node-llama-cpp 3.0.0-beta.11 → 3.0.0-beta.12

package/llamaBins/mac-arm64-metal/ggml-metal.metal

@@ -392,7 +392,7 @@ kernel void kernel_soft_max(
     float lmax = -INFINITY;
 
     for (int i00 = tpitg; i00 < ne00; i00 += ntg) {
-        lmax = MAX(lmax, psrc0[i00]*scale + (pmask ? pmask[i00] : 0.0f) + slope*ppos[i00]);
+        lmax = MAX(lmax, psrc0[i00]*scale + (pmask ? pmask[i00] : 0.0f) + (ppos ? slope*ppos[i00] : 0.0f));
     }
 
     // find the max value in the block
@@ -417,7 +417,7 @@ kernel void kernel_soft_max(
     // parallel sum
     float lsum = 0.0f;
     for (int i00 = tpitg; i00 < ne00; i00 += ntg) {
-        const float exp_psrc0 = exp((psrc0[i00]*scale + (pmask ? pmask[i00] : 0.0f) + slope*ppos[i00]) - max_val);
+        const float exp_psrc0 = exp((psrc0[i00]*scale + (pmask ? pmask[i00] : 0.0f) + (ppos ? slope*ppos[i00] : 0.0f)) - max_val);
         lsum += exp_psrc0;
         pdst[i00] = exp_psrc0;
     }
@@ -495,7 +495,7 @@ kernel void kernel_soft_max_4(
     float4 lmax4 = -INFINITY;
 
     for (int i00 = tpitg; i00 < ne00/4; i00 += ntg) {
-        lmax4 = fmax(lmax4, psrc4[i00]*scale + (pmask ? pmask[i00] : 0.0f) + slope*ppos[i00]);
+        lmax4 = fmax(lmax4, psrc4[i00]*scale + (pmask ? pmask[i00] : 0.0f) + (ppos ? slope*ppos[i00] : 0.0f));
     }
 
     const float lmax = MAX(MAX(lmax4[0], lmax4[1]), MAX(lmax4[2], lmax4[3]));
@@ -521,7 +521,7 @@ kernel void kernel_soft_max_4(
     // parallel sum
     float4 lsum4 = 0.0f;
     for (int i00 = tpitg; i00 < ne00/4; i00 += ntg) {
-        const float4 exp_psrc4 = exp((psrc4[i00]*scale + (pmask ? pmask[i00] : 0.0f) + slope*ppos[i00]) - max_val);
+        const float4 exp_psrc4 = exp((psrc4[i00]*scale + (pmask ? pmask[i00] : 0.0f) + (ppos ? slope*ppos[i00] : 0.0f)) - max_val);
         lsum4 += exp_psrc4;
         pdst4[i00] = exp_psrc4;
     }
@@ -2525,12 +2525,32 @@ typedef struct {
 } block_iq3_xxs;
 // 98 bytes / block for QK_K = 256, so 3.0625 bpw
 
+// 3.4375 bpw
+#if QK_K == 64
+#define IQ3S_N_SCALE 2
+#else
+#define IQ3S_N_SCALE QK_K/64
+#endif
+typedef struct {
+    half d;
+    uint8_t qs[QK_K/4];
+    uint8_t qh[QK_K/32];
+    uint8_t signs[QK_K/8];
+    uint8_t scales[IQ3S_N_SCALE];
+} block_iq3_s;
+
 typedef struct {
     half d;
     uint8_t qs[QK_K/8];
     uint8_t scales[QK_K/16];
 } block_iq1_s;
 
+// Non-linear quants
+#define QK4_NL 32
+typedef struct {
+    half    d;
+    uint8_t qs[QK4_NL/2];
+} block_iq4_nl;
 
 //====================================== dot products =========================
 
@@ -3789,6 +3809,73 @@ constexpr constant static uint32_t iq3xxs_grid[256] = {
     0x3e1c1c1c, 0x3e1c3404, 0x3e24140c, 0x3e24240c, 0x3e2c0404, 0x3e2c0414, 0x3e2c1424, 0x3e341c04,
 };
 
