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

Merge branch 'bert' of github.com:iamlemec/llama.cpp into bert

Browse source
This commit is contained in:
Douglas Hanley 2024-02-08 23:05:45 -06:00
commit 3a1895d786
1 changed files with 6 additions and 3 deletions
Download patch file Download diff file Expand all files Collapse all files

9
llama.cpp
View file

@ -3893,9 +3893,9 @@ static bool llm_load_tensors(
} break; } break;
case LLM_ARCH_BLOOM: case LLM_ARCH_BLOOM:
{ {
model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
model.tok_norm = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight"), {n_embd}); model.tok_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight"), {n_embd});
model.tok_norm_b = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD_NORM, "bias"), {n_embd}); model.tok_norm_b = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD_NORM, "bias"), {n_embd});
// output // output
{ {
@ -5757,6 +5757,7 @@ struct llm_build_context {
// embed layer norm // embed layer norm
inpL = llm_build_norm(ctx0, inpL, hparams, model.tok_norm, model.tok_norm_b, LLM_NORM, cb, -1); inpL = llm_build_norm(ctx0, inpL, hparams, model.tok_norm, model.tok_norm_b, LLM_NORM, cb, -1);
cb(inpL, "inp_norm", -1);
// KQ_mask (mask for 1 head, it will be broadcasted to all heads) // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
struct ggml_tensor * KQ_mask = ggml_view_2d(ctx0, lctx.inp_KQ_mask, n_kv, n_tokens, n_kv*ggml_type_size(lctx.inp_KQ_mask->type), 0); struct ggml_tensor * KQ_mask = ggml_view_2d(ctx0, lctx.inp_KQ_mask, n_kv, n_tokens, n_kv*ggml_type_size(lctx.inp_KQ_mask->type), 0);
@ -5793,6 +5794,7 @@ struct llm_build_context {
cur = llm_build_norm(ctx0, cur, hparams, model.layers[il].attn_norm, model.layers[il].attn_norm_b, LLM_NORM, cb, il); cur = llm_build_norm(ctx0, cur, hparams, model.layers[il].attn_norm, model.layers[il].attn_norm_b, LLM_NORM, cb, il);
struct ggml_tensor * ffn_inp = cur; struct ggml_tensor * ffn_inp = cur;
cb(ffn_inp, "ffn_inp", il);
// feed-forward network // feed-forward network
cur = llm_build_ffn(ctx0, cur, cur = llm_build_ffn(ctx0, cur,
@ -5801,6 +5803,7 @@ struct llm_build_context {
model.layers[il].ffn_down, model.layers[il].ffn_down_b, model.layers[il].ffn_down, model.layers[il].ffn_down_b,
NULL, NULL,
LLM_FFN_GELU, LLM_FFN_SEQ, cb, il); LLM_FFN_GELU, LLM_FFN_SEQ, cb, il);
cb(cur, "ffn_out", il);
// attentions bypass the intermediate layer // attentions bypass the intermediate layer
cur = ggml_add(ctx0, cur, ffn_inp); cur = ggml_add(ctx0, cur, ffn_inp);