@fugood/llama.node 0.3.13 → 0.3.14

package/src/llama.cpp/ggml/src/ggml-cpu/ggml-cpu.c

@@ -112,7 +112,8 @@ struct ggml_arm_arch_features_type {
     int has_i8mm;
     int has_sve;
     int sve_cnt;
-} ggml_arm_arch_features = {-1, -1, -1, -1, 0};
+    int has_sme;
+} ggml_arm_arch_features = {-1, -1, -1, -1, 0, -1};
 #endif
 
 
@@ -236,6 +237,8 @@ typedef pthread_t ggml_thread_t;
 #else
 #if defined(__POWER9_VECTOR__)
 #define CACHE_LINE_SIZE 128
+#elif defined(__VXE__) || defined(__VXE2__)
+#define CACHE_LINE_SIZE 256
 #else
 #define CACHE_LINE_SIZE 64
 #endif
@@ -244,9 +247,9 @@ typedef pthread_t ggml_thread_t;
 static const size_t CACHE_LINE_SIZE_F32 = CACHE_LINE_SIZE/sizeof(float);
 
 
-static void ggml_vec_dot_f32(int n, float * restrict s, size_t bs, const float * restrict x, size_t bx, const float * restrict y, size_t by, int nrc);
-static void ggml_vec_dot_f16(int n, float * restrict s, size_t bs, ggml_fp16_t * restrict x, size_t bx, ggml_fp16_t * restrict y, size_t by, int nrc);
-static void ggml_vec_dot_bf16(int n, float * restrict s, size_t bs, ggml_bf16_t * restrict x, size_t bx, ggml_bf16_t * restrict y, size_t by, int nrc);
+static void ggml_vec_dot_f32(int n, float * GGML_RESTRICT s, size_t bs, const float * GGML_RESTRICT x, size_t bx, const float * GGML_RESTRICT y, size_t by, int nrc);
+static void ggml_vec_dot_f16(int n, float * GGML_RESTRICT s, size_t bs, ggml_fp16_t * GGML_RESTRICT x, size_t bx, ggml_fp16_t * GGML_RESTRICT y, size_t by, int nrc);
+static void ggml_vec_dot_bf16(int n, float * GGML_RESTRICT s, size_t bs, ggml_bf16_t * GGML_RESTRICT x, size_t bx, ggml_bf16_t * GGML_RESTRICT y, size_t by, int nrc);
 
 static const struct ggml_type_traits_cpu type_traits_cpu[GGML_TYPE_COUNT] = {
     [GGML_TYPE_F32] = {
@@ -1210,6 +1213,87 @@ static inline void __lsx_f16x4_store(ggml_fp16_t * x, __m128 y) {
 #define GGML_F16_VEC_MUL             GGML_F32Cx4_MUL
 #define GGML_F16_VEC_REDUCE          GGML_F32Cx4_REDUCE
 
+#elif defined(__VXE__) || defined(__VXE2__)
+
+#define GGML_SIMD
+
+// F32 s390x
+
+#define GGML_F32_STEP 32
+#define GGML_F32_EPR  4
+
+#define GGML_F32x4              __vector float
+#define GGML_F32x4_ZERO         vec_splats(0.0f)
+#define GGML_F32x4_SET1         vec_splats
+#define GGML_F32x4_LOAD(p)      vec_xl(0, p)
+#define GGML_F32x4_STORE(p, r)  vec_xst(r, 0, p)
+#define GGML_F32x4_FMA(a, b, c) vec_madd(b, c, a)
+#define GGML_F32x4_ADD          vec_add
+#define GGML_F32x4_MUL          vec_mul
+#define GGML_F32x4_REDUCE(res, x)                   \
+{                                                   \
+    int offset = GGML_F32_ARR >> 1;                 \
+    for (int i = 0; i < offset; ++i) {              \
+        x[i] = vec_add(x[i], x[offset + i]);        \
+    }                                               \
+    offset >>= 1;                                   \
+    for (int i = 0; i < offset; ++i) {              \
+        x[i] = vec_add(x[i], x[offset + i]);        \
+    }                                               \
+    offset >>= 1;                                   \
+    for (int i = 0; i < offset; ++i) {              \
+        x[i] = vec_add(x[i], x[offset + i]);        \
+    }                                               \
+    res = vec_extract(x[0], 0) +                    \
+          vec_extract(x[0], 1) +                    \
+          vec_extract(x[0], 2) +                    \
+          vec_extract(x[0], 3);                     \
+}
+
+#define GGML_F32_VEC        GGML_F32x4
+#define GGML_F32_VEC_ZERO   GGML_F32x4_ZERO
+#define GGML_F32_VEC_SET1   GGML_F32x4_SET1
+#define GGML_F32_VEC_LOAD   GGML_F32x4_LOAD
+#define GGML_F32_VEC_STORE  GGML_F32x4_STORE
+#define GGML_F32_VEC_FMA    GGML_F32x4_FMA
+#define GGML_F32_VEC_ADD    GGML_F32x4_ADD
+#define GGML_F32_VEC_MUL    GGML_F32x4_MUL
+#define GGML_F32_VEC_REDUCE GGML_F32x4_REDUCE
+
+// F16 s390x
+#define GGML_F16_STEP GGML_F32_STEP
+#define GGML_F16_EPR  GGML_F32_EPR
+
+static inline __vector float __lzs_f16cx4_load(const ggml_fp16_t * x) {
+    float tmp[4];
+
+    for (int i = 0; i < 4; i++) {
+        tmp[i] = GGML_FP16_TO_FP32(x[i]);
+    }
+
+    return vec_xl(0, tmp);
+}
+
+static inline void __lzs_f16cx4_store(ggml_fp16_t * x, __vector float y) {
+    float arr[4];
+
+    vec_xst(y, 0, arr);
+
+    for (int i = 0; i < 4; i++) {
+        x[i] = GGML_FP32_TO_FP16(arr[i]);
+    }
+}
+
+#define GGML_F16_VEC                GGML_F32x4
+#define GGML_F16_VEC_ZERO           GGML_F32x4_ZERO
+#define GGML_F16_VEC_SET1           GGML_F32x4_SET1
+#define GGML_F16_VEC_LOAD(p, i)     __lzs_f16cx4_load(p)
+#define GGML_F16_VEC_STORE(p, r, i) __lzs_f16cx4_store(p, r[i])
+#define GGML_F16_VEC_FMA            GGML_F32x4_FMA
+#define GGML_F16_VEC_ADD            GGML_F32x4_ADD
+#define GGML_F16_VEC_MUL            GGML_F32x4_MUL
+#define GGML_F16_VEC_REDUCE         GGML_F32x4_REDUCE
+
 #endif
 
 // GGML_F32_ARR / GGML_F16_ARR
@@ -1331,17 +1415,43 @@ inline static void ggml_vec_cpy_i32(const int n, int32_t * y, const int32_t * x)
 inline static void ggml_vec_set_f16(const int n, ggml_fp16_t * x, const int32_t v) { for (int i = 0; i < n; ++i) x[i] = v; }
 inline static void ggml_vec_set_bf16(const int n, ggml_bf16_t * x, const ggml_bf16_t v) { for (int i = 0; i < n; ++i) x[i] = v; }
 inline static void ggml_vec_add_f32 (const int n, float * z, const float * x, const float * y) { for (int i = 0; i < n; ++i) z[i]  = x[i] + y[i]; }
+inline static void ggml_vec_add_f16 (const int n, ggml_fp16_t * z, const ggml_fp16_t * x, const ggml_fp16_t * y) {
+    for (int i = 0; i < n; ++i) {
+        z[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(x[i]) + GGML_FP16_TO_FP32(y[i]));
+    }
+}
 inline static void ggml_vec_add1_f32(const int n, float * z, const float * x, const float   v) { for (int i = 0; i < n; ++i) z[i]  = x[i] + v;    }
 inline static void ggml_vec_acc_f32 (const int n, float * y, const float * x)                  { for (int i = 0; i < n; ++i) y[i] += x[i];        }
 inline static void ggml_vec_acc1_f32(const int n, float * y, const float   v)                  { for (int i = 0; i < n; ++i) y[i] += v;           }
 inline static void ggml_vec_sub_f32 (const int n, float * z, const float * x, const float * y) { for (int i = 0; i < n; ++i) z[i]  = x[i] - y[i]; }
+inline static void ggml_vec_sub_f16 (const int n, ggml_fp16_t * z, const ggml_fp16_t * x, const ggml_fp16_t * y) {
+    for (int i = 0; i < n; ++i) {
+        z[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(x[i]) - GGML_FP16_TO_FP32(y[i]));
+    }
+}
 inline static void ggml_vec_set_f32 (const int n, float * x, const float   v)                  { for (int i = 0; i < n; ++i) x[i]  = v;           }
 inline static void ggml_vec_cpy_f32 (const int n, float * y, const float * x)                  { for (int i = 0; i < n; ++i) y[i]  = x[i];        }
 inline static void ggml_vec_neg_f32 (const int n, float * y, const float * x)                  { for (int i = 0; i < n; ++i) y[i]  = -x[i];       }
+inline static void ggml_vec_neg_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
+    for (int i = 0; i < n; ++i) {
+        y[i] = GGML_FP32_TO_FP16(-GGML_FP16_TO_FP32(x[i]));
+    }
+}
+
 inline static void ggml_vec_mul_f32 (const int n, float * z, const float * x, const float * y) { for (int i = 0; i < n; ++i) z[i]  = x[i]*y[i];   }
+inline static void ggml_vec_mul_f16 (const int n, ggml_fp16_t * z, const ggml_fp16_t * x, const ggml_fp16_t * y) {
+    for (int i = 0; i < n; ++i) {
+        z[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(x[i]) * GGML_FP16_TO_FP32(y[i]));
+    }
+}
 inline static void ggml_vec_div_f32 (const int n, float * z, const float * x, const float * y) { for (int i = 0; i < n; ++i) z[i]  = x[i]/y[i];   }
+inline static void ggml_vec_div_f16 (const int n, ggml_fp16_t * z, const ggml_fp16_t * x, const ggml_fp16_t * y) {
+    for (int i = 0; i < n; ++i) {
+        z[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(x[i]) / GGML_FP16_TO_FP32(y[i]));
+    }
+}
 
-static void ggml_vec_dot_f32(int n, float * restrict s, size_t bs, const float * restrict x, size_t bx, const float * restrict y, size_t by, int nrc) {
+static void ggml_vec_dot_f32(int n, float * GGML_RESTRICT s, size_t bs, const float * GGML_RESTRICT x, size_t bx, const float * GGML_RESTRICT y, size_t by, int nrc) {
    assert(nrc == 1);
    UNUSED(nrc);
    UNUSED(bx);
@@ -1384,7 +1494,7 @@ static void ggml_vec_dot_f32(int n, float * restrict s, size_t bs, const float *
     *s = sumf;
 }
 