+constexpr constant static uint32_t iq3xs_grid[512] = {
+    0x04040404, 0x0404040c, 0x04040414, 0x0404042c, 0x0404043e, 0x04040c04, 0x04040c0c, 0x04040c14,
+    0x04040c24, 0x04040c34, 0x04041404, 0x0404140c, 0x0404142c, 0x04041c1c, 0x04042404, 0x04042414,
+    0x0404242c, 0x0404243e, 0x04042c0c, 0x04042c1c, 0x04043404, 0x04043414, 0x04043e0c, 0x04043e24,
+    0x04043e3e, 0x040c0404, 0x040c040c, 0x040c0414, 0x040c0424, 0x040c0c04, 0x040c0c0c, 0x040c0c2c,
+    0x040c1404, 0x040c141c, 0x040c143e, 0x040c1c0c, 0x040c1c2c, 0x040c2424, 0x040c340c, 0x040c342c,
+    0x040c3e14, 0x04140404, 0x0414040c, 0x0414042c, 0x0414043e, 0x04140c04, 0x04140c1c, 0x04140c34,
+    0x0414140c, 0x0414142c, 0x04141c04, 0x04141c24, 0x04142414, 0x0414242c, 0x0414243e, 0x04142c0c,
+    0x04142c1c, 0x04143e04, 0x04143e1c, 0x041c041c, 0x041c0c0c, 0x041c0c2c, 0x041c1404, 0x041c1414,
+    0x041c1c0c, 0x041c1c1c, 0x041c1c34, 0x041c2424, 0x041c2c04, 0x041c2c14, 0x041c343e, 0x041c3e0c,
+    0x041c3e2c, 0x04240404, 0x04240c1c, 0x04240c3e, 0x0424140c, 0x04241424, 0x04241c14, 0x04242404,
+    0x0424241c, 0x04242c0c, 0x04243e04, 0x042c0414, 0x042c0424, 0x042c1404, 0x042c1414, 0x042c1434,
+    0x042c1c1c, 0x042c240c, 0x042c242c, 0x042c243e, 0x042c3434, 0x042c3e1c, 0x04340434, 0x04340c0c,
+    0x04340c1c, 0x04341c0c, 0x04342c14, 0x04343e0c, 0x043e0404, 0x043e0414, 0x043e0424, 0x043e1404,
+    0x043e1414, 0x043e1434, 0x043e1c1c, 0x043e2c04, 0x043e2c24, 0x0c040404, 0x0c04040c, 0x0c040414,
+    0x0c040424, 0x0c040c04, 0x0c040c0c, 0x0c040c1c, 0x0c040c2c, 0x0c040c3e, 0x0c041404, 0x0c041414,
+    0x0c041c0c, 0x0c041c24, 0x0c041c34, 0x0c042c24, 0x0c042c34, 0x0c04340c, 0x0c043e14, 0x0c0c0404,
+    0x0c0c040c, 0x0c0c041c, 0x0c0c0434, 0x0c0c0c04, 0x0c0c0c24, 0x0c0c140c, 0x0c0c1c04, 0x0c0c1c1c,
+    0x0c0c240c, 0x0c0c2c04, 0x0c0c2c14, 0x0c0c3e04, 0x0c0c3e34, 0x0c140404, 0x0c140c14, 0x0c140c2c,
+    0x0c140c3e, 0x0c141404, 0x0c141424, 0x0c141c14, 0x0c142404, 0x0c14241c, 0x0c142c2c, 0x0c143404,
+    0x0c143e14, 0x0c1c040c, 0x0c1c0424, 0x0c1c043e, 0x0c1c0c04, 0x0c1c0c1c, 0x0c1c140c, 0x0c1c143e,
+    0x0c1c1c04, 0x0c1c1c24, 0x0c1c240c, 0x0c1c3414, 0x0c1c3e04, 0x0c24041c, 0x0c24042c, 0x0c240c14,
+    0x0c240c24, 0x0c241c0c, 0x0c241c1c, 0x0c242414, 0x0c242434, 0x0c242c04, 0x0c242c24, 0x0c2c040c,
+    0x0c2c0c04, 0x0c2c0c1c, 0x0c2c140c, 0x0c2c1c04, 0x0c2c1c14, 0x0c2c2c0c, 0x0c341404, 0x0c341424,
+    0x0c34143e, 0x0c342424, 0x0c342434, 0x0c3e040c, 0x0c3e041c, 0x0c3e0c04, 0x0c3e0c14, 0x0c3e140c,
+    0x0c3e1c2c, 0x0c3e240c, 0x0c3e3414, 0x0c3e3e04, 0x14040404, 0x1404040c, 0x1404041c, 0x1404042c,
+    0x1404043e, 0x14040c04, 0x14040c14, 0x14040c24, 0x14040c34, 0x1404140c, 0x1404141c, 0x1404143e,
+    0x14041c04, 0x14041c14, 0x1404240c, 0x1404241c, 0x1404242c, 0x14042c04, 0x14042c14, 0x1404343e,
+    0x14043e04, 0x14043e1c, 0x14043e2c, 0x140c0404, 0x140c0414, 0x140c0c04, 0x140c0c1c, 0x140c0c3e,
+    0x140c1414, 0x140c142c, 0x140c1c0c, 0x140c1c24, 0x140c2414, 0x140c2c0c, 0x1414040c, 0x14140424,
+    0x1414043e, 0x1414140c, 0x1414141c, 0x14141c04, 0x14141c3e, 0x1414240c, 0x14142c1c, 0x14142c3e,
+    0x14143e0c, 0x14143e24, 0x141c0404, 0x141c0414, 0x141c042c, 0x141c0c0c, 0x141c1414, 0x141c1424,
+    0x141c1c0c, 0x141c1c1c, 0x141c2414, 0x141c2c04, 0x141c3434, 0x1424040c, 0x1424043e, 0x14241404,
+    0x1424141c, 0x14241c14, 0x14241c2c, 0x1424240c, 0x14243e14, 0x14243e2c, 0x142c0424, 0x142c0c0c,
+    0x142c1414, 0x142c1c3e, 0x142c2404, 0x142c2c1c, 0x142c3e04, 0x14340404, 0x14340414, 0x1434043e,
