llama_cpp 0.5.2 → 0.6.0

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

@@ -24,12 +24,59 @@ typedef struct {
     int8_t  qs[QK8_0]; // quants
 } block_q8_0;
 
+// general-purpose kernel for addition of two tensors
+// pros: works for non-contiguous tensors, supports broadcast across dims 1, 2 and 3
+// cons: not very efficient
 kernel void kernel_add(
-        device const float4 * src0,
-        device const float4 * src1,
-        device       float4 * dst,
-        uint tpig[[thread_position_in_grid]]) {
-    dst[tpig] = src0[tpig] + src1[tpig];
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        constant  int64_t & ne00,
+        constant  int64_t & ne01,
+        constant  int64_t & ne02,
+        constant  int64_t & ne03,
+        constant  int64_t & nb00,
+        constant  int64_t & nb01,
+        constant  int64_t & nb02,
+        constant  int64_t & nb03,
+        constant  int64_t & ne10,
+        constant  int64_t & ne11,
+        constant  int64_t & ne12,
+        constant  int64_t & ne13,
+        constant  int64_t & nb10,
+        constant  int64_t & nb11,
+        constant  int64_t & nb12,
+        constant  int64_t & nb13,
+        constant  int64_t & ne0,
+        constant  int64_t & ne1,
+        constant  int64_t & ne2,
+        constant  int64_t & ne3,
+        constant  int64_t & nb0,
+        constant  int64_t & nb1,
+        constant  int64_t & nb2,
+        constant  int64_t & nb3,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {
+    const int64_t i03 = tgpig.z;
+    const int64_t i02 = tgpig.y;
+    const int64_t i01 = tgpig.x;
+
+    const int64_t i13 = i03 % ne13;
+    const int64_t i12 = i02 % ne12;
+    const int64_t i11 = i01 % ne11;
+
+    device const char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01 + tpitg.x*nb00;
+    device const char * src1_ptr = src1 + i13*nb13 + i12*nb12 + i11*nb11 + tpitg.x*nb10;
+    device       char * dst_ptr  = dst  + i03*nb3  + i02*nb2  + i01*nb1  + tpitg.x*nb0;
+
+    for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) {
+        ((device float *)dst_ptr)[0] = ((device float *)src0_ptr)[0] + ((device float *)src1_ptr)[0];
+
+        src0_ptr += ntg.x*nb00;
+        src1_ptr += ntg.x*nb10;
+        dst_ptr  += ntg.x*nb0;
+    }
 }
 
 // assumption: src1 is a row
@@ -38,7 +85,7 @@ kernel void kernel_add_row(
         device const float4 * src0,
         device const float4 * src1,
         device       float4 * dst,
-        constant   int64_t & nb,
+        constant    int64_t & nb [[buffer(27)]],
         uint tpig[[thread_position_in_grid]]) {
     dst[tpig] = src0[tpig] + src1[tpig % nb];
 }
@@ -118,7 +165,7 @@ kernel void kernel_soft_max(
     device       float * pdst  = dst  + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
 
     // parallel max
-    float lmax = psrc0[tpitg[0]];
+    float lmax = tpitg[0] < ne00 ? psrc0[tpitg[0]] : -INFINITY;
     for (int i00 = tpitg[0] + ntg[0]; i00 < ne00; i00 += ntg[0]) {
         lmax = MAX(lmax, psrc0[i00]);
     }
@@ -158,7 +205,7 @@ kernel void kernel_soft_max_4(
     device       float4 * pdst4 = (device       float4 *)(dst  + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00);
 
     // parallel max
-    float4 lmax4 = psrc4[tpitg[0]];
+    float4 lmax4 = tpitg[0] < ne00/4 ? psrc4[tpitg[0]] : -INFINITY;
     for (int i00 = tpitg[0] + ntg[0]; i00 < ne00/4; i00 += ntg[0]) {
         lmax4 = fmax(lmax4, psrc4[i00]);
     }
@@ -523,6 +570,79 @@ kernel void kernel_mul_mat_q8_0_f32(
     }
 }
 
