From aa6738a8f400e31c1939fb26644e2b88bf03c986 Mon Sep 17 00:00:00 2001 From: Iwan Kawrakow Date: Wed, 21 May 2025 17:49:39 +0300 Subject: [PATCH] Just in case, set computed wkv_b vew source to null --- src/llama.cpp | 1 + 1 file changed, 1 insertion(+) diff --git a/src/llama.cpp b/src/llama.cpp index b7534420..12467a97 100644 --- a/src/llama.cpp +++ b/src/llama.cpp @@ -7025,6 +7025,7 @@ static void llm_prepare_mla(llama_model & model, int mla) { auto name = std::string{"blk."} + std::to_string(il) + ".attn_kv_b.weight"; l.computed_wkv_b = std::make_unique(*wkv_b); + l.computed_wkv_b->view_src = nullptr; l.computed_wkv_b->buffer = ggml_backend_buft_alloc_buffer(ggml_backend_buffer_get_type(l.wk_b->buffer), ggml_nbytes(wkv_b)); l.computed_wkv_b->data = ggml_backend_buffer_get_base(l.computed_wkv_b->buffer); l.computed_wkv_b->op = GGML_OP_NONE; // we absolutely need to do this, else the backend will attempt to find the parents