only support n_seq=1 for now

Author: am17an

src/llama-graph.cpp

@@ -824,15 +824,7 @@ bool llm_graph_input_dsv4::can_reuse(const llm_graph_params & params) {
 
     bool res = true;
 
-    if (inp_raw->self_k_idxs && inp_raw->self_k_idxs->buffer) {
-        res &= inp_raw->self_k_idxs->ne[0] == params.ubatch.n_tokens;
-        res &= can_reuse_kq_mask(inp_raw->self_kq_mask, mctx->get_raw()->get_base(), params.ubatch, params.cparams);
-    }
-
-    if (inp_raw->self_k_idxs_swa && inp_raw->self_k_idxs_swa->buffer) {
-        res &= inp_raw->self_k_idxs_swa->ne[0] == params.ubatch.n_tokens;
-        res &= can_reuse_kq_mask(inp_raw->self_kq_mask_swa, mctx->get_raw()->get_swa(), params.ubatch, params.cparams);
-    }
+    res &= inp_raw->can_reuse(params);
 
     res &= dsv4_can_reuse_comp_input(inp_csa, mctx->get_csa_plan(params.ubatch), params.ubatch.n_tokens);
     res &= dsv4_can_reuse_comp_input(inp_hca, mctx->get_hca_plan(params.ubatch), params.ubatch.n_tokens);

src/llama-kv-cache-dsv4.cpp

@@ -722,6 +722,10 @@ llama_kv_cache_dsv4::llama_kv_cache_dsv4(
     hparams_hca(model.hparams),
     hparams_lid(model.hparams) {
 
+    if (n_seq_max > 1) {
+        throw std::runtime_error("DSV4 KV cache currently supports only single-sequence decoding");
+    }
+
     const layer_filter_cb filter_raw = [&](int32_t il) {
         if (filter && !filter(il)) {
             return false;