+#define N_F32_F32 4
+
+kernel void kernel_mul_mat_f32_f32(
+        device const  char * src0,
+        device const  char * 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,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint tiisg[[thread_index_in_simdgroup]]) {
+
+    const int64_t r0 = tgpig.x;
+    const int64_t rb = tgpig.y*N_F32_F32;
+    const int64_t im = tgpig.z;
+
+    device const float * x = (device const float *) (src0 + r0*nb01 + im/(ne12/ne02)*nb02);
+
+    if (ne00 < 128) {
+        for (int row = 0; row < N_F32_F32; ++row) {
+            int r1 = rb + row;
+            if (r1 >= ne11) {
+                break;
+            }
+
+            device const float * y = (device const float *) (src1 + r1*nb11 + im*nb12);
+
+            float sumf = 0;
+            for (int i = tiisg; i < ne00; i += 32) {
+                sumf += (float) x[i] * (float) y[i];
+            }
+
+            float all_sum = simd_sum(sumf);
+            if (tiisg == 0) {
+                dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
+            }
+        }
+    } else {
+        device const float4 * x4 = (device const float4 *)x;
+        for (int row = 0; row < N_F32_F32; ++row) {
+            int r1 = rb + row;
+            if (r1 >= ne11) {
+                break;
+            }
+
+            device const float  * y  = (device const float  *) (src1 + r1*nb11 + im*nb12);
+            device const float4 * y4 = (device const float4 *) y;
+
+            float sumf = 0;
+            for (int i = tiisg; i < ne00/4; i += 32) {
+                for (int k = 0; k < 4; ++k) sumf += (float) x4[i][k] * y4[i][k];
+            }
+
+            float all_sum = simd_sum(sumf);
+            if (tiisg == 0) {
+                for (int i = 4*(ne00/4); i < ne00; ++i) all_sum += (float) x[i] * y[i];
+                dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
+            }
+        }
+    }
+}
+
 kernel void kernel_mul_mat_f16_f32_1row(
         device const  char * src0,
         device const  char * src1,
@@ -733,30 +853,61 @@ kernel void kernel_alibi_f32(
     }
 }
 
+typedef void (rope_t)(
+        device const    void * src0,
+        device const int32_t * src1,
+        device         float * dst,
+        constant     int64_t & ne00,
+        constant     int64_t & ne01,
+        constant     int64_t & ne02,
+        constant     int64_t & ne03,
+        constant    uint64_t & nb00,
+        constant    uint64_t & nb01,
+        constant    uint64_t & nb02,
+        constant    uint64_t & nb03,
+        constant     int64_t & ne0,
+        constant     int64_t & ne1,
+        constant     int64_t & ne2,
+        constant     int64_t & ne3,
+        constant    uint64_t & nb0,
+        constant    uint64_t & nb1,
+        constant    uint64_t & nb2,
+        constant    uint64_t & nb3,
+        constant         int & n_past,
+        constant         int & n_dims,
+        constant         int & mode,
+        constant       float & freq_base,
+        constant       float & freq_scale,
+        uint  tiitg[[thread_index_in_threadgroup]],
+        uint3 tptg[[threads_per_threadgroup]],
+        uint3 tgpig[[threadgroup_position_in_grid]]);
+
+template<typename T>
 kernel void kernel_rope(
-        device const  void * src0,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant   int64_t & ne03,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant  uint64_t & nb03,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   int64_t & ne2,
-        constant   int64_t & ne3,
-        constant  uint64_t & nb0,
-        constant  uint64_t & nb1,
-        constant  uint64_t & nb2,
-        constant  uint64_t & nb3,
-        constant       int & n_past,
-        constant       int & n_dims,
-        constant       int & mode,
-        constant     float & freq_base,
-        constant     float & freq_scale,
+        device const    void * src0,
+        device const int32_t * src1,
+        device         float * dst,
+        constant     int64_t & ne00,
+        constant     int64_t & ne01,
+        constant     int64_t & ne02,
+        constant     int64_t & ne03,
+        constant    uint64_t & nb00,
+        constant    uint64_t & nb01,
+        constant    uint64_t & nb02,
+        constant    uint64_t & nb03,
+        constant     int64_t & ne0,
+        constant     int64_t & ne1,
+        constant     int64_t & ne2,
+        constant     int64_t & ne3,
+        constant    uint64_t & nb0,
+        constant    uint64_t & nb1,
+        constant    uint64_t & nb2,
+        constant    uint64_t & nb3,
+        constant         int & n_past,
+        constant         int & n_dims,
+        constant         int & mode,
+        constant       float & freq_base,
+        constant       float & freq_scale,
         uint  tiitg[[thread_index_in_threadgroup]],
         uint3 tptg[[threads_per_threadgroup]],
         uint3 tgpig[[threadgroup_position_in_grid]]) {
@@ -766,7 +917,9 @@ kernel void kernel_rope(
 
