fix

orca-zhang · orca-zhang · commit 2dd8afc4ed01 · 2025-02-27T18:03:53.000+08:00
diff --git a/src/llama.cpp b/src/llama.cpp
@@ -608,7 +608,7 @@ static struct ggml_tensor * llm_build_kqv(
             cur = ggml_cont(ctx, ggml_view_3d(ctx, cur, n_embd_head_v, n_head, n_tokens,
                             ggml_row_size(cur->type, n_embd_head_v_out),
                             ggml_row_size(cur->type, n_embd_head_v_out * n_head),
-                            0));
+                            ggml_row_size(cur->type, n_embd_head_v_out * n_head * (n_embd_head_k - n_embd_head_v))));
         }
 
         cur = ggml_reshape_2d(ctx, cur, n_embd_head_v*n_head, n_tokens);

Original file line number	Diff line number	Diff line change
`@@ -608,7 +608,7 @@ static struct ggml_tensor * llm_build_kqv(`
`608`	`608`	`cur = ggml_cont(ctx, ggml_view_3d(ctx, cur, n_embd_head_v, n_head, n_tokens,`
`609`	`609`	`ggml_row_size(cur->type, n_embd_head_v_out),`
`610`	`610`	`ggml_row_size(cur->type, n_embd_head_v_out * n_head),`
`611`		`- 0));`
	`611`	`+ ggml_row_size(cur->type, n_embd_head_v_out * n_head * (n_embd_head_k - n_embd_head_v))));`
`612`	`612`	`}`
`613`	`613`
`614`	`614`	`cur = ggml_reshape_2d(ctx, cur, n_embd_head_v*n_head, n_tokens);`