@fugood/llama.node 0.3.14 → 0.3.16

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

@@ -8158,7 +8158,156 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
     const int nb = n / QK_K;
 
-#ifdef __ARM_NEON
+#ifdef __ARM_FEATURE_SVE
+    const int vector_length = ggml_cpu_get_sve_cnt()*8;
+    float sum = 0;
+    svuint8_t m4b = svdup_n_u8(0xf);
+    svint32_t vzero = svdup_n_s32(0);
+    svuint8_t mone = svdup_n_u8(0x30);
+    svint8_t q6bytes_1, q6bytes_2, q6bytes_3, q6bytes_4;
+    svuint8_t q6h_1, q6h_2, q6h_3, q6h_4;
+
+    for (int i = 0; i < nb; ++i) {
+        const float d_all = GGML_FP16_TO_FP32(x[i].d);
+
+        const uint8_t * GGML_RESTRICT q6 = x[i].ql;
+        const uint8_t * GGML_RESTRICT qh = x[i].qh;
+        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
+
+        const int8_t * GGML_RESTRICT scale = x[i].scales;
+
+        const svbool_t pg16_8 = svptrue_pat_b16(SV_VL8);
+        const svint16_t q8sums_1 = svld1_s16(pg16_8, y[i].bsums);
+        const svint16_t q8sums_2 = svld1_s16(pg16_8, y[i].bsums + 8);
+        const svint16_t q6scales_1 = svunpklo_s16(svld1_s8(svptrue_pat_b8(SV_VL8), scale));
+        const svint16_t q6scales_2 = svunpklo_s16(svld1_s8(svptrue_pat_b8(SV_VL8), scale + 8));
+        const svint64_t prod = svdup_n_s64(0);
+        int32_t isum_mins = svaddv_s64(svptrue_b64(), svadd_s64_x(svptrue_b64(), svdot_s64(prod, q8sums_1, q6scales_1),
+                                                                                 svdot_s64(prod, q8sums_2, q6scales_2)));
+        int32_t isum = 0;
+
+        switch (vector_length) {
+            case 128:
+                {
+                    const svbool_t pg32_4 = svptrue_pat_b32(SV_VL4);
+                    const svbool_t pg8_16 = svptrue_pat_b8(SV_VL16);
+                    svint32_t isum_tmp = svdup_n_s32(0);
+                    for (int j = 0; j < QK_K/128; ++j) {
+                        svuint8_t qhbits_1 = svld1_u8(pg8_16, qh);
+                        svuint8_t qhbits_2 = svld1_u8(pg8_16, qh+16);
+                        qh += 32;
+                        svuint8_t q6bits_1 = svld1_u8(pg8_16, q6);
+                        svuint8_t q6bits_2 = svld1_u8(pg8_16, q6+16);
+                        svuint8_t q6bits_3 = svld1_u8(pg8_16, q6+32);
+                        svuint8_t q6bits_4 = svld1_u8(pg8_16, q6+48);
+                        q6 += 64;
+                        svint8_t q8bytes_1 = svld1_s8(pg8_16, q8);
+                        svint8_t q8bytes_2 = svld1_s8(pg8_16, q8+16);
+                        svint8_t q8bytes_3 = svld1_s8(pg8_16, q8+32);
+                        svint8_t q8bytes_4 = svld1_s8(pg8_16, q8+48);
+                        q8 += 64;
+
+                        q6h_1 = svand_u8_x(pg16_8, mone, svlsl_n_u8_x(pg16_8, qhbits_1, 4));
+                        q6h_2 = svand_u8_x(pg16_8, mone, svlsl_n_u8_x(pg16_8, qhbits_2, 4));
+                        q6h_3 = svand_u8_x(pg16_8, mone, svlsl_n_u8_x(pg16_8, qhbits_1, 2));
+                        q6h_4 = svand_u8_x(pg16_8, mone, svlsl_n_u8_x(pg16_8, qhbits_2, 2));
+                        q6bytes_1 = svreinterpret_s8_u8(svorr_u8_x(pg8_16, svand_u8_x(pg8_16, q6bits_1, m4b), q6h_1));
+                        q6bytes_2 = svreinterpret_s8_u8(svorr_u8_x(pg8_16, svand_u8_x(pg8_16, q6bits_2, m4b), q6h_2));