-static void ggml_vec_dot_bf16(int n, float * restrict s, size_t bs, ggml_bf16_t * restrict x, size_t bx, ggml_bf16_t * restrict y, size_t by, int nrc) {
+static void ggml_vec_dot_bf16(int n, float * GGML_RESTRICT s, size_t bs, ggml_bf16_t * GGML_RESTRICT x, size_t bx, ggml_bf16_t * GGML_RESTRICT y, size_t by, int nrc) {
     assert(nrc == 1);
     UNUSED(nrc);
     UNUSED(bx);
@@ -1452,7 +1562,7 @@ static void ggml_vec_dot_bf16(int n, float * restrict s, size_t bs, ggml_bf16_t
     *s = sumf;
 }
 
-static void ggml_vec_dot_f16(int n, float * restrict s, size_t bs, ggml_fp16_t * restrict x, size_t bx, ggml_fp16_t * restrict y, size_t by, int nrc) {
+static void ggml_vec_dot_f16(int n, float * GGML_RESTRICT s, size_t bs, ggml_fp16_t * GGML_RESTRICT x, size_t bx, ggml_fp16_t * GGML_RESTRICT y, size_t by, int nrc) {
     assert(nrc == 1);
     UNUSED(nrc);
     UNUSED(bx);
@@ -1496,10 +1606,10 @@ static void ggml_vec_dot_f16(int n, float * restrict s, size_t bs, ggml_fp16_t *
 
 // compute GGML_VEC_DOT_UNROLL dot products at once
 // xs - x row stride in bytes
-inline static void ggml_vec_dot_f16_unroll(const int n, const int xs, float * restrict s, void * restrict xv, ggml_fp16_t * restrict y) {
+inline static void ggml_vec_dot_f16_unroll(const int n, const int xs, float * GGML_RESTRICT s, void * GGML_RESTRICT xv, ggml_fp16_t * GGML_RESTRICT y) {
     ggml_float sumf[GGML_VEC_DOT_UNROLL] = { 0.0 };
 
-    ggml_fp16_t * restrict x[GGML_VEC_DOT_UNROLL];
+    ggml_fp16_t * GGML_RESTRICT x[GGML_VEC_DOT_UNROLL];
 
     for (int i = 0; i < GGML_VEC_DOT_UNROLL; ++i) {
         x[i] = (ggml_fp16_t *) ((char *) xv + i*xs);
@@ -1549,7 +1659,7 @@ inline static void ggml_vec_dot_f16_unroll(const int n, const int xs, float * re
     }
 }
 
-inline static void ggml_vec_mad_f32(const int n, float * restrict y, const float * restrict x, const float v) {
+inline static void ggml_vec_mad_f32(const int n, float * GGML_RESTRICT y, const float * GGML_RESTRICT x, const float v) {
 #if defined(GGML_SIMD)
     const int np = (n & ~(GGML_F32_STEP - 1));
 
@@ -1580,7 +1690,7 @@ inline static void ggml_vec_mad_f32(const int n, float * restrict y, const float
 #endif
 }
 
-inline static void ggml_vec_mad_f16(const int n, ggml_fp16_t * restrict y, const ggml_fp16_t * restrict x, const float v) {
+inline static void ggml_vec_mad_f16(const int n, ggml_fp16_t * GGML_RESTRICT y, const ggml_fp16_t * GGML_RESTRICT x, const float v) {
 #if defined(GGML_SIMD)
     const int np = (n & ~(GGML_F16_STEP - 1));
 
@@ -1612,10 +1722,10 @@ inline static void ggml_vec_mad_f16(const int n, ggml_fp16_t * restrict y, const
 }
 
 // xs and vs are byte strides of x and v
-inline static void ggml_vec_mad_f32_unroll(const int n, const int xs, const int vs, float * restrict y, const float * restrict xv, const float * restrict vv) {
+inline static void ggml_vec_mad_f32_unroll(const int n, const int xs, const int vs, float * GGML_RESTRICT y, const float * GGML_RESTRICT xv, const float * GGML_RESTRICT vv) {
 
-    const float * restrict x[GGML_VEC_MAD_UNROLL];
-    const float * restrict v[GGML_VEC_MAD_UNROLL];
+    const float * GGML_RESTRICT x[GGML_VEC_MAD_UNROLL];
+    const float * GGML_RESTRICT v[GGML_VEC_MAD_UNROLL];
 
     for (int i = 0; i < GGML_VEC_MAD_UNROLL; ++i) {
         x[i] = (const float *) ((const char *) xv + i*xs);
@@ -1726,22 +1836,107 @@ inline static void ggml_vec_scale_f16(const int n, ggml_fp16_t * y, const float
 
 inline static void ggml_vec_norm_f32 (const int n, float * s, const float * x) { ggml_vec_dot_f32(n, s, 0, x, 0, x, 0, 1); *s = sqrtf(*s);   }
 inline static void ggml_vec_sqr_f32  (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = x[i]*x[i];   }
+inline static void ggml_vec_sqr_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
+    for (int i = 0; i < n; ++i) {
+        float v = GGML_FP16_TO_FP32(x[i]);
+        y[i] = GGML_FP32_TO_FP16(v*v);
+    }
+}
 inline static void ggml_vec_sqrt_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = sqrtf(x[i]); }
+inline static void ggml_vec_sqrt_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
+    for (int i = 0; i < n; ++i) {
+        y[i] = GGML_FP32_TO_FP16(sqrtf(GGML_FP16_TO_FP32(x[i])));
+    }
+}
 inline static void ggml_vec_log_f32  (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = logf(x[i]);  }
+inline static void ggml_vec_log_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
+    for (int i = 0; i < n; ++i) {
+        y[i] = GGML_FP32_TO_FP16(logf(GGML_FP16_TO_FP32(x[i])));
+    }
+}
 inline static void ggml_vec_sin_f32  (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = sinf(x[i]);  }
+inline static void ggml_vec_sin_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
+    for (int i = 0; i < n; ++i) {
+        y[i] = GGML_FP32_TO_FP16(sinf(GGML_FP16_TO_FP32(x[i])));
+    }
+}
 inline static void ggml_vec_cos_f32  (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = cosf(x[i]);  }
+inline static void ggml_vec_cos_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
+    for (int i = 0; i < n; ++i) {
+        y[i] = GGML_FP32_TO_FP16(cosf(GGML_FP16_TO_FP32(x[i])));
+    }
+}
 inline static void ggml_vec_abs_f32  (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = fabsf(x[i]); }
+inline static void ggml_vec_abs_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
+    for (int i = 0; i < n; ++i) {
+        y[i] = GGML_FP32_TO_FP16(fabsf(GGML_FP16_TO_FP32(x[i])));
+    }
+}
 inline static void ggml_vec_sgn_f32  (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? 1.f : ((x[i] < 0.f) ? -1.f : 0.f); }
+inline static void ggml_vec_sgn_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
+    for (int i = 0; i < n; ++i) {
+        float v = GGML_FP16_TO_FP32(x[i]);
+        y[i] = GGML_FP32_TO_FP16((v > 0.f) ? 1.f : ((v < 0.f) ? -1.f : 0.f));
+    }
+}
 inline static void ggml_vec_step_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? 1.f : 0.f; }
+inline static void ggml_vec_step_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
+    for (int i = 0; i < n; ++i) {
+        y[i] = GGML_FP32_TO_FP16((GGML_FP16_TO_FP32(x[i]) > 0.f) ? 1.f : 0.f);
+    }
+}
 inline static void ggml_vec_tanh_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = tanhf(x[i]);  }
+inline static void ggml_vec_tanh_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
+    for (int i = 0; i < n; ++i) {
+        y[i] = GGML_FP32_TO_FP16(tanhf(GGML_FP16_TO_FP32(x[i])));
+    }
+}
 inline static void ggml_vec_elu_f32  (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? x[i] : expm1f(x[i]); }
+inline static void ggml_vec_elu_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
+    for (int i = 0; i < n; ++i) {
+        y[i] = GGML_FP32_TO_FP16(expm1f(GGML_FP16_TO_FP32(x[i])));
+    }
+}
 inline static void ggml_vec_relu_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? x[i] : 0.f; }
+inline static void ggml_vec_relu_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
+    for (int i = 0; i < n; ++i) {
+        float v = GGML_FP16_TO_FP32(x[i]);
+        y[i] = GGML_FP32_TO_FP16((v > 0.f) ? v : 0.f);
+    }
+}
 inline static void ggml_vec_leaky_relu_f32 (const int n, float * y, const float * x, const float ns) { for (int i = 0; i < n; ++i) y[i] = ((x[i] > 0.f) ? x[i] : 0.f) + ns * ((x[i] < 0.0f) ? x[i] : 0.f); }
+inline static void ggml_vec_leaky_relu_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x, const float ns) {
+    for (int i = 0; i < n; ++i) {
+        float v = GGML_FP16_TO_FP32(x[i]);
+        y[i] = GGML_FP32_TO_FP16(((v > 0.f) ? v : 0.f) + ns * ((v < 0.0f) ? v : 0.f));
+    }
+}
 inline static void ggml_vec_sigmoid_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = 1.f / (1.f + expf(-x[i])); }
+inline static void ggml_vec_sigmoid_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
+    for (int i = 0; i < n; ++i) {
+        y[i] = GGML_FP32_TO_FP16(1.f / (1.f + expf(-GGML_FP16_TO_FP32(x[i]))));
+    }
+}
 // TODO: optimize performance
 inline static void ggml_vec_hardswish_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = x[i] * fminf(1.0f, fmaxf(0.0f, (x[i] + 3.0f) / 6.0f)); }
+inline static void ggml_vec_hardswish_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
+    for (int i = 0; i < n; ++i) {
+        float v = GGML_FP16_TO_FP32(x[i]);
+        y[i] = GGML_FP32_TO_FP16(v * fminf(1.0f, fmaxf(0.0f, (v + 3.0f) / 6.0f)));
+    }
+}
 inline static void ggml_vec_hardsigmoid_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = fminf(1.0f, fmaxf(0.0f, (x[i] + 3.0f) / 6.0f)); }
+inline static void ggml_vec_hardsigmoid_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
+    for (int i = 0; i < n; ++i) {
+        y[i] = GGML_FP32_TO_FP16(fminf(1.0f, fmaxf(0.0f, (GGML_FP16_TO_FP32(x[i]) + 3.0f) / 6.0f)));
+    }
+}
 inline static void ggml_vec_exp_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = expf(x[i]); }
+inline static void ggml_vec_exp_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
+    for (int i = 0; i < n; ++i) {
+        y[i] = GGML_FP32_TO_FP16(expf(GGML_FP16_TO_FP32(x[i])));
+    }
+}
 
 static const float GELU_COEF_A     = 0.044715f;
 static const float GELU_QUICK_COEF = -1.702f;
@@ -1809,10 +2004,21 @@ inline static void ggml_vec_gelu_quick_f32(const int n, float * y, const float *
 }
 #endif
 
