diff --git a/src/llama-model.cpp b/src/llama-model.cpp
index ad7a6ccb2f..a7b39262ef 100644
--- a/src/llama-model.cpp
+++ b/src/llama-model.cpp
@@ -663,7 +663,7 @@ struct ggml_backend_meta_split_state llama_meta_device_get_split_state(const str
             if (std::regex_match(tensor_name, pattern_attn_sinks)) {
                 GGML_ASSERT(segments.size() == 1);
                 if (ud->model->arch == LLM_ARCH_DEEPSEEK4) {
-                    return {1};
+                    return {hparams.n_head(il) / hparams.dsv4_o_group_count};
                 }
                 return {std::lcm(n_embd_q, blck_size_perf)/n_embd_q * n_gqa};
             }
@@ -671,7 +671,9 @@ struct ggml_backend_meta_split_state llama_meta_device_get_split_state(const str
             if (ud->model->arch == LLM_ARCH_DEEPSEEK4) {
                 if (std::regex_match(tensor_name, pattern_attn_q_b_weight)) {
                     GGML_ASSERT(segments.size() == 1);
-                    return {hparams.n_embd_head_k(il)};
+                    // the grouped output projection requires each device to hold whole groups of heads
+                    const int64_t n_head_group = hparams.n_head(il) / hparams.dsv4_o_group_count;
+                    return {n_head_group * hparams.n_embd_head_k(il)};
                 }
                 if (std::regex_match(tensor_name, pattern_attn_out_a_weight) ||
                         std::regex_match(tensor_name, pattern_attn_out_b_weight)) {