+                        q6bytes_3 = svreinterpret_s8_u8(svorr_u8_x(pg8_16, svand_u8_x(pg8_16, q6bits_3, m4b), q6h_3));
+                        q6bytes_4 = svreinterpret_s8_u8(svorr_u8_x(pg8_16, svand_u8_x(pg8_16, q6bits_4, m4b), q6h_4));
+                        isum_tmp = svmla_n_s32_x(pg32_4, isum_tmp, svdot_s32(vzero, q6bytes_1, q8bytes_1), scale[0]);
+                        isum_tmp = svmla_n_s32_x(pg32_4, isum_tmp, svdot_s32(vzero, q6bytes_2, q8bytes_2), scale[1]);
+                        isum_tmp = svmla_n_s32_x(pg32_4, isum_tmp, svdot_s32(vzero, q6bytes_3, q8bytes_3), scale[2]);
+                        isum_tmp = svmla_n_s32_x(pg32_4, isum_tmp, svdot_s32(vzero, q6bytes_4, q8bytes_4), scale[3]);
+
+                        scale += 4;
+                        q8bytes_1 = svld1_s8(pg8_16, q8);
+                        q8bytes_2 = svld1_s8(pg8_16, q8+16);
+                        q8bytes_3 = svld1_s8(pg8_16, q8+32);
+                        q8bytes_4 = svld1_s8(pg8_16, q8+48);
+                        q8 += 64;
+
+                        q6h_1 = svand_u8_x(pg16_8, mone, qhbits_1);
+                        q6h_2 = svand_u8_x(pg16_8, mone, qhbits_2);
+                        q6h_3 = svand_u8_x(pg16_8, mone, svlsr_n_u8_x(pg16_8, qhbits_1, 2));
+                        q6h_4 = svand_u8_x(pg16_8, mone, svlsr_n_u8_x(pg16_8, qhbits_2, 2));
+                        q6bytes_1 = svreinterpret_s8_u8(svorr_u8_x(pg8_16, svlsr_n_u8_x(pg8_16, q6bits_1, 4), q6h_1));
+                        q6bytes_2 = svreinterpret_s8_u8(svorr_u8_x(pg8_16, svlsr_n_u8_x(pg8_16, q6bits_2, 4), q6h_2));
+                        q6bytes_3 = svreinterpret_s8_u8(svorr_u8_x(pg8_16, svlsr_n_u8_x(pg8_16, q6bits_3, 4), q6h_3));
+                        q6bytes_4 = svreinterpret_s8_u8(svorr_u8_x(pg8_16, svlsr_n_u8_x(pg8_16, q6bits_4, 4), q6h_4));
+                        isum_tmp = svmla_n_s32_x(pg32_4, isum_tmp, svdot_s32(vzero, q6bytes_1, q8bytes_1), scale[0]);
+                        isum_tmp = svmla_n_s32_x(pg32_4, isum_tmp, svdot_s32(vzero, q6bytes_2, q8bytes_2), scale[1]);
+                        isum_tmp = svmla_n_s32_x(pg32_4, isum_tmp, svdot_s32(vzero, q6bytes_3, q8bytes_3), scale[2]);
+                        isum_tmp = svmla_n_s32_x(pg32_4, isum_tmp, svdot_s32(vzero, q6bytes_4, q8bytes_4), scale[3]);
+                        scale += 4;
+                    }
+                    isum += svaddv_s32(pg32_4, isum_tmp);
+                    sum += d_all * y[i].d * (isum - 32 * isum_mins);
+                }
+                break;
+            case 256:
+            case 512:
+                {
+                    const svbool_t pg8_2 = svptrue_pat_b8(SV_VL2);
+                    const svbool_t pg32_8 = svptrue_pat_b32(SV_VL8);
+                    const svbool_t pg8_32 = svptrue_pat_b8(SV_VL32);
+                    svint32_t isum_tmp = svdup_n_s32(0);
+                    for (int j = 0; j < QK_K/128; j++) {
+                        svuint8_t qhbits_1 = svld1_u8(pg8_32, qh);
+                        qh += 32;
+                        svuint8_t q6bits_1 = svld1_u8(pg8_32, q6);
+                        svuint8_t q6bits_2 = svld1_u8(pg8_32, q6+32);
+                        q6 += 64;
+                        svint8_t q8bytes_1 = svld1_s8(pg8_32, q8);
+                        svint8_t q8bytes_2 = svld1_s8(pg8_32, q8+32);
