whispercpp 1.3.2 → 1.3.3

data/ext/sources/ggml/src/ggml-metal/ggml-metal.metal

@@ -35,6 +35,17 @@ 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
 };
 
+static inline int best_index_int8(int n, constant float * val, float x) {
+    if (x <= val[0]) return 0;
+    if (x >= val[n-1]) return n-1;
+    int ml = 0, mu = n-1;
+    while (mu-ml > 1) {
+        int mav = (ml+mu)/2;
+        if (x < val[mav]) mu = mav; else ml = mav;
+    }
+    return x - val[mu-1] < val[mu] - x ? mu-1 : mu;
+}
+
 // NOTE: this is not dequantizing - we are simply fitting the template
 template <typename type4x4>
 void dequantize_f32(device const float4x4 * src, short il, thread type4x4 & reg) {
@@ -97,6 +108,176 @@ void dequantize_q4_0_t4(device const block_q4_0 * xb, short il, thread type4 & r
     }
 }
 
+void quantize_q4_0(device const float * src, device block_q4_0 & dst) {
+    float amax = 0.0f; // absolute max
+    float max  = 0.0f;
+
+    for (int j = 0; j < QK4_0; j++) {
+        const float v = src[j];
+        if (amax < fabs(v)) {
+            amax = fabs(v);
+            max  = v;
+        }
+    }
+
+    const float d = max / -8;
+    const float id = d ? 1.0f/d : 0.0f;
+
+    dst.d = d;
+
+    for (int j = 0; j < QK4_0/2; ++j) {
+        const float x0 = src[0       + j]*id;
+        const float x1 = src[QK4_0/2 + j]*id;
+
+        const uint8_t xi0 = MIN(15, (int8_t)(x0 + 8.5f));
+        const uint8_t xi1 = MIN(15, (int8_t)(x1 + 8.5f));
+
+        dst.qs[j]  = xi0;
+        dst.qs[j] |= xi1 << 4;
+    }
+}
+
+void quantize_q4_1(device const float * src, device block_q4_1 & dst) {
+#pragma METAL fp math_mode(safe)
+    float min = FLT_MAX;
+    float max = -FLT_MAX;
+
+    for (int j = 0; j < QK4_1; j++) {
+        const float v = src[j];
+        if (min > v) min = v;
+        if (max < v) max = v;
+    }
+
+    const float d = (max - min) / ((1 << 4) - 1);
+    const float id = d ? 1.0f/d : 0.0f;
+
+    dst.d = d;
+    dst.m = min;
+
+    for (int j = 0; j < QK4_1/2; ++j) {
+        const float x0 = (src[0       + j] - min)*id;
+        const float x1 = (src[QK4_1/2 + j] - min)*id;
+
+        const uint8_t xi0 = MIN(15, (int8_t)(x0 + 0.5f));
+        const uint8_t xi1 = MIN(15, (int8_t)(x1 + 0.5f));
+
+        dst.qs[j]  = xi0;
+        dst.qs[j] |= xi1 << 4;
+    }
+}
+
+void quantize_q5_0(device const float * src, device block_q5_0 & dst) {
+    float amax = 0.0f; // absolute max
+    float max  = 0.0f;
+
+    for (int j = 0; j < QK5_0; j++) {
+        const float v = src[j];
+        if (amax < fabs(v)) {
+            amax = fabs(v);
+            max  = v;
+        }
+    }
+
+    const float d = max / -16;
+    const float id = d ? 1.0f/d : 0.0f;
+
+    dst.d = d;
+
+    uint32_t qh = 0;
+    for (int j = 0; j < QK5_0/2; ++j) {
+        const float x0 = src[0       + j]*id;
+        const float x1 = src[QK5_0/2 + j]*id;
+
+        const uint8_t xi0 = MIN(31, (int8_t)(x0 + 16.5f));
+        const uint8_t xi1 = MIN(31, (int8_t)(x1 + 16.5f));
+
+        dst.qs[j] = (xi0 & 0xf) | ((xi1 & 0xf) << 4);
+        qh |= ((xi0 & 0x10u) >> 4) << (j + 0);
+        qh |= ((xi1 & 0x10u) >> 4) << (j + QK5_0/2);
+    }
+
+    thread const uint8_t * qh8 = (thread const uint8_t *)&qh;
+
+    for (int j = 0; j < 4; ++j) {
+        dst.qh[j] = qh8[j];
+    }
+}
+
+void quantize_q5_1(device const float * src, device block_q5_1 & dst) {
+#pragma METAL fp math_mode(safe)
+    float max = src[0];
+    float min = src[0];
+
+    for (int j = 1; j < QK5_1; j++) {
+        const float v = src[j];
+        min = v < min ? v : min;
+        max = v > max ? v : max;
+    }
+
+    const float d = (max - min) / 31;
+    const float id = d ? 1.0f/d : 0.0f;
+
+    dst.d = d;
+    dst.m = min;
+
+    uint32_t qh = 0;
+    for (int j = 0; j < QK5_1/2; ++j) {
+        const float x0 = (src[0       + j] - min)*id;
+        const float x1 = (src[QK5_1/2 + j] - min)*id;
+
+        const uint8_t xi0 = (uint8_t)(x0 + 0.5f);
+        const uint8_t xi1 = (uint8_t)(x1 + 0.5f);
+
+        dst.qs[j] = (xi0 & 0xf) | ((xi1 & 0xf) << 4);
+        qh |= ((xi0 & 0x10u) >> 4) << (j + 0);
+        qh |= ((xi1 & 0x10u) >> 4) << (j + QK5_1/2);
+    }
+
+    thread const uint8_t * qh8 = (thread const uint8_t *)&qh;
+
+    for (int j = 0; j < 4; ++j) {
+        dst.qh[j] = qh8[j];
+    }
+}
+
+void quantize_iq4_nl(device const float * src, device block_iq4_nl & dst) {
+#pragma METAL fp math_mode(safe)
+    float amax = 0.0f; // absolute max
+    float max  = 0.0f;
+
+    for (int j = 0; j < QK4_NL; j++) {
+        const float v = src[j];
+        if (amax < fabs(v)) {
+            amax = fabs(v);
+            max  = v;
+        }
+    }
+
+    const float d = max / kvalues_iq4nl_f[0];
+    const float id = d ? 1.0f/d : 0.0f;
+
+    float sumqx = 0, sumq2 = 0;
+    for (int j = 0; j < QK4_NL/2; ++j) {
+        const float x0 = src[0        + j]*id;
+        const float x1 = src[QK4_NL/2 + j]*id;
+
+        const uint8_t xi0 = best_index_int8(16, kvalues_iq4nl_f, x0);
+        const uint8_t xi1 = best_index_int8(16, kvalues_iq4nl_f, x1);
+
+        dst.qs[j] = xi0 | (xi1 << 4);
+
+        const float v0 = kvalues_iq4nl_f[xi0];
+        const float v1 = kvalues_iq4nl_f[xi1];
+        const float w0 = src[0        + j]*src[0        + j];
+        const float w1 = src[QK4_NL/2 + j]*src[QK4_NL/2 + j];
+        sumqx += w0*v0*src[j] + w1*v1*src[QK4_NL/2 + j];
+        sumq2 += w0*v0*v0 + w1*v1*v1;
+
+    }
+
+    dst.d = sumq2 > 0 ? sumqx/sumq2 : d;
+}
+
 template <typename type4x4>
 void dequantize_q4_1(device const block_q4_1 * xb, short il, thread type4x4 & reg) {
     device const uint16_t * qs = ((device const uint16_t *)xb + 2);
@@ -279,6 +460,26 @@ void dequantize_q8_0_t4(device const block_q8_0 *xb, short il, thread type4 & re
     }
 }
 
