diff --git a/src/graphs/build_deepseek2.cpp b/src/graphs/build_deepseek2.cpp
index 9d9bc558..0c4a2f70 100644
--- a/src/graphs/build_deepseek2.cpp
+++ b/src/graphs/build_deepseek2.cpp
@@ -157,10 +157,8 @@ ggml_tensor * llm_build_context::build_deepseek2_tp_attention(
             // Per-rank wk_b/wv_b slices already exist from distribute_mla_tensors:
             //   wk_b_local_pp: [n_embd_head_qk_nope, kv_lora_rank, n_head_local]
             //   wv_b_local_pp: [kv_lora_rank, n_embd_head_v,    n_head_local]
-            auto wk_b_pp_split_raw = (const ggml_split_tensor_t *)model.layers[il].wk_b->extra;
             auto wv_b_pp_split_raw = (const ggml_split_tensor_t *)model.layers[il].wv_b->extra;
-            GGML_ASSERT(wk_b_pp_split_raw && wv_b_pp_split_raw);
-            ggml_tensor * wk_b_local_pp = wk_b_pp_split_raw->splits[id];
+            GGML_ASSERT(wv_b_pp_split_raw);
             ggml_tensor * wv_b_local_pp = wv_b_pp_split_raw->splits[id];
 
             ggml_tensor * kv_cache_nope = ggml_view_2d(ctx0, cache_local,
diff --git a/src/llama-load-tensors.cpp b/src/llama-load-tensors.cpp
index b27884d8..1fb84421 100644
--- a/src/llama-load-tensors.cpp
+++ b/src/llama-load-tensors.cpp
@@ -3812,8 +3812,6 @@ static void distribute_mla_tensors_for_split_mode_graph(
     const int n_head        = hparams.n_head(il);
     const int n_embd_head_k = hparams.n_embd_head_k(il);
     const int n_embd_head_v = hparams.n_embd_head_v(il);
-    const int qk_rope       = hparams.n_rot;
-    const int qk_nope       = n_embd_head_k - qk_rope;
 
     // granularity=4: keeps wo row blocks K-quant-aligned (% 256) and gqa_ratio % 4 == 0 for FA-MMA.
     auto split_heads = create_split(n_head, 4, cur_splits, mem_used);