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

rem tabs

Browse source
This commit is contained in:
nopperl 2024-07-17 17:18:09 +02:00
parent 758612a984
commit 90766e15e2
1 changed files with 20 additions and 18 deletions
Download patch file Download diff file Expand all files Collapse all files

38
src/llama.cpp
View file

@ -5244,7 +5244,7 @@ static void llm_load_hparams(
case LLM_ARCH_CHAMELEON: case LLM_ARCH_CHAMELEON:
{ {
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
hparams.f_norm_eps = 1e-5; // eps for qk-norm, torch default hparams.f_norm_eps = 1e-5; // eps for qk-norm, torch default
ml.get_key(LLM_KV_SWIN_NORM, hparams.swin_norm); ml.get_key(LLM_KV_SWIN_NORM, hparams.swin_norm);
switch (hparams.n_layer) { switch (hparams.n_layer) {
@ -13718,11 +13718,11 @@ struct llm_build_context {
struct ggml_tensor * inpSA = inpL; struct ggml_tensor * inpSA = inpL;
// norm // norm
if (!hparams.swin_norm) { if (!hparams.swin_norm) {
cur = llm_build_norm(ctx0, inpL, hparams, cur = llm_build_norm(ctx0, inpL, hparams,
model.layers[il].attn_norm, NULL, model.layers[il].attn_norm, NULL,
LLM_NORM_RMS, cb, il); LLM_NORM_RMS, cb, il);
} }
cb(cur, "attn_norm", il); cb(cur, "attn_norm", il);
// self-attention // self-attention
@ -13780,11 +13780,11 @@ struct llm_build_context {
model.layers[il].wo, nullptr, model.layers[il].wo, nullptr,
Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
if (hparams.swin_norm) { if (hparams.swin_norm) {
cur = llm_build_norm(ctx0, cur, hparams, cur = llm_build_norm(ctx0, cur, hparams,
model.layers[il].attn_norm, NULL, model.layers[il].attn_norm, NULL,
LLM_NORM_RMS, cb, il); LLM_NORM_RMS, cb, il);
} }
} }
if (il == n_layer - 1) { if (il == n_layer - 1) {
@ -13799,12 +13799,12 @@ struct llm_build_context {
cb(ffn_inp, "ffn_inp", il); cb(ffn_inp, "ffn_inp", il);
// feed-forward network // feed-forward network
if (!hparams.swin_norm) { if (!hparams.swin_norm) {
cur = llm_build_norm(ctx0, ffn_inp, hparams, cur = llm_build_norm(ctx0, ffn_inp, hparams,
model.layers[il].ffn_norm, NULL, model.layers[il].ffn_norm, NULL,
LLM_NORM_RMS, cb, il); LLM_NORM_RMS, cb, il);
cb(cur, "ffn_norm", il); cb(cur, "ffn_norm", il);
} }
cur = llm_build_ffn(ctx0, cur, cur = llm_build_ffn(ctx0, cur,
model.layers[il].ffn_up, NULL, NULL, model.layers[il].ffn_up, NULL, NULL,
@ -13814,12 +13814,12 @@ struct llm_build_context {
LLM_FFN_SILU, LLM_FFN_PAR, cb, il); LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
cb(cur, "ffn_out", il); cb(cur, "ffn_out", il);
if (hparams.swin_norm) { if (hparams.swin_norm) {
cur = llm_build_norm(ctx0, cur, hparams, cur = llm_build_norm(ctx0, cur, hparams,
model.layers[il].ffn_norm, NULL, model.layers[il].ffn_norm, NULL,
LLM_NORM_RMS, cb, il); LLM_NORM_RMS, cb, il);
cb(cur, "ffn_norm", il); cb(cur, "ffn_norm", il);
} }
cur = ggml_add(ctx0, cur, ffn_inp); cur = ggml_add(ctx0, cur, ffn_inp);
cb(cur, "ffn_out", il); cb(cur, "ffn_out", il);
@ -13842,13 +13842,15 @@ struct llm_build_context {
cur = ggml_mul_mat(ctx0, model.output, cur); cur = ggml_mul_mat(ctx0, model.output, cur);
cb(cur, "result_output_with_img_logits", -1); cb(cur, "result_output_with_img_logits", -1);
int img_token_end_idx = 8196; // TODO: this suppresses the output of image tokens, which is required to enable text-only outputs.
int img_token_start_idx = 4; // Needs to be removed once image outputs are supported.
int num_img_tokens = img_token_end_idx - img_token_start_idx; int img_token_end_idx = 8196;
struct ggml_tensor * img_logits = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, num_img_tokens); int img_token_start_idx = 4;
img_logits = ggml_add1(ctx0, img_logits, ggml_new_f32(ctx0, -FLT_MAX)); int num_img_tokens = img_token_end_idx - img_token_start_idx;
cb(img_logits, "img_logits", -1); struct ggml_tensor * img_logits = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, num_img_tokens);
cur = ggml_set_1d(ctx0, cur, img_logits, ggml_element_size(cur) * img_token_start_idx); img_logits = ggml_add1(ctx0, img_logits, ggml_new_f32(ctx0, -FLT_MAX));
cb(img_logits, "img_logits", -1);
cur = ggml_set_1d(ctx0, cur, img_logits, ggml_element_size(cur) * img_token_start_idx);
cb(cur, "result_output", -1); cb(cur, "result_output", -1);
ggml_build_forward_expand(gf, cur); ggml_build_forward_expand(gf, cur);
@ -15713,8 +15715,8 @@ struct llm_tokenizer_bpe {
break; break;
case LLAMA_VOCAB_PRE_TYPE_CHAMELEON: case LLAMA_VOCAB_PRE_TYPE_CHAMELEON:
regex_exprs = { regex_exprs = {
"<sentinel:[0-9]+>", // Sentinel tokens "<sentinel:[0-9]+>", // Sentinel tokens
"(IMGIMG)((A|B|C|D|E|F|G|H|I){1,4})Z", // Image tokens "(IMGIMG)((A|B|C|D|E|F|G|H|I){1,4})Z", // Image tokens
"([\t\n]| | )", // directly from tokenizer.json "([\t\n]| | )", // directly from tokenizer.json
"\\p{N}", // Individual digits "\\p{N}", // Individual digits
"[\\p{P}\\$\\+<=>\\^~\\|`]+", // Punctuation "[\\p{P}\\$\\+<=>\\^~\\|`]+", // Punctuation