+    0x1434140c, 0x14342c2c, 0x1434340c, 0x143e042c, 0x143e0c0c, 0x143e1434, 0x143e1c04, 0x143e241c,
+    0x143e2c04, 0x1c040414, 0x1c040c0c, 0x1c040c1c, 0x1c040c2c, 0x1c040c3e, 0x1c041414, 0x1c041c0c,
+    0x1c041c1c, 0x1c041c2c, 0x1c042414, 0x1c042424, 0x1c04243e, 0x1c042c0c, 0x1c04341c, 0x1c043e0c,
+    0x1c0c040c, 0x1c0c041c, 0x1c0c042c, 0x1c0c0c24, 0x1c0c140c, 0x1c0c141c, 0x1c0c2404, 0x1c0c3404,
+    0x1c0c3e14, 0x1c0c3e34, 0x1c140404, 0x1c140c14, 0x1c141404, 0x1c141c14, 0x1c141c24, 0x1c142c04,
+    0x1c1c040c, 0x1c1c0c04, 0x1c1c0c24, 0x1c1c140c, 0x1c1c141c, 0x1c1c143e, 0x1c1c1c04, 0x1c1c240c,
+    0x1c1c241c, 0x1c1c243e, 0x1c1c2c2c, 0x1c1c3e1c, 0x1c24041c, 0x1c240c0c, 0x1c240c34, 0x1c241414,
+    0x1c241c0c, 0x1c242c14, 0x1c243404, 0x1c243424, 0x1c2c040c, 0x1c2c0c04, 0x1c2c0c14, 0x1c2c142c,
+    0x1c2c1c14, 0x1c2c2424, 0x1c2c2c34, 0x1c2c3e1c, 0x1c340c34, 0x1c34240c, 0x1c3e040c, 0x1c3e041c,
+    0x1c3e1404, 0x1c3e1414, 0x1c3e1c2c, 0x24040404, 0x24040424, 0x24040c14, 0x24041404, 0x24041424,
+    0x2404143e, 0x24041c14, 0x2404240c, 0x24042c04, 0x24043e04, 0x240c0414, 0x240c043e, 0x240c0c0c,
+    0x240c0c1c, 0x240c1414, 0x240c1c04, 0x240c1c2c, 0x240c241c, 0x240c2c0c, 0x240c2c2c, 0x2414040c,
+    0x2414041c, 0x24140c04, 0x24140c2c, 0x2414140c, 0x24141c1c, 0x24142404, 0x24142c3e, 0x24143414,
+    0x24143e04, 0x241c0424, 0x241c0c0c, 0x241c0c1c, 0x241c1404, 0x241c1414, 0x241c1c0c, 0x241c1c2c,
+    0x24240404, 0x24240414, 0x24241424, 0x24241c3e, 0x24242404, 0x24243e0c, 0x242c042c, 0x242c043e,
+    0x242c140c, 0x242c3414, 0x24340c1c, 0x24341c24, 0x24343404, 0x243e0c04, 0x243e0c2c, 0x243e1c04,
+    0x243e241c, 0x243e2c0c, 0x2c040414, 0x2c040c04, 0x2c040c24, 0x2c041414, 0x2c042404, 0x2c042424,
+    0x2c04243e, 0x2c042c14, 0x2c043434, 0x2c043e24, 0x2c0c040c, 0x2c0c041c, 0x2c0c042c, 0x2c0c0c14,
+    0x2c0c140c, 0x2c0c1c14, 0x2c0c3e14, 0x2c140404, 0x2c140c0c, 0x2c14141c, 0x2c141c04, 0x2c141c34,
+    0x2c142c1c, 0x2c1c0414, 0x2c1c043e, 0x2c1c0c04, 0x2c1c143e, 0x2c1c2424, 0x2c1c2c0c, 0x2c1c342c,
+    0x2c1c3e1c, 0x2c24040c, 0x2c240424, 0x2c241404, 0x2c241c14, 0x2c242434, 0x2c2c0c14, 0x2c2c1434,
+    0x2c2c2c0c, 0x2c2c2c1c, 0x2c342414, 0x2c3e0414, 0x2c3e0424, 0x2c3e1414, 0x34040c0c, 0x34040c1c,
+    0x34040c2c, 0x34041c0c, 0x34041c1c, 0x34043404, 0x340c0404, 0x340c1404, 0x340c143e, 0x340c3424,
+    0x34140c14, 0x34141c24, 0x34142414, 0x34142c2c, 0x34143414, 0x34143e04, 0x341c0404, 0x341c0c24,
+    0x341c140c, 0x341c2404, 0x3424142c, 0x3424241c, 0x34243414, 0x342c0404, 0x342c041c, 0x342c1c24,
+    0x342c3404, 0x3434042c, 0x34342404, 0x343e0c0c, 0x343e0c1c, 0x3e040404, 0x3e040424, 0x3e04043e,
+    0x3e041404, 0x3e041414, 0x3e041c34, 0x3e042404, 0x3e042c24, 0x3e043414, 0x3e0c0414, 0x3e0c0c0c,
+    0x3e0c1424, 0x3e0c241c, 0x3e0c242c, 0x3e14040c, 0x3e140424, 0x3e140c04, 0x3e140c34, 0x3e14140c,
+    0x3e141c04, 0x3e142c0c, 0x3e1c0414, 0x3e1c1c14, 0x3e1c1c2c, 0x3e1c2c1c, 0x3e24040c, 0x3e24042c,
+    0x3e240c1c, 0x3e241404, 0x3e242c04, 0x3e2c1414, 0x3e2c2414, 0x3e340414, 0x3e341c0c, 0x3e3e0404,
+};
+
 #define NGRID_IQ1S 512
 constexpr constant static uint64_t iq1s_grid[NGRID_IQ1S] = {
     0xffffffffffff0101, 0xffffffffff01ff00, 0xffffffffff010100, 0xffffffff00000000,
@@ -4027,7 +4114,10 @@ void kernel_mul_mv_iq2_xxs_f32_impl(
         y4 += 32 * 32;
     }
 #else
-    // TODO
+    (void) x;
+    (void) y;
+    (void) yl;
+    (void) nb32;
 #endif
 