+                        svint8_t q8bytes_3 = svld1_s8(pg8_32, q8+64);
+                        svint8_t q8bytes_4 = svld1_s8(pg8_32, q8+96);
+                        q8 += 128;
+                        q6h_1 = svand_u8_x(pg8_32, mone, svlsl_n_u8_x(pg8_32, qhbits_1, 4));
+                        q6h_2 = svand_u8_x(pg8_32, mone, svlsl_n_u8_x(pg8_32, qhbits_1, 2));
+                        q6h_3 = svand_u8_x(pg8_32, mone, qhbits_1);
+                        q6h_4 = svand_u8_x(pg8_32, mone, svlsr_n_u8_x(pg8_32, qhbits_1, 2));
+                        q6bytes_1 = svreinterpret_s8_u8(svorr_u8_x(pg8_32, svand_u8_x(pg8_32, q6bits_1, m4b), q6h_1));
+                        q6bytes_2 = svreinterpret_s8_u8(svorr_u8_x(pg8_32, svand_u8_x(pg8_32, q6bits_2, m4b), q6h_2));
+                        q6bytes_3 = svreinterpret_s8_u8(svorr_u8_x(pg8_32, svlsr_n_u8_x(pg8_32, q6bits_1, 4), q6h_3));
+                        q6bytes_4 = svreinterpret_s8_u8(svorr_u8_x(pg8_32, svlsr_n_u8_x(pg8_32, q6bits_2, 4), q6h_4));
+
+                        svint8_t scale_lane_1_tmp = svld1_s8(pg8_2, scale);
+                        scale_lane_1_tmp= svzip1_s8(scale_lane_1_tmp, scale_lane_1_tmp);
+                        scale_lane_1_tmp= svzip1_s8(scale_lane_1_tmp, scale_lane_1_tmp);
+                        svint8_t scale_lane_2_tmp = svld1_s8(pg8_2, scale+2);
+                        scale_lane_2_tmp = svzip1_s8(scale_lane_2_tmp, scale_lane_2_tmp);
+                        scale_lane_2_tmp = svzip1_s8(scale_lane_2_tmp, scale_lane_2_tmp);
+                        svint8_t scale_lane_3_tmp = svld1_s8(pg8_2, scale+4);
+                        scale_lane_3_tmp = svzip1_s8(scale_lane_3_tmp, scale_lane_3_tmp);
+                        scale_lane_3_tmp = svzip1_s8(scale_lane_3_tmp, scale_lane_3_tmp);
+                        svint8_t scale_lane_4_tmp = svld1_s8(pg8_2, scale+6);
+                        scale_lane_4_tmp = svzip1_s8(scale_lane_4_tmp, scale_lane_4_tmp);
+                        scale_lane_4_tmp = svzip1_s8(scale_lane_4_tmp, scale_lane_4_tmp);
+                        svint32_t scale_lane_1 = svunpklo_s32(svunpklo_s16(scale_lane_1_tmp));
+                        svint32_t scale_lane_2 = svunpklo_s32(svunpklo_s16(scale_lane_2_tmp));
+                        svint32_t scale_lane_3 = svunpklo_s32(svunpklo_s16(scale_lane_3_tmp));
+                        svint32_t scale_lane_4 = svunpklo_s32(svunpklo_s16(scale_lane_4_tmp));
+
+                        isum_tmp = svmla_s32_x(pg32_8, isum_tmp, svdot_s32(vzero, q6bytes_1, q8bytes_1), scale_lane_1);
+                        isum_tmp = svmla_s32_x(pg32_8, isum_tmp, svdot_s32(vzero, q6bytes_2, q8bytes_2), scale_lane_2);
+                        isum_tmp = svmla_s32_x(pg32_8, isum_tmp, svdot_s32(vzero, q6bytes_3, q8bytes_3), scale_lane_3);
+                        isum_tmp = svmla_s32_x(pg32_8, isum_tmp, svdot_s32(vzero, q6bytes_4, q8bytes_4), scale_lane_4);
+                        scale += 8;
+                    }
+                    isum += svaddv_s32(pg32_8, isum_tmp);
+                    sum += d_all * y[i].d * (isum - 32 * isum_mins);
+                }
+                break;
+            default:
+                assert(false && "Unsupported vector length");
+                break;
+        }
+    }
+
+    *s = sum;
+
+#elif __ARM_NEON
     float sum = 0;
 
