Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Support attention_bias on LLaMA architecture #4283

Merged
merged 3 commits into from
Dec 1, 2023
Merged

Support attention_bias on LLaMA architecture #4283

Changes from 1 commit
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
c48679a
Support attention_bias on LLaMA architecture
CausalLM Dec 1, 2023
e192572
check existence of qkvo bias while loading llama models
CausalLM Dec 1, 2023
b1efaed
Update llama.cpp
CausalLM Dec 1, 2023
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
29 changes: 25 additions & 4 deletions llama.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1248,6 +1248,9 @@ struct llama_layer {
struct ggml_tensor * wqkv;

// attention bias
struct ggml_tensor * bq;
struct ggml_tensor * bk;
struct ggml_tensor * bv;
struct ggml_tensor * bo;
struct ggml_tensor * bqkv;

Expand Down Expand Up @@ -2781,6 +2784,11 @@ static void llm_load_tensors(
layer.wk = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_gqa}, backend_split);
layer.wv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_gqa}, backend_split);
layer.wo = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, backend_split);

layer.bq = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_Q, "bias", i), {n_embd}, backend);
layer.bk = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_K, "bias", i), {n_embd_gqa}, backend);
layer.bv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_V, "bias", i), {n_embd_gqa}, backend);
layer.bo = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, backend);
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I think you need to crease these tensors only for the models that have them, otherwise this breaks llama models.

Copy link
Owner

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Correct

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

fixed, tested on llama2, cuda & cpu


layer.ffn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, backend);

Expand All @@ -2791,8 +2799,9 @@ static void llm_load_tensors(
if (backend == GGML_BACKEND_GPU) {
vram_weights +=
ggml_nbytes(layer.attn_norm) + ggml_nbytes(layer.wq) + ggml_nbytes(layer.wk) +
ggml_nbytes(layer.wv) + ggml_nbytes(layer.wo) + ggml_nbytes(layer.ffn_norm) +
ggml_nbytes(layer.ffn_gate) + ggml_nbytes(layer.ffn_down) + ggml_nbytes(layer.ffn_up);
ggml_nbytes(layer.wv) + ggml_nbytes(layer.wo) + ggml_nbytes(layer.bq) +
ggml_nbytes(layer.bk) + ggml_nbytes(layer.bv) + ggml_nbytes(layer.bo) +
ggml_nbytes(layer.ffn_norm) + ggml_nbytes(layer.ffn_gate) + ggml_nbytes(layer.ffn_down) + ggml_nbytes(layer.ffn_up);
}
}
} break;
Expand Down Expand Up @@ -3891,13 +3900,25 @@ struct llm_build_context {
// compute Q and K and RoPE them
struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].wq, cur);
cb(Qcur, "Qcur", il);
if (model.layers[il].bq) {
Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
cb(Qcur, "Qcur", il);
}

struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].wk, cur);
cb(Kcur, "Kcur", il);
if (model.layers[il].bk) {
Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
cb(Kcur, "Kcur", il);
}

struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur);
cb(Vcur, "Vcur", il);

if (model.layers[il].bv) {
Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
cb(Vcur, "Vcur", il);
}

Qcur = ggml_rope_custom(
ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos,
n_embd_head, 0, 0, n_orig_ctx, freq_base, freq_scale,
Expand All @@ -3915,7 +3936,7 @@ struct llm_build_context {
llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il);

cur = llm_build_kqv(ctx0, hparams, kv_self,
model.layers[il].wo, NULL,
model.layers[il].wo, model.layers[il].bo,
Qcur, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, cb, il);
cb(cur, "kqv_out", il);
}
Expand Down