     for (int row = 0; row < N_DST; ++row) {
@@ -4170,7 +4260,10 @@ void kernel_mul_mv_iq2_xs_f32_impl(
         y4 += 32 * 32;
     }
 #else
-    // TODO
+    (void) x;
+    (void) y;
+    (void) yl;
+    (void) nb32;
 #endif
 
     for (int row = 0; row < N_DST; ++row) {
@@ -4306,7 +4399,10 @@ void kernel_mul_mv_iq3_xxs_f32_impl(
         y4 += 32 * 32;
     }
 #else
-    // TODO
+    (void) x;
+    (void) y;
+    (void) yl;
+    (void) nb32;
 #endif
 
     for (int row = 0; row < N_DST; ++row) {
@@ -4346,7 +4442,7 @@ kernel void kernel_mul_mv_iq3_xxs_f32(
     kernel_mul_mv_iq3_xxs_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, shared_values, tgpig, tiisg, sgitg);
 }
 
-void kernel_mul_mv_iq1_s_f32_impl(
+void kernel_mul_mv_iq3_s_f32_impl(
         device const  void * src0,
         device const float * src1,
         device       float * dst,
@@ -4359,6 +4455,7 @@ void kernel_mul_mv_iq1_s_f32_impl(
         constant   int64_t & ne1,
         constant   uint    & r2,
         constant   uint    & r3,
+        threadgroup int8_t * shared_values [[threadgroup(0)]],
         uint3 tgpig[[threadgroup_position_in_grid]],
         uint  tiisg[[thread_index_in_simdgroup]],
         uint  sgitg[[simdgroup_index_in_threadgroup]]) {
@@ -4376,6 +4473,134 @@ void kernel_mul_mv_iq1_s_f32_impl(
 