     const uint8x16_t m4b = vdupq_n_u8(0xF);

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

@@ -3110,17 +3110,17 @@ static void ggml_compute_forward_dup_same_cont(
     const int ith = params->ith; // thread index
     const int nth = params->nth; // number of threads
 
-    // parallelize by elements
-    const int ne = ggml_nelements(dst);
-    const int dr = (ne + nth - 1) / nth;
-    const int ie0 = dr * ith;
-    const int ie1 = MIN(ie0 + dr, ne);
+    // parallelize by blocks
+    const int nk = ggml_nelements(src0)/ggml_blck_size(src0->type);
+    const int dr = (nk + nth - 1) / nth;
+    const int k0 = dr * ith;
+    const int k1 = MIN(k0 + dr, nk);
 
-    if (ie0 < ie1) {
+    if (k0 < k1) {
         memcpy(
-            ((char *)  dst->data + ie0*nb0),
-            ((char *) src0->data + ie0*nb0),
-            (ie1 - ie0) * nb0);
+            ((char *)  dst->data + k0*nb0),
+            ((char *) src0->data + k0*nb0),
+            (k1 - k0) * nb0);
     }
 }
 
@@ -4055,7 +4055,6 @@ static void ggml_compute_forward_dup_f32(
 static void ggml_compute_forward_dup_bytes(
         const struct ggml_compute_params * params,
         struct ggml_tensor * dst) {
-
     const struct ggml_tensor * src0 = dst->src[0];
 
     GGML_ASSERT(ggml_nelements(dst) == ggml_nelements(src0));
@@ -4069,10 +4068,10 @@ static void ggml_compute_forward_dup_bytes(
     }
 
     const size_t type_size = ggml_type_size(src0->type);
+
     const int ith = params->ith; // thread index
     const int nth = params->nth; // number of threads
 
-
     // parallelize by rows
     const int nr = ne01;
     // number of rows per thread
@@ -4082,10 +4081,10 @@ static void ggml_compute_forward_dup_bytes(
     const int ir1 = MIN(ir0 + dr, nr);
 
     if (src0->type == dst->type &&
-        ne00 == ne0 &&
+        ggml_are_same_shape(src0, dst) &&
         nb00 == type_size && nb0 == type_size) {
         // copy by rows
-        const size_t rs = ne00 * type_size;
+        const size_t rs = ggml_row_size(src0->type, ne00);
         for (int64_t i03 = 0; i03 < ne03; i03++) {
             for (int64_t i02 = 0; i02 < ne02; i02++) {
                 for (int64_t i01 = ir0; i01 < ir1; i01++) {
@@ -4140,17 +4139,20 @@ static void ggml_compute_forward_dup_bytes(
     }
 
     // dst counters
-
-    int64_t i10 = 0;
+    int64_t k10 = 0;
     int64_t i11 = 0;
     int64_t i12 = 0;
     int64_t i13 = 0;
 
+    // number of blocks in a row
+    const int64_t nk00 = ne00 / ggml_blck_size(src0->type);
+    const int64_t nk0  = ne0  / ggml_blck_size(dst->type);
+
     for (int64_t i03 = 0; i03 < ne03; i03++) {
         for (int64_t i02 = 0; i02 < ne02; i02++) {
-            i10 += ne00 * ir0;
-            while (i10 >= ne0) {
-                i10 -= ne0;
+            k10 += nk00 * ir0;
+            while (k10 >= nk0) {
+                k10 -= nk0;
                 if (++i11 == ne1) {
                     i11 = 0;
                     if (++i12 == ne2) {
@@ -4162,14 +4164,14 @@ static void ggml_compute_forward_dup_bytes(
                 }
             }
             for (int64_t i01 = ir0; i01 < ir1; i01++) {
-                for (int64_t i00 = 0; i00 < ne00; i00++) {
-                    const char * src0_ptr = ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03);
-                          char * dst_ptr  = ((char *)  dst->data + i10*nb0  + i11*nb1  + i12*nb2  + i13*nb3);
+                for (int64_t k00 = 0; k00 < nk00; k00++) {
+                    const char * src0_ptr = ((char *) src0->data + k00*nb00 + i01*nb01 + i02*nb02 + i03*nb03);
+                          char * dst_ptr  = ((char *)  dst->data + k10*nb0  + i11*nb1  + i12*nb2  + i13*nb3);
 
                     memcpy(dst_ptr, src0_ptr, type_size);
 
