.f / 256.f * acc1u[3]) * sc8[1] * 1.f / 16.f +
                                   (acc2u[0] + 1.f / 256.f * acc2u[1]) * sc8[4] +
                                   (acc2u[2] + 1.f / 256.f * acc2u[3]) * sc8[5] * 1.f / 16.f) -
                         dhu[1] * (sumy[0] * sc8[2] + sumy[1] * sc8[3] +
                                   sumy[2] * sc8[6] + sumy[3] * sc8[7]);

            qg1 += args.nb01 / 2;
            scg += args.nb01 / 2;
            dhg += args.nb01 / 2;
            qu1 += args.nb01 / 2;
            scu += args.nb01 / 2;
            dhu += args.nb01 / 2;
        }

        y4 += 4 * QK_K;
    }

    for (int row = 0; row < N_R0_Q4_K && first_row + row < args.ne0; ++row) {
        const float gate = simd_sum(sumg[row]);
        const float up = simd_sum(sumu[row]);
        if (tiisg == 0) {
            const uint out_row = first_row + row;
            float g = gate;
            float u = up;
            if (c > 1.0e-6f) {
                g = min(g, c);
                u = clamp(u, -c, c);
            }
            gate_f32[out_row] = gate;
            up_f32[out_row] = up;
            const float silu = g / (1.0f + exp(-g));
            mid_f32[out_row] = silu * u * route_weight;
        }
    }

    (void)tiitg;
}

kernel void kernel_mul_mv_addr_q4_K_pair_swiglu_f32(
        constant ds4_metal_args_mul_mv_id & args,
        constant ds4_metal_dsv4_moe_swiglu_weight_args & act,
        device const ulong * gate_addrs,
        device const ulong * up_addrs,
        device const char * src1,
        device       char * dst_gate,
        device       char * dst_up,
        device       char * dst_mid,
        device const char * ids,
        device const char * weights,
        threadgroup  char * shmem [[threadgroup(0)]],
        uint3  tgpig[[threadgroup_position_in_grid]],
        ushort tiitg[[thread_index_in_threadgroup]],
        ushort tiisg[[thread_index_in_simdgroup]],
        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
    const int iid1 = tgpig.z / args.nei0;
    const int idx  = tgpig.z % args.nei0;

    tgpig.z = 0;

    const int32_t i02 = ((device const int32_t *)(ids + iid1 * args.nbi1))[idx];
    if (i02 < 0 || i02 >= args.ne02 || i02 >= 384) {
        return;
    }
    const int64_t i11 = idx % args.ne11;
    const int64_t i12 = iid1;

    device const char *src0_gate_cur =
        reinterpret_cast<device const char *>(gate_addrs[(uint)i02]);
    device const char *src0_up_cur =
        reinterpret_cast<device const char *>(up_addrs[(uint)i02]);
    device const char *src1_cur = src1 + i11 * args.nb11 + i12 * args.nb12;

    device char *dst_gate_cur = dst_gate + (idx * args.ne0 + i12 * args.ne1 * args.ne0) * sizeof(float);
    device char *dst_up_cur   = dst_up   + (idx * args.ne0 + i12 * args.ne1 * args.ne0) * sizeof(float);

    ds4_metal_args_mul_mv args0 = {
        args.ne00, args.ne01, 1,
        args.nb00, args.nb01, args.nb02, args.nb02,
        args.ne10, 1, 1,
        args.nb10, args.nb11, args.nb12, args.nb12,
        args.ne0, 1, args.nr0, 1, 1,
    };

    kernel_mul_mv_q4_K_f32_impl<N_R0_Q4_K>(
        args0,
        src0_gate_cur,
        src1_cur,
        dst_gate_cur,
        shmem,
        tgpig,
        tiisg,
        sgitg);
    kernel_mul_mv_q4_K_f32_impl<N_R0_Q4_K>(
        args0,
        src0_up_cur,
        src1_cur,
        dst_up_cur,
        shmem,
        tgpig,
        tiisg,
        sgitg);

    const short NSG = FC_mul_mv_nsg;
    const int first_row = (tgpig.x * NSG + sgitg) * N_R0_Q4_K;
    device float *gate_f32 = (device float *)dst_gate_cur;
    device float *up_f32 = (device float *)dst_up_cur;
    const uint64_t pair_row = (uint64_t)i12 * (uint64_t)args.nei0 + (uint64_t)idx;
    device float *mid_f32 = (device float *)(dst_mid + pair_row * act.mid_row_stride);
    device const float *route_w = (device const float *)(weights + pair_row * act.weight_stride);
    const float c = act.clamp_value;
    const float route_weight = route_w[0];

    if (tiisg == 0) {
        for (int row = 0; row < N_R0_Q4_K && first_row + row < args.ne0; ++row) {
            const uint out_row = first_row + row;
            float g = gate_f32[out_row];
            float u = up_f32[out_row];
            if (c > 1.0e-6f) {
                g = min(g, c);
                u = clamp(u, -c, c);
            }
            const float silu = g / (1.0f + exp(-g));
            mid_f32[out_row] = silu * u * route_weight;
        }
    }

    (void)tiitg;
}

kernel void kernel_q4_gather_slots6(
        constant ds4_metal_q4_gather_slots6_args &args,
        device const char *src_group0,
        device const char *src_group1,
        device const char *src_group2,
        device const char *src_group3,
        device const char *src_group4,
        device const char *src_group5,
        device const int32_t *ids,
        device char *dst,
        uint3 tgpig [[threadgroup_position_in_grid]],
        uint tiitg [[thread_index_in_threadgroup]]) {
    const uint slot = tgpig.y;
    if (slot >= args.n_slots || args.group_size == 0) return;

    const int32_t expert = ids[slot];
    if (expert < 0) return;

    const uint expert_u = (uint)expert;
    const uint group = expert_u / args.group_size;
    if (group >= 6) return;

    const uint local_expert = expert_u - group * args.group_size;
    device const char *src_group = src_group0;
    switch (group) {
    case 1: src_group = src_group1; break;
    case 2: src_group = src_group2; break;
    case 3: src_group = src_group3; break;
    case 4: src_group = src_group4; break;
    case 5: src_group = src_group5; break;
    default: break;
    }

    const uint64_t chunk = (uint64_t)tgpig.x * 256ul + (uint64_t)tiitg;
    const uint64_t n_chunks = args.expert_bytes >> 4;
    if (chunk >= n_chunks) return;

    device const uint4 *src =
        (device const uint4 *)(src_group + (uint64_t)local_expert * args.expert_bytes);
    device uint4 *out =
        (device uint4 *)(dst + (uint64_t)slot * args.expert_bytes);
    out[chunk] = src[chunk];
}

kernel void kernel_mul_mv_slots6_q4_K_pair_swiglu_f32(
        constant ds4_metal_args_mul_mv_id & args,
        constant ds4_metal_dsv4_moe_swiglu_weight_args & act,
        device const char * src0_gate0,
        device const char * src0_gate1,
        device const char * src0_gate2,
        device const char * src0_gate3,
        device const char * src0_gate4,
        device const char * src0_gate5,
        device const char * src0_up0,
        device const char * src0_up1,
        device const char * src0_up2,
        device const char * src0_up3,
        device const char * src0_up4,
        device const char * src0_up5,
        device const char * src1,
        device       char * dst_gate,
        device       char * dst_up,
        device       char * dst_mid,
        device const char * weights,
        threadgroup  char * shmem [[threadgroup(0)]],
        uint3  tgpig[[threadgroup_position_in_grid]],
        ushort tiitg[[thread_index_in_threadgroup]],
        ushort tiisg[[thread_index_in_simdgroup]],
        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
    const int iid1 = tgpig.z / args.nei0;
    const int idx  = tgpig.z % args.nei0;

    tgpig.z = 0;

    const int64_t i11 = idx % args.ne11;
    const int64_t i12 = iid1;

    device const char *src0_gate_cur = src0_gate0;
    device const char *src0_up_cur = src0_up0;
    switch (idx) {
    case 1: src0_gate_cur = src0_gate1; src0_up_cur = src0_up1; break;
    case 2: src0_gate_cur = src0_gate2; src0_up_cur = src0_up2; break;
    case 3: src0_gate_cur = src0_gate3; src0_up_cur = src0_up3; break;
    case 4: src0_gate_cur = src0_gate4; src0_up_cur = src0_up4; break;
    case 5: src0_gate_cur = src0_gate5; src0_up_cur = src0_up5; break;
    default: break;
    }

    device const char *src1_cur = src1 + i11 * args.nb11 + i12 * args.nb12;

    device char *dst_gate_cur = dst_gate + (idx * args.ne0 + i12 * args.ne1 * args.ne0) * sizeof(float);
    device char *dst_up_cur   = dst_up   + (idx * args.ne0 + i12 * args.ne1 * args.ne0) * sizeof(float);

    ds4_metal_args_mul_mv args0 = {
        args.ne00, args.ne01, 1,
        args.nb00, args.nb01, args.nb02, args.nb02,
        args.ne10, 1, 1,
        args.nb10, args.nb11, args.nb12, args.nb12,
        args.ne0, 1, args.nr0, 1, 1,
    };

    kernel_mul_mv_q4_K_f32_impl<N_R0_Q4_K>(
        args0,
        src0_gate_cur,
        src1_cur,
        dst_gate_cur,
        shmem,
        tgpig,
        tiisg,
        sgitg);
    kernel_mul_mv_q4_K_f32_impl<N_R0_Q4_K>(
        args0,
        src0_up_cur,
        src1_cur,
        dst_up_cur,
        shmem,
        tgpig,
        tiisg,
        sgitg);

    const short NSG = FC_mul_mv_nsg;
    const int first_row = (tgpig.x * NSG + sgitg) * N_R0_Q4_K;
    device float *gate_f32 = (device float *)dst_gate_cur;
    device float *up_f32 = (device float *)dst_up_cur;
    const uint64_t pair_row = (uint64_t)i12 * (uint64_t)args.nei0 + (uint64_t)idx;
    device float *mid_f32 = (device float *)(dst_mid + pair_row * act.mid_row_stride);
    device const float *route_w = (device const float *)(weights + pair_row * act.weight_stride);
    const float c = act.clamp_value;
    const float route_weight = route_w[0];

    if (tiisg == 0) {
        for (int row = 0; row < N_R0_Q4_K && first_row + row < args.ne0; ++row) {
            const uint out_row = first_row + row;
            float g = gate_f32[out_row];
            float u = up_f32[out_row];
            if (c > 1.0e-6f) {
                g = min(g, c);
                u = clamp(u, -c, c);
            }
            const float silu = g / (1.0f + exp(-g));
            mid_f32[out_row] = silu * u * route_weight;
        }
    }

    (void)tiitg;
}

static inline device const char *ds4_q4_group24_select(
        uint32_t group_id,
        device const char *src00,
        device const char *src01,
        device const char *src02,
        device const char *src03,
        device const char *src04,
        device const char *src05,
        device const char *src06,
        device const char *src07,
        device const char *src08,
        device const char *src09,
        device const char *src10,
        device const char *src11,
        device const char *src12,
        device const char *src13,
        device const char *src14,
        device const char *src15,
        device const char *src16,
        device const char *src17,
        device const char *src18,
        device const char *src19,
        device const char *src20,
        device const char *src21,
        device const char *src22,
        device const char *src23) {
    switch (group_id) {
    case 1:  return src01;
    case 2:  return src02;
    case 3:  return src03;
    case 4:  return src04;
    case 5:  return src05;
    case 6:  return src06;
    case 7:  return src07;
    case 8:  return src08;
    case 9:  return src09;
    case 10: return src10;
    case 11: return src11;
    case 12: return src12;
    case 13: return src13;
    case 14: return src14;
    case 15: return src15;
    case 16: return src16;
    case 17: return src17;
    case 18: return src18;
    case 19: return src19;
    case 20: return src20;
    case 21: return src21;
    case 22: return src22;
    case 23: return src23;
    default: return src00;
    }
}

kernel void kernel_mul_mv_group6_q4_K_pair_swiglu_f32(
        constant ds4_metal_args_mul_mv_id & args,
        constant ds4_metal_dsv4_moe_swiglu_weight_args & act,
        device const char * src0_gate0,
        device const char * src0_gate1,
        device const char * src0_gate2,
        device const char * src0_gate3,
        device const char * src0_gate4,
        device const char * src0_gate5,
        device const char * src0_up0,
        device const char * src0_up1,
        device const char * src0_up2,
        device const char * src0_up3,
        device const char * src0_up4,
        device const char * src0_up5,
        device const char * src1,
        device       char * dst_gate,
        device       char * dst_up,
        device       char * dst_mid,
        device const char * ids,
        device const char * weights,
        threadgroup  char * shmem [[threadgroup(0)]],
        uint3  tgpig[[threadgroup_position_in_grid]],
        ushort tiitg[[thread_index_in_threadgroup]],
        ushort tiisg[[thread_index_in_simdgroup]],
        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
    constexpr uint32_t expert_group_size = 64;
    const int iid1 = tgpig.z / args.nei0;
    const int idx  = tgpig.z % args.nei0;

    tgpig.z = 0;

    const int32_t expert = ((device const int32_t *)(ids + iid1 * args.nbi1))[idx];
    if (expert < 0) {
        return;
    }
    const uint32_t expert_u = (uint32_t)expert;
    const uint32_t group_id = expert_u / expert_group_size;
    if (group_id >= 6) {
        return;
    }
    const uint32_t expert_local = expert_u - group_id * expert_group_size;

    device const char *src0_gate_cur = src0_gate0;
    device const char *src0_up_cur = src0_up0;
    switch (group_id) {
    case 1: src0_gate_cur = src0_gate1; src0_up_cur = src0_up1; break;
    case 2: src0_gate_cur = src0_gate2; src0_up_cur = src0_up2; break;
    case 3: src0_gate_cur = src0_gate3; src0_up_cur = src0_up3; break;
    case 4: src0_gate_cur = src0_gate4; src0_up_cur = src0_up4; break;
    case 5: src0_gate_cur = src0_gate5; src0_up_cur = src0_up5; break;
    default: break;
    }

    const int64_t i11 = idx % args.ne11;
    const int64_t i12 = iid1;

    src0_gate_cur += (uint64_t)expert_local * args.nb02;
    src0_up_cur   += (uint64_t)expert_local * args.nb02;
    device const char *src1_cur = src1 + i11 * args.nb11 + i12 * args.nb12;

    device char *dst_gate_cur = dst_gate + (idx * args.ne0 + i12 * args.ne1 * args.ne0) * sizeof(float);
    device char *dst_up_cur   = dst_up   + (idx * args.ne0 + i12 * args.ne1 * args.ne0) * sizeof(float);

    ds4_metal_args_mul_mv args0 = {
        args.ne00, args.ne01, 1,
        args.nb00, args.nb01, args.nb02, args.nb02,
        args.ne10, 1, 1,
        args.nb10, args.nb11, args.nb12, args.nb12,
        args.ne0, 1, args.nr0, 1, 1,
    };

    kernel_mul_mv_q4_K_f32_impl<N_R0_Q4_K>(
        args0,
        src0_gate_cur,
        src1_cur,
        dst_gate_cur,
        shmem,
        tgpig,
        tiisg,
        sgitg);
    kernel_mul_mv_q4_K_f32_impl<N_R0_Q4_K>(
        args0,
        src0_up_cur,
        src1_cur,
        dst_up_cur,
        shmem,
        tgpig,
        tiisg,
        sgitg);

    const short NSG = FC_mul_mv_nsg;
    const int first_row = (tgpig.x * NSG + sgitg) * N_R0_Q4_K;
    device float *gate_f32 = (device float *)dst_gate_cur;
    device float *up_f32 = (device float *)dst_up_cur;
    const uint64_t pair_row = (uint64_t)i12 * (uint64_t)args.nei0 + (uint64_t)idx;
    device float *mid_f32 = (device float *)(dst_mid + pair_row * act.mid_row_stride);
    device const float *route_w = (device const float *)(weights + pair_row * act.weight_stride);
    const float c = act.clamp_value;
    const float route_weight = route_w[0];

    if (tiisg == 0) {
        for (int row = 0; row < N_R0_Q4_K && first_row + row < args.ne0; ++row) {
            const uint out_row = first_row + row;
            float g = gate_f32[out_row];
            float u = up_f32[out_row];
            if (c > 1.0e-6f) {
                g = min(g, c);
                u = clamp(u, -c, c);
            }
            const float silu = g / (1.0f + exp(-g));
            mid_f32[out_row] = silu * u * route_weight;
        }
    }