     const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
 
+    device const block_iq3_s * x = (device const block_iq3_s *) src0 + ib_row + offset0;
+    device const float       * y = (device const float       *) src1 + r1*ne10 + im*ne00*ne1;
+
+    float yl[32];
+    float sumf[N_DST]={0.f}, all_sum;
+
+    const int nb32 = nb * (QK_K / 32);
+
+    threadgroup uint32_t * values = (threadgroup uint32_t *)shared_values;
+    {
+        int nval = 8;
+        int pos  = (32*sgitg + tiisg)*nval;
+        for (int i = 0; i < nval; ++i) values[pos + i] = iq3xs_grid[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 (int 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_iq3_s * xr = x + ibl;
+        device const uint8_t * qs = xr->qs + 8 * ib;
+        device const uint8_t * qh = xr->qh + ib;
+        device const uint8_t * sc = xr->scales + (ib/2);
+        device const uint8_t * signs = xr->signs + 4 * ib;
+        device const half * dh = &xr->d;
+
+        for (int row = 0; row < N_DST; row++) {
+
+            const float db = dh[0];
+            const float d = db * (0.5f + ((sc[0] >> 4*(ib%2)) & 0xf));
+
+            float2 sum = {0};
+            for (int l = 0; l < 4; ++l) {
+                const threadgroup uint8_t * grid1 = (const threadgroup uint8_t *)(values + (qs[2*l+0] | ((qh[0] << (8-2*l)) & 256)));
+                const threadgroup uint8_t * grid2 = (const threadgroup uint8_t *)(values + (qs[2*l+1] | ((qh[0] << (7-2*l)) & 256)));
+                for (int j = 0; j < 4; ++j) {
+                    sum[0] += yl[8*l + j + 0] * grid1[j] * select(1, -1, signs[l] & kmask_iq2xs[j+0]);
+                    sum[1] += yl[8*l + j + 4] * grid2[j] * select(1, -1, signs[l] & kmask_iq2xs[j+4]);
+                }
+            }
+            sumf[row] += d * (sum[0] + sum[1]);
+
+            dh  += nb*sizeof(block_iq3_s)/2;
+            qs  += nb*sizeof(block_iq3_s);
+            qh  += nb*sizeof(block_iq3_s);
+            sc  += nb*sizeof(block_iq3_s);
+            signs += nb*sizeof(block_iq3_s);
+        }
+
+        y4 += 32 * 32;
+    }
+
+    for (int row = 0; row < N_DST; ++row) {
+        all_sum = simd_sum(sumf[row]);
+        if (tiisg == 0) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum * 0.5f;
+        }
+    }
+}
+
+[[host_name("kernel_mul_mv_iq3_s_f32")]]
+kernel void kernel_mul_mv_iq3_s_f32(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant  uint64_t & nb00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb02,
+        constant   int64_t & ne10,
+        constant   int64_t & ne11,
+        constant   int64_t & ne12,
+        constant  uint64_t & nb10,
+        constant  uint64_t & nb11,
+        constant  uint64_t & nb12,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   uint    & r2,
+        constant   uint    & r3,
+        threadgroup int8_t * shared_values [[threadgroup(0)]],
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
+
+    kernel_mul_mv_iq3_s_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, shared_values, tgpig, tiisg, sgitg);
+}
+
+void kernel_mul_mv_iq1_s_f32_impl(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant   int64_t & ne10,
+        constant   int64_t & ne12,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   uint    & r2,
+        constant   uint    & r3,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
+
+    const int nb = ne00/QK_K;
+    const int r0 = tgpig.x;
+    const int r1 = tgpig.y;
+    const int im = tgpig.z;
+
+    const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST;
+    const int ib_row = first_row * nb;
+
+    const uint i12 = im%ne12;
+    const uint i13 = im/ne12;
+
+    const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
     device const block_iq1_s * x = (device const block_iq1_s *) src0 + ib_row + offset0;
     device const float       * y = (device const float       *) src1 + r1*ne10 + im*ne00*ne1;
 
@@ -4424,7 +4649,10 @@ void kernel_mul_mv_iq1_s_f32_impl(
         y4 += 16 * 32;
     }
 #else
-    // TODO
+    (void) x;
+    (void) y;
+    (void) yl;
+    (void) nb32;
 #endif
 
     for (int row = 0; row < N_DST; ++row) {
@@ -4435,6 +4663,103 @@ void kernel_mul_mv_iq1_s_f32_impl(
     }
 }
 
+constexpr constant static float kvalues_iq4nl_f[16] = {
+    -127.f, -104.f, -83.f, -65.f, -49.f, -35.f, -22.f, -10.f, 1.f, 13.f, 25.f, 38.f, 53.f, 69.f, 89.f, 113.f
+};
+
+void kernel_mul_mv_iq4_nl_f32_impl(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant   int64_t & ne10,
+        constant   int64_t & ne12,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   uint    & r2,
+        constant   uint    & r3,
+        threadgroup float  * shared_values [[threadgroup(0)]],
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
+
+    const int nb = ne00/QK4_NL;
+    const int r0 = tgpig.x;
+    const int r1 = tgpig.y;
+    const int im = tgpig.z;
+    const int first_row = (r0 * 2 + sgitg) * 2;
+    const int ib_row = first_row * nb;
+
+    const uint i12 = im%ne12;
+    const uint i13 = im/ne12;
+
+    const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
+    device const block_iq4_nl * x = (device const block_iq4_nl *) src0 + ib_row + offset0;
+    device const float        * y = (device const float        *) src1 + r1*ne10 + im*ne00*ne1;
+
+    const int ix = tiisg/2;  // 0...15
+    const int it = tiisg%2;  // 0 or 1
+
+    shared_values[tiisg] = kvalues_iq4nl_f[tiisg%16];
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    float4 yl[4];
+    float sumf[2]={0.f}, all_sum;
+
+    device const float * yb = y + ix * QK4_NL + it * 8;
+
+    uint32_t aux32[2];
+    thread const uint8_t * q8 = (thread const uint8_t *)aux32;
+
+    float4 qf1, qf2;
+
+    for (int ib = ix; ib < nb; ib += 16) {
+
+        device const float4 * y4 = (device const float4 *)yb;
+        yl[0] = y4[0]; yl[1] = y4[4]; yl[2] = y4[1]; yl[3] = y4[5];
+
+        for (int row = 0; row < 2; ++row) {
+
+            device const block_iq4_nl & xb = x[row*nb + ib];
+            device const uint16_t * q4 = (device const uint16_t *)(xb.qs + 8*it);
+
+            float4 acc1 = {0.f}, acc2 = {0.f};
+
+            aux32[0] = q4[0] | (q4[1] << 16);
+            aux32[1] = (aux32[0] >> 4) & 0x0f0f0f0f;
+            aux32[0] &= 0x0f0f0f0f;
+            qf1 = {shared_values[q8[0]], shared_values[q8[1]], shared_values[q8[2]], shared_values[q8[3]]};
+            qf2 = {shared_values[q8[4]], shared_values[q8[5]], shared_values[q8[6]], shared_values[q8[7]]};
+            acc1 += yl[0] * qf1;
+            acc2 += yl[1] * qf2;
+
+            aux32[0] = q4[2] | (q4[3] << 16);
+            aux32[1] = (aux32[0] >> 4) & 0x0f0f0f0f;
+            aux32[0] &= 0x0f0f0f0f;
+            qf1 = {shared_values[q8[0]], shared_values[q8[1]], shared_values[q8[2]], shared_values[q8[3]]};
+            qf2 = {shared_values[q8[4]], shared_values[q8[5]], shared_values[q8[6]], shared_values[q8[7]]};
+            acc1 += yl[2] * qf1;
+            acc2 += yl[3] * qf2;
+
+            acc1 += acc2;
+
+            sumf[row] += (float)xb.d * (acc1[0] + acc1[1] + acc1[2] + acc1[3]);
+
+        }
+
+        yb += 16 * QK4_NL;
+    }
+
+    for (int row = 0; row < 2; ++row) {
+        all_sum = simd_sum(sumf[row]);
+        if (tiisg == 0) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum;
+        }
+    }
+}
+
 [[host_name("kernel_mul_mv_iq1_s_f32")]]
 kernel void kernel_mul_mv_iq1_s_f32(
         device const  void * src0,
@@ -4463,6 +4788,34 @@ kernel void kernel_mul_mv_iq1_s_f32(
     kernel_mul_mv_iq1_s_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, tgpig, tiisg, sgitg);
 }
 