-                    if (++i10 == ne0) {
-                        i10 = 0;
+                    if (++k10 == nk0) {
+                        k10 = 0;
                         if (++i11 == ne1) {
                             i11 = 0;
                             if (++i12 == ne2) {
@@ -4182,9 +4184,9 @@ static void ggml_compute_forward_dup_bytes(
                     }
                 }
             }
-            i10 += ne00 * (ne01 - ir1);
-            while (i10 >= ne0) {
-                i10 -= ne0;
+            k10 += nk00 * (ne01 - ir1);
+            while (k10 >= nk0) {
+                k10 -= nk0;
                 if (++i11 == ne1) {
                     i11 = 0;
                     if (++i12 == ne2) {
@@ -8548,6 +8550,69 @@ static void ggml_compute_forward_group_norm(
     }
 }
 
+// ggml_compute_forward_l2_norm
+
+static void ggml_compute_forward_l2_norm_f32(
+    const struct ggml_compute_params * params,
+    struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    GGML_ASSERT(ggml_are_same_shape(src0, dst));
+
+    GGML_ASSERT(src0->nb[0] == sizeof(float));
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    GGML_TENSOR_UNARY_OP_LOCALS
+
+    float eps;
+    memcpy(&eps, dst->op_params, sizeof(float));
+
+    GGML_ASSERT(eps >= 0.0f);
+
+    // TODO: optimize
+    for (int64_t i03 = 0; i03 < ne03; i03++) {
+        for (int64_t i02 = 0; i02 < ne02; i02++) {
+            for (int64_t i01 = ith; i01 < ne01; i01 += nth) {
+                const float * x = (float *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03);
+
+                ggml_float sum = 0.0;
+                for (int64_t i00 = 0; i00 < ne00; i00++) {
+                    sum += (ggml_float)(x[i00] * x[i00]);
+                }
+
+                float * y = (float *) ((char *) dst->data + i01*nb1 + i02*nb2 + i03*nb3);
+
+                memcpy(y, x, ne00 * sizeof(float));
+
+                const float scale = 1.0f/fmaxf(sqrtf(sum), eps);
+
+                ggml_vec_scale_f32(ne00, y, scale);
+            }
+        }
+    }
+}
+
+static void ggml_compute_forward_l2_norm(
+    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_l2_norm_f32(params, dst);
+            } break;
+        default:
+            {
+                GGML_ABORT("fatal error");
+            }
+    }
+}
+
 // ggml_compute_forward_mul_mat
 
 static void ggml_compute_forward_mul_mat_one_chunk(
@@ -13604,6 +13669,184 @@ static void ggml_compute_forward_gla(
     }
 }
 