     const bool is_neox = mode & 2;
 
-    const int64_t p = ((mode & 1) == 0 ? n_past + i2 : i2);
+    device const int32_t * pos = src1;
+
+    const int64_t p = pos[i2];
 
     const float theta_0 = freq_scale * (float)p;
     const float inv_ndims = -1.f/n_dims;
@@ -778,11 +931,11 @@ kernel void kernel_rope(
             const float cos_theta = cos(theta);
             const float sin_theta = sin(theta);
 
-            device const float * const src = (device float *)((device char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
-            device       float * dst_data  = (device float *)((device char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
+            device const T * const src = (device T *)((device char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
+            device       T * dst_data  = (device T *)((device char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
 
-            const float x0 = src[0];
-            const float x1 = src[1];
+            const T x0 = src[0];
+            const T x1 = src[1];
 
             dst_data[0] = x0*cos_theta - x1*sin_theta;
             dst_data[1] = x0*sin_theta + x1*cos_theta;
@@ -797,8 +950,8 @@ kernel void kernel_rope(
 
                 const int64_t i0 = ib*n_dims + ic/2;
 
-                device const float * const src = (device float *)((device char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
-                device       float * dst_data  = (device float *)((device char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
+                device const T * const src = (device T *)((device char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
+                device       T * dst_data  = (device T *)((device char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
 
                 const float x0 = src[0];
                 const float x1 = src[n_dims/2];
@@ -810,6 +963,9 @@ kernel void kernel_rope(
     }
 }
 
+template [[host_name("kernel_rope_f32")]] kernel rope_t kernel_rope<float>;
+template [[host_name("kernel_rope_f16")]] kernel rope_t kernel_rope<half>;
+
 kernel void kernel_cpy_f16_f16(
         device const half * src0,
         device       half * dst,
@@ -1200,8 +1356,8 @@ kernel void kernel_mul_mat_q3_K_f32(
 
     float yl[32];
 
-    const uint16_t kmask1 = 0x3030;
-    const uint16_t kmask2 = 0x0f0f;
+    //const uint16_t kmask1 = 0x3030;
+    //const uint16_t kmask2 = 0x0f0f;
 
     const int tid = tiisg/4;
     const int ix  = tiisg%4;
@@ -1321,7 +1477,6 @@ kernel void kernel_mul_mat_q3_K_f32(
             dst[r1*ne0 + r2*ne0*ne1 + first_row + row] = sumf1[row];
         }
     }
-
 }
 #else
 kernel void kernel_mul_mat_q3_K_f32(
@@ -1400,13 +1555,13 @@ kernel void kernel_mul_mat_q4_K_f32(
         device const float * src1,
         device       float * dst,
         constant   int64_t & ne00,
-        constant   int64_t & ne01[[buffer(4)]],
-        constant   int64_t & ne02[[buffer(5)]],
-        constant   int64_t & ne10[[buffer(9)]],
-        constant   int64_t & ne12[[buffer(11)]],
-        constant   int64_t & ne0[[buffer(15)]],
-        constant   int64_t & ne1[[buffer(16)]],
-        constant   uint    & gqa[[buffer(17)]],
+        constant   int64_t & ne01 [[buffer(4)]],
+        constant   int64_t & ne02 [[buffer(5)]],
+        constant   int64_t & ne10 [[buffer(9)]],
+        constant   int64_t & ne12 [[buffer(11)]],
+        constant   int64_t & ne0  [[buffer(15)]],
+        constant   int64_t & ne1  [[buffer(16)]],
+        constant   uint    & gqa  [[buffer(17)]],
         uint3 tgpig[[threadgroup_position_in_grid]],
         uint tiisg[[thread_index_in_simdgroup]],
         uint sgitg[[simdgroup_index_in_threadgroup]]) {
@@ -1865,6 +2020,15 @@ kernel void kernel_mul_mat_q6_K_f32(
 