+[[host_name("kernel_mul_mv_iq4_nl_f32")]]
+kernel void kernel_mul_mv_iq4_nl_f32(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant  uint64_t & nb00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb02,
+        constant   int64_t & ne10,
+        constant   int64_t & ne11,
+        constant   int64_t & ne12,
+        constant  uint64_t & nb10,
+        constant  uint64_t & nb11,
+        constant  uint64_t & nb12,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   uint    & r2,
+        constant   uint    & r3,
+        threadgroup float * shared_values [[threadgroup(0)]],
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint tiisg[[thread_index_in_simdgroup]],
+        uint sgitg[[simdgroup_index_in_threadgroup]]) {
+
+    kernel_mul_mv_iq4_nl_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, shared_values, tgpig, tiisg, sgitg);
+}
 
 //============================= templates and their specializations =============================
 
@@ -4659,6 +5012,8 @@ void dequantize_q4_K(device const block_q4_K *xb, short il, thread type4x4 & reg
     const float dl = d * sc[0];
     const float ml = min * sc[1];
 #else
+    (void) get_scale_min_k4_just2;
+
     q = q + 16 * (il&1);
     device const uint8_t * s = xb->scales;
     device const half2 * dh = (device const half2 *)xb->d;
@@ -4808,6 +5163,31 @@ void dequantize_iq3_xxs(device const block_iq3_xxs * xb, short il, thread type4x
     }
 }
 
+template <typename type4x4>
+void dequantize_iq3_s(device const block_iq3_s * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const float d = xb->d;
+    const int ib32 = il/2;
+    il = il%2;
+    // il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
+    device const uint8_t * qs = xb->qs + 8*ib32;
+    device const uint8_t * signs = xb->signs + 4*ib32 + 2*il;
+    const uint8_t qh = xb->qh[ib32] >> 4*il;
+    const float dl = d * (0.5f + ((xb->scales[ib32/2] >> 4*(ib32%2)) & 0xf)) * 0.5f;
+    constant uint8_t * grid1 = (constant uint8_t *)(iq3xs_grid + (qs[4*il+0] | ((qh << 8) & 256)));
+    constant uint8_t * grid2 = (constant uint8_t *)(iq3xs_grid + (qs[4*il+1] | ((qh << 7) & 256)));
+    for (int i = 0; i < 4; ++i) {
+        reg[0][i] = dl * grid1[i] * select(1, -1, signs[0] & kmask_iq2xs[i+0]);
+        reg[1][i] = dl * grid2[i] * select(1, -1, signs[0] & kmask_iq2xs[i+4]);
+    }
+    grid1 = (constant uint8_t *)(iq3xs_grid + (qs[4*il+2] | ((qh << 6) & 256)));
+    grid2 = (constant uint8_t *)(iq3xs_grid + (qs[4*il+3] | ((qh << 5) & 256)));
+    for (int i = 0; i < 4; ++i) {
+        reg[2][i] = dl * grid1[i] * select(1, -1, signs[1] & kmask_iq2xs[i+0]);
+        reg[3][i] = dl * grid2[i] * select(1, -1, signs[1] & kmask_iq2xs[i+4]);
+    }
+}
+
 template <typename type4x4>
 void dequantize_iq1_s(device const block_iq1_s * xb, short il, thread type4x4 & reg) {
     // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
@@ -4824,6 +5204,21 @@ void dequantize_iq1_s(device const block_iq1_s * xb, short il, thread type4x4 &
     }
 }
 
