kv-cache : support non-FA case

ggml-ci

Author: ggerganov

src/llama-kv-cache-unified.cpp

@@ -777,18 +777,27 @@ ggml_tensor * llama_kv_cache_unified::cpy_v(ggml_context * ctx, ggml_tensor * v_
 
     if (!v_trans) {
         if (kv_idxs) {
-            return ggml_set_rows(ctx, v, ggml_reshape_2d(ctx, v_cur, v->ne[0], n_tokens), kv_idxs);
+            return ggml_set_rows(ctx, v, v_cur, kv_idxs);
         }
 
         v_view = ggml_view_1d(ctx, v,
                 n_tokens*hparams.n_embd_v_gqa(il),
                 ggml_row_size(v->type, hparams.n_embd_v_gqa(il))*head_cur);
     } else {
+        // note: the V cache is transposed when not using flash attention
         if (kv_idxs) {
-            GGML_ABORT("TODO: implement kv_idxs for transposed V cache -- for now use flash attention");
+            // the row becomes a single element and we repeat the KV indices d_head times
+            // TODO: this seems not very optimal - can we do something better?
+            v_view = ggml_view_3d(ctx, v, 1, v->ne[1], hparams.n_embd_v_gqa(il),
+                    ggml_element_size(v),
+                    (v->ne[1])*ggml_element_size(v),
+                    0);
+
+            v_cur = ggml_reshape_3d(ctx, ggml_cont(ctx, ggml_transpose(ctx, v_cur)), 1, n_tokens, hparams.n_embd_v_gqa(il));
+
+            return ggml_set_rows(ctx, v_view, v_cur, ggml_repeat_4d(ctx, kv_idxs, n_tokens, hparams.n_embd_v_gqa(il), 1, 1));
         }
 
-        // note: the V cache is transposed when not using flash attention
         v_view = ggml_view_2d(ctx, v, n_tokens, hparams.n_embd_v_gqa(il),
                 (v->ne[1])*ggml_element_size(v),
                 (head_cur)*ggml_element_size(v));