 //============================= templates and their specializations =============================
 
+// 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) {
+    float4x4 temp = *(((device float4x4 *)src));
+    for (int i = 0; i < 16; i++){
+        reg[i/4][i%4] = temp[i/4][i%4];
+    }
+}
+
 template <typename type4x4>
 void dequantize_f16(device const half4x4 * src, short il, thread type4x4 & reg) {
     half4x4 temp = *(((device half4x4 *)src));
@@ -1875,7 +2039,6 @@ void dequantize_f16(device const half4x4 * src, short il, thread type4x4 & reg)
 
 template <typename type4x4>
 void dequantize_q4_0(device const block_q4_0 *xb, short il, thread type4x4 & reg) {
-
     device const uint16_t * qs = ((device const uint16_t *)xb + 1);
     const float d1 = il ? (xb->d / 16.h) : xb->d;
     const float d2 = d1 / 256.f;
@@ -1887,12 +2050,10 @@ void dequantize_q4_0(device const block_q4_0 *xb, short il, thread type4x4 & reg
         reg[i/2][2*(i%2)+0] = d1 * (qs[i] & mask0) + md;
         reg[i/2][2*(i%2)+1] = d2 * (qs[i] & mask1) + md;
     }
-
 }
 
 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);
     const float d1 = il ? (xb->d / 16.h) : xb->d;
     const float d2 = d1 / 256.f;
@@ -1964,7 +2125,6 @@ void dequantize_q3_K(device const block_q3_K *xb, short il, thread type4x4 & reg
     for (int i = 0; i < 16; ++i) {
         reg[i/4][i%4] = dl * (q[i] & mask) - (h[i] & m ? 0 : ml);
     }
-
 #else
     float    kcoef = il&1 ? 1.f/16.f : 1.f;
     uint16_t kmask = il&1 ? 0xF0     : 0x0F;
@@ -2008,7 +2168,6 @@ void dequantize_q4_K(device const block_q4_K *xb, short il, thread type4x4 & reg
     for (int i = 0; i < 16; ++i) {
         reg[i/4][i%4] = dl * (q[i] & mask) - ml;
     }
-
 }
 
 template <typename type4x4>
@@ -2110,22 +2269,25 @@ kernel void kernel_get_rows(
 // each block_q contains 16*nl weights
 template<typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread half4x4 &)>
 kernel void kernel_mul_mm(device const  uchar * src0,
-                           device const  float * src1,
-                           device        float * dst,
-                           constant    int64_t & ne00,
-                           constant    int64_t & ne02,
-                           constant    int64_t & nb01,
-                           constant    int64_t & nb02,
-                           constant    int64_t & ne12,
-                           constant    int64_t & ne0,
-                           constant    int64_t & ne1,
-                           constant    uint & gqa,
-                           threadgroup   uchar * shared_memory [[threadgroup(0)]],
-                           uint3                 tgpig[[threadgroup_position_in_grid]],
-                           uint                  tiitg[[thread_index_in_threadgroup]],
-                           uint                  sgitg[[simdgroup_index_in_threadgroup]]) {
-
-    threadgroup half * sa = ((threadgroup half *)shared_memory);
+                          device const  uchar * src1,
+                          device        float * dst,
+                          constant    int64_t & ne00,
+                          constant    int64_t & ne02,
+                          constant    int64_t & nb01,
+                          constant    int64_t & nb02,
+                          constant    int64_t & ne12,
+                          constant    int64_t & nb10,
+                          constant    int64_t & nb11,
+                          constant    int64_t & nb12,
+                          constant    int64_t & ne0,
+                          constant    int64_t & ne1,
+                          constant       uint & gqa,
+                          threadgroup   uchar * shared_memory [[threadgroup(0)]],
+                          uint3                 tgpig[[threadgroup_position_in_grid]],
+                          uint                  tiitg[[thread_index_in_threadgroup]],
+                          uint                  sgitg[[simdgroup_index_in_threadgroup]]) {
+
+    threadgroup half  * sa = (threadgroup half  *)(shared_memory);
     threadgroup float * sb = (threadgroup float *)(shared_memory + 4096);
 