+template <typename type4x4>
+void dequantize_iq4_nl(device const block_iq4_nl * xb, short il, thread type4x4 & reg) {
+    device const uint16_t * q4 = (device const uint16_t *)xb->qs;
+    const float d = xb->d;
+    uint32_t aux32;
+    thread const uint8_t * q8 = (thread const uint8_t *)&aux32;
+    for (int i = 0; i < 4; ++i) {
+        aux32 = ((q4[2*i] | (q4[2*i+1] << 16)) >> 4*il) & 0x0f0f0f0f;
+        reg[i][0] = d * kvalues_iq4nl_f[q8[0]];
+        reg[i][1] = d * kvalues_iq4nl_f[q8[1]];
+        reg[i][2] = d * kvalues_iq4nl_f[q8[2]];
+        reg[i][3] = d * kvalues_iq4nl_f[q8[3]];
+    }
+}
+
 template<typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread float4x4 &)>
 kernel void kernel_get_rows(
         device const  void * src0,
@@ -5366,7 +5761,9 @@ template [[host_name("kernel_get_rows_q6_K")]] kernel get_rows_t kernel_get_rows
 template [[host_name("kernel_get_rows_iq2_xxs")]] kernel get_rows_t kernel_get_rows<block_iq2_xxs, QK_NL, dequantize_iq2_xxs>;
 template [[host_name("kernel_get_rows_iq2_xs")]]  kernel get_rows_t kernel_get_rows<block_iq2_xs,  QK_NL, dequantize_iq2_xs>;
 template [[host_name("kernel_get_rows_iq3_xxs")]] kernel get_rows_t kernel_get_rows<block_iq3_xxs, QK_NL, dequantize_iq3_xxs>;
+template [[host_name("kernel_get_rows_iq3_s")]]   kernel get_rows_t kernel_get_rows<block_iq3_s,   QK_NL, dequantize_iq3_s>;
 template [[host_name("kernel_get_rows_iq1_s")]]   kernel get_rows_t kernel_get_rows<block_iq1_s,   QK_NL, dequantize_iq1_s>;
+template [[host_name("kernel_get_rows_iq4_nl")]]  kernel get_rows_t kernel_get_rows<block_iq4_nl,  2, dequantize_iq4_nl>;
 
 //
 // matrix-matrix multiplication
@@ -5406,7 +5803,9 @@ template [[host_name("kernel_mul_mm_q6_K_f32")]] kernel mat_mm_t kernel_mul_mm<b
 template [[host_name("kernel_mul_mm_iq2_xxs_f32")]] kernel mat_mm_t kernel_mul_mm<block_iq2_xxs, QK_NL, dequantize_iq2_xxs>;
 template [[host_name("kernel_mul_mm_iq2_xs_f32")]]  kernel mat_mm_t kernel_mul_mm<block_iq2_xs,  QK_NL, dequantize_iq2_xs>;
 template [[host_name("kernel_mul_mm_iq3_xxs_f32")]] kernel mat_mm_t kernel_mul_mm<block_iq3_xxs, QK_NL, dequantize_iq3_xxs>;
+template [[host_name("kernel_mul_mm_iq3_s_f32")]]   kernel mat_mm_t kernel_mul_mm<block_iq3_s,   QK_NL, dequantize_iq3_s>;
 template [[host_name("kernel_mul_mm_iq1_s_f32")]]   kernel mat_mm_t kernel_mul_mm<block_iq1_s,   QK_NL, dequantize_iq1_s>;
+template [[host_name("kernel_mul_mm_iq4_nl_f32")]]  kernel mat_mm_t kernel_mul_mm<block_iq4_nl,  2, dequantize_iq4_nl>;
 
 //
 // indirect matrix-matrix multiplication
@@ -5458,7 +5857,9 @@ template [[host_name("kernel_mul_mm_id_q6_K_f32")]] kernel mat_mm_id_t kernel_mu
 template [[host_name("kernel_mul_mm_id_iq2_xxs_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_iq2_xxs, QK_NL, dequantize_iq2_xxs>;
 template [[host_name("kernel_mul_mm_id_iq2_xs_f32")]]  kernel mat_mm_id_t kernel_mul_mm_id<block_iq2_xs,  QK_NL, dequantize_iq2_xs>;
 template [[host_name("kernel_mul_mm_id_iq3_xxs_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_iq3_xxs, QK_NL, dequantize_iq3_xxs>;
+template [[host_name("kernel_mul_mm_id_iq3_s_f32")]]   kernel mat_mm_id_t kernel_mul_mm_id<block_iq3_s,   QK_NL, dequantize_iq3_s>;
 template [[host_name("kernel_mul_mm_id_iq1_s_f32")]]   kernel mat_mm_id_t kernel_mul_mm_id<block_iq1_s,   QK_NL, dequantize_iq1_s>;
+template [[host_name("kernel_mul_mm_id_iq4_nl_f32")]]  kernel mat_mm_id_t kernel_mul_mm_id<block_iq4_nl,  2, dequantize_iq4_nl>;
 