+void quantize_q8_0(device const float * src, device block_q8_0 & dst) {
+    float amax = 0.0f; // absolute max
+
+    for (int j = 0; j < QK8_0; j++) {
+        const float v = src[j];
+        amax = MAX(amax, fabs(v));
+    }
+
+    const float d = amax / ((1 << 7) - 1);
+    const float id = d ? 1.0f/d : 0.0f;
+
+    dst.d = d;
+
+    for (int j = 0; j < QK8_0; ++j) {
+        const float x0 = src[j]*id;
+
+        dst.qs[j] = round(x0);
+    }
+}
+
 template <typename type4x4>
 void dequantize_q2_K(device const block_q2_K *xb, short il, thread type4x4 & reg) {
     const float d = xb->d;
@@ -993,31 +1194,125 @@ kernel void kernel_neg(
     dst[tpig] = -src0[tpig];
 }
 
+kernel void kernel_reglu(
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        constant ggml_metal_kargs_glu & args,
+        uint tgpig[[threadgroup_position_in_grid]],
+        uint tpitg[[thread_position_in_threadgroup]],
+        uint   ntg[[threads_per_threadgroup]]) {
+    device const float * src0_row = (device const float *) ((device const char *) src0 + tgpig*args.nb01) + args.i00;
+    device const float * src1_row = (device const float *) ((device const char *) src1 + tgpig*args.nb11) + args.i10;
+    device       float * dst_row  = (device       float *) ((device       char *) dst  + tgpig*args.nb1);
+
+    for (int i0 = tpitg; i0 < args.ne0; i0 += ntg) {
+        const float x0 = src0_row[i0];
+        const float x1 = src1_row[i0];
+
+        dst_row[i0] = x0*x1*(x0 > 0.0f);
+    }
+}
+
+kernel void kernel_geglu(
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        constant ggml_metal_kargs_glu & args,
+        uint tgpig[[threadgroup_position_in_grid]],
+        uint tpitg[[thread_position_in_threadgroup]],
+        uint   ntg[[threads_per_threadgroup]]) {
+    device const float * src0_row = (device const float *) ((device const char *) src0 + tgpig*args.nb01) + args.i00;
+    device const float * src1_row = (device const float *) ((device const char *) src1 + tgpig*args.nb11) + args.i10;
+    device       float * dst_row  = (device       float *) ((device       char *) dst  + tgpig*args.nb1);
+
+    for (int i0 = tpitg; i0 < args.ne0; i0 += ntg) {
+        const float x0 = src0_row[i0];
+        const float x1 = src1_row[i0];
+
+        const float gelu = 0.5f*x0*(1.0f + precise::tanh(SQRT_2_OVER_PI*x0*(1.0f + GELU_COEF_A*x0*x0)));
+
+        dst_row[i0] = gelu*x1;
+    }
+}
+
+kernel void kernel_swiglu(
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        constant ggml_metal_kargs_glu & args,
+        uint tgpig[[threadgroup_position_in_grid]],
+        uint tpitg[[thread_position_in_threadgroup]],
+        uint   ntg[[threads_per_threadgroup]]) {
+    device const float * src0_row = (device const float *) ((device const char *) src0 + tgpig*args.nb01) + args.i00;
+    device const float * src1_row = (device const float *) ((device const char *) src1 + tgpig*args.nb11) + args.i10;
+    device       float * dst_row  = (device       float *) ((device       char *) dst  + tgpig*args.nb1);
+
+    for (int i0 = tpitg; i0 < args.ne0; i0 += ntg) {
+        const float x0 = src0_row[i0];
+        const float x1 = src1_row[i0];
+
+        const float silu = x0 / (1.0f + exp(-x0));
+
+        dst_row[i0] = silu*x1;
+    }
+}
+
+template <bool norm>
 kernel void kernel_sum_rows(
+        constant ggml_metal_kargs_sum_rows & args,
         device const float * src0,
         device       float * dst,
-        constant ggml_metal_kargs_sum_rows & args,
-        uint3 tpig[[thread_position_in_grid]]) {
-    int64_t i3 = tpig.z;
-    int64_t i2 = tpig.y;
-    int64_t i1 = tpig.x;
+        threadgroup  float * shmem_f32 [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    int64_t i3 = tgpig.z;
+    int64_t i2 = tgpig.y;
+    int64_t i1 = tgpig.x;
 
     if (i3 >= args.ne03 || i2 >= args.ne02 || i1 >= args.ne01) {
         return;
     }
 
+    if (sgitg == 0) {
+        shmem_f32[tiisg] = 0.0f;
+    }
+
     device const float * src_row = (device const float *) ((device const char *) src0 + i1*args.nb01 + i2*args.nb02 + i3*args.nb03);
     device       float * dst_row = (device       float *) ((device       char *) dst  + i1*args.nb1  + i2*args.nb2  + i3*args.nb3);
 
-    float row_sum = 0;
+    float sumf = 0;
 
-    for (int64_t i0 = 0; i0 < args.ne00; i0++) {
-        row_sum += src_row[i0];
+    for (int64_t i0 = tpitg.x; i0 < args.ne00; i0 += ntg.x) {
+        sumf += src_row[i0];
     }
 
-    dst_row[0] = row_sum;
+    sumf = simd_sum(sumf);
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    if (tiisg == 0) {
+        shmem_f32[sgitg] = sumf;
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    sumf = shmem_f32[tiisg];
+    sumf = simd_sum(sumf);
+
+    if (tpitg.x == 0) {
+        dst_row[0] = norm ? sumf / args.ne00 : sumf;
+    }
 }
 