    (void)tiitg;
}

kernel void kernel_mul_mv_group8_q4_K_pair_swiglu_f32(
        constant ds4_metal_args_mul_mv_id & args,
        constant ds4_metal_dsv4_moe_swiglu_weight_args & act,
        device const char * src0_gate0,
        device const char * src0_gate1,
        device const char * src0_gate2,
        device const char * src0_gate3,
        device const char * src0_gate4,
        device const char * src0_gate5,
        device const char * src0_gate6,
        device const char * src0_gate7,
        device const char * src0_up0,
        device const char * src0_up1,
        device const char * src0_up2,
        device const char * src0_up3,
        device const char * src0_up4,
        device const char * src0_up5,
        device const char * src0_up6,
        device const char * src0_up7,
        device const char * src1,
        device       char * dst_gate,
        device       char * dst_up,
        device       char * dst_mid,
        device const char * ids,
        device const char * weights,
        threadgroup  char * shmem [[threadgroup(0)]],
        uint3  tgpig[[threadgroup_position_in_grid]],
        ushort tiitg[[thread_index_in_threadgroup]],
        ushort tiisg[[thread_index_in_simdgroup]],
        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
    constexpr uint32_t expert_group_size = 48;
    const int iid1 = tgpig.z / args.nei0;
    const int idx  = tgpig.z % args.nei0;

    tgpig.z = 0;

    const int32_t expert = ((device const int32_t *)(ids + iid1 * args.nbi1))[idx];
    if (expert < 0) {
        return;
    }
    const uint32_t expert_u = (uint32_t)expert;
    const uint32_t group_id = expert_u / expert_group_size;
    if (group_id >= 8) {
        return;
    }
    const uint32_t expert_local = expert_u - group_id * expert_group_size;

    device const char *src0_gate_cur = src0_gate0;
    device const char *src0_up_cur = src0_up0;
    switch (group_id) {
    case 1: src0_gate_cur = src0_gate1; src0_up_cur = src0_up1; break;
    case 2: src0_gate_cur = src0_gate2; src0_up_cur = src0_up2; break;
    case 3: src0_gate_cur = src0_gate3; src0_up_cur = src0_up3; break;
    case 4: src0_gate_cur = src0_gate4; src0_up_cur = src0_up4; break;
    case 5: src0_gate_cur = src0_gate5; src0_up_cur = src0_up5; break;
    case 6: src0_gate_cur = src0_gate6; src0_up_cur = src0_up6; break;
    case 7: src0_gate_cur = src0_gate7; src0_up_cur = src0_up7; break;
    default: break;
    }

    const int64_t i11 = idx % args.ne11;
    const int64_t i12 = iid1;

    src0_gate_cur += (uint64_t)expert_local * args.nb02;
    src0_up_cur   += (uint64_t)expert_local * args.nb02;
    device const char *src1_cur = src1 + i11 * args.nb11 + i12 * args.nb12;

    device char *dst_gate_cur = dst_gate + (idx * args.ne0 + i12 * args.ne1 * args.ne0) * sizeof(float);
    device char *dst_up_cur   = dst_up   + (idx * args.ne0 + i12 * args.ne1 * args.ne0) * sizeof(float);

    ds4_metal_args_mul_mv args0 = {
        args.ne00, args.ne01, 1,
        args.nb00, args.nb01, args.nb02, args.nb02,
        args.ne10, 1, 1,
        args.nb10, args.nb11, args.nb12, args.nb12,
        args.ne0, 1, args.nr0, 1, 1,
    };

    kernel_mul_mv_q4_K_f32_impl<N_R0_Q4_K>(
        args0,
        src0_gate_cur,
        src1_cur,
        dst_gate_cur,
        shmem,
        tgpig,
        tiisg,
        sgitg);
    kernel_mul_mv_q4_K_f32_impl<N_R0_Q4_K>(
        args0,
        src0_up_cur,
        src1_cur,
        dst_up_cur,
        shmem,
        tgpig,
        tiisg,
        sgitg);

    const short NSG = FC_mul_mv_nsg;
    const int first_row = (tgpig.x * NSG + sgitg) * N_R0_Q4_K;
    device float *gate_f32 = (device float *)dst_gate_cur;
    device float *up_f32 = (device float *)dst_up_cur;
    const uint64_t pair_row = (uint64_t)i12 * (uint64_t)args.nei0 + (uint64_t)idx;
    device float *mid_f32 = (device float *)(dst_mid + pair_row * act.mid_row_stride);
    device const float *route_w = (device const float *)(weights + pair_row * act.weight_stride);
    const float c = act.clamp_value;
    const float route_weight = route_w[0];

    if (tiisg == 0) {
        for (int row = 0; row < N_R0_Q4_K && first_row + row < args.ne0; ++row) {
            const uint out_row = first_row + row;
            float g = gate_f32[out_row];
            float u = up_f32[out_row];
            if (c > 1.0e-6f) {
                g = min(g, c);
                u = clamp(u, -c, c);
            }
            const float silu = g / (1.0f + exp(-g));
            mid_f32[out_row] = silu * u * route_weight;
        }
    }

    (void)tiitg;
}

kernel void kernel_mul_mv_group24_q4_K_id_f32(
        constant ds4_metal_args_mul_mv_id & args,
        device const char * src00,
        device const char * src01,
        device const char * src02,
        device const char * src03,
        device const char * src04,
        device const char * src05,
        device const char * src06,
        device const char * src07,
        device const char * src08,
        device const char * src09,
        device const char * src10,
        device const char * src11,
        device const char * src12,
        device const char * src13,
        device const char * src14,
        device const char * src15,
        device const char * src16,
        device const char * src17,
        device const char * src18,
        device const char * src19,
        device const char * src20,
        device const char * src21,
        device const char * src22,
        device const char * src23,
        device const char * src1,
        device       char * dst,
        device const char * ids,
        threadgroup  char * shmem [[threadgroup(0)]],
        uint3  tgpig[[threadgroup_position_in_grid]],
        ushort tiitg[[thread_index_in_threadgroup]],
        ushort tiisg[[thread_index_in_simdgroup]],
        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
    constexpr uint32_t expert_group_size = 16;
    const int iid1 = tgpig.z / args.nei0;
    const int idx  = tgpig.z % args.nei0;

    tgpig.z = 0;

    const int32_t expert = ((device const int32_t *)(ids + iid1 * args.nbi1))[idx];
    if (expert < 0) {
        return;
    }
    const uint32_t expert_u = (uint32_t)expert;
    const uint32_t group_id = expert_u / expert_group_size;
    if (group_id >= 24) {
        return;
    }
    const uint32_t expert_local = expert_u - group_id * expert_group_size;

    device const char *src0_cur = ds4_q4_group24_select(group_id,
                                                        src00, src01, src02, src03,
                                                        src04, src05, src06, src07,
                                                        src08, src09, src10, src11,
                                                        src12, src13, src14, src15,
                                                        src16, src17, src18, src19,
                                                        src20, src21, src22, src23);
    const int64_t i11 = idx % args.ne11;
    const int64_t i12 = iid1;

    src0_cur += (uint64_t)expert_local * args.nb02;
    device const char *src1_cur = src1 + i11 * args.nb11 + i12 * args.nb12;
    device char *dst_cur = dst + (idx * args.ne0 + i12 * args.ne1 * args.ne0) * sizeof(float);

    ds4_metal_args_mul_mv args0 = {
        args.ne00, args.ne01, 1,
        args.nb00, args.nb01, args.nb02, args.nb02,
        args.ne10, 1, 1,
        args.nb10, args.nb11, args.nb12, args.nb12,
        args.ne0, 1, args.nr0, 1, 1,
    };

    kernel_mul_mv_q4_K_f32_impl<N_R0_Q4_K>(
        args0,
        src0_cur,
        src1_cur,
        dst_cur,
        shmem,
        tgpig,
        tiisg,
        sgitg);

    (void)tiitg;
}

kernel void kernel_mul_mv_group_q4_K_pair_swiglu_f32(
        constant ds4_metal_args_mul_mv_id & args,
        constant ds4_metal_dsv4_moe_swiglu_weight_args & act,
        constant ds4_metal_moe_expert_group_args & group,
        device const char * src0_gate,
        device const char * src0_up,
        device const char * src1,
        device       char * dst_gate,
        device       char * dst_up,
        device       char * dst_mid,
        device const char * ids,
        device const char * weights,
        threadgroup  char * shmem [[threadgroup(0)]],
        uint3  tgpig[[threadgroup_position_in_grid]],
        ushort tiitg[[thread_index_in_threadgroup]],
        ushort tiisg[[thread_index_in_simdgroup]],
        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
    const int iid1 = tgpig.z / args.nei0;
    const int idx  = tgpig.z % args.nei0;

    tgpig.z = 0;

    const int32_t expert_global = ((device const int32_t *)(ids + iid1 * args.nbi1))[idx];
    if (expert_global < 0) {
        return;
    }
    const uint32_t expert_u = (uint32_t)expert_global;
    if (expert_u < group.expert_base ||
        expert_u >= group.expert_base + group.expert_count) {
        return;
    }
    const uint32_t expert_local = expert_u - group.expert_base;

    const int64_t i11 = idx % args.ne11;
    const int64_t i12 = iid1;

    device const char *src0_gate_cur = src0_gate + (uint64_t)expert_local * args.nb02;
    device const char *src0_up_cur   = src0_up   + (uint64_t)expert_local * args.nb02;
    device const char *src1_cur      = src1      + i11 * args.nb11 + i12 * args.nb12;

    device char *dst_gate_cur = dst_gate + (idx * args.ne0 + i12 * args.ne1 * args.ne0) * sizeof(float);
    device char *dst_up_cur   = dst_up   + (idx * args.ne0 + i12 * args.ne1 * args.ne0) * sizeof(float);

    ds4_metal_args_mul_mv args0 = {
        args.ne00, args.ne01, 1,
        args.nb00, args.nb01, args.nb02, args.nb02,
        args.ne10, 1, 1,
        args.nb10, args.nb11, args.nb12, args.nb12,
        args.ne0, 1, args.nr0, 1, 1,
    };

    kernel_mul_mv_q4_K_f32_impl<N_R0_Q4_K>(
        args0,
        src0_gate_cur,
        src1_cur,
        dst_gate_cur,
        shmem,
        tgpig,
        tiisg,
        sgitg);
    kernel_mul_mv_q4_K_f32_impl<N_R0_Q4_K>(
        args0,
        src0_up_cur,
        src1_cur,
        dst_up_cur,
        shmem,
        tgpig,
        tiisg,
        sgitg);

    const short NSG = FC_mul_mv_nsg;
    const int first_row = (tgpig.x * NSG + sgitg) * N_R0_Q4_K;
    device float *gate_f32 = (device float *)dst_gate_cur;
    device float *up_f32 = (device float *)dst_up_cur;
    const uint64_t pair_row = (uint64_t)i12 * (uint64_t)args.nei0 + (uint64_t)idx;
    device float *mid_f32 = (device float *)(dst_mid + pair_row * act.mid_row_stride);
    device const float *route_w = (device const float *)(weights + pair_row * act.weight_stride);
    const float c = act.clamp_value;
    const float route_weight = route_w[0];

    if (tiisg == 0) {
        for (int row = 0; row < N_R0_Q4_K && first_row + row < args.ne0; ++row) {
            const uint out_row = first_row + row;
            float g = gate_f32[out_row];
            float u = up_f32[out_row];
            if (c > 1.0e-6f) {
                g = min(g, c);
                u = clamp(u, -c, c);
            }
            const float silu = g / (1.0f + exp(-g));
            mid_f32[out_row] = silu * u * route_weight;
        }
    }

    (void)tiitg;
}


/* IQ2_XXS down projection summed over the token's selected experts (the
 * GLM 5.2 routed blobs are IQ2_XXS end to end, unlike DS4 Flash's Q2_K
 * down).  Same contract as the q2_K sum kernel: mid rows already carry
 * silu(gate)*up*route_weight, expert ownership honors the TP split, and
 * add_in folds the shared-expert partial when tp_addend is set. */
kernel void kernel_mul_mv_id_iq2_xxs_sum6_f32(
        constant ds4_metal_args_mul_mv_id & args,
        device const char * src0s,
        device const char * src1,
        device       char * dst,
        device const char * ids,
        device const char * add_in,
        threadgroup  char * shmem [[threadgroup(0)]],
        uint3  tgpig[[threadgroup_position_in_grid]],
        ushort tiitg[[thread_index_in_threadgroup]],
        ushort tiisg[[thread_index_in_simdgroup]],
        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
    const short NSG = FC_mul_mv_nsg;
    const short nr0 = N_R0_IQ2_XXS;
    const int nb = args.ne00/QK_K;
    const int nb32 = nb * (QK_K / 32);
    const int first_row = (tgpig.x * NSG + sgitg) * nr0;
    const uint token = tgpig.y;
    device const int32_t *token_ids = (device const int32_t *)(ids + (uint64_t)token * args.nbi1);
    device const char *token_src1 = src1 + (uint64_t)token * args.nb12;

    float sumf[nr0] = {0.f};
    float yl[32];

    threadgroup uint64_t *svalues = (threadgroup uint64_t *)(shmem);
    threadgroup uint8_t  *ssigns  = (threadgroup uint8_t *)(svalues + 256);
    {
        int nval = 4;
        int pos = (32 * sgitg + tiisg) * nval;
        for (int i = 0; i < nval; ++i) svalues[pos + i] = ds4_metal_iq2xxs_grid[pos + i];
        nval = 2;
        pos = (32 * sgitg + tiisg) * nval;
        for (int i = 0; i < nval; ++i) ssigns[pos + i] = ds4_metal_ksigns_iq2xs[pos + i];
        threadgroup_barrier(mem_flags::mem_threadgroup);
    }

    const int ix = tiisg;

    for (int expert_slot = 0; expert_slot < args.nei0; expert_slot++) {
        const int32_t expert = token_ids[expert_slot];
        if (!ds4_tp_owns_expert(expert, args.ne02, args.tp_rank, args.tp_world)) continue;
        device const block_iq2_xxs *x =
            (device const block_iq2_xxs *)(src0s + (int64_t)(expert - args.tp_expert_base)*args.nb02 + first_row*args.nb01);
        device const float *y = (device const float *)(token_src1 + expert_slot*args.nb11);
        device const float *y4 = y + 32 * ix;

        for (int ib32 = ix; ib32 < nb32; ib32 += 32) {
            for (short i = 0; i < 32; ++i) {
                yl[i] = y4[i];
            }

            const int ibl = ib32 / (QK_K / 32);
            const int ib  = ib32 % (QK_K / 32);

            device const block_iq2_xxs *xr = x + ibl;
            device const uint16_t *q2 = xr->qs + 4 * ib;
            device const half *dh = &xr->d;

            for (short row = 0; row < nr0; row++) {
                if (first_row + row < args.ne0) {
                    device const uint8_t *aux8 = (device const uint8_t *)q2;
                    const uint32_t aux32 = q2[2] | (q2[3] << 16);
                    const float d = (float)dh[0] * (0.5f + (aux32 >> 28));

                    float s = 0;
                    for (short l = 0; l < 4; ++l) {
                        const threadgroup uint8_t *grid = (const threadgroup uint8_t *)(svalues + aux8[l]);
                        const uint8_t sign = ssigns[(aux32 >> 7 * l) & 127];
                        for (short j = 0; j < 8; ++j) {
                            s += yl[8 * l + j] * grid[j] *
                                 (sign & ds4_metal_kmask_iq2xs[j] ? -1.f : 1.f);
                        }
                    }
                    sumf[row] += d * s;
                }
                dh += args.nb01 / 2;
                q2 += args.nb01 / 2;
            }

            y4 += 32 * 32;
        }
    }

    device float * dst_f32 = (device float *)(dst + (uint64_t)token * args.nb1);
    for (int row = 0; row < nr0 && first_row + row < args.ne0; row++) {
        const float sum_all = simd_sum(sumf[row]) * 0.25f;
        if (tiisg == 0) {
            float outv = sum_all;
            if (args.tp_addend) {
                outv += ((device const float *) add_in)[first_row + row];
            }
            dst_f32[first_row + row] = outv;
        }
    }