+// ggml_compute_forward_rwkv_wkv7
+
+static void ggml_compute_forward_rwkv_wkv7_f32(
+        const struct ggml_compute_params * params,
+        struct ggml_tensor * dst) {
+    const int64_t T = dst->src[1]->ne[2];
+    const int64_t C = dst->ne[0];
+    const int64_t HEADS = dst->src[1]->ne[1];
+    const int64_t n_seqs = dst->src[6]->ne[1];
+    const int64_t head_size = C / HEADS;
+
+    float * dst_data = (float *) dst->data;
+    float * state = ((float *) dst->data) + C * T;
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    if (ith >= HEADS) {
+        return;
+    }
+
+    const int h_start = (HEADS * ith) / nth;
+    const int h_end = ((HEADS * (ith + 1)) / nth < HEADS) ?
+                (HEADS * (ith + 1)) / nth : HEADS;
+
+    float * r = (float *) dst->src[0]->data;
+    float * w = (float *) dst->src[1]->data;
+    float * k = (float *) dst->src[2]->data;
+    float * v = (float *) dst->src[3]->data;
+    float * a = (float *) dst->src[4]->data;
+    float * b = (float *) dst->src[5]->data;
+
+    int64_t t_stride = HEADS * head_size; // Same to C
+
+    int64_t h_stride = C / HEADS;
+    GGML_ASSERT(C % HEADS == 0); // C must be divisible by HEADS
+    int64_t h_stride_2d = head_size * head_size;
+
+    #if defined(GGML_SIMD)
+        for (int64_t t = 0; t < T; t++) {
+            int64_t t_offset = t * t_stride;
+            int64_t state_offset = head_size * C * (t / (T / n_seqs));
+            float * state_cur = state + state_offset;
+            float * state_prev = t % (T / n_seqs) ? state_cur : (float*)dst->src[6]->data + state_offset;
+
+            for (int64_t h = h_start; h < h_end; h++) {
+                int64_t h_offset = h * h_stride;
+                int64_t t_h_offset = t_offset + h_offset;
+                int64_t h_2d_offset = h * h_stride_2d;
+
+                for (int64_t ii = 0; ii < head_size; ii++) {
+                    int64_t t_h_i_offset = t_h_offset + ii;
+                    int64_t h_2d_i_offset = h_2d_offset + ii * h_stride;
+
+                    GGML_F32_VEC v_vec = GGML_F32_VEC_SET1(v[t_h_i_offset]);
+
+                    float sa = 0;
+                    {
+                        GGML_F32_VEC sum[GGML_F32_ARR] = { GGML_F32_VEC_ZERO };
+                        GGML_F32_VEC ax[GGML_F32_ARR];
+                        GGML_F32_VEC ay[GGML_F32_ARR];
+                        for (int64_t j = 0; j < head_size; j += GGML_F32_STEP) {
+                            for (int64_t kk = 0; kk < GGML_F32_ARR; kk++) {
+                                ax[kk] = GGML_F32_VEC_LOAD(&a[t_h_offset + j + kk * GGML_F32_EPR]);
+                                ay[kk] = GGML_F32_VEC_LOAD(&state_prev[h_2d_i_offset + j + kk * GGML_F32_EPR]);
+                                sum[kk] = GGML_F32_VEC_FMA(sum[kk], ax[kk], ay[kk]);
+                            }
+                        }
+                        GGML_F32_VEC_REDUCE(sa, sum);
+                    }
+
+                    GGML_F32_VEC sa_vec = GGML_F32_VEC_SET1(sa);
+
+                    int64_t j = 0;
+                    GGML_F32_VEC result_vec[GGML_F32_ARR] = { GGML_F32_VEC_ZERO };
+                    for (; j < head_size; j += GGML_F32_STEP) {
+                        for (int64_t kk = 0; kk < GGML_F32_ARR; kk++) {
+                            int64_t t_h_j_offset = t_h_offset + j + kk * GGML_F32_EPR;
+                            int64_t h_2d_i_j_offset = h_2d_i_offset + j + kk * GGML_F32_EPR;
+
+                            GGML_F32_VEC r_vec = GGML_F32_VEC_LOAD(&r[t_h_j_offset]);
+                            GGML_F32_VEC w_vec = GGML_F32_VEC_LOAD(&w[t_h_j_offset]);
+                            GGML_F32_VEC k_vec = GGML_F32_VEC_LOAD(&k[t_h_j_offset]);
+                            GGML_F32_VEC b_vec = GGML_F32_VEC_LOAD(&b[t_h_j_offset]);
+
+                            k_vec = GGML_F32_VEC_MUL(v_vec, k_vec);
+
+                            GGML_F32_VEC state_vec = GGML_F32_VEC_LOAD(&state_prev[h_2d_i_j_offset]);
+                            // kv + s * decay + sa * b
+                            state_vec = GGML_F32_VEC_FMA(k_vec, state_vec, w_vec);
+                            state_vec = GGML_F32_VEC_FMA(state_vec, sa_vec, b_vec);
+                            GGML_F32_VEC_STORE(&state_cur[h_2d_i_j_offset], state_vec);
+
+                            result_vec[kk] = GGML_F32_VEC_FMA(result_vec[kk], state_vec, r_vec);
+                        }
+                    }
+                    GGML_F32_VEC_REDUCE(dst_data[t_h_i_offset], result_vec);
+
+                    // There shouldn't be left-overs though.
+                    for (; j < head_size; j++) {
+                        int64_t t_h_j_offset = t_h_offset + j;
+                        int64_t h_2d_i_j_offset = h_2d_i_offset + j;
+
+                        float r_val = r[t_h_j_offset];
+                        float w_val = w[t_h_j_offset];
+                        float k_val = k[t_h_j_offset];
+                        float b_val = b[t_h_j_offset];
+                        float kv_val = v[t_h_i_offset] * k_val;
+
+                        float prev_state_val = state_prev[h_2d_i_j_offset];
+                        state_cur[h_2d_i_j_offset] = prev_state_val * w_val + kv_val + sa * b_val;
+                        dst_data[t_h_i_offset] += state_cur[h_2d_i_j_offset] * r_val;
+                    }
+                }
+            }
+        }
+    #else
+        for (int64_t t = 0; t < T; t++) {
+            int64_t t_offset = t * t_stride;
+            int64_t state_offset = head_size * C * (t / (T / n_seqs));
+            float * state_cur = state + state_offset;
+            float * state_prev = t % (T / n_seqs) ? state_cur : (float*)dst->src[6]->data + state_offset;
+
+            for (int64_t h = h_start; h < h_end; h++) {
+                int64_t h_offset = h * h_stride;
+                int64_t t_h_offset = t_offset + h_offset;
+                int64_t h_2d_offset = h * h_stride_2d;
+
+                for (int64_t i = 0; i < head_size; i++) {
+                    int64_t t_h_i_offset = t_h_offset + i;
+                    int64_t h_2d_i_offset = h_2d_offset + i * h_stride;
+
+                    float v_val = v[t_h_i_offset];
+
+                    float sa = 0, result = 0;
+                    for (int64_t j = 0; j < head_size; j++) {
+                        sa += a[t_h_offset + j] * state_prev[h_2d_i_offset + j];
+                    }
+
+                    for (int64_t j = 0; j < head_size; j++) {
+                        int64_t t_h_j_offset = t_h_offset + j;
+                        int64_t h_2d_i_j_offset = h_2d_i_offset + j;
+
+                        float r_val = r[t_h_j_offset];
+                        float w_val = w[t_h_j_offset];
+                        float k_val = k[t_h_j_offset];
+                        float b_val = b[t_h_j_offset];
+                        float kv_val = v_val * k_val;
+                        float prev_state_val = state_prev[h_2d_i_j_offset];
+                        state_cur[h_2d_i_j_offset] = prev_state_val * w_val + kv_val + sa * b_val;
+                        result += state_cur[h_2d_i_j_offset] * r_val;
+                    }
+                    dst_data[t_h_i_offset] = result;
+                }
+            }
+        }
+    #endif
+}
+
+
+static void ggml_compute_forward_rwkv_wkv7(
+        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_rwkv_wkv7_f32(params, dst);
+            } break;
+        default:
+            {
+                GGML_ABORT("fatal error");
+            }
+    }
+}
+
 // ggml_compute_forward_map_unary
 