+inline static void ggml_vec_gelu_quick_f16(const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
+    for (int i = 0; i < n; ++i) {
+        float v = GGML_FP16_TO_FP32(x[i]);
+        y[i] = GGML_FP32_TO_FP16(v*(1.0f/(1.0f+expf(GELU_QUICK_COEF*v))));
+    }
+}
+
 // Sigmoid Linear Unit (SiLU) function
 inline static float ggml_silu_f32(float x) {
     return x/(1.0f + expf(-x));
 }
+inline static ggml_fp16_t ggml_silu_f16(ggml_fp16_t x) {
+    float v = GGML_FP16_TO_FP32(x);
+    return GGML_FP32_TO_FP16(v/(1.0f + expf(-v)));
+}
 
 #if __FINITE_MATH_ONLY__
 #error "some routines in ggml.c require non-finite math arithmetics -- pass -fno-finite-math-only to the compiler to fix"
@@ -2036,6 +2242,12 @@ static void ggml_vec_silu_f32(const int n, float * y, const float * x) {
     }
 }
 
+inline static void ggml_vec_silu_f16(const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
+    for (int i = 0; i < n; ++i) {
+        y[i] = ggml_silu_f16(x[i]);
+    }
+}
+
 static ggml_float ggml_vec_soft_max_f32(const int n, float * y, const float * x, float max) {
     int i = 0;
     ggml_float sum = 0;
@@ -2107,12 +2319,24 @@ inline static float ggml_silu_backward_f32(float x, float dy) {
     return dy*s*(1.0f + x*(1.0f - s));
 }
 
+inline static ggml_fp16_t ggml_silu_backward_f16(ggml_fp16_t x, ggml_fp16_t dy) {
+    const float v = GGML_FP16_TO_FP32(x);
+    const float s = 1.0f/(1.0f + expf(-v));
+    return GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(dy)*s*(1.0f + v*(1.0f - s)));
+}
+
 inline static void ggml_vec_silu_backward_f32(const int n, float * dx, const float * x, const float * dy) {
     for (int i = 0; i < n; ++i) {
         dx[i] = ggml_silu_backward_f32(x[i], dy[i]);
     }
 }
 
+inline static void ggml_vec_silu_backward_f16(const int n, ggml_fp16_t * dx, const ggml_fp16_t * x, const ggml_fp16_t * dy) {
+    for (int i = 0; i < n; ++i) {
+        dx[i] = ggml_silu_backward_f16(x[i], dy[i]);
+    }
+}
+
 inline static void ggml_vec_sum_f32(const int n, float * s, const float * x) {
 #ifndef GGML_USE_ACCELERATE
     ggml_float sum = 0.0;
@@ -2381,15 +2605,20 @@ bool ggml_is_numa(void) {
 #define HWCAP2_I8MM (1 << 13)
 #endif
 
+#if !defined(HWCAP2_SME)
+#define HWCAP2_SME (1 << 23)
+#endif
+
 static void ggml_init_arm_arch_features(void) {
 #if defined(__linux__) && defined(__aarch64__)
     uint32_t hwcap = getauxval(AT_HWCAP);
     uint32_t hwcap2 = getauxval(AT_HWCAP2);
 
-    ggml_arm_arch_features.has_neon = !!(hwcap & HWCAP_ASIMD);
+    ggml_arm_arch_features.has_neon    = !!(hwcap & HWCAP_ASIMD);
     ggml_arm_arch_features.has_dotprod = !!(hwcap & HWCAP_ASIMDDP);
-    ggml_arm_arch_features.has_i8mm = !!(hwcap2 & HWCAP2_I8MM);
-    ggml_arm_arch_features.has_sve  = !!(hwcap & HWCAP_SVE);
+    ggml_arm_arch_features.has_i8mm    = !!(hwcap2 & HWCAP2_I8MM);
+    ggml_arm_arch_features.has_sve     = !!(hwcap & HWCAP_SVE);
+    ggml_arm_arch_features.has_sme     = !!(hwcap2 & HWCAP2_SME);
 
 #if defined(__ARM_FEATURE_SVE)
     ggml_arm_arch_features.sve_cnt = PR_SVE_VL_LEN_MASK & prctl(PR_SVE_GET_VL);
@@ -2412,6 +2641,11 @@ static void ggml_init_arm_arch_features(void) {
     }
     ggml_arm_arch_features.has_i8mm = oldp;
 
+    if (sysctlbyname("hw.optional.arm.FEAT_SME", &oldp, &size, NULL, 0) != 0) {
+        oldp = 0;
+    }
+    ggml_arm_arch_features.has_sme = oldp;
+
     ggml_arm_arch_features.has_sve = 0;
     ggml_arm_arch_features.sve_cnt = 0;
 #else
@@ -2435,6 +2669,12 @@ static void ggml_init_arm_arch_features(void) {
     ggml_arm_arch_features.has_sve = 0;
     ggml_arm_arch_features.sve_cnt = 0;
 #endif
+
+#if defined(__ARM_FEATURE_SME) || defined(__ARM_FEATURE_SME2)
+    ggml_arm_arch_features.has_sme = 1;
+#else
+    ggml_arm_arch_features.has_sme = 0;
+#endif
 #endif
 }
 #endif
@@ -4279,7 +4519,7 @@ static void ggml_compute_forward_add_f16_f16(
     const struct ggml_tensor * src0 = dst->src[0];
     const struct ggml_tensor * src1 = dst->src[1];
 
-    GGML_ASSERT(ggml_are_same_shape(src0, src1) && ggml_are_same_shape(src0, dst));
+    GGML_ASSERT(ggml_can_repeat(src1, src0) && ggml_are_same_shape(src0, dst));
 
     const int ith = params->ith;
     const int nth = params->nth;
@@ -4304,17 +4544,22 @@ static void ggml_compute_forward_add_f16_f16(
 
     if (nb10 == sizeof(ggml_fp16_t)) {
         for (int ir = ir0; ir < ir1; ++ir) {
-            // src0, src1 and dst are same shape => same indices
-            const int i3 = ir/(ne2*ne1);
-            const int i2 = (ir - i3*ne2*ne1)/ne1;
-            const int i1 = (ir - i3*ne2*ne1 - i2*ne1);
+            // src1 is broadcastable across src0 and dst in i1, i2, i3
+            const int64_t i03 = ir/(ne02*ne01);
+            const int64_t i02 = (ir - i03*ne02*ne01)/ne01;
+            const int64_t i01 = (ir - i03*ne02*ne01 - i02*ne01);
+
+            const int64_t i13 = i03 % ne13;
+            const int64_t i12 = i02 % ne12;
+            const int64_t i11 = i01 % ne11;
+            const int64_t nr0 = ne00 / ne10;
 
-            ggml_fp16_t * dst_ptr  = (ggml_fp16_t *) ((char *) dst->data  + i3*nb3  + i2*nb2  + i1*nb1);
-            ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01);
-            ggml_fp16_t * src1_ptr = (ggml_fp16_t *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11);
+            ggml_fp16_t * dst_ptr  = (ggml_fp16_t *) ((char *) dst->data  + i03*nb3  + i02*nb2  + i01*nb1 );
+            ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01);
+            ggml_fp16_t * src1_ptr = (ggml_fp16_t *) ((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11);
 
-            for (int i = 0; i < ne0; i++) {
-                dst_ptr[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(src0_ptr[i]) + GGML_FP16_TO_FP32(src1_ptr[i]));
+            for (int64_t r = 0; r < nr0; ++r) {
+                ggml_vec_add_f16(ne10, dst_ptr + r*ne10, src0_ptr + r*ne10, src1_ptr);
             }
         }
     }
@@ -5102,6 +5347,62 @@ static void ggml_compute_forward_sub_f32(
     }
 }
 
+static void ggml_compute_forward_sub_f16(
+    const struct ggml_compute_params * params,
+    struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
+    assert(ggml_can_repeat(src1, src0) && ggml_are_same_shape(src0, dst));
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    const int nr  = ggml_nrows(src0);
+
+    GGML_TENSOR_BINARY_OP_LOCALS
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F16);
+    GGML_ASSERT(src1->type == GGML_TYPE_F16);
+    GGML_ASSERT(dst->type  == GGML_TYPE_F16);
+
+    GGML_ASSERT( nb0 == sizeof(ggml_fp16_t));
+    GGML_ASSERT(nb00 == sizeof(ggml_fp16_t));
+
+    // rows per thread
+    const int dr = (nr + nth - 1)/nth;
+
+    // row range for this thread
+    const int ir0 = dr*ith;
+    const int ir1 = MIN(ir0 + dr, nr);
+
+    if (nb10 == sizeof(ggml_fp16_t)) {
+        for (int ir = ir0; ir < ir1; ++ir) {
+            // src1 is broadcastable across src0 and dst in i1, i2, i3
+            const int64_t i03 = ir/(ne02*ne01);
+            const int64_t i02 = (ir - i03*ne02*ne01)/ne01;
+            const int64_t i01 = (ir - i03*ne02*ne01 - i02*ne01);
+
+            const int64_t i13 = i03 % ne13;
+            const int64_t i12 = i02 % ne12;
+            const int64_t i11 = i01 % ne11;
+            const int64_t nr0 = ne00 / ne10;
+
+            ggml_fp16_t * dst_ptr  = (ggml_fp16_t *) ((char *) dst->data  + i03*nb3  + i02*nb2  + i01*nb1 );
+            ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01);
+            ggml_fp16_t * src1_ptr = (ggml_fp16_t *) ((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11);
+
+            for (int64_t r = 0; r < nr0; ++r) {
+                ggml_vec_sub_f16(ne10, dst_ptr + r*ne10, src0_ptr + r*ne10, src1_ptr);
+            }
+        }
+    } else {
+        // src1 is not contiguous
+        GGML_ABORT("unimplemented error");
+    }
+}
+
 static void ggml_compute_forward_sub(
         const struct ggml_compute_params * params,
         struct ggml_tensor * dst) {
@@ -5113,6 +5414,10 @@ static void ggml_compute_forward_sub(
             {
                 ggml_compute_forward_sub_f32(params, dst);
             } break;
+        case GGML_TYPE_F16:
+            {
+                ggml_compute_forward_sub_f16(params, dst);
+            } break;
         default:
             {
                 GGML_ABORT("fatal error");
@@ -5193,6 +5498,55 @@ static void ggml_compute_forward_mul_f32(
     }
 }
 
