@fugood/llama.node 1.4.8 → 1.4.9

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

@@ -692,6 +692,100 @@ void ggml_gemv_iq4_nl_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs
     }
 }
 
+void ggml_gemv_q8_0_4x4_q8_0_generic(int                        n,
+                                     float * GGML_RESTRICT      s,
+                                     size_t                     bs,
+                                     const void * GGML_RESTRICT vx,
+                                     const void * GGML_RESTRICT vy,
+                                     int                        nr,
+                                     int                        nc) {
+    const int qk                = QK8_0;
+    const int nb                = n / qk;
+    const int ncols_interleaved = 4;
+    const int blocklen          = 4;
+
+    assert(nr == 1);
+    assert(n % qk == 0);
+    assert(nc % ncols_interleaved == 0);
+
+    UNUSED(bs);
+    UNUSED(nr);
+
+    float sumf[4];
+    int   sumi;
+
+    const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
+    for (int x = 0; x < nc / ncols_interleaved; x++) {
+        const block_q8_0x4 * b_ptr = (const block_q8_0x4 *) vx + (x * nb);
+
+        for (int j = 0; j < ncols_interleaved; j++) {
+            sumf[j] = 0.0;
+        }
+        for (int l = 0; l < nb; l++) {
+            for (int k = 0; k < (qk / blocklen); k++) {
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    sumi = 0;
+                    for (int i = 0; i < blocklen; ++i) {
+                        const int v0 = b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i];
+                        sumi += v0 * a_ptr[l].qs[k * blocklen + i];
+                    }
+                    sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
+                }
+            }
+        }
+        for (int j = 0; j < ncols_interleaved; j++) {
+            s[x * ncols_interleaved + j] = sumf[j];
+        }
+    }
+}
+
+void ggml_gemv_q8_0_4x8_q8_0_generic(int                        n,
+                                     float * GGML_RESTRICT      s,
+                                     size_t                     bs,
+                                     const void * GGML_RESTRICT vx,
+                                     const void * GGML_RESTRICT vy,
+                                     int                        nr,
+                                     int                        nc) {
+    const int qk                = QK8_0;
+    const int nb                = n / qk;
+    const int ncols_interleaved = 4;
+    const int blocklen          = 8;
+
+    assert(nr == 1);
+    assert(n % qk == 0);
+    assert(nc % ncols_interleaved == 0);
+
+    UNUSED(bs);
+    UNUSED(nr);
+
+    float sumf[4];
+    int   sumi;
+
+    const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
+    for (int x = 0; x < nc / ncols_interleaved; x++) {
+        const block_q8_0x4 * b_ptr = (const block_q8_0x4 *) vx + (x * nb);
+
+        for (int j = 0; j < ncols_interleaved; j++) {
+            sumf[j] = 0.0;
+        }
+        for (int l = 0; l < nb; l++) {
+            for (int k = 0; k < (qk / blocklen); k++) {
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    sumi = 0;
+                    for (int i = 0; i < blocklen; ++i) {
+                        const int v0 = b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i];
+                        sumi += v0 * a_ptr[l].qs[k * blocklen + i];
+                    }
+                    sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
+                }
+            }
+        }
+        for (int j = 0; j < ncols_interleaved; j++) {
+            s[x * ncols_interleaved + j] = sumf[j];
+        }
+    }
+}
+
 void ggml_gemm_q4_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
     const int qk = QK8_0;
     const int nb = n / qk;
@@ -1219,8 +1313,129 @@ void ggml_gemm_iq4_nl_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs
     }
 }
 