    (void)shmem;
    (void)tiitg;
    (void)tgpig;
}

kernel void kernel_mul_mv_id_q2_K_sum6_f32(
        constant ds4_metal_args_mul_mv_id & args,
        device const char * src0s,
        device const char * src1,
        device       char * dst,
        device const char * ids,
        device const char * add_in,
        threadgroup  char * shmem [[threadgroup(0)]],
        uint3  tgpig[[threadgroup_position_in_grid]],
        ushort tiitg[[thread_index_in_threadgroup]],
        ushort tiisg[[thread_index_in_simdgroup]],
        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
    const short NSG = FC_mul_mv_nsg;
    const short nr0 = N_R0_Q2_K;
    const int nb = args.ne00/QK_K;
    const int first_row = (tgpig.x * NSG + sgitg) * nr0;
    const uint token = tgpig.y;
    device const int32_t *token_ids = (device const int32_t *)(ids + (uint64_t)token * args.nbi1);
    device const char *token_src1 = src1 + (uint64_t)token * args.nb12;

    float sumf[nr0] = {0.f};

    const short ix = tiisg/8;
    const short it = tiisg%8;
    const short iq = it/4;
    const short ir = it%4;
    const short is = (8*ir)/16;

    for (int expert_slot = 0; expert_slot < args.nei0; expert_slot++) {
        const int32_t expert = token_ids[expert_slot];
        if (!ds4_tp_owns_expert(expert, args.ne02, args.tp_rank, args.tp_world)) continue;
        device const block_q2_K * x = (device const block_q2_K *)(src0s + (int64_t)(expert - args.tp_expert_base)*args.nb02 + first_row*args.nb01);
        device const float * y = (device const float *)(token_src1 + expert_slot*args.nb11);
        device const float * y4 = y + ix * QK_K + 128 * iq + 8 * ir;

        for (int ib = ix; ib < nb; ib += 4) {
            float yl[32];
            float4 sumy = {0.f, 0.f, 0.f, 0.f};
            for (short i = 0; i < 8; ++i) {
                yl[i+ 0] = y4[i+ 0]; sumy[0] += yl[i+ 0];
                yl[i+ 8] = y4[i+32]; sumy[1] += yl[i+ 8];
                yl[i+16] = y4[i+64]; sumy[2] += yl[i+16];
                yl[i+24] = y4[i+96]; sumy[3] += yl[i+24];
            }

            device const uint8_t  * sc = (device const uint8_t  *)x[ib].scales + 8*iq + is;
            device const uint16_t * qs = (device const uint16_t *)x[ib].qs + 16 * iq + 4 * ir;
            device const half     * dh = &x[ib].d;

            for (short row = 0; row < nr0; row++) {
                if (first_row + row < args.ne0) {
                    float4 acc1 = {0.f, 0.f, 0.f, 0.f};
                    float4 acc2 = {0.f, 0.f, 0.f, 0.f};
                    for (int i = 0; i < 8; i += 2) {
                        acc1[0] += yl[i+ 0] * (qs[i/2] & 0x0003);
                        acc2[0] += yl[i+ 1] * (qs[i/2] & 0x0300);
                        acc1[1] += yl[i+ 8] * (qs[i/2] & 0x000c);
                        acc2[1] += yl[i+ 9] * (qs[i/2] & 0x0c00);
                        acc1[2] += yl[i+16] * (qs[i/2] & 0x0030);
                        acc2[2] += yl[i+17] * (qs[i/2] & 0x3000);
                        acc1[3] += yl[i+24] * (qs[i/2] & 0x00c0);
                        acc2[3] += yl[i+25] * (qs[i/2] & 0xc000);
                    }
                    float dall = dh[0];
                    float dmin = dh[1] * 1.f/16.f;
                    sumf[row] += dall * ((acc1[0] + 1.f/256.f * acc2[0]) * (sc[0] & 0xF) * 1.f/ 1.f +
                                         (acc1[1] + 1.f/256.f * acc2[1]) * (sc[2] & 0xF) * 1.f/ 4.f +
                                         (acc1[2] + 1.f/256.f * acc2[2]) * (sc[4] & 0xF) * 1.f/16.f +
                                         (acc1[3] + 1.f/256.f * acc2[3]) * (sc[6] & 0xF) * 1.f/64.f) -
                                 dmin * (sumy[0] * (sc[0] & 0xF0) + sumy[1] * (sc[2] & 0xF0) +
                                         sumy[2] * (sc[4] & 0xF0) + sumy[3] * (sc[6] & 0xF0));
                }

                qs += args.nb01/2;
                sc += args.nb01;
                dh += args.nb01/2;
            }

            y4 += 4 * QK_K;
        }
    }

    device float * dst_f32 = (device float *)(dst + (uint64_t)token * args.nb1);
    for (int row = 0; row < nr0 && first_row + row < args.ne0; row++) {
        const float sum_all = simd_sum(sumf[row]);
        if (tiisg == 0) {
            float outv = sum_all;
            if (args.tp_addend) {
                outv += ((device const float *) add_in)[first_row + row];
            }
            dst_f32[first_row + row] = outv;
        }
    }

    (void)shmem;
    (void)tiitg;
    (void)tgpig;
}

kernel void kernel_mul_mv_slots6_q2_K_sum6_f32(
        constant ds4_metal_args_mul_mv_id & args,
        device const char * src00,
        device const char * src01,
        device const char * src02,
        device const char * src03,
        device const char * src04,
        device const char * src05,
        device const char * src1,
        device       char * dst,
        threadgroup  char * shmem [[threadgroup(0)]],
        uint3  tgpig[[threadgroup_position_in_grid]],
        ushort tiitg[[thread_index_in_threadgroup]],
        ushort tiisg[[thread_index_in_simdgroup]],
        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
    const short NSG = FC_mul_mv_nsg;
    const short nr0 = N_R0_Q2_K;
    const int nb = args.ne00 / QK_K;
    const int first_row = (tgpig.x * NSG + sgitg) * nr0;
    const uint token = tgpig.y;
    device const char *token_src1 = src1 + (uint64_t)token * args.nb12;

    float sumf[nr0] = {0.f};

    const short ix = tiisg / 8;
    const short it = tiisg % 8;
    const short iq = it / 4;
    const short ir = it % 4;
    const short is = (8 * ir) / 16;

    for (int expert_slot = 0; expert_slot < 6; expert_slot++) {
        device const char *src0_cur = src00;
        switch (expert_slot) {
        case 1: src0_cur = src01; break;
        case 2: src0_cur = src02; break;
        case 3: src0_cur = src03; break;
        case 4: src0_cur = src04; break;
        case 5: src0_cur = src05; break;
        default: break;
        }
        device const block_q2_K *x =
            (device const block_q2_K *)(src0_cur + first_row * args.nb01);
        device const float *y = (device const float *)(token_src1 + expert_slot * args.nb11);
        device const float *y4 = y + ix * QK_K + 128 * iq + 8 * ir;

        for (int ib = ix; ib < nb; ib += 4) {
            float yl[32];
            float4 sumy = {0.f, 0.f, 0.f, 0.f};
            for (short i = 0; i < 8; ++i) {
                yl[i +  0] = y4[i +  0]; sumy[0] += yl[i +  0];
                yl[i +  8] = y4[i + 32]; sumy[1] += yl[i +  8];
                yl[i + 16] = y4[i + 64]; sumy[2] += yl[i + 16];
                yl[i + 24] = y4[i + 96]; sumy[3] += yl[i + 24];
            }

            device const uint8_t  *sc = (device const uint8_t *)x[ib].scales + 8 * iq + is;
            device const uint16_t *qs = (device const uint16_t *)x[ib].qs + 16 * iq + 4 * ir;
            device const half     *dh = &x[ib].d;

            for (short row = 0; row < nr0; row++) {
                if (first_row + row < args.ne0) {
                    float4 acc1 = {0.f, 0.f, 0.f, 0.f};
                    float4 acc2 = {0.f, 0.f, 0.f, 0.f};
                    for (int i = 0; i < 8; i += 2) {
                        acc1[0] += yl[i +  0] * (qs[i / 2] & 0x0003);
                        acc2[0] += yl[i +  1] * (qs[i / 2] & 0x0300);
                        acc1[1] += yl[i +  8] * (qs[i / 2] & 0x000c);
                        acc2[1] += yl[i +  9] * (qs[i / 2] & 0x0c00);
                        acc1[2] += yl[i + 16] * (qs[i / 2] & 0x0030);
                        acc2[2] += yl[i + 17] * (qs[i / 2] & 0x3000);
                        acc1[3] += yl[i + 24] * (qs[i / 2] & 0x00c0);
                        acc2[3] += yl[i + 25] * (qs[i / 2] & 0xc000);
                    }
                    float dall = dh[0];
                    float dmin = dh[1] * 1.f / 16.f;
                    sumf[row] += dall * ((acc1[0] + 1.f / 256.f * acc2[0]) * (sc[0] & 0xF) * 1.f /  1.f +
                                         (acc1[1] + 1.f / 256.f * acc2[1]) * (sc[2] & 0xF) * 1.f /  4.f +
                                         (acc1[2] + 1.f / 256.f * acc2[2]) * (sc[4] & 0xF) * 1.f / 16.f +
                                         (acc1[3] + 1.f / 256.f * acc2[3]) * (sc[6] & 0xF) * 1.f / 64.f) -
                                 dmin * (sumy[0] * (sc[0] & 0xF0) + sumy[1] * (sc[2] & 0xF0) +
                                         sumy[2] * (sc[4] & 0xF0) + sumy[3] * (sc[6] & 0xF0));
                }

                qs += args.nb01 / 2;
                sc += args.nb01;
                dh += args.nb01 / 2;
            }

            y4 += 4 * QK_K;
        }
    }

    device float *dst_f32 = (device float *)(dst + (uint64_t)token * args.nb1);
    for (int row = 0; row < nr0 && first_row + row < args.ne0; row++) {
        const float sum_all = simd_sum(sumf[row]);
        if (tiisg == 0) dst_f32[first_row + row] = sum_all;
    }

    (void)shmem;
    (void)tiitg;
    (void)tgpig;
}

kernel void kernel_mul_mv_addr_q2_K_sum6_f32(
        constant ds4_metal_args_mul_mv_id & args,
        device const uint64_t * addrs,
        device const char * src1,
        device       char * dst,
        device const char * ids,
        threadgroup  char * shmem [[threadgroup(0)]],
        uint3  tgpig[[threadgroup_position_in_grid]],
        ushort tiitg[[thread_index_in_threadgroup]],
        ushort tiisg[[thread_index_in_simdgroup]],
        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
    const short NSG = FC_mul_mv_nsg;
    const short nr0 = N_R0_Q2_K;
    const int nb = args.ne00 / QK_K;
    const int first_row = (tgpig.x * NSG + sgitg) * nr0;
    const uint token = tgpig.y;
    device const char *token_src1 = src1 + (uint64_t)token * args.nb12;
    device const int32_t *token_ids =
        (device const int32_t *)(ids + (uint64_t)token * args.nbi1);

    float sumf[nr0] = {0.f};

    const short ix = tiisg / 8;
    const short it = tiisg % 8;
    const short iq = it / 4;
    const short ir = it % 4;
    const short is = (8 * ir) / 16;

    for (int expert_slot = 0; expert_slot < 6; expert_slot++) {
        const int32_t expert = token_ids[expert_slot];
        if (expert < 0 || expert >= args.ne02 || expert >= 384) {
            continue;
        }
        const uint64_t addr = addrs[(uint)expert];
        if (addr == 0) {
            continue;
        }
        device const char *src0_cur =
            reinterpret_cast<device const char *>(addr);
        device const block_q2_K *x =
            (device const block_q2_K *)(src0_cur + first_row * args.nb01);
        device const float *y = (device const float *)(token_src1 + expert_slot * args.nb11);
        device const float *y4 = y + ix * QK_K + 128 * iq + 8 * ir;

        for (int ib = ix; ib < nb; ib += 4) {
            float yl[32];
            float4 sumy = {0.f, 0.f, 0.f, 0.f};
            for (short i = 0; i < 8; ++i) {
                yl[i +  0] = y4[i +  0]; sumy[0] += yl[i +  0];
                yl[i +  8] = y4[i + 32]; sumy[1] += yl[i +  8];
                yl[i + 16] = y4[i + 64]; sumy[2] += yl[i + 16];
                yl[i + 24] = y4[i + 96]; sumy[3] += yl[i + 24];
            }

            device const uint8_t  *sc = (device const uint8_t *)x[ib].scales + 8 * iq + is;
            device const uint16_t *qs = (device const uint16_t *)x[ib].qs + 16 * iq + 4 * ir;
            device const half     *dh = &x[ib].d;

            for (short row = 0; row < nr0; row++) {
                if (first_row + row < args.ne0) {
                    float4 acc1 = {0.f, 0.f, 0.f, 0.f};
                    float4 acc2 = {0.f, 0.f, 0.f, 0.f};
                    for (int i = 0; i < 8; i += 2) {
                        acc1[0] += yl[i +  0] * (qs[i / 2] & 0x0003);
                        acc2[0] += yl[i +  1] * (qs[i / 2] & 0x0300);
                        acc1[1] += yl[i +  8] * (qs[i / 2] & 0x000c);
                        acc2[1] += yl[i +  9] * (qs[i / 2] & 0x0c00);
                        acc1[2] += yl[i + 16] * (qs[i / 2] & 0x0030);
                        acc2[2] += yl[i + 17] * (qs[i / 2] & 0x3000);
                        acc1[3] += yl[i + 24] * (qs[i / 2] & 0x00c0);
                        acc2[3] += yl[i + 25] * (qs[i / 2] & 0xc000);
                    }
                    float dall = dh[0];
                    float dmin = dh[1] * 1.f / 16.f;
                    sumf[row] += dall * ((acc1[0] + 1.f / 256.f * acc2[0]) * (sc[0] & 0xF) * 1.f /  1.f +
                                         (acc1[1] + 1.f / 256.f * acc2[1]) * (sc[2] & 0xF) * 1.f /  4.f +
                                         (acc1[2] + 1.f / 256.f * acc2[2]) * (sc[4] & 0xF) * 1.f / 16.f +
                                         (acc1[3] + 1.f / 256.f * acc2[3]) * (sc[6] & 0xF) * 1.f / 64.f) -
                                 dmin * (sumy[0] * (sc[0] & 0xF0) + sumy[1] * (sc[2] & 0xF0) +
                                         sumy[2] * (sc[4] & 0xF0) + sumy[3] * (sc[6] & 0xF0));
                }
                qs += args.nb01 / 2;
                sc += args.nb01;
                dh += args.nb01 / 2;
            }
            y4 += 4 * QK_K;
        }
    }

    device float * dst_f32 = (device float *)(dst + (uint64_t)token * args.nb1);
    for (int row = 0; row < nr0 && first_row + row < args.ne0; row++) {
        const float sum_all = simd_sum(sumf[row]);
        if (tiisg == 0) dst_f32[first_row + row] = sum_all;
    }

    (void)shmem;
    (void)tiitg;
    (void)tgpig;
}

kernel void kernel_mul_mv_addr_q2_K_sum6_masked_f32(
        constant ds4_metal_args_mul_mv_id & args,
        constant ds4_metal_stream_expert_split_args & split,
        device const uint64_t * addrs,
        device const char * src1,
        device       char * dst,
        device const char * ids,
        threadgroup  char * shmem [[threadgroup(0)]],
        uint3  tgpig[[threadgroup_position_in_grid]],
        ushort tiitg[[thread_index_in_threadgroup]],
        ushort tiisg[[thread_index_in_simdgroup]],
        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
    const short NSG = FC_mul_mv_nsg;
    const short nr0 = N_R0_Q2_K;
    const int nb = args.ne00 / QK_K;
    const int first_row = (tgpig.x * NSG + sgitg) * nr0;
    const uint token = tgpig.y;
    device const char *token_src1 = src1 + (uint64_t)token * args.nb12;
    device const int32_t *token_ids =
        (device const int32_t *)(ids + (uint64_t)token * args.nbi1);

    float sumf[nr0] = {0.f};

    const short ix = tiisg / 8;
    const short it = tiisg % 8;
    const short iq = it / 4;
    const short ir = it % 4;
    const short is = (8 * ir) / 16;

    for (int expert_slot = 0; expert_slot < 6; expert_slot++) {
        if ((split.active_mask & (1u << (uint)expert_slot)) == 0) {
            continue;
        }
        const int32_t expert = token_ids[expert_slot];
        if (expert < 0 || expert >= args.ne02 || expert >= 384) {
            continue;
        }
        const uint64_t addr = addrs[(uint)expert];
        if (addr == 0) {
            continue;
        }
        device const char *src0_cur =
            reinterpret_cast<device const char *>(addr);
        device const block_q2_K *x =
            (device const block_q2_K *)(src0_cur + first_row * args.nb01);
        device const float *y = (device const float *)(token_src1 + expert_slot * args.nb11);
        device const float *y4 = y + ix * QK_K + 128 * iq + 8 * ir;

        for (int ib = ix; ib < nb; ib += 4) {
            float yl[32];
            float4 sumy = {0.f, 0.f, 0.f, 0.f};
            for (short i = 0; i < 8; ++i) {
                yl[i +  0] = y4[i +  0]; sumy[0] += yl[i +  0];
                yl[i +  8] = y4[i + 32]; sumy[1] += yl[i +  8];
                yl[i + 16] = y4[i + 64]; sumy[2] += yl[i + 16];
                yl[i + 24] = y4[i + 96]; sumy[3] += yl[i + 24];
            }

            device const uint8_t  *sc = (device const uint8_t *)x[ib].scales + 8 * iq + is;
            device const uint16_t *qs = (device const uint16_t *)x[ib].qs + 16 * iq + 4 * ir;
            device const half     *dh = &x[ib].d;

            for (short row = 0; row < nr0; row++) {
                if (first_row + row < args.ne0) {
                    float4 acc1 = {0.f, 0.f, 0.f, 0.f};
                    float4 acc2 = {0.f, 0.f, 0.f, 0.f};
                    for (int i = 0; i < 8; i += 2) {
                        acc1[0] += yl[i +  0] * (qs[i / 2] & 0x0003);
                        acc2[0] += yl[i +  1] * (qs[i / 2] & 0x0300);
                        acc1[1] += yl[i +  8] * (qs[i / 2] & 0x000c);
                        acc2[1] += yl[i +  9] * (qs[i / 2] & 0x0c00);
                        acc1[2] += yl[i + 16] * (qs[i / 2] & 0x0030);
                        acc2[2] += yl[i + 17] * (qs[i / 2] & 0x3000);
                        acc1[3] += yl[i + 24] * (qs[i / 2] & 0x00c0);
                        acc2[3] += yl[i + 25] * (qs[i / 2] & 0xc000);
                    }
                    float dall = dh[0];
                    float dmin = dh[1] * 1.f / 16.f;
                    sumf[row] += dall * ((acc1[0] + 1.f / 256.f * acc2[0]) * (sc[0] & 0xF) * 1.f /  1.f +
                                         (acc1[1] + 1.f / 256.f * acc2[1]) * (sc[2] & 0xF) * 1.f /  4.f +
                                         (acc1[2] + 1.f / 256.f * acc2[2]) * (sc[4] & 0xF) * 1.f / 16.f +
                                         (acc1[3] + 1.f / 256.f * acc2[3]) * (sc[6] & 0xF) * 1.f / 64.f) -
                                 dmin * (sumy[0] * (sc[0] & 0xF0) + sumy[1] * (sc[2] & 0xF0) +
                                         sumy[2] * (sc[4] & 0xF0) + sumy[3] * (sc[6] & 0xF0));
                }
                qs += args.nb01 / 2;
                sc += args.nb01;
                dh += args.nb01 / 2;
            }
            y4 += 4 * QK_K;
        }
    }

    device float * dst_f32 = (device float *)(dst + (uint64_t)token * args.nb1);
    for (int row = 0; row < nr0 && first_row + row < args.ne0; row++) {
        const float sum_all = simd_sum(sumf[row]);
        if (tiisg == 0) {
            if (split.accumulate) {
                dst_f32[first_row + row] += sum_all;
            } else {
                dst_f32[first_row + row] = sum_all;
            }
        }
    }

