Mirrors/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2025-01-30 13:53:03 +01:00
Code Issues Packages Projects Releases Wiki Activity

convert : pad vocab size to multiple of 64, not 8

Browse Source
This commit is contained in:
Jared Van Bortel 2024-02-12 16:47:00 -05:00
parent 48a7ef6ebc
commit b8ff85efe0
1 changed files with 1 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
convert-hf-to-gguf.py
View File

@ -1763,9 +1763,7 @@ class NomicBertModel(BertModel):
for name, data in super().get_tensors(): for name, data in super().get_tensors():
# Nomic Embed's token embeddings tensor is padded, but llama.cpp wants tensor sizes to match exactly. # Nomic Embed's token embeddings tensor is padded, but llama.cpp wants tensor sizes to match exactly.
if name == 'embeddings.word_embeddings.weight' and data.shape[1] != self.vocab_size: if name == 'embeddings.word_embeddings.weight' and data.shape[1] != self.vocab_size:
rounded_vocab_size = (self.vocab_size + 7) // 8 * 8 rounded_vocab_size = (self.vocab_size + 63) // 64 * 64
print(data.shape)
print(rounded_vocab_size, self.hparams["n_embd"])
assert data.shape == (rounded_vocab_size, self.hparams["n_embd"]) assert data.shape == (rounded_vocab_size, self.hparams["n_embd"])
data = data[:self.vocab_size, :] data = data[:self.vocab_size, :]
yield name, data yield name, data