+void ggml_gemm_q8_0_4x4_q8_0_generic(int                        n,
+                                     float * GGML_RESTRICT      s,
+                                     size_t                     bs,
+                                     const void * GGML_RESTRICT vx,
+                                     const void * GGML_RESTRICT vy,
+                                     int                        nr,
+                                     int                        nc) {
+    const int qk                = QK8_0;
+    const int nb                = n / qk;
+    const int ncols_interleaved = 4;
+    const int blocklen          = 4;
+
+    assert(n % qk == 0);
+    assert(nr % 4 == 0);
+    assert(nc % ncols_interleaved == 0);
+
+    float sumf[4][4];
+    int   sumi;
+
+    for (int y = 0; y < nr / 4; y++) {
+        const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_q8_0x4 * b_ptr = (const block_q8_0x4 *) vx + (x * nb);
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    sumf[m][j] = 0.0;
+                }
+            }
+            for (int l = 0; l < nb; l++) {
+                for (int k = 0; k < (qk / blocklen); k++) {
+                    for (int m = 0; m < 4; m++) {
+                        for (int j = 0; j < ncols_interleaved; j++) {
+                            sumi = 0;
+                            for (int i = 0; i < blocklen; ++i) {
+                                const int v0 = b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i];
+                                sumi += v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i];
+                            }
+                            sumf[m][j] +=
+                                sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
+                        }
+                    }
+                }
+            }
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j];
+                }
+            }
+        }
+    }
+}
+
+void ggml_gemm_q8_0_4x8_q8_0_generic(int                        n,
+                                     float * GGML_RESTRICT      s,
+                                     size_t                     bs,
+                                     const void * GGML_RESTRICT vx,
+                                     const void * GGML_RESTRICT vy,
+                                     int                        nr,
+                                     int                        nc) {
+    const int qk                = QK8_0;
+    const int nb                = n / qk;
+    const int ncols_interleaved = 4;
+    const int blocklen          = 8;
+
+    assert(n % qk == 0);
+    assert(nr % 4 == 0);
+    assert(nc % ncols_interleaved == 0);
+
+    float sumf[4][4];
+    int   sumi;
+
+    for (int y = 0; y < nr / 4; y++) {
+        const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_q8_0x4 * b_ptr = (const block_q8_0x4 *) vx + (x * nb);
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    sumf[m][j] = 0.0;
+                }
+            }
+            for (int l = 0; l < nb; l++) {
+                for (int k = 0; k < (qk / blocklen); k++) {
+                    for (int m = 0; m < 4; m++) {
+                        for (int j = 0; j < ncols_interleaved; j++) {
+                            sumi = 0;
+                            for (int i = 0; i < blocklen; ++i) {
+                                const int v0 = b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i];
+                                sumi += v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i];
+                            }
+                            sumf[m][j] +=
+                                sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
+                        }
+                    }
+                }
+            }
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j];
+                }
+            }
+        }
+    }
+}
+
 } // extern "C"
 
+static block_q8_0x4 make_block_q8_0x4(block_q8_0 * in, unsigned int blck_size_interleave) {
+    block_q8_0x4 out;
+
+    for (int i = 0; i < 4; i++) {
+        out.d[i] = in[i].d;
+    }
+
+    const int end = QK8_0 * 4 / blck_size_interleave;
+    for (int i = 0; i < end; ++i) {
+        int src_id     = i % 4;
+        int src_offset = (i / 4) * blck_size_interleave;
+        int dst_offset = i * blck_size_interleave;
+        memcpy(&out.qs[dst_offset], &in[src_id].qs[src_offset], blck_size_interleave);
+    }
+    return out;
+}
+
 static block_q4_0x4 make_block_q4_0x4(block_q4_0 * in, unsigned int blck_size_interleave) {
     block_q4_0x4 out;
 
@@ -1534,6 +1749,38 @@ static int repack_q4_0_to_q4_0_8_bl(struct ggml_tensor * t, int interleave_block
     GGML_UNUSED(data_size);
 }
 
+static int repack_q8_0_to_q8_0_4_bl(struct ggml_tensor *       t,
+                                    int                        interleave_block,
+                                    const void * GGML_RESTRICT data,
+                                    size_t                     data_size) {
+    GGML_ASSERT(t->type == GGML_TYPE_Q8_0);
+    GGML_ASSERT(interleave_block == 4 || interleave_block == 8);
+    constexpr int nrows_interleaved = 4;
+
+    block_q8_0x4 *     dst = (block_q8_0x4 *) t->data;
+    const block_q8_0 * src = (const block_q8_0 *) data;
+    block_q8_0         dst_tmp[4];
+    int                nrow    = ggml_nrows(t);
+    int                nblocks = t->ne[0] / QK8_0;
+
+    GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_q8_0));
+
+    if (t->ne[1] % nrows_interleaved != 0 || t->ne[0] % 8 != 0) {
+        return -1;
+    }
+
+    for (int b = 0; b < nrow; b += nrows_interleaved) {
+        for (int64_t x = 0; x < nblocks; x++) {
+            for (int i = 0; i < nrows_interleaved; i++) {
+                dst_tmp[i] = src[x + i * nblocks];
+            }
+            *dst++ = make_block_q8_0x4(dst_tmp, interleave_block);
+        }
+        src += nrows_interleaved * nblocks;
+    }
+    return 0;
+}
+
 static block_iq4_nlx4 make_block_iq4_nlx4(block_iq4_nl * in, unsigned int blck_size_interleave) {
     block_iq4_nlx4 out;
 
