From 215a69b5af958d82dcf09042aeed999801646157 Mon Sep 17 00:00:00 2001 From: James Ravenscroft Date: Sat, 26 Aug 2023 16:16:01 +0100 Subject: [PATCH] update clblast code in gpt-j model --- src/gptj.cpp | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-) diff --git a/src/gptj.cpp b/src/gptj.cpp index 8aafa7d..dba98a2 100644 --- a/src/gptj.cpp +++ b/src/gptj.cpp @@ -584,7 +584,9 @@ bool GPTJModel::load_model(std::string fname) { layer.c_mlp_proj_w->backend = GGML_BACKEND_GPU; #if defined(GGML_USE_CLBLAST) - ggml_cl_transform_tensor(layer.c_attn_attn_w->data,layer.c_attn_attn_w); vram_total += ggml_nbytes(layer.c_attn_attn_w); + ggml_cl_transform_tensor(layer.c_attn_q_proj_w->data,layer.c_attn_q_proj_w); vram_total += ggml_nbytes(layer.c_attn_q_proj_w); + ggml_cl_transform_tensor(layer.c_attn_k_proj_w->data,layer.c_attn_k_proj_w); vram_total += ggml_nbytes(layer.c_attn_k_proj_w); + ggml_cl_transform_tensor(layer.c_attn_v_proj_w->data,layer.c_attn_v_proj_w); vram_total += ggml_nbytes(layer.c_attn_v_proj_w); ggml_cl_transform_tensor(layer.c_attn_proj_w->data,layer.c_attn_proj_w); vram_total += ggml_nbytes(layer.c_attn_proj_w); ggml_cl_transform_tensor(layer.c_mlp_fc_w->data,layer.c_mlp_fc_w); vram_total += ggml_nbytes(layer.c_mlp_fc_w); ggml_cl_transform_tensor(layer.c_mlp_proj_w->data,layer.c_mlp_proj_w); vram_total += ggml_nbytes(layer.c_mlp_proj_w);