 static void ggml_compute_forward_map_unary_f32(
@@ -14067,7 +14310,9 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
     }
 
     // extra_buffer op?
-    if (ggml_cpu_extra_compute_forward(params, tensor)) return;
+    if (ggml_cpu_extra_compute_forward(params, tensor)) {
+        return;
+    }
 
     switch (tensor->op) {
         case GGML_OP_DUP:
@@ -14170,6 +14415,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
             {
                 ggml_compute_forward_group_norm(params, tensor);
             } break;
+        case GGML_OP_L2_NORM:
+            {
+                ggml_compute_forward_l2_norm(params, tensor);
+            } break;
         case GGML_OP_MUL_MAT:
             {
                 ggml_compute_forward_mul_mat(params, tensor);
@@ -14357,6 +14606,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
             {
                 ggml_compute_forward_gla(params, tensor);
             } break;
+        case GGML_OP_RWKV_WKV7:
+            {
+                ggml_compute_forward_rwkv_wkv7(params, tensor);
+            } break;
         case GGML_OP_MAP_UNARY:
             {
                 ggml_unary_op_f32_t fun;
@@ -14582,6 +14835,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
         case GGML_OP_NORM:
         case GGML_OP_RMS_NORM:
         case GGML_OP_RMS_NORM_BACK:
+        case GGML_OP_L2_NORM:
         case GGML_OP_GROUP_NORM:
         case GGML_OP_CONCAT:
         case GGML_OP_MUL_MAT:
@@ -14648,14 +14902,15 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
         case GGML_OP_FLASH_ATTN_BACK:
         case GGML_OP_SSM_CONV:
         case GGML_OP_SSM_SCAN:
+        case GGML_OP_RWKV_WKV6:
+        case GGML_OP_GATED_LINEAR_ATTN:
+        case GGML_OP_RWKV_WKV7:
             {
                 n_tasks = n_threads;
             } break;
         case GGML_OP_WIN_PART:
         case GGML_OP_WIN_UNPART:
         case GGML_OP_GET_REL_POS:
-        case GGML_OP_RWKV_WKV6:
-        case GGML_OP_GATED_LINEAR_ATTN:
         case GGML_OP_MAP_UNARY:
         case GGML_OP_MAP_BINARY:
         case GGML_OP_MAP_CUSTOM1_F32:

