diff --git a/ggml/src/ggml-backend-meta.cpp b/ggml/src/ggml-backend-meta.cpp
--- a/ggml/src/ggml-backend-meta.cpp
+++ b/ggml/src/ggml-backend-meta.cpp
@@
         ggml_cgraph * cgraph_main = nullptr;
         int           offset      = 0; // Node offset vs. original graph
         int           offset_end  = 0; // Exclusive node offset vs. original graph
         int           fused_add   = -1;
+        int           fused_add_via = -1;
         int           fused_to    = -1;
@@
                 int i_stop = i;
                 int fused_add = -1;
+                int fused_add_via = -1;
                 int fused_to = -1;
@@
                         fused_add = i + 1;
                         fused_allreduce_add_count++;
                     }
+
+                    if (fused_add < 0 && i + 2 < cgraph->n_nodes &&
+                            next->op == GGML_OP_RESHAPE &&
+                            next->src[0] == node &&
+                            ggml_node_get_use_count(cgraph, i) == 1 &&
+                            ggml_node_get_use_count(cgraph, i + 1) == 1 &&
+                            next->type == GGML_TYPE_F32 &&
+                            node->type == GGML_TYPE_F32 &&
+                            ggml_nbytes(next) == ggml_nbytes(node) &&
+                            ggml_nbytes(node) <= 64*1024 &&
+                            next_split_state.axis == GGML_BACKEND_SPLIT_AXIS_MIRRORED) {
+                        ggml_tensor * add = cgraph->nodes[i + 2];
+                        const bool reshaped_is_src0 = add->op == GGML_OP_ADD && add->src[0] == next;
+                        const bool reshaped_is_src1 = add->op == GGML_OP_ADD && add->src[1] == next;
+                        ggml_tensor * add_residual = reshaped_is_src0 ? add->src[1] : (reshaped_is_src1 ? add->src[0] : nullptr);
+                        const ggml_backend_meta_split_state add_residual_split_state = add_residual ?
+                            ggml_backend_meta_get_split_state(add_residual, /*assume_sync =*/ false) :
+                            ggml_backend_meta_split_state{GGML_BACKEND_SPLIT_AXIS_UNKNOWN, {0}, 1};
+                        const ggml_backend_meta_split_state add_split_state =
+                            ggml_backend_meta_get_split_state(add, /*assume_sync =*/ false);
+
+                        if ((reshaped_is_src0 || reshaped_is_src1) &&
+                                add->type == GGML_TYPE_F32 &&
+                                add_residual != nullptr && add_residual->type == GGML_TYPE_F32 &&
+                                ggml_nbytes(add) == ggml_nbytes(node) &&
+                                ggml_nbytes(add_residual) == ggml_nbytes(node) &&
+                                add_residual_split_state.axis == GGML_BACKEND_SPLIT_AXIS_MIRRORED &&
+                                add_split_state.axis == GGML_BACKEND_SPLIT_AXIS_MIRRORED) {
+                            fused_add = i + 2;
+                            fused_add_via = i + 1;
+                            fused_allreduce_add_count++;
+                        }
+                    }
                 }
@@
-                    fprintf(stderr, " fused_to=%d", fused_to);
+                    fprintf(stderr, " fused_add_via=%d fused_to=%d", fused_add_via, fused_to);
@@
                     bcj.cgraphs[n_subgraphs].offset = i_start;
                     bcj.cgraphs[n_subgraphs].offset_end = i_stop + 1;
                     bcj.cgraphs[n_subgraphs].fused_add = fused_add;
+                    bcj.cgraphs[n_subgraphs].fused_add_via = fused_add_via;
                     bcj.cgraphs[n_subgraphs].fused_to = fused_to;
@@
             ggml_cgraph * cgraph_i0 = backend_ctx->backend_configs[0].cgraphs[i].cgraph_main;
             const int fused_add_i = backend_ctx->backend_configs[0].cgraphs[i].fused_add;
+            const int fused_add_via_i = backend_ctx->backend_configs[0].cgraphs[i].fused_add_via;
             const int fused_to_i = backend_ctx->backend_configs[0].cgraphs[i].fused_to;
@@
                         auto & bcj = backend_ctx->backend_configs[j];
                         ggml_tensor * add_node = bcj.nodes[fused_add_i];
+                        ggml_tensor * add_input = nodes[j];
                         if (add_node->op != GGML_OP_ADD) {
                             valid_fused_add = false;
                             break;
                         }
-                        if (add_node->src[0] == nodes[j]) {
+                        if (fused_add_via_i >= 0) {
+                            ggml_tensor * via_node = bcj.nodes[fused_add_via_i];
+                            if (via_node->op != GGML_OP_RESHAPE ||
+                                    via_node->src[0] != nodes[j] ||
+                                    ggml_nbytes(via_node) != ggml_nbytes(nodes[j])) {
+                                valid_fused_add = false;
+                                break;
+                            }
+                            add_input = via_node;
+                        }
+                        if (add_node->src[0] == add_input) {
                             residuals.push_back(add_node->src[1]);
-                        } else if (add_node->src[1] == nodes[j]) {
+                        } else if (add_node->src[1] == add_input) {
                             residuals.push_back(add_node->src[0]);
                         } else {
                             valid_fused_add = false;
@@
             }
             if (fused_add_i >= 0 && !backend_allreduce_add_success) {
+                if (fused_add_via_i >= 0) {
+                    const ggml_status status = compute_aux_node(fused_add_via_i);
+                    if (status != GGML_STATUS_SUCCESS) {
+                        return status;
+                    }
+                }
                 const ggml_status status = compute_aux_node(fused_add_i);
                 if (status != GGML_STATUS_SUCCESS) {
                     return status;