    (void)shmem;
    (void)tiitg;
    (void)tgpig;
}

kernel void kernel_mul_mv_id_q4_K_sum6_f32(
        constant ds4_metal_args_mul_mv_id & args,
        device const char * src0s,
        device const char * src1,
        device       char * dst,
        device const char * ids,
        device const char * add_in,
        threadgroup  char * shmem [[threadgroup(0)]],
        uint3  tgpig[[threadgroup_position_in_grid]],
        ushort tiitg[[thread_index_in_threadgroup]],
        ushort tiisg[[thread_index_in_simdgroup]],
        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
    const short NSG = FC_mul_mv_nsg;
    const short nr0 = N_R0_Q4_K;
    const int nb = args.ne00 / QK_K;
    const int first_row = (tgpig.x * NSG + sgitg) * nr0;
    const uint token = tgpig.y;
    device const int32_t *token_ids = (device const int32_t *)(ids + (uint64_t)token * args.nbi1);
    device const char *token_src1 = src1 + (uint64_t)token * args.nb12;

    constexpr uint16_t kmask1 = 0x3f3f;
    constexpr uint16_t kmask2 = 0x0f0f;
    constexpr uint16_t kmask3 = 0xc0c0;

    const short ix = tiisg / 8;
    const short it = tiisg % 8;
    const short iq = it / 4;
    const short ir = it % 4;

    float sumf[nr0] = {0.f};
    uint16_t sc16[4];
    thread const uint8_t *sc8 = (thread const uint8_t *)sc16;

    for (int expert_slot = 0; expert_slot < args.nei0; expert_slot++) {
        const int32_t expert = token_ids[expert_slot];
        if (!ds4_tp_owns_expert(expert, args.ne02, args.tp_rank, args.tp_world)) continue;
        device const block_q4_K *x =
            (device const block_q4_K *)(src0s + (int64_t)(expert - args.tp_expert_base) * args.nb02 + first_row * args.nb01);
        device const float *y = (device const float *)(token_src1 + expert_slot * args.nb11);
        device const float *y4 = y + ix * QK_K + 64 * iq + 8 * ir;

        for (int ib = ix; ib < nb; ib += 4) {
            float yl[16];
            float yh[16];
            float4 sumy = {0.f, 0.f, 0.f, 0.f};

            for (short i = 0; i < 8; ++i) {
                yl[i + 0] = y4[i +   0]; sumy[0] += yl[i + 0];
                yl[i + 8] = y4[i +  32]; sumy[1] += yl[i + 8];
                yh[i + 0] = y4[i + 128]; sumy[2] += yh[i + 0];
                yh[i + 8] = y4[i + 160]; sumy[3] += yh[i + 8];
            }

            device const uint16_t *sc = (device const uint16_t *)x[ib].scales + iq;
            device const uint16_t *q1 = (device const uint16_t *)x[ib].qs + 16 * iq + 4 * ir;
            device const half *dh = &x[ib].d;

            for (short row = 0; row < nr0; row++) {
                if (first_row + row < args.ne0) {
                    sc16[0] = sc[0] & kmask1;
                    sc16[1] = sc[2] & kmask1;
                    sc16[2] = ((sc[4] >> 0) & kmask2) | ((sc[0] & kmask3) >> 2);
                    sc16[3] = ((sc[4] >> 4) & kmask2) | ((sc[2] & kmask3) >> 2);

                    device const uint16_t *q2 = q1 + 32;

                    float4 acc1 = {0.f, 0.f, 0.f, 0.f};
                    float4 acc2 = {0.f, 0.f, 0.f, 0.f};

                    FOR_UNROLL (short i = 0; i < 4; ++i) {
                        acc1[0] += yl[2 * i + 0] * (q1[i] & 0x000F);
                        acc1[1] += yl[2 * i + 1] * (q1[i] & 0x0F00);
                        acc1[2] += yl[2 * i + 8] * (q1[i] & 0x00F0);
                        acc1[3] += yl[2 * i + 9] * (q1[i] & 0xF000);
                        acc2[0] += yh[2 * i + 0] * (q2[i] & 0x000F);
                        acc2[1] += yh[2 * i + 1] * (q2[i] & 0x0F00);
                        acc2[2] += yh[2 * i + 8] * (q2[i] & 0x00F0);
                        acc2[3] += yh[2 * i + 9] * (q2[i] & 0xF000);
                    }

                    sumf[row] += dh[0] * ((acc1[0] + 1.f / 256.f * acc1[1]) * sc8[0] +
                                          (acc1[2] + 1.f / 256.f * acc1[3]) * sc8[1] * 1.f / 16.f +
                                          (acc2[0] + 1.f / 256.f * acc2[1]) * sc8[4] +
                                          (acc2[2] + 1.f / 256.f * acc2[3]) * sc8[5] * 1.f / 16.f) -
                                 dh[1] * (sumy[0] * sc8[2] + sumy[1] * sc8[3] +
                                          sumy[2] * sc8[6] + sumy[3] * sc8[7]);
                }

                q1 += args.nb01 / 2;
                sc += args.nb01 / 2;
                dh += args.nb01 / 2;
            }

            y4 += 4 * QK_K;
        }
    }

    device float *dst_f32 = (device float *)(dst + (uint64_t)token * args.nb1);
    for (int row = 0; row < nr0 && first_row + row < args.ne0; row++) {
        const float sum_all = simd_sum(sumf[row]);
        if (tiisg == 0) {
            float outv = sum_all;
            if (args.tp_addend) {
                outv += ((device const float *) add_in)[first_row + row];
            }
            dst_f32[first_row + row] = outv;
        }
    }

    (void)shmem;
    (void)tiitg;
    (void)tgpig;
}

kernel void kernel_mul_mv_group_q4_K_sum6_f32(
        constant ds4_metal_args_mul_mv_id & args,
        constant ds4_metal_moe_expert_group_args & group,
        device const char * src0s,
        device const char * src1,
        device       char * dst,
        device const char * ids,
        threadgroup  char * shmem [[threadgroup(0)]],
        uint3  tgpig[[threadgroup_position_in_grid]],
        ushort tiitg[[thread_index_in_threadgroup]],
        ushort tiisg[[thread_index_in_simdgroup]],
        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
    const short NSG = FC_mul_mv_nsg;
    const short nr0 = N_R0_Q4_K;
    const int nb = args.ne00 / QK_K;
    const int first_row = (tgpig.x * NSG + sgitg) * nr0;
    const uint token = tgpig.y;
    device const int32_t *token_ids = (device const int32_t *)(ids + (uint64_t)token * args.nbi1);
    device const char *token_src1 = src1 + (uint64_t)token * args.nb12;

    constexpr uint16_t kmask1 = 0x3f3f;
    constexpr uint16_t kmask2 = 0x0f0f;
    constexpr uint16_t kmask3 = 0xc0c0;

    const short ix = tiisg / 8;
    const short it = tiisg % 8;
    const short iq = it / 4;
    const short ir = it % 4;

    float sumf[nr0] = {0.f};
    uint16_t sc16[4];
    thread const uint8_t *sc8 = (thread const uint8_t *)sc16;

    for (int expert_slot = 0; expert_slot < 6; expert_slot++) {
        const int32_t expert = token_ids[expert_slot];
        if (expert < 0) {
            continue;
        }
        const uint32_t expert_u = (uint32_t)expert;
        if (expert_u < group.expert_base ||
            expert_u >= group.expert_base + group.expert_count) {
            continue;
        }
        const uint32_t expert_local = expert_u - group.expert_base;

        device const block_q4_K *x =
            (device const block_q4_K *)(src0s + (uint64_t)expert_local * args.nb02 + first_row * args.nb01);
        device const float *y = (device const float *)(token_src1 + expert_slot * args.nb11);
        device const float *y4 = y + ix * QK_K + 64 * iq + 8 * ir;

        for (int ib = ix; ib < nb; ib += 4) {
            float yl[16];
            float yh[16];
            float4 sumy = {0.f, 0.f, 0.f, 0.f};

            for (short i = 0; i < 8; ++i) {
                yl[i + 0] = y4[i +   0]; sumy[0] += yl[i + 0];
                yl[i + 8] = y4[i +  32]; sumy[1] += yl[i + 8];
                yh[i + 0] = y4[i + 128]; sumy[2] += yh[i + 0];
                yh[i + 8] = y4[i + 160]; sumy[3] += yh[i + 8];
            }

            device const uint16_t *sc = (device const uint16_t *)x[ib].scales + iq;
            device const uint16_t *q1 = (device const uint16_t *)x[ib].qs + 16 * iq + 4 * ir;
            device const half *dh = &x[ib].d;

            for (short row = 0; row < nr0; row++) {
                if (first_row + row < args.ne0) {
                    sc16[0] = sc[0] & kmask1;
                    sc16[1] = sc[2] & kmask1;
                    sc16[2] = ((sc[4] >> 0) & kmask2) | ((sc[0] & kmask3) >> 2);
                    sc16[3] = ((sc[4] >> 4) & kmask2) | ((sc[2] & kmask3) >> 2);

                    device const uint16_t *q2 = q1 + 32;

                    float4 acc1 = {0.f, 0.f, 0.f, 0.f};
                    float4 acc2 = {0.f, 0.f, 0.f, 0.f};

                    FOR_UNROLL (short i = 0; i < 4; ++i) {
                        acc1[0] += yl[2 * i + 0] * (q1[i] & 0x000F);
                        acc1[1] += yl[2 * i + 1] * (q1[i] & 0x0F00);
                        acc1[2] += yl[2 * i + 8] * (q1[i] & 0x00F0);
                        acc1[3] += yl[2 * i + 9] * (q1[i] & 0xF000);
                        acc2[0] += yh[2 * i + 0] * (q2[i] & 0x000F);
                        acc2[1] += yh[2 * i + 1] * (q2[i] & 0x0F00);
                        acc2[2] += yh[2 * i + 8] * (q2[i] & 0x00F0);
                        acc2[3] += yh[2 * i + 9] * (q2[i] & 0xF000);
                    }

                    sumf[row] += dh[0] * ((acc1[0] + 1.f / 256.f * acc1[1]) * sc8[0] +
                                          (acc1[2] + 1.f / 256.f * acc1[3]) * sc8[1] * 1.f / 16.f +
                                          (acc2[0] + 1.f / 256.f * acc2[1]) * sc8[4] +
                                          (acc2[2] + 1.f / 256.f * acc2[3]) * sc8[5] * 1.f / 16.f) -
                                 dh[1] * (sumy[0] * sc8[2] + sumy[1] * sc8[3] +
                                          sumy[2] * sc8[6] + sumy[3] * sc8[7]);
                }

                q1 += args.nb01 / 2;
                sc += args.nb01 / 2;
                dh += args.nb01 / 2;
            }

            y4 += 4 * QK_K;
        }
    }

    device float *dst_f32 = (device float *)(dst + (uint64_t)token * args.nb1);
    for (int row = 0; row < nr0 && first_row + row < args.ne0; row++) {
        const float sum_all = simd_sum(sumf[row]);
        if (tiisg == 0) {
            if (group.accumulate) {
                dst_f32[first_row + row] += sum_all;
            } else {
                dst_f32[first_row + row] = sum_all;
            }
        }
    }

    (void)shmem;
    (void)tiitg;
    (void)tgpig;
}

kernel void kernel_mul_mv_table_q4_K_sum6_f32(
        constant ds4_metal_args_mul_mv_id & args,
        device const ds4_metal_q4_expert_table & table,
        device const char * src1,
        device       char * dst,
        device const char * ids,
        threadgroup  char * shmem [[threadgroup(0)]],
        uint3  tgpig[[threadgroup_position_in_grid]],
        ushort tiitg[[thread_index_in_threadgroup]],
        ushort tiisg[[thread_index_in_simdgroup]],
        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
    const short NSG = FC_mul_mv_nsg;
    const short nr0 = N_R0_Q4_K;
    const int nb = args.ne00 / QK_K;
    const int first_row = (tgpig.x * NSG + sgitg) * nr0;
    const uint token = tgpig.y;
    device const int32_t *token_ids = (device const int32_t *)(ids + (uint64_t)token * args.nbi1);
    device const char *token_src1 = src1 + (uint64_t)token * args.nb12;

    constexpr uint16_t kmask1 = 0x3f3f;
    constexpr uint16_t kmask2 = 0x0f0f;
    constexpr uint16_t kmask3 = 0xc0c0;

    const short ix = tiisg / 8;
    const short it = tiisg % 8;
    const short iq = it / 4;
    const short ir = it % 4;

    float sumf[nr0] = {0.f};
    uint16_t sc16[4];
    thread const uint8_t *sc8 = (thread const uint8_t *)sc16;

    for (int expert_slot = 0; expert_slot < 6; expert_slot++) {
        const int32_t expert = token_ids[expert_slot];
        if (expert < 0 || expert >= args.ne02 || expert >= 384) {
            return;
        }
        device const block_q4_K *x =
            (device const block_q4_K *)(table.experts[(uint)expert] + first_row * args.nb01);
        device const float *y = (device const float *)(token_src1 + expert_slot * args.nb11);
        device const float *y4 = y + ix * QK_K + 64 * iq + 8 * ir;

        for (int ib = ix; ib < nb; ib += 4) {
            float yl[16];
            float yh[16];
            float4 sumy = {0.f, 0.f, 0.f, 0.f};

            for (short i = 0; i < 8; ++i) {
                yl[i + 0] = y4[i +   0]; sumy[0] += yl[i + 0];
                yl[i + 8] = y4[i +  32]; sumy[1] += yl[i + 8];
                yh[i + 0] = y4[i + 128]; sumy[2] += yh[i + 0];
                yh[i + 8] = y4[i + 160]; sumy[3] += yh[i + 8];
            }

            device const uint16_t *sc = (device const uint16_t *)x[ib].scales + iq;
            device const uint16_t *q1 = (device const uint16_t *)x[ib].qs + 16 * iq + 4 * ir;
            device const half *dh = &x[ib].d;

            for (short row = 0; row < nr0; row++) {
                if (first_row + row < args.ne0) {
                    sc16[0] = sc[0] & kmask1;
                    sc16[1] = sc[2] & kmask1;
                    sc16[2] = ((sc[4] >> 0) & kmask2) | ((sc[0] & kmask3) >> 2);
                    sc16[3] = ((sc[4] >> 4) & kmask2) | ((sc[2] & kmask3) >> 2);

                    device const uint16_t *q2 = q1 + 32;

                    float4 acc1 = {0.f, 0.f, 0.f, 0.f};
                    float4 acc2 = {0.f, 0.f, 0.f, 0.f};

                    FOR_UNROLL (short i = 0; i < 4; ++i) {
                        acc1[0] += yl[2 * i + 0] * (q1[i] & 0x000F);
                        acc1[1] += yl[2 * i + 1] * (q1[i] & 0x0F00);
                        acc1[2] += yl[2 * i + 8] * (q1[i] & 0x00F0);
                        acc1[3] += yl[2 * i + 9] * (q1[i] & 0xF000);
                        acc2[0] += yh[2 * i + 0] * (q2[i] & 0x000F);
                        acc2[1] += yh[2 * i + 1] * (q2[i] & 0x0F00);
                        acc2[2] += yh[2 * i + 8] * (q2[i] & 0x00F0);
                        acc2[3] += yh[2 * i + 9] * (q2[i] & 0xF000);
                    }

                    sumf[row] += dh[0] * ((acc1[0] + 1.f / 256.f * acc1[1]) * sc8[0] +
                                          (acc1[2] + 1.f / 256.f * acc1[3]) * sc8[1] * 1.f / 16.f +
                                          (acc2[0] + 1.f / 256.f * acc2[1]) * sc8[4] +
                                          (acc2[2] + 1.f / 256.f * acc2[3]) * sc8[5] * 1.f / 16.f) -
                                 dh[1] * (sumy[0] * sc8[2] + sumy[1] * sc8[3] +
                                          sumy[2] * sc8[6] + sumy[3] * sc8[7]);
                }

                q1 += args.nb01 / 2;
                sc += args.nb01 / 2;
                dh += args.nb01 / 2;
            }

            y4 += 4 * QK_K;
        }
    }

    device float *dst_f32 = (device float *)(dst + (uint64_t)token * args.nb1);
    for (int row = 0; row < nr0 && first_row + row < args.ne0; row++) {
        const float sum_all = simd_sum(sumf[row]);
        if (tiisg == 0) dst_f32[first_row + row] = sum_all;
    }

