Apply suggestion from @CISC

Co-authored-by: Sigbjørn Skjæret <[email protected]>

Author: mitmul

src/llama-model.cpp

@@ -15636,21 +15636,14 @@ struct llm_build_plamo2 : public llm_graph_context_mamba {
             const int64_t k_offset = n_embd_head_q * n_head;
             const int64_t v_offset = k_offset + n_embd_head_k * n_head_kv;
 
-            ggml_tensor * Qcur = ggml_view_2d(ctx0, qkv, n_embd_head_q * n_head, n_tokens, qkv->nb[1], q_offset * ggml_element_size(qkv));
-            ggml_tensor * Kcur = ggml_view_2d(ctx0, qkv, n_embd_head_k * n_head_kv, n_tokens, qkv->nb[1], k_offset * ggml_element_size(qkv));
-            ggml_tensor * Vcur = ggml_view_2d(ctx0, qkv, n_embd_head_v * n_head_kv, n_tokens, qkv->nb[1], v_offset * ggml_element_size(qkv));
-
-            // make tensors contiguous before reshape
-            Qcur = ggml_cont(ctx0, Qcur);
-            Kcur = ggml_cont(ctx0, Kcur);
-            Vcur = ggml_cont(ctx0, Vcur);
+            ggml_tensor * Qcur = ggml_view_3d(ctx0, qkv, n_embd_head_q, n_head, n_tokens, n_embd_head_q * sizeof(float), qkv->nb[1], q_offset * ggml_element_size(qkv));
+            ggml_tensor * Kcur = ggml_view_3d(ctx0, qkv, n_embd_head_k, n_head_kv, n_tokens, n_embd_head_k * sizeof(float), qkv->nb[1], k_offset * ggml_element_size(qkv));
+            ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, qkv, n_embd_head_v * n_head_kv, n_tokens, qkv->nb[1], v_offset * ggml_element_size(qkv)));
 
             cb(Qcur, "Qcur", il);
             cb(Kcur, "Kcur", il);
             cb(Vcur, "Vcur", il);
 
-            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head_q, n_head,    n_tokens);
-            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head_k, n_head_kv, n_tokens);
             Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head_v, n_head_kv, n_tokens);
 
             Qcur = build_norm(Qcur, model.layers[il].wq, NULL, LLM_NORM_RMS, il);