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

Fixes to rebase

Browse source
This commit is contained in:
Igor Pissolati 2023-08-07 12:59:11 -03:00
parent f6d5fe3afc
commit 099119f532
2 changed files with 3 additions and 8 deletions
Download patch file Download diff file Expand all files Collapse all files

7
convert.py
View file

@ -1124,13 +1124,8 @@ class OutputFile:
params.n_mult,
params.n_head,
params.n_layer,
<<<<<<< HEAD
params.n_embd // params.n_head, # rot (obsolete)
file_type.value,
=======
params.n_vocab_base | 0xF0000000, # reuse obsolete rot value to store vocab_base
params.file_type.value,
>>>>>>> bfccc62 (Use some tricks to eliminate the necessity for a new format)
file_type.value,
]
self.fout.write(struct.pack("I" * len(values), *values))

4
llama.cpp
View file

@ -555,7 +555,7 @@ struct llama_file_loader {
// LLaMAv2
// TODO: read from header
hparams.n_head_kv = hparams.n_head;
=======
}
void read_vocab() {
vocab.id_to_token.resize(hparams.n_vocab);
@ -1442,7 +1442,7 @@ static struct ggml_cgraph * llama_build_graph(
const int64_t n_embd_head = hparams.n_embd_head();
const int64_t n_embd_gqa = hparams.n_embd_gqa();
LLAMA_ASSERT(n_embd_head == hparams.n_rot);
LLAMA_ASSERT(n_embd_head == hparams.n_embd/hparams.n_head);
const float freq_base = hparams.rope_freq_base;
const float freq_scale = hparams.rope_freq_scale;