+typedef decltype(kernel_sum_rows<false>) kernel_sum_rows_t;
+
+template [[host_name("kernel_sum_rows")]] kernel kernel_sum_rows_t kernel_sum_rows<false>;
+template [[host_name("kernel_mean")]]     kernel kernel_sum_rows_t kernel_sum_rows<true>;
+
 template<typename T>
 kernel void kernel_soft_max(
         device const  char * src0,
@@ -2502,6 +2797,70 @@ template [[host_name("kernel_mul_mv_bf16_f32")]]  kernel mul_mv_t kernel_mul_mv<
 template [[host_name("kernel_mul_mv_bf16_bf16")]] kernel mul_mv_t kernel_mul_mv<bfloat, bfloat4, bfloat, bfloat4>;
 #endif
 
+template<typename T04, typename T14, typename args_t>
+void kernel_mul_mv_c4_impl(
+        args_t args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint3  tgpig,
+        ushort tiisg) {
+    const int r0 = tgpig.x*32 + tiisg;
+    const int rb = tgpig.y*N_MV_T_T;
+    const int im = tgpig.z;
+
+    if (r0 >= args.ne01) {
+        return;
+    }
+
+    const uint i12 = im%args.ne12;
+    const uint i13 = im/args.ne12;
+
+    const uint64_t offset0 = r0*args.nb01 + (i12/args.r2)*args.nb02 + (i13/args.r3)*args.nb03;
+
+    device const T04 * x = (device const T04 *) (src0 + offset0);
+
+    device float * dst_f32 = (device float *) dst + (uint64_t)im*args.ne0*args.ne1;
+
+    for (int row = 0; row < N_MV_T_T; ++row) {
+        int r1 = rb + row;
+        if (r1 >= args.ne11) {
+            break;
+        }
+
+        const uint64_t offset1 = r1*args.nb11 + (i12   )*args.nb12 + (i13   )*args.nb13;
+
+        device const T14 * y = (device const T14 *) (src1 + offset1);
+
+        dst_f32[(uint64_t)r1*args.ne0 + r0] = dot((float4) x[0], (float4) y[0]);
+    }
+}
+
+template<typename T04, typename T14>
+kernel void kernel_mul_mv_c4(
+        constant ggml_metal_kargs_mul_mv & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiisg[[thread_index_in_simdgroup]]) {
+    kernel_mul_mv_c4_impl<T04, T14, constant ggml_metal_kargs_mul_mv &>(
+        args,
+        src0,
+        src1,
+        dst,
+        tgpig,
+        tiisg);
+}
+
+typedef decltype(kernel_mul_mv_c4<half4, half4>) mul_mv_c4_t;
+
+template [[host_name("kernel_mul_mv_f32_f32_c4")]]  kernel mul_mv_c4_t kernel_mul_mv_c4<float4,  float4>;
+template [[host_name("kernel_mul_mv_f16_f32_c4")]]  kernel mul_mv_c4_t kernel_mul_mv_c4<half4,   float4>;
+#if defined(GGML_METAL_USE_BF16)
+template [[host_name("kernel_mul_mv_bf16_f32_c4")]] kernel mul_mv_c4_t kernel_mul_mv_c4<bfloat4, float4>;
+#endif
+
 template<typename T, typename T4>
 kernel void kernel_mul_mv_1row(
         constant ggml_metal_kargs_mul_mv & args,
@@ -3328,14 +3687,12 @@ kernel void kernel_flash_attn_ext(
     constexpr short NW  = N_SIMDWIDTH;
     constexpr short SH  = (2*C + Q); // shared memory per simdgroup (s_t == float)
 
-    const short TS = nsg*SH;   // shared memory size per query in (s_t == float)
-    const short T  = DK + 2*TS; // shared memory size per query in (half)
+    const short TS = nsg*SH;      // shared memory size per query in (s_t == float)
+    const short T  = 2*DK + 2*TS; // shared memory size per query in (half)
 
-    threadgroup q_t  * sq  = (threadgroup q_t  *) (shmem_f16 +              0*DK); // holds the query data
-    threadgroup q4_t * sq4 = (threadgroup q4_t *) (shmem_f16 +              0*DK); // same as above but in q4_t
-    threadgroup o_t  * so  = (threadgroup o_t  *) (shmem_f16 +              0*DK); // reuse query data for accumulation
-    threadgroup o4_t * so4 = (threadgroup o4_t *) (shmem_f16 +              0*DK); // same as above but in o4_t
-    threadgroup s_t  * ss  = (threadgroup s_t  *) (shmem_f16 + 2*sgitg*SH + Q*DK); // scratch buffer for attention, mask and diagonal matrix
+    threadgroup q_t  * sq  = (threadgroup q_t  *) (shmem_f16 +                0*DK); // holds the query data
+    threadgroup q4_t * sq4 = (threadgroup q4_t *) (shmem_f16 +                0*DK); // same as above but in q4_t
+    threadgroup s_t  * ss  = (threadgroup s_t  *) (shmem_f16 + 2*sgitg*SH + 2*Q*DK); // scratch buffer for attention, mask and diagonal matrix
 
     threadgroup k_t    * sk    = (threadgroup k_t    *) (shmem_f16 + sgitg*(4*16*KV) + Q*T); // scratch buffer to load K in shared memory
     threadgroup k4x4_t * sk4x4 = (threadgroup k4x4_t *) (shmem_f16 + sgitg*(4*16*KV) + Q*T); // same as above but in k4x4_t
@@ -3354,7 +3711,7 @@ kernel void kernel_flash_attn_ext(
             if (iq1 + j < args.ne01) {
                 sq4[j*DK4 + i] = (q4_t) q4[i];
             } else {
-                sq4[j*DK4 + i] = (q4_t) 0.0f;
+                sq4[j*DK4 + i] = 0;
             }
         }
     }
@@ -3548,20 +3905,20 @@ kernel void kernel_flash_attn_ext(
 
             // O = diag(ms)*O
             {
-                s8x8_t mm;
-                simdgroup_load(mm, ss + 2*C, TS, 0, false);
+                s8x8_t ms;
+                simdgroup_load(ms, ss + 2*C, TS, 0, false);
 
                 #pragma unroll(DV8)
                 for (short i = 0; i < DV8; ++i) {
-                    simdgroup_multiply(lo[i], mm, lo[i]);
+                    simdgroup_multiply(lo[i], ms, lo[i]);
                 }
             }
 
             // O = O + (Q*K^T)*V
             {
                 for (short cc = 0; cc < C/8; ++cc) {
-                    s8x8_t ms;
-                    simdgroup_load(ms, ss + 8*cc, TS, 0, false);
+                    s8x8_t vs;
+                    simdgroup_load(vs, ss + 8*cc, TS, 0, false);
 
                     if (is_same<vd4x4_t, v4x4_t>::value) {
                         // we can read directly from global memory
@@ -3572,7 +3929,7 @@ kernel void kernel_flash_attn_ext(
                             v8x8_t mv;
                             simdgroup_load(mv, pv + i*8, args.nb21/sizeof(v_t), 0, false); // TODO: use ne20
 