    (void)shmem;
    (void)tiitg;
    (void)tgpig;
}

kernel void kernel_mul_mv_addr_q4_K_sum6_f32(
        constant ds4_metal_args_mul_mv_id & args,
        device const ulong * addrs,
        device const char * src1,
        device       char * dst,
        device const char * ids,
        threadgroup  char * shmem [[threadgroup(0)]],
        uint3  tgpig[[threadgroup_position_in_grid]],
        ushort tiitg[[thread_index_in_threadgroup]],
        ushort tiisg[[thread_index_in_simdgroup]],
        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
    const short NSG = FC_mul_mv_nsg;
    const short nr0 = N_R0_Q4_K;
    const int nb = args.ne00 / QK_K;
    const int first_row = (tgpig.x * NSG + sgitg) * nr0;
    const uint token = tgpig.y;
    device const int32_t *token_ids = (device const int32_t *)(ids + (uint64_t)token * args.nbi1);
    device const char *token_src1 = src1 + (uint64_t)token * args.nb12;

    constexpr uint16_t kmask1 = 0x3f3f;
    constexpr uint16_t kmask2 = 0x0f0f;
    constexpr uint16_t kmask3 = 0xc0c0;

    const short ix = tiisg / 8;
    const short it = tiisg % 8;
    const short iq = it / 4;
    const short ir = it % 4;

    float sumf[nr0] = {0.f};
    uint16_t sc16[4];
    thread const uint8_t *sc8 = (thread const uint8_t *)sc16;

    for (int expert_slot = 0; expert_slot < 6; expert_slot++) {
        const int32_t expert = token_ids[expert_slot];
        if (expert < 0 || expert >= args.ne02 || expert >= 384) {
            return;
        }
        device const char *expert_base =
            reinterpret_cast<device const char *>(addrs[(uint)expert]);
        device const block_q4_K *x =
            (device const block_q4_K *)(expert_base + first_row * args.nb01);
        device const float *y = (device const float *)(token_src1 + expert_slot * args.nb11);
        device const float *y4 = y + ix * QK_K + 64 * iq + 8 * ir;

        for (int ib = ix; ib < nb; ib += 4) {
            float yl[16];
            float yh[16];
            float4 sumy = {0.f, 0.f, 0.f, 0.f};

            for (short i = 0; i < 8; ++i) {
                yl[i + 0] = y4[i +   0]; sumy[0] += yl[i + 0];
                yl[i + 8] = y4[i +  32]; sumy[1] += yl[i + 8];
                yh[i + 0] = y4[i + 128]; sumy[2] += yh[i + 0];
                yh[i + 8] = y4[i + 160]; sumy[3] += yh[i + 8];
            }

            device const uint16_t *sc = (device const uint16_t *)x[ib].scales + iq;
            device const uint16_t *q1 = (device const uint16_t *)x[ib].qs + 16 * iq + 4 * ir;
            device const half *dh = &x[ib].d;

            for (short row = 0; row < nr0; row++) {
                if (first_row + row < args.ne0) {
                    sc16[0] = sc[0] & kmask1;
                    sc16[1] = sc[2] & kmask1;
                    sc16[2] = ((sc[4] >> 0) & kmask2) | ((sc[0] & kmask3) >> 2);
                    sc16[3] = ((sc[4] >> 4) & kmask2) | ((sc[2] & kmask3) >> 2);

                    device const uint16_t *q2 = q1 + 32;

                    float4 acc1 = {0.f, 0.f, 0.f, 0.f};
                    float4 acc2 = {0.f, 0.f, 0.f, 0.f};

                    FOR_UNROLL (short i = 0; i < 4; ++i) {
                        acc1[0] += yl[2 * i + 0] * (q1[i] & 0x000F);
                        acc1[1] += yl[2 * i + 1] * (q1[i] & 0x0F00);
                        acc1[2] += yl[2 * i + 8] * (q1[i] & 0x00F0);
                        acc1[3] += yl[2 * i + 9] * (q1[i] & 0xF000);
                        acc2[0] += yh[2 * i + 0] * (q2[i] & 0x000F);
                        acc2[1] += yh[2 * i + 1] * (q2[i] & 0x0F00);
                        acc2[2] += yh[2 * i + 8] * (q2[i] & 0x00F0);
                        acc2[3] += yh[2 * i + 9] * (q2[i] & 0xF000);
                    }

                    sumf[row] += dh[0] * ((acc1[0] + 1.f / 256.f * acc1[1]) * sc8[0] +
                                          (acc1[2] + 1.f / 256.f * acc1[3]) * sc8[1] * 1.f / 16.f +
                                          (acc2[0] + 1.f / 256.f * acc2[1]) * sc8[4] +
                                          (acc2[2] + 1.f / 256.f * acc2[3]) * sc8[5] * 1.f / 16.f) -
                                 dh[1] * (sumy[0] * sc8[2] + sumy[1] * sc8[3] +
                                          sumy[2] * sc8[6] + sumy[3] * sc8[7]);
                }

                q1 += args.nb01 / 2;
                sc += args.nb01 / 2;
                dh += args.nb01 / 2;
            }

            y4 += 4 * QK_K;
        }
    }

    device float *dst_f32 = (device float *)(dst + (uint64_t)token * args.nb1);
    for (int row = 0; row < nr0 && first_row + row < args.ne0; row++) {
        const float sum_all = simd_sum(sumf[row]);
        if (tiisg == 0) dst_f32[first_row + row] = sum_all;
    }

    (void)shmem;
    (void)tiitg;
    (void)tgpig;
}

kernel void kernel_mul_mv_slots6_q4_K_sum6_f32(
        constant ds4_metal_args_mul_mv_id & args,
        device const char * src00,
        device const char * src01,
        device const char * src02,
        device const char * src03,
        device const char * src04,
        device const char * src05,
        device const char * src1,
        device       char * dst,
        threadgroup  char * shmem [[threadgroup(0)]],
        uint3  tgpig[[threadgroup_position_in_grid]],
        ushort tiitg[[thread_index_in_threadgroup]],
        ushort tiisg[[thread_index_in_simdgroup]],
        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
    const short NSG = FC_mul_mv_nsg;
    const short nr0 = N_R0_Q4_K;
    const int nb = args.ne00 / QK_K;
    const int first_row = (tgpig.x * NSG + sgitg) * nr0;
    const uint token = tgpig.y;
    device const char *token_src1 = src1 + (uint64_t)token * args.nb12;

    constexpr uint16_t kmask1 = 0x3f3f;
    constexpr uint16_t kmask2 = 0x0f0f;
    constexpr uint16_t kmask3 = 0xc0c0;

    const short ix = tiisg / 8;
    const short it = tiisg % 8;
    const short iq = it / 4;
    const short ir = it % 4;

    float sumf[nr0] = {0.f};
    uint16_t sc16[4];
    thread const uint8_t *sc8 = (thread const uint8_t *)sc16;

    for (int expert_slot = 0; expert_slot < 6; expert_slot++) {
        device const char *src0_cur = src00;
        switch (expert_slot) {
        case 1: src0_cur = src01; break;
        case 2: src0_cur = src02; break;
        case 3: src0_cur = src03; break;
        case 4: src0_cur = src04; break;
        case 5: src0_cur = src05; break;
        default: break;
        }
        device const block_q4_K *x =
            (device const block_q4_K *)(src0_cur + first_row * args.nb01);
        device const float *y = (device const float *)(token_src1 + expert_slot * args.nb11);
        device const float *y4 = y + ix * QK_K + 64 * iq + 8 * ir;

        for (int ib = ix; ib < nb; ib += 4) {
            float yl[16];
            float yh[16];
            float4 sumy = {0.f, 0.f, 0.f, 0.f};

            for (short i = 0; i < 8; ++i) {
                yl[i + 0] = y4[i +   0]; sumy[0] += yl[i + 0];
                yl[i + 8] = y4[i +  32]; sumy[1] += yl[i + 8];
                yh[i + 0] = y4[i + 128]; sumy[2] += yh[i + 0];
                yh[i + 8] = y4[i + 160]; sumy[3] += yh[i + 8];
            }

            device const uint16_t *sc = (device const uint16_t *)x[ib].scales + iq;
            device const uint16_t *q1 = (device const uint16_t *)x[ib].qs + 16 * iq + 4 * ir;
            device const half *dh = &x[ib].d;

            for (short row = 0; row < nr0; row++) {
                if (first_row + row < args.ne0) {
                    sc16[0] = sc[0] & kmask1;
                    sc16[1] = sc[2] & kmask1;
                    sc16[2] = ((sc[4] >> 0) & kmask2) | ((sc[0] & kmask3) >> 2);
                    sc16[3] = ((sc[4] >> 4) & kmask2) | ((sc[2] & kmask3) >> 2);

                    device const uint16_t *q2 = q1 + 32;

                    float4 acc1 = {0.f, 0.f, 0.f, 0.f};
                    float4 acc2 = {0.f, 0.f, 0.f, 0.f};

                    FOR_UNROLL (short i = 0; i < 4; ++i) {
                        acc1[0] += yl[2 * i + 0] * (q1[i] & 0x000F);
                        acc1[1] += yl[2 * i + 1] * (q1[i] & 0x0F00);
                        acc1[2] += yl[2 * i + 8] * (q1[i] & 0x00F0);
                        acc1[3] += yl[2 * i + 9] * (q1[i] & 0xF000);
                        acc2[0] += yh[2 * i + 0] * (q2[i] & 0x000F);
                        acc2[1] += yh[2 * i + 1] * (q2[i] & 0x0F00);
                        acc2[2] += yh[2 * i + 8] * (q2[i] & 0x00F0);
                        acc2[3] += yh[2 * i + 9] * (q2[i] & 0xF000);
                    }

                    sumf[row] += dh[0] * ((acc1[0] + 1.f / 256.f * acc1[1]) * sc8[0] +
                                          (acc1[2] + 1.f / 256.f * acc1[3]) * sc8[1] * 1.f / 16.f +
                                          (acc2[0] + 1.f / 256.f * acc2[1]) * sc8[4] +
                                          (acc2[2] + 1.f / 256.f * acc2[3]) * sc8[5] * 1.f / 16.f) -
                                 dh[1] * (sumy[0] * sc8[2] + sumy[1] * sc8[3] +
                                          sumy[2] * sc8[6] + sumy[3] * sc8[7]);
                }

                q1 += args.nb01 / 2;
                sc += args.nb01 / 2;
                dh += args.nb01 / 2;
            }

            y4 += 4 * QK_K;
        }
    }

    device float *dst_f32 = (device float *)(dst + (uint64_t)token * args.nb1);
    for (int row = 0; row < nr0 && first_row + row < args.ne0; row++) {
        const float sum_all = simd_sum(sumf[row]);
        if (tiisg == 0) dst_f32[first_row + row] = sum_all;
    }

    (void)shmem;
    (void)tiitg;
    (void)tgpig;
}

kernel void kernel_mul_mv_group6_q4_K_sum6_f32(
        constant ds4_metal_args_mul_mv_id & args,
        device const char * src00,
        device const char * src01,
        device const char * src02,
        device const char * src03,
        device const char * src04,
        device const char * src05,
        device const char * src1,
        device       char * dst,
        device const char * ids,
        threadgroup  char * shmem [[threadgroup(0)]],
        uint3  tgpig[[threadgroup_position_in_grid]],
        ushort tiitg[[thread_index_in_threadgroup]],
        ushort tiisg[[thread_index_in_simdgroup]],
        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
    constexpr uint32_t expert_group_size = 64;
    const short NSG = FC_mul_mv_nsg;
    const short nr0 = N_R0_Q4_K;
    const int nb = args.ne00 / QK_K;
    const int first_row = (tgpig.x * NSG + sgitg) * nr0;
    const uint token = tgpig.y;
    device const int32_t *token_ids = (device const int32_t *)(ids + (uint64_t)token * args.nbi1);
    device const char *token_src1 = src1 + (uint64_t)token * args.nb12;

    constexpr uint16_t kmask1 = 0x3f3f;
    constexpr uint16_t kmask2 = 0x0f0f;
    constexpr uint16_t kmask3 = 0xc0c0;

    const short ix = tiisg / 8;
    const short it = tiisg % 8;
    const short iq = it / 4;
    const short ir = it % 4;

    float sumf[nr0] = {0.f};
    uint16_t sc16[4];
    thread const uint8_t *sc8 = (thread const uint8_t *)sc16;

    for (int expert_slot = 0; expert_slot < 6; expert_slot++) {
        const int32_t expert = token_ids[expert_slot];
        if (expert < 0) {
            continue;
        }
        const uint32_t expert_u = (uint32_t)expert;
        const uint32_t group_id = expert_u / expert_group_size;
        if (group_id >= 6) {
            continue;
        }
        const uint32_t expert_local = expert_u - group_id * expert_group_size;

        device const char *src0_cur = src00;
        switch (group_id) {
        case 1: src0_cur = src01; break;
        case 2: src0_cur = src02; break;
        case 3: src0_cur = src03; break;
        case 4: src0_cur = src04; break;
        case 5: src0_cur = src05; break;
        default: break;
        }

        device const block_q4_K *x =
            (device const block_q4_K *)(src0_cur + (uint64_t)expert_local * args.nb02 + first_row * args.nb01);
        device const float *y = (device const float *)(token_src1 + expert_slot * args.nb11);
        device const float *y4 = y + ix * QK_K + 64 * iq + 8 * ir;

        for (int ib = ix; ib < nb; ib += 4) {
            float yl[16];
            float yh[16];
            float4 sumy = {0.f, 0.f, 0.f, 0.f};

            for (short i = 0; i < 8; ++i) {
                yl[i + 0] = y4[i +   0]; sumy[0] += yl[i + 0];
                yl[i + 8] = y4[i +  32]; sumy[1] += yl[i + 8];
                yh[i + 0] = y4[i + 128]; sumy[2] += yh[i + 0];
                yh[i + 8] = y4[i + 160]; sumy[3] += yh[i + 8];
            }

            device const uint16_t *sc = (device const uint16_t *)x[ib].scales + iq;
            device const uint16_t *q1 = (device const uint16_t *)x[ib].qs + 16 * iq + 4 * ir;
            device const half *dh = &x[ib].d;

            for (short row = 0; row < nr0; row++) {
                if (first_row + row < args.ne0) {
                    sc16[0] = sc[0] & kmask1;
                    sc16[1] = sc[2] & kmask1;
                    sc16[2] = ((sc[4] >> 0) & kmask2) | ((sc[0] & kmask3) >> 2);
                    sc16[3] = ((sc[4] >> 4) & kmask2) | ((sc[2] & kmask3) >> 2);

                    device const uint16_t *q2 = q1 + 32;

                    float4 acc1 = {0.f, 0.f, 0.f, 0.f};
                    float4 acc2 = {0.f, 0.f, 0.f, 0.f};

                    FOR_UNROLL (short i = 0; i < 4; ++i) {
                        acc1[0] += yl[2 * i + 0] * (q1[i] & 0x000F);
                        acc1[1] += yl[2 * i + 1] * (q1[i] & 0x0F00);
                        acc1[2] += yl[2 * i + 8] * (q1[i] & 0x00F0);
                        acc1[3] += yl[2 * i + 9] * (q1[i] & 0xF000);
                        acc2[0] += yh[2 * i + 0] * (q2[i] & 0x000F);
                        acc2[1] += yh[2 * i + 1] * (q2[i] & 0x0F00);
                        acc2[2] += yh[2 * i + 8] * (q2[i] & 0x00F0);
                        acc2[3] += yh[2 * i + 9] * (q2[i] & 0xF000);
                    }

                    sumf[row] += dh[0] * ((acc1[0] + 1.f / 256.f * acc1[1]) * sc8[0] +
                                          (acc1[2] + 1.f / 256.f * acc1[3]) * sc8[1] * 1.f / 16.f +
                                          (acc2[0] + 1.f / 256.f * acc2[1]) * sc8[4] +
                                          (acc2[2] + 1.f / 256.f * acc2[3]) * sc8[5] * 1.f / 16.f) -
                                 dh[1] * (sumy[0] * sc8[2] + sumy[1] * sc8[3] +
                                          sumy[2] * sc8[6] + sumy[3] * sc8[7]);
                }

                q1 += args.nb01 / 2;
                sc += args.nb01 / 2;
                dh += args.nb01 / 2;
            }

            y4 += 4 * QK_K;
        }
    }

    device float *dst_f32 = (device float *)(dst + (uint64_t)token * args.nb1);
    for (int row = 0; row < nr0 && first_row + row < args.ne0; row++) {
        const float sum_all = simd_sum(sumf[row]);
        if (tiisg == 0) dst_f32[first_row + row] = sum_all;
    }