+static void ggml_compute_forward_mul_f16(
+    const struct ggml_compute_params * params,
+    struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
+    GGML_ASSERT(ggml_can_repeat(src1, src0) && ggml_are_same_shape(src0, dst));
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    const int64_t nr = ggml_nrows(src0);
+
+    GGML_TENSOR_BINARY_OP_LOCALS
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F16);
+    GGML_ASSERT(src1->type == GGML_TYPE_F16);
+    GGML_ASSERT(dst->type  == GGML_TYPE_F16);
+
+    GGML_ASSERT( nb0 == sizeof(ggml_fp16_t));
+    GGML_ASSERT(nb00 == sizeof(ggml_fp16_t));
+
+    if (nb10 == sizeof(ggml_fp16_t)) {
+        for (int64_t ir = ith; ir < nr; ir += nth) {
+            // src0 and dst are same shape => same indices
+            const int64_t i03 = ir/(ne02*ne01);
+            const int64_t i02 = (ir - i03*ne02*ne01)/ne01;
+            const int64_t i01 = (ir - i03*ne02*ne01 - i02*ne01);
+
+            const int64_t i13 = i03 % ne13;
+            const int64_t i12 = i02 % ne12;
+            const int64_t i11 = i01 % ne11;
+            const int64_t nr0 = ne00 / ne10;
+
+            ggml_fp16_t * dst_ptr  = (ggml_fp16_t *) ((char *) dst->data  + i03*nb3  + i02*nb2  + i01*nb1 );
+            ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01);
+            ggml_fp16_t * src1_ptr = (ggml_fp16_t *) ((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11);
+
+            for (int64_t r = 0 ; r < nr0; ++r) {
+                ggml_vec_mul_f16(ne10, dst_ptr + r*ne10, src0_ptr + r*ne10, src1_ptr);
+            }
+        }
+    } else {
+        // src1 is not contiguous
+        GGML_ABORT("unimplemented error");
+    }
+}
+
 static void ggml_compute_forward_mul(
         const struct ggml_compute_params * params,
         struct ggml_tensor * dst) {
@@ -5200,13 +5554,17 @@ static void ggml_compute_forward_mul(
     const struct ggml_tensor * src0 = dst->src[0];
     const struct ggml_tensor * src1 = dst->src[1];
 
-    GGML_ASSERT(src1->type == GGML_TYPE_F32 && "only f32 src1 supported for now");
+    GGML_ASSERT((src1->type == GGML_TYPE_F32 || src1->type == GGML_TYPE_F16) && "only f32/f16 src1 supported for now");
 
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
                 ggml_compute_forward_mul_f32(params, dst);
             } break;
+        case GGML_TYPE_F16:
+            {
+                ggml_compute_forward_mul_f16(params, dst);
+            } break;
         default:
             {
                 GGML_ABORT("fatal error");
@@ -5287,18 +5645,71 @@ static void ggml_compute_forward_div_f32(
     }
 }
 
-static void ggml_compute_forward_div(
-        const struct ggml_compute_params * params,
-        struct ggml_tensor * dst) {
+static void ggml_compute_forward_div_f16(
+    const struct ggml_compute_params * params,
+    struct ggml_tensor * dst) {
 
     const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
 
-    switch (src0->type) {
-        case GGML_TYPE_F32:
-            {
-                ggml_compute_forward_div_f32(params, dst);
-            } break;
-        default:
+    GGML_ASSERT(ggml_can_repeat(src1, src0) && ggml_are_same_shape(src0, dst));
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    const int64_t nr = ggml_nrows(src0);
+
+    GGML_TENSOR_BINARY_OP_LOCALS
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F16);
+    GGML_ASSERT(src1->type == GGML_TYPE_F16);
+    GGML_ASSERT(dst->type  == GGML_TYPE_F16);
+
+    GGML_ASSERT( nb0 == sizeof(ggml_fp16_t));
+    GGML_ASSERT(nb00 == sizeof(ggml_fp16_t));
+
+    if (nb10 == sizeof(ggml_fp16_t)) {
+        for (int64_t ir = ith; ir < nr; ir += nth) {
+            // src0 and dst are same shape => same indices
+            const int64_t i03 = ir/(ne02*ne01);
+            const int64_t i02 = (ir - i03*ne02*ne01)/ne01;
+            const int64_t i01 = (ir - i03*ne02*ne01 - i02*ne01);
+
+            const int64_t i13 = i03 % ne13;
+            const int64_t i12 = i02 % ne12;
+            const int64_t i11 = i01 % ne11;
+            const int64_t nr0 = ne00 / ne10;
+
+            ggml_fp16_t * dst_ptr  = (ggml_fp16_t *) ((char *) dst->data  + i03*nb3  + i02*nb2  + i01*nb1 );
+            ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01);
+            ggml_fp16_t * src1_ptr = (ggml_fp16_t *) ((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11);
+
+            for (int64_t r = 0; r < nr0; ++r) {
+                ggml_vec_div_f16(ne10, dst_ptr + r*ne10, src0_ptr + r*ne10, src1_ptr);
+            }
+        }
+    } else {
+        // src1 is not contiguous
+        GGML_ABORT("unimplemented error");
+    }
+}
+
+static void ggml_compute_forward_div(
+        const struct ggml_compute_params * params,
+        struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    switch (src0->type) {
+        case GGML_TYPE_F32:
+            {
+                ggml_compute_forward_div_f32(params, dst);
+            } break;
+        case GGML_TYPE_F16:
+            {
+                ggml_compute_forward_div_f16(params, dst);
+            } break;
+        default:
             {
                 GGML_ABORT("fatal error");
             }
@@ -5332,6 +5743,31 @@ static void ggml_compute_forward_sqr_f32(
     }
 }
 
+static void ggml_compute_forward_sqr_f16(
+    const struct ggml_compute_params * params,
+    struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    if (params->ith != 0) {
+        return;
+    }
+
+    assert(ggml_are_same_shape(src0, dst));
+
+    const int n     = ggml_nrows(src0);
+    const int nc    = src0->ne[0];
+
+    assert( dst->nb[0] == sizeof(ggml_fp16_t));
+    assert(src0->nb[0] == sizeof(ggml_fp16_t));
+
+    for (int i = 0; i < n; i++) {
+        ggml_vec_sqr_f16(nc,
+                (ggml_fp16_t *) ((char *) dst->data  + i*( dst->nb[1])),
+                (ggml_fp16_t *) ((char *) src0->data + i*(src0->nb[1])));
+    }
+}
+
 static void ggml_compute_forward_sqr(
         const struct ggml_compute_params * params,
         struct ggml_tensor * dst) {
@@ -5343,6 +5779,10 @@ static void ggml_compute_forward_sqr(
             {
                 ggml_compute_forward_sqr_f32(params, dst);
             } break;
+        case GGML_TYPE_F16:
+            {
+                ggml_compute_forward_sqr_f16(params, dst);
+            } break;
         default:
             {
                 GGML_ABORT("fatal error");
@@ -5377,6 +5817,31 @@ static void ggml_compute_forward_sqrt_f32(
     }
 }
 
+static void ggml_compute_forward_sqrt_f16(
+    const struct ggml_compute_params * params,
+    struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    if (params->ith != 0) {
+        return;
+    }
+
+    assert(ggml_are_same_shape(src0, dst));
+
+    const int n  = ggml_nrows(src0);
+    const int nc = src0->ne[0];
+
+    assert( dst->nb[0] == sizeof(ggml_fp16_t));
+    assert(src0->nb[0] == sizeof(ggml_fp16_t));
+
+    for (int i = 0; i < n; i++) {
+        ggml_vec_sqrt_f16(nc,
+                (ggml_fp16_t *) ((char *) dst->data  + i*( dst->nb[1])),
+                (ggml_fp16_t *) ((char *) src0->data + i*(src0->nb[1])));
+    }
+}
+
 static void ggml_compute_forward_sqrt(
         const struct ggml_compute_params * params,
         struct ggml_tensor * dst) {
@@ -5388,6 +5853,10 @@ static void ggml_compute_forward_sqrt(
             {
                 ggml_compute_forward_sqrt_f32(params, dst);
             } break;
+        case GGML_TYPE_F16:
+            {
+                ggml_compute_forward_sqrt_f16(params, dst);
+            } break;
         default:
             {
                 GGML_ABORT("fatal error");
@@ -5422,6 +5891,31 @@ static void ggml_compute_forward_log_f32(
     }
 }
 
+static void ggml_compute_forward_log_f16(
+    const struct ggml_compute_params * params,
+    struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    if (params->ith != 0) {
+        return;
+    }
+
+    GGML_ASSERT(ggml_are_same_shape(src0, dst));
+
+    const int n  = ggml_nrows(src0);
+    const int nc = src0->ne[0];
+
+    GGML_ASSERT( dst->nb[0] == sizeof(ggml_fp16_t));
+    GGML_ASSERT(src0->nb[0] == sizeof(ggml_fp16_t));
+
+    for (int i = 0; i < n; i++) {
+        ggml_vec_log_f16(nc,
+                (ggml_fp16_t *) ((char *) dst->data  + i*( dst->nb[1])),
+                (ggml_fp16_t *) ((char *) src0->data + i*(src0->nb[1])));
+    }
+}
+
 static void ggml_compute_forward_log(
         const struct ggml_compute_params * params,
         struct ggml_tensor * dst) {
@@ -5433,6 +5927,10 @@ static void ggml_compute_forward_log(
             {
                 ggml_compute_forward_log_f32(params, dst);
             } break;
+        case GGML_TYPE_F16:
+            {
+                ggml_compute_forward_log_f16(params, dst);
+            } break;
         default:
             {
                 GGML_ABORT("fatal error");
@@ -5467,6 +5965,31 @@ static void ggml_compute_forward_sin_f32(
     }
 }
 
+static void ggml_compute_forward_sin_f16(
+    const struct ggml_compute_params * params,
+    struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    if (params->ith != 0) {
+        return;
+    }
+
+    GGML_ASSERT(ggml_are_same_shape(src0, dst));
+
+    const int n  = ggml_nrows(src0);
+    const int nc = src0->ne[0];
+
+    GGML_ASSERT( dst->nb[0] == sizeof(ggml_fp16_t));
+    GGML_ASSERT(src0->nb[0] == sizeof(ggml_fp16_t));
+
+    for (int i = 0; i < n; i++) {
+        ggml_vec_sin_f16(nc,
+                (ggml_fp16_t *) ((char *) dst->data  + i*( dst->nb[1])),
+                (ggml_fp16_t *) ((char *) src0->data + i*(src0->nb[1])));
+    }
+}
+
 static void ggml_compute_forward_sin(
         const struct ggml_compute_params * params,
         struct ggml_tensor * dst) {
@@ -5478,6 +6001,10 @@ static void ggml_compute_forward_sin(
             {
                 ggml_compute_forward_sin_f32(params, dst);
             } break;
+        case GGML_TYPE_F16:
+            {
+                ggml_compute_forward_sin_f16(params, dst);
+            } break;
         default:
             {
                 GGML_ABORT("fatal error");
@@ -5512,6 +6039,31 @@ static void ggml_compute_forward_cos_f32(
     }
 }
 