-                            simdgroup_multiply_accumulate(lo[i], ms, mv, lo[i]);
+                            simdgroup_multiply_accumulate(lo[i], vs, mv, lo[i]);
                         }
                     } else {
                         for (short ii = 0; ii < DV16; ii += 4) {
@@ -3593,10 +3950,10 @@ kernel void kernel_flash_attn_ext(
                                     v8x8_t mv;
 
                                     simdgroup_load(mv, sv + 16*k + 0*8, 4*16, 0, false);
-                                    simdgroup_multiply_accumulate(lo[2*(ii + k) + 0], ms, mv, lo[2*(ii + k) + 0]);
+                                    simdgroup_multiply_accumulate(lo[2*(ii + k) + 0], vs, mv, lo[2*(ii + k) + 0]);
 
                                     simdgroup_load(mv, sv + 16*k + 1*8, 4*16, 0, false);
-                                    simdgroup_multiply_accumulate(lo[2*(ii + k) + 1], ms, mv, lo[2*(ii + k) + 1]);
+                                    simdgroup_multiply_accumulate(lo[2*(ii + k) + 1], vs, mv, lo[2*(ii + k) + 1]);
                                 }
                             } else {
                                 if (ii + tx < DV16) {
@@ -3611,10 +3968,10 @@ kernel void kernel_flash_attn_ext(
                                     v8x8_t mv;
 
                                     simdgroup_load(mv, sv + 16*k + 0*8, 4*16, 0, false);
-                                    simdgroup_multiply_accumulate(lo[2*(ii + k) + 0], ms, mv, lo[2*(ii + k) + 0]);
+                                    simdgroup_multiply_accumulate(lo[2*(ii + k) + 0], vs, mv, lo[2*(ii + k) + 0]);
 
                                     simdgroup_load(mv, sv + 16*k + 1*8, 4*16, 0, false);
-                                    simdgroup_multiply_accumulate(lo[2*(ii + k) + 1], ms, mv, lo[2*(ii + k) + 1]);
+                                    simdgroup_multiply_accumulate(lo[2*(ii + k) + 1], vs, mv, lo[2*(ii + k) + 1]);
                                 }
                             }
                         }
@@ -3624,93 +3981,89 @@ kernel void kernel_flash_attn_ext(
         }
 
         // these are needed for reducing the results from the simdgroups (reuse the ss buffer)
-        for (short j = 0; j < Q; ++j) {
-            if (tiisg == 0) {
-                ss[j*TS + 0] = S[j];
-                ss[j*TS + 1] = M[j];
-            }
+        for (short j = tiisg; j < Q; j += NW) {
+            ss[j*TS + 0] = S[j];
+            ss[j*TS + 1] = M[j];
         }
     }
 
-    // reduce the warps sequentially
-    for (ushort sg = 1; sg < nsg; ++sg) {
-        float S = { 0.0f };
-        float M = { -__FLT_MAX__/2 };
+    threadgroup_barrier(mem_flags::mem_threadgroup);
 
-        threadgroup_barrier(mem_flags::mem_threadgroup);
+    threadgroup float  * so  = (threadgroup float  *) (shmem_f16 + 0*DK); // reuse query data for accumulation
+    threadgroup float4 * so4 = (threadgroup float4 *) (shmem_f16 + 0*DK);
 
-        // each simdgroup stores its output to shared memory, reusing sq
-        if (sgitg == sg) {
-            for (short i = 0; i < DV8; ++i) {
-                simdgroup_store(lo[i], so + i*8, DV, 0, false);
-            }
+    // store result to shared memory in F32
+    if (sgitg == 0) {
+        for (short i = 0; i < DV8; ++i) {
+            //simdgroup_store(lo[i], so + i*8, DV, 0, false);
+            simdgroup_float8x8 t(1.0f);
+            simdgroup_multiply(t, lo[i], t);
+            simdgroup_store(t, so + i*8, DV, 0, false);
         }
+    }
 
-        threadgroup_barrier(mem_flags::mem_threadgroup);
+    threadgroup_barrier(mem_flags::mem_threadgroup);
 
-        // the first simdgroup accumulates the results from the other simdgroups
-        if (sgitg == 0) {
-            for (short j = 0; j < Q; ++j) {
-                const float S0 = ss[j*TS +         0];
-                const float S1 = ss[j*TS + sg*SH + 0];
+    // reduce the warps sequentially
+    for (ushort sg = 1; sg < nsg; ++sg) {
+        if (sgitg == sg) {
+            for (short j = tiisg; j < Q; j += NW) {
+                const float S0 = ss[j*TS - 1*SH + 0];
+                const float S1 = ss[j*TS        + 0];
 
-                const float M0 = ss[j*TS +         1];
-                const float M1 = ss[j*TS + sg*SH + 1];
+                const float M0 = ss[j*TS - 1*SH + 1];
+                const float M1 = ss[j*TS        + 1];
 
-                M = max(M0, M1);
+                const float M = max(M0, M1);
 
-                const float ms0 = exp(M0 - M);
-                const float ms1 = exp(M1 - M);
+                float ms0 = exp(M0 - M);
+                float ms1 = exp(M1 - M);
 
-                S = S0*ms0 + S1*ms1;
+                const float S = S0*ms0 + S1*ms1;
 
-                if (tiisg == 0) {
-                    ss[j*TS + 0] = S;
-                    ss[j*TS + 1] = M;
+                ss[j*TS + 0] = S;
+                ss[j*TS + 1] = M;
 
-                    ss[j*TS + 2*C + j        ] = ms0;
-                    ss[j*TS + 2*C + j + sg*SH] = ms1;
-                }
+                ss[j*TS + 2*C + j - 1*SH] = ms0;
+                ss[j*TS + 2*C + j       ] = ms1;
             }
 
+            //simdgroup_barrier(mem_flags::mem_threadgroup);
+
             // O_0 = diag(ms0)*O_0 + diag(ms1)*O_1
             {
                 s8x8_t ms0;
                 s8x8_t ms1;
 
-                simdgroup_load(ms0, ss + 2*C,         TS, 0, false);
-                simdgroup_load(ms1, ss + 2*C + sg*SH, TS, 0, false);
+                simdgroup_load(ms0, ss + 2*C - 1*SH, TS, 0, false);
+                simdgroup_load(ms1, ss + 2*C,        TS, 0, false);
 
                 #pragma unroll(DV8)
                 for (short i = 0; i < DV8; ++i) {
-                    o8x8_t t;
+                    simdgroup_float8x8 t;
 
                     simdgroup_load    (t, so + i*8, DV, 0, false);
-                    simdgroup_multiply(t, ms1, t);
+                    simdgroup_multiply(t, ms0, t);
 