    (void)shmem;
    (void)tiitg;
    (void)tgpig;
}

kernel void kernel_mul_mv_group8_q4_K_sum6_f32(
        constant ds4_metal_args_mul_mv_id & args,
        device const char * src00,
        device const char * src01,
        device const char * src02,
        device const char * src03,
        device const char * src04,
        device const char * src05,
        device const char * src06,
        device const char * src07,
        device const char * src1,
        device       char * dst,
        device const char * ids,
        threadgroup  char * shmem [[threadgroup(0)]],
        uint3  tgpig[[threadgroup_position_in_grid]],
        ushort tiitg[[thread_index_in_threadgroup]],
        ushort tiisg[[thread_index_in_simdgroup]],
        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
    constexpr uint32_t expert_group_size = 48;
    const short NSG = FC_mul_mv_nsg;
    const short nr0 = N_R0_Q4_K;
    const int nb = args.ne00 / QK_K;
    const int first_row = (tgpig.x * NSG + sgitg) * nr0;
    const uint token = tgpig.y;
    device const int32_t *token_ids = (device const int32_t *)(ids + (uint64_t)token * args.nbi1);
    device const char *token_src1 = src1 + (uint64_t)token * args.nb12;

    constexpr uint16_t kmask1 = 0x3f3f;
    constexpr uint16_t kmask2 = 0x0f0f;
    constexpr uint16_t kmask3 = 0xc0c0;

    const short ix = tiisg / 8;
    const short it = tiisg % 8;
    const short iq = it / 4;
    const short ir = it % 4;

    float sumf[nr0] = {0.f};
    uint16_t sc16[4];
    thread const uint8_t *sc8 = (thread const uint8_t *)sc16;

    for (int expert_slot = 0; expert_slot < 6; expert_slot++) {
        const int32_t expert = token_ids[expert_slot];
        if (expert < 0) {
            continue;
        }
        const uint32_t expert_u = (uint32_t)expert;
        const uint32_t group_id = expert_u / expert_group_size;
        if (group_id >= 8) {
            continue;
        }
        const uint32_t expert_local = expert_u - group_id * expert_group_size;

        device const char *src0_cur = src00;
        switch (group_id) {
        case 1: src0_cur = src01; break;
        case 2: src0_cur = src02; break;
        case 3: src0_cur = src03; break;
        case 4: src0_cur = src04; break;
        case 5: src0_cur = src05; break;
        case 6: src0_cur = src06; break;
        case 7: src0_cur = src07; break;
        default: break;
        }

        device const block_q4_K *x =
            (device const block_q4_K *)(src0_cur + (uint64_t)expert_local * args.nb02 + first_row * args.nb01);
        device const float *y = (device const float *)(token_src1 + expert_slot * args.nb11);
        device const float *y4 = y + ix * QK_K + 64 * iq + 8 * ir;

        for (int ib = ix; ib < nb; ib += 4) {
            float yl[16];
            float yh[16];
            float4 sumy = {0.f, 0.f, 0.f, 0.f};

            for (short i = 0; i < 8; ++i) {
                yl[i + 0] = y4[i +   0]; sumy[0] += yl[i + 0];
                yl[i + 8] = y4[i +  32]; sumy[1] += yl[i + 8];
                yh[i + 0] = y4[i + 128]; sumy[2] += yh[i + 0];
                yh[i + 8] = y4[i + 160]; sumy[3] += yh[i + 8];
            }

            device const uint16_t *sc = (device const uint16_t *)x[ib].scales + iq;
            device const uint16_t *q1 = (device const uint16_t *)x[ib].qs + 16 * iq + 4 * ir;
            device const half *dh = &x[ib].d;

            for (short row = 0; row < nr0; row++) {
                if (first_row + row < args.ne0) {
                    sc16[0] = sc[0] & kmask1;
                    sc16[1] = sc[2] & kmask1;
                    sc16[2] = ((sc[4] >> 0) & kmask2) | ((sc[0] & kmask3) >> 2);
                    sc16[3] = ((sc[4] >> 4) & kmask2) | ((sc[2] & kmask3) >> 2);

                    device const uint16_t *q2 = q1 + 32;

                    float4 acc1 = {0.f, 0.f, 0.f, 0.f};
                    float4 acc2 = {0.f, 0.f, 0.f, 0.f};

                    FOR_UNROLL (short i = 0; i < 4; ++i) {
                        acc1[0] += yl[2 * i + 0] * (q1[i] & 0x000F);
                        acc1[1] += yl[2 * i + 1] * (q1[i] & 0x0F00);
                        acc1[2] += yl[2 * i + 8] * (q1[i] & 0x00F0);
                        acc1[3] += yl[2 * i + 9] * (q1[i] & 0xF000);
                        acc2[0] += yh[2 * i + 0] * (q2[i] & 0x000F);
                        acc2[1] += yh[2 * i + 1] * (q2[i] & 0x0F00);
                        acc2[2] += yh[2 * i + 8] * (q2[i] & 0x00F0);
                        acc2[3] += yh[2 * i + 9] * (q2[i] & 0xF000);
                    }

                    sumf[row] += dh[0] * ((acc1[0] + 1.f / 256.f * acc1[1]) * sc8[0] +
                                          (acc1[2] + 1.f / 256.f * acc1[3]) * sc8[1] * 1.f / 16.f +
                                          (acc2[0] + 1.f / 256.f * acc2[1]) * sc8[4] +
                                          (acc2[2] + 1.f / 256.f * acc2[3]) * sc8[5] * 1.f / 16.f) -
                                 dh[1] * (sumy[0] * sc8[2] + sumy[1] * sc8[3] +
                                          sumy[2] * sc8[6] + sumy[3] * sc8[7]);
                }

                q1 += args.nb01 / 2;
                sc += args.nb01 / 2;
                dh += args.nb01 / 2;
            }

            y4 += 4 * QK_K;
        }
    }

    device float *dst_f32 = (device float *)(dst + (uint64_t)token * args.nb1);
    for (int row = 0; row < nr0 && first_row + row < args.ne0; row++) {
        const float sum_all = simd_sum(sumf[row]);
        if (tiisg == 0) dst_f32[first_row + row] = sum_all;
    }

    (void)shmem;
    (void)tiitg;
    (void)tgpig;
}

kernel void kernel_mul_mv_group24_q4_K_sum6_f32(
        constant ds4_metal_args_mul_mv_id & args,
        device const char * src00,
        device const char * src01,
        device const char * src02,
        device const char * src03,
        device const char * src04,
        device const char * src05,
        device const char * src06,
        device const char * src07,
        device const char * src08,
        device const char * src09,
        device const char * src10,
        device const char * src11,
        device const char * src12,
        device const char * src13,
        device const char * src14,
        device const char * src15,
        device const char * src16,
        device const char * src17,
        device const char * src18,
        device const char * src19,
        device const char * src20,
        device const char * src21,
        device const char * src22,
        device const char * src23,
        device const char * src1,
        device       char * dst,
        device const char * ids,
        threadgroup  char * shmem [[threadgroup(0)]],
        uint3  tgpig[[threadgroup_position_in_grid]],
        ushort tiitg[[thread_index_in_threadgroup]],
        ushort tiisg[[thread_index_in_simdgroup]],
        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
    constexpr uint32_t expert_group_size = 16;
    const short NSG = FC_mul_mv_nsg;
    const short nr0 = N_R0_Q4_K;
    const int nb = args.ne00 / QK_K;
    const int first_row = (tgpig.x * NSG + sgitg) * nr0;
    const uint token = tgpig.y;
    device const int32_t *token_ids = (device const int32_t *)(ids + (uint64_t)token * args.nbi1);
    device const char *token_src1 = src1 + (uint64_t)token * args.nb12;

    constexpr uint16_t kmask1 = 0x3f3f;
    constexpr uint16_t kmask2 = 0x0f0f;
    constexpr uint16_t kmask3 = 0xc0c0;

    const short ix = tiisg / 8;
    const short it = tiisg % 8;
    const short iq = it / 4;
    const short ir = it % 4;

    float sumf[nr0] = {0.f};
    uint16_t sc16[4];
    thread const uint8_t *sc8 = (thread const uint8_t *)sc16;

    for (int expert_slot = 0; expert_slot < 6; expert_slot++) {
        const int32_t expert = token_ids[expert_slot];
        if (expert < 0) {
            continue;
        }
        const uint32_t expert_u = (uint32_t)expert;
        const uint32_t group_id = expert_u / expert_group_size;
        if (group_id >= 24) {
            continue;
        }
        const uint32_t expert_local = expert_u - group_id * expert_group_size;

        device const char *src0_cur = ds4_q4_group24_select(group_id,
                                                            src00, src01, src02, src03,
                                                            src04, src05, src06, src07,
                                                            src08, src09, src10, src11,
                                                            src12, src13, src14, src15,
                                                            src16, src17, src18, src19,
                                                            src20, src21, src22, src23);
        device const block_q4_K *x =
            (device const block_q4_K *)(src0_cur + (uint64_t)expert_local * args.nb02 + first_row * args.nb01);
        device const float *y = (device const float *)(token_src1 + expert_slot * args.nb11);
        device const float *y4 = y + ix * QK_K + 64 * iq + 8 * ir;

        for (int ib = ix; ib < nb; ib += 4) {
            float yl[16];
            float yh[16];
            float4 sumy = {0.f, 0.f, 0.f, 0.f};

            for (short i = 0; i < 8; ++i) {
                yl[i + 0] = y4[i +   0]; sumy[0] += yl[i + 0];
                yl[i + 8] = y4[i +  32]; sumy[1] += yl[i + 8];
                yh[i + 0] = y4[i + 128]; sumy[2] += yh[i + 0];
                yh[i + 8] = y4[i + 160]; sumy[3] += yh[i + 8];
            }

            device const uint16_t *sc = (device const uint16_t *)x[ib].scales + iq;
            device const uint16_t *q1 = (device const uint16_t *)x[ib].qs + 16 * iq + 4 * ir;
            device const half *dh = &x[ib].d;

            for (short row = 0; row < nr0; row++) {
                if (first_row + row < args.ne0) {
                    sc16[0] = sc[0] & kmask1;
                    sc16[1] = sc[2] & kmask1;
                    sc16[2] = ((sc[4] >> 0) & kmask2) | ((sc[0] & kmask3) >> 2);
                    sc16[3] = ((sc[4] >> 4) & kmask2) | ((sc[2] & kmask3) >> 2);

                    device const uint16_t *q2 = q1 + 32;

                    float4 acc1 = {0.f, 0.f, 0.f, 0.f};
                    float4 acc2 = {0.f, 0.f, 0.f, 0.f};

                    FOR_UNROLL (short i = 0; i < 4; ++i) {
                        acc1[0] += yl[2 * i + 0] * (q1[i] & 0x000F);
                        acc1[1] += yl[2 * i + 1] * (q1[i] & 0x0F00);
                        acc1[2] += yl[2 * i + 8] * (q1[i] & 0x00F0);
                        acc1[3] += yl[2 * i + 9] * (q1[i] & 0xF000);
                        acc2[0] += yh[2 * i + 0] * (q2[i] & 0x000F);
                        acc2[1] += yh[2 * i + 1] * (q2[i] & 0x0F00);
                        acc2[2] += yh[2 * i + 8] * (q2[i] & 0x00F0);
                        acc2[3] += yh[2 * i + 9] * (q2[i] & 0xF000);
                    }

                    sumf[row] += dh[0] * ((acc1[0] + 1.f / 256.f * acc1[1]) * sc8[0] +
                                          (acc1[2] + 1.f / 256.f * acc1[3]) * sc8[1] * 1.f / 16.f +
                                          (acc2[0] + 1.f / 256.f * acc2[1]) * sc8[4] +
                                          (acc2[2] + 1.f / 256.f * acc2[3]) * sc8[5] * 1.f / 16.f) -
                                 dh[1] * (sumy[0] * sc8[2] + sumy[1] * sc8[3] +
                                          sumy[2] * sc8[6] + sumy[3] * sc8[7]);
                }

                q1 += args.nb01 / 2;
                sc += args.nb01 / 2;
                dh += args.nb01 / 2;
            }

            y4 += 4 * QK_K;
        }
    }

    device float *dst_f32 = (device float *)(dst + (uint64_t)token * args.nb1);
    for (int row = 0; row < nr0 && first_row + row < args.ne0; row++) {
        const float sum_all = simd_sum(sumf[row]);
        if (tiisg == 0) dst_f32[first_row + row] = sum_all;
    }

    (void)shmem;
    (void)tiitg;
    (void)tgpig;
}

#define QK_NL 16

// Builds the compact per-expert work map used by batched MoE matmul. DS4 routes
// each token to a small fixed top-k list, so this turns token-major ids into
// expert-major slices that the tiled matmul can consume.
template<short ne20>
kernel void kernel_mul_mm_id_map0(
        constant ds4_metal_args_mul_mm_id_map0 & args,
        device  const char * src2,
        device        char * htpe,
        device        char * hids,
        threadgroup   char * shmem [[threadgroup(0)]],
        ushort tpitg[[thread_position_in_threadgroup]],
        ushort   ntg[[threads_per_threadgroup]]) {
    const short ide = tpitg;

    uint32_t n_all = 0;

    device int32_t * ids_i32 = (device int32_t *) hids + ide*args.ne21;

    for (int i21 = 0; i21 < args.ne21; i21 += ntg) {
        if (i21 + tpitg < args.ne21) {
            device const int32_t * src2_i32 = (device const int32_t *) (src2 + (i21 + tpitg)*args.nb21);

            threadgroup uint16_t * sids = (threadgroup uint16_t *) shmem + tpitg*ne20;

            #pragma unroll(ne20)
            for (short i20 = 0; i20 < ne20; i20++) {
                sids[i20] = src2_i32[i20];
            }
        }

        threadgroup_barrier(mem_flags::mem_threadgroup);

        for (short t = 0; t < ntg; t++) {
            if (i21 + t >= args.ne21) {
                break;
            }

            threadgroup const uint16_t * sids = (threadgroup const uint16_t *) shmem + t*ne20;

            short sel = 0;
            #pragma unroll(ne20)
            for (short i20 = 0; i20 < ne20; i20++) {
                sel += (sids[i20] == ide)*(i20 + 1);
            }

            ids_i32[n_all] = (i21 + t)*ne20 + sel - 1;

            n_all += sel > 0;
        }

        threadgroup_barrier(mem_flags::mem_threadgroup);
    }

    device uint32_t * tpe_u32 = (device uint32_t *) (htpe);
    tpe_u32[ide] = n_all;
}

typedef decltype(kernel_mul_mm_id_map0<1>) kernel_mul_mm_id_map0_t;

// Host-visible map builders for the routed-expert counts used by DS4 graph
// shapes. Some arities are generic leftovers retained for nearby batch sizes.
template [[host_name("kernel_mul_mm_id_map0_ne20_1" )]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<1>;
template [[host_name("kernel_mul_mm_id_map0_ne20_2" )]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<2>;
template [[host_name("kernel_mul_mm_id_map0_ne20_4" )]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<4>;
template [[host_name("kernel_mul_mm_id_map0_ne20_5" )]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<5>;
template [[host_name("kernel_mul_mm_id_map0_ne20_6" )]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<6>;
template [[host_name("kernel_mul_mm_id_map0_ne20_8" )]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<8>;
template [[host_name("kernel_mul_mm_id_map0_ne20_10")]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<10>;
template [[host_name("kernel_mul_mm_id_map0_ne20_16")]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<16>;
template [[host_name("kernel_mul_mm_id_map0_ne20_22")]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<22>;

// Batched routed-expert matmul. It reads the expert-major map produced above,
// loads selected expert weights, and writes results back to token-major slots
// so the DS4 FFN can apply SwiGLU, weighting, and the down projection.
template<short NR1, typename S0, typename S0_4x4, typename S0_8x8, typename S1, typename S1_2x4, typename S1_8x8, typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread S0_4x4 &), typename T0, typename T0_4x4, typename T1, typename T1_2x4>
kernel void kernel_mul_mm_id(
        constant ds4_metal_args_mul_mm_id & args,
        device const char * src0,
        device const char * src1,
        device const char * htpe,
        device const char * hids,
        device       char * dst,
        threadgroup  char * shmem [[threadgroup(0)]],
        uint3  tgpig[[threadgroup_position_in_grid]],
        ushort tiitg[[thread_index_in_threadgroup]],
        ushort tiisg[[thread_index_in_simdgroup]],
        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
    constexpr int NR0 = 64;
    static_assert(NR1 == 32, "kernel_mul_mm_id accumulator layout supports only 32 routed rows");

    constexpr int NK  = 32;
    constexpr int NL0 = NK/16;
    constexpr int NL1 = NK/8;
    constexpr int SA_BYTES = NR0 * NR1 * (int)sizeof(S0);

    threadgroup S0 * sa = (threadgroup S0 *)(shmem);
    threadgroup S1 * sb = (threadgroup S1 *)(shmem + SA_BYTES);

    const int im = tgpig.z;
    const int r0 = tgpig.y*NR0;
    const int r1 = tgpig.x*NR1;

    device const uint32_t * tpe_u32 = (device const uint32_t *) (htpe);
    device const int32_t  * ids_i32 = (device const int32_t  *) (hids);

    const int32_t neh1 = tpe_u32[im];

    if (r1 >= neh1) {
        return;
    }

    const short nr0 = (args.ne0 - r0 < NR0) ? (args.ne0 - r0) : NR0;
    const short nr1 = (    neh1 - r1 < NR1) ? (    neh1 - r1) : NR1;

    if (!ds4_tp_owns_expert(im, args.ne02, args.tp_rank, args.tp_world)) {
        /* Unowned expert under the TP split: zero this tile's output rows so
         * the downstream swiglu/sum stages stay unchanged. Each (token,slot)
         * row belongs to exactly one expert, so nothing else writes them. */
        for (short j = sgitg; j < nr1; j += 4) {
            const int idj = ids_i32[im*args.ne21 + r1 + j];

            const short ide = idj % args.ne20;
            const short idt = idj / args.ne20;

            device float * D = (device float *) dst + r0 + ide*args.ne0 + idt*args.ne1*args.ne0;

            for (int i = tiisg; i < nr0; i += 32) {
                D[i] = 0.0f;
            }
        }
        return;
    }

    const short lr0 = ((short)tiitg/NL0) < nr0 ? ((short)tiitg/NL0) : nr0 - 1;
    const short lr1 = ((short)tiitg/NL1) < nr1 ? ((short)tiitg/NL1) : nr1 - 1;

    const short il0 = (tiitg % NL0);

    short il = il0;

    const int id = ids_i32[im*args.ne21 + r1 + lr1];

    const short i11 = (id % args.ne20) % args.ne11;
    const short i12 = (id / args.ne20);
    const short i13 = 0;

    const uint64_t offset0 = (uint64_t)(im - args.tp_expert_base)*args.nb02 + i13*args.nb03;
    const short    offset1 = il0/nl;

    device const block_q * x = (device const block_q *)(src0 + args.nb01*(r0 + lr0) + offset0) + offset1;

    const short iy = 8*(tiitg % NL1);

    device const T1 * y = (device const T1 *)(src1
        + args.nb13*i13
        + args.nb12*i12
        + args.nb11*i11
        + args.nb10*iy);