+static void ggml_compute_forward_cos_f16(
+    const struct ggml_compute_params * params,
+    struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    if (params->ith != 0) {
+        return;
+    }
+
+    GGML_ASSERT(ggml_are_same_shape(src0, dst));
+
+    const int n  = ggml_nrows(src0);
+    const int nc = src0->ne[0];
+
+    GGML_ASSERT( dst->nb[0] == sizeof(ggml_fp16_t));
+    GGML_ASSERT(src0->nb[0] == sizeof(ggml_fp16_t));
+
+    for (int i = 0; i < n; i++) {
+        ggml_vec_cos_f16(nc,
+                (ggml_fp16_t *) ((char *) dst->data  + i*( dst->nb[1])),
+                (ggml_fp16_t *) ((char *) src0->data + i*(src0->nb[1])));
+    }
+}
+
 static void ggml_compute_forward_cos(
         const struct ggml_compute_params * params,
         struct ggml_tensor * dst) {
@@ -5523,6 +6075,10 @@ static void ggml_compute_forward_cos(
             {
                 ggml_compute_forward_cos_f32(params, dst);
             } break;
+        case GGML_TYPE_F16:
+            {
+                ggml_compute_forward_cos_f16(params, dst);
+            } break;
         default:
             {
                 GGML_ABORT("fatal error");
@@ -6092,14 +6648,14 @@ static void ggml_compute_forward_repeat_back(
 
 // ggml_compute_forward_concat
 
-static void ggml_compute_forward_concat_f32(
+static void ggml_compute_forward_concat_any(
     const struct ggml_compute_params * params,
     struct ggml_tensor * dst) {
 
     const struct ggml_tensor * src0 = dst->src[0];
     const struct ggml_tensor * src1 = dst->src[1];
 
-    GGML_ASSERT(src0->nb[0] == sizeof(float));
+    const size_t len = ggml_type_size(src0->type);
 
     const int ith = params->ith;
     const int nth = params->nth;
@@ -6113,7 +6669,7 @@ static void ggml_compute_forward_concat_f32(
     int64_t o[4] = {0, 0, 0, 0};
     o[dim] = src0->ne[dim];
 
-    const float * x;
+    const char * x;
 
     // TODO: smarter multi-theading
     for (int i3 = 0; i3 < ne3; i3++) {
@@ -6121,101 +6677,268 @@ static void ggml_compute_forward_concat_f32(
             for (int i1 = 0; i1 < ne1; i1++) {
                 for (int i0 = 0; i0 < ne0; i0++) {
                     if (i0 < ne00 && i1 < ne01 && i2 < ne02 && i3 < ne03) {
-                        x = (const float *) ((const char *)src0->data + (i0       )*nb00 + (i1       )*nb01 + (i2       )*nb02 + (i3       )*nb03);
+                        x = (const char *)src0->data + (i0       )*nb00 + (i1       )*nb01 + (i2       )*nb02 + (i3       )*nb03;
                     } else {
-                        x = (const float *) ((const char *)src1->data + (i0 - o[0])*nb10 + (i1 - o[1])*nb11 + (i2 - o[2])*nb12 + (i3 - o[3])*nb13);
+                        x = (const char *)src1->data + (i0 - o[0])*nb10 + (i1 - o[1])*nb11 + (i2 - o[2])*nb12 + (i3 - o[3])*nb13;
                     }
 
-                    float * y = (float *)((char *)dst->data + i0*nb0 + i1*nb1 + i2*nb2 + i3*nb3);
+                    char * y = (char *)dst->data + i0*nb0 + i1*nb1 + i2*nb2 + i3*nb3;
 
-                    *y = *x;
+                    memcpy(y, x, len);
                 }
             }
         }
     }
 }
 
-static void ggml_compute_forward_concat(
+static void ggml_compute_forward_concat_i8(
     const struct ggml_compute_params * params,
     struct ggml_tensor * dst) {
 
     const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
 
-    switch (src0->type) {
-        case GGML_TYPE_F32:
-        case GGML_TYPE_I32:
-            {
-                ggml_compute_forward_concat_f32(params, dst);
-            } break;
-        default:
-            {
-                GGML_ABORT("fatal error");
-            }
-    }
-}
-
-// ggml_compute_forward_abs
+    GGML_ASSERT(ggml_type_size(src0->type) == sizeof(int8_t));
 
-static void ggml_compute_forward_abs_f32(
-        const struct ggml_compute_params * params,
-        struct ggml_tensor * dst) {
+    const int ith = params->ith;
+    const int nth = params->nth;
 
-    const struct ggml_tensor * src0 = dst->src[0];
+    GGML_TENSOR_BINARY_OP_LOCALS
 
-    if (params->ith != 0) {
-        return;
-    }
+    const int32_t dim = ggml_get_op_params_i32(dst, 0);
 
-    assert(ggml_is_contiguous_1(src0));
-    assert(ggml_is_contiguous_1(dst));
-    assert(ggml_are_same_shape(src0, dst));
+    GGML_ASSERT(dim >= 0 && dim < 4);
 
-    const int n  = ggml_nrows(src0);
-    const int nc = src0->ne[0];
+    int64_t o[4] = {0, 0, 0, 0};
+    o[dim] = src0->ne[dim];
 
-    for (int i = 0; i < n; i++) {
-        ggml_vec_abs_f32(nc,
-                (float *) ((char *) dst->data  + i*( dst->nb[1])),
-                (float *) ((char *) src0->data + i*(src0->nb[1])));
-    }
-}
+    const int8_t * x;
 
-static void ggml_compute_forward_abs(
-        const struct ggml_compute_params * params,
-        struct ggml_tensor * dst) {
+    // TODO: smarter multi-theading
+    for (int i3 = 0; i3 < ne3; i3++) {
+        for (int i2 = ith; i2 < ne2; i2 += nth) {
+            for (int i1 = 0; i1 < ne1; i1++) {
+                for (int i0 = 0; i0 < ne0; i0++) {
+                    if (i0 < ne00 && i1 < ne01 && i2 < ne02 && i3 < ne03) {
+                        x = (const int8_t *) ((const char *)src0->data + (i0       )*nb00 + (i1       )*nb01 + (i2       )*nb02 + (i3       )*nb03);
+                    } else {
+                        x = (const int8_t *) ((const char *)src1->data + (i0 - o[0])*nb10 + (i1 - o[1])*nb11 + (i2 - o[2])*nb12 + (i3 - o[3])*nb13);
+                    }
 
-    const struct ggml_tensor * src0 = dst->src[0];
+                    int8_t * y = (int8_t *)((char *)dst->data + i0*nb0 + i1*nb1 + i2*nb2 + i3*nb3);
 
-    switch (src0->type) {
-        case GGML_TYPE_F32:
-            {
-                ggml_compute_forward_abs_f32(params, dst);
-            } break;
-        default:
-            {
-                GGML_ABORT("fatal error");
+                    *y = *x;
+                }
             }
+        }
     }
 }
 
-// ggml_compute_forward_sgn
-
-static void ggml_compute_forward_sgn_f32(
-        const struct ggml_compute_params * params,
-        struct ggml_tensor * dst) {
+static void ggml_compute_forward_concat_f16(
+    const struct ggml_compute_params * params,
+    struct ggml_tensor * dst) {
 
     const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
 
-    if (params->ith != 0) {
-        return;
-    }
+    GGML_ASSERT(ggml_type_size(src0->type) == sizeof(ggml_fp16_t));
 
-    assert(ggml_is_contiguous_1(src0));
-    assert(ggml_is_contiguous_1(dst));
-    assert(ggml_are_same_shape(src0, dst));
+    const int ith = params->ith;
+    const int nth = params->nth;
 
-    const int n  = ggml_nrows(src0);
-    const int nc = src0->ne[0];
+    GGML_TENSOR_BINARY_OP_LOCALS
+
+    const int32_t dim = ggml_get_op_params_i32(dst, 0);
+
+    GGML_ASSERT(dim >= 0 && dim < 4);
+
+    int64_t o[4] = {0, 0, 0, 0};
+    o[dim] = src0->ne[dim];
+
+    const ggml_fp16_t * x;
+
+    // TODO: smarter multi-theading
+    for (int i3 = 0; i3 < ne3; i3++) {
+        for (int i2 = ith; i2 < ne2; i2 += nth) {
+            for (int i1 = 0; i1 < ne1; i1++) {
+                for (int i0 = 0; i0 < ne0; i0++) {
+                    if (i0 < ne00 && i1 < ne01 && i2 < ne02 && i3 < ne03) {
+                        x = (const ggml_fp16_t *) ((const char *)src0->data + (i0       )*nb00 + (i1       )*nb01 + (i2       )*nb02 + (i3       )*nb03);
+                    } else {
+                        x = (const ggml_fp16_t *) ((const char *)src1->data + (i0 - o[0])*nb10 + (i1 - o[1])*nb11 + (i2 - o[2])*nb12 + (i3 - o[3])*nb13);
+                    }
+
+                    ggml_fp16_t * y = (ggml_fp16_t *)((char *)dst->data + i0*nb0 + i1*nb1 + i2*nb2 + i3*nb3);
+
+                    *y = *x;
+                }
+            }
+        }
+    }
+}
+
+static void ggml_compute_forward_concat_f32(
+    const struct ggml_compute_params * params,
+    struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
+    GGML_ASSERT(ggml_type_size(src0->type) == sizeof(float));
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    GGML_TENSOR_BINARY_OP_LOCALS
+
+    const int32_t dim = ggml_get_op_params_i32(dst, 0);
+
+    GGML_ASSERT(dim >= 0 && dim < 4);
+
+    int64_t o[4] = {0, 0, 0, 0};
+    o[dim] = src0->ne[dim];
+
+    const float * x;
+
+    // TODO: smarter multi-theading
+    for (int i3 = 0; i3 < ne3; i3++) {
+        for (int i2 = ith; i2 < ne2; i2 += nth) {
+            for (int i1 = 0; i1 < ne1; i1++) {
+                for (int i0 = 0; i0 < ne0; i0++) {
+                    if (i0 < ne00 && i1 < ne01 && i2 < ne02 && i3 < ne03) {
+                        x = (const float *) ((const char *)src0->data + (i0       )*nb00 + (i1       )*nb01 + (i2       )*nb02 + (i3       )*nb03);
+                    } else {
+                        x = (const float *) ((const char *)src1->data + (i0 - o[0])*nb10 + (i1 - o[1])*nb11 + (i2 - o[2])*nb12 + (i3 - o[3])*nb13);
+                    }
+
+                    float * y = (float *)((char *)dst->data + i0*nb0 + i1*nb1 + i2*nb2 + i3*nb3);
+
+                    *y = *x;
+                }
+            }
+        }
+    }
+}
+
+static void ggml_compute_forward_concat(
+    const struct ggml_compute_params * params,
+    struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    switch (src0->type) {
+        case GGML_TYPE_F16:
+        case GGML_TYPE_BF16:
+        case GGML_TYPE_I16:
+            {
+                ggml_compute_forward_concat_f16(params, dst);
+            } break;
+        case GGML_TYPE_I8:
+            {
+                ggml_compute_forward_concat_i8(params, dst);
+            } break;
+        case GGML_TYPE_F32:
+        case GGML_TYPE_I32:
+            {
+                ggml_compute_forward_concat_f32(params, dst);
+            } break;
+        default:
+            {
+                ggml_compute_forward_concat_any(params, dst);
+            }
+    }
+}
+
+// ggml_compute_forward_abs
+
+static void ggml_compute_forward_abs_f32(
+        const struct ggml_compute_params * params,
+        struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    if (params->ith != 0) {
+        return;
+    }
+
+    assert(ggml_is_contiguous_1(src0));
+    assert(ggml_is_contiguous_1(dst));
+    assert(ggml_are_same_shape(src0, dst));
+
+    const int n  = ggml_nrows(src0);
+    const int nc = src0->ne[0];
+
+    for (int i = 0; i < n; i++) {
+        ggml_vec_abs_f32(nc,
+                (float *) ((char *) dst->data  + i*( dst->nb[1])),
+                (float *) ((char *) src0->data + i*(src0->nb[1])));
+    }
+}
+
+static void ggml_compute_forward_abs_f16(
+    const struct ggml_compute_params * params,
+    struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    if (params->ith != 0) {
+        return;
+    }
+
+    assert(ggml_is_contiguous_1(src0));
+    assert(ggml_is_contiguous_1(dst));
+    assert(ggml_are_same_shape(src0, dst));
+
+    const int n  = ggml_nrows(src0);
+    const int nc = src0->ne[0];
+
+    for (int i = 0; i < n; i++) {
+        ggml_vec_abs_f16(nc,
+                (ggml_fp16_t *) ((char *) dst->data  + i*( dst->nb[1])),
+                (ggml_fp16_t *) ((char *) src0->data + i*(src0->nb[1])));
+    }
+}
+
+static void ggml_compute_forward_abs(
+        const struct ggml_compute_params * params,
+        struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    switch (src0->type) {
+        case GGML_TYPE_F32:
+            {
+                ggml_compute_forward_abs_f32(params, dst);
+            } break;
+        case GGML_TYPE_F16:
+            {
+                ggml_compute_forward_abs_f16(params, dst);
+            } break;
+        default:
+            {
+                GGML_ABORT("fatal error");
+            }
+    }
+}
+
+// ggml_compute_forward_sgn
+
+static void ggml_compute_forward_sgn_f32(
+        const struct ggml_compute_params * params,
+        struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    if (params->ith != 0) {
+        return;
+    }
+
+    assert(ggml_is_contiguous_1(src0));
+    assert(ggml_is_contiguous_1(dst));
+    assert(ggml_are_same_shape(src0, dst));
+
+    const int n  = ggml_nrows(src0);
+    const int nc = src0->ne[0];
 