-                    simdgroup_multiply_accumulate(lo[i], ms0, lo[i], t);
+                    simdgroup_multiply_accumulate(t, ms1, lo[i], t);
+                    simdgroup_store(t, so + i*8, DV, 0, false);
                 }
             }
         }
-    }
 
-    // store result to shared memory (reuse sq)
-    if (sgitg == 0) {
-        for (short i = 0; i < DV8; ++i) {
-            simdgroup_store(lo[i], so + i*8, DV, 0, false);
-        }
+        threadgroup_barrier(mem_flags::mem_threadgroup);
     }
 
-    device float4 * dst4 = (device float4 *) dst;
+    threadgroup s_t * sf = (threadgroup s_t *) (shmem_f16 + 2*(nsg-1)*SH + 2*Q*DK);
 
     // final rescale with 1/S and store to global memory
-    if (sgitg == 0) {
-        for (short j = 0; j < Q && iq1 + j < args.ne01; ++j) {
-            const float S = ss[j*TS + 0];
+    for (short j = sgitg; j < Q && iq1 + j < args.ne01; j += nsg) {
+        const float S = 1.0f/sf[j*TS + 0];
 
-            for (short i = tiisg; i < DV4; i += NW) {
-                dst4[((uint64_t)iq3*args.ne2*args.ne1 + iq2 + (uint64_t)(iq1 + j)*args.ne1)*DV4 + i] = (float4) so4[j*DV4 + i]/S;
-            }
+        device float4 * dst4 = (device float4 *) dst + ((uint64_t)iq3*args.ne2*args.ne1 + iq2 + (uint64_t)(iq1 + j)*args.ne1)*DV4;
+
+        for (short i = tiisg; i < DV4; i += NW) {
+            dst4[i] = (float4) so4[j*DV4 + i]*S;
         }
     }
 }
@@ -3719,12 +4072,22 @@ kernel void kernel_flash_attn_ext(
 //       template to be able to explore different combinations
 //
 #define FA_TYPES \
-    half,  half4,   simdgroup_half8x8,  \
-    half,  half4x4, simdgroup_half8x8,  \
-    half,  half4x4, simdgroup_half8x8,  \
-    float,          simdgroup_float8x8, \
-    float,          simdgroup_float8x8, \
-    half,  half4,   simdgroup_half8x8
+    float,  float4,    simdgroup_float8x8, \
+    half,   half4x4,   simdgroup_half8x8,  \
+    half,   half4x4,   simdgroup_half8x8,  \
+    float,             simdgroup_float8x8, \
+    float,             simdgroup_float8x8, \
+    half,   half4,     simdgroup_half8x8
+    //float,  float4,    simdgroup_float8x8
+
+#define FA_TYPES_BF \
+    bfloat, bfloat4,   simdgroup_bfloat8x8, \
+    bfloat, bfloat4x4, simdgroup_bfloat8x8, \
+    bfloat, bfloat4x4, simdgroup_bfloat8x8, \
+    float,             simdgroup_float8x8,  \
+    float,             simdgroup_float8x8,  \
+    half,   half4,     simdgroup_half8x8
+    //float,  float4,    simdgroup_float8x8
 
 typedef decltype(kernel_flash_attn_ext<FA_TYPES, half4x4, 1, dequantize_f16, half4x4, 1, dequantize_f16, 64, 64>) flash_attn_ext_t;
 
@@ -3739,15 +4102,15 @@ template [[host_name("kernel_flash_attn_ext_f16_h256")]]         kernel flash_at
 template [[host_name("kernel_flash_attn_ext_f16_hk576_hv512")]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  576, 512>;
 
 #if defined(GGML_METAL_USE_BF16)
-template [[host_name("kernel_flash_attn_ext_bf16_h64" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 64,  64>;
-template [[host_name("kernel_flash_attn_ext_bf16_h80" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 80,  80>;
-template [[host_name("kernel_flash_attn_ext_bf16_h96" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 96,  96>;
-template [[host_name("kernel_flash_attn_ext_bf16_h112")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 112, 112>;
-template [[host_name("kernel_flash_attn_ext_bf16_h128")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 128, 128>;
-template [[host_name("kernel_flash_attn_ext_bf16_h192")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 192, 192>;
-template [[host_name("kernel_flash_attn_ext_bf16_hk192_hv128")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 192, 128>;
-template [[host_name("kernel_flash_attn_ext_bf16_h256")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 256, 256>;
-template [[host_name("kernel_flash_attn_ext_bf16_hk576_hv512")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 576, 512>;
+template [[host_name("kernel_flash_attn_ext_bf16_h64" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 64,  64>;
+template [[host_name("kernel_flash_attn_ext_bf16_h80" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 80,  80>;
+template [[host_name("kernel_flash_attn_ext_bf16_h96" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 96,  96>;
+template [[host_name("kernel_flash_attn_ext_bf16_h112")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 112, 112>;
+template [[host_name("kernel_flash_attn_ext_bf16_h128")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 128, 128>;
+template [[host_name("kernel_flash_attn_ext_bf16_h192")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 192, 192>;
+template [[host_name("kernel_flash_attn_ext_bf16_hk192_hv128")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 192, 128>;
+template [[host_name("kernel_flash_attn_ext_bf16_h256")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 256, 256>;
+template [[host_name("kernel_flash_attn_ext_bf16_hk576_hv512")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 576, 512>;
 #endif
 
 template [[host_name("kernel_flash_attn_ext_q4_0_h64" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 64,  64>;
@@ -3801,6 +4164,7 @@ template [[host_name("kernel_flash_attn_ext_q8_0_h256")]]        kernel flash_at
 template [[host_name("kernel_flash_attn_ext_q8_0_hk576_hv512")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 576, 512>;
 
 #undef FA_TYPES
+#undef FA_TYPES_BF
 
 template<
     typename q4_t,  // query types in shared memory
@@ -3847,12 +4211,12 @@ kernel void kernel_flash_attn_ext_vec(
 
     const short T = DK + nsg*SH; // shared memory size per query in (half)
 