    S0_8x8 ma[4];
    S1_8x8 mb[2];

    simdgroup_float8x8 mc[8];

    for (short i = 0; i < 8; i++){
        mc[i] = make_filled_simdgroup_matrix<float, 8>(0.f);
    }

    for (int loop_k = 0; loop_k < args.ne00; loop_k += NK) {
        if (is_same<T0_4x4, block_q>::value && FC_mul_mm_bc_inp) {
            threadgroup_barrier(mem_flags::mem_threadgroup);

            for (short i = 0; i < 16; i++) {
                const short sx = 2*il0 + i/8;
                const short sy = (tiitg/NL0)/8;

                const short lx = (tiitg/NL0)%8;
                const short ly = i%8;

                const short ib = 8*sx + sy;

                *(sa + 64*ib + 8*ly + lx) = loop_k + 16*il + i < args.ne00 ? *((device T0 *) x + i) : 0;
            }
        } else {
            S0_4x4 temp_a;
            dequantize_func(x, il, temp_a);

            threadgroup_barrier(mem_flags::mem_threadgroup);

            FOR_UNROLL (short i = 0; i < 16; i++) {
                const short sx = 2*il0 + i/8;
                const short sy = (tiitg/NL0)/8;

                const short lx = (tiitg/NL0)%8;
                const short ly = i%8;

                const short ib = 8*sx + sy;

                *(sa + 64*ib + 8*ly + lx) = temp_a[i/4][i%4];
            }
        }

        if (FC_mul_mm_bc_inp) {
            for (short i = 0; i < 8; ++i) {
                const short sx = (tiitg%NL1);
                const short sy = (tiitg/NL1)/8;

                const short lx = i;
                const short ly = (tiitg/NL1)%8;

                const short ib = 4*sx + sy;

                *(sb + 64*ib + 8*ly + lx) = loop_k + iy + i < args.ne00 ? (S1) *((device T1 *) y + i) : 0;
            }
        } else {
            const short sx = (tiitg%NL1);
            const short sy = (tiitg/NL1)/8;

            const short ly = (tiitg/NL1)%8;

            const short ib = 4*sx + sy;

            *(threadgroup S1_2x4 *)(sb + 64*ib + 8*ly) = (S1_2x4)(*((device T1_2x4 *) y));
        }

        il = (il + 2 < nl) ? il + 2 : il % 2;
        x  = (il < 2) ? x + (2 + nl - 1)/nl : x;

        y += NK;

        threadgroup_barrier(mem_flags::mem_threadgroup);

        threadgroup const S0 * lsma = (sa + 4*64*(sgitg%2));
        threadgroup const S1 * lsmb = (sb + 2*64*(sgitg/2));

        FOR_UNROLL (short ik = 0; ik < NK/8; ik++) {
            simdgroup_barrier(mem_flags::mem_none);

            FOR_UNROLL (short i = 0; i < 4; i++) {
                simdgroup_load(ma[i], lsma + 64*i, 8, 0, false);
            }

            simdgroup_barrier(mem_flags::mem_none);

            FOR_UNROLL (short i = 0; i < 2; i++) {
                simdgroup_load(mb[i], lsmb + 64*i, 8, 0, false);
            }

            simdgroup_barrier(mem_flags::mem_none);

            FOR_UNROLL (short i = 0; i < 8; i++){
                simdgroup_multiply_accumulate(mc[i], mb[i/4], ma[i%4], mc[i]);
            }

            lsma += 8*64;
            lsmb += 4*64;
        }
    }

    threadgroup_barrier(mem_flags::mem_threadgroup);

    threadgroup float * temp_str = ((threadgroup float *) shmem) + 32*(sgitg&1) + (16*(sgitg >> 1))*NR0;

    for (short i = 0; i < 8; i++) {
        simdgroup_store(mc[i], temp_str + 8*(i%4) + 8*NR0*(i/4), NR0, 0, false);
    }

    threadgroup_barrier(mem_flags::mem_threadgroup);

    for (short j = sgitg; j < nr1; j += 4) {
        const int idj = ids_i32[im*args.ne21 + r1 + j];

        const short ide = idj % args.ne20;
        const short idt = idj / args.ne20;

        device float  * D  = (device float  *) dst + r0 + ide*args.ne0 + idt*args.ne1*args.ne0;
        device float4 * D4 = (device float4 *) D;

        threadgroup float  * C  = (threadgroup float  *) shmem + j*NR0;
        threadgroup float4 * C4 = (threadgroup float4 *) C;

        int i = tiisg;
        for (; i < nr0/4; i += 32) {
            *(D4 + i) = *(C4 + i);
        }

        i = (4*(nr0/4)) + tiisg;
        for (; i < nr0; i += 32) {
            *(D + i) = *(C + i);
        }
    }
}

// Address-table variant used by SSD streaming.  The routing ids remain the
// model's original expert ids, but each expert's resident buffer is found via a
// GPU-address table instead of a contiguous full-layer tensor.
template<short NR1, typename S0, typename S0_4x4, typename S0_8x8, typename S1, typename S1_2x4, typename S1_8x8, typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread S0_4x4 &), typename T0, typename T0_4x4, typename T1, typename T1_2x4>
kernel void kernel_mul_mm_id_addr(
        constant ds4_metal_args_mul_mm_id & args,
        device const uint64_t * src0_addrs,
        device const char * src1,
        device const char * htpe,
        device const char * hids,
        device       char * dst,
        threadgroup  char * shmem [[threadgroup(0)]],
        uint3  tgpig[[threadgroup_position_in_grid]],
        ushort tiitg[[thread_index_in_threadgroup]],
        ushort tiisg[[thread_index_in_simdgroup]],
        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
    constexpr int NR0 = 64;
    static_assert(NR1 == 32, "kernel_mul_mm_id_addr accumulator layout supports only 32 routed rows");

    constexpr int NK  = 32;
    constexpr int NL0 = NK/16;
    constexpr int NL1 = NK/8;
    constexpr int SA_BYTES = NR0 * NR1 * (int)sizeof(S0);

    threadgroup S0 * sa = (threadgroup S0 *)(shmem);
    threadgroup S1 * sb = (threadgroup S1 *)(shmem + SA_BYTES);

    const int im = tgpig.z;
    const int r0 = tgpig.y*NR0;
    const int r1 = tgpig.x*NR1;

    device const uint32_t * tpe_u32 = (device const uint32_t *) (htpe);
    device const int32_t  * ids_i32 = (device const int32_t  *) (hids);

    const int32_t neh1 = tpe_u32[im];

    if (r1 >= neh1) {
        return;
    }

    const uint64_t base_addr = src0_addrs[im];
    if (base_addr == 0) {
        return;
    }
    device const char * src0 = reinterpret_cast<device const char *>(base_addr);

    const short nr0 = (args.ne0 - r0 < NR0) ? (args.ne0 - r0) : NR0;
    const short nr1 = (    neh1 - r1 < NR1) ? (    neh1 - r1) : NR1;

    const short lr0 = ((short)tiitg/NL0) < nr0 ? ((short)tiitg/NL0) : nr0 - 1;
    const short lr1 = ((short)tiitg/NL1) < nr1 ? ((short)tiitg/NL1) : nr1 - 1;

    const short il0 = (tiitg % NL0);

    short il = il0;

    const int id = ids_i32[im*args.ne21 + r1 + lr1];

    const short i11 = (id % args.ne20) % args.ne11;
    const short i12 = (id / args.ne20);
    const short i13 = 0;

    const short offset1 = il0/nl;

    device const block_q * x = (device const block_q *)(src0 + args.nb01*(r0 + lr0)) + offset1;

    const short iy = 8*(tiitg % NL1);

    device const T1 * y = (device const T1 *)(src1
        + args.nb13*i13
        + args.nb12*i12
        + args.nb11*i11
        + args.nb10*iy);

    S0_8x8 ma[4];
    S1_8x8 mb[2];

    simdgroup_float8x8 mc[8];

    for (short i = 0; i < 8; i++){
        mc[i] = make_filled_simdgroup_matrix<float, 8>(0.f);
    }

    for (int loop_k = 0; loop_k < args.ne00; loop_k += NK) {
        if (is_same<T0_4x4, block_q>::value && FC_mul_mm_bc_inp) {
            threadgroup_barrier(mem_flags::mem_threadgroup);

            for (short i = 0; i < 16; i++) {
                const short sx = 2*il0 + i/8;
                const short sy = (tiitg/NL0)/8;

                const short lx = (tiitg/NL0)%8;
                const short ly = i%8;

                const short ib = 8*sx + sy;

                *(sa + 64*ib + 8*ly + lx) = loop_k + 16*il + i < args.ne00 ? *((device T0 *) x + i) : 0;
            }
        } else {
            S0_4x4 temp_a;
            dequantize_func(x, il, temp_a);

            threadgroup_barrier(mem_flags::mem_threadgroup);

            FOR_UNROLL (short i = 0; i < 16; i++) {
                const short sx = 2*il0 + i/8;
                const short sy = (tiitg/NL0)/8;

                const short lx = (tiitg/NL0)%8;
                const short ly = i%8;

                const short ib = 8*sx + sy;

                *(sa + 64*ib + 8*ly + lx) = temp_a[i/4][i%4];
            }
        }

        if (FC_mul_mm_bc_inp) {
            for (short i = 0; i < 8; ++i) {
                const short sx = (tiitg%NL1);
                const short sy = (tiitg/NL1)/8;

                const short lx = i;
                const short ly = (tiitg/NL1)%8;

                const short ib = 4*sx + sy;

                *(sb + 64*ib + 8*ly + lx) = loop_k + iy + i < args.ne00 ? (S1) *((device T1 *) y + i) : 0;
            }
        } else {
            const short sx = (tiitg%NL1);
            const short sy = (tiitg/NL1)/8;

            const short ly = (tiitg/NL1)%8;

            const short ib = 4*sx + sy;

            *(threadgroup S1_2x4 *)(sb + 64*ib + 8*ly) = (S1_2x4)(*((device T1_2x4 *) y));
        }

        il = (il + 2 < nl) ? il + 2 : il % 2;
        x  = (il < 2) ? x + (2 + nl - 1)/nl : x;

        y += NK;

        threadgroup_barrier(mem_flags::mem_threadgroup);

        threadgroup const S0 * lsma = (sa + 4*64*(sgitg%2));
        threadgroup const S1 * lsmb = (sb + 2*64*(sgitg/2));

        FOR_UNROLL (short ik = 0; ik < NK/8; ik++) {
            simdgroup_barrier(mem_flags::mem_none);

            FOR_UNROLL (short i = 0; i < 4; i++) {
                simdgroup_load(ma[i], lsma + 64*i, 8, 0, false);
            }

            simdgroup_barrier(mem_flags::mem_none);

            FOR_UNROLL (short i = 0; i < 2; i++) {
                simdgroup_load(mb[i], lsmb + 64*i, 8, 0, false);
            }

            simdgroup_barrier(mem_flags::mem_none);

            FOR_UNROLL (short i = 0; i < 8; i++){
                simdgroup_multiply_accumulate(mc[i], mb[i/4], ma[i%4], mc[i]);
            }

            lsma += 8*64;
            lsmb += 4*64;
        }
    }

    threadgroup_barrier(mem_flags::mem_threadgroup);

    threadgroup float * temp_str = ((threadgroup float *) shmem) + 32*(sgitg&1) + (16*(sgitg >> 1))*NR0;

    for (short i = 0; i < 8; i++) {
        simdgroup_store(mc[i], temp_str + 8*(i%4) + 8*NR0*(i/4), NR0, 0, false);
    }

    threadgroup_barrier(mem_flags::mem_threadgroup);

    for (short j = sgitg; j < nr1; j += 4) {
        const int idj = ids_i32[im*args.ne21 + r1 + j];

        const short ide = idj % args.ne20;
        const short idt = idj / args.ne20;

        device float  * D  = (device float  *) dst + r0 + ide*args.ne0 + idt*args.ne1*args.ne0;
        device float4 * D4 = (device float4 *) D;

        threadgroup float  * C  = (threadgroup float  *) shmem + j*NR0;
        threadgroup float4 * C4 = (threadgroup float4 *) C;

        int i = tiisg;
        for (; i < nr0/4; i += 32) {
            *(D4 + i) = *(C4 + i);
        }

        i = (4*(nr0/4)) + tiisg;
        for (; i < nr0; i += 32) {
            *(D + i) = *(C + i);
        }
    }
}

// Fused routed gate+up grouped matmul with the SwiGLU epilogue, generic over
// the expert quant block.  Both A streams share one staged B tile per k-step;
// each output keeps the exact MMA accumulation order of the separate GEMMs,
// and the epilogue matches kernel_dsv4_moe_swiglu_weight_f16, so the fused
// result is bit-identical to the unfused path.
template<typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread half4x4 &)>
kernel void kernel_mul_mm_id_pair_swiglu_f16_impl(
        constant ds4_metal_args_mul_mm_id & args,
        constant ds4_metal_dsv4_moe_swiglu_weight_args & act,
        device const char * src0_gate,
        device const char * src0_up,
        device const char * src1,
        device const char * htpe,
        device const char * hids,
        device       char * dst_mid,
        device const char * weights,
        threadgroup  char * shmem [[threadgroup(0)]],
        uint3  tgpig[[threadgroup_position_in_grid]],
        ushort tiitg[[thread_index_in_threadgroup]],
        ushort tiisg[[thread_index_in_simdgroup]],
        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
    threadgroup half *sa_gate = (threadgroup half *)(shmem);
    threadgroup half *sa_up   = (threadgroup half *)(shmem + 4096);
    threadgroup half *sb      = (threadgroup half *)(shmem + 8192);

    constexpr int NR0 = 64;
    constexpr int NR1 = 32;
    constexpr int NK  = 32;
    constexpr int NL0 = NK/16;
    constexpr int NL1 = NK/8;

    const int im = tgpig.z;
    const int r0 = tgpig.y*NR0;
    const int r1 = tgpig.x*NR1;

    device const uint32_t * tpe_u32 = (device const uint32_t *) (htpe);
    device const int32_t  * ids_i32 = (device const int32_t  *) (hids);

    const int32_t neh1 = tpe_u32[im];

    if (r1 >= neh1) {
        return;
    }

    const short nr0 = (args.ne0 - r0 < NR0) ? (args.ne0 - r0) : NR0;
    const short nr1 = (    neh1 - r1 < NR1) ? (    neh1 - r1) : NR1;

    const short lr0 = ((short)tiitg/NL0) < nr0 ? ((short)tiitg/NL0) : nr0 - 1;
    const short lr1 = ((short)tiitg/NL1) < nr1 ? ((short)tiitg/NL1) : nr1 - 1;

    const short il0 = (tiitg % NL0);
    short il = il0;

    const int id = ids_i32[im*args.ne21 + r1 + lr1];

    const short i11 = (id % args.ne20) % args.ne11;
    const short i12 = (id / args.ne20);
    const short i13 = 0;

    const uint64_t offset0 = im*args.nb02 + i13*args.nb03;
    const short    offset1 = il0/nl;

    device const block_q * xg =
        (device const block_q *)(src0_gate + args.nb01*(r0 + lr0) + offset0) + offset1;
    device const block_q * xu =
        (device const block_q *)(src0_up + args.nb01*(r0 + lr0) + offset0) + offset1;

    const short iy = 8*(tiitg % NL1);

    device const float * y = (device const float *)(src1
        + args.nb13*i13
        + args.nb12*i12
        + args.nb11*i11
        + args.nb10*iy);

    simdgroup_half8x8 ma_g[4];
    simdgroup_half8x8 ma_u[4];
    simdgroup_half8x8 mb[2];

    simdgroup_float8x8 mc_gate[8];
    simdgroup_float8x8 mc_up[8];

    for (short i = 0; i < 8; i++) {
        mc_gate[i] = make_filled_simdgroup_matrix<float, 8>(0.f);
        mc_up[i] = make_filled_simdgroup_matrix<float, 8>(0.f);
    }

    for (int loop_k = 0; loop_k < args.ne00; loop_k += NK) {
        half4x4 temp_gate;
        dequantize_func(xg, il, temp_gate);
        half4x4 temp_up;
        dequantize_func(xu, il, temp_up);

        threadgroup_barrier(mem_flags::mem_threadgroup);

        // Stage the shared B tile and both expert A tiles in one pass: two
        // threadgroup barriers per k-step total instead of two per matrix.
        {
            const short sx_b = (tiitg%NL1);
            const short sy_b = (tiitg/NL1)/8;
            const short ly_b = (tiitg/NL1)%8;
            const short ib_b = 4*sx_b + sy_b;
            *(threadgroup half2x4 *)(sb + 64*ib_b + 8*ly_b) =
                (half2x4)(*((device float2x4 *) y));
        }