@@ -1702,6 +1949,14 @@ template <> int repack<block_iq4_nl, 8, 8>(struct ggml_tensor * t, const void *
     return repack_iq4_nl_to_iq4_nl_8_bl(t, 8, data, data_size);
 }
 
+template <> int repack<block_q8_0, 4, 4>(struct ggml_tensor * t, const void * data, size_t data_size) {
+    return repack_q8_0_to_q8_0_4_bl(t, 4, data, data_size);
+}
+
+template <> int repack<block_q8_0, 8, 4>(struct ggml_tensor * t, const void * data, size_t data_size) {
+    return repack_q8_0_to_q8_0_4_bl(t, 8, data, data_size);
+}
+
 // gemv
 template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PARAM_TYPE>
 void gemv(int, float *, size_t, const void *, const void *, int, int);
@@ -1738,6 +1993,14 @@ template <> void gemv<block_iq4_nl, 8, 8, GGML_TYPE_Q8_0>(int n, float * s, size
     ggml_gemv_iq4_nl_8x8_q8_0(n, s, bs, vx, vy, nr, nc);
 }
 
+template <> void gemv<block_q8_0, 4, 4, GGML_TYPE_Q8_0>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemv_q8_0_4x4_q8_0(n, s, bs, vx, vy, nr, nc);
+}
+
+template <> void gemv<block_q8_0, 8, 4, GGML_TYPE_Q8_0>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemv_q8_0_4x8_q8_0(n, s, bs, vx, vy, nr, nc);
+}
+
 // gemm
 template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PARAM_TYPE>
 void gemm(int, float *, size_t, const void *, const void *, int, int);
@@ -1774,6 +2037,14 @@ template <> void gemm<block_iq4_nl, 8, 8, GGML_TYPE_Q8_0>(int n, float * s, size
     ggml_gemm_iq4_nl_8x8_q8_0(n, s, bs, vx, vy, nr, nc);
 }
 
+template <> void gemm<block_q8_0, 4, 4, GGML_TYPE_Q8_0>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemm_q8_0_4x4_q8_0(n, s, bs, vx, vy, nr, nc);
+}
+
+template <> void gemm<block_q8_0, 8, 4, GGML_TYPE_Q8_0>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemm_q8_0_4x8_q8_0(n, s, bs, vx, vy, nr, nc);
+}
+
 class tensor_traits_base : public ggml::cpu::tensor_traits {
   public:
     virtual int repack(struct ggml_tensor * t, const void * data, size_t data_size) = 0;
@@ -2168,6 +2439,10 @@ static const ggml::cpu::tensor_traits * ggml_repack_get_optimal_repack_type(cons
     static const ggml::cpu::repack::tensor_traits<block_iq4_nl, 4, 4, GGML_TYPE_Q8_0> iq4_nl_4x4_q8_0;
     static const ggml::cpu::repack::tensor_traits<block_iq4_nl, 8, 8, GGML_TYPE_Q8_0> iq4_nl_8x8_q8_0;
 
