model: mistral le chonk (#31376)

* llama: extend DSA and sdpark to new model

* fix ci

src/llama-graph.cpp

@@ -3429,13 +3429,20 @@ ggml_tensor * llm_graph_context::build_attn(
     // by doing so, the number of splits in the graph is reduced
     // expand k later to enable rope fusion which directly writes into k-v cache
     ggml_build_forward_expand(gf, q_cur);
-    ggml_build_forward_expand(gf, v_cur);
-    ggml_build_forward_expand(gf, k_cur);
+
+    if (v_cur) {
+        ggml_build_forward_expand(gf, v_cur);
+    }
+
+    if (k_cur) {
+        ggml_build_forward_expand(gf, k_cur);
+    }
 
     const auto * mctx_cur = inp->mctx->get_mla();
 
     // store to KV cache
-    {
+    // note: k_cur is null for layers that reuse the KV cache of another layer
+    if (k_cur) {
         const auto & k_idxs = inp->get_k_idxs_mla();
 
         ggml_build_forward_expand(gf, mctx_cur->cpy_k(ctx0, k_cur, k_idxs, il));
@@ -3481,7 +3488,10 @@ ggml_tensor * llm_graph_context::build_attn(
 
     ggml_tensor * q = q_cur;
     ggml_tensor * k = mctx_cur->get_k(ctx0, il);
-    ggml_tensor * v = ggml_view_4d(ctx0, k, v_cur->ne[0], k->ne[1], k->ne[2], k->ne[3], k->nb[1], k->nb[2], k->nb[3], 0);
+
+    // MLA: V is a view of the first n_embd_v elements of K
+    const int64_t n_embd_v = v_cur ? v_cur->ne[0] : v_mla->ne[0];
+    ggml_tensor * v = ggml_view_4d(ctx0, k, n_embd_v, k->ne[1], k->ne[2], k->ne[3], k->nb[1], k->nb[2], k->nb[3], 0);
 
     ggml_tensor * cur = build_attn_mha(q, k, v, kq_b, kq_mask_top_k, sinks, v_mla, top_k->ne[0], kq_scale, il);
     cb(cur, "kqv_out", il);