     for (int i = 0; i < n; i++) {
         ggml_vec_sgn_f32(nc,
@@ -6224,6 +6947,30 @@ static void ggml_compute_forward_sgn_f32(
     }
 }
 
+static void ggml_compute_forward_sgn_f16(
+    const struct ggml_compute_params * params,
+    struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    if (params->ith != 0) {
+        return;
+    }
+
+    assert(ggml_is_contiguous_1(src0));
+    assert(ggml_is_contiguous_1(dst));
+    assert(ggml_are_same_shape(src0, dst));
+
+    const int n  = ggml_nrows(src0);
+    const int nc = src0->ne[0];
+
+    for (int i = 0; i < n; i++) {
+        ggml_vec_sgn_f16(nc,
+                (ggml_fp16_t *) ((char *) dst->data  + i*( dst->nb[1])),
+                (ggml_fp16_t *) ((char *) src0->data + i*(src0->nb[1])));
+    }
+}
+
 static void ggml_compute_forward_sgn(
         const struct ggml_compute_params * params,
         struct ggml_tensor * dst) {
@@ -6235,6 +6982,10 @@ static void ggml_compute_forward_sgn(
             {
                 ggml_compute_forward_sgn_f32(params, dst);
             } break;
+        case GGML_TYPE_F16:
+            {
+                ggml_compute_forward_sgn_f16(params, dst);
+            } break;
         default:
             {
                 GGML_ABORT("fatal error");
@@ -6268,6 +7019,30 @@ static void ggml_compute_forward_neg_f32(
     }
 }
 
+static void ggml_compute_forward_neg_f16(
+    const struct ggml_compute_params * params,
+    struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    if (params->ith != 0) {
+        return;
+    }
+
+    assert(ggml_is_contiguous_1(src0));
+    assert(ggml_is_contiguous_1(dst));
+    assert(ggml_are_same_shape(src0, dst));
+
+    const int n  = ggml_nrows(src0);
+    const int nc = src0->ne[0];
+
+    for (int i = 0; i < n; i++) {
+        ggml_vec_neg_f16(nc,
+                (ggml_fp16_t *) ((char *) dst->data  + i*( dst->nb[1])),
+                (ggml_fp16_t *) ((char *) src0->data + i*(src0->nb[1])));
+    }
+}
+
 static void ggml_compute_forward_neg(
         const struct ggml_compute_params * params,
         struct ggml_tensor * dst) {
@@ -6279,6 +7054,10 @@ static void ggml_compute_forward_neg(
             {
                 ggml_compute_forward_neg_f32(params, dst);
             } break;
+        case GGML_TYPE_F16:
+            {
+                ggml_compute_forward_neg_f16(params, dst);
+            } break;
         default:
             {
                 GGML_ABORT("fatal error");
@@ -6312,6 +7091,30 @@ static void ggml_compute_forward_step_f32(
     }
 }
 
+static void ggml_compute_forward_step_f16(
+    const struct ggml_compute_params * params,
+    struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    if (params->ith != 0) {
+        return;
+    }
+
+    assert(ggml_is_contiguous_1(src0));
+    assert(ggml_is_contiguous_1(dst));
+    assert(ggml_are_same_shape(src0, dst));
+
+    const int n  = ggml_nrows(src0);
+    const int nc = src0->ne[0];
+
+    for (int i = 0; i < n; i++) {
+        ggml_vec_step_f16(nc,
+                (ggml_fp16_t *) ((char *) dst->data  + i*( dst->nb[1])),
+                (ggml_fp16_t *) ((char *) src0->data + i*(src0->nb[1])));
+    }
+}
+
 static void ggml_compute_forward_step(
         const struct ggml_compute_params * params,
         struct ggml_tensor * dst) {
@@ -6323,6 +7126,10 @@ static void ggml_compute_forward_step(
             {
                 ggml_compute_forward_step_f32(params, dst);
             } break;
+        case GGML_TYPE_F16:
+            {
+                ggml_compute_forward_step_f16(params, dst);
+            } break;
         default:
             {
                 GGML_ABORT("fatal error");
@@ -6356,6 +7163,30 @@ static void ggml_compute_forward_tanh_f32(
     }
 }
 
+static void ggml_compute_forward_tanh_f16(
+    const struct ggml_compute_params * params,
+    struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    if (params->ith != 0) {
+        return;
+    }
+
+    assert(ggml_is_contiguous_1(src0));
+    assert(ggml_is_contiguous_1(dst));
+    assert(ggml_are_same_shape(src0, dst));
+
+    const int n  = ggml_nrows(src0);
+    const int nc = src0->ne[0];
+
+    for (int i = 0; i < n; i++) {
+        ggml_vec_tanh_f16(nc,
+                (ggml_fp16_t *) ((char *) dst->data  + i*( dst->nb[1])),
+                (ggml_fp16_t *) ((char *) src0->data + i*(src0->nb[1])));
+    }
+}
+
 static void ggml_compute_forward_tanh(
         const struct ggml_compute_params * params,
         struct ggml_tensor * dst) {
@@ -6367,6 +7198,10 @@ static void ggml_compute_forward_tanh(
             {
                 ggml_compute_forward_tanh_f32(params, dst);
             } break;
+        case GGML_TYPE_F16:
+            {
+                ggml_compute_forward_tanh_f16(params, dst);
+            } break;
         default:
             {
                 GGML_ABORT("fatal error");
@@ -6400,6 +7235,30 @@ static void ggml_compute_forward_elu_f32(
     }
 }
 
+static void ggml_compute_forward_elu_f16(
+    const struct ggml_compute_params * params,
+    struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    if (params->ith != 0) {
+        return;
+    }
+
+    assert(ggml_is_contiguous_1(src0));
+    assert(ggml_is_contiguous_1(dst));
+    assert(ggml_are_same_shape(src0, dst));
+
+    const int n  = ggml_nrows(src0);
+    const int nc = src0->ne[0];
+
+    for (int i = 0; i < n; i++) {
+        ggml_vec_elu_f16(nc,
+                (ggml_fp16_t *) ((char *) dst->data  + i*( dst->nb[1])),
+                (ggml_fp16_t *) ((char *) src0->data + i*(src0->nb[1])));
+    }
+}
+
 static void ggml_compute_forward_elu(
         const struct ggml_compute_params * params,
         struct ggml_tensor * dst) {
@@ -6411,6 +7270,10 @@ static void ggml_compute_forward_elu(
             {
                 ggml_compute_forward_elu_f32(params, dst);
             } break;
+        case GGML_TYPE_F16:
+            {
+                ggml_compute_forward_elu_f16(params, dst);
+            } break;
         default:
             {
                 GGML_ABORT("fatal error");
@@ -6444,6 +7307,30 @@ static void ggml_compute_forward_relu_f32(
     }
 }
 
+static void ggml_compute_forward_relu_f16(
+    const struct ggml_compute_params * params,
+    struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    if (params->ith != 0) {
+        return;
+    }
+
+    assert(ggml_is_contiguous_1(src0));
+    assert(ggml_is_contiguous_1(dst));
+    assert(ggml_are_same_shape(src0, dst));
+
+    const int n  = ggml_nrows(src0);
+    const int nc = src0->ne[0];
+
+    for (int i = 0; i < n; i++) {
+        ggml_vec_relu_f16(nc,
+                (ggml_fp16_t *) ((char *) dst->data  + i*( dst->nb[1])),
+                (ggml_fp16_t *) ((char *) src0->data + i*(src0->nb[1])));
+    }
+}
+
 static void ggml_compute_forward_relu(
         const struct ggml_compute_params * params,
         struct ggml_tensor * dst) {
@@ -6455,6 +7342,10 @@ static void ggml_compute_forward_relu(
             {
                 ggml_compute_forward_relu_f32(params, dst);
             } break;
+        case GGML_TYPE_F16:
+            {
+                ggml_compute_forward_relu_f16(params, dst);
+            } break;
         default:
             {
                 GGML_ABORT("fatal error");
@@ -6488,6 +7379,30 @@ static void ggml_compute_forward_sigmoid_f32(
     }
 }
 