+    // instance for Q8_0
+    static const ggml::cpu::repack::tensor_traits<block_q8_0, 4, 4, GGML_TYPE_Q8_0> q8_0_4x4_q8_0;
+    static const ggml::cpu::repack::tensor_traits<block_q8_0, 8, 4, GGML_TYPE_Q8_0> q8_0_4x8_q8_0;
+
     if (cur->type == GGML_TYPE_Q4_0) {
         if (ggml_cpu_has_avx2() || (ggml_cpu_has_sve() && ggml_cpu_has_matmul_int8() && ggml_cpu_get_sve_cnt() == QK8_0)
             || (ggml_cpu_has_riscv_v() && (ggml_cpu_get_rvv_vlen() >= QK4_0))) {
@@ -2218,6 +2493,17 @@ static const ggml::cpu::tensor_traits * ggml_repack_get_optimal_repack_type(cons
                 return &iq4_nl_4x4_q8_0;
             }
         }
+    } else if (cur->type == GGML_TYPE_Q8_0) {
+        if (ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) {
+            if (cur->ne[1] % 4 == 0) {
+                return &q8_0_4x8_q8_0;
+            }
+        }
+        if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) {
+            if (cur->ne[1] % 4 == 0) {
+                return &q8_0_4x4_q8_0;
+            }
+        }
     }
 
     return nullptr;

package/src/llama.cpp/ggml/src/ggml-cpu/repack.h

@@ -98,6 +98,10 @@ void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
 void ggml_gemm_q2_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_iq4_nl_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q8_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q8_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q8_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q8_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 
 // Native implementations
 void ggml_quantize_mat_q8_0_4x4_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
@@ -120,6 +124,10 @@ void ggml_gemm_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
 void ggml_gemm_q2_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_iq4_nl_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q8_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q8_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q8_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q8_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 
 #if defined(__cplusplus)
 } // extern "C"

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

@@ -195,8 +195,48 @@ void ggml_vec_dot_bf16(int n, float * GGML_RESTRICT s, size_t bs, ggml_bf16_t *
     sumf += (ggml_float)_mm_cvtss_f32(g);
 
 #undef LOAD
-#endif
+#elif defined(__riscv_v_intrinsic) && defined(__riscv_zvfbfwma)
+    size_t vl = __riscv_vsetvlmax_e32m4();
+
+    // initialize accumulators to all zeroes
+    vfloat32m4_t vsum0 = __riscv_vfmv_v_f_f32m4(0.0f, vl);
+    vfloat32m4_t vsum1 = __riscv_vfmv_v_f_f32m4(0.0f, vl);
+
+    // calculate step size
+    const size_t epr = __riscv_vsetvlmax_e16m2();
+    const size_t step = epr * 2;
+    const int np = (n & ~(step - 1));
+
+    // unroll by 2
+    for (; i < np; i += step) {
+        vbfloat16m2_t ax0 = __riscv_vle16_v_bf16m2((const __bf16 *)&x[i], epr);
+        vbfloat16m2_t ay0 = __riscv_vle16_v_bf16m2((const __bf16 *)&y[i], epr);
+        vsum0 = __riscv_vfwmaccbf16_vv_f32m4(vsum0, ax0, ay0, epr);
+        __asm__ __volatile__ ("" ::: "memory");
+
+        vbfloat16m2_t ax1 = __riscv_vle16_v_bf16m2((const __bf16 *)&x[i + epr], epr);
+        vbfloat16m2_t ay1 = __riscv_vle16_v_bf16m2((const __bf16 *)&y[i + epr], epr);
+        vsum1 = __riscv_vfwmaccbf16_vv_f32m4(vsum1, ax1, ay1, epr);
+        __asm__ __volatile__ ("" ::: "memory");
+    }
 
+    // accumulate in 1 register
+    vsum0 = __riscv_vfadd_vv_f32m4(vsum0, vsum1, vl);
+
+    // leftovers
+    for (i = np; i < n; i += vl) {
+        vl = __riscv_vsetvl_e16m2(n - i);
+        vbfloat16m2_t ax0 = __riscv_vle16_v_bf16m2((const __bf16 *)&x[i], vl);
+        vbfloat16m2_t ay0 = __riscv_vle16_v_bf16m2((const __bf16 *)&y[i], vl);
+        vsum0 = __riscv_vfwmaccbf16_vv_f32m4(vsum0, ax0, ay0, vl);
+    }
+
+    // reduce
+    vl = __riscv_vsetvlmax_e32m4();
+    vfloat32m1_t redsum = __riscv_vfredusum_vs_f32m4_f32m1(vsum0, __riscv_vfmv_v_f_f32m1(0.0f, 1), vl);
+    sumf += __riscv_vfmv_f_s_f32m1_f32(redsum);
+
+#endif
     for (; i < n; ++i) {
         sumf += (ggml_float)(GGML_BF16_TO_FP32(x[i]) *
                              GGML_BF16_TO_FP32(y[i]));