-  //threadgroup q_t   * sq  = (threadgroup q_t   *) (shmem_f16 +                  0*DK); // holds the query data
-    threadgroup q4_t  * sq4 = (threadgroup q4_t  *) (shmem_f16 +                  0*DK); // same as above but in q4_t
-    threadgroup s_t   * ss  = (threadgroup s_t   *) (shmem_f16 + sgitg*SH       + Q*DK); // scratch buffer for attention
-    threadgroup s4_t  * ss4 = (threadgroup s4_t  *) (shmem_f16 + sgitg*SH       + Q*DK); // same as above but in s4_t
-    threadgroup float * sm  = (threadgroup float *) (shmem_f16 + sgitg*SH + 2*C + Q*DK); // scratch buffer for mask
-    threadgroup o4_t  * sr4 = (threadgroup o4_t  *) (shmem_f16 + sgitg*DV       + Q*T);  // scratch buffer for the results
+  //threadgroup q_t   * sq  = (threadgroup q_t   *) (shmem_f16 +                    0*DK); // holds the query data
+    threadgroup q4_t  * sq4 = (threadgroup q4_t  *) (shmem_f16 +                    0*DK); // same as above but in q4_t
+    threadgroup s_t   * ss  = (threadgroup s_t   *) (shmem_f16 +   sgitg*SH       + Q*DK); // scratch buffer for attention
+    threadgroup s4_t  * ss4 = (threadgroup s4_t  *) (shmem_f16 +   sgitg*SH       + Q*DK); // same as above but in s4_t
+    threadgroup float * sm  = (threadgroup float *) (shmem_f16 +   sgitg*SH + 2*C + Q*DK); // scratch buffer for mask
+    threadgroup o4_t  * sr4 = (threadgroup o4_t  *) (shmem_f16 + 2*sgitg*DV       + Q*T);  // scratch buffer for the results
 
     // store the result for all queries in local memory (the O matrix from the paper)
     o4_t lo[DV4/NL];
@@ -4157,7 +4521,7 @@ kernel void kernel_flash_attn_ext_vec(
            half4,  \
     float,         \
     float, float4, \
-           half4
+           float4
 
 typedef decltype(kernel_flash_attn_ext_vec<FA_TYPES, half4, 1, dequantize_f16_t4, half4, 1, dequantize_f16_t4, 128, 128, 4>) flash_attn_ext_vec_t;
 
@@ -4271,11 +4635,16 @@ kernel void kernel_cpy(
         device  const char * src0,
         device        char * dst,
         uint3   tgpig[[threadgroup_position_in_grid]],
+        uint    tiitg[[thread_index_in_threadgroup]],
         ushort3 tpitg[[thread_position_in_threadgroup]],
-        ushort3   ntg[[threads_per_threadgroup]]) {
+        ushort3  tptg[[threads_per_threadgroup]]) {
     const int i03 = tgpig[2];
     const int i02 = tgpig[1];
-    const int i01 = tgpig[0];
+    const int i01 = tgpig[0]*tptg.y + tiitg/tptg.x;
+
+    if (i01 >= args.ne01) {
+        return;
+    }
 
     const int64_t n = i03*args.ne02*args.ne01*args.ne00 + i02*args.ne01*args.ne00 + i01*args.ne00;
 
@@ -4286,7 +4655,7 @@ kernel void kernel_cpy(
 
     device T1 * dst_data = (device T1 *) (dst + i3*args.nb3 + i2*args.nb2 + i1*args.nb1 + i0*args.nb0);
 
-    for (int64_t i00 = tpitg.x; i00 < args.ne00; i00 += ntg.x) {
+    for (int64_t i00 = tiitg%tptg.x; i00 < args.ne00; i00 += tptg.x) {
         device const T0 * src = (device T0 *)(src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01 + i00*args.nb00);
         dst_data[i00] = (T1) src[0];
     }
@@ -4306,6 +4675,7 @@ template [[host_name("kernel_cpy_bf16_f32")]]  kernel kernel_cpy_t kernel_cpy<bf
 template [[host_name("kernel_cpy_bf16_bf16")]] kernel kernel_cpy_t kernel_cpy<bfloat, bfloat>;
 #endif
 
+// TODO: templetify these kernels
 kernel void kernel_cpy_f32_q8_0(
         constant ggml_metal_kargs_cpy & args,
         device const char * src0,
@@ -4329,23 +4699,7 @@ kernel void kernel_cpy_f32_q8_0(
     for (int64_t i00 = tpitg.x*QK8_0; i00 < args.ne00; i00 += ntg.x*QK8_0) {
         device const float * src = (device float *)(src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01 + i00*args.nb00);
 
-        float amax = 0.0f; // absolute max
-
-        for (int j = 0; j < QK8_0; j++) {
-            const float v = src[j];
-            amax = MAX(amax, fabs(v));
-        }
-
-        const float d = amax / ((1 << 7) - 1);
-        const float id = d ? 1.0f/d : 0.0f;
-
-        dst_data[i00/QK8_0].d = d;
-
-        for (int j = 0; j < QK8_0; ++j) {
-            const float x0 = src[j]*id;
-
-            dst_data[i00/QK8_0].qs[j] = round(x0);
-        }
+        quantize_q8_0(src, dst_data[i00/QK8_0]);
     }
 }
 
@@ -4372,32 +4726,7 @@ kernel void kernel_cpy_f32_q4_0(
     for (int64_t i00 = tpitg.x*QK4_0; i00 < args.ne00; i00 += ntg.x*QK4_0) {
         device const float * src = (device float *)(src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01 + i00*args.nb00);
 
-        float amax = 0.0f; // absolute max
-        float max  = 0.0f;
-
-        for (int j = 0; j < QK4_0; j++) {
-            const float v = src[j];
-            if (amax < fabs(v)) {
-                amax = fabs(v);
-                max  = v;
-            }
-        }
-
-        const float d = max / -8;
-        const float id = d ? 1.0f/d : 0.0f;
-
-        dst_data[i00/QK4_0].d = d;
-
-        for (int j = 0; j < QK4_0/2; ++j) {
-            const float x0 = src[0       + j]*id;
-            const float x1 = src[QK4_0/2 + j]*id;
-
-            const uint8_t xi0 = MIN(15, (int8_t)(x0 + 8.5f));
-            const uint8_t xi1 = MIN(15, (int8_t)(x1 + 8.5f));
-
-            dst_data[i00/QK4_0].qs[j]  = xi0;
-            dst_data[i00/QK4_0].qs[j] |= xi1 << 4;
-        }
+        quantize_q4_0(src, dst_data[i00/QK4_0]);
     }
 }
 
@@ -4424,31 +4753,7 @@ kernel void kernel_cpy_f32_q4_1(
     for (int64_t i00 = tpitg.x*QK4_1; i00 < args.ne00; i00 += ntg.x*QK4_1) {
         device const float * src = (device float *)(src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01 + i00*args.nb00);
 
-        float min = FLT_MAX;
-        float max = -FLT_MAX;
-
-        for (int j = 0; j < QK4_1; j++) {
-            const float v = src[j];
-            if (min > v) min = v;
-            if (max < v) max = v;
-        }
-
-        const float d = (max - min) / ((1 << 4) - 1);
-        const float id = d ? 1.0f/d : 0.0f;
-
-        dst_data[i00/QK4_1].d = d;
-        dst_data[i00/QK4_1].m = min;
-
-        for (int j = 0; j < QK4_1/2; ++j) {
-            const float x0 = (src[0       + j] - min)*id;
-            const float x1 = (src[QK4_1/2 + j] - min)*id;
-
-            const uint8_t xi0 = MIN(15, (int8_t)(x0 + 0.5f));
-            const uint8_t xi1 = MIN(15, (int8_t)(x1 + 0.5f));
-
-            dst_data[i00/QK4_1].qs[j]  = xi0;
-            dst_data[i00/QK4_1].qs[j] |= xi1 << 4;
-        }
+        quantize_q4_1(src, dst_data[i00/QK4_1]);
     }
 }
 