+static void ggml_compute_forward_sigmoid_f16(
+    const struct ggml_compute_params * params,
+    struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    if (params->ith != 0) {
+        return;
+    }
+
+    assert(ggml_is_contiguous_1(src0));
+    assert(ggml_is_contiguous_1(dst));
+    assert(ggml_are_same_shape(src0, dst));
+
+    const int n  = ggml_nrows(src0);
+    const int nc = src0->ne[0];
+
+    for (int i = 0; i < n; i++) {
+        ggml_vec_sigmoid_f16(nc,
+                (ggml_fp16_t *) ((char *) dst->data  + i*( dst->nb[1])),
+                (ggml_fp16_t *) ((char *) src0->data + i*(src0->nb[1])));
+    }
+}
+
 static void ggml_compute_forward_sigmoid(
         const struct ggml_compute_params * params,
         struct ggml_tensor * dst) {
@@ -6499,6 +7414,113 @@ static void ggml_compute_forward_sigmoid(
             {
                 ggml_compute_forward_sigmoid_f32(params, dst);
             } break;
+        case GGML_TYPE_F16:
+            {
+                ggml_compute_forward_sigmoid_f16(params, dst);
+            } break;
+        default:
+            {
+                GGML_ABORT("fatal error");
+            }
+    }
+}
+
+// ggml_compute_forward_gelu
+
+static void ggml_compute_forward_gelu_f32(
+        const struct ggml_compute_params * params,
+        struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    assert(ggml_is_contiguous_1(src0));
+    assert(ggml_is_contiguous_1(dst));
+    assert(ggml_are_same_shape(src0, dst));
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    const int nc = src0->ne[0];
+    const int nr = ggml_nrows(src0);
+
+    // rows per thread
+    const int dr = (nr + nth - 1)/nth;
+
+    // row range for this thread
+    const int ir0 = dr*ith;
+    const int ir1 = MIN(ir0 + dr, nr);
+
+    for (int i1 = ir0; i1 < ir1; i1++) {
+        ggml_vec_gelu_f32(nc,
+                (float *) ((char *) dst->data  + i1*( dst->nb[1])),
+                (float *) ((char *) src0->data + i1*(src0->nb[1])));
+
+#ifndef NDEBUG
+        for (int k = 0; k < nc; k++) {
+            const float x = ((float *) ((char *) dst->data + i1*( dst->nb[1])))[k];
+            UNUSED(x);
+            assert(!isnan(x));
+            assert(!isinf(x));
+        }
+#endif
+    }
+}
+
+static void ggml_compute_forward_gelu_f16(
+    const struct ggml_compute_params * params,
+    struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    assert(ggml_is_contiguous_1(src0));
+    assert(ggml_is_contiguous_1(dst));
+    assert(ggml_are_same_shape(src0, dst));
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    const int nc = src0->ne[0];
+    const int nr = ggml_nrows(src0);
+
+    // rows per thread
+    const int dr = (nr + nth - 1)/nth;
+
+    // row range for this thread
+    const int ir0 = dr*ith;
+    const int ir1 = MIN(ir0 + dr, nr);
+
+    for (int i1 = ir0; i1 < ir1; i1++) {
+        ggml_vec_gelu_f16(nc,
+                (ggml_fp16_t *) ((char *) dst->data  + i1*( dst->nb[1])),
+                (ggml_fp16_t *) ((char *) src0->data + i1*(src0->nb[1])));
+
+#ifndef NDEBUG
+        for (int k = 0; k < nc; k++) {
+            const ggml_fp16_t x = ((ggml_fp16_t *) ((char *) dst->data + i1*( dst->nb[1])))[k];
+            const float v = GGML_FP16_TO_FP32(x);
+            UNUSED(v);
+            assert(!isnan(v));
+            assert(!isinf(v));
+        }
+#endif
+    }
+}
+
+static void ggml_compute_forward_gelu(
+        const struct ggml_compute_params * params,
+        struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    switch (src0->type) {
+        case GGML_TYPE_F32:
+            {
+                ggml_compute_forward_gelu_f32(params, dst);
+            } break;
+        case GGML_TYPE_F16:
+            {
+                ggml_compute_forward_gelu_f16(params, dst);
+            } break;
         default:
             {
                 GGML_ABORT("fatal error");
@@ -6506,11 +7528,50 @@ static void ggml_compute_forward_sigmoid(
     }
 }
 
-// ggml_compute_forward_gelu
-
-static void ggml_compute_forward_gelu_f32(
-        const struct ggml_compute_params * params,
-        struct ggml_tensor * dst) {
+// ggml_compute_forward_gelu_quick
+
+static void ggml_compute_forward_gelu_quick_f32(
+        const struct ggml_compute_params * params,
+        struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    assert(ggml_is_contiguous_1(src0));
+    assert(ggml_is_contiguous_1(dst));
+    assert(ggml_are_same_shape(src0, dst));
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    const int nc = src0->ne[0];
+    const int nr = ggml_nrows(src0);
+
+    // rows per thread
+    const int dr = (nr + nth - 1)/nth;
+
+    // row range for this thread
+    const int ir0 = dr*ith;
+    const int ir1 = MIN(ir0 + dr, nr);
+
+    for (int i1 = ir0; i1 < ir1; i1++) {
+        ggml_vec_gelu_quick_f32(nc,
+                (float *) ((char *) dst->data  + i1*( dst->nb[1])),
+                (float *) ((char *) src0->data + i1*(src0->nb[1])));
+
+#ifndef NDEBUG
+        for (int k = 0; k < nc; k++) {
+            const float x = ((float *) ((char *) dst->data + i1*( dst->nb[1])))[k];
+            UNUSED(x);
+            assert(!isnan(x));
+            assert(!isinf(x));
+        }
+#endif
+    }
+}
+
+static void ggml_compute_forward_gelu_quick_f16(
+    const struct ggml_compute_params * params,
+    struct ggml_tensor * dst) {
 
     const struct ggml_tensor * src0 = dst->src[0];
 
@@ -6532,22 +7593,23 @@ static void ggml_compute_forward_gelu_f32(
     const int ir1 = MIN(ir0 + dr, nr);
 
     for (int i1 = ir0; i1 < ir1; i1++) {
-        ggml_vec_gelu_f32(nc,
-                (float *) ((char *) dst->data  + i1*( dst->nb[1])),
-                (float *) ((char *) src0->data + i1*(src0->nb[1])));
+        ggml_vec_gelu_quick_f16(nc,
+                (ggml_fp16_t *) ((char *) dst->data  + i1*( dst->nb[1])),
+                (ggml_fp16_t *) ((char *) src0->data + i1*(src0->nb[1])));
 
 #ifndef NDEBUG
         for (int k = 0; k < nc; k++) {
-            const float x = ((float *) ((char *) dst->data + i1*( dst->nb[1])))[k];
-            UNUSED(x);
-            assert(!isnan(x));
-            assert(!isinf(x));
+            const ggml_fp16_t x = ((ggml_fp16_t *) ((char *) dst->data + i1*( dst->nb[1])))[k];
+            const float v = GGML_FP16_TO_FP32(x);
+            UNUSED(v);
+            assert(!isnan(v));
+            assert(!isinf(v));
         }
 #endif
     }
 }
 
-static void ggml_compute_forward_gelu(
+static void ggml_compute_forward_gelu_quick(
         const struct ggml_compute_params * params,
         struct ggml_tensor * dst) {
 
@@ -6556,7 +7618,11 @@ static void ggml_compute_forward_gelu(
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_gelu_f32(params, dst);
+                ggml_compute_forward_gelu_quick_f32(params, dst);
+            } break;
+        case GGML_TYPE_F16:
+            {
+                ggml_compute_forward_gelu_quick_f16(params, dst);
             } break;
         default:
             {
@@ -6565,9 +7631,9 @@ static void ggml_compute_forward_gelu(
     }
 }
 
-// ggml_compute_forward_gelu_quick
+// ggml_compute_forward_silu
 
-static void ggml_compute_forward_gelu_quick_f32(
+static void ggml_compute_forward_silu_f32(
         const struct ggml_compute_params * params,
         struct ggml_tensor * dst) {
 
@@ -6591,13 +7657,13 @@ static void ggml_compute_forward_gelu_quick_f32(
     const int ir1 = MIN(ir0 + dr, nr);
 
     for (int i1 = ir0; i1 < ir1; i1++) {
-        ggml_vec_gelu_quick_f32(nc,
+        ggml_vec_silu_f32(nc,
                 (float *) ((char *) dst->data  + i1*( dst->nb[1])),
                 (float *) ((char *) src0->data + i1*(src0->nb[1])));
 
 #ifndef NDEBUG
         for (int k = 0; k < nc; k++) {
-            const float x = ((float *) ((char *) dst->data + i1*( dst->nb[1])))[k];
+            const float x = ((float *) ((char *) dst->data + i1*(dst->nb[1])))[k];
             UNUSED(x);
             assert(!isnan(x));
             assert(!isinf(x));
@@ -6606,29 +7672,9 @@ static void ggml_compute_forward_gelu_quick_f32(
     }
 }
 
-static void ggml_compute_forward_gelu_quick(
-        const struct ggml_compute_params * params,
-        struct ggml_tensor * dst) {
-
-    const struct ggml_tensor * src0 = dst->src[0];
-
-    switch (src0->type) {
-        case GGML_TYPE_F32:
-            {
-                ggml_compute_forward_gelu_quick_f32(params, dst);
-            } break;
-        default:
-            {
-                GGML_ABORT("fatal error");
-            }
-    }
-}
-
-// ggml_compute_forward_silu
-
-static void ggml_compute_forward_silu_f32(
-        const struct ggml_compute_params * params,
-        struct ggml_tensor * dst) {
+static void ggml_compute_forward_silu_f16(
+    const struct ggml_compute_params * params,
+    struct ggml_tensor * dst) {
 
     const struct ggml_tensor * src0 = dst->src[0];
 
@@ -6650,16 +7696,17 @@ static void ggml_compute_forward_silu_f32(
     const int ir1 = MIN(ir0 + dr, nr);
 
     for (int i1 = ir0; i1 < ir1; i1++) {
-        ggml_vec_silu_f32(nc,
-                (float *) ((char *) dst->data  + i1*( dst->nb[1])),
-                (float *) ((char *) src0->data + i1*(src0->nb[1])));
+        ggml_vec_silu_f16(nc,
+                (ggml_fp16_t *) ((char *) dst->data  + i1*( dst->nb[1])),
+                (ggml_fp16_t *) ((char *) src0->data + i1*(src0->nb[1])));
 
 #ifndef NDEBUG
         for (int k = 0; k < nc; k++) {
-            const float x = ((float *) ((char *) dst->data + i1*(dst->nb[1])))[k];
-            UNUSED(x);
-            assert(!isnan(x));
-            assert(!isinf(x));
+            const ggml_fp16_t x = ((ggml_fp16_t *) ((char *) dst->data + i1*(dst->nb[1])))[k];
+            const float v = GGML_FP16_TO_FP32(x);
+            UNUSED(v);
+            assert(!isnan(v));
+            assert(!isinf(v));
         }
 #endif
     }
@@ -6676,6 +7723,10 @@ static void ggml_compute_forward_silu(
             {
                 ggml_compute_forward_silu_f32(params, dst);
             } break;
+        case GGML_TYPE_F16:
+            {
+                ggml_compute_forward_silu_f16(params, dst);
+            } break;
         default:
             {
                 GGML_ABORT("fatal error");
@@ -6714,6 +7765,36 @@ static void ggml_compute_forward_leaky_relu_f32(
     }
 }
 