        FOR_UNROLL (short i = 0; i < 16; i++) {
            const short sx = 2*il0 + i/8;
            const short sy = (tiitg/NL0)/8;
            const short lx = (tiitg/NL0)%8;
            const short ly = i%8;
            const short ib = 8*sx + sy;
            *(sa_gate + 64*ib + 8*ly + lx) = temp_gate[i/4][i%4];
            *(sa_up   + 64*ib + 8*ly + lx) = temp_up[i/4][i%4];
        }

        threadgroup_barrier(mem_flags::mem_threadgroup);

        threadgroup const half * lsma_gate = (sa_gate + 4*64*(sgitg%2));
        threadgroup const half * lsma_up   = (sa_up   + 4*64*(sgitg%2));
        threadgroup const half * lsmb      = (sb      + 2*64*(sgitg/2));

        FOR_UNROLL (short ik = 0; ik < NK/8; ik++) {
            simdgroup_barrier(mem_flags::mem_none);

            FOR_UNROLL (short i = 0; i < 4; i++) {
                simdgroup_load(ma_g[i], lsma_gate + 64*i, 8, 0, false);
                simdgroup_load(ma_u[i], lsma_up   + 64*i, 8, 0, false);
            }
            FOR_UNROLL (short i = 0; i < 2; i++) {
                simdgroup_load(mb[i], lsmb + 64*i, 8, 0, false);
            }

            simdgroup_barrier(mem_flags::mem_none);

            FOR_UNROLL (short i = 0; i < 8; i++) {
                simdgroup_multiply_accumulate(mc_gate[i], mb[i/4], ma_g[i%4], mc_gate[i]);
                simdgroup_multiply_accumulate(mc_up[i],   mb[i/4], ma_u[i%4], mc_up[i]);
            }

            lsma_gate += 8*64;
            lsma_up   += 8*64;
            lsmb += 4*64;
        }

        il = (il + 2 < nl) ? il + 2 : il % 2;
        xg = (il < 2) ? xg + (2 + nl - 1)/nl : xg;
        xu = (il < 2) ? xu + (2 + nl - 1)/nl : xu;
        y += NK;
    }

    threadgroup_barrier(mem_flags::mem_threadgroup);

    threadgroup float * temp_gate = (threadgroup float *) shmem;
    threadgroup float * temp_up = temp_gate + NR0*NR1;
    threadgroup float * temp_gate_str =
        temp_gate + 32*(sgitg&1) + (16*(sgitg >> 1))*NR0;
    threadgroup float * temp_up_str =
        temp_up + 32*(sgitg&1) + (16*(sgitg >> 1))*NR0;

    for (short i = 0; i < 8; i++) {
        simdgroup_store(mc_gate[i], temp_gate_str + 8*(i%4) + 8*NR0*(i/4), NR0, 0, false);
        simdgroup_store(mc_up[i],   temp_up_str   + 8*(i%4) + 8*NR0*(i/4), NR0, 0, false);
    }

    threadgroup_barrier(mem_flags::mem_threadgroup);

    const float c = act.clamp_value;
    for (short j = sgitg; j < nr1; j += 4) {
        const int idj = ids_i32[im*args.ne21 + r1 + j];

        const short ide = idj % args.ne20;
        const short idt = idj / args.ne20;

        device half *D = (device half *)(dst_mid +
            ((uint64_t)idt*args.ne1 + (uint64_t)ide)*act.mid_row_stride) + r0;
        device const float *w = (device const float *)(weights + (uint64_t)idj*act.weight_stride);
        const float route_weight = w[0];

        threadgroup float *Cg = temp_gate + j*NR0;
        threadgroup float *Cu = temp_up   + j*NR0;

        int i = tiisg;
        for (; i < nr0; i += 32) {
            float g = Cg[i];
            float u = Cu[i];
            if (c > 1.0e-6f) {
                g = min(g, c);
                u = clamp(u, -c, c);
            }
            const float silu = g / (1.0f + exp(-g));
            D[i] = (half)(silu * u * route_weight);
        }
    }
}

typedef decltype(kernel_mul_mm_id_pair_swiglu_f16_impl<block_iq2_xxs, QK_NL, dequantize_iq2_xxs>) mul_mm_id_pair_swiglu_f16_iq2;
typedef decltype(kernel_mul_mm_id_pair_swiglu_f16_impl<block_q4_K, QK_NL, dequantize_q4_K>) mul_mm_id_pair_swiglu_f16_q4;
typedef decltype(kernel_mul_mm_id_pair_swiglu_f16_impl<block_q5_K, QK_NL, dequantize_q5_K>) mul_mm_id_pair_swiglu_f16_q5;

// Host-visible fused routed pair matmuls for the DS4 expert quant formats.
template [[host_name("kernel_mul_mm_id_iq2_xxs_pair_swiglu_f16")]] kernel mul_mm_id_pair_swiglu_f16_iq2 kernel_mul_mm_id_pair_swiglu_f16_impl<block_iq2_xxs, QK_NL, dequantize_iq2_xxs>;
template [[host_name("kernel_mul_mm_id_q4_K_pair_swiglu_f16")]] kernel mul_mm_id_pair_swiglu_f16_q4 kernel_mul_mm_id_pair_swiglu_f16_impl<block_q4_K, QK_NL, dequantize_q4_K>;
template [[host_name("kernel_mul_mm_id_q5_K_pair_swiglu_f16")]] kernel mul_mm_id_pair_swiglu_f16_q5 kernel_mul_mm_id_pair_swiglu_f16_impl<block_q5_K, QK_NL, dequantize_q5_K>;

typedef decltype(kernel_mul_mm_id<32, half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_q2_K, QK_NL, dequantize_q2_K, float, float4x4, float, float2x4>) mul_mm_id;
typedef decltype(kernel_mul_mm_id<32, half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_q2_K, QK_NL, dequantize_q2_K, half, half4x4, half, half2x4>) mul_mm_id_f16_rhs;
typedef decltype(kernel_mul_mm_id<32, float, float4x4, simdgroup_float8x8, float, float2x4, simdgroup_float8x8, block_q2_K, QK_NL, dequantize_q2_K, float, float4x4, float, float2x4>) mul_mm_id_ff32;
typedef decltype(kernel_mul_mm_id_addr<32, half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_q2_K, QK_NL, dequantize_q2_K, float, float4x4, float, float2x4>) mul_mm_id_addr;
typedef decltype(kernel_mul_mm_id_addr<32, half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_q2_K, QK_NL, dequantize_q2_K, half, half4x4, half, half2x4>) mul_mm_id_addr_f16_rhs;

// Host-visible batched MoE matmul variants for the DS4 quant formats.
template [[host_name("kernel_mul_mm_id_q8_0_f32")]]         kernel mul_mm_id kernel_mul_mm_id<32, half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_q8_0,    2,     dequantize_q8_0,    float, float4x4, float, float2x4>;
template [[host_name("kernel_mul_mm_id_q2_K_f32")]]         kernel mul_mm_id kernel_mul_mm_id<32, half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_q2_K,    QK_NL, dequantize_q2_K,    float, float4x4, float, float2x4>;
template [[host_name("kernel_mul_mm_id_q4_K_f32")]]         kernel mul_mm_id kernel_mul_mm_id<32, half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_q4_K,    QK_NL, dequantize_q4_K,    float, float4x4, float, float2x4>;
template [[host_name("kernel_mul_mm_id_q5_K_f32")]]         kernel mul_mm_id kernel_mul_mm_id<32, half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_q5_K,    QK_NL, dequantize_q5_K,    float, float4x4, float, float2x4>;
template [[host_name("kernel_mul_mm_id_q6_K_f32")]]         kernel mul_mm_id kernel_mul_mm_id<32, half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_q6_K,    QK_NL, dequantize_q6_K,    float, float4x4, float, float2x4>;
template [[host_name("kernel_mul_mm_id_iq2_xxs_f32")]]      kernel mul_mm_id kernel_mul_mm_id<32, half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_iq2_xxs, QK_NL, dequantize_iq2_xxs, float, float4x4, float, float2x4>;
template [[host_name("kernel_mul_mm_id_q8_0_f16")]]         kernel mul_mm_id_f16_rhs kernel_mul_mm_id<32, half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_q8_0,    2,     dequantize_q8_0,    half, half4x4, half, half2x4>;
template [[host_name("kernel_mul_mm_id_q2_K_f16")]]         kernel mul_mm_id_f16_rhs kernel_mul_mm_id<32, half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_q2_K,    QK_NL, dequantize_q2_K,    half, half4x4, half, half2x4>;
template [[host_name("kernel_mul_mm_id_q4_K_f16")]]         kernel mul_mm_id_f16_rhs kernel_mul_mm_id<32, half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_q4_K,    QK_NL, dequantize_q4_K,    half, half4x4, half, half2x4>;
template [[host_name("kernel_mul_mm_id_q5_K_f16")]]         kernel mul_mm_id_f16_rhs kernel_mul_mm_id<32, half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_q5_K,    QK_NL, dequantize_q5_K,    half, half4x4, half, half2x4>;
template [[host_name("kernel_mul_mm_id_q6_K_f16")]]         kernel mul_mm_id_f16_rhs kernel_mul_mm_id<32, half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_q6_K,    QK_NL, dequantize_q6_K,    half, half4x4, half, half2x4>;
template [[host_name("kernel_mul_mm_id_iq2_xxs_f16")]]      kernel mul_mm_id_f16_rhs kernel_mul_mm_id<32, half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_iq2_xxs, QK_NL, dequantize_iq2_xxs, half, half4x4, half, half2x4>;
template [[host_name("kernel_mul_mm_id_q4_K_ff32")]]        kernel mul_mm_id_ff32 kernel_mul_mm_id<32, float, float4x4, simdgroup_float8x8, float, float2x4, simdgroup_float8x8, block_q4_K, QK_NL, dequantize_q4_K, float, float4x4, float, float2x4>;
template [[host_name("kernel_mul_mm_id_q5_K_ff32")]]        kernel mul_mm_id_ff32 kernel_mul_mm_id<32, float, float4x4, simdgroup_float8x8, float, float2x4, simdgroup_float8x8, block_q5_K, QK_NL, dequantize_q5_K, float, float4x4, float, float2x4>;
template [[host_name("kernel_mul_mm_id_q6_K_ff32")]]        kernel mul_mm_id_ff32 kernel_mul_mm_id<32, float, float4x4, simdgroup_float8x8, float, float2x4, simdgroup_float8x8, block_q6_K, QK_NL, dequantize_q6_K, float, float4x4, float, float2x4>;

template [[host_name("kernel_mul_mm_id_addr_q2_K_f32")]]    kernel mul_mm_id_addr kernel_mul_mm_id_addr<32, half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_q2_K, QK_NL, dequantize_q2_K, float, float4x4, float, float2x4>;
template [[host_name("kernel_mul_mm_id_addr_q4_K_f32")]]    kernel mul_mm_id_addr kernel_mul_mm_id_addr<32, half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_q4_K, QK_NL, dequantize_q4_K, float, float4x4, float, float2x4>;
template [[host_name("kernel_mul_mm_id_addr_q2_K_f16")]]    kernel mul_mm_id_addr_f16_rhs kernel_mul_mm_id_addr<32, half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_q2_K, QK_NL, dequantize_q2_K, half, half4x4, half, half2x4>;
template [[host_name("kernel_mul_mm_id_addr_q4_K_f16")]]    kernel mul_mm_id_addr_f16_rhs kernel_mul_mm_id_addr<32, half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_q4_K, QK_NL, dequantize_q4_K, half, half4x4, half, half2x4>;

#ifdef DS4_METAL_HAS_TENSOR
// Attention-output low-rank projection retained for Metal4 prefill.  It uses
// the same direct-RHS idea as dense matmul: dequantize the Q8_0 low projection
// weights to a half tile, then let TensorOps read the dense head activations
// directly.  Only the 64-token direct-RHS instantiation is exported because the
// staged-RHS and 32-token variants were benchmark-only experiments.
//
// Full tiles (the host dispatch guarantee for aligned batches) skip all bounds
// work.  The weight tile is double-buffered: the next k-step's dequant
// overlaps the current cooperative matmul, so the k-loop needs one
// threadgroup barrier per step instead of two.
template<short NR1>
kernel void kernel_attn_out_low_q8_0_mpp_direct_rhs(
        constant ds4_metal_args_mul_mm_id & args,
        device const char * srcA,
        device const char * srcB,
        device       char * dst,
        threadgroup  char * shmem [[threadgroup(0)]],
        uint3  tgpig [[threadgroup_position_in_grid]],
        ushort tiitg [[thread_index_in_threadgroup]],
        ushort sgitg [[simdgroup_index_in_threadgroup]]) {
    (void) sgitg;

    constexpr int NR0 = 64;
    constexpr int NK  = 32;
    constexpr int NL  = NK/16;
    constexpr int NUM_THREADS = 128;

    const int K = args.ne00;
    const int M = args.ne0;
    const int N = args.ne21;
    const int G = args.ne1;
    const int group = tgpig.z;
    const int r0 = tgpig.y*NR0;
    const int r1 = tgpig.x*NR1;
    const bool full_tile = r0 + NR0 <= M && r1 + NR1 <= N && (K % NK) == 0;

    threadgroup half *sa = (threadgroup half *)shmem;
    auto tA0 = tensor(sa,          dextents<int32_t, 2>(NK, NR0));
    auto tA1 = tensor(sa + NR0*NK, dextents<int32_t, 2>(NK, NR0));

    device float *ptrB = (device float *)(srcB + args.nb11*group);
    const int strideB = args.nb12/sizeof(float);
    auto tB = tensor(ptrB, dextents<int32_t, 2>(K, N), array<int, 2>({1, strideB}));

    matmul2d<
        matmul2d_descriptor(NR1, NR0, NK, false, true, true,
            matmul2d_descriptor::mode::multiply_accumulate),
        execution_simdgroups<4>> mm;

    auto cT = mm.template get_destination_cooperative_tensor<decltype(tB), decltype(tA0), float>();

    #pragma unroll
    for (uint16_t i = 0; i < cT.get_capacity(); ++i) {
        if (cT.is_valid_element(i)) {
            cT[i] = 0.0f;
        }
    }

    auto stage_tile = [&](const int loop_k, threadgroup half *buf) {
        for (int work = tiitg; work < NR0*NL; work += NUM_THREADS) {
            const int row = work/NL;
            const int k_chunk = work%NL;
            const int k_pos = loop_k + k_chunk*16;
            const short k_base = k_chunk*16;

            if (full_tile) {
                device const block_q8_0 *row_ptr =
                    (device const block_q8_0 *)(srcA + args.nb01*(r0 + row) + group*args.nb02);

                half4x4 temp_a;
                dequantize_q8_0_pairs(row_ptr + k_pos/32, (k_pos/16)%2, temp_a);
                threadgroup half4 *dst4 = (threadgroup half4 *)(buf + row*NK + k_base);
                dst4[0] = temp_a[0];
                dst4[1] = temp_a[1];
                dst4[2] = temp_a[2];
                dst4[3] = temp_a[3];
            } else if (r0 + row < M) {
                const int block_idx = k_pos/32;
                const short il = (k_pos/16)%2;
                device const block_q8_0 *row_ptr =
                    (device const block_q8_0 *)(srcA + args.nb01*(r0 + row) + group*args.nb02);

                half4x4 temp_a;
                dequantize_q8_0_pairs(row_ptr + block_idx, il, temp_a);
                FOR_UNROLL (short i = 0; i < 16; i++) {
                    buf[row*NK + k_base + i] = (k_pos + i < K) ? temp_a[i/4][i%4] : (half)0;
                }
            } else {
                FOR_UNROLL (short i = 0; i < 16; i++) {
                    buf[row*NK + k_base + i] = (half)0;
                }
            }
        }
    };

    stage_tile(0, sa);
    threadgroup_barrier(mem_flags::mem_threadgroup);

    uint buf_sel = 0;
    for (int loop_k = 0; loop_k < K; loop_k += NK) {
        auto mA = (buf_sel ? tA1 : tA0).slice(0, 0);
        auto mB = tB.slice(loop_k, r1);
        mm.run(mB, mA, cT);

        const int next_k = loop_k + NK;
        if (next_k < K) {
            buf_sel ^= 1u;
            stage_tile(next_k, buf_sel ? sa + NR0*NK : sa);
        }
        threadgroup_barrier(mem_flags::mem_threadgroup);
    }

    device float *dst_group = (device float *)dst + group*M;
    if (full_tile) {
        device float *dst_tile = dst_group + r0 + (uint64_t)r1*G*M;
        auto tD = tensor(dst_tile, dextents<int32_t, 2>(NR0, NR1), array<int, 2>({1, G*M}));
        cT.store(tD);
    } else {
        auto tD = tensor(dst_group, dextents<int32_t, 2>(M, N), array<int, 2>({1, G*M}));
        auto mD = tD.slice(r0, r1);
        cT.store(mD);
    }
}

typedef decltype(kernel_attn_out_low_q8_0_mpp_direct_rhs<64>) attn_out_low_q8_0_mpp_direct_rhs_n64_t;

template [[host_name("kernel_attn_out_low_q8_0_mpp_direct_rhs_n64")]] kernel attn_out_low_q8_0_mpp_direct_rhs_n64_t kernel_attn_out_low_q8_0_mpp_direct_rhs<64>;

#endif

#undef QK_NL
#undef kmask_iq2xs
#undef ksigns_iq2xs
#undef iq2xxs_grid
#undef QK_K
#undef N_R0_Q2_K
#undef N_R0_Q4_K
#undef N_R0_Q8_K
#undef N_R0_GLM_Q4_PAIR2_K
#undef N_R0_GLM_Q4_PAIR_K
#undef N_R0_Q5_PAIR_K
#undef N_R0_Q5_K
#undef N_R0_Q6_K
#undef N_R0_IQ2_XXS