vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

hard-code token_type = 0

Browse source
This commit is contained in:
Douglas Hanley 2024-02-08 23:03:33 -06:00
parent 68758083d6
commit d080bebcc6
1 changed files with 2 additions and 3 deletions
Download patch file Download diff file Expand all files Collapse all files

5
llama.cpp
View file

@ -5746,14 +5746,13 @@ struct llm_build_context {
struct ggml_tensor * inpL; struct ggml_tensor * inpL;
// get input vectors with right size // get input vectors with right size
struct ggml_tensor * inp_type = ggml_view_1d(ctx0, lctx.inp_type, n_tokens, 0);
struct ggml_tensor * inp_pos = ggml_view_1d(ctx0, lctx.inp_pos, n_tokens, 0); struct ggml_tensor * inp_pos = ggml_view_1d(ctx0, lctx.inp_pos, n_tokens, 0);
struct ggml_tensor * inp_sum = ggml_view_1d(ctx0, lctx.inp_sum, n_tokens, 0); struct ggml_tensor * inp_sum = ggml_view_1d(ctx0, lctx.inp_sum, n_tokens, 0);
// construct input embeddings (token, type, position) // construct input embeddings (token, type, position)
inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, lctx.inp_tokens, lctx.inp_embd, cb); inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, lctx.inp_tokens, lctx.inp_embd, cb);
inpL = ggml_add(ctx0, ggml_get_rows(ctx0, model.type_embd, inp_type), inpL); inpL = ggml_add(ctx0, inpL, ggml_view_1d(ctx0, model.type_embd, n_embd, 0)); // since type = 0
inpL = ggml_add(ctx0, ggml_get_rows(ctx0, model.pos_embd, inp_pos), inpL); inpL = ggml_add(ctx0, inpL, ggml_get_rows(ctx0, model.pos_embd, inp_pos));
cb(inpL, "inp_embd", -1); cb(inpL, "inp_embd", -1);
// embed layer norm // embed layer norm