+static void ggml_compute_forward_leaky_relu_f16(
+    const struct ggml_compute_params * params,
+    struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    if (params->ith != 0) {
+        return;
+    }
+
+    assert(ggml_is_contiguous_1(src0));
+    assert(ggml_is_contiguous_1(dst));
+    assert(ggml_are_same_shape(src0, dst));
+
+    const int n  = ggml_nrows(src0);
+    const int nc = src0->ne[0];
+
+    float negative_slope;
+    memcpy(&negative_slope, dst->op_params, sizeof(float));
+
+    assert(dst->nb[0]  == sizeof(ggml_fp16_t));
+    assert(src0->nb[0] == sizeof(ggml_fp16_t));
+
+    for (int i = 0; i < n; i++) {
+        ggml_vec_leaky_relu_f16(nc,
+                (ggml_fp16_t *) ((char *) dst->data  + i*( dst->nb[1])),
+                (ggml_fp16_t *) ((char *) src0->data + i*(src0->nb[1])), negative_slope);
+    }
+}
+
 static void ggml_compute_forward_leaky_relu(
         const struct ggml_compute_params * params,
         struct ggml_tensor * dst) {
@@ -6725,6 +7806,10 @@ static void ggml_compute_forward_leaky_relu(
             {
                 ggml_compute_forward_leaky_relu_f32(params, dst);
             } break;
+        case GGML_TYPE_F16:
+            {
+                ggml_compute_forward_leaky_relu_f16(params, dst);
+            } break;
         default:
             {
                 GGML_ABORT("fatal error");
@@ -6777,6 +7862,50 @@ static void ggml_compute_forward_silu_back_f32(
     }
 }
 
+static void ggml_compute_forward_silu_back_f16(
+    const struct ggml_compute_params * params,
+    struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * grad = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
+    assert(ggml_is_contiguous_1(grad));
+    assert(ggml_is_contiguous_1(src1));
+    assert(ggml_is_contiguous_1(dst));
+    assert(ggml_are_same_shape(src1, dst));
+    assert(ggml_are_same_shape(src1, grad));
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    const int nc = src1->ne[0];
+    const int nr = ggml_nrows(src1);
+
+    // rows per thread
+    const int dr = (nr + nth - 1)/nth;
+
+    // row range for this thread
+    const int ir0 = dr*ith;
+    const int ir1 = MIN(ir0 + dr, nr);
+
+    for (int i1 = ir0; i1 < ir1; i1++) {
+        ggml_vec_silu_backward_f16(nc,
+                (ggml_fp16_t *) ((char *) dst->data  + i1*( dst->nb[1])),
+                (ggml_fp16_t *) ((char *) src1->data + i1*(src1->nb[1])),
+                (ggml_fp16_t *) ((char *) grad->data + i1*(grad->nb[1])));
+
+    #ifndef NDEBUG
+        for (int k = 0; k < nc; k++) {
+            const float x = ((ggml_fp16_t *) ((char *) dst->data + i1*( dst->nb[1])))[k];
+            const float v = GGML_FP16_TO_FP32(x);
+            UNUSED(v);
+            assert(!isnan(v));
+            assert(!isinf(v));
+        }
+    #endif
+    }
+}
+
 static void ggml_compute_forward_silu_back(
         const struct ggml_compute_params * params,
         struct ggml_tensor * dst) {
@@ -6788,6 +7917,10 @@ static void ggml_compute_forward_silu_back(
             {
                 ggml_compute_forward_silu_back_f32(params, dst);
             } break;
+        case GGML_TYPE_F16:
+            {
+                ggml_compute_forward_silu_back_f16(params, dst);
+            } break;
         default:
             {
                 GGML_ABORT("fatal error");
@@ -6795,7 +7928,6 @@ static void ggml_compute_forward_silu_back(
     }
 }
 
-
 static void ggml_compute_forward_hardswish_f32(
         const struct ggml_compute_params * params,
         struct ggml_tensor * dst) {
@@ -6819,6 +7951,31 @@ static void ggml_compute_forward_hardswish_f32(
                 (float *) ((char *) src0->data + i*(src0->nb[1])));
     }
 }
+
+static void ggml_compute_forward_hardswish_f16(
+    const struct ggml_compute_params * params,
+    struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    if (params->ith != 0) {
+        return;
+    }
+
+    assert(ggml_is_contiguous_1(src0));
+    assert(ggml_is_contiguous_1(dst));
+    assert(ggml_are_same_shape(src0, dst));
+
+    const int n  = ggml_nrows(src0);
+    const int nc = src0->ne[0];
+
+    for (int i = 0; i < n; i++) {
+        ggml_vec_hardswish_f16(nc,
+                (ggml_fp16_t *) ((char *) dst->data  + i*( dst->nb[1])),
+                (ggml_fp16_t *) ((char *) src0->data + i*(src0->nb[1])));
+    }
+}
+
 static void ggml_compute_forward_hardswish(
         const struct ggml_compute_params * params,
         struct ggml_tensor * dst) {
@@ -6830,6 +7987,10 @@ static void ggml_compute_forward_hardswish(
             {
                 ggml_compute_forward_hardswish_f32(params, dst);
             } break;
+        case GGML_TYPE_F16:
+            {
+                ggml_compute_forward_hardswish_f16(params, dst);
+            } break;
         default:
             {
                 GGML_ABORT("fatal error");
@@ -6861,6 +8022,30 @@ static void ggml_compute_forward_hardsigmoid_f32(
     }
 }
 
+static void ggml_compute_forward_hardsigmoid_f16(
+    const struct ggml_compute_params * params,
+    struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    if (params->ith != 0) {
+        return;
+    }
+
+    assert(ggml_is_contiguous_1(src0));
+    assert(ggml_is_contiguous_1(dst));
+    assert(ggml_are_same_shape(src0, dst));
+
+    const int n  = ggml_nrows(src0);
+    const int nc = src0->ne[0];
+
+    for (int i = 0; i < n; i++) {
+        ggml_vec_hardsigmoid_f16(nc,
+                (ggml_fp16_t *) ((char *) dst->data  + i*( dst->nb[1])),
+                (ggml_fp16_t *) ((char *) src0->data + i*(src0->nb[1])));
+    }
+}
+
 static void ggml_compute_forward_hardsigmoid(
         const struct ggml_compute_params * params,
         struct ggml_tensor * dst) {
@@ -6872,6 +8057,10 @@ static void ggml_compute_forward_hardsigmoid(
             {
                 ggml_compute_forward_hardsigmoid_f32(params, dst);
             } break;
+        case GGML_TYPE_F16:
+            {
+                ggml_compute_forward_hardsigmoid_f16(params, dst);
+            } break;
         default:
             {
                 GGML_ABORT("fatal error");
@@ -6903,6 +8092,30 @@ static void ggml_compute_forward_exp_f32(
     }
 }
 
+static void ggml_compute_forward_exp_f16(
+    const struct ggml_compute_params * params,
+    struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    if (params->ith != 0) {
+        return;
+    }
+
+    assert(ggml_is_contiguous_1(src0));
+    assert(ggml_is_contiguous_1(dst));
+    assert(ggml_are_same_shape(src0, dst));
+
+    const int n  = ggml_nrows(src0);
+    const int nc = src0->ne[0];
+
+    for (int i = 0; i < n; i++) {
+        ggml_vec_exp_f16(nc,
+                (ggml_fp16_t *) ((char *) dst->data  + i*( dst->nb[1])),
+                (ggml_fp16_t *) ((char *) src0->data + i*(src0->nb[1])));
+    }
+}
+
 static void ggml_compute_forward_exp(
         const struct ggml_compute_params * params,
         struct ggml_tensor * dst) {
@@ -6914,6 +8127,10 @@ static void ggml_compute_forward_exp(
             {
                 ggml_compute_forward_exp_f32(params, dst);
             } break;
+        case GGML_TYPE_F16:
+            {
+                ggml_compute_forward_exp_f16(params, dst);
+            } break;
         default:
             {
                 GGML_ABORT("fatal error");
@@ -9198,6 +10415,43 @@ static void ggml_compute_forward_clamp_f32(
     }
 }
 
+static void ggml_compute_forward_clamp_f16(
+    const struct ggml_compute_params * params,
+    struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    float min;
+    float max;
+    memcpy(&min, (float *) dst->op_params + 0, sizeof(float));
+    memcpy(&max, (float *) dst->op_params + 1, sizeof(float));
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    const int n  = ggml_nrows(src0);
+    const int nc = src0->ne[0];
+
+    const size_t nb00 = src0->nb[0];
+    const size_t nb01 = src0->nb[1];
+
+    const size_t nb0 = dst->nb[0];
+    const size_t nb1 = dst->nb[1];
+
+    GGML_ASSERT( nb0 == sizeof(ggml_fp16_t));
+    GGML_ASSERT(nb00 == sizeof(ggml_fp16_t));
+
+    for (int j = ith; j < n; j += nth) {
+        ggml_fp16_t * dst_ptr  = (ggml_fp16_t *) ((char *)  dst->data + j*nb1);
+        ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + j*nb01);
+
+        for (int i = 0; i < nc; i++) {
+            float v = GGML_FP16_TO_FP32(src0_ptr[i]);
+            dst_ptr[i] = GGML_FP32_TO_FP16(MAX(MIN(v, max), min));
+        }
+    }
+}
+
 static void ggml_compute_forward_clamp(
         const struct ggml_compute_params * params,
         struct ggml_tensor * dst) {
@@ -9210,6 +10464,9 @@ static void ggml_compute_forward_clamp(
                 ggml_compute_forward_clamp_f32(params, dst);
             } break;
         case GGML_TYPE_F16:
+            {
+                ggml_compute_forward_clamp_f16(params, dst);
+            } break;
         case GGML_TYPE_BF16:
         case GGML_TYPE_Q4_0:
         case GGML_TYPE_Q4_1:
@@ -14322,6 +15579,14 @@ int ggml_cpu_has_amx_int8(void) {
 #endif
 }
 
+int ggml_cpu_has_bmi2(void) {
+#if defined(__BMI2__)
+    return 1;
+#else
+    return 0;
+#endif
+}
+
 int ggml_cpu_has_fma(void) {
 #if defined(__FMA__)
     return 1;
@@ -14402,6 +15667,14 @@ int ggml_cpu_has_vsx(void) {
 #endif
 }
 
+int ggml_cpu_has_vxe(void) {
+#if defined(__VXE__) || defined(__VXE2__)
+    return 1;
+#else
+    return 0;
+#endif
+}
+
 int ggml_cpu_has_neon(void) {
 #if defined(__ARM_ARCH) && defined(__ARM_NEON)
     return ggml_arm_arch_features.has_neon;
@@ -14442,6 +15715,14 @@ int ggml_cpu_get_sve_cnt(void) {
 #endif
 }
 
+int ggml_cpu_has_sme(void) {
+#if defined(__ARM_ARCH) && defined(__ARM_FEATURE_SME)
+    return ggml_arm_arch_features.has_sme;
+#else
+    return 0;
+#endif
+}
+
 void ggml_cpu_init(void) {
     // needed to initialize f16 tables
     {