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

Update llama.cpp

Browse source
This commit is contained in:
CausalLM 2023-12-02 01:09:17 +08:00 • committed by GitHub
parent e192572d21
commit b1efaed381
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
1 changed files with 19 additions and 19 deletions
Download patch file Download diff file Expand all files Collapse all files

38
llama.cpp
View file

@ -2784,29 +2784,29 @@ 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.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.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.wo = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, backend_split);
try { try {
layer.bq = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_Q, "bias", i), {n_embd}, backend); layer.bq = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_Q, "bias", i), {n_embd}, backend);
} catch (const std::runtime_error& e) { } catch (const std::runtime_error& e) {
if (std::string(e.what()).find("not found") != std::string::npos) layer.bq = NULL; else throw; if (std::string(e.what()).find("not found") != std::string::npos) layer.bq = NULL; else throw;
} }
try { try {
layer.bk = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_K, "bias", i), {n_embd_gqa}, backend); layer.bk = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_K, "bias", i), {n_embd_gqa}, backend);
} catch (const std::runtime_error& e) { } catch (const std::runtime_error& e) {
if (std::string(e.what()).find("not found") != std::string::npos) layer.bk = NULL; else throw; if (std::string(e.what()).find("not found") != std::string::npos) layer.bk = NULL; else throw;
} }
try { try {
layer.bv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_V, "bias", i), {n_embd_gqa}, backend); layer.bv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_V, "bias", i), {n_embd_gqa}, backend);
} catch (const std::runtime_error& e) { } catch (const std::runtime_error& e) {
if (std::string(e.what()).find("not found") != std::string::npos) layer.bv = NULL; else throw; if (std::string(e.what()).find("not found") != std::string::npos) layer.bv = NULL; else throw;
} }
try { try {
layer.bo = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, backend); layer.bo = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, backend);
} catch (const std::runtime_error& e) { } catch (const std::runtime_error& e) {
if (std::string(e.what()).find("not found") != std::string::npos) layer.bo = NULL; else throw; if (std::string(e.what()).find("not found") != std::string::npos) layer.bo = NULL; else throw;
} }
layer.ffn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, backend); layer.ffn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, backend);
@ -3941,7 +3941,7 @@ struct llm_build_context {
Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
cb(Vcur, "Vcur", il); cb(Vcur, "Vcur", il);
} }
Qcur = ggml_rope_custom( Qcur = ggml_rope_custom(
ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, 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, n_embd_head, 0, 0, n_orig_ctx, freq_base, freq_scale,
@ -3959,7 +3959,7 @@ struct llm_build_context {
llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il); 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, cur = llm_build_kqv(ctx0, hparams, kv_self,
model.layers[il].wo, model.layers[il].bo, model.layers[il].wo, model.layers[il].bo,
Qcur, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, cb, il); Qcur, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, cb, il);
cb(cur, "kqv_out", il); cb(cur, "kqv_out", il);
} }