package/src/llama.cpp/ggml/src/ggml-cpu/vec.h

@@ -224,13 +224,71 @@ inline static void ggml_vec_dot_f16_unroll(const int n, const int xs, float * GG
         }
         GGML_F16x_VEC_REDUCE(sumf[0], sum_00, sum_01, sum_02, sum_03);
         GGML_F16x_VEC_REDUCE(sumf[1], sum_10, sum_11, sum_12, sum_13);
-    #elif defined(__riscv_v_intrinsic)
-      // todo: RVV impl
-      for (int i = 0; i < n; ++i) {
-          for (int j = 0; j < GGML_VEC_DOT_UNROLL; ++j) {
-              sumf[j] += (ggml_float)(GGML_CPU_FP16_TO_FP32(x[j][i])*GGML_CPU_FP16_TO_FP32(y[i]));
-          }
-      }
+
+    #elif defined(__riscv_v_intrinsic) && defined(__riscv_zvfh)
+        size_t vl = __riscv_vsetvlmax_e32m4();
+
+        // initialize accumulators to all zeroes
+        vfloat32m4_t vsum0_0 = __riscv_vfmv_v_f_f32m4(0.0f, vl);
+        vfloat32m4_t vsum0_1 = __riscv_vfmv_v_f_f32m4(0.0f, vl);
+        vfloat32m4_t vsum1_0 = __riscv_vfmv_v_f_f32m4(0.0f, vl);
+        vfloat32m4_t vsum1_1 = __riscv_vfmv_v_f_f32m4(0.0f, vl);
+
+        // calculate step size
+        const size_t epr = __riscv_vsetvlmax_e16m2();
+        const size_t step = epr * 2;
+        const int np = (n & ~(step - 1));
+
+        // unroll by 2 along the row dimension
+        for (int i = 0; i < np; i += step) {
+            vfloat16m2_t ay0 = __riscv_vle16_v_f16m2((const _Float16 *)(y + i), epr);
+            vfloat16m2_t ax0_0 = __riscv_vle16_v_f16m2((const _Float16 *)(x[0] + i), epr);
+            vfloat16m2_t ax1_0 = __riscv_vle16_v_f16m2((const _Float16 *)(x[1] + i), epr);
+            vsum0_0 = __riscv_vfwmacc_vv_f32m4(vsum0_0, ax0_0, ay0, epr);
+            vsum1_0 = __riscv_vfwmacc_vv_f32m4(vsum1_0, ax1_0, ay0, epr);
+
+            vfloat16m2_t ay1 = __riscv_vle16_v_f16m2((const _Float16 *)(y + i + epr), epr);
+            vfloat16m2_t ax0_1 = __riscv_vle16_v_f16m2((const _Float16 *)(x[0] + i + epr), epr);
+            vfloat16m2_t ax1_1 = __riscv_vle16_v_f16m2((const _Float16 *)(x[1] + i + epr), epr);
+            vsum0_1 = __riscv_vfwmacc_vv_f32m4(vsum0_1, ax0_1, ay1, epr);
+            vsum1_1 = __riscv_vfwmacc_vv_f32m4(vsum1_1, ax1_1, ay1, epr);
+        }
+
+        vfloat32m4_t vsum0 = __riscv_vfadd_vv_f32m4(vsum0_0, vsum0_1, vl);
+        vfloat32m4_t vsum1 = __riscv_vfadd_vv_f32m4(vsum1_0, vsum1_1, vl);
+
+        // leftovers
+        for (int i = np; i < n; i += vl) {
+            vl = __riscv_vsetvl_e16m2(n - i);
+            vfloat16m2_t ay = __riscv_vle16_v_f16m2((const _Float16 *)(y + i), vl);
+            vfloat16m2_t ax0 = __riscv_vle16_v_f16m2((const _Float16 *)(x[0] + i), vl);
+            vfloat16m2_t ax1 = __riscv_vle16_v_f16m2((const _Float16 *)(x[1] + i), vl);
+
+            vsum0 = __riscv_vfwmacc_vv_f32m4(vsum0, ax0, ay, vl);
+            vsum1 = __riscv_vfwmacc_vv_f32m4(vsum1, ax1, ay, vl);
+        }
+
+        // reduce
+        vl = __riscv_vsetvlmax_e32m2();
+        vfloat32m2_t acc0_0 = __riscv_vfadd_vv_f32m2(__riscv_vget_v_f32m4_f32m2(vsum0, 0),
+                                    __riscv_vget_v_f32m4_f32m2(vsum0, 1), vl);
+        vl = __riscv_vsetvlmax_e32m1();
+        vfloat32m1_t acc0_1 = __riscv_vfadd_vv_f32m1(__riscv_vget_v_f32m2_f32m1(acc0_0, 0),
+        __riscv_vget_v_f32m2_f32m1(acc0_0, 1), vl);
+        vfloat32m1_t redsum0 = __riscv_vfredusum_vs_f32m1_f32m1(
+                                    acc0_1, __riscv_vfmv_v_f_f32m1(0.0f, 1), vl);
+
+        vl = __riscv_vsetvlmax_e32m2();
+        vfloat32m2_t acc1_0 = __riscv_vfadd_vv_f32m2(__riscv_vget_v_f32m4_f32m2(vsum1, 0),
+                                    __riscv_vget_v_f32m4_f32m2(vsum1, 1), vl);
+        vl = __riscv_vsetvlmax_e32m1();
+        vfloat32m1_t acc1_1 = __riscv_vfadd_vv_f32m1(__riscv_vget_v_f32m2_f32m1(acc1_0, 0),
+                                    __riscv_vget_v_f32m2_f32m1(acc1_0, 1), vl);
+        vfloat32m1_t redsum1 = __riscv_vfredusum_vs_f32m1_f32m1(
+                                    acc1_1, __riscv_vfmv_v_f_f32m1(0.0f, 1), vl);
+        sumf[0] = __riscv_vfmv_f_s_f32m1_f32(redsum0);
+        sumf[1] = __riscv_vfmv_f_s_f32m1_f32(redsum1);
+
     #else
         const int np = (n & ~(GGML_F16_STEP - 1));
 