     const uint r0 = tgpig.y;
@@ -2138,7 +2300,7 @@ kernel void kernel_mul_mm(device const  uchar * src0,
     short thread_row = ((short)tiitg/THREAD_PER_ROW) < n_rows ? ((short)tiitg/THREAD_PER_ROW) : n_rows - 1;
     short thread_col = ((short)tiitg/THREAD_PER_COL) < n_cols ? ((short)tiitg/THREAD_PER_COL) : n_cols - 1;
 
-    simdgroup_half8x8 ma[4];
+    simdgroup_half8x8  ma[4];
     simdgroup_float8x8 mb[2];
     simdgroup_float8x8 c_res[8];
     for (int i = 0; i < 8; i++){
@@ -2146,10 +2308,15 @@ kernel void kernel_mul_mm(device const  uchar * src0,
     }
 
     short il = (tiitg % THREAD_PER_ROW);
-    uint offset0 = im/gqa*nb02; ushort offset1 = il/nl;
-    device const block_q  * x = (device const block_q  *)(src0 + (r0 * BLOCK_SIZE_M + thread_row) * nb01 + offset0) + offset1;
-    device const float * y = src1 + (r1 * BLOCK_SIZE_N + thread_col) * ne00 \
-                             + BLOCK_SIZE_K / THREAD_PER_COL * (tiitg % THREAD_PER_COL) + im * ne00 * ne1;
+
+    uint   offset0 = im/gqa*nb02;
+    ushort offset1 = il/nl;
+
+    device const block_q * x = (device const block_q *)(src0 + (r0 * BLOCK_SIZE_M + thread_row) * nb01 + offset0) + offset1;
+    device const float   * y = (device const float   *)(src1
+        + nb12 * im
+        + nb11 * (r1 * BLOCK_SIZE_N + thread_col)
+        + nb10 * (BLOCK_SIZE_K / THREAD_PER_COL * (tiitg % THREAD_PER_COL)));
 
     for (int loop_k = 0; loop_k < ne00; loop_k += BLOCK_SIZE_K) {
         //load data and store to threadgroup memory
@@ -2229,6 +2396,7 @@ kernel void kernel_mul_mm(device const  uchar * src0,
 typedef void (get_rows_t)(device const void *, device const int *, device float *, constant int64_t &, \
                           constant uint64_t &, constant uint64_t &, uint, uint, uint);
 
+template [[host_name("kernel_get_rows_f32")]]  kernel get_rows_t kernel_get_rows<float4x4,   1, dequantize_f32>;
 template [[host_name("kernel_get_rows_f16")]]  kernel get_rows_t kernel_get_rows<half4x4,    1, dequantize_f16>;
 template [[host_name("kernel_get_rows_q4_0")]] kernel get_rows_t kernel_get_rows<block_q4_0, 2, dequantize_q4_0>;
 template [[host_name("kernel_get_rows_q4_1")]] kernel get_rows_t kernel_get_rows<block_q4_1, 2, dequantize_q4_1>;
@@ -2239,14 +2407,28 @@ template [[host_name("kernel_get_rows_q4_K")]] kernel get_rows_t kernel_get_rows
 template [[host_name("kernel_get_rows_q5_K")]] kernel get_rows_t kernel_get_rows<block_q5_K, QK_NL, dequantize_q5_K>;
 template [[host_name("kernel_get_rows_q6_K")]] kernel get_rows_t kernel_get_rows<block_q6_K, QK_NL, dequantize_q6_K>;
 
