From 1eee98f01fca721a889defac3d38e9ada7abb617 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Stanis=C5=82aw=20Szymczyk?= Date: Mon, 27 Jan 2025 09:32:25 +0100 Subject: [PATCH] llama : removed unnecessary code in DeepSeek V2 implementation --- src/llama.cpp | 3 --- 1 file changed, 3 deletions(-) diff --git a/src/llama.cpp b/src/llama.cpp index 3df989692..a4c78240b 100644 --- a/src/llama.cpp +++ b/src/llama.cpp @@ -6544,9 +6544,6 @@ struct llm_build_context { struct ggml_tensor * q_pe_perm = ggml_permute(ctx0, q_pe, 0, 3, 2, 1); cb(q_pe_perm, "q_pe_perm", il); - struct ggml_tensor * kr_cache_perm = ggml_permute(ctx0, kr_cache, 0, 2, 3, 1); - cb(kr_cache_perm, "kr_cache_perm", il); - struct ggml_tensor * kq_pe = ggml_mul_mat(ctx0, kr_cache, q_pe); cb(kq_pe, "kq_pe", il);