package/src/llama.cpp/ggml/src/ggml-cuda/vendors/hip.h

@@ -112,7 +112,7 @@
 #define cudaGraphExecDestroy hipGraphExecDestroy
 #define cudaGraphLaunch hipGraphLaunch
 #define cudaErrorGraphExecUpdateFailure hipErrorGraphExecUpdateFailure
-#define cudaGraphExecUpdateResultInfo hipGraphExecUpdateResult
+#define cudaGraphExecUpdateResult hipGraphExecUpdateResult
 #define cudaGraphNodeType hipGraphNodeType
 #define cudaGraphNodeTypeKernel hipGraphNodeTypeKernel
 #define cudaGraphInstantiate hipGraphInstantiate
@@ -129,6 +129,7 @@
 #define cudaGraph_t hipGraph_t
 #define cudaStream_t hipStream_t
 #define cudaSuccess hipSuccess
+#define cudaOccupancyMaxActiveBlocksPerMultiprocessor hipOccupancyMaxActiveBlocksPerMultiprocessor
 #define __trap() do { abort(); __builtin_unreachable(); } while(0)
 #define CUBLAS_STATUS_SUCCESS HIPBLAS_STATUS_SUCCESS
 #define CUBLAS_STATUS_NOT_INITIALIZED HIPBLAS_STATUS_NOT_INITIALIZED

package/src/llama.cpp/ggml/src/ggml-cuda/vendors/musa.h

@@ -119,7 +119,7 @@
 #define cudaGraphExecDestroy musaGraphExecDestroy
 #define cudaGraphExec_t musaGraphExec_t
 #define cudaGraphExecUpdate musaGraphExecUpdate
-#define cudaGraphExecUpdateResultInfo musaGraphExecUpdateResult
+#define cudaGraphExecUpdateResult musaGraphExecUpdateResult
 #define cudaGraphGetNodes musaGraphGetNodes
 #define cudaGraphInstantiate musaGraphInstantiate
 #define cudaGraphKernelNodeGetParams musaGraphKernelNodeGetParams
@@ -132,6 +132,8 @@
 #define cudaGraph_t musaGraph_t
 #define cudaKernelNodeParams musaKernelNodeParams
 #define cudaStreamCaptureModeRelaxed musaStreamCaptureModeRelaxed
+#define cudaStreamBeginCapture musaStreamBeginCapture
 #define cudaStreamEndCapture musaStreamEndCapture
+#define cudaOccupancyMaxActiveBlocksPerMultiprocessor musaOccupancyMaxActiveBlocksPerMultiprocessor
 
 typedef mt_bfloat16 nv_bfloat16;

package/src/llama.cpp/ggml/src/ggml-metal/ggml-metal-impl.h

@@ -285,6 +285,13 @@ typedef struct {
     float    eps;
 } ggml_metal_kargs_rms_norm;
 
+typedef struct {
+    int32_t  ne00;
+    int32_t  ne00_4;
+    uint64_t nb01;
+    float    eps;
+} ggml_metal_kargs_l2_norm;
+
 typedef struct {
     int64_t  ne00;
     int64_t  ne01;

package/src/llama.cpp/ggml/src/ggml-musa/CMakeLists.txt

@@ -67,10 +67,6 @@ if (MUSAToolkit_FOUND)
     add_compile_definitions(GGML_USE_MUSA)
     add_compile_definitions(GGML_CUDA_PEER_MAX_BATCH_SIZE=${GGML_CUDA_PEER_MAX_BATCH_SIZE})
 
-    if (GGML_CUDA_GRAPHS)
-        add_compile_definitions(GGML_CUDA_USE_GRAPHS)
-    endif()
-
     if (GGML_CUDA_FORCE_MMQ)
         add_compile_definitions(GGML_CUDA_FORCE_MMQ)
     endif()