 //
 // matrix-vector multiplication
@@ -6427,6 +6828,71 @@ kernel void kernel_mul_mv_id_iq3_xxs_f32(
         sgitg);
 }
 
+[[host_name("kernel_mul_mv_id_iq3_s_f32")]]
+kernel void kernel_mul_mv_id_iq3_s_f32(
+        device const    char * ids,
+        device const    char * src1,
+        device         float * dst,
+        constant    uint64_t & nbi1,
+        constant     int64_t & ne00,
+        constant     int64_t & ne01,
+        constant     int64_t & ne02,
+        constant    uint64_t & nb00,
+        constant    uint64_t & nb01,
+        constant    uint64_t & nb02,
+        constant     int64_t & ne10,
+        constant     int64_t & ne11,
+        constant     int64_t & ne12,
+        constant     int64_t & ne13,
+        constant    uint64_t & nb10,
+        constant    uint64_t & nb11,
+        constant    uint64_t & nb12,
+        constant     int64_t & ne0,
+        constant     int64_t & ne1,
+        constant    uint64_t & nb1,
+        constant        uint & r2,
+        constant        uint & r3,
+        constant         int & idx,
+        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,
+        threadgroup int8_t   * shared_values [[threadgroup(0)]],
+        uint3                  tgpig[[threadgroup_position_in_grid]],
+        uint                   tiitg[[thread_index_in_threadgroup]],
+        uint                   tiisg[[thread_index_in_simdgroup]],
+        uint                   sgitg[[simdgroup_index_in_threadgroup]]) {
+    device const char * src0[8] = {src00, src01, src02, src03, src04, src05, src06, src07};
+
+    const int64_t bid = tgpig.z/(ne12*ne13);
+
+    tgpig.z = tgpig.z%(ne12*ne13);
+
+    const int32_t id = ((device int32_t *) (ids + bid*nbi1))[idx];
+
+    kernel_mul_mv_iq3_s_f32_impl(
+        src0[id],
+        (device const float *) (src1 + bid*nb11),
+        dst + bid*ne0,
+        ne00,
+        ne01,
+        ne02,
+        ne10,
+        ne12,
+        ne0,
+        ne1,
+        r2,
+        r3,
+        shared_values,
+        tgpig,
+        tiisg,
+        sgitg);
+}
+
 [[host_name("kernel_mul_mv_id_iq1_s_f32")]]
 kernel void kernel_mul_mv_id_iq1_s_f32(
         device const    char * ids,
@@ -6489,3 +6955,68 @@ kernel void kernel_mul_mv_id_iq1_s_f32(
         tiisg,
         sgitg);
 }
+
+[[host_name("kernel_mul_mv_id_iq4_nl_f32")]]
+kernel void kernel_mul_mv_id_iq4_nl_f32(
+        device const    char * ids,
+        device const    char * src1,
+        device         float * dst,
+        constant    uint64_t & nbi1,
+        constant     int64_t & ne00,
+        constant     int64_t & ne01,
+        constant     int64_t & ne02,
+        constant    uint64_t & nb00,
+        constant    uint64_t & nb01,
+        constant    uint64_t & nb02,
+        constant     int64_t & ne10,
+        constant     int64_t & ne11,
+        constant     int64_t & ne12,
+        constant     int64_t & ne13,
+        constant    uint64_t & nb10,
+        constant    uint64_t & nb11,
+        constant    uint64_t & nb12,
+        constant     int64_t & ne0,
+        constant     int64_t & ne1,
+        constant    uint64_t & nb1,
+        constant        uint & r2,
+        constant        uint & r3,
+        constant         int & idx,
+        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,
+        threadgroup float    * shared_values [[threadgroup(0)]],
+        uint3                  tgpig[[threadgroup_position_in_grid]],
+        uint                   tiitg[[thread_index_in_threadgroup]],
+        uint                   tiisg[[thread_index_in_simdgroup]],
+        uint                   sgitg[[simdgroup_index_in_threadgroup]]) {
+    device const char * src0[8] = {src00, src01, src02, src03, src04, src05, src06, src07};
+
+    const int64_t bid = tgpig.z/(ne12*ne13);
+
+    tgpig.z = tgpig.z%(ne12*ne13);
+
+    const int32_t id = ((device int32_t *) (ids + bid*nbi1))[idx];
+
+    kernel_mul_mv_iq4_nl_f32_impl(
+        src0[id],
+        (device const float *) (src1 + bid*nb11),
+        dst + bid*ne0,
+        ne00,
+        ne01,
+        ne02,
+        ne10,
+        ne12,
+        ne0,
+        ne1,
+        r2,
+        r3,
+        shared_values,
+        tgpig,
+        tiisg,
+        sgitg);
+}