@@ -4475,38 +4780,7 @@ kernel void kernel_cpy_f32_q5_0(
     for (int64_t i00 = tpitg.x*QK5_0; i00 < args.ne00; i00 += ntg.x*QK5_0) {
         device const float * src = (device float *)(src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01 + i00*args.nb00);
 
-        float amax = 0.0f; // absolute max
-        float max  = 0.0f;
-
-        for (int j = 0; j < QK5_0; j++) {
-            const float v = src[j];
-            if (amax < fabs(v)) {
-                amax = fabs(v);
-                max  = v;
-            }
-        }
-
-        const float d = max / -16;
-        const float id = d ? 1.0f/d : 0.0f;
-
-        dst_data[i00/QK5_0].d = d;
-
-        uint32_t qh = 0;
-        for (int j = 0; j < QK5_0/2; ++j) {
-            const float x0 = src[0       + j]*id;
-            const float x1 = src[QK5_0/2 + j]*id;
-
-            const uint8_t xi0 = MIN(31, (int8_t)(x0 + 16.5f));
-            const uint8_t xi1 = MIN(31, (int8_t)(x1 + 16.5f));
-
-            dst_data[i00/QK5_0].qs[j] = (xi0 & 0xf) | ((xi1 & 0xf) << 4);
-            qh |= ((xi0 & 0x10u) >> 4) << (j + 0);
-            qh |= ((xi1 & 0x10u) >> 4) << (j + QK5_0/2);
-        }
-        thread const uint8_t * qh8 = (thread const uint8_t *)&qh;
-        for (int j = 0; j < 4; ++j) {
-            dst_data[i00/QK5_0].qh[j] = qh8[j];
-        }
+        quantize_q5_0(src, dst_data[i00/QK5_0]);
     }
 }
 
@@ -4533,49 +4807,8 @@ kernel void kernel_cpy_f32_q5_1(
     for (int64_t i00 = tpitg.x*QK5_1; i00 < args.ne00; i00 += ntg.x*QK5_1) {
         device const float * src = (device float *)(src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01 + i00*args.nb00);
 
-        float max = src[0];
-        float min = src[0];
-
-        for (int j = 1; j < QK5_1; j++) {
-            const float v = src[j];
-            min = v < min ? v : min;
-            max = v > max ? v : max;
-        }
-
-        const float d = (max - min) / 31;
-        const float id = d ? 1.0f/d : 0.0f;
-
-        dst_data[i00/QK5_1].d = d;
-        dst_data[i00/QK5_1].m = min;
-
-        uint32_t qh = 0;
-        for (int j = 0; j < QK5_1/2; ++j) {
-            const float x0 = (src[0       + j] - min)*id;
-            const float x1 = (src[QK5_1/2 + j] - min)*id;
-
-            const uint8_t xi0 = (uint8_t)(x0 + 0.5f);
-            const uint8_t xi1 = (uint8_t)(x1 + 0.5f);
-
-            dst_data[i00/QK5_1].qs[j] = (xi0 & 0xf) | ((xi1 & 0xf) << 4);
-            qh |= ((xi0 & 0x10u) >> 4) << (j + 0);
-            qh |= ((xi1 & 0x10u) >> 4) << (j + QK5_1/2);
-        }
-        thread const uint8_t * qh8 = (thread const uint8_t *)&qh;
-        for (int j = 0; j < 4; ++j) {
-            dst_data[i00/QK5_1].qh[j] = qh8[j];
-        }
-    }
-}
-
-static inline int best_index_int8(int n, constant float * val, float x) {
-    if (x <= val[0]) return 0;
-    if (x >= val[n-1]) return n-1;
-    int ml = 0, mu = n-1;
-    while (mu-ml > 1) {
-        int mav = (ml+mu)/2;
-        if (x < val[mav]) mu = mav; else ml = mav;
+        quantize_q5_1(src, dst_data[i00/QK5_1]);
     }
-    return x - val[mu-1] < val[mu] - x ? mu-1 : mu;
 }
 
 kernel void kernel_cpy_f32_iq4_nl(
@@ -4601,40 +4834,7 @@ kernel void kernel_cpy_f32_iq4_nl(
     for (int64_t i00 = tpitg.x*QK4_NL; i00 < args.ne00; i00 += ntg.x*QK4_NL) {
         device const float * src = (device float *)(src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01 + i00*args.nb00);
 
-        float amax = 0.0f; // absolute max
-        float max  = 0.0f;
-
-        for (int j = 0; j < QK4_NL; j++) {
-            const float v = src[j];
-            if (amax < fabs(v)) {
-                amax = fabs(v);
-                max  = v;
-            }
-        }
-
-        const float d = max / kvalues_iq4nl_f[0];
-        const float id = d ? 1.0f/d : 0.0f;
-
-        float sumqx = 0, sumq2 = 0;
-        for (int j = 0; j < QK4_NL/2; ++j) {
-            const float x0 = src[0        + j]*id;
-            const float x1 = src[QK4_NL/2 + j]*id;
-
-            const uint8_t xi0 = best_index_int8(16, kvalues_iq4nl_f, x0);
-            const uint8_t xi1 = best_index_int8(16, kvalues_iq4nl_f, x1);
-
-            dst_data[i00/QK4_NL].qs[j] = xi0 | (xi1 << 4);
-
-            const float v0 = kvalues_iq4nl_f[xi0];
-            const float v1 = kvalues_iq4nl_f[xi1];
-            const float w0 = src[0        + j]*src[0        + j];
-            const float w1 = src[QK4_NL/2 + j]*src[QK4_NL/2 + j];
-            sumqx += w0*v0*src[j] + w1*v1*src[QK4_NL/2 + j];
-            sumq2 += w0*v0*v0 + w1*v1*v1;
-
-        }
-
-        dst_data[i00/QK4_NL].d = sumq2 > 0 ? sumqx/sumq2 : d;
+        quantize_iq4_nl(src, dst_data[i00/QK4_NL]);
     }
 }
 