@@ -475,15 +533,39 @@ inline static void ggml_vec_mad_f16(const int n, ggml_fp16_t * GGML_RESTRICT y,
     }
     np = n;
 #elif defined(__riscv_zvfh) // implies __riscv_v_intrinsic
-    const int np = n;
-    _Float16 hv = (_Float16)v;
-    for (int i = 0, avl; i < n; i += avl) {
-        avl = __riscv_vsetvl_e16m8(n - i);
-        vfloat16m8_t ax = __riscv_vle16_v_f16m8((const _Float16 *)&x[i], avl);
-        vfloat16m8_t ay = __riscv_vle16_v_f16m8((_Float16 *)&y[i], avl);
-        vfloat16m8_t ny = __riscv_vfmadd_vf_f16m8(ax, hv, ay, avl);
-        __riscv_vse16_v_f16m8((_Float16 *)&y[i], ny, avl);
+    const ggml_fp16_t s = GGML_CPU_FP32_TO_FP16(v);
+    const _Float16 scale = *(const _Float16*)(&s);
+
+    // calculate step size
+    const int epr = __riscv_vsetvlmax_e16m4();
+    const int step = epr * 2;
+    int np = (n & ~(step - 1));
+
+    // unroll by 2
+    for (int i = 0; i < np; i += step) {
+        vfloat16m4_t ax0 = __riscv_vle16_v_f16m4((const _Float16*)x + i, epr);
+        vfloat16m4_t ay0 = __riscv_vle16_v_f16m4((const _Float16*)y + i, epr);
+        ay0 = __riscv_vfmacc_vf_f16m4(ay0, scale, ax0, epr);
+        __riscv_vse16_v_f16m4((_Float16*)y + i, ay0, epr);
+        __asm__ __volatile__ ("" ::: "memory");
+
+        vfloat16m4_t ax1 = __riscv_vle16_v_f16m4((const _Float16*)x + i + epr, epr);
+        vfloat16m4_t ay1 = __riscv_vle16_v_f16m4((const _Float16*)y + i + epr, epr);
+        ay1 = __riscv_vfmacc_vf_f16m4(ay1, scale, ax1, epr);
+        __riscv_vse16_v_f16m4((_Float16*)y + i + epr, ay1, epr);
+        __asm__ __volatile__ ("" ::: "memory");
+    }
+
+    // leftovers
+    int vl;
+    for (int i = np; i < n; i += vl) {
+        vl = __riscv_vsetvl_e16m4(n - i);
+        vfloat16m4_t ax0 = __riscv_vle16_v_f16m4((const _Float16*)x + i, vl);
+        vfloat16m4_t ay0 = __riscv_vle16_v_f16m4((const _Float16*)y + i, vl);
+        ay0 = __riscv_vfmacc_vf_f16m4(ay0, scale, ax0, vl);
+        __riscv_vse16_v_f16m4((_Float16*)y + i, ay0, vl);
     }
+    np = n;
 #elif defined(GGML_SIMD)
     const int np = (n & ~(GGML_F16_STEP - 1));
 