-typedef void (mat_mm_t)(device const uchar *, device const float *, device float *, constant int64_t &,\
-                             constant int64_t &, constant int64_t &, constant int64_t &, constant int64_t &, \
-                             constant int64_t &, constant int64_t &, constant uint &, threadgroup uchar *, uint3, uint, uint);
-
-template [[host_name("kernel_mul_mm_f16_f32")]]  kernel mat_mm_t kernel_mul_mm<half4x4,    1, dequantize_f16>;
-template [[host_name("kernel_mul_mm_q4_0_f32")]] kernel mat_mm_t kernel_mul_mm<block_q4_0, 2, dequantize_q4_0>;
-template [[host_name("kernel_mul_mm_q4_1_f32")]] kernel mat_mm_t kernel_mul_mm<block_q4_1, 2, dequantize_q4_1>;
-template [[host_name("kernel_mul_mm_q8_0_f32")]] kernel mat_mm_t kernel_mul_mm<block_q8_0, 2, dequantize_q8_0>;
+typedef void (mat_mm_t)(
+        device const  uchar * src0,
+        device const  uchar * src1,
+        device        float * dst,
+        constant    int64_t & ne00,
+        constant    int64_t & ne02,
+        constant    int64_t & nb01,
+        constant    int64_t & nb02,
+        constant    int64_t & ne12,
+        constant    int64_t & nb10,
+        constant    int64_t & nb11,
+        constant    int64_t & nb12,
+        constant    int64_t & ne0,
+        constant    int64_t & ne1,
+        constant       uint & gqa,
+        threadgroup uchar *, uint3, uint, uint);
+
+template [[host_name("kernel_mul_mm_f32_f32")]]  kernel mat_mm_t kernel_mul_mm<float4x4,   1,     dequantize_f32>;
+template [[host_name("kernel_mul_mm_f16_f32")]]  kernel mat_mm_t kernel_mul_mm<half4x4,    1,     dequantize_f16>;
+template [[host_name("kernel_mul_mm_q4_0_f32")]] kernel mat_mm_t kernel_mul_mm<block_q4_0, 2,     dequantize_q4_0>;
+template [[host_name("kernel_mul_mm_q4_1_f32")]] kernel mat_mm_t kernel_mul_mm<block_q4_1, 2,     dequantize_q4_1>;
+template [[host_name("kernel_mul_mm_q8_0_f32")]] kernel mat_mm_t kernel_mul_mm<block_q8_0, 2,     dequantize_q8_0>;
 template [[host_name("kernel_mul_mm_q2_K_f32")]] kernel mat_mm_t kernel_mul_mm<block_q2_K, QK_NL, dequantize_q2_K>;
 template [[host_name("kernel_mul_mm_q3_K_f32")]] kernel mat_mm_t kernel_mul_mm<block_q3_K, QK_NL, dequantize_q3_K>;
 template [[host_name("kernel_mul_mm_q4_K_f32")]] kernel mat_mm_t kernel_mul_mm<block_q4_K, QK_NL, dequantize_q4_K>;

data/ext/llama_cpp/src/ggml-opencl.cpp

@@ -847,7 +847,7 @@ std::array<std::string, 2> mul_str_values = {
     "mul_f32", "float"
 };
 
-std::string& replace(std::string& s, const std::string& from, const std::string& to) {
+static std::string& replace(std::string& s, const std::string& from, const std::string& to) {
     size_t pos = 0;
     while ((pos = s.find(from, pos)) != std::string::npos) {
          s.replace(pos, from.length(), to);
@@ -856,7 +856,7 @@ std::string& replace(std::string& s, const std::string& from, const std::string&
     return s;
 }
 
-std::string generate_kernels() {
+static std::string generate_kernels() {
     std::stringstream src;
     src << program_source << '\n';
     src << k_quants_source << '\n';
@@ -1788,7 +1788,7 @@ bool ggml_cl_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tens
     return false;
 }
 
-bool ggml_cl_mul_mat_use_f16(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * /* dst */) {
+static bool ggml_cl_mul_mat_use_f16(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * /* dst */) {
     // If device doesn't support FP16
     if (!fp16_support) {
         return false;