@@ -6315,10 +6515,10 @@ kernel void kernel_mul_mv_iq4_xs_f32(
 
 template<typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread float4x4 &)>
 kernel void kernel_get_rows_q(
+        constant ggml_metal_kargs_get_rows & args,
         device const  void * src0,
         device const  void * src1,
         device       float * dst,
-        constant ggml_metal_kargs_get_rows & args,
         uint3                tgpig[[threadgroup_position_in_grid]],
         uint                 tiitg[[thread_index_in_threadgroup]],
         uint3                tptg [[threads_per_threadgroup]]) {
@@ -6338,10 +6538,10 @@ kernel void kernel_get_rows_q(
 
 template<typename T>
 kernel void kernel_get_rows_f(
+        constant ggml_metal_kargs_get_rows & args,
         device const  void * src0,
         device const  void * src1,
         device       float * dst,
-        constant ggml_metal_kargs_get_rows & args,
         uint3                tgpig[[threadgroup_position_in_grid]],
         uint                 tiitg[[thread_index_in_threadgroup]],
         uint3                tptg [[threads_per_threadgroup]]) {
@@ -6359,10 +6559,10 @@ kernel void kernel_get_rows_f(
 }
 
 kernel void kernel_get_rows_i32(
+        constant ggml_metal_kargs_get_rows & args,
         device const  void * src0,
         device const  void * src1,
         device     int32_t * dst,
-        constant ggml_metal_kargs_get_rows & args,
         uint3                tgpig[[threadgroup_position_in_grid]],
         uint                 tiitg[[thread_index_in_threadgroup]],
         uint3                tptg [[threads_per_threadgroup]]) {
@@ -6379,6 +6579,67 @@ kernel void kernel_get_rows_i32(
     }
 }
 
+template<typename block_q, void (*quantize_func)(device const float *, device block_q &)>
+kernel void kernel_set_rows_q32(
+        constant ggml_metal_kargs_set_rows & args,
+        device const  void * src0,
+        device const  void * src1,
+        device       float * dst,
+        uint3                tgpig[[threadgroup_position_in_grid]],
+        uint                 tiitg[[thread_index_in_threadgroup]],
+        uint3                tptg [[threads_per_threadgroup]]) {
+    const int32_t i03 = tgpig.z;
+    const int32_t i02 = tgpig.y;
+
+    const int32_t i12 = i03%args.ne12;
+    const int32_t i11 = i02%args.ne11;
+
+    const int32_t i01 = tgpig.x*tptg.y + tiitg/tptg.x;
+    if (i01 >= args.ne01) {
+        return;
+    }
+
+    const int32_t i10 = i01;
+    const int64_t i1 = ((const device int64_t *) ((const device char *) src1 + i10*args.nb10 + i11*args.nb11 + i12*args.nb12))[0];
+
+          device block_q * dst_row = (      device block_q *) ((      device char *) dst  +  i1*args.nb1  + i02*args.nb2  + i03*args.nb3);
+    const device float   * src_row = (const device float   *) ((const device char *) src0 + i01*args.nb01 + i02*args.nb02 + i03*args.nb03);
+
+    for (int ind = tiitg%tptg.x; ind < args.nk0; ind += tptg.x) {
+        quantize_func(src_row + 32*ind, dst_row[ind]);
+    }
+}
+
+template<typename T>
+kernel void kernel_set_rows_f(
+        constant ggml_metal_kargs_set_rows & args,
+        device const  void * src0,
+        device const  void * src1,
+        device       float * dst,
+        uint3                tgpig[[threadgroup_position_in_grid]],
+        uint                 tiitg[[thread_index_in_threadgroup]],
+        uint3                tptg [[threads_per_threadgroup]]) {
+    const int32_t i03 = tgpig.z;
+    const int32_t i02 = tgpig.y;
+
+    const int32_t i12 = i03%args.ne12;
+    const int32_t i11 = i02%args.ne11;
+
+    const int32_t i01 = tgpig.x*tptg.y + tiitg/tptg.x;
+    if (i01 >= args.ne01) {
+        return;
+    }
+
+    const int32_t i10 = i01;
+    const int64_t i1 = ((const device int64_t *) ((const device char *) src1 + i10*args.nb10 + i11*args.nb11 + i12*args.nb12))[0];
+
+    device T     * dst_row = (      device T     *) ((      device char *) dst  +  i1*args.nb1  + i02*args.nb2  + i03*args.nb3);
+    const device float * src_row = (const device float *) ((const device char *) src0 + i01*args.nb01 + i02*args.nb02 + i03*args.nb03);
+
+    for (int ind = tiitg%tptg.x; ind < args.nk0; ind += tptg.x) {
+        dst_row[ind] = (T) src_row[ind];
+    }
+}
 
 #define BLOCK_SIZE_M 64 // 8 simdgroup matrices from matrix A
 #define BLOCK_SIZE_N 32 // 4 simdgroup matrices from matrix B
@@ -6802,6 +7063,27 @@ template [[host_name("kernel_get_rows_iq1_m")]]   kernel get_rows_q_t kernel_get
 template [[host_name("kernel_get_rows_iq4_nl")]]  kernel get_rows_q_t kernel_get_rows_q<block_iq4_nl,  2,     dequantize_iq4_nl>;
 template [[host_name("kernel_get_rows_iq4_xs")]]  kernel get_rows_q_t kernel_get_rows_q<block_iq4_xs,  QK_NL, dequantize_iq4_xs>;
 
+//
+// set rows
+//
+
+typedef decltype(kernel_set_rows_f<float>) set_rows_f_t;
+
+template [[host_name("kernel_set_rows_f32")]]  kernel set_rows_f_t kernel_set_rows_f<float>;
+template [[host_name("kernel_set_rows_f16")]]  kernel set_rows_f_t kernel_set_rows_f<half>;
+#if defined(GGML_METAL_USE_BF16)
+template [[host_name("kernel_set_rows_bf16")]] kernel set_rows_f_t kernel_set_rows_f<bfloat>;
+#endif
+
+typedef decltype(kernel_set_rows_q32<block_q8_0, quantize_q8_0>) set_rows_q32_t;
+
+template [[host_name("kernel_set_rows_q8_0")]]   kernel set_rows_q32_t kernel_set_rows_q32<block_q8_0,   quantize_q8_0>;
+template [[host_name("kernel_set_rows_q4_0")]]   kernel set_rows_q32_t kernel_set_rows_q32<block_q4_0,   quantize_q4_0>;
+template [[host_name("kernel_set_rows_q4_1")]]   kernel set_rows_q32_t kernel_set_rows_q32<block_q4_1,   quantize_q4_1>;
+template [[host_name("kernel_set_rows_q5_0")]]   kernel set_rows_q32_t kernel_set_rows_q32<block_q5_0,   quantize_q5_0>;
+template [[host_name("kernel_set_rows_q5_1")]]   kernel set_rows_q32_t kernel_set_rows_q32<block_q5_1,   quantize_q5_1>;
+template [[host_name("kernel_set_rows_iq4_nl")]] kernel set_rows_q32_t kernel_set_rows_q32<block_iq4_nl, quantize_iq4_nl>;
+
 //
 // matrix-matrix multiplication
 //