@@ -724,13 +806,34 @@ inline static void ggml_vec_scale_f16(const int n, ggml_fp16_t * y, const float
         svst1_f16(pg, (__fp16 *)(y + np), out);
     }
 #elif defined(__riscv_v_intrinsic) && defined(__riscv_zvfh)
-    for (int i = 0, vl; i < n; i += vl) {
-        vl = __riscv_vsetvl_e16m2(n - i);
-        vfloat16m2_t vy = __riscv_vle16_v_f16m2((_Float16 *)&y[i], vl);
-        vfloat32m4_t vy32 = __riscv_vfwcvt_f_f_v_f32m4(vy, vl);
-        vy32 = __riscv_vfmul_vf_f32m4(vy32, v, vl);
-        vy = __riscv_vfncvt_f_f_w_f16m2(vy32, vl);
-        __riscv_vse16_v_f16m2((_Float16 *)&y[i], vy, vl);
+    const ggml_fp16_t s = GGML_CPU_FP32_TO_FP16(v);
+    const _Float16 scale = *(const _Float16*)(&s);
+
+    // calculate step size
+    const int epr = __riscv_vsetvlmax_e16m4();
+    const int step = epr * 2;
+    const int np = (n & ~(step - 1));
+
+    // unroll by 2
+    for (int i = 0; i < np; i += step) {
+        vfloat16m4_t ay0 = __riscv_vle16_v_f16m4((const _Float16*)y + i, epr);
+        ay0 = __riscv_vfmul_vf_f16m4(ay0, scale, epr);
+        __riscv_vse16_v_f16m4((_Float16*)y + i, ay0, epr);
+        __asm__ __volatile__ ("" ::: "memory");
+
+        vfloat16m4_t ay1 = __riscv_vle16_v_f16m4((const _Float16*)y + i + epr, epr);
+        ay1 = __riscv_vfmul_vf_f16m4(ay1, scale, epr);
+        __riscv_vse16_v_f16m4((_Float16*)y + i + epr, ay1, epr);
+        __asm__ __volatile__ ("" ::: "memory");
+    }
+
+    // leftovers
+    int vl;
+    for (int i = np; i < n; i += vl) {
+        vl = __riscv_vsetvl_e16m4(n - i);
+        vfloat16m4_t ay0 = __riscv_vle16_v_f16m4((const _Float16*)y + i, vl);
+        ay0 = __riscv_vfmul_vf_f16m4(ay0, scale, vl);
+        __riscv_vse16_v_f16m4((_Float16*)y + i, ay0, vl);
     }
 #elif defined(GGML_SIMD)
     const int np = (n & ~(GGML_F16_STEP - 1));

package/src/llama.cpp/src/llama-arch.cpp

@@ -2055,7 +2055,7 @@ static std::set<llm_tensor> llm_get_tensor_names(llm_arch arch) {
                 LLM_TENSOR_SHORTCONV_INPROJ,
                 LLM_TENSOR_SHORTCONV_OUTPROJ,
                 LLM_TENSOR_TOKEN_EMBD,
-                LLM_TENSOR_OUTPUT_NORM,
+                LLM_TENSOR_OUTPUT_NORM_LFM2,
                 LLM_TENSOR_FFN_GATE_INP,
                 LLM_TENSOR_FFN_GATE_EXPS,
                 LLM_TENSOR_FFN_DOWN_EXPS,